Proton exchange membrane fuel cell thermal management method based on fuzzy neural network

Abstract

The application discloses a proton exchange membrane fuel cell thermal management method based on a fuzzy neural network and belongs to the technical field of fuel cell thermal management. The method comprises the following steps: obtaining an output current value and a cooling water inlet temperature value of a proton exchange membrane fuel cell stack at a current moment; inputting the output current value and the cooling water inlet temperature value into a fuzzy neural network model constructed in advance, and driving the fuzzy neural network model to perform forward calculation and output a cooling water outlet target temperature value corresponding to the current moment; inputting a temperature deviation signal into a preset multilayer feedback regulation network, and generating a fan rotating speed control instruction and a water pump flow control instruction of the current moment according to the temperature deviation signal; and adjusting the rotating speed of a radiator fan according to the fan rotating speed control instruction, and adjusting the flow of a cooling water pump according to the water pump flow control instruction, so as to change the heat dissipation of the stack.

Description

Technical Field

[0001] This invention relates to the field of fuel cell thermal management technology, specifically to a proton exchange membrane fuel cell thermal management method based on fuzzy neural networks. Background Technology

[0002] During operation, the output power of a proton exchange membrane fuel cell stack is closely related to the rate of heat generation from internal chemical reactions. Excessive or insufficient stack temperature can affect the performance and lifespan of the membrane electrode assembly (MEA). Maintaining a suitable and stable operating temperature is crucial for ensuring efficient system operation. Existing thermal management methods generally employ strategies based on traditional proportional-integral-derivative (PID) control or single-rule lookup tables. These methods directly adjust fan speed or water pump flow based on preset fixed control logic by detecting the actual temperature of the cooling water outlet. This approach exhibits significant lag in temperature control response when faced with drastic fluctuations in the operating conditions of fuel cell vehicles. When the stack output current rapidly increases, heat generation surges, and the fixed control parameters cannot quickly detect changes in operating conditions and adjust heat dissipation in time, leading to significant overshoot in stack temperature, directly threatening the safety of the MEA. Conversely, when the current drops sharply, excessive heat dissipation can easily cause temperature overshoot, leading to membrane drying or flooding. Furthermore, existing technologies typically control the fan and water pump as independent loops or use simple linked proportional control, failing to effectively decouple and collaboratively optimize the dynamic response of the two actuators, resulting in high system energy consumption and insufficient temperature stability. While fuzzy neural networks have demonstrated their ability to handle nonlinearity and uncertainty in some industrial process control applications, a significant challenge remains in the field of proton exchange membrane fuel cell thermal management: accurately and in real-time determining the target cooling water outlet temperature under different combinations of current and inlet water temperature, given the highly dynamic characteristics of the fuel cell stack's thermal load. Furthermore, after obtaining the target temperature deviation, another critical technical bottleneck is the development of a multi-level feedback regulation mechanism capable of differentiated and parallel optimization of the fan and pump's individual response characteristics, eliminating temperature deviations while simultaneously considering system energy consumption. Summary of the Invention

[0003] The purpose of this invention is to provide a thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks. The method uses a fuzzy neural network model to dynamically generate the target cooling water outlet temperature value in real time based on the output current and the cooling water inlet temperature. The method also generates fan speed control commands and water pump flow control commands through a multi-layer feedback control network containing two independent control sub-networks, thereby achieving precise and rapid control of the fuel cell stack temperature.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks, comprising:

[0005] Obtain the output current value and cooling water inlet temperature value of the proton exchange membrane fuel cell stack at the current moment.

[0006] The output current value and the cooling water inlet temperature value are input into a pre-built fuzzy neural network model, and the fuzzy neural network model is driven to perform forward calculations to output the target cooling water outlet temperature value corresponding to the current moment.

[0007] The target temperature value of the cooling water outlet is compared with the actual temperature value of the cooling water outlet collected in real time to generate a temperature deviation signal.

[0008] The temperature deviation signal is input to a preset multi-layer feedback control network, which generates a fan speed control command and a water pump flow control command for the current moment based on the temperature deviation signal.

[0009] The fan speed is adjusted according to the fan speed control command, and the flow rate of the cooling water pump is adjusted according to the water pump flow control command, so as to change the heat dissipation of the fuel cell stack.

[0010] As a technical solution of the present invention, obtaining the output current value and cooling water inlet temperature value of the proton exchange membrane fuel cell stack at the current moment specifically includes: acquiring the main circuit current signal at the output end of the stack through a current sensor, and obtaining the output current value after analog-to-digital conversion of the main circuit current signal; acquiring the inlet water temperature analog signal through a thermocouple installed at the cooling water inlet pipe of the stack, and obtaining the cooling water inlet temperature value after analog-to-digital conversion and filtering of the inlet water temperature analog signal. Preferably, the sampling frequency of the analog-to-digital conversion is 100Hz, and the filtering process uses a moving average filtering algorithm to smooth multiple continuously acquired inlet water temperature analog signals, thereby effectively suppressing on-site noise interference and ensuring the stability and accuracy of the input signal.

[0011] As a technical solution of the present invention, the output current value and the cooling water inlet temperature value are input into a pre-constructed fuzzy neural network model, and the fuzzy neural network model is driven to perform forward calculation to output the target cooling water outlet temperature value corresponding to the current moment. Specifically, this includes: using the output current value and the cooling water inlet temperature value as input quantities to the first input node and the second input node of the fuzzy neural network model, respectively, and performing fuzzification processing on the input quantities of the first input node and the second input node to obtain a current membership vector and a temperature membership vector; matching the current membership vector and the temperature membership vector with the first fuzzy rule layer node and the second fuzzy rule layer node in the fuzzy neural network model, respectively, to activate the corresponding fuzzy rule consequent parameters; inputting the activated fuzzy rule consequent parameters into the output layer node of the fuzzy neural network model, and performing weighted summation calculation on the fuzzy rule consequent parameters through the output layer node to output the target cooling water outlet temperature value.

[0012] Preferably, fuzzification processing is performed on the input values ​​of the first input node and the second input node to obtain a current membership vector and a temperature membership vector. This includes: for the first input node, calculating the current membership value of the output current value on multiple current fuzzy subsets using a preset current membership function, and combining all the current membership values ​​into the current membership vector, wherein the number of current fuzzy subsets is at least three; for the second input node, calculating the temperature membership value of the cooling water inlet temperature value on multiple temperature fuzzy subsets using a preset temperature membership function, and combining all the temperature membership values ​​into the temperature membership vector, wherein the number of temperature fuzzy subsets is at least three. More preferably, both the current membership function and the temperature membership function are Gaussian membership functions to smoothly describe the membership relationships of the input values ​​and enhance the model's adaptability to changes in operating conditions.

[0013] Preferably, the current membership vector and the temperature membership vector are matched with the first and second fuzzy rule layer nodes in the fuzzy neural network model to activate the corresponding fuzzy rule consequent parameters. This includes: comparing each current membership value in the current membership vector with the current prerequisite parameters stored in the first fuzzy rule layer node, and selecting the current fuzzy subset corresponding to the current membership value with the highest matching degree as the activated current subset; comparing each temperature membership value in the temperature membership vector with the temperature prerequisite parameters stored in the second fuzzy rule layer node, and selecting the temperature fuzzy subset corresponding to the temperature membership value with the highest matching degree as the activated temperature subset; and calling the fuzzy rule consequent parameter corresponding to the combined index of the activated current subset and the activated temperature subset from the rule consequent parameter library of the fuzzy neural network model, and using the called fuzzy rule consequent parameter as the activated fuzzy rule consequent parameter. In this way, the model can automatically and accurately correspond to the optimal fuzzy rule under the current operating condition, realizing refined reasoning for the target temperature value of the cooling water outlet.

