Abstract:

To address the challenges of poor nonlinear adaptability and inaccurate interval estimation in substation short-term load prediction,an interval prediction method integrating adaptive diffusion kernel density estimation (ADKDE) with long short-term memory networks (LSTM) is proposed.Historical load and meteorological data are fused,where ADKDE method analyzes error distributions and LSTM network temporal features to construct prediction intervals at a 95% confidence level.Experimental results based on a 220 kV substation dataset demonstrate that the proposed model achieves an average prediction interval coverage probability (PICP) of 0.914 across four datasets,while reducing the prediction interval average width (PIAW) by 20%-30% compared to the comparison models.The proposed method effectively quantifies load uncertainty,providing reliable interval predictions to support power grid planning.

Key words: data fusion, ADKDE-LSTM, interval load prediction, adaptive diffusion kernel density estimation (ADKDE)