Active intelligent thermal management system, method and equipment for proton exchange membrane fuel cell and medium

Abstract

The invention discloses an active intelligent thermal management system, method, equipment and medium for a proton exchange membrane fuel cell, the system comprises a hardware layer and a control layer, the hardware layer is composed of an electric pile module, a cooling actuator and a sensor network; a bipolar plate is arranged on the electric pile module and is used for optimizing a heat conduction path; the cooling actuator comprises a frequency conversion fan, a liquid cooling pump, a three-way valve and a phase change heat storage header. The sensor network comprises a distributed temperature sensor, a humidity sensor and a current sensor; the control layer is composed of an edge controller and a cloud server. By constructing the complete AI-driven intelligent thermal management closed-loop control method, the problem that a traditional thermal management system lacks intelligent decision-making ability is fundamentally solved.

Description

Technical Field

[0001] This invention relates to the field of thermal management technology for proton exchange membrane fuel cells, and in particular to an active intelligent thermal management system, method, device and medium for proton exchange membrane fuel cells. Background Technology

[0002] During PEMFC operation, approximately 50% of the energy is dissipated as heat, necessitating strict control of the operating temperature (50-80℃) to prevent membrane dehydration and catalyst deactivation. Traditional passive thermal management (e.g., heat pipes, phase change materials) relies on fixed heat dissipation structures, making it difficult to dynamically adapt to load changes, and its waste heat recovery efficiency is low. Existing active thermal management (e.g., liquid cooling) can regulate temperature, but it relies on preset threshold control and lacks intelligent decision-making capabilities, resulting in high energy consumption and sluggish response. PEMFC thermal management presents significant challenges.

[0003] Meanwhile, existing active thermal management controls have technical shortcomings. Passive systems have a fixed heat dissipation capacity and cannot cope with sudden heat loads or environmental changes. Traditional active systems' rule-based PID control struggles to optimize complex nonlinear thermal processes and fails to fully utilize waste heat recovery potential. Currently, AI is rarely used in PEMFC thermal management, lacking intelligent control schemes that combine real-time data with predictive models, indicating a deficiency in AI applications.

[0004] The performance of PEMFCs is highly dependent on operating temperature, and traditional passive thermal management systems suffer from slow response speed and low temperature control accuracy. While active thermal management systems can improve efficiency, existing control strategies (such as PID control) are difficult to adapt to dynamic heat load changes under complex operating conditions. The introduction of artificial intelligence technologies (such as adaptive PID, fuzzy logic, and neural networks) can significantly improve the robustness and intelligence of the system. Summary of the Invention

[0005] To overcome the technical defects of active and passive thermal management control, this invention proposes an active intelligent thermal management system, method, equipment and medium for proton exchange membrane fuel cells. Through multi-sensor data fusion, dynamic thermal model prediction and intelligent algorithm optimization, precise control of PEMFC stack temperature is achieved.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an active intelligent thermal management system for a proton exchange membrane fuel cell, comprising a hardware layer and a control layer. The hardware layer consists of a fuel cell stack module, a cooling actuator, and a sensor network. The fuel cell stack module is equipped with bipolar plates for optimizing the heat conduction path. The cooling actuator includes a variable frequency fan, a liquid-cooled pump, a three-way valve, and a phase change thermal storage manifold. The sensor network includes distributed temperature sensors, humidity sensors, and current sensors. The control layer consists of an edge controller and a cloud server.

[0007] As a preferred embodiment of the active intelligent thermal management system for proton exchange membrane fuel cells described in this invention, the edge controller deploys a lightweight AI model for processing data and generating control commands; the cloud server stores historical data, updates the lightweight AI model, and provides remote monitoring.

[0008] Through the coordinated operation of the hardware and control layers, intelligent thermal management of proton exchange membrane fuel cells (PEMFCs) is achieved. The hardware layer integrates optimized stack modules, multimodal cooling actuators, and a distributed sensor network, enabling comprehensive monitoring and precise control of the stack temperature. The control layer employs an edge-cloud collaborative architecture. Edge controllers deploy lightweight AI models for millisecond-level real-time response, while cloud servers handle model optimization and remote monitoring, ensuring continuous system learning and performance improvement. This system addresses the issues of lag and low control precision in traditional thermal management systems, thereby improving the operating efficiency and lifespan of fuel cells.

