A wavelet-PSO-ANN integrated framework for modeling thermal management of PCM-based PEM fuel cell stacks in cold climates

The implementation of passive thermal management strategies employing phase change materials (PCM) and insulation for maintaining optimal operating temperatures within polymer electrolyte membrane (PEM) fuel cell stacks under cold weather conditions is crucial for efficient power generation and durability, especially for electric vehicles. This study addresses a critical gap by proposing a novel framework for accurately predicting operating time and stack average temperature. The framework integrates stationary wavelet transforms (SWT), particle swarm optimization (PSO), and multilayer perceptron neural networks (MLPNN). Key input parameters considered include PCM type, ambient temperature, heat transfer coefficient, and the thicknesses of both PCM and insulation layers. To investigate the effect of wavelet transforms, models were developed with and without SWT for time and temperature prediction. Based on statistical criteria in testing stage, the PSO/MLPNN model (R = 0.99992), achieved superior performance in temperature prediction, while the PSO/SWT/MLPNN model with a decomposition level of 2 (R = 0.99833) excelled in time prediction. PSO effectively identified optimal hyperparameters, including wavelet elements, network architecture, and training parameters, for each specific model. Visualization and error analysis techniques revealed valuable insights into model accuracy. This research offers a significant advancement for optimizing PEM fuel cell performance by reducing experimental costs and simulation computations, enabling improved design and control strategies for clean energy devices such as electric vehicles operating in cold climates. • A novel PSO/SWT/ANN framework predicts PEM fuel cell performance in cold climates. • Key inputs: PCM type, ambient factors, and PCM/insulation layer thicknesses. • PSO optimized wavelet elements, network architecture, and training parameters. • PSO/ANN model achieved high accuracy in stack temperature prediction (R = 0.99992). • PSO/ANN/SWT (decomposition level 2) excelled in operating time prediction (R = 0.99833).