Vehicle-mounted scene fuel cell dynamics loss prediction method, device and medium

Abstract

The application discloses a fuel cell dynamics loss prediction method and device in a vehicle scene, and a medium, and relates to the field of new energy fuel cells.The method comprises the following steps: collecting target time domain signals and target frequency domain signals of a fuel cell in a vehicle in real time, wherein the target frequency domain signals comprise fixed-frequency impedance information at two target frequencies, the two target frequencies are 1.9953 Hz and 199.53 Hz respectively, and the fixed-frequency impedance information comprises the amplitude and phase angle of an impedance spectrum; taking the target time domain signals and the target frequency domain signals in a real-time time window as the input of a deep learning model, and obtaining the current prediction value of at least one dynamics loss of the fuel cell in the vehicle through the deep learning model. Compared with the case where only the target time domain signals are taken as the input of the deep learning model, the deep learning model has more comprehensive prediction basis, and thus more accurate prediction results can be given.

Description

Technical Field

[0001] This application relates to the field of new energy fuel cells, and in particular to a method, device and medium for predicting the dynamic loss of fuel cells in a vehicle-mounted scenario. Background Technology

[0002] Performance evaluation and condition monitoring of fuel cells are crucial for ensuring their efficient and stable operation. Electrochemical impedance spectroscopy (EIS), as a non-destructive electrochemical characterization technique, is widely used for fuel cell condition monitoring and performance analysis. Existing research shows that multiple target impedances can be obtained from the full-frequency impedance spectrum of a fuel cell (including impedance spectra at all frequencies) through analysis using an equivalent circuit model (ECM) or relaxation time distribution (DRT). These target impedances can represent different kinetic losses, such as proton transport loss, ion transport loss, and gas-water mass transport loss. However, in vehicle-mounted applications, there is currently no mature technology to achieve full-frequency impedance spectrum measurement of fuel cells (which can be achieved in a laboratory environment), making it impossible to predict kinetic losses based on the full-frequency impedance spectrum of fuel cells in vehicle-mounted scenarios. Therefore, current technologies tend to use deep learning models to predict multiple target impedances representing different kinetic losses based on the measured time-domain signal information. Since the information input to the deep learning model only contains time-domain signals, the deep learning model struggles to accurately predict the target impedances. Therefore, current methods for predicting kinetic losses in vehicle-mounted fuel cells suffer from insufficient prediction accuracy. Summary of the Invention

[0003] The purpose of this application is to provide a method, device, and medium for predicting the dynamic loss of fuel cells in a vehicle-mounted scenario, which can provide more accurate prediction results.

[0004] To achieve the above objectives, this application provides the following solution:

[0005] Firstly, this application provides a method for predicting the dynamic loss of a fuel cell in a vehicle-mounted scenario, including:

[0006] The system acquires target time-domain and target frequency-domain signals of the fuel cell in the vehicle in real time. The target time-domain signals include multiple parameters such as current, voltage, fuel cell temperature, anode inlet temperature, cathode inlet temperature, anode flow rate, cathode flow rate, anode back pressure, and cathode back pressure. The target frequency-domain signals include fixed-frequency impedance information at two target frequencies, namely 1.9953Hz and 199.53Hz. The fixed-frequency impedance information includes the amplitude and phase angle of the impedance spectrum.

[0007] The target time-domain signal and the target frequency-domain signal within the real-time time window are used as inputs to a deep learning model, and the current predicted value of at least one dynamic loss of the fuel cell in the vehicle is obtained through the deep learning model.

[0008] In a second aspect, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the fuel cell dynamic loss prediction method in the vehicle scenario described above.

[0009] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the fuel cell dynamic loss prediction method in the vehicle-mounted scenario described above.

[0010] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the fuel cell dynamic loss prediction method in the vehicle-mounted scenario described above.

