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 15, Article number: 16371 (2025)
2621
1
5
Metrics details
In this paper, six different types of solar PV technologies are compared in terms of their performances under tropical conditions, using three years of performance data from a 1.2 MW experimental solar farm. The technologies considered include single-crystalline silicon, polycrystalline silicon, microcrystalline silicon, amorphous silicon, copper indium selenium (CIS), and hetero-junction with intrinsic thin layer (HIT). The field performances of these cells were initially assessed using standard performance indices such as Array Yield, Reference Yield, Capture Loss, Performance Ratio, and Efficiency Ratio. Among the technologies studied, amorphous silicon and HIT-based systems demonstrated better performance, showing higher Performance and Efficiency Ratios, along with lower capture losses. This study also modelled the fluctuations in power production from these panels. Under probabilistic modeling, the ramping behavior of the systems was characterized using the Generalized Logistic Distribution. Based on this analysis, CIS PV systems were found to have minimum power ramps, where as the HIT based systems showed the highest power fluctuations. To predict minute-wise and hourly ramping of the PV systems under varying levels of solar insolation, machine learning methods based on Artificial Neural Networks (ANN), Support Vector Machines (SVM), and k-Nearest Neighbors (kNN) were developed. With a Normalized Root Mean Square Error (NRMSE) of over 96%, these models demonstrated high accuracy in capturing the ramping characteristics of the studied PV systems. The results of this study offer valuable insights into the performance of different PV systems under tropical regions, which can be used in efficiently designing and managing solar PV projects.
Owing to its tremendous resource potential, a significant increase in system efficiencies and a constant decline in the system cost, the global installed capacity of solar PV systems has significantly increased in recent years. For example, with an added capacity of 446 GW during the year, the global solar PV installations have reached around 1581 GW by 2023, which is sufficient for 5.5% of global energy consumption1,2. The future prediction indicates that, by 2050, the cumulative installed capacity around the world will exceed 8519 GW which contribute up to 25% of the global energy consumption3. Most of these installations are planned to be grid-connected. With this increasing global presence, solar PV technologies have been considered as one of the major contributors to the future clean energy scenario.
There are various types of solar PV technologies, each differing in materials and structural features. The efficiencies of these systems also vary significantly4, which in turn impacts on the cost of generation. Hence, understanding the comparative performance of different PV technologies under the operating environment of a prospective project is crucial for planning and developing a successful solar PV program.
Similarly, one of the major challenges that the power community may face with large PV installations is the frequent fluctuations in the power produced by solar PV plants. Due to the stochastic nature of solar irradiance, power output from the solar plants would vary significantly with time. This may pose significant challenges to power system operators and planners as sufficient reserves, with quick ramping-up capabilities, are to be made readily available to handle these power fluctuations. Hence, a better understanding of these performance variations of the systems and the corresponding power ramps is essential for the efficient design and management of the PV-integrated power systems.
In recent years, several studies have been conducted to understand the short-term output fluctuations of solar PV systems5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21. Most of these studies adopt statistical approaches under which the Probability Density Function (PDF) and Cumulative Distribution Function (CDF) of the power variability are used to define the solar power fluctuation13,15,18,20,21 In another attempt, transfer functions based on Discrete Fourier Transform (DFT) and Fast Fourier Transform (FFT) algorithms are used for predicting the power output fluctuations9,17. Machine learning methods like Artificial Neural Networks (ANN), Convolution Neural Networks (CNN), Support Vector Machine (SVM), k-nearest neighbor (kNN), and Random Forest (RF) are also being tried in a few studies for analyzing solar power variability6,7,8,10,11,14,16,19,22.
Several studies have investigated the performance of various photovoltaic (PV) technologies in different climatic conditions23,24,25,26,27. A study conducted in Hungary reported that polycrystalline silicon (pc-Si) performed well in that region based on a dataset collected over a two-year period with a 10-min resolution24. Similarly, based on the performance analysis of five different PV technologies in a humid tropical climate in Ghana, it is concluded that HIT-based systems were the most suitable for space-constrained environments, whereas CIS technology was the least effective. Their study also found that polycrystalline silicon was the most suitable PV technology, followed by amorphous silicon (a-Si) and HIT27. In another study in Morocco’s hot semi-arid climate, it was found that monocrystalline silicon (m-Si) was the most efficient, although polycrystalline silicon (p-Si) was more cost-effective, making it the preferred choice for the region25.
