A Performance Prediction Method Applicable to Proton Exchange Membrane Fuel Cells under Dynamic Loads

Abstract

A performance prediction method for a proton exchange membrane fuel cell under dynamic load according to the present invention includes: inputting the voltage of each test step in a cycle of the fuel cell into an autoencoder network, and the target output of the autoencoder network is also the voltage; taking the output of the middle layer of the autoencoder network as the fuel cell health index; using the number of cycles and the fuel cell health index obtained in the first stage as the input of a long short-term memory network to predict the future change in the fuel cell health state; inputting the predicted fuel cell health state into the decoder of the autoencoder network to obtain the predicted voltage. The present invention has good prediction performance under different dynamic load conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of proton exchange membrane fuel cells, and particularly relates to a performance prediction method applicable to proton exchange membrane fuel cells under dynamic loads. Background Art

[0002] A proton exchange membrane fuel cell is an efficient, high-power density, and pollution-free energy conversion device. Although proton exchange membrane fuel cells have been applied in fields such as automobiles and ships, their large-scale popularization and application have been restricted by durability. During the operation of a proton exchange membrane fuel cell, its core component, the proton exchange membrane, will suffer irreversible aging. In order to formulate a reasonable life extension control strategy, it is first necessary to accurately predict the future performance changes of the fuel cell. Currently, the performance prediction methods applied to fuel cells under dynamic loads can be divided into two categories, namely direct prediction and indirect prediction.

[0003] The direct prediction method directly uses measured data (including current, voltage, temperature, etc.) for future voltage prediction. Zuo (Method 1) et al. proposed a voltage prediction method under dynamic load conditions based on deep learning, in which the voltage data of the fuel cell was directly applied to a gated recurrent neural network introducing an attention mechanism. Under dynamic loads, the fluctuations in battery voltage bring great difficulties to the training of the neural network, which will affect the prediction accuracy of the direct prediction method.

[0004] For the indirect prediction method, it is first necessary to extract the health index of the fuel cell from the measurement data, and then predict the future health state of the battery through a prediction method. Yue (Method 2) et al. proposed a health index extraction method based on a polarization curve model, and then predicted the future health state of the battery through an echo state network. However, this health index extraction method contains a large number of assumptions, which will cause the extracted health index not to fully reflect the health state of the battery, thus affecting the final prediction effect. Summary of the Invention

[0005] In order to overcome the deficiencies of the above-mentioned existing performance prediction methods, the purpose of the present invention is to propose a performance prediction method applicable to proton exchange membrane fuel cells under dynamic loads, which is a new fuel cell performance prediction method and can have higher prediction accuracy compared with Method 1 and Method 2 under different dynamic load conditions.

[0006] To achieve the above purpose, the present invention provides the following technical solutions:

[0007] A performance prediction method applicable to proton exchange membrane fuel cells under dynamic loads, comprising the following steps:

[0008] Step 1: Input the voltage X of each test step in a cycle of the fuel cell into the autoencoder network, and the target output of the autoencoder network is also the voltage X; take the output of the middle layer of the autoencoder network as the fuel cell health indicator;

[0009] Step 2: Use the number of cycles and the fuel cell health indicator obtained in the first stage as the input of the long short-term memory network to predict the future change of the fuel cell health state;

[0010] Step 3: Input the predicted fuel cell health state into the decoder of the autoencoder network to obtain the predicted voltage.

[0011] Furthermore, Step 1 includes:

[0012] (1) Extract the last voltage value in each test step from the original data as the first data set, and separately extract the 0A voltage and the maximum load voltage in each cycle in the first data set to form the second data set;

[0013] (2) Use the data from 0 to 600 hours as the training data, and set the time window length to 50 hours;

[0014] (3) Design an autoencoder network A1 with the number of neurons 35-20-10-1-10-20-35 and an autoencoder network A2 with the number of neurons 2-20-10-1-10-20-2 for the training data of the first data set and the second data set respectively; extract a fuel cell health indicator from each cycle through the trained autoencoder.

[0015] Furthermore, Step 2 includes: Take the moving window length as 50, train the fuel cell health indicators extracted from the first and second data sets with two long short-term memory networks respectively. The prediction process adopts iterative prediction, and the network is updated by the moving window method to obtain the health indicator prediction value H1 representing the fuel cell health state and the health indicator prediction value H2.

