Using Data-Driven Simulation Models for Deep Reinforcement Learning Based HVAC Control

Analysis of Meta-Reinforcement Learning on Transfer Learning for HVAC Control

Deploying reinforcement learning (RL) agents for heating, ventilation, and air-conditioning (HVAC) control across climatically diverse buildings is challenging, as policies trained in one climate often fail to generalize to others. We investigate transfer and meta-reinforcement learning for building climate control using Sinergym with a custom EnergyPlus model of a real room, comparing Double Deep Q-Network (DDQN) and Model-Agnostic Meta Learning DDQN (MAML-DDQN) across three transfer scenarios on a target very hot climate. Fine-tuned DDQN reduces temperature violations by a factor of 35 over the from-scratch baseline and improves reward by 86% over zero-shot transfer, while MAML-DDQN converges in less than half the fine-tuning episodes and exhibits substantially lower variance across seeds. Results highlight a key trade-off: MAML-DDQN is preferable at adaptation-focused deployments, while fine-tuned DDQN is superior when maximum performance is the objective.

Safe Deep Reinforcement Learning for Energy-Efficient HVAC Control in Multi-Zone Residential Buildings

HVAC systems represent a major share of building energy consumption. Traditional control strategies are limited in coordinating energy-comfort tradeoffs across multiple zones simultaneously. Reinforcement learning (RL) offers adaptive, data-driven control that optimizes performance over time. However, deploying learned neural network controllers in safety-critical building systems remains challenging due to lack of formal safety guarantees. We propose a safety-certified deep RL framework for multi-zone residential HVAC control. Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) agents are trained in an EnergyPlus/Sinergym simulation to minimize energy consumption while maintaining thermal comfort. Post-training safety certification is performed on the PPO policy using Lipschitz-based forward invariance analysis, building on existing tools for the computation of Lipschitz constants for neural networks, to guarantee constraint satisfaction. Both agents are evaluated over an annual simulation cycle in an eight-zone variable refrigerant flow (VRF) testbed. The PPO agent achieves 67\% comfort violation reduction compared to rule-based control, while the SAC agent achieves 27.6\% energy savings. The PPO policy satisfies formal safety certification with a margin of $2.003^\circ$C. These results demonstrate the feasibility of combining reinforcement learning with post-training safety verification for multi-zone building control.

Verifier-Based Reinforcement Fine-Tuning of Reasoning Models for Thermal Energy Storage Control

Buildings are expected to shift cooling loads in response to grid conditions. Thermal energy storage (TES) enables this shift, but scheduling it well requires planning hours ahead under storage constraints. Model predictive control (MPC) and reinforcement learning are difficult to scale across buildings. This study instead adapts an open-weight reasoning model through reinforcement learning with verifiable rewards (RLVR). We convert exact offline dynamic-programming (DP) action values into dense rewards for every candidate action. Using only 30 training prompts, reinforcement fine-tuning (RFT) trains the model as an upper-level scheduler that outputs hourly heat-pump setpoints from text-based states and forecasts. Evaluation uses a deliberately simple office-building TES benchmark where exact DP is tractable and the optimum is known. RFT reduces the open-weight model's emissions from 70.5 to 61.2 kg-CO2, close to the DP optimum of 60.8 kg-CO2. GPT-5 nearly matches DP and MPC without task-specific training, while GPT-4o, a non-reasoning LLM, produces higher emissions than the no-storage baseline, so inference-time reasoning appears important. Trace analysis shows that RFT mainly stabilizes observable planning patterns (candidate comparison, look-ahead, and feasibility checking) rather than creating a new strategy. Robustness and generalization tests clarify what transfers: the reinforced planning patterns persist under forecast errors and an unseen TES condition and carry over to a battery task, but its different structure limits the gains. DP-based verifiable rewards offer a practical way to adapt open-weight reasoning models to building storage scheduling. These results motivate higher-fidelity tests of whole-building control and scalable verifiers for city-scale energy management.

Exact Action Values Are Not Enough: Rollout-Verified Reinforcement Fine-Tuning of a Reasoning Model for Multi-Zone VAV Control

Multi-zone variable-air-volume control must balance thermal comfort, indoor air quality, and electricity use across several continuous actuators. Model predictive control and reinforcement learning are widely studied, but deployment typically requires building-specific modeling or training, limiting scalability. We first test whether a frontier reasoning model (an LLM trained to use additional inference-time computation) can achieve competitive VAV control from text without building-specific training. With that capability established, we then test whether TD3-guided reinforcement fine-tuning (RFT) can transfer control knowledge into a locally deployable open-weight model. Five controllers are evaluated over three summer days in a physics-based four-zone emulator. Relative to a Guideline 36-based baseline, TD3 reduced HVAC electricity by 4.5% while improving temperature and CO$_2$ compliance. Without building-specific training, GPT-5 achieved the largest reduction (6.2%) but reduced the ventilation margin. For RFT, deterministic rollouts restore a saved state, apply one candidate, and follow TD3 to score each action. Auditing a learned critic against these rollouts exposed a failure hidden by its near-perfect across-time correlation ($r=0.9998$): within-state ranking was unreliable; the critic selected the rollout-best candidate in only 5 of 10 states. Even with the rollout verifier, 200 RFT steps produced no sustained improvement in sampled-action return; the open-weight controller used more electricity than the baseline before and after training, and its five-minute predictions remained worse than persistence. GPT-5 predicted transitions far better. Exact rollout scores rank sampled actions but reveal neither next-state effects nor an improvement direction. The unchanged transition errors motivate transition-focused supervised fine-tuning before value-based RFT.

RL-DTNet: Reinforcement Learning Driven Deep Temporal Network for Accurate State-of-Charge Estimation in Lithium-Ion Batteries

Accurate state-of-charge (SoC) estimation is essential for ensuring the reliability and operation of lithium-ion batteries in electric vehicles. Traditional data-driven models failed to generalize across varying load profiles. Also, they do not handle nonlinear electrochemical behavior and long-term degradation effects of batteries. Recently, reinforcement learning–based SoC correction techniques have shown promising results in SoC predictions. But these models are based on synthetic or simplified feedback, which is unstable for real-world conditions. The goal of this work is to develop a novel approach for improving SoC prediction in real-time battery management systems. To achieve this, in this work, a reinforcement learning driven deep temporal network (RL-DTNet) is proposed for SoC prediction. The RL-DTNet consists of four major stages: (1) an adaptive feature extraction layer to construct an enriched multidomain representation; (2) a temporal attention gated recurrent unit (GRU) module to focus on the most informative timesteps under varying load disturbances; (3) a reinforcement learning self-correction module using a deep Q-network to learn how to minimize prediction errors by adjusting outputs based on real-time error feedback; and (4) a long-term degradation modeling layer to adjust SoC estimates for aging effects using a cycle-aware degradation factor. The novel contribution of this work is the integration of reinforcement learning for self-correction, temporal attention to handle dynamic dependencies, and a degradation-aware model for long-term prediction accuracy. Experimental results on real-time datasets show that RL-DTNet achieves better results of the root mean squared error, mean absolute error, and coefficient of determination ( R 2 ) of 0.6377, 0.5056, and R 2 , respectively.

A Hybrid Classifier-Guided Deep Q-Learning Approach for Multi-Objective Building Environment Management