A neural network-based fuel cell thermal management method, device and system

Abstract

The application provides a fuel cell thermal management method based on a neural network, which comprises the following steps: operating a test fuel cell stack on a test bench, obtaining an initial data set in a time sequence, and forming a training data set and a test data set; establishing an echo state neural network regression model, training the echo state neural network regression model by using the training data set, and obtaining an echo state neural network prediction model; testing by using the test data set to verify the prediction accuracy of the model, and obtaining a trained neural network prediction model; collecting data of a fuel cell stack in a running state in real time, obtaining a prediction signal by using the trained neural network prediction model, and controlling the running state of a cooling water pump and a radiator according to the prediction signal. The application further discloses corresponding devices and systems. By implementing the application, the accuracy and efficiency of fuel cell thermal management can be improved.

Description

Technical Field

[0001] This invention relates to the field of thermal management technology for proton exchange membrane fuel cells (PEMFCs), and in particular to a neural network-based thermal management method for fuel cells. Background Technology

[0002] Hydrogen energy is an abundant, green, low-carbon, and widely applicable secondary energy source, which is of great significance for building a clean, low-carbon, safe, and efficient energy system and achieving carbon peaking and carbon neutrality goals. Hydrogen fuel cell vehicles have attracted widespread attention due to their high efficiency and cleanliness. Proton exchange membrane fuel cells (PEMFCs) have advantages such as high energy conversion efficiency, low-temperature operation capability, high reliability, and zero emissions, making them a promising candidate for application in the field of new energy vehicles.

[0003] A Proton Exchange Membrane Cell (PEMFC) is a device that converts chemical energy into electrical energy. Its main components include a proton exchange membrane, anode, cathode, and bipolar plates. A PEMFC is a complex, nonlinear system with multiple physics fields and parameters coupled together. Its operating temperature is a critical factor affecting its output performance and lifespan. Excessive operating temperature can cause liquid water evaporation, leading to membrane dryness failure; excessively low temperature can cause flooding of the cathode channel, preventing oxygen from penetrating the gas diffusion layer. The normal operating temperature range of a PEMFC is 60–80°C; however, it generates a significant amount of heat during operation, necessitating effective thermal management. Improper thermal management can lead to an irreversible drop in the PEMFC's output voltage, accelerating its aging process.

[0004] Currently, the main method for thermal management of PEMFCs is to control the cooling water flow rate based on a temperature model. Control methods include PID control, state feedback control, model predictive control, and fuzzy control. While these methods are simple in principle and easy to use, they suffer from drawbacks such as slow response speed and long settling time. The inherent nonlinear characteristics and parameter uncertainties of fuel cells, along with the sensitivity and large fluctuations in temperature of high-power fuel cell stacks for vehicles, make the application of these control methods challenging.

[0005] As research deepens, several novel control methods have emerged. For example, some researchers have designed fuzzy control methods for PEMFC thermal management, controlling the PEMFC temperature by adjusting fan speed. Others have employed an improved particle swarm optimization fuzzy PID fuel cell temperature control method, utilizing improved particle swarm optimization fuzzy PID control with the control strategy set based on empirical control rules. Furthermore, a model reference adaptive control method has been designed, controlling the fuel cell stack temperature and circulating coolant inlet temperature by adjusting the coolant mass flow rate and bypass valve opening coefficient. This invention is simple and efficient, easily applicable to fuel cell systems of various power levels, enabling real-time and effective temperature control. Additionally, considering the drawbacks of single control methods, some researchers have disclosed a proton exchange membrane fuel cell temperature control method. This method uses a PID controller as a general feedback controller to stabilize the fuel cell stack temperature, achieving initial control. Then, based on the predictive function of a grey model, the temperature change trend of the fuel cell system is obtained. The fuzzy controller uses the obtained predictive information to compensate for system uncertainties and external disturbances, further improving the system control accuracy.

[0006] However, in practical applications, especially in industrial process control, due to the severe nonlinearity of the controlled object, the uncertainty of the mathematical model, and the drastic changes in the system operating point, existing control theory-based methods have serious and irreparable defects, greatly limiting their effectiveness. Furthermore, most methods use step load signals to verify the control method, but hydrogen fuel cell vehicles experience acceleration, constant speed, and deceleration during actual driving. The frequent changes in operating conditions make fuel cell temperature control more complex, and extreme temperatures may occur. The aforementioned control theory-based methods at present all have the following shortcomings: insufficient accuracy in controlling the temperature to the target value, and large fluctuations at the target temperature value. Summary of the Invention

[0007] The technical problem to be solved by this invention is to propose a fuel cell thermal management method, device and system based on neural networks, which can efficiently and accurately stabilize the inlet and outlet temperatures of the PEMFC fuel cell stack at the target temperature value.

