Evolving LLM-Derived Control Policies for Residential EV Charging and Vehicle-to-Grid Energy Optimization

This research presents a novel application of Evolutionary Computation to the domain of residential electric vehicle (EV) energy management. While reinforcement learning (RL) achieves high performance in vehicle-to-grid (V2G) optimization, it typically produces opaque “black-box” neural networks that are difficult for consumers and regulators to audit. Addressing this interpretability gap, we propose a program search framework that leverages Large Language Models (LLMs) as intelligent mutation operators within an iterative prompt-evaluation-repair loop. Utilizing the high-fidelity EV2Gym simulation environment as a fitness function, the system undergoes successive refinement cycles to synthesize executable Python policies that balance profit maximization, user comfort, and physical safety constraints. We benchmark four prompting strategies: Imitation, Reasoning, Hybrid and Runtime, evaluating their ability to discover adaptive control logic. Results demonstrate that the Hybrid strategy produces concise, human-readable heuristics that achieve 118% of the baseline profit, effectively discovering complex behaviors like anticipatory arbitrage and hysteresis without explicit programming. This work establishes LLM-driven Evolutionary Computation as a practical approach for generating EV charging control policies that are transparent, inspectable, and suitable for real residential deployment.

💡 Research Summary

This paper introduces a novel framework that uses large language models (LLMs) as intelligent mutation operators within an evolutionary program‑search loop to automatically synthesize interpretable residential electric‑vehicle (EV) charging and vehicle‑to‑grid (V2G) control policies. The authors argue that while deep reinforcement‑learning (RL) agents achieve high economic performance, their neural‑network policies are opaque, hindering auditability for consumers and regulators. To bridge this “interpretability gap,” the work adopts a “code‑as‑policy” paradigm: the LLM directly writes a Python function that maps the current state (price forecast, state‑of‑charge, photovoltaic generation, household load, time‑to‑departure, etc.) to a signed power set‑point every five minutes.

The core methodology is a six‑stage evolutionary pipeline. First, a compact dataset of state‑action examples (derived from a baseline heuristic) is assembled. Second, a structured prompt is crafted that includes physical guardrails (SoC limits, charger power caps) and a request to generate a function decide_power(state). Third, the LLM (GPT‑4o) produces the code, which is parsed and injected into the high‑fidelity EV2Gym‑Residential simulator. Fourth, the policy is evaluated over multi‑day rollouts, yielding total profit, SoC violation counts, and battery‑degradation proxies. Fifth, quantitative feedback is automatically summarized, together with targeted counter‑examples. Sixth, the previous code and feedback are appended to a new prompt, instructing the LLM to revise the policy. This loop repeats for several generations, progressively improving both performance and readability.

Four prompting strategies are compared: (1) Imitation, which simply asks the LLM to reproduce an existing heuristic; (2) Reasoning, which encourages the model to reason about price‑SoC‑time relationships; (3) Runtime, which queries the LLM at every control step to obtain an action; and (4) Hybrid, which blends supervised exemplars, reward‑driven objectives, and explicit error‑correction feedback. Experiments under identical New South Wales (NSW) household traces show that the Hybrid approach yields the most concise, human‑readable if‑else logic while achieving 118 % of the baseline profit and virtually eliminating SoC violations. The resulting policy explicitly encodes anticipatory arbitrage (price‑threshold triggers), hysteresis to avoid rapid charge‑discharge cycling, and battery‑health safeguards—behaviors that emerged without hand‑coding.

Technical contributions include: (i) a reproducible 6‑stage pipeline that integrates LLM code synthesis with simulation‑based fitness evaluation; (ii) a methodological bridge from classic genetic programming (which mutates abstract syntax trees) to high‑level Python code, preserving auditability; (iii) an empirical benchmark demonstrating that LLM‑driven evolution can match or surpass state‑of‑the‑art RL agents on the EV2Gym benchmark while delivering transparent policies; and (iv) a discussion of regulatory implications, showing that the generated code can be inspected, verified, and certified by grid operators.

Limitations are acknowledged. The study relies on the EV2Gym simulator, so real‑world deployment may reveal distributional shifts. LLM outputs can contain syntax errors or logical contradictions, necessitating robust parsing and static‑analysis safeguards. Future work will explore online, data‑stream‑driven evolution, multi‑household coordination, incorporation of more detailed battery degradation models, and the impact of newer LLM architectures (e.g., GPT‑5) on policy quality.

In sum, the paper demonstrates that LLM‑guided evolutionary computation offers a practical pathway to generate high‑performing, interpretable, and regulator‑friendly residential EV charging and V2G policies, positioning LLMs as powerful design agents beyond traditional code‑completion or pure reinforcement‑learning roles.