Multivariable-based fuel cell residual life prediction method and system

Abstract

The invention discloses a multivariable fuel cell residual life prediction method and system, and the method comprises the steps: collecting voltage, current, temperature and residual life data through a sensor, and decomposing the data into a plurality of intrinsic mode functions through an empirical reduced-order variational mode decomposition algorithm; carrying out time-frequency transformation on the first intrinsic mode function to generate a time-frequency matrix, and converting the time-frequency matrix into a two-dimensional image of a Gramer angle field; inputting the intrinsic mode function and the two-dimensional image information into a time perception graph convolutional recursive network model for training, optimizing hyper-parameters in combination with a neural population dynamic optimization algorithm to obtain a prediction model, and finally displaying a prediction result through a front end; according to the method, the complex signal decomposition precision is improved, the non-stationary and non-linear signal processing capability is enhanced, and high-precision fuel cell life prediction is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of fuel cells, and particularly relates to a method and system for predicting the remaining life of a fuel cell based on multiple variables. Background Art

[0002] As an efficient and environmentally friendly new energy power generation technology, fuel cells show broad application prospects in fields such as transportation and distributed energy. With the rapid development of fuel cell technology, its operation safety and life prediction have become the focus of the industry. In recent years, research institutions and enterprises at home and abroad have actively explored methods for evaluating the life of fuel cells to improve their reliability and economy. In industrial applications, predicting the remaining life of fuel cells is of great significance for optimizing maintenance strategies and reducing operating costs, and has become one of the important directions of industry technology research and development.

[0003] Currently, there are mainly three methods for predicting the remaining life of fuel cells: data-driven, model-driven, and hybrid models. The data-driven method uses machine learning algorithms (such as LSTM, CNN, UPF, etc.) to analyze operation data and mine the law of life decline; the model-driven method is based on electrochemistry mechanism modeling and predicts life through parameter changes (such as electrochemically active area, exchange current density); the hybrid model combines the advantages of data-driven and model-driven to improve prediction accuracy through semi-empirical models or data fusion. In addition, emerging technologies such as RVMD, Gramian Angular Field, and TGCRN models have further improved the spatio-temporal data processing ability, providing new technical approaches for life prediction.

[0004] Although the existing methods have made certain progress, the parameters of most prediction models are fixed and lack the ability of dynamic adjustment. In actual operation, the working conditions of fuel cells are complex and changeable, and traditional models are difficult to adaptively optimize parameters, resulting in limited prediction accuracy and robustness, which restricts the real-time performance and accuracy of fuel cell life prediction. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to provide a method for predicting the remaining life of a fuel cell based on multiple variables that can improve prediction accuracy and robustness; on the other hand, to provide a system for predicting the remaining life of a fuel cell based on multiple variables.

[0006] Technical Solution: The method for predicting the remaining life of a fuel cell based on multiple variables according to the present invention includes the following steps:

[0007] (1) Use sensors to collect voltage, current, temperature, and remaining life data of the fuel cell respectively, providing comprehensive and high-precision original data support for subsequent analysis, and ensuring the integrity and reliability of the input information of the prediction model;

[0008] (2) The empirical reduced-order variational mode decomposition algorithm is used to decompose the voltage, current, temperature, and remaining life data variables, obtaining multiple intrinsic mode functions, which can effectively improve the decomposition accuracy of the algorithm and enhance the adaptability to complex signals;

[0009] (3) Perform time-frequency transformation on the first intrinsic mode function of each variable to generate a time-frequency distribution matrix, enhancing the distinguishability of key degradation information and laying a foundation for subsequent feature fusion;

[0010] (4) Use the Gram angular field to transform the time-frequency distribution matrix into a Gram matrix and generate two-dimensional image information to achieve the visual representation of time series signals, facilitating the extraction of richer spatio-temporal degradation features by the deep learning model and improving the pattern recognition ability;

[0011] (5) Input the intrinsic mode functions and two-dimensional image information into the time-aware graph convolutional recurrent network model for training, and use the neural population dynamics optimization algorithm to optimize the hyperparameters of the model to obtain a fuel cell remaining life prediction model, significantly improving the prediction ability of the model;

[0012] (6) Based on the trained and optimized prediction model, output the fuel cell remaining life prediction result and display it through the front-end interface to support users' real-time monitoring and decision-making, improving the operation and maintenance efficiency of the fuel cell system.

