Learning-based Tuning Rules for Data-Driven Control of Energy Hubs

Abstract

Buildings account for a large share of global energy consumption, and coordinating their heating, storage, and energy conversion components effectively requires a controller that can react to changing occupancy, weather, and pricing conditions. Model Predictive Control has been widely used for this purpose, but its performance depends on an explicit dynamical model of the building, which is costly to obtain and must be rebuilt whenever the building changes. Data-Enabled Predictive Control (DeePC) offers an alternative. A controller can be designed directly from measured input-output data, without identifying an explicit model first. This comes at a cost of its own. DeePC's performance depends strongly on a small number of regularization settings, and no systematic procedure currently exists for choosing them for a new building without repeating a full, building-specific tuning process from scratch.

This thesis develops and evaluates a learned mapping from a building's physical properties and the current season to a recommended regularization setting, using Gaussian process regression. Two Gaussian process models are trained separately for each of four seasons on a grid of regularization settings, one predicting closed-loop comfort violation and one predicting energy cost. Each combination is evaluated in closed-loop simulation across three training buildings differing in size, insulation, and thermal time constant. A recommended setting is selected by identifying the candidates with the lowest predicted violation within a fixed tolerance, then choosing the cheapest among them. The same procedure is applied to the fitted Gaussian process predictions and compared against the measured sweep data.

The fitted models recover each training building's own known optimal setting in most cases, and extend to buildings outside the training set through both interpolation and extrapolation in the physical feature space. No single regularization setting performs acceptably across every building and season at once. Applying a setting outside the building it was obtained for can more than double the resulting comfort violation, confirming the need for a building-specific mapping rather than a single fixed rule.

This thesis shows that a small set of training buildings is enough to learn regularization settings that transfer to buildings and seasons not seen during training, reducing the need for a separate manual tuning procedure for every new building. The results also identify the conditions under which this learned mapping is most reliable, providing a basis for extending the approach to a wider range of buildings and operating conditions in future work.