Accurate short-term forecasting of electricity consumption is essential for maintaining balance in the power grid and minimizing imbalance costs for energy suppliers. This thesis addresses the improvement of the energy consumption forecasting model used by Frank Energie, a rapidly growing green energy provider in the Netherlands. The current forecasting approach, based on a HistGradientBoostingRegressor machine learning model, is first evaluated to establish a performance baseline. We then investigate several enhancements: (1) incorporating additional predictive features and applying dimensionality reduction techniques (Principal Component Analysis and Lasso regression) to refine feature selection; (2) developing separate forecasting models for daytime and nighttime consumption to account for distinct daily usage patterns; and (3) implementing a neural network model to capture complex non-linear relationships in the consumption data. Each proposed modification is tested and compared
against the baseline using historical consumption data, with model performance assessed primarily by the mean squared error of predictions. The experimental results indicate that both feature engineering and the day–night modeling approach yield improved forecasting accuracy, significantly reducing prediction error compared to the original model. The neural network model further demonstrates the potential to enhance accuracy, albeit with increased computational complexity and the need for careful tuning. These improvements in forecast precision can help reduce imbalance volumes and associated costs for the energy supplier. In conclusion, the study finds that a combination of advanced feature engineering and tailored modeling techniques can contribute substantially to more reliable and cost-effective energy
consumption forecasts. The thesis ends with recommendations for integrating these improved methods into the company’s operations and suggestions for future research to build upon these results.