[0013] Preferably, step 2 includes:

[0014] (201) Use the reduced-order variational mode decomposition algorithm to decompose the original signal into the preliminary mode u k and the central frequency ω k , and the reduced-order variational mode decomposition process is:

[0015]

[0016] where x(t) is the original signal, u k (t) is the k-th mode, ω k is the central frequency, K is the number of decomposition layers, and the preliminary mode set

[0017] (202) Perform empirical mode decomposition on each preliminary mode to obtain the refined mode and the residual, and the empirical mode decomposition process is:

[0018]

[0019] where e max (t) is the maximum envelope, e min (t) is the minimum envelope, m k (t) is the mean envelope, As a temporary candidate intrinsic mode function, if is satisfied, the formula loop stops. If not, repeat the above formula steps until the residual is a monotonic function or a constant, and decompose each into and the residual r k (t);

[0020] (203) Fuse the decomposition results of the reduced-order variational mode and the empirical mode to construct a hybrid optimization objective function:

[0021]

[0022] where α is the bandwidth parameter of the reduced-order variational mode, β is the sparse constraint term of the empirical mode, γ is the consistency weight of the decomposition results of the reduced-order variational mode and the empirical mode to prevent over-decomposition, and M is the number of layers of the empirical mode decomposition;

[0023] (204) Based on the fused modal reconstructed signal, introduce an adaptive weighting coefficient:

[0024]

[0025] where, is the final reconstructed signal, and ω c is the introduced adaptive weighting coefficient.

[0026] By introducing the bandwidth constraint of the reduced-order variational mode decomposition, the sparsity constraint of the empirical mode decomposition, and the modal consistency constraint, the problems of modal aliasing and over-decomposition in the traditional method are effectively solved. Finally, an adaptive weighting coefficient is used for signal reconstruction, which not only retains the mathematical rigor of the reduced-order variational mode decomposition but also integrates the ability of the empirical mode decomposition to capture non-linear features, enabling the intrinsic mode function components after decomposition to more accurately represent the characteristic information of different degradation stages of the fuel cell, laying a high-quality signal foundation for subsequent time-frequency analysis and life prediction.

[0027] Preferably, step 3 includes:

[0028] (301) Perform a Fourier transform on the first intrinsic mode function to obtain a discrete-time signal x(t), that is, x[n]. The expression of the discrete time-frequency transform of this time series is:

[0029]

[0030] where n = 0, 1, 2, … N-1, N is the number of sampling points; Δt is the sampling interval; t n = nΔt and τ m = mΔt are the discrete time and time offset respectively, n = 0, 1, 2, …, N-1, m = 0, 1, 2, …, N-1; fk = kΔt is the frequency parameter after discretization, where k = 0, 1, 2, …, N - 1;

[0031] (302) The calculated S[m, k] values will be filled into the position of the k-th row and the m-th column to form an N×N matrix M, and this matrix is the time-frequency distribution matrix of the time-frequency transformation.

[0032] Parameters such as the number of sampling points and the sampling interval ensure the accuracy of time-frequency analysis. Finally, the calculation results are filled to form the time-frequency distribution matrix M, which completely retains the joint characteristics of the signal in the time domain and the frequency domain, not only reflecting the dynamic change law of the fuel cell degradation process, but also highlighting the key fault characteristic frequencies, providing a high-resolution time-frequency feature expression for the subsequent Gramian angular field transformation, and significantly enhancing the ability of the life prediction model to capture weak degradation characteristics.

[0033] Preferably, step 4 includes:

[0034] (401) Normalize the row data of the time-frequency distribution matrix and scale the data to the range [-1, 1]. The formula is:

[0035]

[0036] where Mi represents the i-th row of matrix M, and represents the scaled matrix; is the normalized sequence;

[0037] (402) Extract the normalized one-dimensional data sequence from the matrix row by row and convert it to polar coordinate representation. The formula is:

[0038]

[0039] where: t i is the timestamp; N is the constant factor of the regularization polar coordinate generation space; θ i is the i-th inverse cosine function value; r is the polar coordinate radius;

[0040] (403) Generate the Gram matrix through the element calculation formula G GASF = cos(θ 1 + ν 1 ). The formula is:

[0041]

[0042] Use the Gram matrix as the pixel value of the image to generate a two-dimensional image.

[0043] This design not only effectively encodes the spatio-temporal correlation characteristics of the time-frequency matrix, but also forms a visual feature expression suitable for processing by deep learning models, significantly enhancing the feature extraction ability of the model for fuel cell degradation modes.

[0044] Preferably, the training process of the time-aware graph convolutional recurrent network model described in step 5 includes:

[0045] (501) Input the intrinsic mode function and two-dimensional image information into the time-aware graph convolutional recurrent network model, and perform a normalized mapping on the image pixel values:

[0046]

[0047] where X norm is the original pixel value; X max , X min are the maximum and minimum values respectively;

[0048] (502) Adjust the data to the input format expected by the time-aware graph convolutional recurrent network model, and the formula is:

[0049]

[0050] where represents all time series data at time t p ; N represents that the input is N spatially correlated time series, each time series covers P time steps and has d-dimensional features;

[0051] (503) Combine the node embedding E v and the time encoding Φ(t), and calculate the correlation between nodes:

[0052]

[0053] where represents the static spatial correlation between nodes i and j; <Φ(t 1 ), Φ(t 2 )> represents the time evolution of the graph structure; is the node embedding, d N is the node embedding dimension; is the time encoding function, d T is the time embedding dimension; is a vector that fuses node and time information.

