Real-Time Machine Learning-Based Physical Model for Estimating Photovoltaic Output Power Using Meteorological Data

Accurately estimating the output power of photovoltaic (PV) systems is challenging because of missing meteorological data and variable environmental conditions. This paper introduces a novel approach for the real-time estimation of PV output power using a machine learning-based physical model with real-site data validation. The model integrates key meteorological inputs into physical expressions, while missing data are imputed accurately using a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i>-Nearest Neighbor (kNN) model. The model was optimized using an Adam-based gradient descent with an early stopping technique to prevent overfitting. This optimization process improves the model’s ability to capture complex environmental interactions for real-time power estimation with Root Mean Square Error (RMSE) of 21.124W, a 67.5% improvement over the unoptimized model’s 64.575 W, and a 71.5% reduction in Mean Absolute Error (MAE) (from 40.011 W to 11.413 W). When evaluated against a hyperparameter-optimized XGBoost model on a new dataset, the proposed model showed superior efficiency, maintaining stable performance under varying conditions, whereas XGBoost’s errors increase significantly in fluctuating weather scenarios. The robustness of the model can support multiple practical applications, such as detecting suspicious behavior for fault identification and optimizing energy dispatch for efficient grid integration, providing a computationally efficient tool to improve the performance of the photovoltaic system under real-world operations.