Abstract:

Against the backdrop of the new power systems, the penetration rate of distributed energy resources in distribution networks is rising steadily, and the load characteristics are becoming increasingly diversified. Existing short-term load forecasting methods thus fail to effectively capture the high-dimensional nonlinear temporal characteristics of load data. To address this issue, this paper proposes a short-term load forecasting method for distribution networks based on Transformer and ensemble learning. First, a multi-dimensional feature embedding layer is constructed to fuse the temporal and periodic characteristics of loads as well as environmental variables. Second, a multi-head self-attention mechanism is adopted to establish dynamic cross-time interval correlations, thereby extracting the spatiotemporal coupling characteristics of loads accurately. Third, a hierarchical randomized feedforward network is designed, with the Dropout technique integrated to enhance the multimodal representation capability of the model’s latent space. Finally, multiple differentiated Dropout-based models are ensembled, and Bayesian evaluation of forecasting uncertainty is realized through sampling with multiple forward propagations. Experimental results demonstrate that the proposed method outperforms state-of-the-art benchmark models in both forecasting accuracy and stability, and can thus provide effective technical support for the optimal dispatching of distribution networks.

Key words: short-term load forecasting, Transformer, ensemble learning, Dropout, forward propagation sampling