Adaptive Algorithm for Forecasting of Medium-term District Heating Demand

Sammanfattning: District heating is one of the most sustainable ways of producing and distributing heat to residential and industrial buildings. District heating load forecasting in the medium- to long-term have an important role in production planning and the strategic development of the district heating market. Inaccurate load forecasts lead to a mismatch in supply and demand, imposing the use of alternative heat sources with higher greenhouse gas emissions. Previous approaches for medium-term load forecasting assumes a static environment and therefore neglect the potential impact of changes in the heat load caused by renovations, such as replacing windows, or drastic changes in social behaviour. When such changes occur, it is desirable to update the forecast. This thesis shows that the static environment assumption should not be made, and propose an adaptive algorithm for probabilistic load forecasting to handle such changes automatically that is based on a dynamic regression ensemble which can add and reweight its models based on their recent performance. It is found that the proposed adaptive approach gains improvement compared to previous work when concept drift is present, and due to only adapting when concept drift occurs it has the same accuracy as previous approaches in static environments. The base model for the ensemble is found by evaluating multiple machine learning models for its effectiveness in predicting heat load.