[0009] Secondly, this invention provides an active intelligent thermal management method for proton exchange membrane fuel cells, comprising the following steps: acquiring the actual stack temperature through a sensor network. Current and ambient temperature The average temperature of the fuel cell stack was obtained by fusing data from distributed temperature sensors using a weighted average method. The average temperature Tavg and current of the fuel cell stack are controlled by an edge controller. Ambient temperature As input sequence, the LSTM model is used to predict the stack temperature at future times based on historical data. ; the future stack temperature With target temperature The system compares and calculates the temperature deviation. Based on the current state and the reward function, it generates preliminary control actions using a reinforcement learning algorithm. The preliminary control action Using these as initial values, the cooling parameters are optimized using model predictive control to obtain the optimal control parameters. To minimize the temperature deviation in the prediction time domain; through the optimal control parameters The variable frequency fan and liquid cooling pump drive the cooling actuator, while simultaneously controlling the phase change thermal storage manifold to store or release heat; the actual stack temperature after execution is... With future time stack temperature The prediction error is calculated and fed back to the edge controller to update the parameters of the lightweight AI model.

[0010] As a preferred embodiment of the active intelligent thermal management system for proton exchange membrane fuel cells described in this invention, the LSTM model adopts a 3-layer LSTM structure, with each layer containing 128 neurons and a fully connected layer; the input features include current. Ambient temperature The output includes the inlet humidity (φ) and coolant flow rate; the output is the fuel cell stack temperature at a future time N seconds in advance. The loss function uses mean squared error (MSE), and the calculation steps are as follows: obtain the actual stack temperature. LSTM model predicts future stack temperature Calculate the prediction error at each time point i, i.e., the actual stack temperature. Subtract the predicted temperature Square the prediction error at each time point; sum the squared errors of all n time points; divide the sum by the number of samples n to obtain the mean squared error (MSE); train the LSTM model by minimizing the MSE.

[0011] As a preferred embodiment of the active intelligent thermal management system for proton exchange membrane fuel cells described in this invention, wherein: when fusing distributed temperature sensor data using a weighted average method, the weighting coefficients are... The relationship between sensor location and fuel cell heat flux density is determined by the following formula: ; In the formula, β is the distance sensitivity coefficient; Let n be the distance between sensor i and the hotspot, n be the number of temperature measurement points, and j be the summation index variable.

[0012] The beneficial effects of this preferred technical solution are as follows: by using a distance-sensitive weighted fusion algorithm, the temperature sensor closer to the hot spot is given a higher weight, which accurately reflects the true thermal distribution of the fuel cell stack, avoids the temperature estimation deviation caused by simple averaging, and improves the accuracy of temperature monitoring.

[0013] As a preferred embodiment of the active intelligent thermal management system for proton exchange membrane fuel cells described in this invention, the initial control action is generated through a reinforcement learning algorithm based on the current state and the reward function. During the process, the current state includes the actual stack temperature at the current time t. Coolant flow rate Phase change heat storage The preliminary control action Including fan speed Liquid cooling pump flow rate regulation, valve opening degree v; the reward function The calculation method is as follows: for the maximum temperature difference inside the fuel cell stack Average temperature deviation of fuel cell stack Cooling system energy consumption After weighted summation and taking the negative value, the calculation formula is as follows: ; In the formula, Let be the reward value at time t; Energy consumption of the cooling system; The maximum temperature difference weighting coefficient; The average temperature difference weighting coefficient; t represents the energy consumption weighting coefficient; t represents the time step index.

[0014] The beneficial effects of this preferred technical solution are as follows: by designing a multi-objective weighted reward function, it can reduce temperature fluctuations while taking into account energy consumption optimization, avoiding energy waste caused by single-objective control, and enabling the system to achieve the optimal balance between temperature control performance and economy.

[0015] As a preferred embodiment of the active intelligent thermal management system for proton exchange membrane fuel cells described in this invention, in the process of optimizing cooling parameters using model predictive control, the optimal control parameter at is obtained by constructing and solving an optimization problem. The optimization problem aims to minimize the temperature deviation and control input variation within the future prediction time domain, and the calculation formula is as follows: ; And satisfy the following constraints: Predicted stack temperature at the k-th step Must be within the allowable temperature range to between; Fan speed Must be within the rated range to between; Phase change heat storage Within the capacity range to between; In the formula: J is the objective function value; α is the control smoothing coefficient; Δa(k) is the change in control input at step k; To predict the number of time-domain steps; To control the number of time-domain steps; k is the prediction step index; and These are the minimum and maximum allowable temperatures for the fuel cell stack, respectively. and These are the lower and upper limits of the fan speed, respectively, in r / min; and These are the lower and upper limits for phase change heat storage, respectively.

