A Prognostics and Health Monitoring Framework for Self-Humidified Proton Exchange Membrane Fuel Cell Stacks

Abstract

Proton exchange membrane fuel cells (PEMFCs) systems are of considerable interest in clean energy research because of their advantages. Self-humidified PEMFCs, specifically, are noteworthy since they can operate without many of the auxiliary components present in traditional PEMFC configurations. However, cost, performance issues, and short lifetimes limit their widespread usage, prompting the need for a prognostics and health monitoring (PHM) framework. Novel methods are developed and compared against those in literature to address issues specific to self-humidified PEMFCs. The framework is developed using the following steps: -Find suitable features for health diagnosis: Health monitoring (HM) strategies applied to externally humidified PEMFC systems are compared to determine their suitability to self-humidified systems. A relative health scale is also developed to address the absence of suitable state-of-health definitions for self-humidified PEMFCs operating in real-world applications. -Develop an HM framework for the stack: A data-driven framework based on Gaussian process regression (GPR) is developed. The framework provides estimations for the system’s current health. The framework can provide uncertainty measurements, and variational learning is used to reduce the associated computational cost. An alternative model is also developed that can work with less training data. -Develop a hybrid probabilistic prognostics methodology: Data from the HM framework is repurposed to build a steady-state diagnostics (SSD) model of the stack. The SSD model can adapt to the highly fluctuating performance of self-humidified stacks. This SSD model is then used to provide basic prognostics predictions, which, when combined with a modified Gaussian process–Long short-term memory (GP-LSTM) network, forms a powerful generative prognostics model suitable for self-humidified PEMFC systems operating under dynamic loads. The hybrid model also provides an uncertainty estimate.