Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Distributed ADMM-based Deep Learning Approach for Thermal Control in Multi-Zone Buildings under Demand Response Events

Vincent Taboga, Hanane Dagdougui

TL;DR

This work addresses DR-driven HVAC control in multi-zone buildings by marrying non-convex, data-driven ADMM with deep zone models. It introduces Distributed Planning Networks (DPN), comprising two-layer coordination and learnable dynamics (SSM/RSSM) to plan zone setpoint changes under a total power constraint, with two algorithmic variants: deterministic (DDPN) and stochastic (SDPN). A theoretical convergence argument for non-convex ADMM is provided, and a practical surrogate formulation enables scalable, parallelized coordination. Experiments on an 18-zone EnergyPlus building show accurate horizon predictions (mean absolute errors on the order of a few percent) and successful peak-power reduction during DR events, with convergence and computation-time characteristics favoring SDPN for larger-scale deployments. The results demonstrate a scalable, data-driven approach to DR-enabled BEMS, offering a robust path toward real-world energy savings while maintaining user comfort.

Abstract

The increasing electricity use and reliance on intermittent renewable energy sources challenge power grid management during peak demand, making Demand Response programs and energy conservation measures essential. This research combines distributed optimization using ADMM with deep learning models to plan indoor temperature setpoints effectively. A two-layer hierarchical structure is used, with a central building coordinator at the upper layer and local controllers at the thermal zone layer. The coordinator must limit the building's maximum power by translating the building's total power to local power targets for each zone. Local controllers can modify the temperature setpoints to meet the local power targets. While most algorithms are either centralized or require prior knowledge about the building's structure, our approach is distributed and fully data-driven. The proposed algorithm, called Distributed Planning Networks, is designed to be both adaptable and scalable to many types of buildings, tackling two of the main challenges in the development of such systems. The proposed approach is tested on an 18-zone building modeled in EnergyPlus. The algorithm successfully manages Demand Response peak events.

A Distributed ADMM-based Deep Learning Approach for Thermal Control in Multi-Zone Buildings under Demand Response Events

TL;DR

This work addresses DR-driven HVAC control in multi-zone buildings by marrying non-convex, data-driven ADMM with deep zone models. It introduces Distributed Planning Networks (DPN), comprising two-layer coordination and learnable dynamics (SSM/RSSM) to plan zone setpoint changes under a total power constraint, with two algorithmic variants: deterministic (DDPN) and stochastic (SDPN). A theoretical convergence argument for non-convex ADMM is provided, and a practical surrogate formulation enables scalable, parallelized coordination. Experiments on an 18-zone EnergyPlus building show accurate horizon predictions (mean absolute errors on the order of a few percent) and successful peak-power reduction during DR events, with convergence and computation-time characteristics favoring SDPN for larger-scale deployments. The results demonstrate a scalable, data-driven approach to DR-enabled BEMS, offering a robust path toward real-world energy savings while maintaining user comfort.

Abstract

The increasing electricity use and reliance on intermittent renewable energy sources challenge power grid management during peak demand, making Demand Response programs and energy conservation measures essential. This research combines distributed optimization using ADMM with deep learning models to plan indoor temperature setpoints effectively. A two-layer hierarchical structure is used, with a central building coordinator at the upper layer and local controllers at the thermal zone layer. The coordinator must limit the building's maximum power by translating the building's total power to local power targets for each zone. Local controllers can modify the temperature setpoints to meet the local power targets. While most algorithms are either centralized or require prior knowledge about the building's structure, our approach is distributed and fully data-driven. The proposed algorithm, called Distributed Planning Networks, is designed to be both adaptable and scalable to many types of buildings, tackling two of the main challenges in the development of such systems. The proposed approach is tested on an 18-zone building modeled in EnergyPlus. The algorithm successfully manages Demand Response peak events.
Paper Structure (22 sections, 10 theorems, 43 equations, 9 figures, 4 tables, 3 algorithms)

This paper contains 22 sections, 10 theorems, 43 equations, 9 figures, 4 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Description of the Control Problem
  3. Problem Formulation
  4. Distributed Sharing Algorithm
  5. Distributed Planning Networks
  6. Deep Learning Models for Thermal Zones
  7. State Space Models
  8. Reccurent State Space Model
  9. Distributed Planning Networks Algorithms
  10. Deterministic DPN (DDPN)
  11. Stochastic DPN (SDPN)
  12. Maximum power constraint conversion
  13. Experiments
  14. Description of the test environment
  15. Prediction accuracy of the zones' models
  16. ...and 7 more sections

Key Result

Theorem 2.1

The Algorithm alg:non_convex_admm converges to the set of stationary solutions of Problem (opt:admm_problem): where $Z^*$ is the set of primal-dual stationary solutions of the problem.

Figures (9)

  • Figure 1: Two layers hierarchical control architecture. The LCs observe the state of their zone and plan for the best setpoints while communicating with the coordinator to enforce a maximum constraint on power usage. Setpoints are sent to the existing setpoint trackers that act on each zone.
  • Figure 2: Distributed Planning Network - For each iteration of the ADMM algorithm, the coordinator sends power constraints to the LCs. Using the last observation, the LCs plan for the best trajectory of actions, either by shooting or random search, and send the predicted powers to the coordinator. Once the ADMM has converged, only the first action of the horizon is applied to the zones.
  • Figure 3: Multi-step ahead power predictions with weather forecasts
  • Figure 4: Mean Average Error of the SSM and RSSM in each zone for different horizons. Each model is trained with 5 different training seeds and a dot represents the mean score on the test set for a training seed.
  • Figure 5: Example of power predictions compared to the actual power consumption for one zone.
  • ...and 4 more figures

Theorems & Definitions (18)

  • Theorem 2.1
  • Proposition A.1
  • Proposition A.2
  • proof
  • Proposition A.3
  • proof
  • Lemma A.1
  • proof
  • Lemma A.2
  • proof
  • ...and 8 more