Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Scientific Reports volume 16, Article number: 22267 (2026) Cite this article
471 Accesses
Metrics details
This study proposes a fault diagnosis method for grid-connected photovoltaic (PV) plants based on dual-channel spatiotemporal feature fusion. First, the Gramian Angular Field (GAF) transformation is employed to convert raw monitoring data into two-dimensional structured feature matrices with inherent temporal correlations, forming a two-dimensional feature processing branch. Meanwhile, the original one-dimensional current and voltage time series are retained as a numerical feature branch, establishing a parallel dual-channel processing architecture. A convolutional neural network (CNN) is used to extract local spatial patterns from the two-dimensional matrices, while a bidirectional long short-term memory network (BiLSTM) captures global temporal dependencies from the one-dimensional sequences, and a self-attention mechanism is introduced to dynamically weight key features. Through a bimodal feature complementarity mechanism, this method effectively overcomes the limitation of single-modality feature representation encountered by traditional approaches under complex operating conditions, and verifies its feasibility for engineering applications. Experimental results show that the method achieves an accuracy of 94.6875% in diagnosing seven typical fault types, demonstrating the effectiveness of the proposed approach in detecting and classifying various PV faults.
Amid the current global energy revolution, the ‘dual carbon’ strategic goals are reshaping the entire energy sector, spurring the rapid development of the renewable energy industry-characterized by low-carbon practices-and driving the green transition of society and the economy. According to data from the International Energy Agency (IEA), the global share of renewable energy in electricity generation has risen from 18.4% in 2000 to 32.1% in 2024,and is projected to reach 45.6% of total electricity generation by 2030. Solar and wind power account for almost all of the growth in new renewable energy capacity, and this significant growth trend has laid the foundation for the rapid development of the solar photovoltaic industry. Figure 1 illustrates the trend in annual new solar photovoltaic (PV) capacity additions worldwide from 2010 to 2025 (with 2025 figures being projections), clearly demonstrating the rapid growth of the PV industry.
Global Annual Solar PV Capacity Additions, 2010-2025. Data from IEA Renewable Energy Progress Tracker (IEA, 2025) under Creative Commons Attribution 4.0 International (CC BY 4.0) license. For full license terms and disclaimer, see Data availability section.
Photovoltaic power stations are typically situated in harsh external environments, including extreme weather conditions^1,2,3, prolonged exposure to ultraviolet radiation⁴, erosion and accumulation caused by wind and sand^5,6,7, and seawater corrosion⁸. These severe environmental conditions can trigger various faults across the entire system, thereby accelerating the ageing of photovoltaic modules, reducing the power output and overall efficiency of the power station, and consequently affecting its availability. Consequently, early fault detection and diagnosis are crucial for ensuring the power output and long-term reliable operation of the entire photovoltaic power station.
In recent years, a wide range of technical approaches have emerged in the field of photovoltaic module fault detection and diagnosis. Depending on the input data and detection principles, existing fault classification methods can be divided into two main categories: vision and thermal-based methods (VTM) and electrical-based methods (EBM)⁹. Among these, VTM performs fault diagnosis by analysing the visual appearance and thermal imaging characteristics of photovoltaic modules, such as panel damage, surface soiling, and colour abnormalities; whereas EBM primarily achieves fault detection by analysing changes in electrical parameters during the operation of the photovoltaic system, such as voltage, current, and power. EBM can be further categorised into five major groups: statistical and signal processing methods¹⁰, I–V characteristic analysis^11,12, power loss analysis¹³, voltage and current measurement¹⁴, and artificial intelligence techniques^15,16. Compared to VTM, EBM has garnered widespread attention in recent years, with its core advantages lying in its versatility, broad scope of application, and diverse range of potential inputs⁹. Statistical and signal processing methods include time-domain reflectometry, ground capacitance measurement and spread-spectrum time-domain reflectometry, as well as some more traditional statistical methods, such as the Kruskal-Wallis test (KWM), analysis of variance (ANOVA) and the exponentially weighted moving average (EWMA)^10,17. I–V characteristic analysis enables fault diagnosis by extracting the I–V curve characteristics of the PV system, such as through simulation modelling of the PV system to compare actual operational data with expected parameters^18,19,20. Power loss analysis methods achieve fault localisation by quantifying the extent of power loss in PV modules, strings, and arrays²¹. Voltage and current measurements involve collecting electrical parameters from DC output terminals at various levels, with fault determination based on abnormal fluctuations in the measured values²². Artificial intelligence (AI) technologies rely on various machine learning or deep learning models to perform fault detection and classification through feature extraction and pattern recognition of the collected data²³.
There is also a rapidly growing trend in the application of artificial intelligence technologies for photovoltaic fault classification²⁴. A wide variety of machine learning algorithms have been employed in photovoltaic fault detection and diagnosis; currently, the mainstream methods primarily include Recurrent Neural Networks (RNN), Support Vector Machines (SVM) and Random Forests (RF)^25,26. Among these, Zhong et al.²⁷ proposed a method based on time series analysis and SVM to effectively distinguish between fault types with similar electrical characteristics, such as random shading, fixed shading and ageing degradation, achieving an accuracy rate of 99.5% for fault feature classification. However, the simulation setup in this study utilised an LED light source to simulate solar irradiance, which differs from actual conditions. Madeti et al.²⁸ proposed a k-Nearest Neighbors (kNN) algorithm capable of detecting and classifying open-circuit faults, line-to-line faults and bypass diode faults, achieving an average fault classification accuracy of 98.70%. Ghaedi et al.²⁹ proposes a method based on K-Means clustering and ensemble learning vector quantisation (LVQ) to effectively identify faults in photovoltaic systems under complex conditions, such as high impedance and low mismatch. It achieves an average accuracy of 99.26%; however, the use of static maximum power points alone may result in the failure to detect dynamic faults. Eskandari et al.³⁰ proposes an ensemble learning model comprising SVM, Naive Bayes (NB) and kNN, achieving an accuracy of 99.5%; however, this method is limited to line-to-line faults in PV systems. Chen et al.³¹ proposes a RF-based method for classifying partial shading, degradation, open-circuit and short-circuit faults, utilising only the PV array voltage and the current of each PV string. This study yields good results, but is based solely on weather conditions operating within a limited range. Amiri et al.³² utilises a Random Forest Classifier (RFC) to detect and monitor the performance of PV systems, but the model is not comprehensive. Liu et al.³³ proposed a method based on the Gaussian kernel fuzzy C-means clustering algorithm to distinguish between eight common single or composite faults, but did not fully account for dynamic characteristics. Furthermore, methods such as Decision Trees^34,35, Neuro-Fuzzy Classifiers³⁵ and Extremely Randomised Trees³⁶ have also demonstrated application potential. Experiments indicate that machine learning methods can achieve high-precision differentiation between faults with similar I–V curves, thereby avoiding the misclassifications associated with traditional algorithms³⁷. Although these machine learning methods are effective, they also have drawbacks, particularly concerning large datasets that can lead to overfitting³⁸. In addition, machine learning methods exhibit limitations in representing the features of complex high-dimensional data³⁹, and there is still room for improvement in measuring contributions during the dynamic selection of classifier ensembles⁴⁰.
