Forecasting electricity consumption in the power grid : An analysis of the effect of using large granular datasets on prediction performance

Sammanfattning

Electricity consumption has increased over the last decade, driven by electrification and the adoption of new technologies. While this development contributes to the United Nations’ Sustainable Development Goals, it also places growing stress on power grid infrastructure. To ensure reliability and optimal utilization, grid operators require accurate short-term load forecasts to support proactive decision-making. While system-level forecasting using machine learning is common, few studies explore granular forecasts, as they demand large datasets and significant computational resources. This thesis investigates whether short-term electricity consumption in the Stockholm region can be forecast more accurately using granular customer metering point data compared to aggregated data from primary substations at the system level. Approximately 2.5 years of hourly consumption data from 500000 metering points were collected and clustered to create more homogeneous groups. XGBoost models were trained on both granular and system-level datasets using GPU-accelerated distributed computing to handle the scale. The results show that, despite the availability of more detailed features, models trained at the customer metering point level performed worse than models trained on system-level data across all prediction weeks. This suggests that increased granularity does not automatically improve forecasting accuracy, particularly when zero consumption values and limited feature engineering are present. The findings highlight both the potential and the limitations of bottom-up forecasting. While large scale GPU-accelerated training is technically feasible, future work should focus on improved feature engineering, handling zero values, refining clustering strategies, and testing advanced models such as transformers. This study provides practical insights for researchers and grid operators considering granular approaches to short-term load forecasting.