[0054] This design enables the model to adaptively capture the non-linear interactions between multiple variables such as voltage, current, and temperature at different time scales, significantly improving the integrity of degradation feature extraction and the time-series modeling ability of life prediction.

[0055] Preferably, the training process of the time-aware graph convolutional recurrent network model described in step 5 further includes model parameter update:

[0056] (504) The adjacency matrix A of the time perception graph at time t t is normalized to obtain Meanwhile, the node embedding E v and the time vector at time t are

[0057] (505) Calculate the update gate z t , reset gate r t and candidate activation vector

[0058]

[0059] where χ :t contains the information of the picture at the intermediate step, h t-1 is the hidden state at the previous moment, W z , b z are the weight matrix and bias vector, σ is the sigmoid function for controlling the information inflow, tanh is the hyperbolic tangent function, and ⊙ is the Hadamard product;

[0060] (506) Update the hidden state at the current moment to

[0061] (507) Based on step (53), capture the spatio-temporal features in the picture data, calculate the prediction value by combining the weight parameters, and the final learning objective formula and loss function are:

[0062] L = L error + λL time

[0063]

[0064] where λ is an adjustable hyperparameter, L time is the time-related error term, L error is the prediction error term, y i is the true value, is the prediction value, and N is the number of samples.

[0065] This design enables the model to not only learn the long-term dependence relationship of the fuel cell degradation process but also adaptively adjust the prediction weights at different time steps, thereby significantly improving the temporal modeling accuracy and generalization ability of the remaining life prediction.

[0066] Preferably, the hyperparameters for optimizing the model by using the neural population dynamics optimization algorithm described in step 5 include:

[0067] (508) Use the Sobol sequence to replace the traditional initialization of the neural population dynamic optimization algorithm for the neural population state, so that the parameter search points are more evenly distributed in the solution space, laying a good foundation for subsequent optimization:

[0068]

[0069] Among them, P n represents the position of the nth neural population state generated, and UB and LB are the upper and lower bounds of the value range of the global solution respectively. is the ith random number generated by the Sobol sequence, and its value range is [0, 1];

[0070] (509) Optimize the hyperparameters d N , d T , define (d N , d T ) as a set of parameter groups in a two-dimensional search space that includes the node embedding dimension d N and the time embedding dimension;

[0071] (510) Introduce the discoverer update mechanism based on sine-cosine and inertia weight in the sparrow search algorithm and the attractor trend strategy to improve the neural population dynamic optimization algorithm. Guide the parameters towards the optimal solution according to the attractor convergence strategy formula, increase the diversity of the population, and help jump out of the local optimum:

[0072]

[0073] Among them, ω is the inertia weight, t is the current iteration number, g 1 is the amplitude adjustment coefficient, g 2 ∈ [0, 2π], g 3 ∈ [-2, 2], R 2 ∈ [0, 1] and ST ∈ [0.5, 1] are all uniformly distributed random variables; att k is a randomly selected attractor; z attract,i is the current neural population.

[0074] This design not only improves the accuracy of parameter configuration, but also optimizes the overall search efficiency, provides a better parameter optimization scheme for the fuel cell life prediction model, and comprehensively improves the model performance.

[0075] Preferably, the hyperparameters for optimizing the model by using the neural population dynamic optimization algorithm further include:

[0076] (511) Calculate the additive coupling term and the diffusion coupling term according to the additive coupling formula and the diffusion coupling formula, combine the additive coupling term and the diffusion coupling term to obtain the coupling perturbation, effectively enhance the diversity of the parameter group, avoid the algorithm from falling into local optimum, apply the coupling perturbation to the current parameter group, generate a new parameter group, and increase the diversity of the parameter group. The formula is:

[0077] z couple,i =a·(z add,i +z dif,i )

[0078]

[0079] Among them, c 2 , c 3 are random numbers within [0, 0.5], a is the coupling interference scaling factor, z add,i is the additive coupling term of the i-th neural population, z dif,i is the diffusion coupling term of the i-th neural population, and z couple,i is the coupling perturbation;

[0080] (512) Calculate the information transfer term according to the attractor-related information transfer formula and the coupling perturbation-related information transfer formula, realizing the intelligent sharing of optimized experiences among populations. The formula is:

[0081] z C_attract,i =C attract,i ·z attract,i (t + 1)

