Neural Network Surrogate for Multi-Criteria Coordinated Model Predictive Control of Vanadium Redox Flow Battery Stations With Bound Tightening

Abstract

State-of-charge (SOC) inconsistency among units in vanadium redox flow battery (VRFB) stations causes voltage-limit violations and premature charge/discharge cutoffs, thereby degrading station-level energy utilization and dispatch performance. This paper proposes a station-level multi-objective coordinated model predictive control (MPC) strategy using a mixed-integer-representable ReLU neural surrogate and hybrid bound tightening. The main innovation is to integrate OCV-reference selection, SOC-equalising station-level MPC and feasibility-based/optimization-based bound tightening (FBBT + OBBT) into a unified control-oriented framework that preserves terminal-voltage safety whilst reducing online computational burden. First, an open-circuit-voltage (OCV) reference trajectory is selected at the single-stack level to operate in a high flow-rate sensitivity region and mitigate concentration polarization. Second, a station-level MPC formulation jointly tracks the OCV reference, equalises inter-cabin SOC and satisfies dispatch requirements under voltage, concentration, current and power constraints. Third, a ReLU-based surrogate replaces the nonlinear electrochemical model and is embedded into the MPC through Big-M constraints, whilst the bilinear power term is handled by McCormick envelopes. The hybrid FBBT + OBBT scheme tightens neuron bounds by 28.64%–59.24% and reduces preprocessing time by 14% compared with pure OBBT. Simulation results show that the proposed strategy shortens charging time by about 10%, improves voltage efficiency from 75.59% to 79.95%, improves energy efficiency from 71.20% to 74.64% and reduces station-level charge/discharge actions from 31 to 29 under the same 100 MWh dispatch target, demonstrating improved safety, efficiency, SOC consistency and operational economy.