A Horizon-Aware Hybrid Time-Frequency Framework for Direct Multi-Step Hourly Electricity Demand Forecasting

Abstract

In this paper, a horizon-aware time-frequency hybrid direct multi-step prediction framework is presented for hourly-level power data. First, the original signal is denoised using adaptive noise-complete set empirical mode decomposition (CEEMDAN), and then time-frequency fused features are constructed using short-time Fourier transform (STFT). Horizon-aware attention is introduced to adaptively allocate weights to historical time-series data and frequency-domain features according to the prediction horizon. Meanwhile, a BiMamba2-TCN hybrid backbone concurrently captures both global-local dependencies and local dynamic patterns. Additionally, dynamic linear correction is employed to balance nonlinear modeling with global linear extrapolation. Moreover, during the training process, peak-sensitive optimization is adopted to enhance the fitting accuracy during high-load periods. Experiments on the Ontario IESO dataset show strong long-horizon robustness and stable seasonal performance, supporting reliable short-term load forecasting.