Deep Learning-based Serialized Model and Feature Extraction Algorithm for Solid Oxide Fuel Cell State Prediction

Deep learning-based methods have shown their wide application prospects in the field of solid oxide fuel cell(SOFC) prediction. However, the irrationality of the prediction object and the lack of prediction accuracy have hindered the progress of the prediction methods for SOFC systems. Therefore, two algorithms are proposed in this study, a feature selection algorithm based on maximum information coefficient and Bayesian information criterion for identifying key variables that are sensitive to changes in the SOFC system, and a prediction model based on a genetic algorithm optimized for sequence to sequence architecture. The M/BIC algorithm reduces the number of features by 62.5% (from 40 to 15), which contributes to the acceleration of the algorithm training and the subsequent sensor layout optimization. The prediction model incorporates a two-dimensional convolutional neural network and a multi-head attention mechanism to capture more information in the input sequence, an encoder-decoder architecture and a Gated recurrent unit to better predict performance trends. Tests with real experimental data show significant changes in feature dimensionality reduction and prediction accuracy.