[0008] The technical solution adopted in this invention is to provide a fuel cell thermal management method based on neural networks, which includes at least the following steps:

[0009] Step S10: Run the test fuel cell stack on the test bench, obtain the initial dataset in time sequence, and form the training dataset and the test dataset;

[0010] Step S11: Establish an echo state network (ESN) regression model;

[0011] Step S12: Train the established echo state neural network regression model based on the training dataset to obtain the echo state neural network prediction model.

[0012] Step S13: Test the trained echo state neural network prediction model on the test dataset to verify the model's prediction accuracy and obtain the trained neural network prediction model.

[0013] Step S14: Real-time acquisition of data from the fuel cell stack in operation, obtaining prediction signals using the trained neural network prediction model, the prediction signals including cooling water flow rate and cooling air volume, controlling the operation of the cooling water pump and radiator according to the prediction signals, and controlling the temperature of the fuel cell stack at a reasonable level.

[0014] Preferably, step S10 further includes:

[0015] The test fuel cell stack was run on the test bench, and data on the target temperature of the fuel cell, the temperature of the reactant gas, the humidity of the reactant gas, the pressure of the reactant gas, the flow rate of the reactant gas, the current of the fuel cell, the flow rate of the cooling water and the cooling air volume were obtained at each time step through multiple sensors to form an initial dataset.

[0016] The initial dataset is preprocessed, including standardization and normalization.

[0017] The preprocessed initial dataset is divided into a training dataset and a test dataset.

[0018] Preferably, step S11 includes at least:

[0019] An echo state neural network regression model is established, which has an input layer, a reservoir, and an output layer; the input layer takes a time series data as input, and the output layer outputs another time series data related to the input.

[0020] Initializing the echo state neural network regression model includes:

[0021] Determine the number of neuron nodes in the reserve pool;

[0022] Randomly generate the internal connectivity matrix W res The internal connection matrix W res This represents the connection state between each neuron in the reservoir, including the direction and weight of the connection;

[0023] Determine the weights of the input connection matrix of the input layer and the spectral radius of the internal connection matrix.

[0024] At the end of the initialization process, it is necessary to set the weights of the input connections and the spectral radius of the internal connection matrix.

[0025] Preferably, step S12 further includes:

[0026] The initialized echo state neural network regression model is preheated to form an internal state with time series characteristics;

[0027] The data from the training dataset is selected and trained according to the time series, with the target temperature of the fuel cell, the temperature of the reactant gas, the humidity of the reactant gas, the pressure of the reactant gas, the flow rate of the reactant gas, and the current of the fuel cell as inputs, and the cooling water flow rate and the cooling air volume as outputs. The echo state neural network regression model is trained accordingly.

[0028] When the training result of the echo state neural network regression model is not overfitting and exhibits high accuracy, that is, when the root mean square error (RMSE) between the predicted and expected values ​​is close to 0 and the goodness of fit (R²) is high, the model is considered to be overfitting. 2 When the condition is close to 1, training ends and the echo state neural network prediction model is obtained.

[0029] Preferably, training the echo state neural network regression model specifically includes:

[0030] During the sampling phase, input samples need to be fed into the ESN's reservoir to generate corresponding internal states; and after the sampling phase is completed, a set of reservoir states corresponding to each input sample can be obtained.

[0031] During the weight calculation phase, a linear regression method is used to calculate the output weight matrix of the ESN to map the internal state to the desired output.

[0032] Preferably, the weight calculation stage further includes:

[0033] Determine the relationships between the internal pool and the previous pool and input:

[0034] x(t+1)=f[W res ·x(t)+W LR ·u(t)]

[0035] Among them, W LR W is the input connection weight matrix, used to represent the mapping relationship between the input and the reserve pool; RO The output connection weight matrix represents the mapping relationship from the reservoir to the output; f is the activation function, and W... res This is the internal connection weight matrix, used to represent the connection relationships among the elements in the vector x(t) of the reserve pool;

[0036] Based on the system state matrix and sample data collected during the sampling phase, the output connection weight matrix W is calculated. RO ;include:

[0037] The objective function is obtained as follows:

[0038]

[0039] Wherein, there is a linear relationship between the state variable x(t) and the predicted output y(t), where y(t) is the predicted output. The expected output;

[0040] The output connection weight matrix W is calculated to minimize the system mean square error. RO :

[0041]

[0042] Among them, W RO x(t) is the network output.