Further, simulation tools such as PVsyst, PVWATTS, TRNSYS, HOMER, SolarPro, were used for predicting the performance of a 1 MW PV power plant under India’s arid climatic conditions23. It is reported that CIS modules yielded the highest annual energy output of 2471 MWh, with a performance ratio (PR) of 86.1%, while mono-Si exhibited the highest annual temperature-related energy loss (14.4%). In addition, evaluation of PV performance under similar outdoor conditions in India reported that a-Si outperformed p-Si during summer due to its thermal annealing effect but underperformed in winter28. Meanwhile, HIT consistently showed better performance than p-Si throughout the year. A comparative study conducted at Nanjing, China, assessed multiple PV technologies and identified HIT as the best-performing option, while CdTe and p-Si were the least suitable26.
Overall, these studies indicate that performances of PV technologies are highly dependent on the climatic conditions under they operate. Further, though these investigations could provide valuable information about the variability of solar PV systems under changing insolation levels at the study location, most of these studies consider only one specific type of PV technology. Similarly, several of these studies rely on smaller, stand-alone systems with a limited capacity of a few kilowatts. Furthermore, many of these studies rely on data collected during short durations and hence do not fully capture the long-term performance characteristics.
The specific contributions of this paper are as follows: (1) A comprehensive analysis and comparison of the performance and ramping behavior of six different solar PV systems operating at the same site and under identical environmental conditions. (2) Three years of data are utilized for analysis and model development to minimize annual variations and model errors. (3) Data from a 1.2 MW experimental PV power plant is used. The larger plant size, and consequently the higher data volume, enhances the statistical significance of the analysis compared to studies based on smaller systems. This approach also helps average out potential noise in the datasets. Additionally, the larger plant size aids in gaining a better understanding of the systems’ interactions with the grid.
Following the introductory section and a description of the experimental solar PV plant, the results of this study are presented in two parts. In the first part, the performance of six different solar PV technologies is quantified and compared in a representative tropical environment using standard performance metrics. In the second part, the analysis is extended to model the ramping behavior of these systems. This includes the probabilistic modeling of minute-by-minute and hourly ramping, as well as predictive modeling of power ramps using machine learning methods. Such power ramp models can assist system managers in optimizing power dispatch strategies. Furthermore, when these models are integrated with forecasts of solar insolation, they can be useful for plant owners and utilities participating in energy markets.
The solar farm selected for this analysis is the Tenaga Suria Brunei (TSB) power plant situated in Seria, Belait District, Brunei Darussalam. The plant location has a latitude of 4°37′ N and a longitude of 114°23′ E. The configuration of the solar photovoltaic (PV) power plant and the data monitoring system are shown in Fig. 122,29.
Features of the experimental solar PV power plant.
The TSB power plant is equipped with six different types solar PV systems, all of which have a rated capacity of 200kWp. The PV types are (1) single crystalline silicon (sc-Si), (2) polycrystalline silicon (mc-Si), (3) microcrystalline silicon (nc-Si/a-Si), (4) amorphous silicon (a-Si), (5) copper indium selenium (CIS), and (6) hetero-junction with intrinsic thin layer (HIT). The modules’ technical specifications are provided in Table 122. As evident, different panels have different conversion efficiencies and hence, the number of modules and the required area for a given rated power are different.
The standard indices like the Array Yield, Reference Yield, Capture Loss, Performance Ratio, and Efficiency Ratio are used to quantify the performance of the PV systems considered in this study. These are briefly discussed below.
The array yield represents how effectively the PV array harvests solar energy and converts it into electrical energy. It measures the actual energy output of the PV array relative to its rated power and is expressed as a ratio calculated on a daily, monthly, or yearly basis. By comparing the actual energy output to the rated power, the array yield provides insights into the system’s performance and efficiency. Understanding the array yield helps to evaluate the performance and economic viability of PV systems, aiding in decision-making regarding system design, installation, and optimization.
Array Yield (YA) is calculated using the following equation.
where EDC is the total DC energy generated by the PV array during the specified period, measured in kilowatt-hours (kWh) and P0 is the rated capacity of the PV system represented in kilowatts peak (kWp). The monthly array yields vary across different PV module technologies and months are presented in Fig. 2. The array yields of these systems vary throughout the year, indicating seasonal variations in energy output. The CIS-PV module system had the lowest array yield of 3.12 h per day in December, while the amorphous silicon PV module system had the highest array yield of 4.81 h per day in March shown in Fig. 2.
Average daily energy output of various photovoltaic (PV) modules monthly.
