Aggregated forecasting of the load control responses using a hybrid model that combines a physically based model with machine learning

The value of active demand in the electricity and ancillary service markets depends very much on the predictability of its aggregated control responses. In this work, the authors study electrically heated small houses that have electrical heating with heat storage tanks and remote control via a smart metering system. They integrate a simple physically based model to a machine learning forecasting method thus combining the strengths of the component methods. Now a stacked boosters network, a new deep learning method, is applied and briefly compared with a support vector regression, an earlier machine learning model. The simple physically based model component models the thermal dynamics of the heat storage tank and the outdoor dependent heat demand of the house. Varying types of a heuristic market based dynamic load control were applied during the field trials that comprised an identification period (31 May 2012–31 May 2013) and a verification period (1 January 2015–31 December 2019). Each of the 727 houses was hourly metered and aggregated into two groups. The short-term forecast the power of these dynamically controlled groups. They summarise the results. The hybrid method outperformed its component methods.