[0043] Preferably, in step S13, the test data and training data have the same dimension and are input according to a time series.

[0044] Furthermore, it includes optimizing the parameters of the fuel cell stack by optimizing the reservoir size, reservoir radius, reservoir sparsity, and input cell scale parameters to meet the required fitting accuracy.

[0045] Preferably, the fuel cell stack is a fuel cell stack employing a proton exchange membrane fuel cell.

[0046] Accordingly, as another aspect of the present invention, a fuel cell thermal management device based on a neural network is also provided, which includes at least:

[0047] The dataset acquisition unit is used to run test fuel cell stacks on the test bench, obtain the initial dataset in time sequence, and form training datasets and test datasets;

[0048] The regression model acquisition unit is used to establish an echo state neural network regression model.

[0049] The training processing unit is used to train the established echo state neural network regression model based on the training dataset to obtain the echo state neural network prediction model.

[0050] The test processing unit is used to test the trained echo state neural network prediction model on the test dataset to verify the model's prediction accuracy and obtain the trained neural network prediction model.

[0051] The predictive processing unit is used to collect data of the fuel cell stack in operation in real time, obtain predictive signals using the trained neural network predictive model, the predictive signals including cooling water flow rate and cooling air volume, and control the operation of the cooling water pump and radiator according to the predictive signals to control the temperature of the fuel cell stack at a reasonable level.

[0052] Accordingly, as another aspect of the present invention, a neural network-based fuel cell thermal management system is also provided, comprising: a fuel cell stack, a water tank, a water pump, and a radiator connected in a loop;

[0053] Temperature sensors are installed at the inlet and outlet of the fuel cell stack;

[0054] The controller is connected to the fuel cell stack, water pump, temperature sensor and radiator. The controller is equipped with an operating unit for executing the aforementioned neural network-based fuel cell thermal management method.

[0055] Implementing the embodiments of the present invention has the following beneficial effects:

[0056] This invention provides a neural network-based method, apparatus, and system for fuel cell thermal management. By improving the accuracy of fuel cell thermal management, aiming to reduce the error between the inlet and outlet temperatures of the fuel cell stack and the target temperature, an echo state neural network is applied to the field of fuel cell thermal management. The echo state neural network is trained using a dataset obtained from fuel cell stack testing. The trained prediction model controls the cooling water flow rate and cooling airflow signals, ultimately optimizing the thermal management of the fuel cell system. This improves the accuracy and efficiency of temperature control in the fuel cell stack.

[0057] In this embodiment, since the hidden layer of the echo state neural network is a dynamic reservoir structure with echo state properties, it not only enhances the stability of network prediction, but also simplifies the training process by using only a linear algorithm to obtain the network output weights. At the same time, it overcomes the problems of slow convergence speed and easy getting trapped in local minima in traditional neural networks.

[0058] Meanwhile, the regression model obtained after training shows a significant improvement in the stability and accuracy of temperature control compared to control theory-based methods. It also has better temperature regulation capabilities than traditional neural networks, can better resist external load disturbances, and has a smaller deviation from the set value. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 The main flowchart is shown in one embodiment of a neural network-based thermal management method for fuel cells provided by the present invention.

[0061] Figure 2 This is a schematic diagram illustrating the application environment of the present invention;

[0062] Figure 3 This is a schematic diagram illustrating the overall principle of the construction, training, and application of the echo state neural network involved in this invention;

[0063] Figure 4 This is a schematic diagram of the topology of the echo state neural network involved in this invention;

[0064] Figure 5 This is a schematic diagram illustrating the principle framework of thermal management of a fuel cell stack according to the present invention.

