Medium-Term Jointly Load Forecasting via an Enhanced KAN-Based MTL Framework

Electricity load forecasting involves predicting future power demand based on historical data and related factors. Accurate forecasting is crucial for energy distribution and system management. This study introduces a novel multi-task learning (MTL) framework that integrates Kolmogorov-Arnold Networks (KAN) with a multi-head self-attention (MHSA) with a both hard and soft parameter sharing strategy to effectively merge different types of inputs for joint load forecasting. The KAN layer is tasked with cross-feature learning and output fitting while the MHSA mechanism captures the temporal information from multiple perspectives, aiming to facilitate comprehensive nonlinear feature learning. Experimental results show that the proposed method can attains better forecasting accuracy, with an average mean absolute percentage error of 3.842% for 10-day medium-term forecasts.