[0016] The beneficial effects of this preferred technical solution are: constructing a predictive time-domain optimization problem and setting multiple constraints ensures that the control parameters minimize temperature deviation within a safe range, avoids overshoot and oscillation, and improves the robustness and stability of the system.

[0017] As a preferred embodiment of the active intelligent thermal management system for proton exchange membrane fuel cells described in this invention, in the process of controlling the storage or release of phase change heat storage in the phase change heat storage manifold, the phase change heat storage is composed of both the sensible heat and latent heat of the phase change material, and the calculation steps are as follows: determining the mass of the phase change material. ; Calculate the sensible heat of the phase change material in the solid state, using the solid-state specific heat capacity. With solid-state temperature change The product is used to obtain the sensible heat of the phase change material in the liquid state; this is calculated using the specific heat capacity of the liquid. With the change in liquid temperature Δ The product is obtained; the latent heat of the phase change process is calculated through the mass of the phase change material. The total heat storage capacity Qp is obtained by multiplying the sensible heat of the solid state, the sensible heat of the liquid state, and the latent heat of phase change; the calculation formula is as follows: .

[0018] Thirdly, the present invention provides an electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the active intelligent thermal management system for the proton exchange membrane fuel cell.

[0019] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the active intelligent thermal management system for the proton exchange membrane fuel cell.

[0020] Compared with existing technologies, the beneficial effects of this invention are as follows: By constructing a complete AI-driven intelligent thermal management closed-loop control method, the problem of the lack of intelligent decision-making capabilities in traditional thermal management systems is fundamentally solved. LSTM neural networks are used to predict future temperature change trends, combined with reinforcement learning algorithms to generate control strategies that balance temperature control accuracy and energy consumption optimization. Then, model predictive control dynamically compensates for temperature deviations, achieving precise regulation of complex nonlinear thermal processes. Multi-sensor weighted fusion technology improves the accuracy of temperature monitoring, and the real-time feedback mechanism enables the AI ​​model to continuously learn adaptively, enhancing the system's response to sudden heat loads and environmental changes, overcoming the limitations of traditional PID control. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0022] Figure 1 This is a structural diagram of an active intelligent thermal management system for a proton exchange membrane fuel cell according to an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the overall process of the active intelligent thermal management method for proton exchange membrane fuel cells according to an embodiment of the present invention.

[0024] In the diagram, 1 is the hardware layer; 2 is the control layer; 1-1 is the fuel cell stack module; 1-2 is the cooling actuator; 1-3 is the sensor network; 2-1 is the edge controller; and 2-2 is the cloud server. Detailed Implementation

[0025] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0026] Example 1, referring to Figure 1 According to one embodiment of the present invention, an active intelligent thermal management system for proton exchange membrane fuel cells is provided, comprising a hardware layer 1 and a control layer 2. The hardware layer consists of a fuel cell stack module 1-1, a cooling actuator 1-2, and a sensor network 1-3. The fuel cell stack module 1-1 is equipped with bipolar plates for optimizing the heat conduction path. The cooling actuator 1-2 includes a variable frequency fan, a liquid-cooled pump, a three-way valve, and a phase change thermal storage manifold. The sensor network 1-3 includes distributed temperature sensors, humidity sensors, and current sensors. The control layer 2 consists of an edge controller 2-1 and a cloud server 2-2.

[0027] Furthermore, the edge controller 2-1 deploys a lightweight AI model for processing data and generating control commands; the cloud server 2-2 stores historical data, updates the lightweight AI model, and provides remote monitoring.

[0028] In summary, intelligent thermal management of proton exchange membrane fuel cells (PEMFCs) is achieved through the coordinated operation of the hardware and control layers. The hardware layer integrates optimized stack modules, multimodal cooling actuators, and a distributed sensor network, enabling comprehensive monitoring and precise control of the stack temperature. The control layer employs an edge-cloud collaborative architecture, with edge controllers deploying lightweight AI models to achieve millisecond-level real-time responses, while cloud servers handle model optimization and remote monitoring, ensuring continuous system learning and performance improvement. This system addresses the issues of lag and low control precision inherent in traditional thermal management systems, thereby improving the operating efficiency and lifespan of fuel cells.

