Abstract
Accurate forecasting of solar photovoltaic (PV) power generation is vital for integrating renewable energy into the grid and meeting Environmental, Social, and Governance (ESG) goals. This study compares traditional machine learning models—persistence, support vector regression (SVR), random forest (RF), gradient boosting—with a physics-informed neural network (PINN) to improve PV output predictions. All models were trained and tested on historical PV power data with meteorological inputs under consistent conditions. The PINN architecture embeds known PV physics (e.g., irradiance-power and temperature-power relationships) into its structure or loss function, aiming to constrain predictions to physically realistic values. Results show that the PINN achieved forecasting accuracy on par with the best conventional model (RF) and outperformed other baselines in most error metrics. In particular, the PINN’s normalized root-mean-square error (~0.3% of system capacity) was comparable to RF’s (~0.2%) and substantially lower than SVR’s and persistence’s, while its mean absolute percentage error was markedly smaller than those of less sophisticated models. Although the RF ensemble slightly edged out the PINN in raw accuracy, the PINN provided similarly low errors without sacrificing physical consistency or interpretability. Moreover, the PINN was able to generate probabilistic forecasts via Monte Carlo dropout, offering uncertainty intervals that, while needing calibration, supply valuable confidence estimates absent from standard ML outputs. In conclusion, integrating domain knowledge through a PINN yields solar power forecasts that rival the accuracy of state-of-the-art machine learning models and offer added benefits in terms of physical realism and uncertainty quantification. These findings support the central hypothesis that physics-informed techniques can enhance or at least maintain forecast accuracy relative to traditional approaches, thereby enabling more reliable solar energy integration and better support for ESG-compliant energy management.