• forecasting
• machine learning
• time series

Energy resources are essential to our lives and their global demand is constantly growing, which is why consumption forecasting is critical to enable better resource planning and management. This thesis focuses on natural gas and presents the implementation of several machine learning models with the aim of predicting daily natural gas consumption. This prediction is useful to support the business of an important gas distribution company, which has requested the monitoring of its users' consumption to ensure a good gas allocation service and avoid problems during the phase of gas balancing. Therefore, after a brief review of the literature on the subject and an introduction to the main concept of the gas distribution business, the provided dataset of time series was analyzed and different forecasting methodologies were applied to it in order to obtain the desired prediction. Specifically, models belonging to the Exponential Smoothing family, models derived from the ARMA process, LSTM, and an ensemble model, XGBOOST, were trained and tested on a subset of data. Finally, a comparison of the obtained results was performed, highlighting weaknesses and strengths of each model that have arisen during the implementation, possible improvements, and future developments.