[0029] Example 2, refer to Figure 2 As an embodiment of the present invention, based on the above embodiment, an active intelligent thermal management method for proton exchange membrane fuel cells is provided, comprising the following steps S1 to S6: S1. Obtain the actual stack temperature through sensor network 1-3. Current and ambient temperature The average temperature of the fuel cell stack was obtained by fusing data from distributed temperature sensors using a weighted average method. .

[0030] When fusing distributed temperature sensor data using the weighted average method, the weighting coefficients... The relationship between sensor location and fuel cell heat flux density is determined by the following formula: ; In the formula, β is the distance sensitivity coefficient; Let n be the distance between sensor i and the hotspot, n be the number of temperature measurement points, and j be the summation index variable.

[0031] In this embodiment, sensor network 1-3 uniformly distributes 10 K-type thermocouples (accuracy ±0.5℃) within fuel cell module 1-1, with a measurement point spacing ≤10mm. The location of the fuel cell hotspot is determined by calibration using an infrared thermal imager, and the distance between each sensor and the hotspot is measured. In this embodiment, the distance sensitivity coefficient β is set to 2, and the weight coefficient of each sensor is determined through heat flux density simulation. The specific calculation process is as follows: assuming sensor 1 is 0.05m from the hotspot, sensor 2 is 0.08m, sensor 3 is 0.12m, and so on. The weight of sensor 1 is calculated according to the formula: After normalization, it is approximately 0.15; the weight of sensor 2: After normalization, it is approximately 0.14; The remaining sensors calculate in sequence to ensure During real-time data acquisition, sensor networks 1-3 synchronously acquire the temperature at each measuring point at a frequency of 1Hz. (i=1,2,...,10), total stack current Is (range 0-500A, accuracy ±0.5%), and ambient temperature. (Accuracy ±0.5℃). The average temperature of the fuel cell stack is calculated using edge controller 2-1: We obtain the weighted average temperature that reflects the true thermal state of the fuel cell stack.

[0032] In an optional implementation, when fusing distributed temperature sensor data using the weighted average method in step S1, an adaptive weight adjustment strategy based on heat flux density gradient can also be adopted. Specifically, based on the real-time monitored current density distribution Is(x,y), the instantaneous heat flux density distribution q(x,y) inside the fuel cell stack is calculated using finite element simulation, and the distance sensitivity coefficient β is dynamically updated. When a sudden increase in local heat flux density is detected (such as a power surge), the weight coefficient αi of sensors in hotspot areas is automatically increased, while the weight of sensors far from hotspot areas is decreased. This allows the weighted average temperature Tavg to more sensitively reflect changes in the temperature of dangerous hotspots, making it suitable for applications with frequent dynamic load changes.

[0033] In another optional implementation, when obtaining the average temperature of the fuel cell stack in step S1, a multi-sensor data fusion method based on machine learning can also be used. Specifically, historical operating data (including 10 temperature measurement points, current, ambient temperature, coolant flow rate, humidity, and other multi-dimensional features) and the corresponding real core temperature of the fuel cell stack (measured by an embedded high-precision sensor) are collected, and an end-to-end temperature fusion model is trained using random forest or gradient boosting tree algorithms.

[0034] S2, The average temperature Tavg of the fuel cell stack and the current are controlled by the edge controller 2-1. Ambient temperature As input sequence, the LSTM model is used to predict the stack temperature at future times based on historical data. .

[0035] The LSTM model employs a 3-layer LSTM structure, with each layer containing 128 neurons and a fully connected layer; input features include electrical current. Ambient temperature Inlet humidity φ, coolant flow rate; The output is the stack temperature at a future time N seconds in advance. The loss function uses the mean squared error (MSE), and the calculation steps are S2.1~S2.6: S2.1 Obtain the actual stack temperature LSTM model predicts future stack temperature .

[0036] The parameters are shown in Table 1; Table 1: Actual stack temperature vs. LSTM predicted temperature model for future times.

[0037] S2.2 Calculate the prediction error at each time point i, i.e., the actual stack temperature. Subtract the predicted temperature .

[0038] Calculate the prediction error: ; ; ; ; .