[0065] Figure 6 This is a more detailed schematic diagram of the process principle of the method involved in this invention;

[0066] Figure 7 This is a main flowchart of an embodiment of a neural network-based fuel cell thermal management device provided by the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0068] like Figure 1 The diagram shows the main flowchart of an embodiment of a neural network-based fuel cell thermal management method provided by the present invention; combined with... Figures 2 to 7 As shown, in this embodiment, the method includes at least the following steps:

[0069] Step S10: Run the test fuel cell stack on the test bench, obtain the initial dataset in time sequence, and form the training dataset and the test dataset;

[0070] In this embodiment, on the test bench according to Figure 2 The test environment is constructed based on the following architecture, which includes at least: a fuel cell stack, a water tank, a water pump, and a radiator connected in a loop; temperature sensors (and other sensors) located at the inlet and outlet of the fuel cell stack; and a controller connected to the aforementioned components. The fuel cell stack is a PEMFC-based fuel cell stack. The heat generated by the fuel cell stack is first carried to the water tank by the cooling water pump through controlled cooling water flow, and then the heat is carried to the radiator. The radiator then dissipates the heat into the air by controlling the radiator airflow.

[0071] In this embodiment of the invention, it is assumed that the temperature in the cooling water is uniform, and the cooling water temperature at the fuel cell stack outlet is taken as the temperature of the fuel cell stack. Considering that the cooling water pump and radiator cannot be frequently started and stopped in practical applications, in this embodiment, minimum values ​​for the cooling water flow rate and cooling air volume need to be set.

[0072] More specifically, in this embodiment, for the fuel cell stack under test, its basic parameters such as the number of cells and the reaction area of ​​a single cell are determined, and the fuel cell stack is tested under cyclic operating conditions, and the physical parameters are managed during the test.

[0073] In the test bench, each physical parameter needs to be controlled within the parameter range of Table 1 below;

[0074] Table 1 Physical parameters of the test bench

[0075] parameter Control range Operating temperature 20–80℃ gas temperature 20–80℃ Gas humidity 0–100% (relative humidity) air speed 0–100 L / min hydrogen rate 0–30L / min Gas pressure 0–2 Bar Current 0–300A

[0076] In a specific example, step S10 further includes:

[0077] The test fuel cell stack was run on the test bench, and data on the target temperature of the fuel cell, the temperature of the reactant gas, the humidity of the reactant gas, the pressure of the reactant gas, the flow rate of the reactant gas, the current of the fuel cell, the flow rate of the cooling water and the cooling air volume were obtained at each time step through multiple sensors to form an initial dataset.

[0078] The initial dataset is preprocessed, including standardization and normalization.

[0079] The preprocessed initial dataset is divided into a training dataset and a test dataset.

[0080] Step S11: Establish an echo state neural network regression model;

[0081] It can be understood that the echo-state neural network regression model is constructed using an echo-state neural network, also known as a reservoir computation, which is commonly used to solve time series forecasting problems. This network belongs to recurrent neural networks and uses a reservoir of neurons, composed of randomly sparsely connected internal weight matrices, as hidden layers to represent the input in a high-dimensional, non-linear manner, exhibiting superior non-linear learning capabilities. The hidden layer weights of the echo-state neural network are pre-generated rather than trained, and are trained separately from the weights from the hidden layers to the output layer. Once the neural network structure is fixed, the input weight matrix and recurrent weight matrix remain unchanged; only the output weight matrix is ​​optimized through linear regression, significantly improving the computational efficiency of the echo-state neural network. The overall principles of the construction, training, and application of the echo-state neural network involved in this invention can be found in [reference needed]. Figure 3 As shown.

[0082] In a specific example, step S11 includes at least:

[0083] Establish an echo state neural network regression model, such as Figure 4 As shown, it has an input layer, a storage tank, and an output layer; wherein, the input of the input layer is a time series data, and the output layer outputs another time series data related to the input; in this embodiment, the input is the target temperature of the fuel cell, the temperature of the reactant gas, the humidity of the reactant gas, the pressure of the reactant gas, the flow rate of the reactant gas, and the current of the fuel cell; while the output is the cooling water flow rate and the cooling air volume;

[0084] Initializing the echo state neural network regression model includes:

[0085] Determine the number (scale) of neurons in the reserve pool; it is understood that the more neurons in the reserve pool, the stronger the fitting ability, but the more computational workload will also be; in this embodiment, the number of neurons in the reserve pool can be 10 times the total amount of data collected by the sensor.

[0086] Randomly generate the internal connectivity matrix W res The internal connection matrix W res This represents the connection state between each neuron in the reservoir, including the direction and weight of the connection;

[0087] Determine the weights of the input connection matrix of the input layer and the spectral radius of the internal connection matrix.

