Evolving Dynamic Bayesian Networks for CO(2 )Emissions Forecasting in Multi-Source Power Generation Systems

Global warming is a significant challenge. Among the contributors, CO(2 )emission is the foremost, and almost 40% of global emissions come from electricity generation. In this sense, an accurate prediction of CO(2 )emissions in a multi-source system combining traditional and renewable sources can be used to support the reduction of carbon emissions without affecting the energy demand-supply. Despite the several relevant research in this topic, because of higher uncertainty and variability caused mainly by the intermittent nature of renewable energy, CO(2 )emissions forecasting in multi-source power generation systems is a current challenge. This paper presents CO(2 )emissions forecasting for multi-source power generation systems using evolving discrete Dynamic Bayesian Networks. Our proposal uses an analytical threshold for selecting directed edges by the occurrence frequency as data arrives, allowing a constant adaptation to smoothly converges into a robust forecast model. It was tested using real data from multi-source power generation systems of Belgium, Germany, Portugal, and Spain. Its performance was compared with other forecasting methods. Comparing the results against a traditional DBN that not evolves the structure over time, our proposal was superior highlighting a contribution of performance improvement. The proposed method was better when compared against ANN and XgBoost, with the difference in performance statistically significant.