Hybrid renewable energy communities (RECs) combining photovoltaics, battery storage, and hydrogen systems (electrolyzer and fuel cell) are a promising solution for local, low-carbon energy autonomy. However, their real-time optimal control is challenging due to strong nonlinearities in key components, such as fuel cell (FC) performance, leading to non-convex constraints and MINLP formulations that are computationally intractable for real-time applications. This work addresses this challenge by developing neural network-based surrogate models of nonlinear components tailored for Model Predictive Control, enabling MILP reformulation that ensures real-time feasibility while maintaining high solution quality and accurate system dynamics.