This current thesis was induced by a real industrial problem, aiming to create a natural gas forecasting system based on historical data in Hungary which is capable of predicting future consumption thereby facilitating the gas supplier’s future gas transport plan. To establish a predictive system, I use different machine learning techniques which learn on the available dataset and I evaluate them based on the defined cost function. The data used for learning was based on historical consumption (taking into account residential and industrial users), the temperature and the calendar period.
The different machine learning methods are linear regression, feedforward neural networks and recurrent neural networks (including LSTM networks which manages sequential data competently). Linear regression serves as a reference model while the two neural network types have different advantages that make possible to produce a precise prediction. The purpose of the present paper is to show forecasting of time series data is not only feasible by statistical methods but various modern machine learning techniques.