Short-term Load Forecasting Based on R-SDBN

Abstract

This paper proposes a self-organizing deep belief network with robust locally weighted scatterplot smoothing method (R-SDBN) for short-term load forecasting, which can not only organize the structure of short-term load forecasting automatically but also denoise the input load data. The deep belief network (DBN) integrates deep learning and feature learning. It can quickly analyze a large number of data and has a strong data fitting ability. And the robust locally weighted scatterplot smoothing (RLWSS) method is introduced to weaken the noise and correct the outlies of load data. However, the optimal structure of DBN and the optimal span of RLWSS are different given specific applications. In the proposed model, the structure of DBN and the span of RLWSS are selforganized with minimum root mean square error as objective, not selected by experience, during the training process, and they are modified after each forecasting. Taking the actual load of a certain city in China as an example, the root mean square error (RMSE), the mean absolute percentage error (MAPE), and the maximum absolute error (MAE) of the 24hour forecasting is 4.38MW, 6.54%, and 12.48MW respectively. Compared with the backward propagation neural network, support vector machine (SVM), and DBN, R-SDBN has remarkable advantages.