Attention-Based Multi-Time-Scale Aggregation Network for Short-Term Load Forecasting

Abstract

Precise short-term load forecasting (STLF) is pivotal to household-level demand management and distribution-side operation, yet remains challenging due to the coexistence of fine-grained fluctuations and long-range periodicities across multiple temporal scales. This paper proposes an attention-based multi-time-scale aggregation network (MTAN), a unified end-to-end architecture that integrates heterogeneous neural encoders tailored to three temporal resolutions: a bidirectional long short-term memory network (Bi-LSTM) captures hour-level sequential dynamics, a two-dimensional convolutional neural network (2-D CNN) extracts intra-day spatial regularities, and a three-dimensional convolutional neural network (3-D CNN) mines volumetric inter-day and inter-week dependencies. The scalar branch outputs are projected into scale tokens and adaptively fused through a lightweight multi-head self-attention mechanism, enabling dynamic cross-scale weighting and scale-level diagnostic interpretation. Extensive experiments on a publicly available individual household electric power consumption dataset demonstrate that MTAN mostly outperforms five representative baselines (ARIMA, Bi-LSTM, 2-D CNN, 3-D CNN, and Informer) across ten direct forecasting horizons from 0.5 to 5 hours. The model achieves an average mean squared error (MSE) performance gap of merely 1.36% relative to the best-performing method at each horizon, validating the effectiveness of multi-scale aggregation for short-term household load forecasting while acknowledging that broader system-level generalization requires additional datasets.