A deep surrogate modeling approach for distributed control of aggregated air-conditioning loads
TL;DR
A three-layer “Aggregator-Edge-End user” architecture leveraging a data-driven workflow and a hybrid deep learning model combining a Temporal Convolutional Network (TCN) and a Bidirectional Long Short-Term Memory (BiLSTM) network is introduced.
Abstract
With the increasing penetration of intermittent renewable energy sources, harnessing the flexibility of aggregated air conditioning (AC) loads has become critical for maintaining grid stability. However, existing control strategies often face a trade-off between protecting user privacy and providing a standardized, model-free interface for grid dispatch. This paper proposes a deep surrogate modeling approach that addresses this challenge. We introduce a three-layer “Aggregator-Edge-End user” architecture leveraging a data-driven workflow. At the core of this framework is a hybrid deep learning model combining a Temporal Convolutional Network (TCN) and a Bidirectional Long Short-Term Memory (BiLSTM) network. This model integrates grid regulation signals and a user-willingness factor, enabling it to generate a reliable power-response mapping without requiring sensitive physical parameters. For real-time power allocation, we propose an Improved Distributed Particle Swarm Optimization (IDPSO) algorithm that embeds a LightGBM model to predict thermal dynamics, thereby reducing reliance on traditional thermodynamic models. Large-scale simulations demonstrate that the proposed approach achieves a low power tracking (mean relative error < 0.2%) and robustness to user disturbances. By effectively decoupling control from physical models, this work presents a scalable and market-compatible framework for demand response that reduces direct exposure of user-level data through edge-side aggregation.