On the Suitability of Machine Learning Approaches for Long-horizon Power Demand Forecasting in Gothenburg

Abstract

Gothenburg’s electricity grid is projected to require substantial expansion to accommodate the large-scale electrification of society. Long-horizon forecasts of power peaks are critical for informing energy companies of the infrastructure investment needed to meet future demand. This thesis evaluates the capability of machine learning models to produce such forecasts. Machine learning models and traditional time series models were fitted to two datasets of differing temporal coverage and sampling frequency, incorporating target variable data provided by Göteborg Energi alongside publicly available exogenous features. Under the available data, all models projected a continually constant level of power demand, with traditional time series models demonstrating marginally higher reliability across most forecast horizons. The study was subject to several limitations, including small and sparse datasets, historically flat trends in the target variables, and exogenous features with limited explanatory power. It was concluded that the currently available data restricts the viability of machine learning-based modeling approaches for long-horizon power peak forecasting. Nevertheless, the prospects for such methods remain promising, particularly in the context of short-horizon forecasting and with access to richer, more granular and explanatory features.