Heat pump digital twin: An accurate neural network model for heat pump behaviour prediction

Heat pumps are key in reaching the carbon emission reduction goals and can also be used to provide energy flexibility services. In this context, accurate heat pump control is crucial and requires a detailed heat pump model. Such a model should enable to take into account the latest information of the system in order to determine the expected heat pump behaviour. Several works already proposed a variety of heat pump models, but mainly adopt time steps above fifteen minutes. Given these larger time steps, the effects of the internal heat pump control strategies such as the compressor ramping rate or variable speed pump control logics are neglected. Hence, energy management systems are not able to obtain accurate information on the expected heat pump response when applying a certain control signal. Therefore, this work investigates the role of neural networks in developing a digital twin of a water/water heat pump, which adopts a time step of one minute and uses experimental data from a hardware-in-the-loop set-up. An architecture with six inputs and six outputs is proposed. The inputs contain the measured flow rates and inlet temperatures at both the evaporator and condenser, but also the measured and requested condenser outlet temperatures. The outputs of the model are the thermal capacity and electricity consumption, accompanied by the flow rates and outlet temperatures at both the evaporator and condenser. Two different types of neural networks are investigated, i.e. (1) feedforward neural networks and (2) long short-term memory neural networks using several time horizons regarding the previous heat pump operation. The models are evaluated for both a single time step ahead (monitoring) and multiple time steps ahead (behaviour prediction). Regarding the latter aspect, a long short-term memory neural network using the latest 60 min of operational data achieved the following prediction errors: electrical power consumption within ±10 % for 79.70 % of the time, condenser flow rate within ±10 % for 89.81 % of the time and condenser outlet temperature within ±1 % for 83.89 % of the time. The ability of long short-term memory neural networks to take into account previous control decisions is seen key in achieving reliable multiple time step ahead predictions. Finally, the replicability of the model to be used for a different heat emission system is also evaluated. Results show a poorer performance as the model is trained and tested on a different heat emission system, but indicate the potential for replicability when enriching the training dataset.