Smarter Weather Inputs Cut Day-Ahead Solar Power Forecast Errors in Half
Photovoltaic power is notoriously hard to schedule a day in advance. The sun may be predictable in principle, but clouds, haze, and shifting weather patterns turn solar output into a moving target, and the numerical weather prediction (NWP) data that grid operators rely on are often too coarse to describe conditions at a single power station. A new study published in Results in Engineering tackles this problem from the input side, proposing a two-stage framework that first corrects the weather forecasts themselves and only then predicts how much electricity a solar plant will deliver the next day.
The research team, led by Jian-Hong Zhu of the study’s affiliation network alongside co-authors including Jun Zhang and Chenghong Gu, tested their approach on data from a 20 MW photovoltaic station in Hebei Province, China. The dataset spans 30 June 2018 to 12 June 2019 at a 15-minute sampling interval, producing 96 data points per day and more than 32,900 entries in total. The station uses a fixed south-facing array tilted at 31 degrees with polycrystalline silicon modules, a typical utility-scale fixed-tilt configuration equipped with instruments measuring both global and diffuse irradiance.
The scale mismatch at the heart of the problem is striking. Standard NWP products describe weather averaged over grid cells measured in kilometers, while a megawatt-scale solar plant occupies a footprint far smaller than a single forecast grid square. In the sampled week used for illustration, the mean absolute deviations between NWP values and on-site measurements reached 161.26 W/m² for global horizontal irradiance, 1.79 °C for temperature, and 3.28 m/s for wind speed. Wind speed was systematically overestimated, partly because forecast grids cannot resolve the local terrain and surface roughness that shape the wind field at the station.
The first stage of the proposed framework, called VMD-CNN-BiGRU-AM, does not correct the raw weather variables directly. Instead, it computes the error sequence between each NWP variable and the corresponding measured value, then applies variational mode decomposition (VMD) to break that error sequence into several intrinsic mode functions with distinct frequency characteristics. Each component is modeled separately by a hybrid neural network: a convolutional layer extracts local temporal fluctuation features, a bidirectional gated recurrent unit captures temporal dependencies within the historical window, and an attention mechanism weights the most informative time steps. The predicted components are summed back together, and the reconstructed error is added to the original NWP value to yield corrected irradiance, temperature, and wind speed.
A key design choice is that the correction runs in a rolling manner without peeking at the forecast day. For the first time point of the target day, the model uses the latest six historical error values, corresponding to 1.5 hours of information at the 15-minute resolution. For subsequent points, previously predicted errors are fed back into the input window, so all 96 daily points can be generated using only information available before the forecasting day. Normalization parameters come exclusively from training data, and no measurements from the target day enter the procedure, avoiding future information leakage.
The corrected weather variables then feed a second-stage forecasting model, BOCNN-LSTM, which combines a convolutional feature extractor, two stacked LSTM layers, and Bayesian optimization to automatically tune the number of hidden units, the learning rate, and the L2 regularization coefficient. The authors are careful to note that this stage applies well-established components rather than claiming a new forecasting mechanism; its role is to map corrected meteorological inputs and historical power output to a day-ahead power forecast. The framework produces deterministic point forecasts only, without prediction intervals or probabilistic distributions.
The correction stage delivered substantial improvements on its own terms. Compared with raw NWP data, the complete VMD-CNN-BiGRU-AM model reduced mean absolute error and root-mean-square error by 71.3 percent and 74.5 percent for irradiance, 84.2 percent and 82.0 percent for temperature, and 89.2 percent and 87.5 percent for wind speed. Ablation experiments showed that VMD contributed the largest share of the gain: removing it increased irradiance correction errors by roughly 63 to 68 percent, while removing the CNN or attention modules produced smaller but consistent degradations. The bidirectional recurrent unit was retained in all variants as the common temporal backbone, so its individual contribution was not separately quantified.
Downstream, the effect on power forecasting was equally clear. When forecasts built on CNN-BiGRU-AM-corrected inputs were compared with those built on the full VMD-corrected inputs across four one-month seasonal samples, average normalized RMSE fell from 0.1619 to 0.0759 per unit and MAE from 0.1173 to 0.0580 per unit, reductions of about 53.1 and 50.6 percent. The correlation coefficient rose from 0.9327 to 0.9823, the accuracy rate climbed from 83.81 to 92.41 percent, and the qualification rate, the share of intervals where error stayed within an acceptable band, rose from 88.76 to 99.44 percent. For the 20 MW station, the average normalized bias of 0.0201 per unit corresponds to roughly 0.402 MW of systematic underprediction, a directional tendency the authors flag as relevant to day-ahead scheduling and balancing decisions.
Against selected baselines trained and tested under identical conditions, including support vector regression, LSTM, Transformer-LSTM, and TCN-ECANet-GRU, the proposed framework achieved the lowest errors on the four randomly selected seasonal test days, with RMSE values ranging from 0.0236 to 0.0468 per unit. The added computational cost of the correction stage was modest in absolute terms: the full VMD-based correction averaged about 26.76 seconds per run versus 7.61 seconds without VMD on the study hardware, though the authors stress this covers only the correction stage, not total training or deployment time.
The authors are explicit about the limits of their evidence. Validation rests on roughly one year of data from a single station, with four one-month samples rather than continuous rolling-origin evaluation, so seasonal transitions, rare weather events, and interannual variability remain untested. The benchmark set excludes newer architectures such as PatchTST and TimesNet, and no strictly controlled comparison against an architecturally matched single-stage end-to-end model has yet been performed. Future work, the team writes, will extend the framework toward probabilistic forecasting through quantile, ensemble, or conformal prediction methods, and toward multi-year, multi-site validation to establish whether the gains reported here generalize beyond one plant and one year.
Subject of Research: A two-stage deep learning framework combining NWP error correction with Bayesian-optimized forecasting for day-ahead photovoltaic power prediction.
Article Title: Day-ahead BO CNN-LSTM PV prediction based on input of NWP forecasting reconstruction by VMD and CNN-BiGRU-AM network
Article References: Zhu, J.-H., Zhang, J., Tang, J., Zhang, X., Zhang, Q., Luo, K., Ge, Y., & Gu, C. (2026). Day-ahead BOCNN-LSTM PV prediction based on input of NWP forecasting reconstruction by VMD and CNN-BiGRU-AM network. Results in Engineering, 32, Article 113747. https://doi.org/10.1016/j.rineng.2026.113747
Image Credits: AI Generated
DOI: 10.1016/j.rineng.2026.113747
Keywords: photovoltaic power forecasting, numerical weather prediction, variational mode decomposition, CNN-LSTM, Bayesian optimization, attention mechanism, NWP bias correction, day-ahead scheduling, renewable energy, deep learning, solar power, error correction
News Source: Faith Mcneil. (October 29, 2026). Smarter Weather Inputs Cut Day-Ahead Solar Power Forecast Errors in Half. Scienmag.
We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.
Enter your email address to subscribe to this blog and receive notifications of new posts by email.
Email Address
Subscribe
Bioengineer.org © Copyright 2023 All Rights Reserved.
Login to your account below


Remember Me

Please enter your username or email address to reset your password.


Bioengineer.org © Copyright 2023 All Rights Reserved.

source