Solar Power Forecast using Machine Learning

Accurate prediction of national photovoltaic (PV) generation is essential for efficient energy management and maintaining grid stability. However, it remains a challenge because PV generation is intermittent, and rapidly changing environmental conditions can make its output unpredictable. This study aims to forecast PV generation at a national scale in mainland Portugal using several machine learning models. To achieve this, an hourly national PV dataset spanning from 2020 to 2024 was developed by combining REN's PV generation data with NASA POWER meteorological data, temporal cyclic encodings, and lagged power values. A proportional adjustment factor was applied to address data heterogeneity resulting from merging distinct capacity reporting systems, and linear interpolation was used to account for annual capacity growth patterns within each year. Four machine learning models, namely Decision Tree, Random Forest, XGBoost, and LightGBM, were evaluated in two scenarios, with default and tuned configurations. The results show that the tuned Random Forest with a normalized target (scenario 2) achieved the best overall performance, with a mean absolute error (MAE) of 71.1 MW and a coefficient of determination (R2) of approximately 0.97. Findings show that short-term persistence and irradiance features are the most influential predictors at the national scale. Moreover, that capacity-consistent aggregation improves the representativeness of weather data input for national-scale forecasting.