[0011] According to the specific embodiments provided in this application, the following technical effects are disclosed:

[0012] This application provides a method, device, and medium for predicting the kinetic loss of fuel cells in a vehicle-mounted scenario. By using target time-domain and target frequency-domain signals within a real-time time window as the prediction basis for a deep learning model, the deep learning model can learn the mapping relationship between kinetic loss and multi-source data (target time-domain and target frequency-domain signals). Specifically, the target time-domain signals are selected from fuel cell current, voltage, fuel cell temperature, anode inlet temperature, cathode inlet temperature, anode flow rate, cathode flow rate, anode back pressure, and cathode back pressure, all of which are correlated with kinetic loss. Compared to using only the target time-domain signals as input to the deep learning model, the deep learning model in this application has richer and more comprehensive prediction basis, thus providing more accurate prediction results. Therefore, it solves the problem of insufficient prediction accuracy in current methods for predicting the kinetic loss of vehicle-mounted fuel cells. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1This is a flowchart illustrating a fuel cell dynamic loss prediction method in a vehicle-mounted scenario according to one embodiment of this application.

[0015] Figure 2 This is a schematic diagram of various target time-domain signals under different operating conditions in one embodiment of this application;

[0016] Figure 3 The correlation between the amplitude, phase angle and five types of dynamic losses at 51 frequency points in one embodiment of this application is shown.

[0017] Figure 4 This is a schematic diagram illustrating the processing of time-domain and frequency-domain signals using the sliding window method in one embodiment of this application;

[0018] Figure 5 This is an architecture diagram of a deep learning model in one embodiment of this application;

[0019] Figure 6 This is a schematic diagram illustrating the impact of different input sequence lengths on the model prediction results in one embodiment of this application;

[0020] Figure 7 This is a schematic diagram comparing the prediction results with and without constant-frequency impedance in the input of a fuel cell dynamic loss prediction method in a vehicle-mounted scenario according to an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] This application provides a method for predicting fuel cell kinetic losses in a vehicle-mounted scenario, which can be executed by an on-board electronic device, specifically a computer installed in the vehicle. The on-board electronic device acquires the target time-domain and target frequency-domain signals of the fuel cell in the vehicle in real time, thereby predicting the fuel cell kinetic losses. Alternatively, the method can be executed in conjunction with a server. The on-board electronic device acquires the target time-domain and target frequency-domain signals of the fuel cell in the vehicle in real time and sends them to the server via a communication network. The server then feeds back the fuel cell kinetic loss prediction results to the on-board electronic device, and can also feed them back to the vehicle's remote management platform. However, considering real-time performance, the method for predicting fuel cell kinetic losses in a vehicle-mounted scenario is preferentially executed by the on-board electronic device.

[0024] In one exemplary embodiment, such as Figure 1 As shown, a method for predicting fuel cell dynamic loss in a vehicle-mounted scenario is provided. Taking the execution of this method by an on-board electronic device as an example, it includes the following steps 110 to 120.

[0025] Step 110: Real-time acquisition of target time-domain signals and target frequency-domain signals of the fuel cell in the vehicle. The target time-domain signals include current, voltage, fuel cell temperature, anode inlet temperature, cathode inlet temperature, anode flow rate, cathode flow rate, anode back pressure, and cathode back pressure. The target frequency-domain signals include fixed-frequency impedance information at at least one target frequency. Fixed-frequency impedance refers to the impedance at a single frequency.

[0026] Step 120: Use the target time-domain signal and target frequency-domain signal within the real-time time window as input to the deep learning model, and obtain the current predicted value of at least one dynamic loss of the fuel cell in the vehicle through the deep learning model.