[0016] Furthermore, Step 3 includes: Input the predicted health indicator prediction values H1 and H2 into the decoders of the autoencoder network A1 and the autoencoder network A2 respectively to reconstruct the predicted voltage values.

[0017] Beneficial effects:

[0018] The fuel cell performance prediction method of the present invention can have good prediction performance under different dynamic load conditions. The present invention proposes a new method for constructing a health index based on an autoencoder. Compared with Method 1, this method converts the complex and fluctuating voltage into a stable health index, which can not only better characterize the health state of the fuel cell, but also greatly reduce the training difficulty of the neural network. Compared with Method 2, this method is a data-driven method, and there is no human intervention in the generation process of the health index. The generated index can better reflect the changes in the original data, so the final prediction result is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flowchart of the performance prediction method applicable to proton exchange membrane fuel cells under dynamic load of the present invention.

[0020] Figure 2 It is a schematic diagram of iterative prediction.

[0021] Figure 3 It is a schematic diagram of moving window prediction.

[0022] Figure 4 It is the prediction result of the health index of this method applied to the first data set.

[0023] Figure 5 It is the prediction result of the health index of this method applied to the second data set. DETAILED DESCRIPTION OF THE INVENTION

[0024] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0025] The performance prediction method applicable to proton exchange membrane fuel cells under dynamic load of the present invention extracts the health state of the battery through an autoencoder and uses a long short-term memory recurrent neural network to predict the future voltage of the battery.

[0026] The performance prediction method of the present invention is as Figure 1 shown. First, the health index of the fuel cell is extracted, that is, an autoencoder network is designed. The size of its input and output is equal to the number of different loads in a cycle, and the size of the middle layer of the autoencoder is 1. The voltage values under different loads in a cycle during the operation of the proton exchange membrane fuel cell are used as the input data of the autoencoder, and the output of the middle layer of the autoencoder is used as the health index of the fuel cell.

[0027] Secondly, the predicted health indicators of the fuel cell are predicted based on the long short-term memory network, that is, the extracted fuel cell health indicators and the number of cycles are used for iterative prediction in the long short-term memory network. The iterative prediction is as follows Figure 2 shown, that is, the predicted result of the new fuel cell health indicator will be re-input into the network, and finally the fuel cell health indicator will be converted into the predicted voltage through the decoder. Figure 2 In the figure, HI represents the health indicator, Cycle represents the number of cycles, HI' represents the predicted value of the health indicator, and U' represents the predicted value of the voltage.

[0028] Finally, reconstruct the voltage value. In order to reduce the influence of the fuel cell characterization test on the prediction result, the moving window prediction method shown in Figure 3 is adopted during the prediction, that is, only the data for the next time window (N hours) is predicted after each training. After sampling the data for these N hours, these data will be added to the training set to update the network, and then the voltage for the next N hours will be predicted. All the prediction results are used as the final result of the proposed algorithm.

[0029] Next, the technical solutions in the embodiments of the present invention will be specifically described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0030] The experimental data in the embodiment comes from a public dataset. A single proton exchange membrane fuel cell stack operates under the New European Driving Cycle condition. Each cycle of this condition lasts for 1181 seconds and contains 35 test steps. The durability test of the fuel cell lasts for about 1000 hours, and this original dataset is called the first dataset. In addition, the 0A voltage and the maximum load voltage in each cycle of the first dataset are extracted to simulate the voltage change under the start-stop cycle condition, and this dataset is called the second dataset.

[0031] As Figure 1 shown, the present invention provides a performance prediction method applicable to proton exchange membrane fuel cells under dynamic loads, including the following steps:

[0032] The first stage: Input the voltage X of each test step in a cycle of the fuel cell into the autoencoder network, and the target output of the autoencoder network is also the voltage X; take the output of the middle layer of the autoencoder network as the fuel cell health indicator;

[0033] The second stage: Use the number of cycles and the fuel cell health indicator obtained in the first stage as the input of the long short-term memory network to predict the future change of the fuel cell health state;

[0034] The third stage: Input the predicted fuel cell health state into the decoder of the autoencoder network to obtain the predicted voltage.