[0082] z C_couple,i =C couple,i ·R couple,i ·z couple,i ·c 4 ·(1 - FE / FE max )

[0083] Among them, C couple,i and C attract,i are random binary value adjacency matrices, R couple,i is the communication intensity matrix, c 4 is a random number within [0, 1], FE is the function evaluation times, and FE max is the maximum function evaluation times;

[0084] (513) Combine the information transfer term with the current parameter group to obtain the updated parameter group. According to the information projection strategy, dynamically adjust the parameter update method according to different training stages and parameter performances, guide the parameter group to develop in a better direction, and can dynamically adjust the update method according to the training stage and parameter performance, so that the parameter group evolves along a better search direction. The formula is:

[0085] z new,i = z i + z C_attract,i + z C_couple,i 。

[0086] This optimization method that combines multiple mechanisms not only greatly improves the exploration breadth of the parameter space, but also ensures the stability and efficiency of the optimization process, providing a more accurate hyperparameter configuration scheme for the fuel cell life prediction model.

[0087] A fuel cell remaining life prediction system based on multiple variables, comprising:

[0088] A data acquisition module for real-time collecting voltage, current, temperature and remaining life data of the fuel cell through sensors;

[0089] A signal decomposition module that uses the empirical reduced-order variational mode decomposition algorithm to decompose the voltage, current, temperature and remaining life data, and obtains multiple intrinsic mode functions of each variable;

[0090] A time-frequency analysis module for performing time-frequency transformation on the first intrinsic mode function of each variable to generate a time-frequency distribution matrix;

[0091] An image feature conversion module that uses the Gram angle field to convert the time-frequency distribution matrix into a Gram matrix and generates two-dimensional image information;

[0092] A model training and optimization module, including a time-aware graph convolutional recurrent network model, for receiving the intrinsic mode functions and two-dimensional image information for training, and using the neural population dynamics optimization algorithm to optimize the model hyperparameters to obtain a fuel cell remaining life prediction model;

[0093] A prediction and visualization module for outputting the fuel cell remaining life prediction result based on the trained and optimized prediction model and performing visual display through the front-end interface.

[0094] A computer-readable storage medium, on which a computer program is stored, characterized in that when the computer program is executed by a processor, it implements the virtual power plant multi-agent aggregation regulation ability quantification evaluation method according to any one of claims 1 to 8.

[0095] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: 1. It effectively improves the decomposition accuracy of complex signals and enhances the adaptability of the algorithm in non-stationary and non-linear signal processing; 2. Through the Sobol sequence initialization and the perturbation strategy of the sparrow search algorithm, the problem of local optimum is avoided, the population diversity and the global optimization ability are enhanced, so as to more efficiently optimize the model parameters; 3. It significantly improves the efficiency of signal feature extraction and the accuracy of the prediction model; 4. Through dynamic parameter optimization and spatio-temporal perception mechanism, the training efficiency and generalization performance of the model are improved, and it is applicable to the prediction task of complex time-series data. Description of the Drawings

[0096] Figure 1 is a schematic flow chart of the present invention;

[0097] Figure 2 is a flow chart of the NPDOA algorithm of the present invention. Detailed Embodiments

[0098] The technical solution of the present invention will be further described below with reference to the drawings.

[0099] As Figure 1 shown, the present invention proposes a method for predicting the remaining life of a fuel cell based on multiple variables, which specifically includes the following steps:

[0100] (1) Use a variety of sensors to collect data such as the voltage, current, temperature, and remaining life of the fuel cell.

[0101] (2) Use the empirical reduced-order variational mode decomposition (ERVMD) algorithm to decompose variables such as voltage, current, temperature, and remaining life to obtain the intrinsic mode functions (IMFs). The ERVMD algorithm embeds the local time-frequency characteristics of the empirical mode decomposition (EMD) algorithm into the global optimization of the reduced-order variational mode decomposition (RVMD) algorithm to form a decomposition model algorithm with mixed constraints. The specific steps are as follows:

[0102] (201) Use RVMD to decompose the original signal into the preliminary mode u k and the central frequency ω k . The specific process of the RVMD variational model is as follows:

[0103]

[0104] Among them, x(t) is the original signal, u k (t) is the kth mode, ω k is the central frequency, K is the number of decomposition layers, and the preliminary mode set

[0105] (202) For each RVMD mode Perform EMD decomposition to obtain refined IMFs and residuals. The formula is as follows:

[0106]

[0107] where, e max (t) is the maximum envelope, e min (t) is the minimum envelope, m k (t) is the mean envelope, is the temporary candidate IMF. If is satisfied, the formula loop stops. If not, repeat the above formula steps until the residual is a monotonic function or a constant. Decompose each into and the residual r k (t);

[0108] (203) Fuse the decomposition results of RVMD and EMD to construct a hybrid optimization objective function to suppress mode mixing. The specific process is as follows:

[0109]

[0110] where, α is the RVMD mode bandwidth parameter to control mode sparsity, β is the sparse constraint term of EMD, γ is the consistency weight of the decomposition results of RVMD and EMD to prevent over-decomposition, and M is the number of EMD decomposition layers;

[0111] (204) Reconstruct the signal based on the fused IMFs to suppress noise and mode mixing. The specific process is as follows:

[0112]

[0113] where, is the final reconstructed signal, ω c is the introduced adaptive weighting coefficient.