[0027] In the above steps, by using the target time-domain signal and target frequency-domain signal within the real-time time window as the prediction basis for the deep learning model, the deep learning model can learn the mapping relationship between dynamic loss and multi-source data (target time-domain signal and target frequency-domain signal). In this embodiment, at least several of the fuel cell current, voltage, fuel cell temperature, anode inlet temperature, cathode inlet temperature, anode flow rate, cathode flow rate, anode back pressure, and cathode back pressure, which are related to dynamic loss, are selected as target time-domain signals. These time-domain signals can all be collected by sensors in the vehicle, and all or several of these time-domain signals can be used as input to the deep learning model. Simultaneously, this embodiment also uses at least one fixed-frequency impedance information at a target frequency as the target frequency-domain signal input to the deep learning model. Compared to using only the target time-domain signal as input to the deep learning model, the target frequency-domain signal can also provide the deep learning model with information related to dynamic loss, giving the deep learning model a richer and more comprehensive prediction basis, and thus providing more accurate prediction results. Therefore, this solves the problem of insufficient prediction accuracy in current on-board fuel cell dynamic loss prediction methods.

[0028] In this embodiment, preferably, the target frequency domain signal includes fixed-frequency impedance information at two target frequencies; the two target frequencies are 1.9953Hz and 199.53Hz, and the fixed-frequency impedance information includes the amplitude and phase angle of the fixed-frequency impedance.

[0029] Considering that in automotive scenarios, only one or two specific frequencies of fixed-frequency impedance information are typically available, and the correlation strength between impedance information at different frequencies and dynamic loss varies, selecting impedance information frequencies that are strongly correlated with dynamic loss is crucial. In this embodiment, fixed-frequency impedance information at frequencies of 1.9953Hz and 199.53Hz is selected as the input to the deep learning model.

[0030] In this embodiment, at least one kinetic loss specifically includes ohmic loss, local proton transport loss, global proton transport loss, charge transfer loss, and mass transfer loss. These five kinetic losses are of significant value in fuel cell state monitoring and performance analysis, as they can directly participate in the internal state regulation of the fuel cell. For example, high-frequency ohmic loss is generally considered to be negatively correlated with the water content of the proton exchange membrane. By predicting ohmic loss and combining it with the hydrogen-air-thermal subsystem, the water content of the fuel cell membrane can be controlled, avoiding membrane dryness and flooding failures. Low-frequency mass transfer loss is generally considered to be related to the liquid water content in the gas diffusion layer and catalyst layer. Liquid water hinders gas transport. By predicting mass transfer loss and combining it with the hydrogen-air-thermal subsystem, the liquid water content in the gas diffusion layer and catalyst layer can be controlled, avoiding gas shortage failures.

[0031] In this preferred embodiment, the training dataset of the deep learning model includes target time-domain signals and target frequency-domain signals under different operating conditions; the different operating conditions include steady-state operating conditions, dynamic operating conditions, membrane dry fault operating conditions, water flooding fault operating conditions, and air shortage fault operating conditions. By training the deep learning model with data under different operating conditions, the deep learning model can have strong robustness and can provide high prediction accuracy even when the vehicle is under any of the above operating conditions.

[0032] The following example will illustrate in detail the selection of the target frequency and the construction process of the training dataset.

[0033] (1) Time-frequency signal acquisition and impedance analysis.

[0034] First, offline tests were conducted on the fuel cell under steady-state, dynamic, and three fault conditions (membrane dryness, flooding, and gas shortage). In each test condition, nine time-domain signals of the fuel cell were acquired in real time: current... I ,Voltage V Fuel cell temperature T cell Anode inlet temperature T an Cathode inlet temperature T ca Anode flow rate ξ an Cathode flow rate ξ ca Anode back pressure P an Cathode back pressure P a In addition, an electrochemical workstation was used to acquire amplitude and phase angle data of the impedance spectrum across the entire frequency band from 0.1 to 10 kHz. Figure 2 The figures show different time-domain signals under different operating conditions. Among them, (a), (b), (c), and (d) show the current, temperature (including fuel cell temperature, anode inlet temperature, and cathode inlet temperature), metering ratio (including the metering ratio converted from anode flow rate and cathode flow rate), and gas pressure (including anode back pressure and cathode back pressure) under steady-state conditions, respectively; (e), (f), (g), and (h) show the current, temperature, metering ratio, and gas pressure under dynamic conditions, respectively; (i), (j), (k), and (l) show the current, temperature, metering ratio, and gas pressure under a fault condition, respectively.