[0088] Understandably, during initialization, it's necessary to set the weights of the input connections in the input connection weight matrix corresponding to the input layer. These weights determine the influence of the input signal on the network. If the weights are too large, the input signal may have an excessive impact on the output, leading to overfitting; if the weights are too small, the important features of the input signal may not be captured. Therefore, it's crucial to choose appropriate weight values. Typically, the weights of the input connections are initialized with random numbers and then normalized to ensure full utilization of the input signal's features.

[0089] The spectral radius of the internal connection matrix is ​​a measure of matrix stability; it represents the largest eigenvalue of the matrix. A larger spectral radius indicates higher nonlinearity in the network, but it may also lead to oscillations and divergence. Therefore, the spectral radius needs to be carefully chosen to ensure network stability and generalization ability. In the initialization of an echo-state neural network, the weights of the input connections and the spectral radius of the internal connection matrix need to be appropriately set. The spectral radius of the internal connection matrix can be determined through trial and error or adjustment algorithms to ensure network stability and performance. Common algorithms for adjusting the spectral radius include eigenvalue decomposition and power iteration.

[0090] Step S12: Train the established echo state neural network regression model based on the training dataset to obtain the echo state neural network prediction model.

[0091] In a specific example, step S12 further includes:

[0092] Step S120: The initialized echo state neural network regression model is preheated to form an internal state with time series characteristics.

[0093] Step S121: Select data from the training dataset, and train the echo state neural network regression model according to the time series, with the target temperature of the fuel cell, the temperature of the reactant gas, the humidity of the reactant gas, the pressure of the reactant gas, the flow rate of the reactant gas, and the current of the fuel cell as inputs, and the cooling water flow rate and the cooling air volume as outputs.

[0094] More specifically, the training process of the echo state neural network regression model includes a sampling phase and a weight calculation phase, specifically including:

[0095] During the sampling phase, input samples need to be fed into the ESN's reservoir to generate corresponding internal states; and after the sampling phase is completed, a set of reservoir states corresponding to each input sample can be obtained.

[0096] For example, we first arbitrarily select an initial state for the network. Typically, the initial state is chosen to be 0, i.e., x(0) = 0. In this embodiment, we assume W... OB The input-to-output and output-to-output connection weights are also assumed to be 0.

[0097] Given an input signal sequence:

[0098] u(0), u(1), ..., u(P)

[0099] Training samples (u(t), t = 1, 2, ... P) are processed by the input connection weight matrix W LR It was added to the reserve pool.

[0100] The system state and output y(t) are calculated and collected sequentially:

[0101] y(0), y(1), ..., y(P)

[0102] To calculate the output connection weight matrix, we need to sample the internal state variables starting from a certain time m, and represent them as a vector (x1(i), x2(i), ... x... N (i), (i = m, m+1...P) form a matrix B(P-m+1, N) with rows, and the corresponding sample data y(t) are also collected and form a column vector T(P-m+1, 1).

[0103] During the weight calculation phase, a linear regression method is used to calculate the output weight matrix of the ESN to map the internal state to the desired output.

[0104] In a specific example, the weight calculation stage further includes:

[0105] Determine the relationships between the internal pool and the previous pool and input:

[0106] x(t+1)=f[W res ·x(t)+W LR ·u(t)]

[0107] Among them, W LR W is the input connection weight matrix, used to represent the mapping relationship between the input and the reserve pool; RO The output connection weight matrix represents the mapping relationship from the reservoir to the output; f is the activation function, and W... res This is the internal connection weight matrix, used to represent the connection relationships among the elements in the vector x(t) of the reserve pool;

[0108] Based on the system state matrix and sample data collected during the sampling phase, the output connection weight matrix W is calculated.RO ;include:

[0109] The objective function is obtained as follows:

[0110]

[0111] Wherein, there is a linear relationship between the state variable x(t) and the predicted output y(t), where y(t) is the predicted output. The expected output;

[0112] When the system's mean square error is minimized, i.e., the desired objective is achieved, the output connection weight matrix W is calculated at this point. RO :

[0113]

[0114] Among them, W RO x(t) is the network output.

[0115] Understandably, when making time series forecasts, the output is usually used as the input to continuously make predictions.

[0116] Understandably, in one example, during training, where t = m, m = 1, ..., P, it means that the output results of the first P time steps are used to train the output layer weight matrix. Here, the time step refers to the different states of the regressive state neural network when it receives and processes the input data at different time points.