[0114] (3) Discretize the decomposed modes and perform the discrete S-transform to obtain the time-frequency distribution matrix. The specific steps are as follows:

[0115] (301) Perform Fourier transform on the mode components to obtain the discrete-time signal x(t), i.e., x[n]. The expression form of the discrete S-transform of this time series is:

[0116]

[0117] where, n = 0, 1, 2, … N-1, N is the number of sampling points; Δt is the sampling interval; t n = nΔt and τ m=mΔt are the discrete time and time offset respectively, where n = 0, 1, 2, …, N - 1, m = 0, 1, 2, …, N - 1; f k =kΔt is the discrete frequency parameter, where k = 0, 1, 2, …, N - 1;

[0118] (302) The calculated S[m, k] values will be filled into the position of the k-th row and the m-th column to form an N×N matrix M, and this matrix is the result matrix of the S transform.

[0119] (4) Use the Gram angle and field to transform the time-frequency distribution matrix into a Gram matrix, and then generate two-dimensional image information. The specific steps are as follows:

[0120] (401) For the matrix set obtained by the S transform, it is often necessary to scale each row of the matrix to scale the data between [-1, 1]. The formula is:

[0121]

[0122] Among them, Mi represents the i-th row of matrix M, indicating the scaled matrix; is the normalized sequence;

[0123] (402) Extract the normalized one-dimensional data sequence from the matrix row by row and convert it into polar coordinate representation. The formula is:

[0124]

[0125] where: t i is the time stamp; N is the constant factor of the regularization polar coordinate generation space; θ i is the value of the i-th inverse cosine function; r is the polar coordinate radius;

[0126] (403) Generate the Gram matrix through the element calculation formula G GASF =cos(θ 1 +ν 1 ) The formula is:

[0127]

[0128] Use the angle matrix as the pixel value of the image to generate a two-dimensional image.

[0129] (5) Input the IMFs in step 2 and the two-dimensional image information in step 4 into the Time-Aware Graph Convolutional Recurrent Network (TGCRN) model for training, and use the improved Neural Population Dynamics Optimization Algorithm (NPDOA) to optimize the hyperparameters of the TGCRN model to obtain the fuel cell remaining life prediction model. The specific steps are as follows:

[0130] (501) Input the decomposed IMFs of ERVMD and the two-dimensional images generated by GADF into the TGCRN model for training;

[0131] (502) Map the pixel value X to [0, 1], and the formula is as follows:

[0132]

[0133] where X norm is the original pixel value, X max , X min are the maximum and minimum values respectively;

[0134] (503) Adjust the data to the expected input format of the TGCRN model, and the formula is:

[0135]

[0136] where represents all time series data at time t p ; N indicates that the input is N spatially correlated time series, each time series covers P time steps and has d-dimensional features.

[0137] (504) Put the processed data into the TGCRN model;

[0138] (505) Combine the node embedding E v and the time encoding Φ(t), when calculating the node - to - node correlation, the formula is:

[0139]

[0140] where represents the static spatial correlation between nodes i and j; <Φ(t 1 ), Φ(t 2 )> represents the time evolution of the graph structure; is the node embedding, d N is the node embedding dimension; is the time encoding function, d T is the time embedding dimension; is the vector that fuses node and time information;

[0141] (506) Normalize the adjacency matrix Α t of the time - aware graph at time t to obtain Meanwhile, combine the node embedding E v and the time vector at time t as

[0142] (507) The calculation formulas for the update gate, reset gate, and candidate activation vector are respectively:

[0143]

[0144] Among them, χ :t contains the information of the picture at the same step, h t-1 is the hidden state at the previous moment, W z , b z are the weight matrix and the bias vector, σ is the sigmoid function, used to control the information inflow, tanh is the hyperbolic tangent function, and ⊙ is the Hadamard product;

[0145] (508) The hidden state at the current moment is updated as:

[0146]

[0147] (509) Capture the spatio-temporal features in the picture data through step (505), calculate the predicted value by combining the weight parameters, and the final learning objective formula and loss function are:

[0148] L = L error + λL time

[0149]

[0150] Among them, λ is an adjustable hyperparameter, L time is the error term related to time, L error is the prediction error term, y i is the true value, is the predicted value, and N is the number of samples.