The emergence of deep learning represents a transformative leap in machine learning, attracting widespread attention due to its comprehensive advantages in pattern recognition and data mining. Common deep learning networks include CNN^41,42, deep transfer learning methods⁴³, and Long Short-Term Memory (LSTM) networks^44,45. Qiao et al.⁴⁶ proposed a photovoltaic module fault diagnosis method based on multi-angle feature expansion and GAF visual image deep learning (MA-GCT), which achieves efficient and accurate detection and classification of multiple PV faults by enhancing multi-domain features of voltage, current, and power signals, mapping them to structured two-dimensional images via GAF, and constructing a collaborative CNN-Transformer network; the fault diagnosis accuracy exceeds 99% in both simulation and experimental tests. Hong et al.⁴⁷ earlier combined stacked GAF images with a three-dimensional convolutional neural network, fully validating the effectiveness of two-dimensional GAF representations for PV array fault detection; however, this method only utilizes the image channel and does not integrate the dynamic details of the original signals. Guo et al.⁴⁸ proposed a sub-module open-circuit fault diagnosis method for PV DC collection systems based on CNN-LSTM, in which the CNN automatically extracts local features from capacitor voltage signals and the LSTM captures temporal dependencies to obtain global features, enabling real-time detection and localization of sub-module switching device open-circuit faults under complex conditions with strong robustness to irradiance fluctuations. Attention mechanisms have drawn increasing interest in PV fault diagnosis due to their ability to adaptively focus on critical information. The self-attention architecture proposed by Vaswani et al.⁴⁹ laid an important theoretical foundation for subsequent fault identification methods based on Transformers and attention modules. Reference⁵⁰ introduces a convolutional block attention module combined with a CNN to improve fault diagnosis accuracy for PV arrays with different blocking diode configurations under various operating conditions. Reference⁵¹ presents a method for detecting micro-crack anomalies in PV module cells by designing an attention classification and segmentation network, achieving efficient and accurate anomaly detection. Reference⁵² proposes a fault diagnosis method based on a clustering algorithm and an LSTM network, demonstrating excellent diagnostic performance under different weather conditions. Reference⁵³ designs a method based on Gated Recurrent Units (GRU), using meteorological satellite and inverter measurements as inputs to a stacked GRU model to achieve high-accuracy fault classification and severity estimation. Reference⁵⁴ proposes a method based on a parallel CNN and Bidirectional Gated Recurrent Unit (BiGRU), optimizing fault identification by leveraging the complementary strengths of recurrent and convolutional layers and achieving an accuracy exceeding 99%. Zheng et al.⁵⁵ construct a multi-source domain adaptation network through fine-grained feature decoupling and diversity regularization, achieving excellent generalization performance in cross-domain diagnosis and providing a reference for the problem of insufficient model generalization under varying operating conditions. Zheng et al.⁵⁶ propose a cross-machine fault diagnosis method based on ConvFormer and biconditional domain adaptation, which extracts global–local features in parallel through a linear Transformer and a separable shuffled CNN, and enhances feature transferability using biconditional constraints from both machine domain and fault category information, thereby achieving accurate fault diagnosis. As the above literature shows, current research mainly focuses on improving model generalization performance and enhancing the representation capability of dynamic characteristics; a reliable and accurate PV fault diagnosis model is crucial for fault detection and classification.
To address the challenge of compound fault identification in photovoltaic (PV) systems under non-stationary operating conditions such as partial shading, non-uniform aging, and dynamic weather, researchers have begun to introduce deep learning architectures that integrate multi-source features into this field. Li et al.⁵⁷ constructed a CNN-BiLSTM-Attention based model for PV array fault diagnosis, where the CNN extracts local spatial features, the BiLSTM captures long-term temporal dependencies, and the attention mechanism weights critical features; the Black-Winged Kite Algorithm (BKA) is introduced to automatically tune hyperparameters, achieving an accuracy of 99.11 That work validates the effectiveness of such hybrid deep learning architectures in PV fault identification, yet its feature input remains limited to one-dimensional electrical parameter sequences, underutilizing the implicit two-dimensional spatiotemporal correlation structure within the data. Mo et al.⁵⁸ proposed a lightweight dual-track 1D-2D feature fusion convolutional network (1D_2DIFCNN) for machinery equipment fault diagnosis. This method constructs a parallel dual-track structure: the 1D track takes the raw vibration signal as input and extracts temporal local features using a 1D-CNN; the 2D track converts the signal into a time-frequency image via the continuous wavelet transform (CWT) and then extracts time-frequency distribution features through a 2D-CNN, while incorporating the convolutional block attention mechanism (CBAM) and random overlapping sampling (ROST) to enhance feature focusing and data diversity. The method achieved excellent diagnostic performance on bearing datasets, demonstrating the effectiveness of dual-track heterogeneous feature fusion. However, this method can still be further improved: the 2D branch employs CWT, which preserves temporal correlation structures relatively indirectly; the 1D branch relies solely on CNN, making it difficult to explicitly capture global long-term dependencies; and the dual-track fusion strategy is relatively straightforward, lacking dynamic exploration of cross-track spatiotemporal relationships.
Unlike the dual-channel CNN parallel architecture proposed by Mo et al., this study presents a fault diagnosis method for grid-connected photovoltaic systems based on deep fusion of dual-channel spatiotemporal features. The core improvement lies in replacing the continuous wavelet transform (CWT) with the Gramian Angular Field (GAF) transformation to convert raw one-dimensional monitoring data into two-dimensional structured images that preserve temporal correlations, thereby retaining the fault evolution process more directly. In the one-dimensional branch, a bidirectional long short-term memory network (BiLSTM) and a hybrid self-attention mechanism are introduced to establish global long-range dependencies of the sequence and dynamically relate critical information at different positions. These components are then combined to achieve deep cross-fusion of dual-channel spatiotemporal features. Specifically, the GAF transformation is first applied to construct a two-dimensional feature processing channel, while the raw one-dimensional time-series data, such as current and voltage, are retained to form a numerical feature channel, resulting in a dual-channel parallel processing structure. A convolutional neural network (CNN) extracts local spatial patterns from the two-dimensional matrices, and a BiLSTM captures global temporal dependencies from the one-dimensional sequences, with the hybrid self-attention mechanism relating different positions within the sequence. On this basis, the outputs of the one-dimensional and two-dimensional channels are concatenated and fused, enabling deep spatiotemporal feature integration. Finally, the effectiveness and engineering feasibility of the proposed method are validated using actual operational data from a grid-connected photovoltaic system. The main contributions of this study are as follows:
To address the problem of single feature representation and the difficulty in simultaneously capturing local spatial patterns and global temporal dependencies under complex operating conditions, an asymmetric dual-track parallel processing structure is proposed. Unlike the dual-track 1D-CNN and 2D-CNN fusion of Mo et al., the 1D track in this study directly models the raw current and voltage sequences using BiLSTM, avoiding the smoothing compression of transient fluctuations caused by convolution and thereby preserving the complete dynamic characteristics of the signals; the 2D track employs GAF transformation to convert the 1D sequence into a 2D image and introduces a 2D-CNN to extract spatial correlation patterns. These two heterogeneous paths achieve feature complementarity and enhancement from the dimensions of transient dynamics in the signal domain and spatial textures in the image domain, rather than simple convolutional fusion.
To tackle the issue that a single model struggles to simultaneously mine spatial features and model long-range temporal dependencies, a heterogeneous collaborative analysis architecture integrating 2D-CNN and BiLSTM is constructed. The 2D-CNN fully extracts local spatial patterns from the Gramian Angular Difference Field (GADF) images, while the BiLSTM performs bidirectional long-period modeling of the original 1D sequence. The features from both tracks are mapped to feature spaces of the same dimension and then fused, compelling both sides to learn discriminative representations that complement each other. This design integrates the stable discriminative capability of GADF textures with the complete dynamic information of the original waveform, achieving effective complementarity between texture and dynamic features with reasonable parameter overhead under moderate model complexity, thus avoiding the loss of transient information caused by premature convolutional compression and leading to significantly improved diagnostic accuracy.
In response to the problem that the importance of features varies greatly across different fault types and that critical transient features can easily be submerged by redundant information, a multi-head self-attention mechanism is introduced after the BiLSTM in the 1D time-series track to dynamically weight the features of each time step output. This mechanism adaptively computes attention weights, enabling the model to automatically focus on the critical transient points where faults occur, reinforce the contribution of discriminative features, and simultaneously suppress the interference of irrelevant fluctuations and background noise. Compared with existing CNN-BiLSTM-Attention methods that apply attention only to a single sequence track, this paper enhances attention on the 1D dynamic branch within a dual-track fusion framework, further improving the sensitivity to weak faults and diagnostic robustness under complex operating conditions.
Experimental validation is conducted using actual operational data of grid-connected PV systems covering seven typical PV faults and normal states. The proposed method achieves a fault diagnosis accuracy of 94.6875%, and under the condition of not using a 1D-CNN, excellent results are obtained solely with BiLSTM and the self-attention mechanism. The results demonstrate that this method has advantages in overcoming the problem of single feature representation and achieving high-accuracy diagnosis with a reasonable and effective dual-track architecture, providing a new technical pathway for intelligent operation and maintenance of grid-connected PV systems.
