A novel stacked ensemble framework with the Kolmogorov-Arnold Network for short-term electric load forecasting

To avoid serious disturbances in both smart grids and traditional utility grids due to overloads, the balance between electricity generation and load demand must be optimally maintained. To achieve this, accurate electricity load forecasting offers necessary tools for energy suppliers and stakeholders to increase their profitability from renewable energy resources and meet the ever-growing electricity demand. However, despite extensive research efforts, the nonlinear dynamics of power system and complex load data continue to challenge the forecasting accuracy. This paper presents a novel stacked ensemble framework that integrates, adaptive boosting (AdaBoost), light gradient boosting machine (LGBM) and multi-layer perceptron (MLP) as initial learners, with the Kolmogorov-Arnold Network (KAN) as a meta-learner for short-term electric load forecasting (STLF). The proposed framework generates meta-data from the outputs of the initial learners, which are then used by KAN to produce final predictions. The KAN utilizes learnable, spline-based activation functions, which allow for dynamic adaptation to complex and nonlinear load patterns. Additionally, a fusion-based feature selection (FFS) technique, incorporating grey correlation analysis (GCA) and ReliefF, is developed to capture both correlation-based and instance-based feature importances. This ensures adaptability of the framework to data dimensionality and enhances accuracy. Experimental validation on the ISO-NE dataset demonstrates that the proposed framework achieves better prediction accuracy and reduced error metrics compared to existing advanced frameworks, while showing a modest increase in training time over multiple forecast horizons.