The reference yield (YR) compares daily in-plane radiation received by PV panels to a predefined reference value to assess PV system performance. It is calculated by comparing the total daily in-plane radiation (Hd) to the reference radiation (G). By a standardized reference level, the reference yield measures the capacity of PV panels to capture and transform solar radiation into electrical energy. It measures the PV system’s sunlight utilization compared to the ideal 1 kW/m2 solar radiation.
The reference yield (YR) equation is below.
Here, Hd is the daily in-plane radiation incident on PV panels in kWh/m2 and G is the standardized reference value of solar radiation received in the PV panel plane in kW/m2.
Capture loss measures efficiency losses or deviations from ideal performance in a PV system by comparing the reference yield (YR) to the array yield (YA). As shown below, it is the difference between the PV panels’ reference radiation-based expected energy output and the system’s actual energy output.
Capture losses of different PV systems are shown in Fig. 3. The lowest capture losses are 18.06% for heterojunction with intrinsic thin layer (HIT) modules followed by 18.22% for amorphous silicon (a-Si) modules. Single-crystalline silicon (sc-Si) modules have the highest capture losses (27.15%). Various factors contribute to capture loss, which diminishes the overall efficiency of the PV array in capturing and converting energy. The difference between expected and actual energy generation could be due to shading, soiling, system degradation, or suboptimal system design. Higher cell temperatures cause thermal capture losses, thereby decreasing cell efficiency.
Annual averaged daily array yield and capture losses of different PV modules.
By comparing array yield (YA) to reference yield (YR), the performance ratio (PR) measures photovoltaic (PV) system efficiency. It shows how well the system converts solar energy into electricity. The performance ratio considers shading, soiling, module degradation, and non-ideal conditions to assess system efficiency. It is calculated mathematically as
A lower performance ratio indicates system losses. Amorphous silicon and HIT PV had the highest monthly performance ratios as shown on Fig. 4. Performance ratios were highest in months with lower ambient temperatures. Indicating the temperature sensitivity of the cells.
Monthly performance ratios of different PV modules.
The efficiency ratio (ηR) is a key factor in determining the field performance of a PV system. It reflects the system’s ability to convert array in-plane irradiation into DC energy output under real-world conditions. The efficiency ratio, which compares energy output (EDC) to in-plane irradiation (H) incident on the array, normalized by efficiency at standard test conditions (ηSTC), offer valuable insights into system performance. The efficiency ratio (ηR) is calculated as:
Module efficiency, shading, soiling, and temperature losses are mainly reflected in the efficiency ratio. Figure 5 provides a comparison of the monthly averaged daily efficiency ratios of different PV modules, which indicate how well the PV modules can perform in practical scenarios. From the figure, it is evident that the field performance of CIS modules closely aligns with their rated performance throughout the year. These modules exhibit an annual average efficiency ratio of 0.83, indicating a strong correlation between laboratory and real-world performance. Following this, the amorphous silicon (a-Si) modules could achieve an average efficiency ratio of 0.81, closely followed by HIT modules with a ratio of 0.8. On the other hand, single crystalline silicon (sc-Si) cells exhibited the lowest performance ratio of 0.69.
Comparison of the monthly average daily efficiency ratios of different PV modules.
Figure 6 displays the hourly changes in solar insolation levels and the efficiency ratio of cells on an annual basis. Like crystalline and amorphous silicon, second-generation modules attain their maximum efficiency ratios earlier when exposed to lower levels of insolation. Amorphous silicon and HIT based systems consistently maintain higher efficiency ratios, even in conditions of lower insolation, surpassing other technologies. CIS modules exhibit commendable efficiency ratios, especially in situations with lower levels of solar radiation. The second-generation thin film technologies show relatively better efficiency across a wider range of solar insolation. This adaptability is highly advantageous in regions such as Brunei Darussalam, where frequent cloud cover resulting in substantial variations in solar insolation is prevalent.
Hourly variations in solar insolation levels and the efficiency ratio.
To model the ramping behavior of different solar PV systems, power output and environmental parameters like solar insolation, ambient temperature, and wind speed were collected from the experimental solar farm. These data, which are recorded every second, are then averaged to minute and hourly data to examine the ramp rates. The data was cleaned to remove the outliers and noises. Out of the three years of data, two years were used for model training and development, followed by testing with the remaining one-year dataset.
