Lund thesis matches reinforcement learning, a KNN surrogate, and an LLM semantic workflow to three data conditions across the building life cycle

A model-free reinforcement learning controller deployed as a supervisory heating-control layer in thirteen occupied buildings in Austria over 138 days reduced heating energy consumption by 29.7% relative to a multi-year baseline and by 7.9% relative to the previous year, while maintaining acceptable indoor temperatures and reducing mean district heating return temperature by 3.85 °C. The result comes from a long-duration field deployment in occupied buildings rather than simulation or laboratory settings, moving reinforcement learning heating control from simulated validation toward operational evaluation across multiple real buildings. The field deployment covers thirteen occupied buildings over 138 days of real operation and reports both multi-year and prior-year comparisons alongside indoor temperature and return temperature metrics.

A KNN surrogate trained on four years of measurements from Swedish residential buildings achieved RMSE 9.12 kW and R² 0.9184, serving as a data-driven environment to evaluate alternative control actions offline before physical deployment, predicting 4–7% energy savings, mitigating more than 40% of identified demand peaks, and keeping indoor temperatures within the comfort range during 98.4% of evaluated intervals. The work positions the surrogate as a pre-deployment safety evaluation environment, responding to the practical constraint that testing unproven strategies directly on occupied buildings is unsafe. Surrogate accuracy is reported via RMSE and R², and the offline evaluation reports energy savings range, demand peak mitigation share, and comfort-range interval share.

A multi-agent semantic workflow combining large language models and Sentence-BERT interpreted IFC-standard BIM material descriptions, processing 789 building-envelope components, generating 345 candidate material matches, and validating 172 of them, with the complete workflow taking approximately 13.4 minutes and costing less than USD 5. The workflow targets design or renovation stages where operational data are absent and only design information and digital documentation exist, automatically linking heterogeneous material descriptions to validated thermal, mass, and energy-related properties to support energy and hygrothermal simulation. Results are presented as verifiable process metrics: components processed, candidate matches, validated matches, execution time, and cost.

The thesis's overall argument is that at any stage of the building life cycle, an appropriate AI method can be matched to the data then available to support the relevant building performance objective. The framework organizes real-time control, pre-deployment evaluation, and semantic knowledge extraction as complementary approaches rather than a single general-purpose model. The conclusion is supported by three studies corresponding respectively to operational, historical-data, and design-document conditions, and is a synthesis-level framework claim.

The thesis addresses practitioners and researchers in building performance management: operators can look to reinforcement learning supervisory heating control for energy savings and lower return temperatures in occupied buildings with control infrastructure and continuous sensor data; teams wanting pre-deployment evaluation of control strategies can use the surrogate as a data-driven environment to compare alternative control actions offline when historical operational data are available; and when operational data are lacking at design or renovation stage, the semantic workflow can extract validated material properties from IFC/BIM documentation to support energy and hygrothermal simulation. The three methods correspond to operational, historical-data, and design-document conditions, and their applicability presupposes the availability of the corresponding data.

What is available here is the abstract and bibliographic record rather than the full text, so the specific experimental designs, baseline definitions, statistical treatment, and validation details of the three studies cannot be assessed from this material. A careful reader might still watch: how the reinforcement learning control performs in other climate zones, building types, and heating systems; how surrogate accuracy affects the conclusions of offline evaluation; what the fact that only 172 of 345 candidate matches were validated implies about matching reliability; and how the three methods would connect when used together across a single building's life cycle. These are directions for further work.