Multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method

Abstract

The invention discloses a multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method, and belongs to the technical field of automation. The method comprises the following specific steps: acquiring degradation data as original input data from a fuel cell system according to working condition characteristics of the proton exchange membrane fuel cell, and smoothing by adopting a wavelet denoising method to reduce noise such as peaks in the data; an exponential decomposition module is adopted to decompose the preprocessed data into subsequences of different time scales, and an exponential decomposition sequence set is obtained; and finally, inputting the exponential decomposition sub-sequence set into a trained prediction model to obtain a voltage prediction result of a future time step, and using the voltage prediction result to predict the degradation of the proton exchange membrane fuel cell. According to the method, the data characteristics of the proton exchange membrane fuel cell under the dynamic working condition can be fully mined, and the prediction accuracy of the degradation data is effectively improved.

Description

Technical Field

[0001] The invention belongs to the field of automation technology, and in particular relates to a multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method. Background Art

[0002] Due to the massive consumption of fossil fuels around the world, which has led to serious energy crises and environmental pollution, people's demand for clean and renewable energy is increasing day by day. Hydrogen energy, as a clean and efficient green renewable energy, is considered to be an effective alternative to traditional fossil fuels. In order to make full use of hydrogen energy, people have conducted extensive research on proton exchange membrane fuel cells, an energy device. Proton exchange membrane fuel cells have the advantages of high energy conversion efficiency, low noise, and reliable operation, but their service life is still an important issue in large-scale applications. With the increase of operating time, the internal and external components of proton exchange membrane fuel cells will experience natural aging phenomena, such as membrane rupture, catalyst dissolution, gas erosion, etc. Frequently changing operating conditions, such as starting, stopping, idling, and dynamic load, will also accelerate the aging process. Based on the degradation mechanism of proton exchange membrane fuel cells, performance degradation prediction can prevent sudden failures, help energy storage systems make correct decisions, improve human operating habits, and ensure battery safety.

[0003] Proton exchange membrane fuel cells are complex nonlinear systems involving multiple variables, scales, and stages, so accurate dynamic modeling is challenging. Obviously, data-driven and hybrid methods can avoid complex mechanism modeling, thereby simplifying the establishment of models. However, the current methods still have the following problems: 1. Existing prediction methods cannot accurately predict the remaining life of proton exchange membrane fuel cells under dynamic operating conditions. 2. Existing prediction methods do not fully consider the correlation between the physical structure of proton exchange membrane fuel cells and battery operating variables when modeling. 3. Existing data-driven and hybrid prediction methods usually directly feed data into model learning, lacking analysis of the time level of data. 4. Existing data-driven and hybrid prediction methods tend to have more nonlinear links and multi-layer basic model combinations, poor model interpretability, and high computational complexity. Summary of the invention

[0004] The purpose of the present invention is to solve the defects of the prior art such as low prediction accuracy under dynamic conditions, inability to effectively mine multimodal information in working data, and complex model structure, and provide a multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method through data preprocessing, model training and online prediction to fill the research gaps in prediction accuracy, multi-scale time information mining and model structure optimization. According to the characteristics of degradation data such as nonlinearity and multi-spike waves, the present invention uses wavelet denoising method to reduce the influence of noise such as internal spike waves. In view of the voltage recovery phenomenon, a multi-scale exponential decomposition module is used to extract sequences of different time scales. A historical information fusion module consisting of internal decomposition, multi-scale time information update and multi-scale merging is established to extract deep dynamic nonlinear characteristics. Finally, the target voltage variable is extracted from the output to obtain the final output result.

[0005] In order to achieve the above-mentioned invention object, the present invention specifically adopts the following technical solutions:

[0006] A multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method comprises the following steps:

[0007] S1. using degradation data obtained from the fuel cell system as raw input data;

[0008] S2. Perform wavelet denoising on all variables in the original input data to obtain denoised data and construct an input data set;

[0009] S3. After batch processing and dividing the input data set, a single batch input is obtained, and the single batch input is decomposed using an exponential decomposition module, and each subsequence obtained by the decomposition is merged into an exponential decomposition subsequence set;

[0010] S4. Input the exponential decomposition subsequence set into the trained prediction model to obtain the voltage prediction result of the future time step for proton exchange membrane fuel cell degradation prediction;

[0011] In the prediction model, the coding layer first encodes the second dimension of each subsequence of the exponential decomposition subsequence set to obtain the coding result of each subsequence and construct a coding set. The coding set is processed by two historical information fusion modules in turn, and the outputs of the two historical information fusion modules are decoded and added to obtain the final output of the historical information fusion module. The last variable of the second dimension of the final output is extracted as the voltage prediction result.

