Ενεργειακός σχεδιασμός και βελτιστοποίηση κτιρίου με χρήση μηχανικής μάθησης

Buildings account for a substantial share of global energy consumption, making energy-efficient design a critical research area. Traditional optimization approaches, relying on repeated Building Performance Simulation runs, become computationally prohibitive when a large number of design alternatives must be evaluated. This thesis develops and applies a coupled simulation-machine learning methodology to investigate the influence of window-to-wall ratio (WWR), varied independently per orientation, and glazing type on the heating and cooling energy demand of a two-storey, courtyard-type residential building in Patras, Greece. A dataset of 500 design configurations was generated via Latin Hypercube Sampling in DesignBuilder/EnergyPlus and used to train and compare seven families of machine learning algorithms as surrogate models. The Optimizable Gaussian Process Regression (GPR) model achieved the lowest prediction error (RMSE) for both heating and cooling, outperforming linear regression, regression trees, support vector machines, ensembles of trees, kernel approximation models, and neural networks. Using the trained GPR models as surrogates, a multi-objective optimization was performed across 151,875 candidate design scenarios, aiming to simultaneously minimize total energy consumption and maximize average WWR. Fifty-seven Pareto-optimal solutions were identified, and, using the weighted sum method with weights prioritizing low energy consumption over a high average window-to-wall ratio (WWR), a single optimal design (WWR East=10%, South=25%, West=10%, North=80%, double glazing) was determined and verified to lie on the Pareto front. To interpret the design logic underlying this optimal solution, Explainable AI techniques (SHAP and LIME) were applied to the trained GPR models. The analysis revealed that glazing type dominates feature importance across the overall design space, while orientation-specific WWR variables become more influential locally, at the optimal design point itself, a divergence linked to elevated predictive uncertainty near the boundary of the sampled design space. The analysis further revealed an unexpected confound between glazing pane count and optical properties (tint) among the glazing types investigated, demonstrating the practical value of explainable AI in uncovering non-obvious relationships within building energy surrogate models.