[0014] As a technical solution of the present invention, the temperature deviation signal is input to a preset multi-layer feedback control network. The multi-layer feedback control network generates a fan speed control command and a water pump flow control command at the current moment based on the temperature deviation signal. This includes: inputting the temperature deviation signal to a first control sub-network and a second control sub-network in the multi-layer feedback control network; the first control sub-network performing proportional-integral-derivative (PID) calculation on the temperature deviation signal to obtain a basic fan speed control quantity, and outputting the basic fan speed control quantity as the fan speed control command after amplitude limiting; the second control sub-network performing PID calculation on the temperature deviation signal to obtain a basic water pump flow control quantity, and outputting the basic water pump flow control quantity as the water pump flow control command after amplitude limiting.

[0015] Preferably, the proportional coefficient, integral coefficient, and derivative coefficient in the first and second regulating subnetworks are dynamically updated using a fuzzy rule-based online self-adjustment method, enabling the regulating parameters to be optimized in real time according to the actual operating state of the system, effectively alleviating the problem of insufficient adaptability of traditional fixed parameter controllers in nonlinear, time-varying thermal management systems.

[0016] More preferably, the first regulating sub-network performs proportional-integral-derivative (PID) calculations on the temperature deviation signal to obtain a basic fan speed control quantity, and outputs the basic fan speed control quantity as the fan speed control command after amplitude limiting. Specifically, this includes: simultaneously inputting the temperature deviation signal into the proportional calculation path, integral calculation path, and derivative calculation path of the first regulating sub-network; the proportional calculation path multiplies the temperature deviation signal by a preset first proportional coefficient to obtain a proportional component; the integral calculation path performs time accumulation calculations on the temperature deviation signal and then multiplies it by a preset first integral coefficient to obtain an integral component; the derivative calculation path performs time accumulation calculations on the temperature deviation signal and then multiplies it by a preset first integral coefficient to obtain an integral component; the derivative calculation path performs time accumulation calculations on the temperature deviation signal... The time derivative is calculated and multiplied by a preset first differential coefficient to obtain the differential component. The proportional component, integral component, and differential component are added to obtain the basic fan speed control value. This basic fan speed control value is then compared with preset upper and lower limits for fan speed. If the basic fan speed control value is greater than the upper limit, the upper limit is output as the fan speed control command. If the basic fan speed control value is less than the lower limit, the lower limit is output as the fan speed control command. Otherwise, the basic fan speed control value is directly output as the fan speed control command. Through limiting protection, damage to the actuator due to over-limit commands is effectively prevented, ensuring the safe and reliable operation of the cooling system.

[0017] As a technical solution of the present invention, before inputting the output current value and the cooling water inlet temperature value into the pre-constructed fuzzy neural network model, the method further includes a model construction and training process: acquiring historical operating data of the proton exchange membrane fuel cell stack at multiple steady-state operating points, the historical operating data including historical output current value, historical cooling water inlet temperature value, historical cooling water outlet temperature value, and historical ambient temperature value; using the backpropagation algorithm, using the historical operating data as training samples to iteratively train the initial fuzzy neural network model, adjusting the fuzzy rule prerequisite parameters and fuzzy rule consequent parameters in the initial fuzzy neural network model until the model error meets the preset accuracy requirements, thereby obtaining the pre-constructed fuzzy neural network model.

[0018] Preferably, the step of using the backpropagation algorithm to iteratively train the initial fuzzy neural network model using the historical operating data as training samples includes: selecting a training sample from the historical operating data; inputting the historical output current value and historical cooling water inlet temperature value from the training sample into the initial fuzzy neural network model; performing forward calculation to obtain the predicted cooling water outlet temperature value output by the model; calculating the mean square error between the predicted cooling water outlet temperature value and the historical cooling water outlet temperature value in the training sample; passing the mean square error value layer by layer along the backpropagation path of the initial fuzzy neural network model, and sequentially calculating the parameter gradient values ​​of each layer in the initial fuzzy neural network model; and synchronously updating the fuzzy rule prerequisite parameters and fuzzy rule consequent parameters in the initial fuzzy neural network model using the gradient descent method based on the parameter gradient values.

[0019] Further preferably, the step of passing the mean squared error value layer by layer along the backpropagation path of the initial fuzzy neural network model and sequentially calculating the parameter gradient values ​​of each layer in the initial fuzzy neural network model includes: passing the mean squared error value to the output layer node of the initial fuzzy neural network model, calculating the first gradient value of the output weight of the output layer node with respect to the mean squared error value, and updating the output weight of the output layer node according to the first gradient value; passing the first gradient value to the fuzzy rule layer node of the initial fuzzy neural network model, calculating the second gradient value of the consequent parameter of the fuzzy rule layer node with respect to the mean squared error value, and updating the consequent parameter of the fuzzy rule layer node according to the second gradient value; passing the second gradient value to the fuzzification layer node of the initial fuzzy neural network model, calculating the third gradient value of the membership function center parameter and membership function width parameter of the fuzzification layer node with respect to the mean squared error value, and updating the membership function center parameter and membership function width parameter according to the third gradient value. This training method enables the fuzzy neural network model to automatically learn and extract the mapping rules of the stack's thermal characteristics from historical data, and establish a high-precision nonlinear model between the output current, inlet water temperature and desired outlet water temperature, laying the foundation for the accurate setting of the feedforward target temperature.

[0020] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0021] By constructing and applying a fuzzy neural network model, the fuel cell stack output current and cooling water inlet temperature are used as inputs to drive the model to perform forward calculations and output the target cooling water outlet temperature at the current moment. This model learns and internalizes the optimal temperature distribution of the fuel cell stack under different current loads and inlet thermal boundary conditions from historical operating data. Its internal fuzzy inference mechanism can accurately express the complex and continuous nonlinear mapping relationship between output current, inlet water temperature, and target outlet temperature. When operating conditions change dynamically, the model can calculate the target temperature precisely matching the current heat generation state and heat dissipation capacity within milliseconds based on real-time collected current and inlet water temperature information, replacing the lag-based open-loop setting method that relies on fixed curves or offline tables in traditional technologies. This solution fundamentally solves the problem that inaccurate target temperature settings cause subsequent feedback control to always adjust around an incorrect benchmark, enabling the entire temperature control system to proactively sense changes in heat load in a feedforward manner, creating conditions for rapidly suppressing temperature fluctuations. By designing a multi-layered feedback control network and inputting the temperature deviation signal in parallel to the first and second independent control sub-networks, each sub-network independently executes a complete control law processing including proportional, integral, and derivative calculations. Based on the distinct dynamic response characteristics of the fan cooling loop and the water pump flow loop, and the actuator constraints, it generates fan speed control commands and water pump flow control commands respectively. This network architecture breaks through the limitations of simple linkage or single-input single-output control of two actuators. Employing a dual-channel parallel feedback structure, it enables the fan and the high-heat-capacity cooling water circulation system to generate different time-scale combinations of adjustment actions based on the same deviation signal. This structure allows the fan speed to respond quickly to provide immediate cooling when a temperature deviation occurs, while the water pump flow rate adjusts at a smoother rate to maintain stable heat exchange within the flow channel. This avoids temperature regulation oscillations and overcompensation caused by sudden changes in a single control variable, achieving precise elimination and stable control of fuel cell stack temperature deviations. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0023] Figure 1 This is a flowchart of a thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks;

[0024] Figure 2 This is a flowchart of the process for obtaining the output current and cooling water inlet temperature of a proton exchange membrane fuel cell stack.

