LLM-CECM: A simulation framework for strategic generation behavior in coupled electricity-carbon markets

The deep integration of electricity and carbon markets creates a coupled decision environment in which generators must coordinate energy offers, quota positions, and expectations about policy shocks. Traditional market simulators often either simplify strategic cognition or incur prohibitive computational cost when richer reasoning is introduced. This paper proposes a rolling multi-agent simulation framework in which carbon-order submission, carbon-cost feedback, and electricity dispatch interact at every time step. The framework combines agents driven by large language models with persona, memory, and belief modules; an adaptive decision-frequency mechanism that allocates expensive reasoning only to informative periods; and a fixed-depth counterfactual-reasoning heuristic for anticipating rival responses. In a controlled 118-bus stress-test setting with Australian electricity-market data, event-window diagnostics show that thermal agents increase markups and quota purchases during a combined market-stress window with an embedded carbon-quota tightening signal. A matched efficiency experiment shows that the adaptive mechanism reduces model calls by 80.5% and runtime by 81.5%, while a congested repeated-clearing experiment shows that counterfactual reasoning lowers the rolling price-volatility proxy from 5.49 to 2.69 Chinese yuan per megawatt-hour. These results provide empirical evidence that the rolling simulation framework can capture policy-responsive coupled-market behavior while improving reasoning efficiency and price-path stability.