[0035] Then, the impedance spectra at 51 frequency points from 0.1 to 10 kHz were analyzed using the relaxation time distribution (DRT). In this example, it was found that most fuel cell impedance spectra exhibited four characteristic peaks, representing four kinetic processes. Therefore, a fourth-order RC equivalent circuit was used to fit the impedance spectrum, resulting in...R 0、 R 1. C 1. R 2. C 2. R 3. C 3. R 4 and C There are a total of 9 parameters. Among them, R 0、 R 1. R 2. R 3. R 4 represents ohmic loss. R o Local proton transport loss R pt1 Global proton transport loss R pt2 Charge transfer loss R ct and mass transfer loss R mt Predicting these five types of kinetic losses can lay the foundation for identifying and regulating the internal state of fuel cells.

[0036] (2) Complex correlation analysis between frequency domain signal and dynamic loss.

[0037] Analyze in sequence P One sample (fuel cell in) P Impedance spectra of the entire frequency band from 0.1 to 10 kHz under different operating conditions) M Amplitude, phase angle, and corresponding ohmic loss at different frequency points R o Local proton transport loss R pt1 Global proton transport loss R pt2 Charge transfer loss R ct and mass transfer loss R mt The correlation between the five dynamic losses. To quantify the correlation, a feature tensor was constructed. This tensor contains P One sample in M The amplitude at different frequency points ( x 1) with phase angle ( x 2) For each frequency point Extract the feature matrix of all samples at this frequency:

[0038]

[0039] in, , Representing samples respectively p At frequency point m The amplitude and phase angle.

[0040] Five types of dynamic loss matrices (Including the actual values ​​of the five dynamic losses corresponding to all samples) are then z-score standardized.

[0041]

[0042] in, y The five dynamic loss matrices are standardized by z-score (containing the standard values ​​of the five dynamic losses for all samples). The average of the actual values ​​of the five dynamic losses corresponding to all samples; The standard deviation of the actual values ​​of the five kinetic losses for all samples.

[0043] Then, for each frequency point The characteristic matrix at this frequency X freq,m Compared with the standardized first i Kinetic loss vector (Including the first corresponding to all samples) Least squares correlation analysis was performed on the standard value of the kinetic loss.

[0044]

[0045] in, Representing frequency point The next i The prediction vector for the dynamic loss includes frequency points. The first corresponding to all samples below i Predicted values ​​for dynamic losses.

[0046] Calculate each type of kinetic loss at the frequency point Correlation :

[0047]

[0048] in, Indicates sample The corresponding number Standard value of dynamic loss, Representing frequency point Lower sample The corresponding number i Predicted values ​​of dynamic losses, Represents the first corresponding to all samples iThe standard mean value of the dynamic loss.

[0049] The above calculations yield the correlation between the frequency domain signal (including amplitude and phase angle) at each frequency point and the five types of dynamic losses. . The closer it is to 1, the stronger the correlation between the amplitude and phase angle at the corresponding frequency point and this type of dynamic loss.

[0050] In this example, the correlation between the frequency domain signals at 51 frequency points and the five types of dynamic losses is calculated. Values ​​such as Figure 3 As shown in Table 1. In Figure 3 In the figure, the horizontal axis represents dynamic loss, and the vertical axis represents frequency.

[0051] Table 1. Correlation between frequency domain signals at different frequency points and five types of dynamic losses value

[0052]

[0053] Since existing online impedance measurement equipment can only measure a maximum of two frequency points, 1.9953Hz and 199.53Hz, which have the highest correlation with dynamic loss, were selected as the impedance measurement frequencies, i.e., the target frequencies. The specific reason is that these two frequencies have the highest correlation with high-frequency ohmic loss and low-frequency mass transfer loss, respectively, and also have a high complex correlation with local proton transport loss, global proton transport loss, and charge transfer loss.