Various weather parameters like ambient temperature, irradiance and wind speed could influence the power production from the PV systems. From the input data considered for this study, the relative influence of these parameters was further analyzed by Gradient Boosted Machine (GBM) algorithm30. The results showed that irradiance, with a high correlation of 99.9%, is the major parameter influencing the PV system output, compared to the wind speed (0.07%) and ambient temperature (0.04%). This agrees with the observations in31. Data between 8 A.M and 6 P.M have been considered for developing the models as these are the productive hours for the PV systems at the site. The production data are normalized with the watt peak capacity of each panel. The fluctuations in the power output of the systems are defined as
where ΔPo (t) is the change in power output at an instant time t, Po represents the power output, Δt is the sampling time interval, and PR is the total capacity of each panel type (200 kWp).
The ramp rates of all the PV systems were computed using Eq. 6, which were then fitted with different probability models. The most appropriate distribution, fitting well with the data, were identified using Kolmogorov–Smirnov (K-S) of Goodness of Fit (GOF) test32,33,34. The Kolmogorov–Smirnov (K-S) test assesses the goodness of fit between observed and computed data by comparing the maximum difference in their cumulative distributions. A K-S statistic closer to zero indicates a better fit. Based on the K–S statistic presented in the Table 2, the ramp-up and ramp-down rate of minutes and hourly power fluctuation for all six panels are well described by Generalized Logistic Distribution.
The probability density function f(x) and Cumulative distribution function F(x) of Generalized Logistic Distribution are given by
and
where (z = ,,,frac{x – mu }{sigma }), k is continuous shape parameter, (sigma) is a continuous scale parameter ((sigma) > 0) ,(mu) is a continuous location parameter, and x is the changes in power output.
Table 3 presents the shape, scale, and location parameters for the continuous distributions of both minute and hourly power fluctuations. The continuous shape parameter k can give some indication on the ramp rate of the panels. Higher k values indicate sharper changes in power, indicating higher ramps. Substituting these values into Eqs. 7 and 8 yields the probability density and cumulative distribution of power ramps. Figure 7 shows these modeled probabilities for minute-wise ramp-up and ramp-down of the PV systems.
The probabilities for ramp up and ramp down of the panels (a) single crystalline (b) poly crystalline (c) micro crystalline (d) amorphous (e) CIS (f) HIT.
The 95% confidence interval for the ramps is presented in Table 4. The measured ramp ups and downs are compared with the corresponding modelled ramps in the table. Considering the shape parameter of the probability curve and looking at the measured and modelled ramps at 95% confidence interval, CIS based PV systems showed the lowest power ramps under the considered tropical environment. The highest power fluctuations are observed for the HIT based PV systems. The ramp-up and ramp-down characteristics are one of the important parameters for designing solar PV plants, as it indicates the capacity of the required energy storage solutions for the stable operation of the system35.
Models for predicting the possible power ramps from these PV systems, at prospective sites, are developed using machine learning methods like Artificial Neural Network (ANN), Support Vector Machine (SVM) and k-Nearest Neighbor (kNN) are shown as a flow chart in Fig. 8. General details of these models can be found in36,37,38,39,40. Three years of minute-wise and hourly data on the fluctuations in solar insolation and corresponding PV power fluctuations (Eq. 6) were used for developing these models. These are briefly explained in the following sections.
Various steps in developing the proposed machin learning models.
Artificial Neural Network (ANN) is being widely used for modeling the performance of solar PV systems19,36,37,41,42. Among the various available network algorithms, Resilient Back Propagation with weight backtracking (RPROP +)31 is utilized for the current analysis. In RPROP + , a separate learning rate is used for each weight that can be changed during the training process in contrast with the traditional Back Propagation Network, where one learning rate is used for the entire training process. The weights are adjusted by the following rule
where (w_{j}^{(l)}) jth weight at lth iteration step, (upeta _{j}^{(l)}) is jth learning rate at lth iteration step, (frac{{partial E^{(l)} }}{{partial w_{j}^{(l)} }}) is the partial derivative of error function to weight at lth iteration step. Error function applied in this model is the Sum of Squared Error (SSE) that measures the deviation from the actual output during the training time.
To predict the model an output layer with single node and two hidden layers with five and four nodes respectively are used. The number of layers and node has been chosen based on the root mean square errors. These layers are activated by tangent hyperbolic transfer function. The threshold value for the partial derivative of the error function is set as 0.01 which is considered as a stopping criterion of the entire process. This Neural network model learns an approximation function between irradiance and power fluctuation that can be used to predict power fluctuation with respect to new irradiance data set.
Support Vector Machine (SVM) is another powerful machine learning method that performs well with both linear and nonlinear classification and regression problems. If the inputs are nonlinearly separable, it is first mapped into higher dimensional space, using kernel function and then a linear model on that feature space is constructed.
