Day-ahead aggregated load forecasting based on two-terminal sparse coding and deep neural network fusion

Abstract

The popularity of smart meters makes it possible to carry out “bottom-up” load forecasting, so as to achieve more refined load forecasting by aggregating users of a certain scale. In this paper, a day-ahead aggregated load forecasting method based on two-terminal sparse coding and deep neural network fusion is proposed. Two-terminal coding is implemented by a sparse auto-encoder. At the feature input terminal, historical power curves are transformed into a sparse vector by the encoder. At the output terminal, the sparse vector is used as an intermediate result of the deep neural network (DNN), and it can be transformed into the day-ahead predicted power curve by the decoder. Two-terminal sparse coding can achieve feature extraction and dimensionality reduction in an unsupervised way, which overcomes the challenge caused by high-dimensional data. In the prediction process, the aggregated load is clustered into different prediction groups. Then for each group, DNNs with different structures are used to predict the load of the same group. Then a concatenated layer is added to fuse these DNNs. So structural advantages of different networks are exploited. Case study shows that the two-terminal sparse coding and DNN fusion can effectively improve the accuracy of day-ahead load forecasting.