Electricity load forecasting for distribution network based on long short-term memory recurrent neural network

Electricity load forecasting is important for the planning and operation of power system. As short-term electricity load time series is non-linear and non-stationary, how to improve the forecasting accuracy of electricity load is still challenging. This paper proposes a method based on singular spectrum analysis and long short-term memory (LSTM) network to tackle this issue. The proposed method firstly employs singular spectrum analysis as a preprocessing tool to decompose and reconstruct electricity load time series. Afterwards, the reconstructed components are used as input of different LSTM models for training and forecasting. The proposed method is tested on a set of real data measured by smart meter. The results demonstrate that the proposed method can achieve high forecasting accuracy and outperforms the conventional LSTM model.