[0117] Specifically, during training, the regressive state neural network first warms up with the initial state and input data, generating some internal state values ​​in the process. This warm-up time is typically not used to train the output layer weight matrix.

[0118] After the warm-up process, the E-regression state neural network begins to generate predicted outputs and trains the output layer weight matrix based on the first P predicted outputs and the actual outputs in the training set. In this process, t = m, m = 1, ..., P means that the predicted outputs from the first step (m = 1) to the Pth step (m = P) of the network's prediction output generation are used to train the output layer weight matrix.

[0119] This training method allows the regressive state neural network to learn and predict only the output results of the first P time steps, discarding the prediction output results of subsequent steps. This reduces the influence of past noisy data on the network, improving its prediction accuracy and robustness. Furthermore, training only on the output results of the first P time steps also improves the network's training efficiency. However, it is important to note that in practical applications, the value of P needs to be carefully chosen to balance the network's stability and training effectiveness.

[0120] Step S122: During training and testing, calculate the root mean square error (RMSE) and goodness-of-fit (R-squared) of the output features using the test dataset. 2 ).

[0121] When the training result of the echo state neural network regression model is not overfitting and exhibits high accuracy, that is, when the root mean square error (RMSE) between the predicted and expected values ​​is close to 0 and the goodness of fit (R-squared(R0)) is good enough, the model is considered to have goodness of fit. 2 When the condition approaches 1, training ends, and the echo state neural network prediction model is obtained. After training on the training set data through the above process, the obtained echo state neural network prediction model can be used for specific time series modeling problems.

[0122] Step S13: Test the trained echo state neural network prediction model on the test dataset to verify the model's prediction accuracy and obtain the trained neural network prediction model.

[0123] More specifically, in step S13, the test data has the same dimension as the training data and is input according to a time series.

[0124] Furthermore, it includes optimizing the parameters of the fuel cell stack by optimizing the reservoir size, reservoir radius, reservoir sparsity, and input cell scale parameters to meet the required fitting accuracy.

[0125] Step S14: Real-time data collection of the fuel cell stack in operation, setting target temperature values ​​for the fuel cell stack inlet and outlet, and obtaining prediction signals using the trained neural network prediction model. These prediction signals include cooling water flow rate and cooling airflow. Based on these prediction signals, the operating status of the cooling water pump and radiator is controlled to maintain the fuel cell stack temperature at a reasonable level. It is understood that the application environment of the fuel cell stack in operation is similar to... Figure 2 As shown. This fuel cell stack can be used in applications such as vehicles.

[0126] It is understood that in the method provided by the present invention, the cooling water flow rate and cooling air volume signals are predicted based on the trained echo state neural network model; the cooling water pump and radiator fan are controlled according to the signals, thereby performing thermal management on the fuel cell stack, and finally completing the control of the fuel cell stack cooling water inlet temperature and fuel cell stack cooling water outlet temperature.

[0127] The method provided in this invention improves the accuracy of fuel cell thermal management by minimizing the error between the inlet and outlet temperatures of the fuel cell stack and the target temperature. An echo state neural network is applied to the field of fuel cell thermal management. The echo state neural network is trained using a dataset obtained from fuel cell stack testing. The trained prediction model controls the cooling water flow rate and cooling airflow signals, ultimately optimizing the thermal management of the fuel cell system.

[0128] Meanwhile, the echo state neural network model based on deep learning used in this embodiment is suitable for the control of nonlinear systems, does not require a specific mathematical model, and can cope with situations such as drastic changes in the system's operating point; the optimized controller can better resist changes in external loads, making the error between the inlet and outlet temperatures and the target temperature values ​​smaller, and can be effectively applied to the thermal management of fuel cells in high-power hybrid electric vehicles, with greater advantages in accuracy and stability.

[0129] like Figure 7 The diagram shows a schematic representation of an embodiment of a neural network-based fuel cell thermal management device provided by the present invention. In this embodiment, the neural network-based fuel cell thermal management device 1 includes at least:

[0130] Data set acquisition unit 10 is used to run a test fuel cell stack on a test bench, obtain an initial dataset in time sequence, and form a training dataset and a test dataset;

[0131] Regression model acquisition unit 11 is used to establish an echo state neural network regression model;

[0132] The training processing unit 12 is used to train the established echo state neural network regression model according to the training dataset to obtain the echo state neural network prediction model.

