emission-aware charging control

This study addresses the challenges of peak grid load, voltage instability, and transformer overload caused by uncoordinated electric vehicle (EV) charging, while accounting for the impact of real-time carbon intensity and renewable energy variability on decarbonization potential. The authors propose an emissions-aware Soft Actor-Critic reinforcement learning strategy that, for the first time, integrates real-time carbon intensity forecasts into both the state space and a multi-objective reward function to jointly optimize carbon emissions, curtailment of wind and solar generation, and user charging satisfaction. Evaluated on the EV2Gym platform using EirGrid carbon data and distributed wind–solar models under 50% wind penetration, the approach reduces system carbon intensity to 23.96 gCO₂/kWh—achieving an 87% reduction compared to an uncontrolled baseline—keeps transformer overload below 7 kWh, and attains a combined wind–solar self-consumption rate of 52%.

To address the real-time low-carbon operational demands of data centers, this paper proposes DC-CFR, a multi-agent reinforcement learning (MARL) framework that jointly optimizes carbon emissions, energy consumption, and electricity cost under dynamic weather conditions and time-varying grid carbon intensity. The method integrates renewable-energy-aware IT workload scheduling, coordinated cooling system control, and dynamic uninterruptible power supply (UPS) battery storage management. Evaluated across multiple geographic regions using real-world traces, DC-CFR achieves average annual reductions of 14.5% in carbon emissions, 14.4% in energy consumption, and 13.7% in energy cost compared to the ASHRAE baseline controller. Its core contribution is the first real-time MARL control paradigm for holistic, stack-wide co-optimization of energy efficiency, carbon footprint, and economic cost—overcoming the limitations of conventional static-threshold control strategies—and providing a deployable intelligent decision-making foundation for green data centers.

To address the challenge of real-time low-carbon operation in data centers, this paper proposes a multi-agent reinforcement learning (MARL) framework that jointly optimizes cooling systems, IT load scheduling, and UPS battery energy storage—dynamically responding to weather variations and grid carbon intensity fluctuations. We innovatively formulate a thermal-electric coupling model and integrate a carbon-intensity-aware control mechanism to enable online, simultaneous optimization of carbon emissions, energy consumption, and electricity cost. Evaluated across multiple geographic regions under realistic dynamic grid and meteorological conditions, our approach achieves average annual reductions of 14.5% in carbon emissions, 14.4% in energy consumption, and 13.7% in electricity cost compared to the ASHRAE-standard controller. This work represents the first demonstration of full-stack, real-time, coordinated optimization in production-scale data centers—overcoming the limitations of conventional static, rule-based policies.

Addressing the challenges of coordinated control among electric vehicle charging, HVAC systems, photovoltaics, and energy storage in renewable energy communities (RECs), as well as deployment bottlenecks—including sensor noise, heterogeneous system integration, uncertain user behavior, and unobserved battery state-of-charge (SoC)—this paper proposes EnergAIze, the first end-to-end multi-agent reinforcement learning (MARL) framework tailored for real-world RECs. EnergAIze innovatively integrates noise-robust data preprocessing, a unified synchronization interface for heterogeneous systems, dynamic occupancy modeling, and SoC inference. It employs the MADDPG algorithm enhanced with physics-informed reward shaping and behavior cloning–assisted training, implemented within an edge-cloud cooperative control architecture. Evaluated across a four-residence REC testbed, EnergAIze reduces daily peak load by 9% and overall energy cost by 5%, significantly improving the practicality and economic viability of RL-based control in complex, real-world energy systems.

This work addresses the challenge of coordinated optimization among heterogeneous electric vehicles (EVs) and building energy systems in office park vehicle-to-grid (V2G) scenarios. We propose a reinforcement learning framework integrating Deep Deterministic Policy Gradient (DDPG), action masking, and mixed-integer linear programming (MILP)-guided policy initialization. The method jointly optimizes EV charging/discharging schedules under dynamic and uncertain conditions, simultaneously satisfying user charging requirements, minimizing monthly time-of-use electricity costs, and curtailing net demand peaks. It supports heterogeneous multi-agent coordination, long-horizon optimization with sparse rewards, and generalizable decision-making in continuous action spaces. Trained on real-world EV operational data from an automotive manufacturer, our approach achieves significant electricity cost savings over state-of-the-art baselines and heuristic methods, guarantees 100% user charging satisfaction, and demonstrates strong cross-scenario scalability and engineering deployability.

This study addresses the challenges posed by large-scale electric vehicle (EV) integration—namely, peak load amplification, voltage fluctuations, and curtailment of renewable energy—by proposing a decentralized intelligent charging scheduling framework. Leveraging independent multi-agent reinforcement learning, each EV autonomously makes charging decisions based solely on local information, such as dynamic electricity prices, state of charge, and temporal constraints, while implicitly coordinating with others to jointly minimize user costs and ensure grid safety. For the first time, the performance of contextual combinatorial bandits and policy gradient algorithms is systematically compared within a heterogeneous multi-agent setting under real photovoltaic-driven dynamic pricing. Experimental results demonstrate that both approaches significantly reduce user expenses and alleviate line overloads across varying levels of grid congestion, thereby validating the feasibility, robustness, and practicality of the proposed decentralized strategy.

This work addresses the deep coupling between AI data centers and power systems, highlighting the urgent need for joint optimization of computational and electrical scheduling to reduce carbon emissions. The paper proposes the first unified framework that simultaneously schedules rigid training workloads and elastic inference requests while co-optimizing local generation, energy storage, and bidirectional grid interactions within a microgrid. The approach achieves carbon-aware integrated compute-energy scheduling under constraints on latency, workload continuity, and carbon budgets. Formulated as a mixed-integer linear program, the model jointly optimizes task scheduling, load routing, storage dispatch, and grid interaction strategies. Experiments demonstrate that the proposed method significantly improves operational efficiency and reduces emissions compared to baselines that optimize only computation or energy in isolation, with inference routing flexibility and energy storage playing pivotal roles—particularly when abundant local renewable generation is available.