Overall, the innovation of this study is reflected in the fusion of GAF-based 2D representations with the original 1D sequence through an asymmetric heterogeneous architecture; the 1D branch, with its CNN-free BiLSTM modeling, fully preserves the transient characteristics of PV electrical quantities, distinguishing it from existing parallel dual-track CNN fusion schemes. On this basis, the multi-head self-attention mechanism is organically embedded into the dual-track fusion framework, achieving enhanced focus on critical fault moments. This method achieves competitive diagnostic performance with lower model complexity and offers certain reference value for engineering applications.
The structure of this paper is organized as follows: Theory and methodology section systematically explains the core theory and research methodology; Experiments and results section presents the experimental implementation and key result analysis in detail; Concluding remarks and future perspectives section concludes the research findings and discusses future research directions.
The technical difficulties in fault diagnosis for photovoltaic power stations primarily stem from three aspects: the subtlety of fault features, the complexity of temporal evolution patterns, and the coupling effects of multi-source interferences. Traditional analysis methods based on single-modality time-series data struggle to effectively balance the extraction of local spatial textures and the capture of transient dynamic anomalies under sensor noise and non-stationary operating conditions. To address these issues, this study proposes an asymmetric dual-track spatiotemporal feature fusion diagnostic framework. First, the GAF is employed to transform raw one-dimensional monitoring signals into two-dimensional feature images that preserve temporal correlations, thereby constructing a spatial feature track. Simultaneously, the original one-dimensional current and voltage waveforms are retained to form a temporal feature track. On the two-dimensional track, a CNN is utilized to extract local spatial patterns from the GAF images. On the one-dimensional track, a BiLSTM network is directly applied to model the raw sequences, avoiding the smoothing compression of transient fluctuations caused by convolution operations, while a multi-head self-attention mechanism is introduced to adaptively focus on critical fault moments. After additive fusion of the two heterogeneous feature sets, the steady-state textures from the image domain and the transient dynamics from the signal domain are integrated to construct a highly robust multi-dimensional feature decision system. Experimental results demonstrate that the proposed scheme achieves significant improvements in both classification accuracy and generalization performance.
The Gram’s corner array is capable of converting one-dimensional time-series signals into two-dimensional image data, thereby maximising the preservation of the original signal’s characteristics and preventing an increase in false alarm rates caused by feature loss. Taking the current data collected from a photovoltaic array as an example, the data processing procedure is as follows:
Scale the data from the photovoltaic array using the mapminmax function to facilitate subsequent processing.
Let the time series of photovoltaic array current collected from the photovoltaic power station system be scaled to (I = { {i_1},{i_2},{i_3}, ldots ,{i_n}}), where n is the total number of sampling points for the photovoltaic array current. The obtained photovoltaic array operating parameters are scaled to the interval [−1, 1] using the following method, in accordance with Eq. (1) below.
In the above equation, (widetilde{{{i}_{x}}}) is the scaled value of the sampling point indexed by x under sample I, where ({{i}_{x}}) denotes the sampling point, (max [I]) is the maximum sampling value in sample I, and (min [I]) is the minimum sampling value in sample I. The normalised scaled data obtained after processing is (widetilde{I}={widetilde{{{i}_{1}}},widetilde{{{i}_{2}}},widetilde{{{i}_{3}}},ldots ,widetilde{{{i}_{n}}}}).
Next, the normalised data (widetilde{I} = { widetilde{{i_1}},widetilde{{i_2}},widetilde{{i_3}}, ldots ,widetilde{{i_n}}}) is transformed into polar coordinates using Eq. (2) to convert the time-series current signal into a vector.
In the above equation, ({t_x}) represents the timestamp, (widetilde{{i_x}}) represents the reconstructed time series, N represents the spatial constant factor generated by the normalised polar coordinate system, R represents the resulting polar coordinate radius, and (theta) represents the resulting polar coordinate angle.
The entire encoding process is bijective; that is, the polar coordinate result obtained after encoding a given time series is unique. Furthermore, compared to the Cartesian coordinate system, the polar coordinate system preserves absolute event correlation and can encompass different angular boundaries, providing varying levels of information granularity for the conversion of one-dimensional time series into two-dimensional image information.
The definition of the Gram matrix is shown in Eq. (3) below. The polar coordinates obtained via the aforementioned polar coordinate transformation are organised into a set of vectors based on the time series; the correlation between vectors is characterised by the inner product between corresponding vectors, and the degree of correlation is represented by the angle between the vectors.
In the above equation, G denotes the Gram matrix, and (< {x_1},{x_n}>) denotes the inner product operation between ({x_1}) and ({x_n}).
Next, employing a trigonometric encoding method in conjunction with the Gramian angular field algorithm, the one-dimensional time series data is transformed into the following two types of two-dimensional image feature representations: the Gramian Angular Summation Field (GASF), based on the cosine function, and the Gramian Angular Difference Field (GADF), based on the sine function. The transformation formulas are as follows:
Each time-series segment selected in this paper contains 13 sampling points; therefore, the resolution of the two-dimensional image generated by the GAF transformation is (13 times 13) pixels. The determination of this resolution does not arise from hyperparameter tuning of the image size but is directly dictated by the length of the input time series. Under the definition of the Gramian Angular Field, the image size (N times N) is determined by the number of sampled data points N, and each pixel (G_{i,j}) encodes the angular or angular-difference relationship between time point i and time point j, completely preserving the pairwise correlation information among all data points in the original sequence.
The choice of the (13times 13) resolution results from a trade-off between information fidelity and computational efficiency in the context of photovoltaic fault diagnosis. On the one hand, 13 sampling points are sufficient to cover the critical feature window of transient fluctuations in electrical parameters under typical fault conditions of a PV system, enabling a complete characterization of both local spatial patterns and global structural information in the GAF texture, thereby satisfying the requirements for fault identification at the level of information fidelity. On the other hand, compared to common image resolutions such as (64times 64) or (224times 224), the extremely small size of (13times 13) significantly reduces the input dimensionality of the subsequent 2D-CNN, leading to a substantial decrease in both the computational load and memory consumption of the convolution operations in the two-dimensional track. Together with the one-dimensional BiLSTM track, it constitutes the overall architecture.
A CNN is a feedforward neural network specifically designed to process grid-like data, such as images and videos, and is one of the classic algorithms in the field of deep learning. A CNN typically consists of components such as convolutional layers, pooling layers, activation functions and fully connected layers. It possesses several characteristics, including receptive fields, weight sharing, resolution and network depth, enabling efficient feature learning from input data.
The convolutional layer captures spatial correlation features through a local receptive field mechanism, and its computational process follows Eq. (6), where (beta) represents the feature map output by this convolutional layer; (sigma) is the nonlinear activation function; ({alpha _{i – r,j – c}}) denotes the input data at row ((i – r)) and column ((j – c)), with r and c indicating the center position of the convolution kernel, which depends on the kernel size and convolution stride, typically being odd numbers; ({w_{i,j}}) represents the weight at row i and column j of the convolution kernel, which is a learnable parameter; b is the bias term; and n represents the dimensionality of the input data. In this operation, the convolution kernel traverses the input feature map in a sliding window manner, encoding it into a compact hidden representation through a parameter-sharing mechanism. This design significantly reduces the complexity of the model and enables the network to adapt to the positions of local features.
The pooling layer achieves feature space compression via Eq. (7).
Here, ({beta _{i,j,k}})is the output feature map; ({alpha _{i + p,j + q,k}}) is the input feature map at the row indexed by (i + p), the column indexed by (j + q), and the channel indexed by k; p and q are the coordinate strides within the pooling window. While preserving key information, this operation progressively reduces the dimensions of the feature map, thereby lowering computational resource consumption and mitigating overfitting.
The fully connected layer serves as the classification decision module, and its formulation is expressed in Eq. (8).
