Samenvatting

CO₂ emissions from electricity generation in the Netherlands fluctuate on an hourly basis, depending on which power plants are active in the energy mix. Accurate next-day forecasts of these emissions would enable companies to align energy consumption with low emission periods, but reliable forecasting models for this specific purpose are currently lacking.

This thesis addresses the research question: which machine learning model can best forecast CO₂ emissions, from electricity generation in the Netherlands, for the following day? To answer this question, three sub-questions are investigated: (1) which existing ML models should be compared, (2) which features are needed to make reliable CO₂ emission forecasts, and (3) what makes a CO₂ emissions forecast model "good."

This study takes a comparative experimental approach, evaluating 14 machine learning models across four categories: ensemble methods, deep learning, linear/statistical, and classical time series. The models are trained on 2.75 years of hourly data collected from the Nationaal Energie Dashboard (NED), the Royal Netherlands Meteorological Institute (KNMI), and the European Network of Transmission System Operators for Electricity (ENTSO-E). A total of 39 engineered features is used as input variables. Bayesian hyperparameter optimization (Optuna) is applied to all models, and each model is evaluated across 10 independent runs to ensure statistical robustness. Model quality is assessed using a composite metric that combines statistical accuracy (R², weighted at 60%) with ranking ability (Spearman rank correlation, weighted at 40%), alongside practical factors such as training time, inference speed, and ease of deployment.

The Hybrid CNN-XGBoost model achieves the highest composite score (0.9255, R² = 0.902), while XGBoost is recommended for deployment due to the combination of its impressive performance (composite 0.9195), CPU-only operation, deterministic results, and optimization time of under two minutes. The three most influential features are day-ahead electricity price (10–39% importance across models), maximum wind gust (11–21%), and global radiation (7–13%). Classical time series models such as ARIMAX and VAR perform poorly on this task. XGBoost offers the best balance of accuracy, ranking ability, and practical deployability for next-day CO₂ emission forecasting in the Netherlands.