Implementation of a machine learning supported model predictive control for a 5th generation district heating and cooling energy hub

Fifth-generation district heating and cooling (5GDHC) is a promising technology for supplying low-emission energy to buildings. However, realizing its full potential requires the central energy supply unit to operate optimally. Model predictive control (MPC) is a proven approach for this task. Although machine learning (ML)-based load prediction and MPC-based operational optimization are well-studied fields, their joint application to 5GDHC networks remains rare. Accordingly, this paper develops an MPC framework that integrates ML-based thermal load predictions with a mixed-integer linear program (MILP) to optimize the operation of a 5GDHC energy hub (EH). To ensure comparability, we use coefficient of variation of the root mean squared error (CV-RMSE) as an error metric for the evaluated ML models. XGBoost delivered the most accurate multi-step predictions 24 h ahead, with a 27.78 % error in predicting residual building loads. We also assessed a deep neural network (DNN) using both recursive and multiple-inputs-multiple-outputs (MIMO) forecasting strategies. The MILP incorporates an air-to-water heat pump (HP), a free-cooling unit (FCU), geothermal borehole field (BHF) and thermal energy storage (TES). In simulation, the MPC framework outperformed a rule-based control (RBC) baseline, achieving annual cost and carbon dioxide (CO 2 ) reductions 27.8 % and 17.0 %, respectively. Under perfect load forecast conditions, the MPC gained an additional 3.9 %-points in operational cost savings and 3.7 %-points in CO 2 reduction. Overall, combining ML-based load forecasts with MPC can improve the efficiency and sustainability of 5GDHC operation.