• Electrolyzer's flexibility governed by hydrogen-to-oxygen limit and cell voltage is explicitly modeled. • Neural networks with high precision are proposed to approximate the nonlinear characteristics. • A stochastic day-ahead operation strategy considering price volatility is proposed for microgrid. Alkaline water electrolyzers (AWEs) can provide substantial operational flexibility to power systems by consuming renewable energy sources. However, in existing studies, such flexibility is usually modeled as constants or implicit constraints that cannot offer directly dispatchable information to system operators. To bridge this gap, this paper proposes a stochastic day-ahead operation strategy for integrated electricity-hydrogen microgrids with neural network surrogate model for electrolyzer’s flexibility, which could provide system operators with a dispatchable flexibility model while ensuring the robustness of operation subjected to price volatility. Firstly, the operation region of an AWE is formulated as explicit yet nonlinear upper and lower bounds, determined by the hydrogen‑to‑oxygen impurity limit and the safe cell voltage respectively. Then, a neural network surrogate model with ReLU active functions is proposed to approximate the formulated complex lower bound. Furthermore, the proposed surrogate model is reformulated as a set of mixed‑integer linear constraints through the big‑M method, which are embedded into the stochastic day-ahead operation model of an integrated electricity‑hydrogen microgrid. Case studies demonstrate that compared with existing flexibility models, the proposed operation strategy effectively enlarges the admissible load range of the AWE without compromising its safety. This enhancement improves the utilization of the electrolyzer and reduces unnecessary start-stops, ultimately increasing overall system profit by 32.3%. Moreover, the stochastic day-ahead operation enhances operational robustness under price volatility, ensuring the reliability of the proposed strategy under real-world uncertainties.
Tan et al. (Fri,) studied this question.