An ensemble deep learning model for short-term load forecasting based on ARIMA and LSTM

Electrical load forecasting is an important part of power system planning and operation, which can guide the power enterprises to arrange generation plan reasonably, reduce the cost of power generation, and provide a reference for power grid reconstruction and optimization. However, due to the complicated inner non-linear property and seasonality pattern of electrical load, accurate short-term load forecasting (STLF) is of big challenge. In this paper, we firstly study the large time-span quasi-periodicity of load sequences, including the inner correlation of a short load segment and the quasi-periodicity among the load segments spanning different time duration from a week to a month. Then, an ensemble method is proposed, which combines Auto-regressive Integrated Moving Average (ARIMA) and Long Short Term Memory (LSTM) in order to fully exploit the large time-span quasi-periodicity of the loads. Here, ARIMA model captures the stationary pattern of the load segments, while LSTM extracts the complicated non-linear relations of load segments. The proposed method is evaluated on a data set of load consumption in Toronto, and the results show the proposed method outperforms the existing popular STLF models with a small payload of computational complexity.