[0012] Based on the above scheme, each step can be implemented in the following preferred specific manner.

[0013] Preferably, the specific implementation process of step S2 is as follows:

[0014] S21. Use db3 wavelet to decompose each variable in the original input data to obtain a wavelet coefficient sequence;

[0015] S22. Perform threshold processing on the wavelet coefficient sequence using a threshold function to obtain a processed wavelet coefficient sequence;

[0016] S23. Use the processed wavelet coefficients to perform inverse wavelet transform reconstruction to obtain denoised data, and construct an input data set from the denoised data.

[0017] Preferably, in step S3, an exponential decomposition module is used to decompose the single batch input X, which is implemented by the following formula:

[0018] X E =ExpDecomp(X)

[0019] Among them, the function ExpDecomp() performs sampling decomposition with an exponential of 2 on the first dimension of a single batch input.

[0020] Preferably, in step S4, the specific processing process of the coding layer is as follows:

[0021] The coding layer decomposes the subsequence set X E The second dimension of each subsequence in is encoded using the following formula:

[0022]

[0023] Wherein, Encoding() represents the encoding layer.

[0024] Preferably, in step S4, the specific processing process of the first historical information fusion module is as follows:

[0025] S41. Set the code Decomposition into long-term series and short-term series This is accomplished using the following formula:

[0026]

[0027] Among them, Padding() represents the encoding set The first dimension of the neutron sequence is filled with 0; AvgPool() kerneln Indicates that n different convolution kernels are used to average pool the padded subsequences; mean() means adding and averaging the different average pooling results;

[0028] S42. For the mth element in the long-term scale sequence The following formula is used to implement long-term scale aggregation, which is used to convert the m+1th element in the long-term scale sequence The information is aggregated into the mth element in the long-term scale sequence:

[0029]

[0030] Among them, TrendLinear() represents the long-term information extractor;

[0031] S43. For the mth element in the short-term scale sequence The following formula is used to implement short-term scale aggregation, which is used to aggregate the m-1th element in the short-term scale sequence The information is aggregated into the mth element in the short-term scale sequence:

[0032]

[0033] Among them, LocalLinear() represents a short-term information extractor;

[0034] S44. The long-term scale sequence after long-term scale aggregation and the short-term scale sequence after short-term scale aggregation are added together to obtain a fused sequence. The fused sequence is passed through a linear layer to obtain a fused sequence after dimension conversion. The fused sequence after dimension conversion is added to the encoding set to obtain the output of the first historical information fusion module.

[0035] Furthermore, in step S42, the long-term information extractor is implemented by a fully connected layer, a GELU and the fully connected layer.

[0036] Further, in step S43, the short-term information extractor is implemented by the fully connected layer, GELU and the fully connected layer.

[0037] Preferably, the final output is obtained as follows:

[0038]

[0039]

[0040] Among them, Linear m () represents the mth fully connected layer; Respectively represent the output of the first and second historical information fusion modules The mth subsequence in; M represents the number of subsequences; Y m Represents the decoding result of the mth subsequence.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] In order to overcome the defects of insufficient mining of fuel cell dynamic operating data features and low degradation data prediction accuracy in the prior art, the present invention proposes a multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method. The method of the present invention belongs to a deep learning method, which is based on a wavelet denoising method, a multi-scale exponential decomposition module and a historical information fusion module. Therefore, the method of the present invention can better play the advantages of each module and achieve complementary advantages. Under dynamic operating conditions, the prediction accuracy has been significantly improved. Among them, the wavelet denoising method is used to reduce the influence of internal noise such as internal spikes; the exponential decomposition module is used to extract sequences of different time scales from degradation data with voltage recovery phenomenon, which can assist subsequent modules to extract time information from degradation data of different scales; the historical information fusion module includes three parts: internal decomposition, multi-scale time information update, and multi-scale merging, which can further explore the deep nonlinear characteristics of the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of the steps of the present invention;

[0044] Figure 2 It is a flow chart of the implementation of the wavelet denoising method in the present invention;

[0045] Figure 3 It is a schematic diagram of the decomposition result of the exponential decomposition module in the present invention;

[0046] Figure 4 Schematic diagram of multi-scale time information update of the present invention; wherein (a) represents a schematic diagram of long-term scale aggregation, and (b) represents a schematic diagram of short-term scale aggregation;

[0047] Figure 5 It is a schematic diagram of the overall structure of the method of the present invention;

[0048] Figure 6 A comparison chart of the data results obtained by the method of the present invention and other common models on actual prediction data;

[0049] Figure 7 This is a comparison chart of the error results obtained by the method of the present invention and other common models on actual prediction data. DETAILED DESCRIPTION

[0050] In order to make the above-mentioned purpose, features and advantages of the present invention more obvious and easy to understand, the specific implementation mode of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in each embodiment of the present invention can be combined accordingly without conflicting with each other.

