Automatic Generation of Residential Thermal Network Models for Predictive Control from Smart Thermostat Data

Model Predictive Control (MPC) can help a building achieve specific objectives, such as reducing operating cost, minimizing energy consumption, or implementing demand response measures. As its name indicates, MPC relies on an accurate building model. However, determining the model structure and level of detail can be challenging. An extensive analysis on a building-by-building basis is typically required, involving significant time and cost. The task of creating a model remains a critical hurdle for the large-scale uptake of MPC. \nThis thesis contributes a systematic method to generate control-oriented residential building thermal models with focus on day-ahead predictions for MPC. The method relies on data from smart thermostats, since their widespread adoption provides a unique opportunity to develop advanced control strategies. The method presented here can be categorized into two main approaches: single-zone and multi-zone models. \nWhen detailed data are available for each room, multi-zone models may provide better estimates of comfort and flexibility. Québec presents an excellent opportunity for testing multi-zone models because of its widespread utilization of decentralized electric baseboards that allow for individual room control. This research introduces a novel automatic method for multi-zone model generation and selection. The methodology starts with a very simple model and iteratively increases the complexity of the model until the model quality cannot increase further. It is then applied on data from an unoccupied experimental house in Shawinigan, Québec. The resulting 13th-order model can accurately predict all 9 zone temperatures 24 hours in advance, with a Root Mean Squared Error of less than 0.5 °C and its parameters reflect the layout of the house, previously unknown to the methodology. \nThe method was applied in a real-time MPC framework to the experimental house during demand response events and compared to MPC using low-order models and a “business-as-usual” (BAU) reference approach. The MPC employing the multi-zone model modeled the building thermal mass separately and managed to leverage it better to preheat more before demand response events compared to the low-order models. The MPC controller with the multi-zone model reduced electricity costs by 55% compared to the BAU scenario; it also outperformed the 40% cost reduction achieved by MPC controllers based on low-order models. \nOn the other hand, a single thermal zone representation can produce sufficiently accurate predictions when coupled with (uncertain) weather and occupancy forecasts. Second-order single-zone models of 7,800 houses in Ontario and Québec were used to investigate the most suitable data length, data interval and calibration horizon of building models for use in an MPC framework. Overall, models with a calibration horizon of 24 hours, data length of 7 days and time interval of 15 minutes provided the best balance between accuracy and computational resources. The models were then used to assess the large-scale deployment of MPC strategies under existing time-of-use tariffs and dynamic pricing. Results showed that the adoption of MPC can reduce the daily electricity cost on average by 16% in Ontario and by 31% in Québec, respectively. \nLastly, this thesis used smart thermostat data to model and characterize 60,000 homes across North America (the resulting model parameters have been made publicly available, enabling building archetypes and building-to-building knowledge transfer). The results showed that just modeling the indoor air temperature of the building may not suffice. Instead, single-zone models need additional states (e.g., for effective temperature of the exterior and/or interior building materials) for accurate predictions. The building time constants were computed as a means to assess building thermal storage ability.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.