By estimating the accuracy of minute- wise and hourly solar ramp predictions, SVM with radial kernels was chosen for the analysis. The model features, optimized by tuning the errors, include Cost C = (0.1, 1, 100), gamma (gamma) = (0.01, 1).
The k-Nearest Neighbor is the simplest machine-learning method that can be used for both classification and regression technique. The method for developing the algorithms for kernel-based K-NN is described in the steps below [22,23,24].
1. Consider (G = left{ {z,y} right}_{i = 1}^{N}) be training dataset where z is the variation in irradiance, y is the power fluctuation, let z be a new data whose power fluctuation must be predicted.
2. Find the k + 1 nearest neighbors for new data set z according to a Minkowski distance function
where q is the order between two data points zi and zj , here it is set it as two; p is the total number of training data points.
3. Transform the distance with optimal kernel function into weights wi such that
where d is the Minkowski distance, k is set it as seven which selected iteratively.
Compute the weighted averaged k nearest neighbor for predicting the output of new data set z. For the new dataset, the number of nearest neighbors (k) used to compute the distance function is set to 7.
The performance of the models developed using the above machine learning methods is assessed by comparing the modeled ramps with the ramps calculated from the actual power measurements. The prediction errors are quantified using matrices like Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Normalized Root Mean Square Error (NRMSE). These errors for the minute-wise and hourly ramp predictions are presented in Table 5 and Table 6 respectively. The errors of the models vary with the machine learning methods and the type of panels. For example, for the minute-wise predictions, SVM performs well for single crystalline, whereas ANN-based models were found better for polycrystalline, HIT, CIS, amorphous and microcrystalline panels. Similarly, when comparing the hourly prediction results, models with ANN show the best results for Single Crystalline silicon, Poly Crystalline silicon, Amorphous PV module and CIS, whereas SVM is found better for Micro Crystalline silicon and HIT PV modules. Nevertheless, in spite of these slight variations in the accuracies, in general, the solar ramp prediction models with all the machine learning techniques could achieve impressive accuracies.
Figures 9, 10, 11, 12, 13, 14 compare the power ramps of the systems estimated using different machine learning models with actual measurements. The accuracy of these estimations varies depending on the modeling approach. ANN-based predictions are more accurate due to their ability to learn complex nonlinear patterns in solar power variations. This is achieved through the multi-layer perceptron architecture and the RPROP + algorithm, which efficiently updates and optimizes weights and biases. Support Vector Machines (SVM) also perform well, due to it’s kernel trick, by which the nonlinear relationships are efficiently map the varying irradiance conditions with the PV ramp outputs. In contrast, k-Nearest Neighbors (kNN) exhibits lower accuracy due to its simplistic modeling approach, which struggles to capture the power dynamics of PV systems. Additionally, kNN is sensitive to noisy data and cannot effectively model high-dimensional relationships. ANN and SVM-based models can also generalize from historical data and mitigate the adverse effects of outliers using soft margin techniques (SVM) or resilient backpropagation (ANN).
Comparison between ramps predicted using the machine learning method and the actual power ramp measurements of single crystalline in a day and all year (a) ANN, (b) SVM (c) KNN.
Comparison between ramps predicted using the machine learning method and the actual power ramp measurements of poly crystalline in a day and all year (a) ANN, (b) SVM, (c) KNN.
Comparison between ramps predicted using the machine learning method and the actual power ramp measurements of micro crystalline in a day and all year (a) ANN, (b) SVM, (c) KNN.
Comparison between ramps predicted using the machine learning method and the actual power ramp measurements of amorphous silicon in a day and all year (a) ANN, (b) SVM (c) KNN.
Comparison between ramps predicted using the machine learning method and the actual power ramp measurements of CIS in a day and all year (a) ANN, (b) SVM, (c) KNN.
Comparison between ramps predicted using the machine learning method and the actual power ramp measurements of HIT in a day and all year (a) ANN, (b) SVM, (c) KNN.
The ramp models developed in this study have substantial practical value for the design and management of prospective solar PV plants. By integrating these models with time series data on solar insolation variations at the site, designers gain insights into expected fluctuations in power output from the PV systems. Understanding these ramp patterns prior to the construction and integration of solar farms enables developers to optimally size the system and arrange alternative generation options to effectively manage load during ramp-down periods. Additionally, these insights assist design engineers in determining the necessary capacity of storage solutions for the plant. The hourly predictions from the models can serve as valuable tools for developing optimal dispatch strategies for power systems integrated with PVs. Furthermore, when paired with solar insolation forecasts for the plant location, these models enable system owners to accurately commit their production across various time scales, thereby facilitating efficient participation in energy markets.
