Machine Learning-Assisted Model Predictive Control of Underfloor Heating Systems in Heritage Buildings : Energy efficiency and thermal comfort assessment

Sammanfattning

Hydronic underfloor heating (UFH) systems in heritage buildings present a three-way control challenge that, to the author’s knowledge, no existing framework addresses simultaneously: high thermal inertia makes conventional on-off control ineffective, preservation requirements prohibit invasive alterations, and rapid slab heating induces thermoelastic stresses that can permanently damage century-old concrete. This thesis develops, implements, and evaluates a machine-learning-assisted model predictive control (MPC) system for the hydronic UFH and cooling system of the Riga Central Market, a UNESCO World Heritage Site. A 3R2C grey-box resistance-capacitance thermal model is identified from building data using a fixed daily on/off heating schedule, recovering the three thermal resistance parameters to a mean error of 1.7% (maximum 3.1%) of their nominal values (validation RMSE:0.012◦C for slab temperature, 0.008◦C for room temperature). A long short-term memory (LSTM) neural network supplies a 12-hour outdoor temperature forecast to the MPC as a disturbance predictor, outperforming a Random Forest baseline at the critical 12-hour horizon. The constrained quadratic-programme MPC enforces floor temperature and comfort-band constraints as hard limits; a post-hoc thermoelastic stress analysis verifies compliance with the 1.0MPa heritage slab allowable. Four control strategies are evaluated across 14-day cold-wave, typical winter, heat-wave, and typical summer simulations. MPC-LSTM reduces cold-wave comfort violations by 44.1% (35.4% to 19.8%), exceeding the 15% project target, and reduces heat-wave overheating violations by 36.6% (4.6% to 2.9%); the time-of-use variant achieves a 43.9% reduction (4.6% to 2.6%), automatically replicating and surpassing the best manually-tuned D9.3 pre-cooling strategy without parameter tuning. Peak thermoelastic stress (0.044MPa) is 4.4% of the1.0 MPa heritage allowable, confirming passive structural safety. The time-of-use MPC variant saves 7207EUR per year (41.0%) through thermal load shifting to off-peak price periods.