Here, ({x_i}) is the input feature vector indexed by i; ({y_0}) is the output feature map; n is the dimensionality of the input features; ({omega _i}) is the weight corresponding to each input vector, which is a learnable parameter; and f is the logistic activation function.
While reducing the number of parameters, the convolutional neural network enhances processing efficiency and generalization capability for two-dimensional image data through local perception and weight-sharing mechanisms. Its structure is illustrated in the following Fig. 2.
Convolutional Nerual Network.
The LSTM unit consists of an input gate, a forget gate, an output gate, and a cell state, with its main formulas as follows:
In Eq. (9), (sigma) is the sigmoid activation function; ({W_f}) is the weight; ({b_f}) is the bias. ({C_{t – 1}}) denotes the cell state, which preserves the memory information from the previous time step, and ({f_t}) represents the extent to which the forget gate discards the cell state. New information to be added to the cell state is determined through the input gate, and the cell state is subsequently updated.
In the above equations, ({i_t}) represents the degree to which the input gate updates new information; (widetilde{{C_t}}) is the candidate vector for the current new state information. Next, the cell state is updated, where ({f_t} cdot {C_{t – 1}}) denotes the information to be forgotten, and ({i_t} cdot widetilde{{C_t}}) denotes the information to be retained; ({C_t}) represents the current cell state. The input information is stored in the memory unit, and finally, information is output through the output gate.
In the above equations, ({o_t}) represents the information output from the output gate, and ({h_t}) is the output of the hidden gate. Meanwhile, the output ({h_t}) is also fed into the backward LSTM layer.
Here, the sigmoid function is expressed as
and the tanh function is expressed as
LSTM neural networks are commonly used to address issues such as vanishing and exploding gradients that arise during long-sequence training. In contrast, a BiLSTM network feeds sequences in both forward and reverse directions into two independent networks, allowing it to capture complete contextual information at each time step and fully leverage both past and future information. The raw signal is passed from the input layer to the BiLSTM network layer, where the forward LSTM and backward LSTM each generate a value, which collectively determine the value transmitted to the hidden units. The formulas for the BiLSTM are as follows:
In the above equations, (overrightarrow{{h_t}}) and (overleftarrow{{h_t}}) represent the outputs of the output gates of the forward LSTM and the backward LSTM, respectively; ({overrightarrow{h} _{t – 1}}) and ({overleftarrow{h} _{t – 1}}) denote the total output values of the forward LSTM and the backward LSTM, respectively; ({y_t}) is the hidden layer output of the BiLSTM network; and f is the activation function of different layers. The schematic diagram of the BiLSTM network is shown in Fig. 3.
Bidirectional Long Short-Term Memory.
In the above figure, ({x_{t – 1}}), ({x_t}), and ({x_{t + 1}}) represent the input vectors of the time series at time step (t – 1) (previous moment), t (current moment), and (t + 1) (next moment), respectively; ({y_{t – 1}}), ({y_t}), and ({y_{t + 1}}) denote the contextual representations finally output by the BiLSTM network, which are composed of the forward and backward hidden states.
Multi-Head Self-Attention is essentially the execution of multiple Self-Attention computations, which enables the model to capture features from different representational subspaces across various levels and to obtain richer contextual information from the sentence. Its structural diagram is shown in the following Fig. 4.
Multi-head Self-Attention.
Let the output matrix of the Bi-LSTM layer be (Y in mathbb {R}^{T times d}), where (T) is the sequence length (number of time steps) and (d) is the number of hidden units of the Bi-LSTM (in this code, (d = 256)). The single self-attention computation employs a multiplicative attention mechanism (scaled dot-product attention). Multi-head attention uses (h = 4) parallel heads, each with query and key dimension (d_k = 4) and value dimension (d_v = d/h = 64). For the (i)-th head ((i = 1, 2, 3, 4)), independent learnable weight matrices (W_i^Q, W_i^K in mathbb {R}^{d times d_k}) and (W_i^V in mathbb {R}^{d times d_v}) are used to linearly project the input (Y) to obtain the query matrix (Q_i), key matrix (K_i), and value matrix (V_i) as (Q_i = Y W_i^Q in mathbb {R}^{T times d_k},; K_i = Y W_i^K in mathbb {R}^{T times d_k},; V_i = Y W_i^V in mathbb {R}^{T times d_v}).
The scaled dot-product attention is then computed as
where (A_i) is the attention weight matrix representing the attention strength of each time step towards all time steps in this head. The context vector of a single head is obtained as
The output feature of this head after self-attention is
Multi-head self-attention performs the above process in parallel for (h = 4) heads. The outputs of all heads are concatenated along the feature dimension, yielding a tensor of dimension (T times h d_v = T times d), which is then fused by a learnable output projection matrix (W^O in mathbb {R}^{d times d} = mathbb {R}^{256 times 256}) to obtain the multi-head self-attention representation:
The role of (W^O) is to project the concatenated 256-dimensional features back to the original (d = 256) space, completing the integration of multi-head information. Each row of the output (y_s) still corresponds to the enhanced feature of a time step, and is subsequently passed through a Dropout layer and a fully connected layer to finally produce a 20-dimensional vector for fusion with the 2D branch.
Fault diagnosis of photovoltaic plants faces three major challenges: weak fault features, complex temporal evolution, and coupled multi-source interference. A single-modality method struggles to stably extract local spatial textures while effectively capturing transient dynamic anomalies. To this end, this paper adopts a complementary fusion strategy and integrates the strengths of each module into an asymmetric dual-channel spatiotemporal feature diagnosis framework.
In the two-dimensional channel, the GAF converts the one-dimensional time series into a two-dimensional image through a one-to-one mapping, fully preserving the pairwise correlations between all time points. The amplitude and phase relationships of the original signal are transformed into local textures and global structures in the image, providing the convolutional network with spatial feature inputs that possess clear physical meaning. Subsequently, a CNN leverages local receptive fields and the weight-sharing mechanism to efficiently extract steady-state spatial patterns from the GAF images, effectively suppressing sensor noise and local distortions and enhancing the robustness of spatial features. However, convolution and pooling operations in the image domain may cause some smoothing of transient fluctuations, thereby reducing sensitivity to rapidly changing anomalies.
To compensate for this limitation, the one-dimensional channel directly retains the raw current and voltage waveforms and employs a bidirectional long short-term memory network (BiLSTM) to model long-range dependencies of the sequences from both forward and backward directions. The BiLSTM can fully capture the contextual information before and after fault evolution, completely preserving transient features and avoiding the loss of detail due to spatial aggregation. On this basis, a multi-head self-attention mechanism is introduced to enable the model to adaptively focus on the time steps that are most critical for fault discrimination, thereby strengthening the representation of weak anomalies and suppressing the coupling effects of multi-source interference. Based on the above design, the overall network framework proposed in this paper is shown in the following Fig. 5.
Schematic of the network framework.
Steps 1 and 2 involve processing data from the photovoltaic array and the Gram-Angle field transformation to divide the data into training and test sets; Steps 3 and 4 involve iterative training by feeding the training set into the 1D-2D-CNN-BiLSTM diagnostic model, and testing the model’s actual performance once convergence is achieved. If the test results fall short of expectations, the model is retrained by adjusting the hyperparameters until the training meets the expected criteria. The trained model is subsequently integrated into the pre-designed software for real-time diagnosis. The workflow of the method described in this paper is illustrated in Fig. 6.
Flowchart for the Classification and Diagnosis of Photovoltaic System Faults.
Based on the laboratory fault experimental data from a photovoltaic microgrid application, the data samples for this paper are constructed⁵⁹. The classification of real injected faults in this Grid-connected Photovoltaic (GPV) system is detailed in Table 1. Early-stage faults in PV systems, which may arise from corrosion, cell degradation, or damage to interconnection parts, are often not severe and can typically be prevented through regular preventive maintenance, thereby avoiding failures in the PV power plant. This dataset does not account for degradation faults, as such faults require long-term data with large sampling intervals for support⁵⁹.