[0151] (510) As Figure 2 shown, use the Sobol sequence to replace the traditional initialization of the NPDOA neural population state:

[0152]

[0153] Among them, P n represents the position of the nth neural population state generated, UB and LB are the upper and lower bounds of the value range of the global solution respectively, is the ith random number generated by the Sobol sequence, and the value range is [0,1];

[0154] (511) In the TGCRN parameter optimization, optimize the hyperparameters d N , d T , define (d N , d T ) as a set of parameter groups in a two-dimensional search space containing the node embedding dimension d N and the time embedding dimension;

[0155] (512) In the TGCRN parameter optimization, as Figure 2 shown, the discovery update mechanism based on sine-cosine and inertia weight in the sparrow search algorithm is introduced to improve the attractor trend strategy of NPDOA. According to the attractor convergence strategy formula, the parameters are guided towards the optimal solution:

[0156]

[0157] (513) Calculate the additive coupling term and the diffusion coupling term according to the additive coupling formula and the diffusion coupling formula. Combine the additive coupling term and the diffusion coupling term to obtain the coupling perturbation. Apply the coupling perturbation to the current parameter group to generate a new parameter group, increasing the diversity of the parameter group. The formula is:

[0158] z couple,i =a·(z add,i +z dif,i )

[0159]

[0160] Among them, c 2 , c 3 are random numbers within [0, 0.5], a is the coupling interference scaling factor, z add,i is the additive coupling term of the i-th neural population, z dif,i is the diffusion coupling term of the i-th neural population, and z couple,i is the coupling perturbation;

[0161] (514) Calculate the information transfer term according to the attractor-related information transfer formula and the coupling perturbation-related information transfer formula. The formula is:

[0162] z C_attract,i =C attract,i ·z attract,i (t + 1)

[0163] z C_couple,i =C couple,i ·R couple,i ·z couple,i ·c 4 ·(1 - FE / FE max )

[0164] Among them, C couple,i and C attract,i are random binary value adjacency matrices, R couple,i is the communication intensity matrix, c 4 is a random number within [0, 1], FE is the function evaluation times, and FE max is the maximum function evaluation times;

[0165] (515) Combine the information transfer item with the current parameter group to obtain an updated parameter group. According to the information projection strategy, dynamically adjust the parameter update method according to different training stages and parameter performance, and guide the parameter group to develop in a better direction. The formula is:

[0166] z new,i =z i +z C_attract,i +z C_couple,i .

[0167] (6) Use the fuel cell remaining life prediction model optimized in step 5 to predict the remaining life of the fuel cell, and obtain the final fuel cell remaining life prediction result.

[0168] (7) Display the fuel cell remaining life prediction result obtained in step 6 on the front-end interface for the supervisors to process.

[0169] A fuel cell remaining life prediction system corresponding to the fuel cell remaining life prediction method includes the following modules:

[0170] A data acquisition module, configured to collect the voltage, current, temperature, and remaining life data of the fuel cell in real time through sensors;

[0171] A signal decomposition module, which uses the empirical reduced-order variational mode decomposition algorithm to decompose the voltage, current, temperature, and remaining life data to obtain multiple intrinsic mode functions of each variable;

[0172] A time-frequency analysis module, configured to perform time-frequency transformation on the first intrinsic mode function of each variable to generate a time-frequency distribution matrix;

[0173] An image feature conversion module, which uses the Gram angular field to convert the time-frequency distribution matrix into a Gram matrix and generates two-dimensional image information;

[0174] A model training and optimization module, including a time-aware graph convolutional recurrent network model, configured to receive the intrinsic mode function and two-dimensional image information for training, and use the neural population dynamics optimization algorithm to optimize the model hyperparameters to obtain a fuel cell remaining life prediction model;

[0175] A prediction and visualization module, configured to output the fuel cell remaining life prediction result based on the trained and optimized prediction model and perform visual display through the front-end interface.

[0176] An embodiment of the present invention also discloses a computer-readable storage medium.

[0177] Specifically, a computer-readable storage medium is used to store a computer program. When the computer program is executed by a processor, the methods in the above method embodiments are implemented. Those skilled in the art can understand that to implement all or part of the processes in the above method embodiments of the present application, it can be completed by instructing relevant hardware through a computer program. This program can be stored in a computer-readable storage medium. When this program is executed, it can include the processes of the above method embodiments. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (HDD), or a solid-state drive (SSD), etc.; the storage medium can also include a combination of the above types of memories.