[0039] S2.3. Squaring the prediction error at each time point.

[0040] ; ; ; ; .

[0041] S2.4 Sum the squared errors of all n time points.

[0042] .

[0043] S2.5. Divide the sum by the number of samples n to obtain the mean squared error (MSE).

[0044] .

[0045] S2.6. Train the LSTM model by minimizing the mean square error (MSE).

[0046] During training, the optimizer (such as Adam) continuously adjusts the weight parameters of the LSTM, causing the MSE to change from its initial value (e.g., 2.5℃). 2 Gradually reduce to the target value (<0.8℃) 2 It eventually converged to 0.72℃. 2 .

[0047] In an optional implementation, when using the LSTM model to predict the future stack temperature in step S2, an attention-enhanced LSTM architecture can be employed to improve prediction accuracy. Specifically, a multi-head attention layer is added after the three LSTM layers to automatically learn the contribution weights of different time steps in the input sequence to the future temperature prediction. For example, when the current undergoes a step change at time t-5, the attention mechanism automatically assigns a higher weight (weight value up to 0.35) to that time, while reducing the weight for stationary times (weight value approximately 0.05), enabling the model to more accurately capture the temperature response characteristics under abrupt changes.

[0048] In another optional implementation, when predicting the future stack temperature in step S2, a multi-model fusion strategy using ensemble learning can also be employed. Specifically, three heterogeneous prediction models are deployed in parallel: an LSTM model (excels at capturing long-term temporal dependencies), a GRU model (Gated Recurrent Unit, computationally efficient), and a Transformer model (excels at handling long sequences). The three models infer independently in different threads of edge controller 2-1, outputting their respective temperature predictions. A weighted average fusion strategy is used. This method reduces prediction uncertainty through model complementarity, achieving continuous performance improvement.

[0049] S3, the future stack temperature With target temperature The system compares and calculates the temperature deviation. Based on the current state and the reward function, it generates preliminary control actions using a reinforcement learning algorithm. .

[0050] It should be noted that the current state includes the actual stack temperature at the current time t. Coolant flow rate Phase change heat storage ; The initial control action Including fan speed Liquid cooling pump flow rate regulation and valve opening degree v; The reward function The calculation method is as follows: for the maximum temperature difference inside the fuel cell stack Average temperature deviation of fuel cell stack Cooling system energy consumption After weighted summation and taking the negative value, the calculation formula is as follows: ; In the formula, Let be the reward value at time t; Energy consumption of the cooling system; The maximum temperature difference weighting coefficient; The average temperature difference weighting coefficient; t represents the energy consumption weighting coefficient; t represents the time step index.

[0051] In this embodiment, the process of generating preliminary control actions in step S3 using a reinforcement learning algorithm is as follows: First, the state information at the current time t is obtained from the edge controller 2-1, including the actual stack temperature. Coolant flow rate and phase change heat storage These state parameters constitute the current state vector. , as input for reinforcement learning decision-making.

[0052] Next, the future stack temperature predicted by the LSTM model in step S2 will be... With preset target temperature By comparison, the temperature deviation is calculated as follows: This positive deviation indicates that the stack temperature will fall below the target value; therefore, the control strategy should reduce cooling intensity or release phase change heat storage to raise the temperature. Simultaneously, the system monitors the temperature distribution inside the stack, measuring the hottest point temperature at 72.1℃ and the coldest point temperature at 66.9℃, and calculates the maximum temperature difference. And the deviation between the average temperature of the fuel cell stack and the target temperature. In addition, the current energy consumption of the cooling system is... This value is calculated from the cumulative energy consumption of the variable frequency fan (power 35W) and liquid cooling pump (power 15W) running for 1 hour.

[0053] The Deep Q-Network (DQN) reinforcement learning module deployed on cloud server 2-2 calculates the reward function value based on the aforementioned state information and monitoring data. In this embodiment, the weight coefficients are set to... (Maximum temperature difference weighting coefficient) (Weighting coefficient for average temperature deviation) (Energy consumption weighting coefficient). Based on the reward function formula. The calculated reward value for the current moment is -3.71. A negative reward indicates that the current state is unsatisfactory, with significant temperature non-uniformity and deviation, requiring system optimization of the control strategy to improve thermal management performance.