[0025] Figure 3This is a flowchart of the cooling water outlet target temperature calculation based on a fuzzy neural network;

[0026] Figure 4 This is a schematic diagram of a multi-layer feedback regulation network;

[0027] Figure 5 This is a flowchart of the online update process for fuzzy self-tuning PID control coefficients;

[0028] Figure 6 This is a flowchart of the offline construction process of a fuzzy neural network model for thermal management of a proton exchange membrane fuel cell.

[0029] Figure 7 These are the sampling and filtering curves of the output current and cooling water inlet temperature of the proton exchange membrane fuel cell stack;

[0030] Figure 8 These are curves showing the relationship between the target outlet temperature of the cooling water and the output current at different cooling water inlet temperatures.

[0031] Figure 9 These are the temperature deviation signal of the thermal management system and the response curves of the fan speed and water pump flow control commands;

[0032] Figure 10 This is the dynamic change curve of the PID coefficient of the fan speed regulation subnetwork based on the temperature deviation signal. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] See Figure 1This invention provides a thermal management method for proton exchange membrane fuel cells (PEMFCs) based on a fuzzy neural network. The method acquires the current output current and cooling water inlet temperature of the PEMFC stack at the current moment, inputs these values ​​into a pre-constructed fuzzy neural network model, drives the model to perform forward calculations, and outputs the target cooling water outlet temperature for the current moment. The target cooling water outlet temperature is compared with the real-time acquired actual cooling water outlet temperature to generate a temperature deviation signal. This temperature deviation signal is then input into a pre-defined multi-layer feedback control network. The multi-layer feedback control network generates fan speed control commands and water pump flow control commands based on the temperature deviation signal. The fan speed control command adjusts the radiator fan speed, and the water pump flow control command adjusts the cooling water pump flow rate to change the heat dissipation of the fuel cell stack.

[0035] Example 1:

[0036] In specific implementation, please refer to Figure 2 The process of acquiring the output current value of a proton exchange membrane fuel cell stack is as follows: A current sensor is connected in series to the main circuit at the stack output end. The current sensor collects the main circuit current signal in real time and outputs an analog voltage proportional to the current. The analog voltage is sent to an analog-to-digital converter (ADC), which continuously samples the main circuit current signal at a sampling frequency of 100Hz. After each sampling, the analog voltage is quantized into a digital quantity, which is the output current value at the current moment. The sampling frequency is set to 100Hz because the dynamic frequency components of the stack output current under load changes are mainly concentrated in the range of 0-20Hz. A sampling frequency of 100Hz satisfies the Nyquist sampling theorem requirement for complete signal reconstruction, while avoiding excessive redundant data generation by the ADC due to an excessively high sampling rate, thereby reducing the computational load of subsequent processing units.

[0037] The process of obtaining the cooling water inlet temperature is as follows: A thermocouple is installed on the inner wall of the cooling water inlet pipe of the fuel cell stack. The temperature measuring end of the thermocouple directly contacts the cooling water flow, converting the cooling water temperature into a thermoelectric potential. After cold junction compensation and amplification by a signal conditioning circuit, the thermoelectric potential forms an analog signal of the inlet water temperature. The analog signal of the inlet water temperature is sent to the same analog-to-digital converter or another analog-to-digital converter, and converted to digital at a sampling frequency of 100Hz to obtain a raw digital sequence of inlet water temperatures arranged in chronological order. Considering the possible local turbulence and thermocouple background noise in the cooling water inlet pipe, random fluctuation signals may be superimposed on the raw digital sequence of inlet water temperatures. Therefore, after analog-to-digital conversion, a moving average filtering algorithm is used to smooth the multiple continuously acquired raw digital signals of inlet water temperature to obtain the cooling water inlet temperature value at the current moment. The expression of the moving average filtering algorithm is:

[0038]

[0039] in, Indicates the first The raw digital values ​​of the inlet water temperature obtained after analog-to-digital conversion at each sampling time point This represents the length of the sliding window in the moving average filtering algorithm. This is the index of the sampling point within the sliding window. The range of values ​​is arrive integers, Indicates the first The cooling water inlet temperature value output after moving average filtering at each sampling time. Sliding window length. The value is 8, and the basis for this value is: through offline analysis of different... The root mean square error and response delay time of the filtered temperature signal are taken as values. At the same time, the attenuation of random noise reduces the signal fluctuation range to less than 1 / 3 of the unfiltered state, and the response delay to temperature step changes is less than 0.5 seconds, which can simultaneously meet the requirements of anti-interference capability and real-time performance.

[0040] The output current value and cooling water inlet temperature value obtained through the above steps are used as inputs to the fuzzy neural network model for subsequent forward calculations.

[0041] See Figure 7 In the graph, the horizontal axis represents time in seconds, the left side of the vertical axis represents the output current in amperes, and the right side represents the cooling water inlet temperature in degrees Celsius. The solid blue line represents the change in the output current of the proton exchange membrane fuel cell stack, the dashed orange line represents the original sampled value of the cooling water inlet temperature, and the dotted green line represents the cooling water inlet temperature after processing by the moving average filtering algorithm.

[0042] As shown in the figure, the output current value remains around 30A from 0 seconds to approximately 2.0 seconds, then experiences a significant jump at 2.0 seconds, rapidly rising to approximately 50A and remaining stable. This jump corresponds to the dynamic response of the current sensor and analog-to-digital converter sampling the output current value in Example 1, demonstrating the rapid response of the current value to load changes.

[0043] The original value of the cooling water inlet temperature (orange dashed line) exhibits significant random fluctuations throughout the time range, ranging from approximately 37°C to 54°C, reflecting the characteristics of thermocouple temperature measurement being affected by local turbulence and background noise. After moving average filtering (green dotted line), the inlet temperature curve is significantly smoother, with the fluctuation amplitude reduced to about 1 / 3 of the original signal. Furthermore, the overall temperature value shows a slow upward trend, gradually increasing from approximately 43°C to approximately 50.5°C. This processing conforms to the technical solution in Example 1, which uses a moving average algorithm with a sliding window length of 8 to filter the inlet temperature signal, balancing anti-interference capability and real-time response.

[0044] Example 2:

[0045] In specific implementation, please refer to Figure 3 The fuzzy neural network model is pre-built through training. The structure of the fuzzy neural network model includes a first input node, a second input node, a fuzzification layer, a first fuzzy rule layer node, a second fuzzy rule layer node, a rule consequent parameter library, and an output layer node. The output current value obtained through the processing described in Example 1 is input to the first input node of the fuzzy neural network model, and the cooling water inlet temperature value obtained through the processing described in Example 1 is input to the second input node of the fuzzy neural network model.

