Abstract
Modern energy systems necessitate accurate peak demand predictions to minimize costs. This study presents an adaptive FastICA-Transformer that improves feature extraction and forecasting precision using entropy-guided component selection and a redesigned transformer architecture. The model modifies transformer depth according to the forecast horizon and initializes attention heads using statistically independent FastICA components, resulting in interpretable and efficient networks. In comparison to PCA-Transformer and sophisticated neural architectures (LSTM, TCN, N-BEATS, TFT), the FastICA-Transformer demonstrates enhanced MAE, RMSE, and MAPE in forecasting for Next-Hour, Next-Day, Next-Week, and Next-Quarter intervals. The results indicate that architectural adaptation, informed by temporal and statistical characteristics, efficiently tackles peak demand forecasts in practical power systems.