Multi-Feature Power Load Forecasting Model Based on Hyperparameter Optimization and Ensemble Learning

Abstract

Power load forecasting is crucial for the economic dispatch and safe operation of power grids, yet the fluctuating and unstable nature of power load data often limits the accuracy of traditional single-model approaches. This paper presents an integrated learning model designed to improve forecast-ing accuracy and stability by addressing these challenges. The model in-corporates improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN), a vector-weighted average optimization algorithm (INFO), convolutional neural networks (CNN), bidirectional long short-term memory (BiLSTM), and random forest (RF). By thoroughly analyzing and preprocessing the data, the approach effectively handles its non-linear, non-stationary characteristics. The combination of CNN and BiLSTM enhances the model’s ability to capture temporal and spatial features, while RF strengthens generalization. The INFO algorithm dynamically adjusts weights and parameters during training, resulting in significantly improved predictive performance. Experimental results on data from the Australian electricity market confirm that the proposed model outperforms existing approaches in key performance metrics, showcasing its effectiveness and potential for practical applications in power load forecasting.