[0046] The fuzzification layer performs fuzzification processing on the input values ​​of the first and second input nodes. For the first input node, the fuzzification layer stores a preset current membership function, which is a Gaussian membership function defined by a center parameter and a width parameter. These parameters are determined during the training phase of the fuzzy neural network model using a backpropagation algorithm. The current membership function contains at least three current fuzzy subsets, each corresponding to a set of center and width parameters. The output current value is substituted into the Gaussian membership function corresponding to each current fuzzy subset to calculate the current membership value of the output current value in each subset. All current membership values ​​are combined according to a fixed order of the current fuzzy subsets to form a current membership vector. The total number of current fuzzy subsets is used to... express, At least for .

[0047] For the second input node, the fuzzification layer stores a preset temperature membership function. This temperature membership function also uses a Gaussian membership function, and its center and width parameters are determined during the fuzzy neural network model training phase. The temperature membership function contains at least three fuzzy temperature subsets, each corresponding to a set of center and width parameters. Substituting the cooling water inlet temperature value into the Gaussian membership function corresponding to each fuzzy temperature subset, the temperature membership value of the cooling water inlet temperature value on each fuzzy temperature subset is calculated. All temperature membership values ​​are then combined according to a fixed order of the fuzzy temperature subsets to form a temperature membership vector. The total number of temperature fuzzy subsets is used... express, At least for .

[0048] After fuzzification, the current membership vector and temperature membership vector are matched with the nodes of the first and second fuzzy rule layers, respectively. The first fuzzy rule layer nodes store current prerequisite parameters, which are index identifiers for each current fuzzy subset. During matching, each current membership value in the current membership vector is mapped to the corresponding index identifier of the current fuzzy subset in the first fuzzy rule layer node, and the current fuzzy subset with the largest current membership value is selected as the active current subset. The second fuzzy rule layer nodes store temperature prerequisite parameters, which are index identifiers for each temperature fuzzy subset. Each temperature membership value in the temperature membership vector is mapped to the corresponding index identifier of the temperature fuzzy subset in the second fuzzy rule layer node, and the temperature fuzzy subset with the largest temperature membership value is selected as the active temperature subset.

[0049] Based on the combined index formed by the activation current subset and the activation temperature subset, the fuzzy rule consequent parameters corresponding to the combined index are retrieved from the rule consequent parameter library of the fuzzy neural network model. The combined index is determined by the sequence number of the activation current subset. and the index of the activation temperature subset The structure and rules are stored in the parameter library. Each combination corresponds to a consequent parameter value, which is learned as an adjustable parameter through backpropagation during the training of the fuzzy neural network model. The call operation will index the combination. The value of the consequent parameter is extracted and used as the consequent parameter of the activated fuzzy rule.

[0050] The activated fuzzy rule consequent parameters are input to the output layer nodes of the fuzzy neural network model. The output layer nodes receive fuzzy rule consequent parameter inputs corresponding to all possible combinations, but only the fuzzy rule consequent parameters corresponding to activated combinations are non-zero; the fuzzy rule consequent parameter inputs corresponding to other inactive combinations are considered zero. The output layer nodes perform a weighted summation calculation on the inputs corresponding to all combinations and output the target cooling water outlet temperature value at the current time. The expression for the weighted summation calculation is:

[0051]

[0052] in, Indicates the target temperature value of the cooling water outlet; The index representing the fuzzy subset of the current. The range of values ​​is from arrive integers, This represents the total number of current fuzzy subsets. The minimum value is ; The index representing the fuzzy subset of temperature. The range of values ​​is from arrive integers, This represents the total number of temperature-fuzzy subsets. The minimum value is ; This indicates that it corresponds to the serial number. The current fuzzy subset and the sequence number are The output weights of the combination formed by the temperature fuzzy subsets are determined iteratively through the backpropagation algorithm during the training phase of the fuzzy neural network model and stored in the output layer nodes. This indicates that it corresponds to the serial number. The current fuzzy subset and the sequence number are The fuzzy rule consequent parameter input value of the combination formed by the temperature fuzzy subset, when a certain combination is activated. This is equivalent to retrieving the consequent parameter value corresponding to that combination from the rule consequent parameter library. When a certain combination is not activated... Values In a single forward computation, only one combination is activated, therefore This is equivalent to the product of the output weights corresponding to the activation combinations and the consequent parameter values. The output layer nodes are calculated as follows. Afterwards, The target cooling water outlet temperature value is output as the current moment and used in the subsequent temperature deviation signal generation process.

[0053] See Figure 8In the figure, the horizontal axis represents the output current of the proton exchange membrane fuel cell stack, ranging from 10A to 80A, and the vertical axis represents the target cooling water outlet temperature, in degrees Celsius, ranging from approximately 40℃ to 95℃. The figure shows five sets of target cooling water outlet temperature curves under different cooling water inlet temperatures: 40℃, 50℃, 60℃, 70℃, and 80℃. Each curve is labeled with a different color and line type.

[0054] As the curves in the graph show, the target cooling water outlet temperature consistently increases steadily with the increase of the output current. This aligns with the characteristics of the fuzzy neural network model described in Example 2, which performs forward calculations based on the current output current and the cooling water inlet temperature to output the target cooling water outlet temperature. The absence of significant abrupt changes in the curves indicates that the model output exhibits good continuity and stability.

[0055] Meanwhile, the cooling water inlet temperature has a significant impact on the cooling water outlet target temperature. When the inlet temperature is high (e.g., 80℃ and 70℃), the corresponding cooling water outlet target temperature curves are generally in the high-temperature range, with initial temperatures of approximately 82℃ and 72℃ respectively. As the output current increases, the outlet temperature slowly rises to approximately 93℃ and 83℃. When the inlet temperature is medium (60℃), the outlet target temperature starts at approximately 62℃ and reaches a maximum of approximately 73℃. When the inlet temperature is low (40℃ and 50℃), the outlet target temperature starts at approximately 42℃ and 52℃ respectively, reaching maximums of approximately 53℃ and 63℃. This reflects the fuzzy neural network model's fuzzification and rule-matching processing of the inlet temperature input, enabling it to reasonably adjust the outlet target temperature according to different inlet temperatures.

[0056] All curves exhibit a smooth, gradually increasing characteristic, reflecting the model's nonlinear response capability to the output current. Furthermore, there are no abnormal fluctuations within the output current range, validating the effectiveness of the fuzzy rule consequent parameters and output weights obtained during training. The overall curve distribution conforms to the physical law of cooling water temperature increasing with current load in fuel cell thermal management, and demonstrates the accurate temperature target prediction achieved through the fuzzy neural network model.

[0057] Example 3:

[0058] In specific implementation, please refer to Figure 4 The pre-defined multi-layer feedback control network comprises a first control sub-network and a second control sub-network that operate independently and in parallel. The real-time generated temperature deviation signal is simultaneously introduced into the input ports of both the first and second control sub-networks. The temperature deviation signal is obtained by subtracting the target cooling water outlet temperature from the real-time acquired actual cooling water outlet temperature. The temperature deviation signal is represented by the symbol... It means that, among them, This represents a continuous-time variable.

[0059] The first regulation subnetwork internally constructs three parallel computation paths: a proportional computation path, an integral computation path, and a differential computation path. Temperature deviation signal. They are simultaneously fed into these three computation paths.

[0060] In the proportional calculation path, the temperature deviation signal Multiplied by the preset first proportional coefficient To obtain the proportional components ,Right now First proportionality coefficient It is a constant positive real number. Its value is determined by offline frequency domain analysis and step response tuning based on the rated heat dissipation power of the fuel cell stack, the adjustable range of the radiator fan speed, and the desired system response speed. The typical value range is 0.5 to 5.0. The specific value is fixed and written into the first regulation sub-network during the system debugging stage.

