Machine learning for smart building heating optimization

This master's thesis deals with the optimization of building heating control using machine learning methods. Traditional control systems often fail to efficiently respond to the thermal inertia of buildings and variable external influences. Model Predictive Control (MPC) addresses these shortcomings but faces challenges in identifying and maintaining accurate physical models. Therefore, this thesis proposes and evaluates a hybrid approach in a simulation environment, combining a nominal model predictive controller with a residual correction layer provided by a recurrent neural network (LSTM). An experimental comparison with three reference strategies (On/Off, PID, and pure MPC) across three room types and six climate scenarios showed that the pro- posed hybridcontroller statisticallysignificantly reducesthe timeof thermalcomfort violation. The greatest benefit was observed in spaces with variable occupancy and high nominal model inaccuracy (e.g., meeting rooms), where the time outside the comfort zone decreased by tens of percent while maintaining comparable energy consumption. Conversely, in thermally stable zones with a well-identified model, the differencewasminimal. Theresultsdemonstratethatcombiningphysics-basedMPC with data-driven residual correction represents an effective solution that compensates for model inaccuracies without compromising the stability of the optimization process.