AI-Driven Predictive Modeling for Solar Power Generation Using Real-Time Photovoltaic Sensor Data

Abstract

Solar PV systems, as part of modern power grids, demand accurate forecasting in real time to address the intermittency issues and ensure a stable grid. Current forecasting models, however, underperform when put into practice because of incomplete sensor coverage and cloud-induced ramp events, lack of cross-site generalization, and lack of quantified uncertainty. In this article, we propose a new end-to-end hybrid network, namely Missingness-Aware Physics-Guided Graph Transformer (MAPGFormer), to tackle these practical challenges. The core pillars of MAPGFormer are: introducing a novel sensor tokenization layer to separate out power, weather, physics, temporal and missingness signals; developing a missingness-aware reconstruction encoder to effectively exploit the structured missing patterns for the robust imputation and forecasting; designing a static-dynamic graph learner to fuse historical similarity with topologies learned by real-time frame attention for reconstruction and inverter/site level spatial modeling; designing a multi-scale temporal Transformer to perform multi-resolutions temporal signal reconstruction and prediction; designing a novel weather-regime Mixture-of-Experts module with specific experts for clear-sky, cloudy/ramp, low-irradiance and missing-sensor conditions; finally designing a probabilistic forecasting head that is able to generate calibrated quantile outputs. The framework was carefully tested on two complementary open Kaggle datasets: 1) Solar Power Generation Data for primary inverter-level development and benchmarking against classical, recurrent and Transformer baselines, and 2) UNISOLAR dataset for external multi-site generalization and transfer learning with validation using Transformer baselines. On the primary data set, MAPGFormer produced remarkable performance in predicting 15-minute electricity demand, achieving an MAE of 12.44 kW, RMSE of 18.09 kW, nRMSE of 2.13%, sMAPE of 4.02%, and R² of 0.9827, whereas the Vanilla Transformer gave a MAE of 13.61 kW and reduced the MAE by 18.6%. The accuracy was consistently demonstrated in multi-horizon and cross-plant analyses, and in UNISOLAR leave one out and held out site analysis, the R2s were 0.9791-0.9803. The model was found to have good robustness when missingness was simulated up to 40% and good probabilistic intervals. The contribution of each architectural component was validated using ablation studies and make the model explainable by SHAP analysis. MAPGFormer sets a new standard in practical PV forecasting for uncertainty-aware real-time grid management systems.