Review of Deep Learning Application for Short-Term Household Load Forecasting

Abstract

The load forecasting is important for the distribution system operation and expansion planning. The main methodologies for load forecasting using deep learning are Long Short-Term Memory (LSTM) and Convolution Neural Networks (CNN). LSTM is specialized in sequential data; on the other hand CNN is specialized in image data. The residential consumption can be treated as a time series (sequential data) and a two-dimensional (image data) dataset. Therefore, LSTM and CNN can be used to extract data characteristics from the residential consumption dataset. Thus, this paper reviews and compares the main methodologies for residential load forecasting such as CNN, LSTM, and CNN-LSTM. The mean square error (MSE) and root mean square error (RMSE) are used as metrics. The dataset is from real residential consumers in Ireland. The result shows a similar performance in training and testing. The best results are found when CNN and LSTM are used together.