[0054] (3) Construction of the training dataset for dynamic loss prediction and the input feature matrix of the deep learning model.

[0055] Nine time-domain signals (current) I ,Voltage V Fuel cell temperature T cell Anode inlet temperature T an Cathode inlet temperature T ca Anode flow rate ξ an Cathode flow rate ξ ca Anode back pressure P an Cathode back pressure P ca ) and 4 frequency domain signals (amplitude |Z) of 1.9953Hz 1.9953 | Phase Angle 1.9953 Amplitude of 199.53Hz | Z 199.53 | Phase Angle 199.53 and the corresponding Ohmic loss R o Local proton transport loss R pt1 Global proton transport loss R pt2 Charge transfer loss R ct and mass transfer loss R mt The data are combined to form a training dataset for predicting dynamic loss.

[0056] Then, adopt as follows Figure 4 The sliding window method shown processes 13 time and frequency input signals, with a sliding window width of [value missing]. w (Corresponding to the real-time time window in the actual prediction process), obtain the two-dimensional input feature matrix. X in (As a sample feature, the dynamic loss at the next time step after the sliding window is used as the corresponding sample label):

[0057]

[0058] The above is an example illustrating the selection of the target frequency domain signal and the construction of the training dataset.

[0059] To enable deep learning models to make predictions, they need to be trained using a training dataset. Furthermore, designing a suitable deep learning model plays a crucial role in predictive accuracy.

[0060] In order to extract meaningful features from the input (such as a two-dimensional input feature matrix) and make accurate dynamic loss prediction, this embodiment selects TCN-Transformer as the deep learning model base and innovatively integrates spatial channel squeezing and excitation (scSE) attention mechanism to form scsE-TCN-Transformer model, so as to improve the deep learning model's ability to capture important features.

[0061] Specifically, the deep learning model includes an encoder and a decoder. The encoder consists of two stacked feature extraction modules. Each feature extraction module includes a stacked temporal convolutional network (TCN) and a spatial and channel compression activation module (scSE). The input of the spatial and channel compression activation module is fused with the output of the temporal convolutional network to serve as the output of the feature extraction module. The decoder is a Transformer decoder.

[0062] The following example will be used to explain deep learning models in detail.

[0063] Reference Figure 5 In deep learning models, the model input is first processed by two stacked feature extraction modules (scSE-TCN modules) to extract key features. The extracted feature sequences are then input to the Transformer decoder (Transformer module), which outputs the final dynamic loss prediction result. The input of the previous feature extraction module is the input of the deep learning model, and the input of the next feature extraction module is the output of the previous feature extraction module. The output of the next feature extraction module is then used as the input of the Transformer decoder.

[0064] In the feature extraction module, the input to the temporal convolutional network is also the input to the feature extraction module. The temporal convolutional network uses causal convolution and dilated convolution to process temporal data. Causal convolution ensures that the output depends only on the current and past inputs, avoiding interference from future information; dilated convolution expands the receptive temporal domain, enabling deep learning models to process longer time series. Specifically, the temporal convolutional network includes stacked convolutional layers, normalized layers, dropout layers, and the ReLU activation function.

[0065] The convolutional layer can be represented as:

[0066]

[0067] in, X Represents the original input sequence; s Indicates the index of the input sequence; i Indicates the number of filters; d Indicates the expansion factor; n Indicates the filter size; f ( i ) indicates the first i The weights of each convolutional filter; X s-d·i This represents the input sequence corresponding to the receptive time domain determined by the expansion.

