Physically Consistent Multiple-Step Data-Driven Predictions Using Physics-Based Filters

Ying Zhao Lian, Jicheng Shi, Colin N. JonesIEEE Control. Syst. Lett 2023

TL;DR

This paper introduces a novel "physics-based filter" that integrates known physical laws into data-driven prediction models. This approach significantly improves the reliability and consistency of predictions, especially in complex systems like building climate control and energy demand response, by ensuring predictions adhere to fundamental physical principles.

What Problem Does This Paper Address?

Data-driven control methods offer a fast way to develop controllers, but they often struggle with consistency. When raw data, which is typically noisy and imperfect, is fed directly into these models, the resulting predictions can violate fundamental physical laws. This inconsistency can lead to suboptimal or even nonsensical decisions, particularly in applications like building energy management where physical constraints are critical. For instance, a prediction might suggest a building is cooling down while simultaneously consuming heating energy, which is physically impossible. The paper aims to bridge this gap by ensuring that data-driven predictions are not only accurate but also physically plausible.

Methods

The core contribution is a "physics-based filter" that preprocesses data before it's used in a data-driven prediction framework based on Willems' fundamental lemma. This lemma provides a way to characterize system responses from historical data. The proposed filter reformulates the prediction problem as a bi-level optimization problem, which is then converted into a single-level optimization problem for tractability. This optimization enforces physical consistency by minimizing the difference between the raw data and a "physically consistent" version of the data, subject to constraints derived from known physical rules (e.g., temperature consistency, bidding consistency). The paper also details numerical strategies to make this optimization problem solvable efficiently.

Datasets

The paper uses real-world I/O data collected from a building on the EPFL campus. This dataset includes indoor temperature, HVAC electrical power consumption, and outdoor environmental conditions. The data spans a heating season (December 2021 to January 2022) and is used to evaluate the proposed method in two case studies: indoor temperature control and demand response bidding.

Key Insights and Ideas

  • Data Preprocessing is Key: Instead of directly embedding physics into the prediction model's structure (like physics-informed neural networks), this work shows that preprocessing the data itself to be physically consistent is a powerful and effective approach.
  • Willems' Fundamental Lemma as a Foundation: The paper leverages Willems' fundamental lemma, a powerful tool for data-driven system identification and prediction, and extends it by incorporating prior knowledge through robust optimization.
  • Bi-level to Single-level Optimization: The clever reformulation of the physics-constrained problem from a bi-level to a single-level optimization makes it numerically tractable, a significant practical advancement.
  • Physical Consistency Prevents Undesired Behavior: In the demand response case study, the physically consistent predictions prevented the system from making economically detrimental decisions that a raw-data-driven model might have made, highlighting the importance of physical plausibility for real-world applications.

Key Results

  • Improved Decision-Making: The physics-based filter leads to more reasonable and physically consistent decisions compared to using raw data, particularly in scenarios with longer prediction horizons or specific physical constraints.
  • Robustness in Demand Response: In the demand response bidding scenario, the filtered data approach successfully identified infeasible bidding situations that a raw-data approach missed, preventing potential economic losses.
  • Trade-off in Prediction Accuracy: While the filtered data might show a slight decrease in raw prediction accuracy (e.g., higher Mean Absolute Error in some cases), this loss is deemed acceptable due to the significant improvement in decision-making reliability and physical consistency.
  • Computational Efficiency: The proposed reformulation allows for efficient numerical solutions, making it practical for real-time applications. The paper provides computation times demonstrating this efficiency.

Limitations and Future Work

  • Assumption of Linear Systems: While the paper acknowledges that building dynamics can be nonlinear, the core framework is built upon Willems' fundamental lemma, which is inherently for Linear Time-Invariant (LTI) systems. The paper mentions strategies for handling some nonlinearities (like bilinearity) but a fully nonlinear framework is not presented.
  • Numerical Optimization Still Required: The method still relies on solving optimization problems, which can be computationally intensive, although the paper demonstrates efficient solutions.
  • Future Work: The authors suggest exploring more advanced numerical optimization techniques and further analysis of the impact of sampling times on prediction accuracy and computation time. They also hint at extending the framework to more complex nonlinear systems.

Summary updated

Abstract

(Extended Version) Data-driven control can facilitate the rapid development of controllers, offering an alternative to conventional approaches. In order to maintain consistency between any known underlying physical laws and a data-driven decision-making process, preprocessing of raw data is necessary to account for measurement noise and any inconsistencies it may introduce. In this paper, we present a physics-based filter to achieve this and demonstrate its effectiveness through practical applications, using real-world datasets collected in a building on the Ecole Polytechnique Federale de Lausanne (EPFL) campus. Two distinct use cases are explored: indoor temperature control and demand response bidding.
BibTeX
@article{lian2023physically,
  author = {Ying Zhao Lian and Jicheng Shi and Colin N. Jones},
  title = {{Physically Consistent Multiple-Step Data-Driven Predictions Using Physics-Based Filters}},
  journal = {IEEE Control. Syst. Lett},
  year = {2023},
}
Figure