AI-driven solar energy forecasting using meteorological variables: Insights from a 22 MW PV plant in Nakhchivan

This study presents a systematic comparative evaluation of ten regression-based machine-learning models for day-ahead photovoltaic (PV) energy forecasting under semi-arid climatic conditions. The analysis is conducted using a limited six month dataset (May-October 2024) of real operational production data obtained from a 22 MW grid-connected PV plant in Nakhchivan, Azerbaijan, integrated with key meteorological variables including solar irradiance, air temperature, relative humidity, and wind speed. Linear Regression, Ridge, Lasso, ElasticNet, Support Vector Regression (SVR), Decision Tree, Random Forest, Gradient Boosting, XGBoost, and a Multi-Layer Perceptron (MLP) were benchmarked using mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), coefficient of determination (R²), and five-fold cross-validated R². The results indicate that regularised linear models-particularly Lasso Regression-provide the most consistent balance between predictive accuracy and generalisation stability under moderate data availability, while Random Forest demonstrates strong cross-validated robustness and the MLP exhibits overfitting behaviour, highlighting sensitivity to limited training data. These findings demonstrate that increased model complexity does not necessarily translate into improved forecasting reliability in data-constrained semi-arid environments. The novelty of this work lies in its unified benchmarking framework that simultaneously evaluates predictive accuracy, interpretability, and generalisation performance using real-world utility-scale operational data. By explicitly linking forecasting reliability with sustainability-oriented planning, the study contributes to more reliable life-cycle cost and emission assessment of large-scale PV systems in emerging renewable-energy markets. The findings offer practical guidance for energy planners and policy-makers seeking transparent, computationally efficient forecasting strategies in semi-arid climates.