Using Dynamic Double Machine Learning for Guided District Heating Forecasting and Physical Parameter Extraction

Sammanfattning

This thesis’ main goals were to provide accurate forecasts and informative physical parameter estimates for energy use of the district heating in the Tingbjerg neighborhood, Copenhagen. Our work is aimed as a contribution for future work towards efficient on-demand energy production. We applied Dynamic Double Machine Learning to estimate the causal effects of weather observations on energy use. These results were used as a reference for feature selection for Dense-LSTM and ARMAX models. Dense-LSTM networks were used for 24 (hour) step ahead predictive modeling. ARMAX models were employed for physical characteristic estimation. According to Dynamic DML, we have found the most impactful weather observations to be ambient temperature, solar radiation, wind speed, rainfall and relative humidity, in this order. We found Dense-LSTM networks to be superior to their LSTM counterpart and provide highly accurate predictions. Lastly, by using ARMAX models we were able to extract informative physically interpretable parameters such as heat loss coefficients, solar gain and diurnal curves, all of which describe heat demand of different building groups under 24hr period.