[0133] The test processing unit 13 is used to test the trained echo state neural network prediction model on the test dataset to verify the model's prediction accuracy and obtain the trained neural network prediction model.

[0134] The prediction processing unit 14 is used to collect data of the fuel cell stack in operation in real time, set the target temperature value of the fuel cell stack inlet and outlet, obtain prediction signals using the trained neural network prediction model, the prediction signals include: cooling water flow rate and cooling air volume, and control the operation status of the cooling water pump and radiator according to the prediction signals to control the temperature of the fuel cell stack at a reasonable level.

[0135] For more details, please refer to and combine with the aforementioned points. Figures 1 to 6 The description of that will not be repeated here.

[0136] Accordingly, as another aspect of the present invention, a fuel cell thermal management system based on a neural network is also provided, the structure of which can be referred to Figure 2 As shown, the system includes at least: a fuel cell stack, a water tank, a water pump, and a radiator connected in a loop;

[0137] Temperature sensors are installed at the inlet and outlet of the fuel cell stack;

[0138] The controller is connected to the fuel cell stack, water pump, temperature sensor, and radiator. The controller includes an operating unit for performing the aforementioned operations. Figures 1 to 6 A neural network-based thermal management method for fuel cells is described.

[0139] For more details, please refer to and combine with the aforementioned points. Figures 1 to 6 The description of that will not be repeated here.

[0140] Implementing the embodiments of the present invention has the following beneficial effects:

[0141] This invention provides a neural network-based method, apparatus, and system for fuel cell thermal management. By improving the accuracy of fuel cell thermal management, aiming to reduce the error between the inlet and outlet temperatures of the fuel cell stack and the target temperature, an echo state neural network is applied to the field of fuel cell thermal management. The echo state neural network is trained using a dataset obtained from fuel cell stack testing. The trained prediction model controls the cooling water flow rate and cooling airflow signals, ultimately optimizing the thermal management of the fuel cell system. This improves the accuracy and efficiency of temperature control in the fuel cell stack.

[0142] In this embodiment, since the hidden layer of the echo state neural network is a dynamic reservoir structure with echo state properties, it not only enhances the stability of network prediction, but also simplifies the training process by using only a linear algorithm to obtain the network output weights. At the same time, it overcomes the problems of slow convergence speed and easy getting trapped in local minima in traditional neural networks.

[0143] Meanwhile, the regression model obtained after training shows a significant improvement in the stability and accuracy of temperature control compared to control theory-based methods. It also has better temperature regulation capabilities than traditional neural networks, can better resist external load disturbances, and has a smaller deviation from the set value.

[0144] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0145] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device for specifying modules in one or more boxes.

[0146] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fuel cell thermal management method based on neural networks, characterized in that, It should include at least the following steps: Step S10: Run the test fuel cell stack on the test bench, obtain the initial dataset in time sequence, and form the training dataset and the test dataset; Step S11: Establish an echo state neural network regression model; Step S12: Train the established echo state neural network regression model based on the training dataset to obtain the echo state neural network prediction model. Step S13: Test the trained echo state neural network prediction model on the test dataset to verify the model's prediction accuracy and obtain the trained neural network prediction model. Step S14: Collect data of the fuel cell stack in operation in real time, obtain prediction signals using the trained neural network prediction model, and control the operation of the cooling water pump and radiator according to the prediction signals to control the temperature of the fuel cell stack. Step S10 further includes: The test fuel cell stack was run on the test bench, and data on the target temperature of the fuel cell, the temperature of the reactant gas, the humidity of the reactant gas, the pressure of the reactant gas, the flow rate of the reactant gas, the current of the fuel cell, the flow rate of the cooling water and the cooling air volume were obtained at each time step through multiple sensors to form an initial dataset. Step S12 further includes: The data from the training dataset is selected and trained according to the time series, with the target temperature of the fuel cell, the temperature of the reactant gas, the humidity of the reactant gas, the pressure of the reactant gas, the flow rate of the reactant gas, and the current of the fuel cell as inputs, and the cooling water flow rate and the cooling air volume as outputs. The echo state neural network regression model is trained accordingly.

2. The method as described in claim 1, characterized in that, Step S10 further includes: The initial dataset is preprocessed, including standardization and normalization. The preprocessed initial dataset is divided into a training dataset and a test dataset.

