The Downscaling-Upscaling Method for Load Prediction Based on Feature Prediction and Curve Clustering

Abstract Load forecasting, as a classical problem, exhibits decreased accuracy in existing algorithms when the prediction window lengthens. This paper proposes a novel approach based on a downscaling-upscaling paradigm. In the downscaled process, the daily minimum load and the daily total load sequence are extracted from the load sequence. These two features are predicted using piecewise linear regression. The upscaling process involves clustering the daily load curves into multiple clusters using the K-means clustering method. The centroids of each cluster form a standard load curve library. During prediction, a standard load curve is selected from the library based on the daily load curve of the current day, and the final daily load curve prediction is obtained by applying translation and scaling operations. The translation and scaling parameters are determined by the predicted daily minimum load, predicted daily total load, and the daily load curve of the current day. Experimental results demonstrate that the proposed method achieves superior accuracy compared to existing methods for long prediction window widths.