Claims

1. A method for predicting the remaining life of a fuel cell based on multiple variables, characterized in that: The following steps are involved: (1) Using sensors to collect the voltage, current, temperature and remaining life data of the fuel cell; (2) using an empirical reduced-order variational mode decomposition algorithm to decompose the voltage, current, temperature and remaining life data variables to obtain a plurality of intrinsic mode functions; (3) Performing time-frequency transformation on the first intrinsic mode function of each variable to generate a time-frequency distribution matrix; (4) using the Gram angular field to convert the time-frequency distribution matrix into a Gram matrix and generate two-dimensional image information; (5) inputting the intrinsic mode function and the two-dimensional image information into a time-aware graph convolutional recursive network model for training, and optimizing the hyperparameters of the model using a neural population dynamic optimization algorithm to obtain a fuel cell remaining life prediction model; (6) Based on the trained and optimized prediction model, the remaining life prediction results of the fuel cell are output and displayed through the front-end interface.

2. The prediction method according to claim 1, characterized in that: Step 2 includes: (201) Use the reduced-order variational mode decomposition algorithm to decompose the original signal into preliminary modes u k and the center frequency ω k , the reduced-order variational mode decomposition process is: Among them, x(t) is the original signal, u k (t) is the kth mode, ω k is the center frequency, K is the number of decomposition layers, and the preliminary mode set is obtained (202) For each preliminary mode Perform empirical mode decomposition to obtain refined modes and residuals. The empirical mode decomposition process is: Among them, e max (t) is the maximum envelope, e min (t) is the minimum envelope, m k (t) is the mean envelope, is a temporary candidate eigenmode function, if When the condition is satisfied, the formula loop stops. If not, repeat the above formula steps until the residual is a monotonic function or a constant, and each Decompose into and residual r k (t); (203) The decomposition results of reduced-order variational modes and empirical modes are integrated to construct a hybrid optimization objective function: Among them, α is the bandwidth parameter of the reduced-order variational mode, β is the sparse constraint term of the empirical mode, γ is the consistency weight between the reduced-order variational mode and the empirical mode decomposition result to prevent over-decomposition, and M is the number of empirical mode decomposition layers; (204) Based on the fused modal reconstruction signal, an adaptive weighting coefficient is introduced: in, For the final reconstructed signal, ω c is the introduced adaptive weighting coefficient.

3. The prediction method according to claim 1, characterized in that: Step 3 includes: (301) The first eigenmode function is subjected to Fourier transform to obtain a discrete time signal x(t), i.e., x[n]. The expression of the discrete time-frequency transform of this time series is: Where n = 0, 1, 2, ... N-1, N is the number of sampling points; Δt is the sampling interval; t n = nΔt and τ m =mΔt and time offset after discretization, n=0,1,2,…,N-1,m=0,1,2,…,N-1; f k =kΔt is the frequency parameter after discretization, k = 0, 1, 2, ..., N-1; (302) The calculated S[m,k] value will be filled into the kth row and mth column to form an N×N matrix M, which is the time-frequency distribution matrix of the time-frequency transform.

4. The prediction method according to claim 1, characterized in that: Step 4 includes: (401) Normalize the row data of the time-frequency distribution matrix and scale the data to between [-1, 1]. The formula is: Among them, Mi represents the i-th row of the matrix M represents the scaled matrix; is the normalized sequence; (402) The normalized one-dimensional data sequence is extracted from the matrix row by row and converted into polar coordinate representation. The formula is: Where: t i is the timestamp; N is the constant factor for the regularized polar coordinate generation space; θ i is the value of the i-th arccosine function; r is the polar coordinate radius; (403) Calculate the formula G by elements GASF =cos(θ1+ν1) generates the Gram matrix, the formula is: The Gram matrix is ​​used as the pixel value of the image to generate a two-dimensional image.

5. The prediction method according to claim 1, characterized in that: The training process of the time-aware graph convolutional recursive network model described in step 5 includes: (501) Inputting the intrinsic mode function and the two-dimensional image information into the time-aware graph convolutional recursive network model, normalizing and mapping the image pixel values: Among them, X norm is the original pixel value; X max , X min are the maximum and minimum values ​​respectively; (502) The data is adjusted to the input format expected by the time-aware graph convolutional recurrent network model, and the formula is: X tp =(X1,X2,…X tp )∈R N×P×d in, Represents in t p All time series data at the moment; N means the input is N spatially related time series, each time series covers P time steps and has d-dimensional features; (503) Combined node embedding E v And time code Φ(t), calculate the correlation between nodes: in, represents the static spatial correlation between nodes i and j; <Φ(t1),Φ(t2)> represents the time evolution of the graph structure; is the node embedding, d N Embedding dimension for nodes; is the time encoding function, d T Embed dimension for time; A vector that combines node and time information.

