PREDICTING NATURAL GAS CONSUMPTION BY NEURAL NETWORKS

The aim of the paper is to create a prediction model of natural gas consumption on a regional level by using neural networks, and to analyze the results in order to improve prediction accuracy in further research. The output variable consisted of the next-day gas consumption in hourly intervals, while the input space included previous-day consumption in addition to exogenous variables. After conducting a feature selection procedure, two neural network algorithms were trained and tested: the multilayer perception and the radial basis function network with different activation functions. The dataset consisted of real historical data of a Croatian gas distributor. The best neural network model is selected on the basis of the mean absolute percentage error obtained on the test sample. The results were analyzed, and some critical hours and days were identified. Guidelines were reported that could be valuable to both researchers and practitioners in this area.