Achieving desirable control for non-inverter HVAC systems by dynamic set-temperature approach

• A dynamic set-temperature control strategy is proposed for ON/OFF HVAC systems. • A neural network (NN1) predicts optimal sensor setpoints using fully connected dense and LSTM architectures. • NN1 correlates sensor readings with thermal and hydrodynamic room behaviour. • Dynamic set-temperature control provides a way of controlling a target variable even when the sensor cannot measure it accurately. The conventional control of non-inverter HVAC systems, which operate using an ON/OFF mechanism, is typically based on comparing the sensor-measured temperature with a fixed setpoint. However, this approach is often inaccurate because these sensors cannot precisely measure the target variable (e.g., the room’s mean temperature). To overcome this limitation without physically modifying the sensor’s location or type, this study developed a neural network model for a benchmark test room, equipped with an AC operating in ON/OFF mode. This model captures the dynamic correlation between the hydrodynamic and thermal characteristics of the room and the sensor temperature readings. It predicts a suitable range for the sensor set temperature ( T dset - and T dset+ ) at each time step, based on the desired target temperature range and boundary conditions. The effectiveness of the proposed dynamic control strategy was compared with both the classical and ideal control strategies for six cases featuring randomly distributed boundary conditions. The comparison considered energy consumption as well as the percentage of time the target variable remained within the desired range (i.e., T set- < T average < T set + ). Results showed that the dynamic control strategy improved regulation of the target variable by 50–100 % and reduced energy consumption by approximately 3–10 % compared to the classical method, while closely approximating the performance of the ideal control. Notably, the strategy remained effective even when varying target setpoints and fluctuating boundary conditions were applied. This performance was achieved even though the input data used to train the models were based on a fixed set temperature of 25 ± 0.5 °C.