This paper presents a performance analysis of six different solar PV systems in a tropical environment, based on data from a 1.2 MW experimental solar farm. The PV technologies analysed include single crystalline silicon (sc-Si), polycrystalline silicon (pc-Si), microcrystalline silicon (mc-Si), amorphous silicon (a-Si), copper indium selenium (CIS), and heterojunction with an intrinsic thin layer (HIT). Performance evaluation was conducted using key metrics such as Array Yield, Reference Yield, Capture Loss, Performance Ratio, and Efficiency Ratio.
Among these technologies, amorphous silicon-based systems exhibited the highest performance, achieving an array yield of 4.81 h/day, whereas CIS-based modules recorded the lowest array yield at 3.12 h/day. Capture losses were minimal for HIT (18.06%) and a-Si (18.22%), while sc-Si experienced the highest loss (27.15%). Similarly, HIT and a-Si achieved the highest Performance Ratios (0.80 and 0.81, respectively), whereas sc-Si had the lowest (0.69).
The ramp-up and ramp-down behaviours of these systems were analysed using probabilistic modeling, revealing that all six PV technologies followed the Generalized Logistic Distribution. HIT-based PV systems exhibited the highest power fluctuations, with a 95th percentile hourly ramp-up of 396.6 W and a ramp-down of 364.15 W. In contrast, CIS-based systems had the lowest ramping rates (95th percentile hourly ramp-up: 356.77 W, ramp-down: 323.23 W). These findings provide critical insights for the the appropriate sizing of energy storage solutions for the PV systems.
To predict the ramping behaviour of these PV systems, machine learning models based on Artificial Neural Networks (ANNs), Support Vector Machines (SVMs), and k-Nearest Neighbors (k-NN) were developed. All models achieved high accuracy, exceeding 96%. These predictive models, when integrated with real-time solar insolation forecasts, can help to forecast power ramps, thereby facilitating the efficient grid and market integration of large-scale solar PV power plants under similar operating environments.
All data generated or analyzed during this study are included in this published article.
IEA. Snapshot of Global PV Markets 2024. www.Iea-Pvps.Org 1–16 (2024).
Statista Search Department. Share of Electricity Generation from Solar Energy Sources Worldwide from 2010 to 2023. https://www.statista.com/statistics/1302055/global-solar-energy-share-electricity-mix/#:~:text=Solar%20energy%20accounted%20for%20roughly,percent%20share%20a%20year%20earlier. (2024).
International Renewable Energy Agency. Future of Solar Photovoltaic: Deployment, Investment, Technology, Grid Integration and Socio-Economic Aspects (A Global Energy Transformation: Paper). International Renewable Energy Agency vol. November (2019).
Green, M. A. et al. Solar cell efficiency tables (Version 63). Progr. Photovoltaics: Res. Appl. 32, 3–13. https://doi.org/10.1002/pip.3750 (2024).
Article Google Scholar
Van Haaren, R., Morjaria, M. & Fthenakis, V. Empirical assessment of short-term variability from utility-scale solar PV plants. Prog. Photovoltaics Res. Appl. 22, 548–559 (2014).
Article Google Scholar
Wang, X., Sun, Y., Luo, D. & Peng, J. Comparative study of machine learning approaches for predicting short-term photovoltaic power output based on weather type classification. Energy https://doi.org/10.1016/j.energy.2021.122733 (2022).
Article PubMed Google Scholar
Qu, J., Qian, Z. & Pei, Y. Day-ahead hourly photovoltaic power forecasting using attention-based CNN-LSTM neural network embedded with multiple relevant and target variables prediction pattern. Energy 232, 120996 (2021).
Article Google Scholar
Mellit, A., Pavan, A. M. & Lughi, V. Deep learning neural networks for short-term photovoltaic power forecasting. Renew. Energy 172, 276–288. https://doi.org/10.1016/j.renene.2021.02.166 (2021).
Article Google Scholar
Yan, J. et al. Frequency-domain decomposition and deep learning based solar PV power ultra-short-term forecasting model. IEEE Trans. Ind. Appl. 57, 3282–3295. https://doi.org/10.1109/TIA.2021.3073652 (2021).
Article Google Scholar
Niu, D., Wang, K., Sun, L., Wu, J. & Xu, X. Short-term photovoltaic power generation forecasting based on random forest feature selection and CEEMD: A case study. Appl. Soft Comput. J. https://doi.org/10.1016/j.asoc.2020.106389 (2020).
