Forecasting of Heat Pump Power Consumption using Neural Networks

Abstract: With the increase in electricity generation from renewable sources over recent years, the demand for ancillary services providing balancing support in the grid has risen. Demand side regulation can be performed by regulating consumption of some household devices, which can act as a virtual battery that can be loaded and unloaded into the grid. Among these devices are thermostatically controllable loads, e.g. heat pumps, AC:s and refrigerators. In order to create a virtual battery however, the nominal consumption of the load needs to be known, and in addition, market mechanisms requires this to be known ahead in time. This thesis has investigated predictive models using neural networks for thermostatically controllable loads in the form of heat pumps. The models were compared to two different naive predictors used as baselines. Results indicated that it is possible to beat these baselines, although it is difficult to fully evaluate the performance until put in real operation. Furthermore, it is unknown whether the presented models are optimal, and further work is likely required to find a more optimal model.