Modelling of spatially correlated weather-based electricity forecasting using combined frequency-based signal decomposition with optimized boosting approach

Effective planning of electric energy systems is becoming increasingly vital due to the growing integration of renewable resources and the electrification of end-user industries. However, short-term electrical load forecasting continues to be challenging due to the effects of weather unpredictability and consumer behaviour. This research presents a machine learning-driven forecasting framework that incorporates spatially correlated meteorological data, temporal characteristics, and frequency-based signal decomposition using the Fourier Transform. The primary contribution is a spatially correlation-driven feature selection technique to choose ideal weather input sites, coupled with the extraction of predominant frequency components from the load signal to enhance model input. Three machine learning models are evaluated: XGBoost, AdaBoost, and Multi-Layer Perceptron (MLP) on datasets from two locations in Indonesia: Bali and Jakarta-Banten. XGBoost attained optimal performance with the five most frequent components. For Bali, the model produced an $R^2$ of 0.89, a correlation coefficient (CC) of 0.98, and a root mean square error (RMSE) of 37.83; for Jakarta-Banten, it gave an $R^2$ of 0.90, a CC of 0.95, and an RMSE of 497.99. These findings underscore the advantages of integrating spatial weather relevance with signal decomposition to improve prediction accuracy, which is essential for reliable and efficient power system operations.