Article Google Scholar
Zhou, H. et al. Short-term photovoltaic power forecasting based on long short term memory neural network and attention mechanism. IEEE Access 7, 78063–78074. https://doi.org/10.1109/ACCESS.2019.2923006 (2019).
Article Google Scholar
Marcos, J., de la Parra, Í., García, M. & Marroyo, L. Simulating the variability of dispersed large PV plants. Prog. Photovoltaics Res. Appl. 24, 680–691 (2016).
Article Google Scholar
Lappalainen, K., Wang, G. C. & Kleissl, J. Estimation of the largest expected photovoltaic power ramp rates. Appl Energy 278, 115636 (2020).
Article Google Scholar
Raj, V., Dotse, S.-Q., Sathyajith, M., Petra, M. I. & Yassin, H. Ensemble machine learning for predicting the power output from different solar photovoltaic systems. Energies https://doi.org/10.3390/en16020671 (2023).
Article Google Scholar
Abuella, M. & Chowdhury, B. Forecasting of solar power ramp events: A post-processing approach. Renew. Energy 133, 1380–1392 (2019).
Article Google Scholar
Dadras Javan, F., Campodonico Avendano, I. A., Najafi, B., Rossi, M. & Rinaldi, F. Estimating morning ramp-up duration for the cooling season in a smart building using machine learning: Determining most promising features. Sustain. Energy Technol. Assess. 69, 103911 (2024).
Google Scholar
Wang, L. et al. Accurate solar PV power prediction interval method based on frequency-domain decomposition and LSTM model. Energy https://doi.org/10.1016/j.energy.2022.125592 (2023).
Article Google Scholar
Liu, R. et al. A short-term probabilistic photovoltaic power prediction method based on feature selection and improved LSTM neural network. Electric Power Syst. Research https://doi.org/10.1016/j.epsr.2022.108069 (2022).
Article Google Scholar
Rana, M., Koprinska, I. & Agelidis, V. G. 2D-interval forecasts for solar power production. Sol. Energy 122, 191–203 (2015).
Article ADS Google Scholar
Widén, J. et al. Variability assessment and forecasting of renewables: A review for solar, wind, wave and tidal resources. Renew. Sustain. Energy Rev. 44, 356–375 (2015).
Article Google Scholar
Sobri, S., Koohi-Kamali, S. & Rahim, N. A. Solar photovoltaic generation forecasting methods: A review. Energy Convers Manag 156, 459–497 (2018).
Article ADS Google Scholar
Ismail, H. et al. Comparative Performance of Grid Integrated Solar Photovoltaic Systems Under the Tropical Environment. 1–6 (2016) https://doi.org/10.1109/ISGT-Asia.2013.6698758.
Chawla, S. & Tikkiwal, V. A. Performance evaluation and degradation analysis of different photovoltaic technologies under arid conditions. Int. J. Energy Res. 45, 786–798 (2021).
Article Google Scholar
Atsu, D., Seres, I. & Farkas, I. The state of solar PV and performance analysis of different PV technologies grid-connected installations in Hungary. Renew. Sustain. Energy Rev. 141, 110808 (2021).
Article Google Scholar
Nour-eddine, I. O., Lahcen, B., Fahd, O. H., Amin, B. & Aziz, O. Outdoor performance analysis of different PV technologies under hot semi-arid climate. Energy Rep. 6, 36–48 (2020).
Article Google Scholar
Li, C. Comparative performance analysis of grid-connected PV power systems with different PV technologies in the hot summer and cold winter zone. Int. J. Photoenergy 2018, 1–9 (2018).
Article Google Scholar
Quansah, D. A., Adaramola, M. S., Appiah, G. K. & Edwin, I. A. Performance analysis of different grid-connected solar photovoltaic (PV) system technologies with combined capacity of 20 kW located in humid tropical climate. Int. J. Hydrogen Energy 42, 4626–4635 (2017).
Article ADS CAS Google Scholar
Sharma, V., Kumar, A., Sastry, O. S. & Chandel, S. S. Performance assessment of different solar photovoltaic technologies under similar outdoor conditions. Energy 58, 511–518 (2013).
Article Google Scholar
Femin, V., Petra, M. I., Mathew, S., Hazra, J. & Ismail, H. Modeling the Temporal Variations in the Output of Large Solar PV Power Plants. Energy Procedia 95, 294–301 (2016).