In this grid-connected photovoltaic system, the output of the PV array is generated by the programmable Chroma 62150H-1000S solar array simulator, which allows for the varying effects of environmental conditions (G and T). The programmable AC power supply Chroma 61511 is used as the grid simulator. The control algorithm is implemented in the Dspace 1104 environment, which is also used for data acquisition. Voltage-oriented control technology is combined with space vector pulse width modulation to control active and reactive power based on grid-side signals. The output voltage is synchronised with the grid voltage via a phase-locked loop. The AC load in this work serves for protection purposes whilst simulating actual faults⁵⁹. For more details on the control structure, energy management system, communication, and configuration of this system, see Reference⁶⁰.
To accurately evaluate the performance of the proposed model in this study, this paper assesses the model using metrics such as Accuracy, Precision, Recall, and F1-score. The classification accuracy for each fault type is obtained from the resulting confusion matrix, which further informs the subsequent improvement of the model.
Taking binary classification as an example, the formulas for Accuracy, Precision, and Recall are defined as follows:
Since Precision and Recall are difficult to improve simultaneously, a balance point is sought to reconcile the performance of these two metrics. This balance point is the F1-score, and its calculation formula is as follows:
In the above formulas, TP (True Positive) denotes the number of samples correctly predicted as positive by the model, TN (True Negative) denotes the number of samples correctly predicted as negative, FP (False Positive) denotes the number of samples incorrectly predicted as positive, and FN (False Negative) denotes the number of samples incorrectly predicted as negative. In this study, the final macro-precision and macro-recall are computed as the mean of precision and recall values across all classes, and these are adopted as the ultimate evaluation metrics.
Based on Eqs. (4) and (5), the Gramian Angular Summation Field (GASF) and Gramian Angular Difference Field (GADF) heatmaps are obtained, as shown in Figs. 7 and  8, respectively.
Gram Point and Field Thermal Map.
Gram angle difference field thermogram.
The aforementioned heatmaps enable a more intuitive structural representation of temporal information, facilitating the identification of anomalous points and helping to understand the key regions on which the model focuses. GASF can reflect global trends and coordinated variations, making it suitable for detecting steady-state characteristics, whereas GADF captures local dynamic changes and instantaneous differences, making it more appropriate for capturing transient faults.
To verify the effectiveness of the proposed feature representation method, the sample data are randomly split into a training set and a test set at a ratio of 8:2, with no sample overlap between them. The model is trained using the Adam optimizer, with an initial learning rate of 0.005, a mini-batch size of 128, and a total of 350 epochs. The detailed network architecture of the proposed model is shown in Table 2.
To evaluate the classification performance of this method, this study employs t-SNE dimensionality reduction to perform three-dimensional visualization analysis on the raw data, the feature representations after Gramian angular field transformation, and the model prediction results. As shown in Figs. 9, 10, and 11, the visualized comparison through three-dimensional scatter plots illustrates the distribution changes of data in the feature space at different processing stages: (a) the raw data distribution exhibits nonlinear separable characteristics, with blurred boundaries between different classes and significant noise interference; (b) the feature space after GAF transformation does not completely eliminate the mixing phenomenon between classes, but shows noticeable intra-class aggregation compared to the raw data; (c) the final prediction results demonstrate good inter-class separation in the three-dimensional projection. This multi-level visualization method effectively validates the effectiveness of GAF feature transformation on the feature structure, while intuitively revealing the model’s feature learning process from raw data to classification decisions.
Original feature space.
GAF feature space.
Deep feature space (Prediction results).
To comprehensively validate the effectiveness of the proposed model, we designed comparative experiments that include single-branch ablation and different fusion structures. The two-dimensional GAF-CNN branch (hereinafter referred to as CNN), the one-dimensional BiLSTM branch, and various combinations of these components with attention mechanisms and parallel structures are selected for comparative analysis. Specifically, the evaluated configurations are: CNN, CNN-Attention, 1D BiLSTM, 1D BiLSTM-Attention, Parallel CNN, Parallel CNN-Attention, Parallel CNN-BiLSTM, and four variants derived from the proposed method: This Method-A, in which the multi-head self-attention is replaced by single-head self-attention; This Method-B, in which the BiLSTM in the one-dimensional branch is replaced by a 1D CNN; This Method-C, in which self-attention is applied only to the two-dimensional branch while the one-dimensional branch does not introduce attention; and This Method-D, in which the attention mechanism is moved after the feature fusion stage. The classification accuracy, macro precision, macro recall, F1-score, number of learnable parameters, and training time of each model on the test set are presented in Table 3.
As shown in Table 3, the proposed method achieves 94.6875% accuracy, 94.36% macro precision, 94.23% macro recall, and a macro F1-score of 0.9413, all four metrics outperforming those of all compared models. When the multi-head self-attention is replaced by single-head self-attention (This Method-A), the accuracy drops to 92.3125% and the F1-score to 0.9222, with noticeable performance degradation across all fault categories, especially for F3 (grid voltage sag) and F7 (boost converter controller fault) (Table 4), verifying the necessity of the multi-head design in capturing multi-aspect discriminative features. Further analysis on the placement of attention reveals that moving the attention mechanism after feature fusion (This Method-D) causes the accuracy to drop sharply to 89.0625% and the F1-score to only 0.8893, indicating that applying attention independently within each branch during the feature extraction stage before fusion enables each branch to focus on the critical information of its own modality, thereby preventing the decline in discriminability caused by information mixing after fusion. When self-attention is applied only to the two-dimensional branch and not introduced into the one-dimensional branch (This Method-C), the accuracy is 89.6875% and the F1-score is 0.8998, also substantially lower than the proposed method, demonstrating that attention in the one-dimensional branch is indispensable for enhancing the representation of temporal features. Moreover, replacing the BiLSTM in the one-dimensional branch with a 1D CNN (This Method-B) reduces the accuracy to 88.375% and the F1-score to 0.8834, proving the irreplaceable role of BiLSTM in modeling sequential dependencies and preserving transient dynamic information. In terms of model complexity, the proposed method has 709.1k learnable parameters, slightly higher than the 623.2k of 1D BiLSTM-Attention, yet it yields a 5.375 percentage point improvement in accuracy and a 4.75% gain in F1-score over the latter. Regarding training time, the proposed method consumes 826.18 seconds, the longest among all models. However, considering that PV fault diagnosis models are typically deployed in an offline-training, online-inference manner, the extra training time is acceptable in engineering practice. During inference, all models achieve a single inference time on the order of 1 ms on a personal laptop equipped with 32 GB memory and an NVIDIA RTX 4060 Laptop GPU, with negligible differences, fully meeting the response requirements for online real-time monitoring. The above ablation experiments consistently demonstrate that the multi-head self-attention mechanism, the independent placement of attention in each branch during the feature extraction stage, and the one-dimensional BiLSTM sequence modeling module are all key prerequisites for the excellent performance of the proposed model. The confusion matrices corresponding to the training of each model are shown in Figs. 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22 and 23. The proposed method achieves better diagnostic performance than the other comparative networks for each fault category.
Confusion matrix of This Method on the test set.
Confusion matrix of CNN on the test set.
Confusion matrix of CNN-Attention on the test set.
Confusion matrix of 1D BiLSTM on the test set.
Confusion matrix of 1D BiLSTM-Attention on the test set.
Confusion matrix of Parallel CNN on the test set.
Confusion matrix of Parallel CNN-Attention on the test set.
Confusion matrix of Parallel CNN-BiLSTM on the test set.
Confusion matrix of This Method-A on the test set.
Confusion matrix of This Method-B on the test set.
Confusion matrix of This Method-C on the test set.
Confusion matrix of This Method-D on the test set.
The precision, recall, and F1-score are all obtained by macro-averaging the results across each class. As shown in Fig. 24, after 350 training epochs, the model has reached a converged state, illustrating the changes in the loss curves during the actual training process of each network.
Training loss comparison curve.
Training accuracy comparison curve.
The accuracy comparison curves of each model during the training phase are shown in Fig. 25. After 350 epochs of training, the proposed method achieves a training accuracy of 97.65% and a final test accuracy of 94.6875%, ranking first among all models. This result verifies the good generalization performance of the method.
To further analyze the recognition capability of each model for different fault types, Table 4 lists the precision, recall, and F1-score of each model for every fault category. These results are computed from the confusion matrices of the models on the test set, and all data in the table represent the specific metric values corresponding to each category.
