Short-term Power Load Prediction Based on Multi-kernel Function and High-degree Unscented Kalman Filter

Recently, neural network-based approaches for short-term power load prediction have garnered significant attention and investigation. The accuracy of predictions is significantly influenced by the choice of kernel functions and their parameters. However, load prediction methods based on a single kernel function have limitations and cannot adaptively allocate kernel function weights according to sample characteristics. This paper develops a novel prediction method based on multi-kernel functions and high-degree unscented Kalman filter. First, we assign weight factors to both local and global kernel functions to construct a multi-kernel function neural network model that combines local and global capabilities. Then, we integrate the weight factors with the kernel function parameters to form an augmented parameter vector and establish a nonlinear system parameter estimation model using historical power load data. Additionally, we employ a high-degree unscented Kalman filter to estimate the augmented parameter vector, allowing the multi-kernel function parameters to adapt based on real-time load features. Finally, We perform load prediction using the trained neural network parameters and conduct experimental analysis with actual grid data to verify the effectiveness of the multi-kernel function adaptive fusion algorithm.