Protocol aware machine learning for prediction and selection of flooding recovery strategies in open cathode PEM fuel cells without recovery transient measurements

Abstract

A protocol aware machine learning framework enables prediction and selection of flooding recovery strategies in open cathode PEM fuel cells without measuring the recovery transient. Four recovery modes, step-up, pulse-up, ramp-up, and short-circuit pulse, are characterized over 102 training and 24 blind validation runs. Recovery time is defined as the time to settle within a ±5 mV band of the final value. Using operating and protocol variables and a polarization derived cell state descriptor, a single Random Forest predicts, without mode specific rules, the settling band time (R 2 = 0.887, nMAE = 9.0%), settling voltage (R 2 = 0.897), peak voltage (R 2 = 0.874), and overshoot (R 2 = 0.622). An MLP branch conditioned on these Random Forest metrics visualizes the corresponding recovery curves. Objective weighted TOPSIS and VIKOR then convert the predicted metrics and the fan parasitic energy into a condition dependent recovery protocol recommendation.