Unlocking Full Dynamic Optimization of District Energy Systems through State-Space Model Learning
Keywords: surrogate model, state-space, physics-informed models, optimal control, sustainable energy systems
TL;DR: We present a state-space learning framework with hybrid AI for optimizing multi-source district energy systems, cutting computational time from days to few minutes.
Abstract: Predictive control enables the operation of physical systems along an optimal trajectory based on forecasts and dynamic simulations. However, the complexity of system dynamics and high computational cost of optimization typically restrict the optimization window to short horizons. Thus, any potential benefits from mid- and long-term rewards are withdrawn. This is particularly relevant for optimization of district energy systems using various low-environmental-impact sources. To address this, we present an end-to-end methodological framework for learning state-space representations of such systems that significantly reduce computational load. The proposed approach leverages the implicit graph structure of such systems to develop and train a physics-informed spatio-temporal graph neural network. This methodology is evaluated on a real-world district heating system incorporating thermal solar panels, storage, biomass and natural gas boilers. Through historical time-series data augmentation and hyperparameter optimization, the learned model demonstrates strong generalization ability and high accuracy in predicting system dynamics. Our method reduces simulation time by four orders of magnitude, cutting optimization time from several days to mere minutes, while also lowering operational costs by up to 25%.
Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4617
Loading