Data-Driven Multistep Ahead and Long-Term Probabilistic Predictions for Automotive Fuel Cell Performance Degradation

Abstract

An accurate prediction of performance degradation is crucial for the timely maintenance of fuel cell. The influence of uncertain factors on automotive fuel cell makes it challenging for traditional point estimation methods to achieve reliable long-term predictive results. To address this, we have designed a data-driven probabilistic prediction model called MambaMixer, which enables multistep ahead and the long-term probabilistic prediction of performance degradation for automotive fuel cells without a mechanism model. The model first combines a dimensional-segmentwise(DSW) strategy to protect historical data. Second, the mamba model and multilayer perceptron (MLP) are used to design a two-stage transform (TST) layer. Then, the decomposition strategy is used to extract multiscale historical information, and the long-term probability prediction module is designed based on the student’s t-distribution. In the prediction experiment of relative power-loss rate (RPLR), our model demonstrates superior performance in multistep ahead prediction compared with three advanced models. In the long-term probabilistic prediction experiments, the effectiveness of the proposed algorithm is validated through multiple sets of experiments. In general, our model helps to improve the prediction performance of automotive fuel cell performance degradation, providing a new solution for data-driven model design.