*Important notice: This news reports on an unedited version of an accepted paper and is awaiting final editing. Therefore, the paper should not be regarded as conclusive or treated as established information.
A novel continuous-time reinforcement learning framework has been developed to enhance energy efficiency, comfort, and sustainability in multi-zone building management systems. The framework, termed NODE-RL-BEM, combines neural ordinary differential equations and adaptive control and was first introduced in Scientific Reports.
Study: AI-driven optimization: revolutionizing energy efficiency in modern buildings. Image Credit: Monster Ztudio
Building Energy Optimization Challenges
Conventional building energy management systems typically rely on discrete-time control methods and simulation-dependent strategies, which often fall short in capturing the continuous and nonlinear dynamics of multi-zone building environments. These approaches tend to lack temporal fidelity, adaptability, and generalizability across a wide variety of building types and operational conditions.
The increasing complexity of building systems and the drive toward sustainability necessitate intelligent, data-driven frameworks that can integrate diverse operational data, learn continuously evolving system dynamics, and optimize control policies adaptively in real time.
Download a free PDF copy of this news piece here!
Addressing these challenges, the study proposes a continuous-time reinforcement learning approach that leverages neural ordinary differential equations (neural ODEs) to model building dynamics within an AI-driven optimization framework.
NODE-RL-BEM Framework Design
The research introduces NODE-RL-BEM (Neural Ordinary Differential Equation Reinforcement Learning for Building Energy Management), a unified optimization model designed to improve energy use in complex multi-zone buildings.
The framework integrates heterogeneous data sources, including building energy consumption, environmental conditions, structural characteristics, and occupancy information. This data is preprocessed via normalization and aligned to a consistent time scale before being transformed into latent state representations using a state space encoder and temporal context constructor that capture operational dependencies.
At the core, a neural ordinary differential equation module models the continuous-time dynamics of the building system, enabling smooth and accurate representation of nonlinear and time-varying thermal behavior. This neural ODE module predicts the evolution of latent states over time by integrating differential equations parameterized through neural networks, thereby encapsulating inherent physical relationships and temporal continuity.
Coupled with the dynamic model, a reinforcement learning policy network adaptively outputs control actions, such as HVAC setpoints, by responding to the present system state embeddings. Multi-objective reward signals based on energy efficiency, occupant comfort, and indoor air quality provide feedback to optimize both the dynamic model parameters and the control policies simultaneously via joint continuous-time optimization.
The training procedure involves propagating system states forward in time using the neural ODE dynamics, applying control actions proposed by the policy network, and observing the resulting system response and reward signals.
Optimization steps update both system dynamics and policy parameters in an intertwined manner, ensuring coordinated learning. Additionally, the framework incorporates a generalization evaluation mechanism to assess robustness and transferability across different building configurations and operational scenarios, thereby supporting scalable deployment and adaptive control.
Performance evaluation is conducted on publicly available datasets, namely the ASHRAE Great Energy Predictor III and Intelligent Indoor Environment datasets, which provide real-world energy consumption, environmental, and occupancy measurements from diverse building types.
Continuous-Time Control Outcomes
The NODE-RL-BEM framework demonstrated strong performance gains in multiple key areas related to building energy optimization. Modeling continuous-time system dynamics through neural ODEs effectively captured nonlinear thermal behaviors and allowed smooth state transitions, even under discontinuous or rapidly changing environmental conditions.
This enhanced predictive fidelity resulted in better alignment between predicted and observed energy consumption trajectories across training and testing sets.
In the EnergyPlus-based evaluation, energy savings achieved ranged from 42% to 48% compared to baseline control strategies, representing a substantial reduction in building operational energy demand. Importantly, these savings did not come at the expense of occupant comfort; the framework maintained comfort violation rates below 0.5%, indicating consistent thermal regulation and stable indoor conditions.
Indoor air quality metrics, defined as the reduction of CO2 levels, showed improvement ranging between 28% and 35%, underscoring the framework’s capacity to balance competing objectives effectively.
Reinforcement learning policies trained in this continuous-time setting produced smoother control signals with fewer abrupt changes than discrete-time controllers. This reduction in control "jitter" translates into more predictable operational behavior, critical for real-world building management system (BMS) integration.
A notable contribution is the framework’s strong generalization capability, with a reported generalization score of 0.91 across diverse building types and operational patterns. This figure was calculated using the model’s energy-savings and comfort-retention results.
This finding suggests NODE-RL-BEM is scalable and could have practical applicability in heterogeneous real-world settings, addressing a key limitation in many existing building energy optimization models that struggle to translate between datasets or configurations. However, the framework has not yet been deployed in a physical building, so future work should move beyond the EnergyPlus-based simulation.
Scalable Intelligent Energy Control
This research advances the field of building energy management by developing a continuous-time AI-driven optimization framework, NODE-RL-BEM, which integrates neural ordinary differential equations with reinforcement learning to model and control building energy dynamics adaptively.
Future work aims to extend this methodology to real-world implementations by interfacing with physical building management systems and sensor networks, incorporating uncertainty-aware learning, and exploring hybrid control architectures.
Such developments can further enhance the robustness, interpretability, and grid-interactive capacity of intelligent building energy systems. Overall, this study provides a methodological and practical foundation for next-generation autonomous, sustainable building operations powered by continuous-time AI optimization.
Journal Reference
Alshammari H.H. (2026). AI-driven optimization: revolutionizing energy efficiency in modern buildings. Scientific Reports. https://www.nature.com/articles/s41598-026-60852-9.