Time Series Forecasting

Abstract

The evolving landscape of the energy industry has heightened the demand for precise consumption forecasting tools, traditionally concentrated on electricity. However, the geopolitical significance of natural gas has propelled the need for forecasting tools in the gas sector. This thesis addresses the criticality of accurate forecasting in the natural gas and electricity domains, which helps to overcome logistical and operation optimization challenges. The thesis proposes a methodology focused on multi-step forecasting, which is based on machine and deep learning algorithms tailored for forecasting tasks in the energy domain. The methodology is evaluated on real-world datasets, demonstrating its effectiveness in handling complex forecasting tasks. Comparative experiments between traditional statistical models and machine learning and deep learning approaches reveal the superiority of machine learning models in terms of accuracy for forecasts of natural gas and electricity consumption. Deep learning models present intermediate results between machine learning- and statistical-based models, suggesting their potential as alternatives to traditional machine learning approaches. However, relying on longer input sequences poses a challenge. In contrast, machine learning methods, while requiring less input data, rely more on engineered features, allowing the incorporation of domain knowledge. A novel evaluation metric, called the Change Point Neighborhood Error (CPNE) was defined. The purpose of the metric is to provide a distinctive measure of the forecast accuracy of the proposed models in parts of the time series where a change point or a data drift emerges. Insights gained from analysis of a change point effects on a forecasting error lead to a definition of multi-step forecasting methodology with change point detection integration and incremental (continual) learning capabilities utilizing Hoeffding tree predictors and Pruned Exact Linear Time algorithm in its core. The integration of change point detection enables the selection of a different model collection for successive time frames. The defined methodology was then evaluated for forecasting scenarios with various densities of detected change points. Models based on this methodology were compared with change point-agnostic baseline approaches and deep learning models. The results of the experiment show that the proposed approach provides better results than deep learning models for the evaluated datasets and that fewer change points generally result in a lower forecasting error. The findings contribute to the ongoing discussion on the advantages and limitations of machine learning and deep learning models, paving the way for future research to explore specific conditions that favor their effectiveness in forecasting energy consumption.