Short-Term Prediction of the Photovoltaic Production of a Solar Farm Using Artificial Intelligence and Machine Learning

Abstract

This thesis presents a short-term prediction study of the photovoltaic power production of a solar farm using Artificial Intelligence and Machine Learning techniques. Given Algeria's exceptional solar potential and the growing integration of photovoltaic systems into the national grid, accurate forecasting of PV output has become essential for grid stability and energy management. The study employs a data-driven regression approach using meteorological and operational data, including solar irradiance, ambient temperature, relative humidity, wind speed, wind direction, and sea-level pressure, as input features to predict the target variable Power_PACE. Five Machine Learning models were developed and compared: Linear Regression, Random Forest, Gradient Boosting, XGBoost, and CatBoost. After preprocessing and feature engineering in Python, each model was trained on historical data and evaluated on a held-out test set using MAE, RMSE, and R2 metrics. Results show that Linear Regression achieved the highest coefficient of determination (R2 = 0.9777) with the lowest RMSE (5,953 W), reflecting the dominant linear relationship between solar irradiance and power output in the summer dataset. Random Forest ranked second with the lowest MAE (4,232 W), demonstrating strong robustness and interpretability. SHAP analysis confirmed Solar_Mean as the primary predictor, followed by ambient temperature and hour of day. The findings demonstrate that even classical Machine Learning models can achieve high forecasting accuracy when solar irradiance is the dominant driver, while ensemble methods such as Random Forest offer a more reliable choice for real-world deployment due to their robustness to noisy data and nonlinear behavior.