A Study of Machine Learning Model's Performance for Load Forecasting Based on XAI Techniques

Accurate load forecasting is crucial for efficient management, planning, and operation of modern power systems, especially in an era of increasing electrical demand driven by technological advancements and environmental changes. This paper explores the development of hybrid Transformer-based models to improve load forecasting accuracy by capturing shortterm fluctuations and long-term dependencies in energy demand. Three hybrid models were developed, the RNR-Transformer, the GRU-Transformer, and the LSTM-Transformer, and compared with traditional hybrid models such as LSTM-GRU and GRURNN. Transformer models outperformed standard RNN-based architectures, with an MSE of 0.0066 and a Symmetric Mean Absolute Percentage Error (sMAPE) of 15.09%, highlighting its ability to predict energy demand with higher accuracy. Feature importance analysis, using Random Forest, Permutation Importance, and SHAP, highlighted temperature as the most influential variable affecting demand, followed by solar radiation and humidity. The implications of these findings are significant in this field, as hybrid models can enhance the adaptability of load forecasting. These can be implemented in different geographical regions and market conditions, improving energy management, reducing operational costs, and facilitating the integration of renewable energy sources into power grids.