[0054] The DQN network employs a three-layer fully connected architecture. The input layer receives a 3D state vector, which is processed through two hidden layers (containing 128 and 64 neurons respectively, with ReLU activation function). The output layer generates 27 Q-values, corresponding to 27 discretized control action combinations. The discretization scheme for the action space is as follows: fan speed... The liquid cooling pump flow rate is adjustable in three settings (1000, 2000, 3000 RPM). The valve speed is set to three levels (1, 3, 5 L / min), and the valve opening (v) is set to three levels (0.3, 0.6, 0.9), resulting in a total of 3 × 3 × 3 = 27 possible action combinations. The decision is made using... Greedy strategy, current exploration rate (This value is obtained through exponential decay during the training process), that is, with an 85% probability, the action with the largest Q value is selected (using learned knowledge), and with a 15% probability, the action is randomly selected (exploring unknown strategies).

[0055] In this decision-making process, the Q-value output by the DQN network represents the action. The corresponding Q value is the largest, which is This value represents the expected cumulative reward over the next 10 steps after executing the action. Therefore, the system generates the following initial control actions: reduce the fan speed to 2000 RPM to reduce heat dissipation intensity, maintain the liquid cooling pump flow rate at a moderate level of 3.0 L / min, and set the valve opening to 60% to moderately introduce the heat released by the phase change thermal storage module. This initial control action is then passed to the Model Predictive Control (MPC) module in step S4 for further fine-tuning and optimization.

[0056] To continuously optimize the decision-making capabilities of the DQN model, the system will incorporate the experience gained from this decision-making process. The system stores 100,000 experience replays in a pool. After every 100 steps, it randomly samples 32 experiences from the pool to form a batch and updates the parameters of the DQN main network using batch gradient descent. Simultaneously, the target network synchronizes its parameters with the main network every 100 steps. This dual-network structure effectively reduces numerical oscillations during training, ensuring the stability and convergence of the learning process.

[0057] In an optional implementation, when generating the initial control action using the reinforcement learning algorithm in step S3, a priority experience replay mechanism can also be employed to improve learning efficiency. The core idea of ​​this method is to assign different priority weights to each experience in the experience replay pool, making the training process focus more on samples with larger model prediction errors. Specifically, for each experience in the experience replay pool... The system calculates its timing difference error. The empirical priority weights are defined as follows: This prevents certain experiences from having zero weights and never being sampled. During training, probability sampling is performed based on priority weights. Experiences with large TD errors (i.e., samples that the model is currently predicting inaccurately) are sampled more frequently, thus enabling the network to learn complex and difficult-to-predict conditions more quickly.

[0058] S4. Perform the preliminary control action. Using these as initial values, the cooling parameters are optimized using model predictive control to obtain the optimal control parameters. To minimize temperature deviation in the prediction time domain.

[0059] The optimal control parameter at is obtained by constructing and solving an optimization problem. The optimization problem aims to minimize the temperature deviation and control input variation within the future prediction time domain. The calculation formula is as follows: ; And satisfy the following constraints: Predicted stack temperature at the k-th step Must be within the allowable temperature range to between; Fan speed Must be within the rated range to between; Phase change heat storage Within the capacity range to between; In the formula: J is the objective function value; α is the control smoothing coefficient; Δa(k) is the change in control input at step k; To predict the number of time-domain steps; To control the number of time-domain steps; k is the prediction step index; and These are the minimum and maximum allowable temperatures for the fuel cell stack, respectively. and These are the lower and upper limits of the fan speed, respectively, in r / min; and These are the lower and upper limits for phase change heat storage, respectively.

[0060] In this embodiment, step S4 optimizes the initial control action generated in step S3 using model predictive control. Edge controller 2-1 uses the initial control action output by DQN in step S3 as the initial value for MPC optimization, and constructs a finite-time optimization problem to obtain the optimal control parameters.

[0061] The objective function of the optimization problem is designed to minimize the weighted sum of the temperature tracking deviation and the change in control input over the future prediction time domain. The first term represents the future... Predicting the stack temperature within a step With target temperature The cumulative sum of squared errors between ℃ is used to ensure temperature tracking accuracy; the second term represents the control time domain. Internal control input variation The sum of squares is used to ensure the smoothness of control actions and avoid frequent actuator movements that lead to increased mechanical wear and energy consumption. The control smoothness coefficient is used to balance the weight relationship between the two objectives.

