HeatFlex: Machine learning based data-driven flexibility prediction for individual heat pumps

With their rising adoption and integration into smart grids, heat pumps are becoming an increasingly important source of flexible energy. Heat pump flexibility can be utilized by using controllers to remotely manage their operation while maintaining the temperature within predefined user comfort bounds. Traditional indoor temperature modelling approaches require detailed information about the deployment site, device specific parameters and monitored data, making them inapplicable for the majority of heat pump deployments. This paper proposes a novel data-driven machine learning based method HeatFlex for indoor temperature forecasting and flexibility prediction using only 3 monitored variables: indoor and outdoor temperatures and heat pump power consumption. HeatFlex enables plug-and-play flexibility prediction from heat pumps without requiring exact device and building specifications or installation of additional sensors. This paper also introduces novel flexibility metrics enabling quantitative evaluation of heat pump flexibility prediction performance. HeatFlex is based on deep learning predictive models: Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) recurrent neural networks. Our experimental evaluation compared these networks with traditional multivariate linear regression and SARIMAX time series forecasting model baselines. HeatFlex performance was qualitatively and quantitatively evaluated using data from three real-world heat pump deployments with different building sizes, heat pump types and specifications. Experimental results indicate that HeatFlex is effective to accurately predict over 90% of available potential flexibility.