predictive thermal scheduling

This work addresses the limitation of current large language models (LLMs) in multi-zone HVAC control, which typically lack explicit modeling of building physics and thermodynamic processes. To bridge this gap, the authors propose a knowledge graph that integrates thermodynamic principles with spatial semantics, constructed upon the Brick ontology and enriched with historical environment-controller interaction data to provide LLMs with structured contextual information. This approach represents the first integration of physics-informed spatial semantic graphs into an LLM-based control framework, explicitly capturing inter-zone thermal couplings and building dynamic responses. Evaluated in a five-zone building simulation, the method significantly improves the trade-off between energy efficiency and occupant comfort compared to both conventional and existing LLM-based strategies, achieving the lowest PMV violation rate while maintaining high energy performance.

Traditional building heating control neglects free heat sources; industrial model predictive control (MPC) relies on oversimplified physics-based models, compromising accuracy and interpretability; and purely data-driven models suffer from poor generalizability and lack uncertainty quantification. To address these challenges, this paper proposes an interpretable thermal dynamic modeling framework that integrates Bayesian inference with long short-term memory (LSTM) networks. It is the first work to introduce Bayesian deep learning into building thermal forecasting, enabling high-accuracy indoor temperature prediction alongside calibrated uncertainty quantification. The method balances data-driven performance, physical interpretability, and cross-building generalizability. The resulting model supports robust MPC deployment and is validated across 100 real-world buildings. Results show significantly higher prediction accuracy than industrial-grade physics-based models, along with reliable confidence intervals—enhancing control safety and decision transparency.

This study addresses the inefficiency of conventional residential immersion water heaters, which often operate continuously during winter due to neglecting predictable hot water usage windows and thermal losses. To mitigate this, the work proposes a deadline-aware control strategy that minimizes energy consumption while guaranteeing the target water temperature is reached by a specified time. It introduces, for the first time, deadline-aware reinforcement learning to water heater control, leveraging a Gymnasium-based simulation environment and a first-order thermal loss model. The approach is evaluated against Bang-Bang control, Monte Carlo Tree Search (MCTS), and Proximal Policy Optimization (PPO). Experimental results demonstrate that PPO achieves an average energy consumption of 3.23 kWh over a two-hour horizon, reducing energy use by 26%–69% compared to Bang-Bang control, and outperforming Bang-Bang and MCTS by 54% and 33%, respectively, in typical scenarios, with near-zero inference overhead.

This study addresses the challenges of control design in complex industrial processes characterized by multivariable coupled dynamics by proposing an automated control strategy generation framework that integrates large language models (LLMs) with Bayesian optimization. The approach decomposes control design into structured code generation steps, ensuring physical consistency through execution-based validation and feedback-driven repair. It pioneers the automatic synthesis of decentralized PI controller architectures and their tuning environments directly from dynamic process models. Evaluated on a nonlinear gas preheater benchmark, the generated control schemes—subsequently refined via Bayesian optimization—achieve a 26.5% improvement in closed-loop performance and significantly enhance the transient response of pressure loops, thereby demonstrating the method’s effectiveness and novelty.

This work addresses the limitation of current large language model (LLM) planning agent evaluations, which predominantly focus on task success while neglecting the dynamic impact of other agents’ responses and physical constraints in cyber-physical systems. The authors introduce the first physically verifiable benchmark platform for demand response in smart grids, evaluating the strategic efficacy of four planning architectures—predefined, sequential, hierarchical, and search-based—within a simulated environment comprising 40 heterogeneous prosumers and a radial feeder. They propose an evaluation protocol based on paired forced counterfactuals, common random responses, and event-level deadline feasibility, combined with typed policy declarations and short instruction constraints to explicitly model schedule generation, prosumer dynamics, and power flow computation. Experiments show that three architectures yield feasible, near-optimal solutions; incorporating deadline feasibility prediction reduces average regret from 90.7 to 29.0, outperforming fixed sequential strategies by 61.1%, underscoring the substantial influence of planning architecture and highlighting solution quality selection among feasible outcomes as a key challenge.

Existing building thermal dynamics datasets are predominantly derived from steady-state operation under fixed control policies, which inadequately excite the system’s state space and consequently limit the generalization capability of data-driven models. To address this, this work proposes BuilDyn, a novel toolkit that introduces, for the first time, a control-oriented active excitation mechanism. By combining sampling across building parameter distributions with customizable excitation strategies, BuilDyn generates thermal dynamics data with high state-space coverage and integrates seamlessly into machine learning workflows via a Python API. Experimental results demonstrate that models trained on BuilDyn-generated data significantly outperform those based on conventional datasets in both predictive accuracy and generalization, thereby establishing a robust data foundation for control-oriented modeling, transfer learning, and the development of building-specific foundation models.

This work addresses the challenge of thermal energy storage scheduling, which requires multi-constraint planning several hours ahead and is difficult to generalize across buildings using conventional methods. The authors propose Reinforcement Learning with Verifiable Rewards (RLVR), a novel approach that, for the first time, transforms action values generated by dynamic programming (DP) into dense rewards to fine-tune an open-source large reasoning model with only 30 few-shot examples. This enables the model to function as a high-level scheduler that outputs heat pump setpoints while stabilizing its internal planning mechanisms—such as candidate comparison, lookahead, and feasibility checking. The method reduces carbon emissions from 70.5 to 61.2 kg-CO₂, closely approaching the DP-optimal value of 60.8, and demonstrates strong generalization under prediction errors, novel storage configurations, and cross-task scenarios.