[0051] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for the purpose of distinguishing descriptions, and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features.

[0052] like Figure 1 As shown, in a preferred implementation of the present invention, the above multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method includes the following steps S1 to S4. The specific implementation process is described below.

[0053] S1. The degradation data obtained in the fuel cell system is used as the original input data X raw ∈R t×F , where t is the total length of the original input data and F is the input dimension of the original input data.

[0054] It should be noted that in step S1 of the present invention, the degradation data has been cleaned, and the time interval for collecting the degradation data each time is the same. The second dimension of the degradation data represents the external characteristics of the fuel cell, and the last variable is the voltage.

[0055] The embodiment of the present invention is based on a continuous dynamic operating condition data collected by a fuel cell system. The data set has 8 input dimensions and 3076 discrete time recording points. The 8 input dimensions are: output voltage, output power, operating temperature, operating current, hydrogen output pressure, air output temperature, air input temperature, and hydrogen output temperature.

[0056] S2. For the original input data X raw ∈R t×F All variables in are subjected to wavelet denoising to obtain the denoised data X in ∈R t×F And construct the input dataset.

[0057] It should be noted that if Figure 2 As shown, the specific implementation process of step S2 of the present invention is as follows:

[0058] S21. Use db3 wavelet to transform the original input data X raw For each variable X f ∈R t×1 ,f∈{0,…,F} is decomposed to obtain the wavelet coefficient sequence w fori .

[0059] In this embodiment, the original input data dimension is X raw ∈R 3076×8 , that is, t = 3076, F = 8. Let each variable be X f ∈R 3076×1 , f∈{0,…,8}, use db3 wavelet to transform X f After decomposition, the wavelet coefficient sequence is obtained.

[0060] S22. Use the threshold function to perform threshold processing on the wavelet coefficient sequence to obtain the processed wavelet coefficient sequence w f .

[0061] In this embodiment, the threshold function is selected as a soft threshold.

[0062] S23. Using the processed wavelet coefficients w f Perform inverse wavelet transform reconstruction to obtain the denoised data X in ∈R t×F , the input dataset is constructed from the denoised data.

[0063] In this embodiment, the dimension of the denoised data is X in ∈R 3076×8 .

[0064] S3. After batch processing of the input data set, a single batch input X∈R is obtained I×F , use the exponential decomposition module to decompose the single batch input X, and merge each subsequence obtained by the decomposition into an exponential decomposition subsequence set. Where I represents the first dimension of the single batch input X.

[0065] In the present invention, the exponential decomposition subsequence set X E ={X0,X1,…,X M} consists of M+1 subsequences X0, X1, …, X M The dimension of the mth subsequence is Where m∈{0,…,M} represents the index of each subsequence; M+1 represents the number of subsequences; X0,X1,…,X M represents the 0th,…,Mth subsequence.

[0066] In step S3, the exponential decomposition module is used to decompose the single batch input X, which is implemented by the following formula:

[0067] X E =ExpDecomp(X)

[0068] Among them, the function ExpDecomp() performs sampling decomposition with an exponential of 2 on the first dimension I of a single batch input X.

[0069] In this embodiment, after the input data set is batch processed and divided, a single batch input X∈R is obtained. 256×8 , that is, I = 256, F = 8. After using the exponential decomposition module to decompose the single batch input, a total of 5 subsequences X0, X1, ..., X M＝4 , the dimension of the mth subsequence is

[0070] In this embodiment, taking a sequence with a length of 512 as an example, the decomposition result of the exponential decomposition module is as follows: Figure 3 shown.

[0071] S4. Input the exponential decomposition subsequence set into the trained prediction model to obtain the voltage prediction result of the future time step for proton exchange membrane fuel cell degradation prediction;

[0072] In the prediction model, the coding layer first decomposes the subsequence set X into exponential form. E The second dimension of each subsequence is encoded to obtain the encoding result of each subsequence and construct the encoding set The coded set is processed by two historical information fusion modules in turn, and the outputs of the two historical information fusion modules are decoded and summed to obtain the final output Y∈R of the historical information fusion module P×F , extract the last variable Y of the second dimension of the final output out ∈P×1 is the voltage prediction result for the next P steps. represents the 0th,…,Mth subsequence X0,X1,…,X M The encoding result of the mth subsequence is

[0073] It should be noted that since the entire input data set is divided into several batches, the exponential decomposition module decomposes several inputs in each batch, and the prediction model completes the prediction. This process is repeated continuously to obtain the voltage prediction results of all batches of the input data set.

