Comparative analysis of XGBoost and LightGBM methods for day ahead spot natural gas price forecasting

Abstract

Natural gas is an important part of modern life. Advantages such as being storable, being able to be used in many different places as a primary energy source and being transported in the form of LNG or CNG increase the importance of natural gas. The discovery of new natural gas fields and the importance given to infrastructure, the laws adopted for the liberalization of the market indicate that natural gas will be used in Turkey for many years. In addition, studies aimed at using natural gas with hydrogen in Turkey are promising to increase the life of natural gas. One of the most important criteria for using natural gas in Turkey for many years is predictable prices. Each eligible customer demands predictable pricing and low price. Competition is necessary for liberal markets. Therefore, EPIAS, the market operator, develops the infrastructure of the natural gas market day by day. The weekly product option added to the market after the spot market opened and the opening of the futures gas market are the best examples of this. However, there are very few studies about the spot natural gas market in the literature. This thesis aims increase price predictability in this developing market. In this study, the official reports prepared for the natural gas market were examined first. Later, the dynamics of the spot natural gas market were examined, and a data set was prepared accordingly. The data set is crucial because it is very difficult to get good results without the proper data set. Machine learning applications are very popular and are very effective in solving various problems. In this study, day ahead spot natural gas price prediction was made with two different machine learning algorithms based on decision trees. XGBoost and LightBoost algorithms were utilized and performances of these two algorithms were compared. First, it was concluded that these algorithms over-learn at which value ranges by iterating the number of trees, tree depth and learning rates. A more detailed price prediction was then made using these ranges together with the GridSearch function of the sklearn library. Then the best features for price prediction are determined by the feature selection function of the XGBoost and LightBoost libraries.