Forecasting the output performance of PEMFCs via a novel deep learning framework considering varying operating conditions and time scales

Abstract

Proton exchange membrane fuel cell (PEMFC) represents a significant technology for hydrogen energy conversion and are widely utilized in renewable energy systems. However, their performance tends to degrade over time during operation. Accurate prediction of PEMFCs performance is critical for optimizing hydrogen energy efficiency and ensuring the reliability of renewable energy systems. Meanwhile, the monitoring data collected from PEMFCs exhibit characteristics of diverse types, varying time resolutions, and distinct operating conditions, which complicate accurate predictions. To address this challenge, the feature-fusion and feature-attention blocks are developed to amalgamate interactive information and emphasize key features across various monitoring datasets. Based on the blocks, the feature-fusion and feature-attention deep learning (FFA-DL) framework that incorporates convolutional long short-term memory (ConvLSTM) networks is proposed. To validate the proposed framework, real-world data from two operation conditions, FC1 and FC2, are employed. The results demonstrate that the FFA-DL framework effectively extracts valuable information from complex monitoring data, thereby enhancing the accuracy of PEMFCs performance prediction. FFA-DL significantly enhanced prediction performance of the embedding models for both FC1 and FC2, and the FFA-enhanced ConvLSTM (FFA-ConvLSTM) outperformed other models with R² of 0.9631 and 0.9946 for FC1 and FC2, respectively. Additionally, the FFA-ConvLSTM exhibited excellent robustness and accuracy for data under varying time resolutions, with R² exceeding 0.9200 and 0.9800 for FC1 and FC2, respectively.