[0074] It should be noted that, in step S4 of the present invention, the specific processing process of the coding layer is as follows:

[0075] The coding layer decomposes the subsequence set X EThe second dimension of each subsequence in is encoded using the following formula:

[0076]

[0077] Wherein, Encoding() represents the encoding layer, which is used to change the exponential decomposition subsequence set X through fully connected layers of different lengths. E The second dimension of each subsequence in .

[0078] In this embodiment, the exponential decomposition subsequence set X E The second dimension of each subsequence in is expanded to F enc =32 dimensions, then the encoding dimension of the mth subsequence is

[0079] It should be noted that in step S4 of the present invention, the historical information fusion module includes three parts: internal decomposition, multi-scale time information update, and multi-scale merging. Taking the first historical information fusion module as an example, its specific processing process is as follows:

[0080] 1) The first part is internal decomposition, which is used to decompose the coding set.

[0081] S41. Set the code Decomposition into long-term scale series and short-term scale series is achieved using the following formula:

[0082]

[0083] Among them, Padding() represents the encoding set The first dimension of the neutron sequence is filled with 0; Indicates that n different convolution kernels are used to average pool the padded subsequences; mean() means adding and averaging the different average pooling results; represents the long-term scale series; Represents a short-term time series.

[0084] In this embodiment, F enc =32, so the dimensions of the long-term and short-term series are

[0085]

[0086] 2) The second part is multi-scale time information update, which is used to aggregate the information in the sequence from different directions of long-term and short-term, which specifically includes two steps: long-term scale aggregation and short-term scale aggregation.

[0087] S42. For the mth element in the long-term scale sequence The following formula is used to implement long-term scale aggregation, which is used to convert the m+1th element in the long-term scale sequence The information is aggregated into the mth element in the long-term scale sequence:

[0088]

[0089] Among them, TrendLinear() represents a long-term information extractor.

[0090] It should be noted that, in step S42, if Figure 4 As shown in (a), the long-term information extractor is implemented by a fully connected layer, GELU and the fully connected layer. The input dimension of the long-term information extractor is The output dimension is In this embodiment, F enc =32, so the input dimension of the long-term information extractor is The output dimension is

[0091] S43. For the mth element in the short-term scale sequence The following formula is used to implement short-term scale aggregation, which is used to aggregate the m-1th element in the short-term scale sequence The information is aggregated into the mth element in the short-term scale sequence:

[0092]

[0093] Among them, LocalLinear() represents a short-term information extractor.

[0094] It should be noted that, in step S43, if Figure 4 As shown in (b), the short-term information extractor is implemented by the fully connected layer, GELU and the fully connected layer. The input dimension of the short-term information extractor is The output dimension is In this embodiment, F enc =32, so the input dimension of the short-term information extractor is The output dimension is

[0095] 3) The second part, multi-scale merging, is used to combine long-term information with short-term information.

[0096] S44. Long-term scale series after long-term scale aggregation And the short-term scale series after short-term scale aggregation After adding, we get the fused sequence. After passing the fused sequence through the linear layer, we get the fused sequence after dimension conversion. We then add the fused sequence after dimension conversion to the encoding set. Add together to get the output of the first historical information fusion module Specifically, it can be implemented as follows:

[0097]

[0098] Wherein, Linear() represents the linear layer, which is used to The dimensions are converted to and encoded sets

[0099] Consistent.

[0100] It should be noted that the output of the first historical information fusion module is Then, according to the processing flow of the first historical information fusion module, the output of the first historical information fusion module is used as the input of the second historical information fusion module. After being processed by the second historical information fusion module, the output of the second historical information fusion module is obtained. Among them, HIIM1() and HIIM2() represent the first and second historical information fusion modules respectively; the mth subsequence in the output of the first and second historical information fusion modules is represented as and The two dimensions are

[0101] It should be noted that after obtaining the outputs of the first and second historical information fusion modules, the two can be decoded to obtain the final output, such as Figure 5 As shown, the details are as follows:

[0102]

[0103] Among them, Linear m () represents the mth fully connected layer, which is used to change the output of the first and second historical information fusion modules The mth subsequence in Dimensions; Respectively represent the output of the first and second historical information fusion modules The mth subsequence in; M represents the number of subsequences; Y m ∈R P×F Represents the decoding result of the mth subsequence.