[0061] In the integration calculation path, the temperature deviation signal The cumulative calculation of execution time specifically involves using either rectangular integration or trapezoidal integration to process the temperature deviation signal. Accumulation is performed on the time axis. Within one discretized control cycle, the accumulated value of the integral path can be expressed as... ,in Indicates the first Discrete control time. This indicates the duration of the control cycle. The result of the cumulative time calculation is multiplied by a preset first integral coefficient. The integral components are obtained. First integral coefficient The value of is such that the integral action is neither too strong, causing overshoot, nor too weak, causing excessively long elimination time during the system steady-state error elimination process. The typical value range is 0.05 to 0.5, which is determined during the system debugging phase.

[0062] In the differential calculation path, the temperature deviation signal The time derivative calculation is performed using a backward difference approximation in the discrete implementation. This involves calculating the change in temperature deviation signal between two adjacent control cycles and dividing by the control cycle duration to obtain an approximate differential value. This differential value is then multiplied by a preset first differential coefficient. The differential components are obtained. First differential coefficient The value of is such that the differential action can effectively suppress the rapid change trend of the temperature deviation signal without amplifying the measurement noise. The typical value range is 0.01 to 0.2, which is determined during the system debugging phase.

[0063] Within each control cycle, the first regulating sub-network will adjust the aforementioned proportional components. Integral components and differential components Summing yields the basic control value for fan speed. The expression for generating the basic control variable of fan speed is:

[0064]

[0065] in, Indicates the first The basic control value of the fan speed at any given time; This represents the first proportionality coefficient, with a value ranging from 0.5 to 5.0; Indicates the first Temperature deviation signal at any given time; This represents the first integral coefficient, with a value ranging from 0.05 to 0.5; Indicates from the initial time to the nth time. The cumulative value of the temperature deviation signal over time; This represents the first differential coefficient, with a value ranging from 0.01 to 0.2; Indicates the first Time and the The difference in temperature deviation signals at any given time; This indicates the duration of the control cycle, typically 0.1 or 0.2 seconds, synchronized with or multiplied by the analog-to-digital conversion sampling cycle.

[0066] Obtain the basic control value of fan speed Then, the first regulating sub-network performs limiting processing. The first regulating sub-network has a pre-stored upper limit value for the fan speed. and fan speed lower limit These two limits are set based on the physical speed range of the radiator fan. For example, the upper limit of fan speed corresponds to the control command value corresponding to the maximum allowable fan speed, and the lower limit of fan speed corresponds to the control command value corresponding to the minimum starting speed or zero speed of the fan. The basic control value for fan speed is... and and Comparison: If Greater than the upper limit of fan speed Then set the upper limit of fan speed. As a fan speed control command output; if Less than the lower limit of fan speed Then set the lower limit of fan speed. As a fan speed control command output; if If the value is neither greater than the upper limit of fan speed nor less than the lower limit of fan speed, then the basic control value of fan speed is directly set to... As a fan speed control command output.

[0067] The second adjustment subnetwork adopts the same proportional-integral-differential (PID) calculation architecture and amplitude limiting method as the first adjustment subnetwork. The second adjustment subnetwork internally includes a proportional calculation path, an integral calculation path, and a differential calculation path, each using a preset second proportional coefficient. Second integral coefficient Second differential coefficient Within each control cycle, the second regulating subnetwork will transmit the temperature deviation signal. Multiplying each component by the second proportional coefficient, then by the second integral coefficient after time accumulation, and finally by the second differential coefficient after time derivative calculation, yields the proportional component, integral component, and differential component. Adding these three components together gives the basic control quantity of the pump flow rate. Second proportionality coefficient The value range is from 0.5 to 5.0, and the second integral coefficient... The value range is from 0.05 to 0.5, and the second differential coefficient is... The value range is from 0.01 to 0.2, and each coefficient is independently tuned during the system commissioning phase based on the cooling water pump flow regulation characteristics and system response requirements. The basic control quantity of the water pump flow rate is... With the pre-stored upper limit of water pump flow and the lower limit of water pump flow rate The comparison process is as follows: If the basic control value of the water pump flow rate is greater than the upper limit value, the upper limit value is output as the water pump flow control command; if the basic control value is less than the lower limit value, the lower limit value is output as the water pump flow control command; otherwise, the basic control value is directly output as the water pump flow control command. The upper limit value corresponds to the control command value corresponding to the maximum allowable flow rate of the cooling water pump, and the lower limit value corresponds to the control command value corresponding to the minimum allowable flow rate or zero flow rate of the cooling water pump.

[0068] The fan speed control command output by the first regulation sub-network is sent to the radiator fan drive unit to adjust the radiator fan speed; the water pump flow control command output by the second regulation sub-network is sent to the cooling water pump drive unit to adjust the cooling water pump flow, thereby changing the heat dissipation of the fuel cell stack.

[0069] See Figure 9 In the graph, the horizontal axis represents time in seconds, and the left side of the vertical axis represents the temperature deviation signal. The unit is degrees Celsius (°C). The right side of the vertical axis represents the percentage (%) of the fan speed control command and the water pump flow control command. The legend indicates that the curves correspond to the temperature deviation signals. (Solid blue line), fan speed control command (dashed blue line), and water pump flow control command (dotted orange line).

[0070] As observed in the graph, the temperature deviation signal occurs within the time interval from 0 to approximately 5 seconds. The temperature fluctuates slightly within the range of -2°C to -1°C. The corresponding fan speed control command and water pump flow control command are stable between approximately 30% to 40% and 22% to 27%, respectively. This indicates that the system is in a negative temperature difference state, and both the fan and water pump are operating with lower control commands to maintain heat dissipation of the fuel cell stack.

[0071] At approximately 5 seconds, the temperature deviation signal... A significant jump occurred, rapidly changing from a negative value to a positive value, reaching approximately 3°C. Subsequently, it fluctuated periodically around 3°C, indicating that the deviation between the actual cooling water outlet temperature and the target temperature changed from negative to positive, and the system's temperature sensing status underwent a sudden change.

[0072] Correspondingly, the fan speed control command and the water pump flow control command also jumped simultaneously. The fan speed control command quickly increased from about 30% to about 80%, and the water pump flow control command increased from about 22% to about 65%. Both of them maintained a high level of fluctuation in the following period, reflecting that the multi-layer feedback regulation network adjusted the radiator fan speed and cooling water pump flow in a timely manner according to the temperature deviation signal, thereby enhancing the cooling capacity.

[0073] Overall, the surge in the temperature deviation signal triggered a rapid response in both the fan speed and water pump flow rate. Furthermore, the fan speed control command and the temperature deviation signal showed a largely synchronized trend, demonstrating that the PID control path of the first regulating sub-network effectively regulated the fan's operating state. The change in the water pump flow rate control command was slightly smaller than that of the fan control command, which met the design requirements for the tuning of the second regulating sub-network coefficients and the system response.

