Short-Term Load Forecasts Using LSTM Networks

Abstract

With the increasing load requirements and the sophistication of power stations, knowing in advance about the electrical load not only at short-term periods such as hours or couple of days but also over the longer-term periods such as weeks and months is indispensable for a range of benefits such as important technical and economic impacts. Traditional methods such as ARMA, SARIMA, and ARMAX have been used for decades. In recent years, the artificial intelligence (AI) techniques such as neural networks and deep learning are emerging in the field of time series analysis. Towards this end, the artificial neural networks (ANN) and recurrent neural networks (RNN) are being explored and have shown promises in much better forecasting as compared to traditional methods. Long short-term memory (LSTM) networks are a special kind of RNN that have the capabilities to learn the long-term dependencies. In this work, we have picked up an electrical load data with exogenous variables including temperature, humidity, and wind speed. The data is used to train the LSTM network. For a fair comparison, the data is also used in traditional methods to model the load time series. The trained LSTM network and the developed models are then used to forecast over the horizons of 24 hours, 48 hours, 7 days and 30 days. The forecasts generated by the LSTM are compared with the results of traditional methods using RMSE and MAPE for all the forecast horizons. The results of a number of experiments show that the LSTM based forecast is better than other methods and have the potential to further improve the accuracies of forecasts.