[0104] In this embodiment, the decoding result Y of the mth subsequence m Dimension Y m ∈R 1×8 , Y m The final output Y∈R of the two historical information fusion modules can be obtained by adding 1×8 .

[0105] It should also be noted that the embodiments of the present invention also compare the method of the present invention with other common models. Figure 6 As shown in the figure, the data results obtained by the method of the present invention and other common models on actual prediction data, wherein MTIMNPrediction represents the method of the present invention, and Real represents the real result. Figure 6 The vertical axis is the voltage prediction result. Figure 7 As shown in FIG. 1 , the error results obtained by the method of the present invention and other common models on actual prediction data are shown, wherein MTIMN Error represents the error of the method of the present invention, Figure 7 The ordinate is the voltage error result Error, and the ordinate represents the test data point Test Point.

[0106] The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method, characterized in that: The following steps are involved: S1. using degradation data obtained from the fuel cell system as raw input data; S2. Perform wavelet denoising on all variables in the original input data to obtain denoised data and construct an input data set; S3. After batch processing and dividing the input data set, a single batch input is obtained, and the single batch input is decomposed using an exponential decomposition module, and each subsequence obtained by the decomposition is merged into an exponential decomposition subsequence set; S4. Input the exponential decomposition subsequence set into the trained prediction model to obtain the voltage prediction result of the future time step for proton exchange membrane fuel cell degradation prediction; In the prediction model, the coding layer first encodes the second dimension of each subsequence of the exponential decomposition subsequence set to obtain the coding result of each subsequence and construct a coding set. The coding set is processed by two historical information fusion modules in turn, and the outputs of the two historical information fusion modules are decoded and added to obtain the final output of the historical information fusion module. The last variable of the second dimension of the final output is extracted as the voltage prediction result.

2. The multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method according to claim 1, characterized in that: The specific implementation process of step S2 is as follows: S21. Use db3 wavelet to decompose each variable in the original input data to obtain a wavelet coefficient sequence; S22. Perform threshold processing on the wavelet coefficient sequence using a threshold function to obtain a processed wavelet coefficient sequence; S23. Use the processed wavelet coefficients to perform inverse wavelet transform reconstruction to obtain denoised data, and construct an input data set from the denoised data.

3. The multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method according to claim 1, characterized in that: In step S3, the exponential decomposition module is used to decompose the single batch input X, which is implemented by the following formula: X E =ExpDecomp(X) Among them, the function ExpDecomp() performs sampling decomposition with an exponential of 2 on the first dimension of a single batch input.

4. The multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method according to claim 1, characterized in that: In step S4, the specific processing process of the coding layer is as follows: The coding layer decomposes the subsequence set X E The second dimension of each subsequence in is encoded using the following formula: Wherein, Encoding() represents the encoding layer.

5. The multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method according to claim 1, characterized in that: In step S4, the specific processing process of the first historical information fusion module is as follows: S41. Set the code Decomposition into long-term series and short-term series This is accomplished using the following formula: Among them, Padding() represents the encoding set The first dimension of the neutron sequence is filled with 0; Indicates that n different convolution kernels are used to average pool the padded subsequences; mean() means adding and averaging the different average pooling results; S42. For the mth element in the long-term scale sequence The following formula is used to implement long-term scale aggregation, which is used to convert the m+1th element in the long-term scale sequence The information is aggregated into the mth element in the long-term scale sequence: Among them, TrendLinear() represents the long-term information extractor; S43. For the mth element in the short-term scale sequence The following formula is used to implement short-term scale aggregation, which is used to aggregate the m-1th element in the short-term scale sequence The information is aggregated into the mth element in the short-term scale sequence: Among them, LocalLinear() represents a short-term information extractor; S44. The long-term scale sequence after long-term scale aggregation and the short-term scale sequence after short-term scale aggregation are added together to obtain a fused sequence. The fused sequence is passed through a linear layer to obtain a fused sequence after dimension conversion. The fused sequence after dimension conversion is added to the encoding set to obtain the output of the first historical information fusion module.

6. The multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method according to claim 5, characterized in that: In step S42, the long-term information extractor is implemented by a fully connected layer, a GELU and the fully connected layer.

7. The multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method according to claim 5, characterized in that: In step S43, the short-term information extractor is implemented by the fully connected layer, GELU and the fully connected layer.

8. The multi-scale time information fusion proton exchange membrane fuel cell degradation prediction method according to claim 1, characterized in that: The final output is obtained as follows: Among them, Linear m () represents the mth fully connected layer; Respectively represent the output of the first and second historical information fusion modules The mth subsequence in; M represents the number of subsequences; Y m Represents the decoding result of the mth subsequence.

Images