[0074] Example 4:

[0075] In specific implementation, please refer to Figure 5 The proportional, integral, and derivative coefficients within the first regulating sub-network are dynamically updated using a fuzzy rule-based online self-adjustment method. A fuzzy coefficient adjuster is built into the first regulating sub-network, performing a coefficient update operation at the beginning of each control cycle or after a fixed number of control cycles. The input to the fuzzy coefficient adjuster is the temperature deviation signal. and the rate of change of temperature deviation ,in, Indicates the first Discrete control time points. Temperature deviation rate of change. Calculate the temperature deviation signal of the current control cycle. Temperature deviation signal from the previous control cycle The difference is then divided by the control cycle duration. To obtain, that is .

[0076] The coefficient fuzzy adjuster performs fuzzification processing on the input quantity. Temperature deviation signal. The universe of discourse is divided into several fuzzy subsets, and the rate of change of temperature deviation The universe of discourse is also divided into several fuzzy subsets. This is for temperature deviation signals. The coefficient fuzzy adjuster stores temperature deviation membership functions, which employ either triangular or Gaussian membership functions. Each fuzzy subset of temperature deviation corresponds to a set of membership function parameters. Substituting the membership functions corresponding to each fuzzy subset of temperature deviation, we can calculate... Membership value on each fuzzy subset of temperature deviation. For the rate of change of temperature deviation. The coefficient fuzzy adjuster stores the rate of change membership function, which also adopts a triangular membership function or a Gaussian membership function. Each rate of change fuzzy subset corresponds to a set of membership function parameters. Substituting the membership functions corresponding to each fuzzy subset of change rate, we can calculate... Membership value on each fuzzy subset of the rate of change.

[0077] The coefficient fuzzy adjuster internally stores fuzzy rule tables for proportional coefficient adjustment, integral coefficient adjustment, and derivative coefficient adjustment. Each fuzzy rule table consists of several "if-then" fuzzy rules, and the antecedent of each fuzzy rule is the temperature deviation signal. Fuzzy subset conditions and rate of change of temperature deviation The fuzzy subset conditions are combined using "AND" logic, with the consequent specifying the fuzzy subset of the corresponding coefficient adjustment amount. Within each coefficient adjustment cycle, the coefficient fuzzy adjuster adjusts according to... and The membership values ​​obtained after fuzzification are used to perform fuzzy inference on the three fuzzy rule tables. Fuzzy inference employs either the Mamdani inference method or the Larsen inference method to calculate the activation degree of each fuzzy rule. Taking the Mamdani inference method as an example, the activation degree of a fuzzy rule is the minimum of the two membership values ​​in the antecedent.

[0078] After completing the fuzzy inference, the coefficient fuzzy adjuster defuzzifies the three fuzzy rule tables using the centroid method. Taking the proportional coefficient adjustment fuzzy rule table as an example, the fuzzy subset of the proportional coefficient adjustment amount specified by the consequent of each fuzzy rule is truncated by activation degree. Then, all rule outputs are weighted and synthesized, and the proportional coefficient adjustment amount is calculated using the centroid method. The precise value. Integral coefficient adjustment. and differential coefficient adjustment amount The precise values ​​are extracted from the corresponding fuzzy rule table in the same way.

[0079] After obtaining the precise values ​​of the three adjustment amounts, the coefficient fuzzy adjuster corrects the proportional coefficient, integral coefficient, and derivative coefficient of the first adjustment sub-network. The expression for the correction process is:

[0080]

[0081] in, Indicates the first The first proportional coefficient after the coefficient update; Indicates the first The first proportional coefficient after the second coefficient update, initial value The value is obtained through offline tuning and ranges from 0.5 to 5.0. Indicates the first The proportional coefficient adjustment amount is obtained by defuzzification using fuzzy inference and the centroid method during the secondary coefficient update. The sign is determined by fuzzy rules based on the temperature deviation signal. and the rate of change of temperature deviation The state is determined in real time, and the corresponding absolute value upper limit is 20% of the initial value of the first proportional coefficient obtained by offline tuning, so as to ensure the stability of the coefficient update process; Represents the learning factor. The value is a constant ranging from 0.1 to 0.5, used to adjust the step size of each coefficient update. The value is determined by adjusting it during offline simulation and online debugging, so that the coefficient can quickly converge to a suitable value under varying operating conditions, without causing oscillations in the control system due to excessive adjustment in a single step.

[0082] The first integral coefficient and the first differential coefficient are dynamically updated using the same correction method, with the integral coefficient adjustment amount... The corresponding upper limit of absolute value is 20% of the initial value of the first integral coefficient obtained from offline tuning, and the adjustment amount of the differential coefficient is... The corresponding upper limit of absolute value is 20% of the initial value of the first differential coefficient obtained through offline tuning. The update results of the three coefficients are constrained by the limiter between their respective preset lower and upper limits. If the corrected coefficient value exceeds the upper limit, the upper limit is taken; if it is lower than the lower limit, the lower limit is taken, to ensure that the coefficients are always within the physically realizable and stable parameter range.

[0083] The second regulating sub-network also contains a coefficient fuzzy adjuster. This second sub-network shares the same fuzzy inference and defuzzification structure as the first sub-network, but its pre-stored proportional coefficient adjustment fuzzy rule table, integral coefficient adjustment fuzzy rule table, and differential coefficient adjustment fuzzy rule table are independently set. The content of the fuzzy rules used is determined separately based on the dynamic characteristics of the cooling water pump's flow regulation. The proportional coefficient correction expression of the second regulating sub-network has the same structure as that of the first regulating sub-network, except that each symbol in the expression is replaced with a second proportional coefficient. Initial value of the second proportional coefficient And the corresponding adjustment amount and learning factor, the learning factor also ranging from 0.1 to 0.5, and the second integral coefficient. Second differential coefficient Each also performs updates and limiting according to the same structure. Through the online calculation of the above-mentioned coefficient fuzzy adjuster, the proportional coefficient, integral coefficient, and derivative coefficient of the first adjustment sub-network and the proportional coefficient, integral coefficient, and derivative coefficient of the second adjustment sub-network are adaptively adjusted according to the real-time operating conditions and thermal state changes of the proton exchange membrane fuel cell stack.

[0084] See Figure 10 In the graph, the horizontal axis represents time in seconds, ranging from 0 to 20 seconds; the vertical axis represents the coefficient values, without specifying the unit. The graph shows the dynamic changes of three control coefficients in the first regulating sub-network, specifically the first proportional coefficient. (Blue solid line) First integral coefficient (Orange dashed line) and the first differential coefficient (Green dotted line).

[0085] Judging from the trend of the curve, the first proportional coefficient It exhibits obvious periodic fluctuations throughout the 20-second period, with the value range roughly between 2.0 and 3.0, a peak of about 3.0, and a trough of about 2.0, showing strong dynamic adjustment characteristics, which is consistent with the functional description of the coefficient fuzzy adjuster in Example 4 that automatically adjusts the proportional coefficient according to the temperature deviation signal.

[0086] First integral coefficient The numerical variation range is relatively small, with an overall fluctuation range of approximately 0.1 to 0.3. The trend of change is relatively gentle, accompanied by slight periodic fluctuations, indicating that the integral coefficient remains relatively stable during the control process but has a certain dynamic adaptive adjustment capability.

[0087] First differential coefficient The values ​​are the lowest, with a variation range of approximately 0.04 to 0.12. The curve shows small fluctuations and weak periodicity, which is consistent with the description in Example 4 of the differential coefficient being dynamically adjusted according to the rate of change of temperature deviation.

