Peer-to-peer electricity platforms with endogenous prices: A Multi-Agent Neural Network Control approach

The transition towards decentralised energy systems creates a need for local electricity platforms on which prosumers trade energy at endogenously determined prices. We develop a differentiable multi-agent control framework for such platforms, drawing on neural stochastic control and the pathwise optimisation principle used in Deep Hedging, whose appeal stems from its simplicity, computational efficiency, industrial use, and scalability to large-scale optimisation problems. The proposed approach embeds physical storage dynamics, feasible prosumer actions, and the market-clearing rule in a single computational graph, allowing decentralised agents to train trading policies by backpropagating through aggregate supply, aggregate demand, and cleared prices. A stop-gradient construction at the clearing layer separates two economically distinct learning modes: price-aware learning, in which agents internalise their marginal price impact, and price-taking learning, in which prices are treated as exogenous. This yields a market-mechanism-aware alternative to reward-only multi-agent reinforcement learning methods such as Multi-Agent Proximal Policy Optimization, which treat the market as a black-box environment component. We evaluate the framework on peer-to-peer electricity markets with heterogeneous prosumers, photovoltaic generation, battery storage, and multiple non-discriminatory pricing rules, using both large-scale synthetic simulations and Swiss smart-meter demand data. The experiments show that Multi-Agent Neural Network Control learns effective decentralised policies, reduces community cost relative to rule-based control, and outperforms Multi-Agent Proximal Policy Optimization in the studied settings in both cost and training efficiency. The learned policies exhibit emergent, economically sensible behaviour: they recover standard storage arbitrage and, under price-aware learning, reveal capacity withholding under quantity-sensitive clearing, which benefits flexible agents at the expense of aggregate community cost. Embedding market clearing in the training graph therefore provides both an efficient learning signal and a diagnostic tool for strategic behaviour in local electricity-market design.