摘要
为了进一步提高传统深度学习方法预测燃料电池剩余使用寿命(RUL)的精度,该文提出了一种综合卷积神经网络(CNN)、双向门控循环单元(BiGRU)和注意力机制(AT)的混合模型。利用奇异谱分析对燃料电池衰减数据进行预处理、消除噪声并获取有效信息,CNN-BiGRU提取其时空特征、历史和未来信息,AT进一步探索时空相关性,并采用增强型鲸鱼优化算法(EWOA)对模型超参数进行优化。结果表明,与长短期记忆(LSTM)网络、CNN、GRU、CNN-GRU、CNN-BiGRU、BiGRU-AT、CNN-BiGRU-AT和其他算法优化的CNN-BiGRU-AT相比,在40%训练数据下,EWOA优化的CNN-BiGRU-AT模型其方均根误差(RMSE)、平均绝对误差(MAE)、平均绝对百分比误差(MAPE)和相对误差(RE)均最小,最小值分别为0.2021%、0.1278%、0.033%和0.027%。此外,该模型在缺失数据达60%的情况下仍能保持较强的鲁棒性,其最小RMSE、MAE、MAPE和RE分别为0.3879%、0.2559%、0.0811%和0.32%,具有较好的燃料电池剩余使用寿命预测性能。
Fuel cells(FCs)are widely used due to their high energy conversion rate,low noise level and no pollutant emissions.However,the internal components of FCs will irreversibly degrade over time,and under certain complex operating conditions,their aging rate will limit their long-term applications.Therefore,accurate prediction of the remaining useful life(RUL)of FCs is essential to extend their service time,reduce operating costs and ensure their durability.Currently,RUL prediction for FCs is classified into model-based,data-driven and hybrid prediction methods.Model-based prediction methods use the physicochemical reactions inside the FCs to create models and make predictions,with the advantage of obtaining decay parameters to characterize the internal aging state,but it is difficult to establish an accurate mechanistic model for the complex physicochemical reactions inside the FCs.Data-driven approaches do not rely on mechanistic models and can learn from large amounts of experimental test data to make accurate RUL predictions.However,some traditional deep learning models,such as long short term memory(LSTM)neural networks,recurrent neural networks(RNN),and gated recurrent unit(GRU)have obvious limitations.They rely only on historical data to predict the RUL of FC without considering the before and after information of FC degradation.In addition,they cannot effectively exploit the spatial correlation existing in the test data,and the accuracy of RUL prediction needs to be further improved.The hybrid prediction method is to merge or fuse various models to make full use of the respective advantages of different models to improve the prediction accuracy.However,it relies on the establishment of an accurate model,and the accuracy and robustness of the model are difficult to guarantee due to the complexity of the degradation mechanism of the FCs,coupled with environmental disturbances and measurement noise.In addition,hybrid methods based on multiple data-driven approaches need to annotate a large amount of data,consume a lot of computational resources and time,and it is difficult to ensure the interpretability of the prediction results.To address these challenges,this paper proposed a hybrid model combining with convolutional neural network(CNN),bidirectional gated recurrent unit(BiGRU)and attention mechanism(AT)to further improve the RUL forecasting accuracy of FCs.Firstly,the FCs’aging data recorded by the FCLAB Research Federation were preprocessed using singular spectrum analysis to eliminate noise and obtain effective information,the spatio-temporal features,historical and future information of FCs were extracted with CNN-BiGRU model,the spatio-temporal correlation was explored with AT,and the hyperparameters of the model were optimized with an enhanced whale optimization algorithm(EWOA)to reduce human intervention error.Subsequently,the root mean square error(RMSE),mean absolute error(MAE),mean absolute percentage error(MAPE),and relative error(RE)were designed to evaluate the RUL forecasting accuracy of the proposed hybrid model by comparing with other deep learning models such as LSTM,CNN,GRU,CNN-LSTM,CNN-GRU,CNN-BiGRU,BiGRU-AT,CNN-BiGRU-AT,and CNN-BiGRU-AT optimized with bayesian optimization(BO),WOA,sparrow search algorithm(SSA),grey wolf optimization(GWO)and black widow optimization algorithm(BWOA).The following conclusions can be drawn from the results:(1)Compared with LSTM,CNN,GRU,CNN-LSTM,CNN-GRU,CNN-BiGRU,BiGRU-AT,CNN-BiGRU-AT,BO-CNN-BiGRU-AT,WOA-CNN-BiGRU-AT,SSA-CNN-BiGRU-AT,GWO-CNN-BiGRU-AT and BWO-CNN-BiGRU-AT,the proposed EWOA-optimized CNN-BiGRU-AT model has the smallest RMSE,MAE,MAPE and RE,which is 0.2021%,0.1278%,0.033%and 0.027%,respectively.(2)The proposed model still maintains superior RUL prediction robustness with 60%missing data,the minimum RMSE,MAE,MAPE and RE are 0.3879%,0.2559%,0.0811%and 0.32%,respectively.(3)Compared with the CNN,GRU,CNN-GRU,CNN-BiGRU and CNN-BiGRU models,the EWOA-optimized CNN-BiGRU-AT model can more accurately describe the twelve-,twenty-four-,and forty-eight-step aging curves of the FCs with higher reliability.
作者
全睿
程功
周宇龙
章国光
全琎
Quan Rui;Cheng Gong;Zhou Yulong;Zhang Guoguang;Quan Jin(Hubei Key Laboratory for High-Efficiency Utilization of Solar Energy and Operation Control of Energy Storage System,Hubei University of Technology,Wuhan,430068,China;Wuhan Hyvitech Co.Ltd,Wuhan,430000,China)
出处
《电工技术学报》
北大核心
2025年第19期6342-6358,共17页
Transactions of China Electrotechnical Society
基金
国家自然科学基金(51977061,51407063)
武汉市东湖高新区揭榜挂帅项目(2024KJB336)资助。
关键词
燃料电池
剩余使用寿命
双向门控循环单元
注意力机制
增强型鲸鱼优化算法
Fuel cell
remaining useful life(RUL)
bidirectional gated recurrent unit(BiGRU)
attention mechanism(AT)
enhanced whale optimization algorithm(EWOA)