As shown in Table 4, the proposed method achieves the best precision, recall, and F1-score on the vast majority of fault categories. In particular, it demonstrates outstanding performance on feedback sensor fault (F2) and PV array partial shading mismatch (F4), with F1-scores of 0.9976 and 0.9899, respectively, significantly outperforming all compared models. On MPPT controller fault (F6), the F1-score of the proposed method is 0.9625, slightly lower than those of 1D BiLSTM (0.9927) and 1D BiLSTM-Attention (0.9880), yet still superior to most of the compared models. Looking at the per-category performance of the ablation variants, This Method-A (the single-head self-attention variant) yields an F1-score of 0.9690 on F6, which falls between the full model and 1D BiLSTM, indicating that the adjustment of the number of attention heads has a certain impact on the discrimination of specific faults. The metrics of This Method-D (post-fusion attention variant) and This Method-C (2D-only attention variant) are weaker than those of the full model on most categories; notably, their F1-scores on challenging categories such as Normal and F3 fall below 0.83. For This Method-B (the 1D CNN variant), the recall on F3 is only 67.44% and on F5 is 81.82%, showing noticeably inferior capability in identifying temporally sensitive faults compared to the BiLSTM-based versions. It is worth noting that, on the PV array partial shading mismatch (F4) category, the 1D BiLSTM model suffers from significantly low values across all three metrics (F1-score of only 0.7973) due to its lack of spatial feature extraction capability. In contrast, the proposed method introduces two-dimensional spatial texture features via GAF transformation and realizes heterogeneous complementarity with the one-dimensional BiLSTM channel, raising the F1-score for this category to 0.9899, which fully validates the effectiveness of the asymmetric dual-channel architecture in fusing multi-dimensional fault features. Furthermore, the comparison with the parallel CNN series models demonstrates that the design of the proposed method, which abandons convolution operations in the one-dimensional channel and directly adopts a combination of BiLSTM and multi-head self-attention, effectively preserves the transient dynamic information of the original signal, leading to marked improvements in the recognition accuracy of temporally sensitive fault categories such as PV array open-circuit mismatch (F5) and boost converter controller fault (F7) (F1-scores of 0.9802 and 0.9390, respectively). The above experimental results convincingly prove that the proposed model possesses superior capability in fault feature extraction and classification.
This study proposes a fault diagnosis method based on a dual-channel input architecture for grid-connected PV systems to address abnormalities in PV array current and voltage parameters caused by faults during operation. The method integrates CNN, BiLSTM networks, and a self-attention mechanism. To fully leverage the advantages of CNNs in two-dimensional data processing, the GAF transformation is employed to convert the grid-connected PV system data into a two-dimensional image format. The network architecture adopts a dual-channel parallel processing mode: the two-dimensional channel converts time-series signals into two-dimensional polar coordinate images using the GAF transformation, while the one-dimensional channel retains the original feature vectors. Through feature concatenation, the original data information is deeply fused with abstract image features, thereby significantly enhancing the model’s generalization capability.
On this basis, a BiLSTM network is introduced to capture the forward and backward dependencies in the original feature sequences, leveraging both historical and future information to improve classification performance. Additionally, a self-attention mechanism is incorporated to dynamically capture key features, effectively enhancing the extraction capability of data sequence features under complex operating conditions. Experimental results show that the proposed method achieves an accuracy of 94.6875% in multi-class fault diagnosis for grid-connected PV systems, representing a significant improvement over the best-performing comparison method. This effectively reduces system downtime and maintenance costs, thereby improving operational efficiency.
It should be noted that although the proposed method achieves a significant improvement in single-fault diagnosis accuracy, several limitations remain. First, the current comparative experiments mainly focus on internal variants of the proposed architecture and have not yet included recent advanced models such as pure Transformers. This is not because such methods lack competitiveness, but rather because their structural paradigms differ substantially from the asymmetric dual-channel heterogeneous fusion architecture in this study. A forced comparison would require substantial modifications to the main network; the resulting structural changes would make it difficult to attribute performance differences to the core design concept, preventing a fair comparison. Second, the individual contributions of different design choices (e.g., replacing GAF with CWT) to the final performance have not yet been separately quantified. The reported performance gains are the result of the synergistic effects of multiple design choices. Furthermore, when multiple concurrent faults occur in the system, the overlapping of fault features still affects the model’s discrimination accuracy. To address these limitations, future research will focus on constructing a structurally compatible unified evaluation benchmark, systematically quantifying the individual contribution of each key module, and exploring multi-task learning or feature decoupling strategies to further enhance the model’s adaptability under multiple concurrent faults and dynamic environments.
The experimental data files that support the analyses in this study are available at http://dx.doi.org/10.17632/n76t439f65.1. The present study builds upon this experimental dataset for further analysis. Additionally, the data used to prepare Figure 1 were obtained from the IEA Renewable Energy Progress Tracker (IEA, 2025). These data are publicly accessible under a Creative Commons Attribution 4.0 International (CC BY 4.0) license, with additional terms as specified by the IEA at https://www.iea.org/terms/creative-commons-cc-licenses. OECD/IEA retains ownership of the data. This work is derived from IEA material by the authors, who take sole responsibility for this derivative work. The derived work is not endorsed by the IEA or its Member countries in any manner. The data were accessed on 22 March 2026 via https://www.iea.org/data-and-statistics/data-tools/renewable-energy-progress-tracker.
Okonkwo, P. C. et al. Solar pv systems under weather extremes: Case studies, classification, vulnerability assessment, and adaptation pathways. Energy Rep. 13, 929–959. https://doi.org/10.1016/j.egyr.2024.12.067 (2025).
Article  Google Scholar 
Bošnjaković, M., Stojkov, M., Katinić, M. & Lacković, I. Effects of extreme weather conditions on pv systems. Sustainability 15, 16044. https://doi.org/10.3390/su152216044 (2023).
Article  ADS  CAS  Google Scholar 
Jackson, N. D. & Gunda, T. Evaluation of extreme weather impacts on utility-scale photovoltaic plant performance in the United States. Appl. Energy 302, 117508. https://doi.org/10.1016/j.apenergy.2021.117508 (2021).
Article  Google Scholar 
Lee, H. W., Biswas, S., Choi, H., Lee, Y. & Kim, H. Influence of ultraviolet radiation on the performance of zinc oxide electron transport layer-based inverted organic photovoltaic devices. ACS Appl. Electron. Mater. 7, 95–103. https://doi.org/10.1021/acsaelm.4c01567 (2025).
Article  CAS  Google Scholar 
Huang, M. et al. Quantifying the effects of dust characteristics on the performance of radiative cooling pv systems. Appl. Energy 377, 124672. https://doi.org/10.1016/j.apenergy.2024.124672 (2025).
Article  Google Scholar 
Zhang, K. et al. Numerical simulation study on the impact of wind-blown sand action on the loading of photovoltaic systems. Phys. Fluids https://doi.org/10.1063/5.0216302 (2024).
Article  Google Scholar 
Sahouane, N. et al. Technical and economic study of the sand and dust accumulation impact on the energy performance of photovoltaic system in algerian sahara. Renew. Energy 205, 142–155. https://doi.org/10.1016/j.renene.2023.01.033 (2023).
Article  Google Scholar 
Djalab, A. et al. A comprehensive review of floating photovoltaic systems: Tech advances, marine environmental influences on offshore pv systems, and economic feasibility analysis. Sol. Energy 277, 112711. https://doi.org/10.1016/j.solener.2024.112711 (2024).
Article  Google Scholar 
Hong, Y.-Y. & Pula, R. A. Methods of photovoltaic fault detection and classification: A review. Energy Rep. 8, 5898–5929. https://doi.org/10.1016/j.egyr.2022.04.043 (2022).
Article  Google Scholar 
Zhang, Z. et al. A fault diagnosis method for photovoltaic module current mismatch based on numerical analysis and statistics. Sol. Energy 225, 221–236. https://doi.org/10.1016/j.solener.2021.07.037 (2021).