[0068] The Spatial and Channel Compression Activation Module consists of a Spatial Compression and Channel Activation (cSE) module and a Channel Compression and Spatial Activation (sSE) module. It enhances the feature representation capabilities of the spatial and channel dimensions in the convolutional neural network through compression and activation. The cSE module weights each channel of the feature map using an attention mechanism to learn the importance between channels; the sSE module weights the spatial dimension through convolution operations to learn the importance of spatial locations in the input feature map.

[0069] The calculation formula for the cSE module is:

[0070]

[0071] in, This indicates the output of the cSE module. This represents the output of a temporal convolutional network. and These represent two fully connected layers F respectively. C The weight matrix, This represents the sigmoid activation function.

[0072] The calculation formula for the sSE module is:

[0073]

[0074] in, This indicates the output of the sSE module. This represents the output of a temporal convolutional network. This represents the convolution operation. This represents the sigmoid activation function.

[0075] Output of the space and channel compression excitation module for:

[0076]

[0077] The input of the temporal convolutional network is fused with the output of the spatial and channel compression excitation module through convolutional skip connections, and then used as the output of the feature extraction module, which is the input of the Transformer decoder.

[0078] The core of the Transformer decoder lies in the application of self-attention, multi-head attention, and positional encoding (Pos-emb). Positional encoding enables the deep learning model to understand the relative position of each element in the sequence, thereby enhancing the deep learning model's ability to represent time series. Positional encoding is typically generated using sine and cosine functions. Specifically, the Transformer decoder includes stacked positional encoding layers, self-attention layers, one-dimensional indexing, fully connected layers, ReLU activation functions, and regression output layers. The input to the Transformer decoder, after positional encoding, serves as the input to the first self-attention layer.

[0079] The self-attention layer uses a self-attention mechanism to process its own input.

[0080] For a single-head self-attention mechanism, it uses three matrices—query, key, and value—to calculate the similarity between elements in its input and assign a weight to each element based on this similarity. Then, it uses these weights to perform a weighted sum of the values ​​to obtain the final output.

[0081]

[0082] in, Q For query matrix; K The key matrix; V It is a value matrix; d k This represents the dimension of the key matrix.

[0083] Multi-head self-attention mechanisms, by computing multiple self-attention heads in parallel, enable deep learning models to learn different relationships in different subspaces, thereby enhancing the performance of deep learning models and greatly improving training efficiency. Specifically, the multi-head self-attention mechanism divides the Q, K, and V matrices into multiple subspaces, then computes self-attention in each subspace, and finally concatenates the outputs of each head, generating the final output through a linear transformation. The formula is as follows:

[0084]

[0085]

[0086] in, , and They represent the first i The weights of the query matrix, key matrix, and value matrix. This represents a linear transformation matrix.

[0087] The above is an explanation of deep learning architecture.

[0088] In this embodiment, the training loss of the deep learning model includes the root mean square error between the predicted value of at least one dynamic loss and the label value. and mean absolute percentage error :

[0089]

[0090] in, and They represent the first i The predicted and actual values ​​of the dynamic loss.

[0091] In this embodiment, the preferred sliding window width is... w The width is 90 seconds.

[0092] Since different input sequence lengths may have a certain impact on the performance of deep learning models, in this embodiment, the sliding window width is... wThe input sequence length was gradually increased from 10 s to 120 s. During this process, the scsE-TCN-Transformer model and four benchmark models (LSTM, CNN, ResNet, TCN-Transformer) were trained. The RMSE of the prediction results of each deep learning dynamic loss prediction model under different input sequence lengths is as follows: Figure 6 As shown, for most input sequence lengths, the scSE-TCN-Transformer achieves better prediction performance than the four baseline models, exhibiting a smaller RMSE. Furthermore, the scSE-TCN-Transformer has the lowest RMSE in predicting dynamics loss when the input sequence length is 90 s. Therefore, 90 s is determined as the optimal input sequence length, i.e., the sliding window width. w =90 s.