6. The prediction method according to claim 5, characterized in that: The training process of the time-aware graph convolutional recursive network model described in step 5 also includes model parameter updating: (504) The adjacency matrix A of the time-aware graph at time t t Normalized, we get Simultaneously merge the node embedding E v and the time vector at time t is (505) Update gate z by dynamic adjacency matrix calculation t , reset gate t and candidate activation vectors Among them, χ :t Contains information about the image in the intervening step, h t-1 is the hidden state at the previous moment, W z 、b z is the weight matrix and bias vector, σ is the sigmoid function used to control the inflow of information, tanh is the hyperbolic tangent function, and ⊙ is the Hadamard product; (506) Update the current hidden state to (507) Based on the spatiotemporal features captured in the image data in step (503), the predicted value is calculated in combination with the weight parameters. The final learning objective formula and loss function are: L=L error +λL time Among them, λ is an adjustable hyperparameter, L time is the time-dependent error term, L error is the prediction error term, y i is the true value, is the predicted value, and N is the number of samples.

7. The prediction method according to claim 6, characterized in that: The hyper parameters of the model optimized by the neural population dynamic optimization algorithm described in step 5 include: (508) Use Sobol sequence to replace the traditional initialization of neural population dynamic optimization algorithm neural population state: Among them, P n represents the generated nth neural group state position, UB and LB are the upper and lower bounds of the value range of the global solution, respectively. is the i-th random number generated by the Sobol sequence, ranging from [0,1]; (509) for the hyperparameter d N ,d T To optimize, define (d N ,d T ) is a node embedding dimension d N and a set of parameter groups in a two-dimensional search space of the temporal embedding dimension; (510) The discoverer update mechanism based on sine, cosine and inertia weights in the sparrow search algorithm and the attractor trend strategy of the improved neural population dynamic optimization algorithm are introduced, and the parameters are guided to the optimal solution according to the attractor convergence strategy formula: Among them, ω is the inertia weight, t is the current iteration number, g1 is the amplitude adjustment coefficient, g2∈[0,2π], g3∈[-2,2], R2∈[0,1] and ST∈[0.5,1] are all uniformly distributed random quantities; att k is a randomly selected attractor; z attract,i is the current neural population.

8. The prediction method according to claim 7, characterized in that: The hyper parameters of the model optimized by the neural population dynamic optimization algorithm also include: (511) According to the additive coupling formula and the diffusion coupling formula, the additive coupling term and the diffusion coupling term are calculated, and the additive coupling term and the diffusion coupling term are combined to obtain the coupling disturbance. The coupling disturbance is applied to the current parameter group to generate a new parameter group and increase the diversity of the parameter group. The formula is: from couple,i =a·(z add,i +z dif,i ) Among them, c2 and c3 are random numbers in [0,0.5], a is the coupling interference scaling factor, z add,i is the additive coupling term of the ith neural population, z dif,i is the diffusion coupling term of the ith neural population, z couple,i is the coupled disturbance; (512) The information transfer term is calculated according to the attractor-related information transfer formula and the coupled perturbation-related information transfer formula, which is: z C_attract,i =C attract,i ·z attract,i (t+1) z C_couple,i =C couple,i ·R couple,i ·z couple,i ·c4·(1-FE / FE max ) Among them, C couple,i and C attract,i is a random binary-valued adjacency matrix, R couple,i is the communication intensity matrix, c4 is a random number in [0,1], FE is the number of function evaluations, FE max is the maximum number of function evaluations; (513) The information transfer term is combined with the current parameter group to obtain an updated parameter group. According to the information projection strategy, the parameter update method is dynamically adjusted according to different training stages and parameter performance to guide the parameter group to develop in a better direction. The formula is: With new,i =from i +with C_attract,i +with C_couple,i 。 9. A fuel cell remaining life prediction system based on multiple variables, characterized in that: include: A data acquisition module is used to collect the voltage, current, temperature and remaining life data of the fuel cell in real time through sensors; A signal decomposition module, which uses an empirical reduced-order variational mode decomposition algorithm to decompose the voltage, current, temperature and remaining life data to obtain multiple intrinsic mode functions of each variable; The time-frequency analysis module is used to perform time-frequency transformation on the first intrinsic mode function of each variable to generate a time-frequency distribution matrix; An image feature conversion module converts the time-frequency distribution matrix into a Gram matrix using a Gram angular field and generates two-dimensional image information; A model training and optimization module, including a time-aware graph convolutional recursive network model, for receiving the intrinsic mode function and two-dimensional image information for training, and optimizing model hyperparameters using a neural population dynamic optimization algorithm to obtain a fuel cell remaining life prediction model; The prediction and visualization module is used to output the prediction results of the remaining life of the fuel cell based on the trained and optimized prediction model, and to visualize it through the front-end interface.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the method for quantitatively evaluating the multi-subject aggregate regulation capability of a virtual power plant according to any one of claims 1 to 8.

Images