Closed-loop learning for flexible operation of thermostatically controlled loads

Abstract

Thermostatically controlled loads (TCLs) for temperature regulation have accounted for nearly 40% of global building energy consumption. Along with TCLs' substantial amount, they simultaneously offer inherent flexibility potential. In this context, thermal dynamics—the physical processes governing TCLs' influence on indoor temperature variations—play a pivotal role in unlocking flexibility. Thus, the flexible operation of TCLs fundamentally depends on both high-fidelity thermal dynamics modeling and optimizations. Data-driven approaches have been widely employed to develop accurate models, serving as constraints in the optimization. In this process, the modeling results can directly affect the formulation and solving performance of the optimization problem. However, traditional methodologies follow a unidirectional "model - then - optimize" paradigm, resulting in a disconnect between modeling and optimization. To address this gap, this thesis develops a closed-loop learning framework that bridges modeling and optimization. Building upon thermal dynamics modeling and TCL optimization, the proposed framework introduces two key innovations: a performance evaluation module and a feedback loop for model refinement, to systematically address the fundamental interdependencies between modeling and optimization.

In the closed-loop learning, we first explore improving model components of parameter, variable, and structure from the view of algorithmic enhancement, aiming to improve the optimization-oriented performance. We begin the study by proposing parameter-feedback learning for cost-aware thermal dynamics modeling. While the traditional accuracy oriented modeling makes it hard to mitigate the optimality gap, we develop an end-to-end gradient-based training framework for parameter identification. Simulation results verify that the proposed strategy can effectively reduce TCL operational costs.

We then propose variable-reduced learning for efficient TCL optimization. Computational time of TCL optimization increases with high-dimensional variables in thermal dynamics models. We propose to learn latent variables with lower dimensions, then we develop two dimension-reduced algorithms to obtain the original decision results. Case studies verify that the proposed method can improve optimization efficiency.

Furthermore, we propose structure-adaptive learning for the feasible solution of coordinated TCL optimization. Nonlinear and heterogeneous model structures could result in non-convex and inaccurate operational regions. Thus, we transfer thermal dynamics into the standard virtual storage structure and generate a sampling-based aggregated region. We develop a two-stage optimization strategy to ensure feasible solutions by fine-tuning region boundaries based on disaggregation errors. Case studies verify the feasibility of generating decisions by fine-tuning model structures.

We finally explore the influence of data issues on closed-loop modeling. We specifically propose a collaborative learning framework to mitigate the generalization issue caused by limited data. During the collaboration, dual heterogeneity in model structure and data distribution is tackled to ensure the learning performance during data sharing. Case studies verify the effectiveness of ensuring model accuracy and optimization cost-efficiency under limited data.

In summary, we investigate the potential of leveraging thermal dynamics to enhance TCLs' operation flexibility. From a closed-loop learning perspective, we systematically transform thermal dynamics into cost-effective, computationally efficient, and solution-feasible surrogate models that are optimization-friendly for flexible TCL operation.