[0088] Example 5:

[0089] In specific implementation, please refer to Figure 6 The process of constructing the fuzzy neural network model was completed offline before the proton exchange membrane fuel cell thermal management method was actually put into online operation. Historical operating data of the proton exchange membrane fuel cell stack at multiple steady-state operating points were acquired. This historical operating data included historical output current values, historical cooling water inlet temperature values, historical cooling water outlet temperature values, and historical ambient temperature values. The steady-state operating points covered multiple output current steps of the stack from low load to high load. At each output current step, the cooling water inlet temperature and ambient temperature varied within a preset typical range. Data was recorded through sampling at equal time intervals, with each sampling synchronously recording the historical output current value, historical cooling water inlet temperature value, historical cooling water outlet temperature value, and historical ambient temperature value at the same moment. All collected data samples constituted a training sample set, which was divided into multiple training samples. Each training sample contained a set of historical output current values, historical cooling water inlet temperature values, and corresponding historical cooling water outlet temperature values. Historical ambient temperature values ​​were used as auxiliary information in the data preprocessing stage for filtering training samples and normalization benchmark correction.

[0090] An initial fuzzy neural network model is constructed, employing a four-layer feedforward structure. The four layers, arranged sequentially according to signal flow, consist of an input layer, a fuzzification layer, a fuzzy rule layer, and an output layer. The input layer contains a first input node and a second input node, corresponding to the input ports of historical output current values ​​and historical cooling water inlet temperature values, respectively. The fuzzification layer contains a first fuzzification node group and a second fuzzification node group. Each node in the first fuzzification node group corresponds to a current fuzzy subset and stores the center parameter and width parameter of the current membership function; each node in the second fuzzification node group corresponds to a temperature fuzzy subset and stores the center parameter and width parameter of the temperature membership function. Both the current and temperature membership functions are Gaussian membership functions. The expression for the current membership function is a Gaussian function, determined by the center parameter and width parameter of the current membership function; similarly, the temperature membership function is also a Gaussian function, determined by the center parameter and width parameter of the temperature membership function. In the initial fuzzy neural network model, the center parameter of the current membership function is initialized in an equal or non-equal interval manner within the measurement range of the stack output current. The width parameter of the current membership function is initialized to a value such that the membership values ​​of adjacent current membership function curves at the intersection are in the range of 0.3 to 0.7. The center parameter of the temperature membership function is initialized in an equal or non-equal interval manner within the measurement range of the cooling water inlet temperature. The width parameter of the temperature membership function is initialized to a value such that the membership values ​​of adjacent temperature membership function curves at the intersection are in the range of 0.3 to 0.7.

[0091] The fuzzy rule layer comprises a first fuzzy rule layer node and a second fuzzy rule layer node, which are either fully connected to the first fuzzy node group and the second fuzzy node group or connected via rule matching, respectively. Each fuzzy rule layer node stores current and temperature prerequisite parameters. The current prerequisite parameter is an index identifier for the current fuzzy subset, and the temperature prerequisite parameter is an index identifier for the temperature fuzzy subset. The fuzzy rule layer is then connected to a rule consequent parameter library. Each storage unit in the rule consequent parameter library corresponds to a combination consisting of the current fuzzy subset index and the temperature fuzzy subset index, and stores a fuzzy rule consequent parameter initially set to a random value or zero. The output layer comprises a single output layer node, which stores the output weights corresponding to each combination. The output weights are initialized to random decimal values ​​or empirically set initial values.

[0092] The backpropagation algorithm is used to iteratively train the initial fuzzy neural network model, with the training process aiming to minimize the error between the model output and the historical cooling water outlet temperature value. In each iteration, a training sample is randomly selected or selected sequentially from the training sample set. The historical output current value from the training sample is input to the first input node of the initial fuzzy neural network model, and the historical cooling water inlet temperature value from the training sample is input to the second input node of the initial fuzzy neural network model, driving the initial fuzzy neural network model to perform forward calculations to obtain the predicted cooling water outlet temperature value output by the model.

[0093] Calculate the mean squared error (MSE) between the predicted cooling water outlet temperature and the historical cooling water outlet temperature values ​​included in the training samples. The expression for calculating the MSE is as follows: ,in, This represents the mean square error value. This represents the predicted cooling water outlet temperature output by the model. This represents the historical cooling water outlet temperature values ​​in the training samples. (Coefficient) The coefficients of the squared terms are eliminated during the differentiation process, simplifying the subsequent gradient calculation.

[0094] The mean square error value was calculated. Then, the mean squared error value is passed layer by layer along the backpropagation path of the initial fuzzy neural network model, and the parameter gradient values ​​of each layer are calculated sequentially. Based on the parameter gradient values, the model parameters are updated synchronously using the gradient descent method. The backpropagation path starts at the output layer node and ends at the fuzzification layer node.

[0095] At the output layer nodes, calculate the gradient of the output weights with respect to the mean squared error. For parameters to be updated, The index representing the fuzzy subset of the current. The index representing the fuzzy subset of temperature. and The range of values ​​is determined based on the total number of fuzzy subsets. The calculation of the output weight gradient value is achieved by adjusting the mean square error value. Seeking information about The partial derivatives are used to derive the values. Based on the gradient values ​​of the output weights, the output weights are updated using gradient descent. The update rule is to subtract the learning rate multiplied by the gradient value of the output weights before the update. The learning rate is a constant value in the range of 0.001 to 0.1. The specific value of the learning rate is selected during the training and debugging phase based on the training convergence speed and convergence stability. For example, the learning rate is set to 0.01.

[0096] At the fuzzy rule layer node, gradient information transmitted from the output layer node is received, and the gradient value of the consequent parameters stored in the fuzzy rule layer node with respect to the mean square error value is calculated. This is done using the index corresponding to the current fuzzy subset. and temperature fuzzy subset index Fuzzy rule consequent parameters These are the parameters to be updated. The gradient values ​​of the fuzzy rule consequent parameters are calculated according to the chain rule, multiplying the gradients transmitted from the output layer nodes with the local gradients from the fuzzy rule consequent parameters to the output signal. Based on the gradient values ​​of the fuzzy rule consequent parameters, the gradient descent method is used to update the fuzzy rule consequent parameters. The learning rate can be the same as the output layer learning rate or set separately. If the learning rate is set separately, the range of the learning rate is also 0.001 to 0.1.

[0097] At the fuzzification layer node, gradient information transmitted from the fuzzy rule layer node is received, and the gradient values ​​of the membership function center parameter and membership function width parameter stored in the fuzzification layer node with respect to the mean square error value are calculated. The corresponding current fuzzy subset index in the first fuzzification node group is used as the starting point. The central parameter of the current membership function and the width parameter of the current membership function The parameter to be updated is the corresponding temperature fuzzy subset index in the second fuzzy node group. Temperature membership function central parameter and the width parameter of the temperature membership function These are the parameters to be updated. The gradient values ​​of the center parameter and width parameter are calculated according to the chain rule, multiplying the gradients passed from the nodes of the fuzzy rule layer with the local partial derivatives of each membership function with respect to its respective parameter. The local partial derivatives of the Gaussian membership function have analytical expressions. The center parameter is updated using gradient descent based on its gradient value. and The width parameter is updated using gradient descent based on its gradient value. and The learning rate ranges from 0.001 to 0.1, and can be consistent with the aforementioned learning rate or configured separately for the membership function parameter.