[0062] In this embodiment, the prediction time domain is set to Step (corresponding to the next 10 seconds), control time domain set to Step 2. Edge controller 2-1 calls the LSTM model to perform rolling prediction, based on the state at the current time t. Starting with the assumed control sequence Iteratively predict the stack temperature sequence for the next 10 steps. Assuming the initial control sequence uses the initial action a_t given by DQN, the predicted temperature sequence is [68.8, 69.2, 69.5, 69.7, 69.8, 69.9, 70.0, 70.0, 69.9, 69.8]℃. It can be seen that the target temperature of 70℃ can only be reached in step 7, and there is a slight overshoot.

[0063] Edge controller 2-1 employs a sequential quadratic programming algorithm to solve the aforementioned constrained nonlinear optimization problem. Starting from initial values, the algorithm iteratively calculates the gradient of the objective function and approximates the Hessian matrix, gradually adjusting the control sequence until the Karush-Kuhn-Tucker (KKT) optimality condition is met or the maximum number of iterations (set to 50) is reached. After 18 iterations, the algorithm converges to the optimal solution.

[0064] After optimization, edge controller 2-1 executes only the first action of the control sequence and sends the control command to cooling actuator 1-2. In the next time step, the stack state is remeasured, and the above prediction-optimization process is repeated, forming a rolling optimization closed-loop control. This model predictive control strategy fully utilizes the accurate predictive capabilities of LSTM and achieves optimal decision-making through multi-step prediction via optimization algorithms, significantly improving the speed, accuracy, and smoothness of temperature control.

[0065] S5. Using the optimal control parameters The variable frequency fan and liquid cooling pump drive the cooling actuators 1-2, while controlling the phase change thermal storage manifold to store or release heat.

[0066] During the process of controlling the storage or release of phase change heat in the phase change heat storage manifold, the phase change heat storage consists of both the sensible heat and latent heat of the phase change material. The calculation steps are S5.1~S5.5: S5.1 Determine the mass of the phase change material .

[0067] S5.2 Calculate the sensible heat of the phase change material in the solid state, using the solid-state specific heat capacity. With solid-state temperature change The product is obtained.

[0068] S5.3 Calculate the sensible heat of the phase change material in the liquid state, using the specific heat capacity of the liquid. With the change in liquid temperature Δ The product is obtained.

[0069] S5.4 Calculate the latent heat of the phase change process, using the mass of the phase change material. It is obtained by multiplying the latent heat of phase transition λ.

[0070] S5.5. Add the sensible heat of the solid state, the sensible heat of the liquid state, and the latent heat of phase change to obtain the total heat storage Qp. The calculation formula is as follows: .

[0071] It should be noted that this formula applies to materials undergoing a complete phase transition from solid to liquid state upon heating. In the case of a partial phase transition in this embodiment, a segmented accumulation method is used for calculation: the total heat released by the header in the current cycle is... This heat is transferred to the fuel cell stack via coolant circulation, causing the average stack temperature to rise from 67.5°C to 68.3°C within 10 seconds, at a rate of approximately 0.08°C / s. Simultaneously, the remaining heat storage capacity within the header is monitored. kJ, ensuring that the heat storage capacity is always within the allowable range. kJ within.

[0072] The coordinated action of cooling actuators 1-2 achieves precise temperature regulation. Reducing fan speed decreases convective heat dissipation, decreasing liquid cooling pump flow reduces heat carried away by forced convection, and increasing valve opening increases heat replenishment from phase change thermal storage. The combined effect of these three factors ensures the fuel cell stack temperature rises to the target value according to the expected trajectory. Throughout the process, edge controller 2-1 continuously monitors the actual response status of the actuators. If a fan or liquid cooling pump malfunction is detected (such as loss of speed feedback signal or abnormal flow sensor), the system immediately triggers fault protection mode, switching to a backup scheme of pure air cooling or pure liquid cooling to ensure the fuel cell stack maintains a safe operating temperature under any circumstances.

[0073] S6. The actual stack temperature after execution. With future time stack temperature The prediction error is calculated and fed back to the edge controller 2-1 to update the parameters of the lightweight AI model.

