Multi-agent reinforcement learning for reservoir cascade release scheduling: a climate-analytics evaluation of downstream low-flow equity under forecast uncertainty

Cascade reservoir operators must release water under uncertain inflow forecasts while satisfying competing demands for hydropower, irrigation and downstream environmental flows. The framework is formalised as a constrained Markov game with a lexicographic objective: maximise energy revenue subject to a probabilistic low-flow equity constraint at downstream reaches. A decomposed soft actor-critic architecture with a centralised equity critic is proposed, and a formal analysis establishes conditions under which the learned joint policy satisfies the equity constraint with high probability. The framework is evaluated conceptually against rule-based and single-agent benchmarks using a stylised three-reservoir cascade, with forecast uncertainty represented by ensemble inflow scenarios. The analysis identifies a structural trade-off between revenue and equity that widens as forecast uncertainty increases, and shows that equity-aware credit assignment reduces the frequency of severe low-flow violations relative to revenue-only learning. The contribution is a formalisation and algorithmic design for equity-aware MARL in cascade scheduling, together with a set of testable propositions about the conditions under which decentralised learning can deliver equitable low-flow outcomes.