Toward efficient energy systems based on natural gas consumption prediction with LSTM Recurrent Neural Networks

Finding suitable forecasting methods for an effective management of energy resources is of paramount importance for improving the efficiency in energy consumption and decreasing its impact on the environment. Natural gas is one of the main sources of electrical energy in Algeria and worldwide. To address this demand, this paper introduces a novel hybrid forecasting approach that resolves the two-stage method's deficiency, by designing a Multi Layered Perceptron (MLP) neural network as a nonlinear forecasting monitor. This model estimates the next day gas consumption profile and selects one of several local models to perform the forecast. The study focuses firstly on an analysis and clustering of natural gas daily consumption profiles, and secondly on building a comprehensive Long Short Term Memory (LSTM) recurrent models according to load behavior. The results are compared with four benchmark approaches: the MLP neural network approach, LSTM, seasonal time series with exogenous variables models and multiple linear regression models. Compared with these alternative approaches and their high dependence on historical loads, the proposed approach presents a new efficient functionality. It estimates the next day consumption profile, which leads to a significant improvement of the forecasting accuracy, especially for days with exceptional customers consumption behavior change.