3. The method as described in claim 2, characterized in that, Step S11 includes at least the following: An echo state neural network regression model is established, which has an input layer, a reservoir, and an output layer; the input layer takes a time series data as input, and the output layer outputs another time series data related to the input. Initializing the echo state neural network regression model includes: Determine the number of neuron nodes in the reserve pool; An internal connection matrix is ​​randomly generated, which represents the connection state between each neuron in the reserve pool, including the direction and weight of the connection; Determine the weights of the input connection matrix and the spectral radius of the internal connection matrix of the input layer; At the end of the initialization process, it is necessary to set the weights of the input connections and the spectral radius of the internal connection matrix.

4. The method as described in claim 3, characterized in that, Step S12 further includes: The initialized echo state neural network regression model is preheated to form an internal state with time series characteristics; When the training result of the echo state neural network regression model is that it is not overfitting and has high accuracy, that is, when the root mean square error between the predicted value and the expected value is close to 0 and the goodness of fit is close to 1, the training ends and the echo state neural network prediction model is obtained.

5. The method as described in claim 4, characterized in that, The specific steps of training the echo state neural network regression model include: During the sampling phase, input samples need to be fed into the ESN's reservoir to generate corresponding internal states; and after the sampling phase is completed, a set of reservoir states corresponding to each input sample is obtained. During the weight calculation phase, a linear regression method is used to calculate the output weight matrix of the ESN to map the internal state to the desired output.

6. The method as described in claim 5, characterized in that, The weight calculation stage further includes: Determine the relationships between the internal pool and the previous pool and input: in, The input connection weight matrix is ​​used to represent the mapping relationship between the input and the reserve pool; The output connection weight matrix is ​​used to represent the mapping relationship from the reserve pool to the output; It is an activation function. This is the internal connection weight matrix, used to represent the connection relationships between neurons in the reservoir; Based on the system state matrix and sample data collected during the sampling phase, the output connection weight matrix is ​​calculated. ;include: The objective function is obtained as follows: Among them, state variables and predicted output The relationship between them is linear. To predict the output, The expected output; The output connection weight matrix is ​​calculated to minimize the system's mean square error. : in, For network output.

7. The method as described in claim 4, characterized in that, In step S13, the test data has the same dimension as the training data and is input according to a time series. Furthermore, it includes optimizing the parameters of the fuel cell stack by optimizing the reservoir size, reservoir radius, reservoir sparsity, and input cell scale parameters to meet the required fitting accuracy.

8. The method according to any one of claims 1 to 7, characterized in that, The fuel cell stack is a fuel cell stack that uses a proton exchange membrane fuel cell.

9. A fuel cell thermal management device based on a neural network, characterized in that, At least including: The dataset acquisition unit is used to run the test fuel cell stack on the test bench, obtain the initial dataset in time sequence, and form the training dataset and the test dataset. Specifically, by running the test fuel cell stack on the test bench and using multiple sensors, the initial dataset is formed by obtaining data on the target temperature of the fuel cell, the temperature of the reactant gas, the humidity of the reactant gas, the pressure of the reactant gas, the flow rate of the reactant gas, the current of the fuel cell, the flow rate of the cooling water, and the cooling air volume at each time sequence. The regression model acquisition unit is used to establish an echo state neural network regression model. The training processing unit is used to train the established echo state neural network regression model based on the training dataset to obtain the echo state neural network prediction model. Specifically, by selecting data from the training dataset and training the echo state neural network regression model according to the time series, the target temperature of the fuel cell, the temperature of the reactant gas, the humidity of the reactant gas, the pressure of the reactant gas, the flow rate of the reactant gas, and the current of the fuel cell are used as inputs, and the cooling water flow rate and the cooling air volume are used as outputs. The test processing unit is used to test the trained echo state neural network prediction model on the test dataset to verify the model's prediction accuracy and obtain the trained neural network prediction model. The predictive processing unit is used to collect data of the fuel cell stack in operation in real time, obtain predictive signals using the trained neural network predictive model, and control the operation of the cooling water pump and radiator according to the predictive signals to control the temperature of the fuel cell stack.

10. A fuel cell thermal management system based on a neural network, characterized in that, include: A fuel cell stack, water tank, water pump, and radiator connected in a loop; Temperature sensors are installed at the inlet and outlet of the fuel cell stack; A controller is connected to the fuel cell stack, water pump, temperature sensor and radiator. The controller is equipped with an operating unit for executing the neural network-based fuel cell thermal management method as described in any one of claims 1 to 8.

Images