Deep learning with LSTM networks for medium to long range demand forecasting: Comparison with machine learning benchmark model

Accurate load forecasting has become increasingly important as it can help power companies for better scheduling to reduce excessive electricity production. This study presents an approach that uses long short-term memory (LSTM) neural network to construct a forecasting model for medium to long term aggregate load forecasting. An array of linear and non-linear machine learning models are also built on electric load dataset and the best performing model is used as benchmark for comparison with deep learning approach. Weather related variables temperature, humidity and wind speed are used as exogenous inputs. The proposed methods are applied on a publicly available dataset of electric load consumption in metropolitan France. The goal is to ascertain with what accuracy can aggregated load values in Megawatts can be predicted using machine learning and deep learning approaches. Results show that LSTM based prediction model using only time lags gives low forecast errors and outperformed machine learning models optimized with hyperparameter tuning.