[0035] The specific steps of the method are as follows:

[0036] (1) Extract the last voltage value in each test step from the original data as the first data set, and separately extract the 0A voltage and the maximum load voltage in each cycle in the first data set to form the second data set;

[0037] (2) Use the data from 0 to 600 hours as the training data, and set the time window length to 50 hours;

[0038] (3) Design an autoencoder network A1 with the number of neurons 35-20-10-1-10-20-35, and an autoencoder network A2 with the number of neurons 2-20-10-1-10-20-2, which are respectively used for the training data of the first data set and the second data set; through the trained autoencoder, a fuel cell health index can be extracted from each cycle.

[0039] (4) Take the moving window length as 50, and train the fuel cell health indexes extracted from the first and second data sets with two long short-term memory networks respectively. The prediction process adopts the iterative prediction as shown in Figure 2 and updates the network in the moving window manner as shown in Figure 3 to obtain the health index prediction value H1 and the health index prediction value H2; the moving window manner is specifically: after training with the data from 0 to 600 hours, predict the voltage data for the next 50h through the iterative prediction of Figure 2 , and take this prediction result as part of the final prediction result; then use the data from 0 to 650 hours to train a new network to predict the voltage between 650 and 700 hours, and add this result to the final prediction result; repeat this process until all the prediction results from 600 to 1000 hours are obtained.

[0040] (5) As shown in Figure 1 , Figure 2 , input the predicted health index prediction value H1 and the health index prediction value H2 into the decoders of the autoencoder network A1 and the autoencoder network A2 respectively to reconstruct the predicted voltage value.

[0041] The fuel cell health index prediction results of the first data set and the second data set are as shown in Figure 4 , Figure 5 . It can be seen from the figure that the present invention can accurately predict the performance of proton exchange membrane fuel cells under the new European driving cycle condition and the start-stop cycle condition.

[0042] Table 1 and Table 2 list the comparison of the prediction results of the performance prediction method of the present invention with those of Method 1 and Method 2 under the same data conditions. Since the present invention adopts an autoencoder network with higher robustness and effectiveness to extract the fuel cell health index, higher prediction accuracy can be obtained.

[0043] Table 1

[0044] First data set Method 1 Method 2 This method Root mean square error 0.015518 0.013400 0.007956 Mean absolute error 0.012016 0.010697 0.005240

[0045] Table 2

[0046] Second data set Method 2 This method Root mean square error 0.018553 0.008875 Mean absolute error 0.015480 0.005578

[0047] It is easy for those skilled in the art to understand that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A performance prediction method applicable to proton exchange membrane fuel cells under dynamic loads, characterized in that It includes the following steps: Step 1: Input the voltage X of each test step in a cycle of the fuel cell into the autoencoder network, and the target output of the autoencoder network is also the voltage X; use the output of the middle layer of the autoencoder network as the fuel cell health indicator; Step 2: Use the number of cycles and the fuel cell health indicator obtained in the first stage as the input of the long short-term memory network to predict the future change of the fuel cell health state; Step 3: Input the predicted fuel cell health state into the decoder of the autoencoder network to obtain the predicted voltage.

2. The performance prediction method for a proton exchange membrane fuel cell under dynamic load according to claim 1, wherein The said Step 1 includes: (1) Extract the last voltage value in each test step from the original data as the first data set, and separately extract the 0A voltage and the maximum load voltage in each cycle in the first data set to form the second data set; (2) Use the data from 0 to 600 hours as the training data and set the time window length to 50 hours; (3) Design an autoencoder network A1 with the number of neurons being 35-20-10-1-10-20-35 and an autoencoder network A2 with the number of neurons being 2-20-10-1-10-20-2, which are respectively used for the training data of the first data set and the second data set; extract a fuel cell health indicator from each cycle through the trained autoencoder.

3. A performance prediction method for a proton exchange membrane fuel cell under dynamic load according to claim 2, characterized in that The said Step 2 includes: Take the moving window length as 50, train the fuel cell health indicators extracted from the first and second data sets respectively with two long short-term memory networks, adopt iterative prediction in the prediction process, and update the network in the way of moving window to obtain the predicted health indicator value H1 representing the fuel cell health state and the predicted health indicator value H2.

4. A performance prediction method for a proton exchange membrane fuel cell under dynamic load according to claim 3, characterized in that, The said Step 3 includes: Input the predicted health indicator values H1 and H2 into the decoders of the autoencoder network A1 and the autoencoder network A2 respectively to reconstruct the predicted voltage values.

Images