Article  ADS  Google Scholar 
Karmakar, B. K. & Pradhan, A. K. Detection and classification of faults in solar pv array using thevenin equivalent resistance. IEEE J. Photovoltaics 10, 644–654. https://doi.org/10.1109/JPHOTOV.2019.2959951 (2020).
Article  Google Scholar 
Fadhel, S. et al. Pv shading fault detection and classification based on i–v curve using principal component analysis: Application to isolated pv system. Sol. Energy 179, 1–10. https://doi.org/10.1016/j.solener.2018.12.048 (2019).
Article  ADS  Google Scholar 
Livera, A., Theristis, M., Micheli, L., Stein, J. S. & Georghiou, G. E. Failure diagnosis and trend-based performance losses routines for the detection and classification of incidents in large-scale photovoltaic systems. Progr. Photovoltaics 30, 921–937. https://doi.org/10.1002/pip.3578 (2022).
Article  Google Scholar 
Mansouri, M., Trabelsi, M., Nounou, H. & Nounou, M. Deep learning-based fault diagnosis of photovoltaic systems: A comprehensive review and enhancement prospects. IEEE Access 9, 126286–126306. https://doi.org/10.1109/ACCESS.2021.3110947 (2021).
Article  Google Scholar 
Hao, J., Zhan, H., Xiao, C., Pei, H. & Wang, L. A method of operation status evaluation and fault diagnosis for pv arrays in the scenario of digitalization. Energy Rep. 12, 3020–3033. https://doi.org/10.1016/j.egyr.2024.08.071 (2024).
Article  Google Scholar 
Mellit, A. & Kalogirou, S. Artificial intelligence and internet of things to improve efficacy of diagnosis and remote sensing of solar photovoltaic systems: Challenges, recommendations and future directions. Renew. Sustain. Energy Rev. 143, 110889. https://doi.org/10.1016/j.rser.2021.110889 (2021).
Article  Google Scholar 
Saha, S. et al. Diagnosis and mitigation of voltage and current sensors malfunctioning in a grid connected pv system. Int. J. Electr. Power Energy Syst. 115, 105381. https://doi.org/10.1016/j.ijepes.2019.105381 (2020).
Article  Google Scholar 
Chen, X. et al. Fault prediagnosis, type identification, and degree diagnosis method of the photovoltaic array based on the current–voltage conversion. IEEE Trans. Power Electron. 39, 16708–16719. https://doi.org/10.1109/TPEL.2024.3427649 (2024).
Article  ADS  Google Scholar 
Navid, Q., Hassan, A., Fardoun, A. A., Ramzan, R. & Alraeesi, A. Fault diagnostic methodologies for utility-scale photovoltaic power plants: A state of the art review. Sustainability 13, 1629. https://doi.org/10.3390/su13041629 (2021).
Article  ADS  CAS  Google Scholar 
Chen, Z., Chen, Y., Wu, L., Cheng, S. & Lin, P. Deep residual network based fault detection and diagnosis of photovoltaic arrays using current-voltage curves and ambient conditions. Energy Convers. Manag. 198, 111793. https://doi.org/10.1016/j.enconman.2019.111793 (2019).
Article  Google Scholar 
Satpathy, P. R., Aljafari, B., Thanikanti, S. B. & Madeti, S. R. K. Electrical fault tolerance of photovoltaic array configurations: Experimental investigation, performance analysis, monitoring and detection. Renew. Energy 206, 960–981. https://doi.org/10.1016/j.renene.2023.02.103 (2023).
Article  Google Scholar 
Liu, Y. et al. Intelligent fault diagnosis of photovoltaic array based on variable predictive models and i–v curves. Sol. Energy 237, 340–351. https://doi.org/10.1016/j.solener.2022.03.062 (2022).
Article  ADS  Google Scholar 
Utama, C., Meske, C., Schneider, J., Schlatmann, R. & Ulbrich, C. Explainable artificial intelligence for photovoltaic fault detection: A comparison of instruments. Sol. Energy 249, 139–151. https://doi.org/10.1016/j.solener.2022.11.018 (2023).
Article  ADS  Google Scholar 
Gaviria, J. F., Narváez, G., Guillen, C., Giraldo, L. F. & Bressan, M. Machine learning in photovoltaic systems: A review. Renew. Energy 196, 298–318. https://doi.org/10.1016/j.renene.2022.06.105 (2022).
Article  Google Scholar 
Ren, B. et al. Machine learning applications in health monitoring of renewable energy systems. Renew. Sustain. Energy Rev. 189, 114039. https://doi.org/10.1016/j.rser.2023.114039 (2024).
Article  Google Scholar 
Chang, Z. & Han, T. Prognostics and health management of photovoltaic systems based on deep learning: A state-of-the-art review and future perspectives. Renew. Sustain. Energy Rev. 205, 114861. https://doi.org/10.1016/j.rser.2024.114861 (2024).
Article  Google Scholar 
Zhong, Y., Zhang, B., Ji, X. & Wu, J. Fault diagnosis of pv array based on time series and support vector machine. Int. J. Photoenergy 1–10, 2024. https://doi.org/10.1155/2024/2885545 (2024).
Article  Google Scholar 
Madeti, S. R. & Singh, S. N. Modeling of pv system based on experimental data for fault detection using knn method. Sol. Energy 173, 139–151. https://doi.org/10.1016/j.solener.2018.07.038 (2018).
Article  ADS  Google Scholar 
Ghaedi, P. et al. Ensemble lvq model for photovoltaic line-to-line fault diagnosis using k-means clustering and adagrad. Energies 17, 5269. https://doi.org/10.3390/en17215269 (2024).
Article  Google Scholar 
Eskandari, A., Milimonfared, J. & Aghaei, M. Line-line fault detection and classification for photovoltaic systems using ensemble learning model based on i–v characteristics. Sol. Energy 211, 354–365. https://doi.org/10.1016/j.solener.2020.09.071 (2020).
Article  ADS  Google Scholar 
Chen, Z. et al. Random forest based intelligent fault diagnosis for pv arrays using array voltage and string currents. Energy Convers. Manag. 178, 250–264. https://doi.org/10.1016/j.enconman.2018.10.040 (2018).
Article  ADS  Google Scholar 
Amiri, A. F., Oudira, H., Chouder, A. & Kichou, S. Faults detection and diagnosis of pv systems based on machine learning approach using random forest classifier. Energy Convers. Manag. 301, 118076. https://doi.org/10.1016/j.enconman.2024.118076 (2024).
Article  Google Scholar 
Liu, S., Dong, L., Liao, X., Cao, X. & Wang, X. Photovoltaic array fault diagnosis based on gaussian kernel fuzzy c-means clustering algorithm. Sensors 19, 1520. https://doi.org/10.3390/s19071520 (2019).
Article  ADS  PubMed  PubMed Central  Google Scholar 
Li, B., Delpha, C., Migan-Dubois, A. & Diallo, D. Fault diagnosis of photovoltaic panels using full i–v characteristics and machine learning techniques. Energy Convers. Manag. 248, 114785. https://doi.org/10.1016/j.enconman.2021.114785 (2021).
Article  Google Scholar 
Lin, C.-J., Lin, C.-H. & Lin, F. Bearing fault diagnosis using a vector-based convolutional fuzzy neural network. Appl. Sci. 13, 3337. https://doi.org/10.3390/app13053337 (2023).
Article  CAS  Google Scholar 
Noura, H. N., Allal, Z., Salman, O. & Chahine, K. Explainable artificial intelligence of tree-based algorithms for fault detection and diagnosis in grid-connected photovoltaic systems. Eng. Appl. Artif. Intell. 139, 109503. https://doi.org/10.1016/j.engappai.2024.109503 (2025).
Article  Google Scholar 
Mellit, A. & Kalogirou, S. Assessment of machine learning and ensemble methods for fault diagnosis of photovoltaic systems. Renew. Energy 184, 1074–1090. https://doi.org/10.1016/j.renene.2021.11.125 (2022).
Article  Google Scholar 
Zhu, Y.-H., Zhang, Z.-L. & Liu, W.-A. Iterative under-sampling for delisting risk prediction of chinese listed companies based on financial information. Eng. Appl. Artif. Intell. 153, 110830. https://doi.org/10.1016/j.engappai.2025.110830 (2025).