[0098] After updating the parameters of each layer in one iteration, the next training sample is selected from the training sample set, and the forward computation, error calculation, backpropagation gradient calculation, and parameter update steps are repeated. A training cycle is completed when all training samples in the training sample set have been traversed. After each training cycle, the mean square error (MSE) of all training samples in the training sample set is calculated. If the MSE is less than a preset accuracy threshold, iterative training stops; if the MSE does not meet the preset accuracy requirement, the next training cycle continues iterating. The preset accuracy threshold is set based on the actual requirements of the proton exchange membrane fuel cell thermal management for predicting the target cooling water outlet temperature, for example, on the order of 0.01℃². After iterative training, the trained current membership function center parameters, current membership function width parameters, temperature membership function center parameters, temperature membership function width parameters, fuzzy rule consequent parameters, and output weights are solidified to form a pre-built fuzzy neural network model for online forward computation.

[0099] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks, characterized in that, The method includes: Obtain the output current value and cooling water inlet temperature value of the proton exchange membrane fuel cell stack at the current moment; The output current value and the cooling water inlet temperature value are input into a pre-built fuzzy neural network model, and the fuzzy neural network model is driven to perform forward calculation to output the target cooling water outlet temperature value corresponding to the current moment. The target temperature value of the cooling water outlet is compared with the actual temperature value of the cooling water outlet collected in real time to generate a temperature deviation signal; The temperature deviation signal is input to a preset multi-layer feedback control network, which generates a fan speed control command and a water pump flow control command for the current moment based on the temperature deviation signal. The fan speed is adjusted according to the fan speed control command, and the flow rate of the cooling water pump is adjusted according to the water pump flow control command, so as to change the heat dissipation of the fuel cell stack.

2. The thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks according to claim 1, characterized in that, The acquisition of the output current value and cooling water inlet temperature value of the proton exchange membrane fuel cell stack at the current moment includes: The main circuit current signal at the output terminal of the fuel cell stack is acquired by a current sensor, and the main circuit current signal is converted from analog to digital to obtain the output current value. The inlet water temperature is collected by a thermocouple installed at the cooling water inlet pipe of the fuel cell stack. The inlet water temperature is then converted from analog to digital and filtered to obtain the cooling water inlet temperature value.

3. The thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks according to claim 2, characterized in that, The analog-to-digital conversion uses a sampling frequency of 100Hz, and the filtering process employs a moving average filtering algorithm to smooth multiple continuously acquired inlet water temperature analog signals.

4. The thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks according to claim 1, characterized in that, The output current value and the cooling water inlet temperature value are input into a pre-built fuzzy neural network model, and the fuzzy neural network model is driven to perform forward calculations to output the target cooling water outlet temperature value corresponding to the current moment, including: The output current value and the cooling water inlet temperature value are used as the input values ​​of the first input node and the second input node of the fuzzy neural network model, respectively. Fuzzification processing is performed on the input values ​​of the first input node and the second input node to obtain the current membership vector and the temperature membership vector. The current membership vector and the temperature membership vector are matched with the first fuzzy rule layer node and the second fuzzy rule layer node in the fuzzy neural network model, respectively, to activate the corresponding fuzzy rule consequent parameters. The activated fuzzy rule consequent parameters are input to the output layer node of the fuzzy neural network model. The output layer node performs a weighted summation calculation on the fuzzy rule consequent parameters and outputs the target temperature value of the cooling water outlet.

5. The thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks according to claim 4, characterized in that, The input values ​​of the first input node and the second input node are fuzzified to obtain the current membership vector and the temperature membership vector, including: For the first input node, a preset current membership function is used to calculate the current membership value of the output current value on multiple current fuzzy subsets, and all the current membership values ​​are combined into the current membership vector. The number of the current fuzzy subsets is at least three. For the second input node, a preset temperature membership function is used to calculate the temperature membership value of the cooling water inlet temperature value on multiple temperature fuzzy subsets, and all the temperature membership values ​​are combined into the temperature membership vector. The number of temperature fuzzy subsets is at least three.

6. The thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks according to claim 5, characterized in that, Both the current membership function and the temperature membership function are Gaussian membership functions.

7. The thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks according to claim 4, characterized in that, The current membership vector and the temperature membership vector are matched with the first fuzzy rule layer nodes and the second fuzzy rule layer nodes in the fuzzy neural network model, respectively, to activate the corresponding fuzzy rule consequent parameters, including: Each current membership value in the current membership vector is compared with the current prerequisite parameters stored in the first fuzzy rule layer node, and the current fuzzy subset corresponding to the current membership value with the highest matching degree is selected as the active current subset. Each temperature membership value in the temperature membership vector is compared with the temperature prerequisite parameters stored in the second fuzzy rule layer node, and the temperature fuzzy subset corresponding to the temperature membership value with the highest matching degree is selected as the active temperature subset. Based on the combined index of the activation current subset and the activation temperature subset, the fuzzy rule consequent parameter corresponding to the combined index is called from the rule consequent parameter library of the fuzzy neural network model, and the called fuzzy rule consequent parameter is used as the activated fuzzy rule consequent parameter.

8. The thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks according to claim 1, characterized in that, The temperature deviation signal is input to a preset multi-layer feedback control network. The multi-layer feedback control network generates a fan speed control command and a water pump flow control command for the current moment based on the temperature deviation signal, including: The temperature deviation signal is input to the first regulation subnetwork and the second regulation subnetwork in the multilayer feedback regulation network, respectively; The first regulation sub-network performs proportional-integral-derivative calculation on the temperature deviation signal to obtain the basic control quantity of the fan speed, and outputs the basic control quantity of the fan speed as the fan speed control command after amplitude limiting. The second regulating sub-network performs proportional-integral-derivative calculations on the temperature deviation signal to obtain the basic control quantity of the water pump flow rate, and outputs the basic control quantity of the water pump flow rate as the water pump flow rate control command after amplitude limiting.

9. The thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks according to claim 8, characterized in that, The proportional coefficients, integral coefficients, and derivative coefficients in the first and second regulation subnetworks are all dynamically updated using a fuzzy rule-based online self-adjustment method.

10. The thermal management method for proton exchange membrane fuel cells based on fuzzy neural networks according to claim 8, characterized in that, The first regulating subnetwork performs proportional-integral-derivative (PID) calculations on the temperature deviation signal to obtain a basic fan speed control value, and outputs the basic fan speed control value as the fan speed control command after amplitude limiting, including: The temperature deviation signal is simultaneously input to the proportional calculation path, integral calculation path and derivative calculation path of the first adjustment sub-network. The proportional calculation path multiplies the temperature deviation signal by a preset first proportional coefficient to obtain the proportional component. The integral calculation path performs time-cumulative calculation on the temperature deviation signal and then multiplies it by a preset first integral coefficient to obtain the integral component. The differential calculation path performs time derivative calculation on the temperature deviation signal and then multiplies it by a preset first differential coefficient to obtain the differential component; The proportional component, the integral component, and the derivative component are added together to obtain the basic fan speed control value. The basic fan speed control value is then compared with preset upper and lower limits for fan speed. If the basic fan speed control value is greater than the upper limit for fan speed, the upper limit for fan speed is output as the fan speed control command. If the basic fan speed control value is less than the lower limit for fan speed, the lower limit for fan speed is output as the fan speed control command. Otherwise, the basic fan speed control value is directly output as the fan speed control command.

Images