Model Predictive Control-Based Optimization of a Transcritical CO2 Thermal Compressor Heat Pump Cycle Using a RNN-based Reduced Model

A Thermal Compressor Heat Pumps (TCHP) primarily utilizes thermal energy sources, such as natural gas or waste heat, to drive a compression cycle for efficient heating applications. TCHP that employ carbon dioxide (CO$_2$) as a refrigerant requires special consideration of its transcritical characteristics to maintain optimal performance. The unique properties of CO$_2$, particularly its high operating pressures and critical point behavior, require precise control strategies to ensure efficiency and system stability. To achieve these objectives, a model-based control framework is established here that automates main cycle actuators such as the high-pressure valve and the speed of the burner fan. A dynamic model based on finite volume approach is proposed to capture the transient processes in the TCHP cycle. The supply water temperature, thermal coefficient of performance (COP), and high-pressure outputs are validated with real experimental data. Two control strategies are tested on the dynamic validated model. The first method employs two independent proportional-integral-derivative (PID) controllers, each designed to achieve specific setpoints derived from offline optimal correlations. The second method utilizes model predictive control (MPC) to perform online optimization based on a defined objective function coupled to a vanilla recurrent neural network (RNN) and long short term memory (LSTM) models. The fast to compute vanilla RNN and LSTM models are deduced from real experimental data enriched with data generated from the validated dynamic model that is slow to compute. Both strategies are evaluated on short term, with variations of the supply water temperature setpoint.