Article  Google Scholar 
El-Banby, G. M., Moawad, N. M., Abouzalm, B. A., Abouzaid, W. F. & Ramadan, E. A. Photovoltaic system fault detection techniques: A review. Neural Comput. Appl. 35, 24829–24842. https://doi.org/10.1007/s00521-023-09041-7 (2023).
Article  Google Scholar 
Zhang, Z.-L., Zhu, Y.-H. & Luo, X.-G. Des-as: Dynamic ensemble selection based on algorithm shapley. Pattern Recogn. 157, 110899. https://doi.org/10.1016/j.patcog.2024.110899 (2025).
Article  Google Scholar 
Chen, Z., Gryllias, K. & Li, W. Mechanical fault diagnosis using convolutional neural networks and extreme learning machine. Mech. Syst. Signal Process. 133, 106272. https://doi.org/10.1016/j.ymssp.2019.106272 (2019).
Article  Google Scholar 
Ali, Y. M., Ding, L. & Qin, S. An efficient approach for diagnosing faults in photovoltaic array using 1d-cnn and feature selection techniques. Int. J. Electr. Power Energy Syst. 166, 110526. https://doi.org/10.1016/j.ijepes.2025.110526 (2025).
Article  Google Scholar 
Shao, S., McAleer, S., Yan, R. & Baldi, P. Highly accurate machine fault diagnosis using deep transfer learning. IEEE Trans. Ind. Inf. 15, 2446–2455. https://doi.org/10.1109/TII.2018.2864759 (2019).
Article  Google Scholar 
He, Z. et al. Shading type and severity diagnosis in photovoltaic systems via i–v curve imaging and two-stream deep neural network. Energy Convers. Manag. 324, 119311. https://doi.org/10.1016/j.enconman.2024.119311 (2025).
Article  Google Scholar 
Hadroug, N. et al. Practical implementation based on histogram of oriented gradient descriptor combined with deep learning: Towards intelligent monitoring of a photovoltaic power plant with robust faults predictions. J. Ind. Inf. Integr. 44, 100760. https://doi.org/10.1016/j.jii.2024.100760 (2025).
Article  Google Scholar 
Liu, Q. et al. An efficient fault diagnosis method combining multi-angle feature expansion and visual image neural networks for solar photovoltaic modules. Energy 333, 137484. https://doi.org/10.1016/j.energy.2025.137484 (2025).
Article  Google Scholar 
Hong, Y.-Y. & Pula, R. A. Detection and classification of faults in photovoltaic arrays using a 3d convolutional neural network. Energy 246, 123391. https://doi.org/10.1016/j.energy.2022.123391 (2022).
Article  Google Scholar 
Guo, K., Lu, Z., Liu, P. & Mo, Z. Fault diagnosis method for sub-module open-circuit faults in photovoltaic dc collection systems based on cnn-lstm. Electronics 14, 1205. https://doi.org/10.3390/electronics14061205 (2025).
Article  Google Scholar 
Vaswani, A. et al. Attention is all you need. arXiv preprint End of data reached while scanning argument https://doi.org/10.48550/arxiv.1706.03762 (2017)
Qu, J., Sun, Q., Qian, Z., Wei, L. & Zareipour, H. Fault diagnosis for pv arrays considering dust impact based on transformed graphical features of characteristic curves and convolutional neural network with cbam modules. Appl. Energy 355, 122252. https://doi.org/10.1016/j.apenergy.2023.122252 (2024).
Article  Google Scholar 
Jiang, Y. & Zhao, C. Attention classification-and-segmentation network for micro-crack anomaly detection of photovoltaic module cells. Sol. Energy 238, 291–304. https://doi.org/10.1016/j.solener.2022.04.012 (2022).
Article  ADS  Google Scholar 
Guo, H., Hu, S., Wang, F. & Zhang, L. A novel method for quantitative fault diagnosis of photovoltaic systems based on data-driven. Electric Power Syst. Res. 210, 108121. https://doi.org/10.1016/j.epsr.2022.108121 (2022).
Article  Google Scholar 
Van Gompel, J., Spina, D. & Develder, C. Satellite based fault diagnosis of photovoltaic systems using recurrent neural networks. Appl. Energy 305, 117874. https://doi.org/10.1016/j.apenergy.2021.117874 (2022).
Article  Google Scholar 
Amiri, A. F., Kichou, S., Oudira, H., Chouder, A. & Silvestre, S. Fault detection and diagnosis of a photovoltaic system based on deep learning using the combination of a convolutional neural network (cnn) and bidirectional gated recurrent unit (bi-gru). Sustainability 16, 1012. https://doi.org/10.3390/su16031012 (2024).
Article  ADS  Google Scholar 
Zheng, X. et al. A fine-grained feature decoupling based multi-source domain adaptation network for rotating machinery fault diagnosis. Reliab. Eng. Syst. Saf. 243, 109892. https://doi.org/10.1016/j.ress.2023.109892 (2024).
Article  Google Scholar 
Zheng, X., Nie, J., He, Z. & Gao, M. Exploring informative and highly-transferable features for cross-machine fault diagnosis by convformer-based biconditional domain adaptation method. IEEE Trans. Ind. Inf. 21, 3107–3116. https://doi.org/10.1109/TII.2024.3523549 (2025).
Article  Google Scholar 
Geng, L., Yang, Y., Yang, G. & Hu, J. Photovoltaic array fault diagnosis based on cnn-bilstm-attention optimized by black-winged kite algorithm. 537–543 (IEEE, 2025).
Mo, C., Huang, K. & Ji, H. A lightweight and precision dual track 1d and 2d feature fusion convolutional network for machinery equipment fault diagnosis. Sci. Rep. 14, 31666. https://doi.org/10.1038/s41598-024-81118-2 (2024).
Article  ADS  PubMed  PubMed Central  Google Scholar 
Bakdi, A., Bounoua, W., Guichi, A. & Mekhilef, S. Real-time fault detection in pv systems under mppt using pmu and high-frequency multi-sensor data through online pca-kde-based multivariate kl divergence. Int. J. Electr. Power Energy Syst. 125, 106457. https://doi.org/10.1016/j.ijepes.2020.106457 (2021).
Article  Google Scholar 
Guichi, A., Talha, A., Berkouk, E. M., Mekhilef, S. & Gassab, S. A new method for intermediate power point tracking for pv generator under partially shaded conditions in hybrid system. Sol. Energy 170, 974–987. https://doi.org/10.1016/j.solener.2018.06.027 (2018).
Article  ADS  Google Scholar 
Download references
Correspondence and requests for materials should be addressed to Z.L.
This work was supported by the key research and development program of Xinjiang Uygur Autonomous Region under Grant (No. 2024B04002-1).
School of Intelligence Science and Technology, Xinjiang University, Urumqi, 830047, China
Peikun Deng
School of Electrical Engineering, Xinjiang University, Urumqi, 830047, China
Zhenen Li & Zhengxing Huang
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Z.L. conceived the core idea, supervised the project, and was responsible for project administration and funding acquisition. Z.H. developed the methodology, designed the experiments, and participated in the validation. P.D. performed the formal analysis, investigation, data curation, wrote the original draft, and prepared the visualizations. Z.L., Z.H., and P.D. participated in the validation. Z.L. and Z.H. reviewed and edited the manuscript.
Correspondence to Zhenen Li.
The authors declare no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
Deng, P., Li, Z. & Huang, Z. Research on a fault diagnosis method for photovoltaic power plants based on a dual-channel 1D-2D-CNN-BiLSTM spatiotemporal feature fusion network. Sci Rep 16, 22267 (2026). https://doi.org/10.1038/s41598-026-60430-z
Download citation
Received: 30 March 2026
Accepted: 29 June 2026
Published: 15 July 2026
Version of record: 15 July 2026
DOI: https://doi.org/10.1038/s41598-026-60430-z
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative
Advertisement
Scientific Reports (Sci Rep)
ISSN 2045-2322 (online)
© 2026 Springer Nature Limited
Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.

source