Article CAS Google Scholar
Natekin, A. & Knoll, A. Gradient boosting machines, a tutorial. Front .Neurorobot. 7, 21 (2013).
Article PubMed PubMed Central Google Scholar
Sharma, N., Sharma, P., Irwin, D. & Shenoy, P. Predicting solar generation from weather forecasts using machine learning. in 2011 IEEE International Conference on Smart Grid Communications, SmartGridComm 2011 528–533 (2011) https://doi.org/10.1109/SmartGridComm.2011.6102379.
Razali, N. M. & Wah, Y. B. Power comparisons of shapiro-wilk, kolmogorov-smirnov, lilliefors and anderson-darling tests. J. Stat. Model. Anal. 2, 21–33 (2011).
Google Scholar
Lilliefors, H. W. On the Kolmogorov-Smirnov test for normality with mean and variance unknown. J. Am. Stat. Assoc. 62, 399–402 (1967).
Article Google Scholar
Mantel, N. Chi-square tests with one degree of freedom; extensions of the Mantel-Haenszel procedure. J. Am. Stat. Assoc. 58, 690–700 (1963).
MathSciNet Google Scholar
Denholm, P. & Margolis, R. M. Evaluating the limits of solar photovoltaics (PV) in electric power systems utilizing energy storage and other enabling technologies. Energy Policy 35, 4424–4433 (2007).
Article Google Scholar
Tayal, D. Achieving high renewable energy penetration in Western Australia using data digitisation and machine learning. Renew. Sustain. Energy Rev. 80, 1537–1543 (2017).
Article Google Scholar
Omar, M. et al. Day-ahead forecasting for photovoltaic power using artificial neural networks ensembles. in Renewable Energy Research and Applications (ICRERA), 2016 IEEE International Conference on 1152–1157 (IEEE, 2016).
Samworth, R. J. Optimal weighted nearest neighbour classifiers. Ann. Stat. 40, 2733–2763 (2012).
Article MathSciNet Google Scholar
Hechenbichler, K. & Schliep, K. Weighted k-nearest-neighbor techniques and ordinal classification. Collaborative Research Center 386, Discussion Paper 399 (2004).
Abuella, M. & Chowdhury, B. Improving combined solar power forecasts using estimated ramp rates: Data-driven postprocessing approach. IET Renew. Power Gener. 12, 1127–1137. https://doi.org/10.1049/iet-rpg.2017.0447 (2018).
Article Google Scholar
Singh, V. P., Vijay, V., Siddhartha Bhatt, M. & Chaturvedi, D. K. Generalized neural network methodology for short term solar power forecasting. in 2013 13th International Conference on Environment and Electrical Engineering (EEEIC) 58–62 (2013) https://doi.org/10.1109/EEEIC-2.2013.6737883.
Singh, V. P., Vaibhav, K. & Chaturvedi, D. K. Solar power forecasting modeling using soft computing approach. in Engineering (NUiCONE), 2012 Nirma University International Conference on 1–5 (IEEE, 2012).
Download references
The authors are grateful to the University of Brunei Darussalam and the Brunei Research Council for supporting this research project. The first author would like to thank the Universiti Brunei Darussalam for providing a graduate research fellowship for her doctoral study.
Open access funding provided by University of Agder
The authors would like to acknowledge the University of Agder, Norway, for providing financial assistance for the publication of this article.
Universiti Brunei Darussalam, Jalan Tungku Link, Gadong, BE, 1410, Brunei Darussalam
V. Femin, R. Veena & M. I. Petra
Faculty of Engineering and Science, University of Agder, Jon Lilletunsvei 9, 4879, Grimstad, Norway
S. Mathew
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
PubMed Google Scholar
F.V.: Writing—original draft, Visualization, Validation, Methodology, Investigation, Formal analysis, Conceptualization. V.R.: Writing—review & editing, Visualization, Validation. P.M.I. and M.S.: editing, Visualization, Validation, Supervision, Methodology, Investigation, Formal analysis, Conceptualization.
Correspondence to S. Mathew.
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 4.0 International License, which permits use, sharing, adaptation, 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 changes were made. 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/4.0/.
Reprints and permissions
Femin, V., Veena, R., Petra, M.I. et al. Comparative analysis of different PV technologies under the tropical environments. Sci Rep 15, 16371 (2025). https://doi.org/10.1038/s41598-025-99958-x
Download citation
Received:
Accepted:
Published:
Version of record:
DOI: https://doi.org/10.1038/s41598-025-99958-x
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 newsletter — what matters in science, free to your inbox daily.