[0093] Finally, to verify the advantages of fusing fixed-frequency impedance (impedance spectrum information at the target frequency) with time-domain signal input, the dynamic loss prediction performance of fusing 9 time-domain signals and 4 frequency-domain signals was compared with the dynamic loss prediction performance of using only 9 time-domain signal inputs. The results are as follows: Figure 7 As shown.

[0094] In this verification, by fusing the amplitude and phase angle of the fixed-frequency impedances at 1.9953Hz and 199.53Hz with the time-domain signal as input, the accuracy of dynamic loss prediction was significantly improved. Specifically, as... Figure 7 As shown, the MAPE of the five dynamic losses was reduced by 64.7%, 47.1%, 27.5%, 52.6% and 66.4% respectively, further demonstrating the advantages of fixed-frequency impedance and time-domain signal fusion input.

[0095] Furthermore, in this embodiment, nine time-domain signals are fused with two fixed-frequency impedance signals in different combinations and input into the proposed deep learning model. The RMSE of the dynamic loss prediction values ​​of the validation set is shown in Table 2. It can be found that the selected 1.9953 Hz and 199.53 Hz have significant advantages in improving the accuracy of dynamic loss prediction.

[0096] Table 2. Prediction accuracy of dynamic loss when nine time-domain signals are input to two fixed-frequency impedance signal models with different combinations.

[0097]

[0098] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0099] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0100] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0101] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0102] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0103] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchain. The processors involved in the embodiments provided in this application may be, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc.

[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting the dynamic loss of a fuel cell in a vehicle-mounted scenario, characterized in that, include: The system acquires target time-domain and target frequency-domain signals of the fuel cell in the vehicle in real time. The target time-domain signals include multiple parameters such as current, voltage, fuel cell temperature, anode inlet temperature, cathode inlet temperature, anode flow rate, cathode flow rate, anode back pressure, and cathode back pressure. The target frequency-domain signals include fixed-frequency impedance information at two target frequencies, namely 1.9953Hz and 199.53Hz. The fixed-frequency impedance information includes the amplitude and phase angle of the impedance spectrum. The target time-domain signal and the target frequency-domain signal within the real-time time window are used as inputs to a deep learning model, and the current predicted value of at least one dynamic loss of the fuel cell in the vehicle is obtained through the deep learning model. The at least one dynamic loss includes ohmic loss, local proton transport loss, global proton transport loss, charge transfer loss, and mass transfer loss.

2. The fuel cell dynamic loss prediction method in a vehicle-mounted scenario according to claim 1, characterized in that, The training dataset of the deep learning model includes target time-domain signals and target frequency-domain signals under different operating conditions; The different operating conditions include steady-state operating conditions, dynamic operating conditions, membrane dryness failure operating conditions, water flooding failure conditions, and air shortage failure conditions.

3. The fuel cell dynamic loss prediction method in a vehicle-mounted scenario according to claim 1, characterized in that, The deep learning model includes an encoder and a decoder. The encoder includes two stacked feature extraction modules. Each feature extraction module includes a stacked temporal convolutional network and a spatial and channel compression excitation module. The output of the spatial and channel compression excitation module is fused with the input of the temporal convolutional network to serve as the output of the feature extraction module.

4. The fuel cell dynamic loss prediction method in a vehicle-mounted scenario according to claim 3, characterized in that, The decoder is a Transformer decoder.

5. The fuel cell dynamic loss prediction method in a vehicle-mounted scenario according to claim 3, characterized in that, The training loss of the deep learning model includes the root mean square error and the mean absolute percentage error between the predicted value and the label value of the at least one dynamic loss.

6. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the fuel cell dynamic loss prediction method in a vehicle-mounted scenario as described in any one of claims 1-5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the fuel cell dynamic loss prediction method in the vehicle-mounted scenario as described in any one of claims 1-5.

8. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the fuel cell dynamic loss prediction method in the vehicle-mounted scenario as described in any one of claims 1-5.

Images