Dynamic adaptive hierarchical TCN driven by IHOA-VMD optimization for short term load forecasting

To address the challenges of high stochasticity, non-stationary fluctuations and multi-scale temporal dependencies in short-term power load forecasting (STLF), this paper proposes a dynamic adaptive hierarchical temporal convolutional network (TCN) model combined with variational modal decomposition (VMD) optimized by an Improved Hiking Optimization Algorithm (IHOA) for data preprocessing. By adaptively adjusting the VMD parameters through IHOA, the original load sequence is decomposed into multiple modal components, effectively extracting multi-scale frequency features. The core innovation lies in the design of the TCN structure of a Layer-wise Dynamic Fusion Mechanism (LDFM), which leverages a learnable weight generator and adaptive gating units to dynamically integrate feature representations from different layers, thereby enhancing the model’s ability to capture complex temporal dependencies. Experimental results based on four real-world electric load datasets show that the proposed model outperforms existing state-of-the-art methods. It demonstrates superior prediction accuracy and generalization performance, significantly reducing load forecasting errors under non-stationary conditions.