[0074] In summary, by constructing a complete AI-driven intelligent thermal management closed-loop control method, the problem of the lack of intelligent decision-making capabilities in traditional thermal management systems is fundamentally solved. LSTM neural networks are used to predict future temperature change trends, combined with reinforcement learning algorithms to generate control strategies that balance temperature control accuracy and energy consumption optimization. Then, model predictive control dynamically compensates for temperature deviations, achieving precise regulation of complex nonlinear thermal processes. Multi-sensor weighted fusion technology improves temperature monitoring accuracy, and the real-time feedback mechanism enables the AI ​​model to continuously learn adaptively, enhancing the system's response to sudden heat loads and environmental changes, overcoming the limitations of traditional PID control.

[0075] Example 3 illustrates a schematic scheme for an active intelligent thermal management method for a proton exchange membrane fuel cell. It should be noted that the technical solution of this active intelligent thermal management method for a proton exchange membrane fuel cell belongs to the same concept as the technical solution of the active intelligent thermal management system for a proton exchange membrane fuel cell described above. Details not described in detail in this embodiment can be found in the description of the active intelligent thermal management system for a proton exchange membrane fuel cell described above.

[0076] This embodiment also provides an electronic device suitable for implementing active intelligent thermal management of proton exchange membrane fuel cells, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the active intelligent thermal management system for proton exchange membrane fuel cells as proposed in the above embodiment.

[0077] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the active intelligent thermal management system for proton exchange membrane fuel cells as proposed in the above embodiments.

[0078] The storage medium proposed in this embodiment and the active intelligent thermal management system for proton exchange membrane fuel cells proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0079] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0080] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An active intelligent thermal management system for a proton exchange membrane fuel cell, comprising a hardware layer (1) and a control layer (2), characterized in that, The hardware layer consists of an electric stack module (1-1), a cooling actuator (1-2), and a sensor network (1-3); The fuel cell module (1-1) is equipped with a bipolar plate to optimize the heat conduction path; The cooling actuator (1-2) includes a variable frequency fan, a liquid cooling pump, a three-way valve, and a phase change heat storage manifold; The sensor network (1-3) includes distributed temperature sensors, humidity sensors, and current sensors; The control layer (2) consists of an edge controller (2-1) and a cloud server (2-2).

2. The active intelligent thermal management system for proton exchange membrane fuel cells as described in claim 1, characterized in that, The edge controller (2-1) deploys a lightweight AI model to process data and generate control commands; The cloud server (2-2) is used to store historical data, update the lightweight AI model, and provide remote monitoring.

3. A method for active intelligent thermal management of a proton exchange membrane fuel cell, using the system described in claim 1 or 2, characterized in that, Includes the following steps: The actual stack temperature is obtained through a sensor network (1-3). Current and ambient temperature The average temperature of the fuel cell stack was obtained by fusing data from distributed temperature sensors using a weighted average method. ; The average temperature Tavg and current of the fuel cell stack are controlled by the edge controller (2-1). Ambient temperature As input sequence, the LSTM model is used to predict the stack temperature at future times based on historical data. ; The future stack temperature With target temperature The system compares and calculates the temperature deviation. Based on the current state and the reward function, it generates preliminary control actions using a reinforcement learning algorithm. ; The initial control action Using these as initial values, the cooling parameters are optimized using model predictive control to obtain the optimal control parameters. To minimize the temperature deviation in the prediction time domain; Through the optimal control parameters The variable frequency fan and liquid cooling pump drive the cooling actuator (1-2) and simultaneously control the phase change thermal storage manifold to store or release heat; The actual stack temperature after execution With future time stack temperature The prediction error is calculated and fed back to the edge controller (2-1) to update the parameters of the lightweight AI model.

4. The active intelligent thermal management method for proton exchange membrane fuel cells as described in claim 3, characterized in that, The LSTM model employs a 3-layer LSTM structure, with each layer containing 128 neurons and a fully connected layer; input features include electrical current. Ambient temperature Inlet humidity φ, coolant flow rate; The output is the stack temperature at a future time N seconds in advance. ; The loss function uses the mean squared error (MSE), and the calculation steps are as follows: Obtain the actual stack temperature LSTM model predicts future stack temperature ; Calculate the prediction error at each time point i, i.e., the actual stack temperature. Subtract the predicted temperature ; The prediction error at each time point is squared. Sum the squared errors at all n time points; Divide the sum by the number of samples n to obtain the mean squared error (MSE). The LSTM model is trained by minimizing the mean squared error (MSE).

5. The active intelligent thermal management method for proton exchange membrane fuel cells as described in claim 4, characterized in that, When fusing distributed temperature sensor data using the weighted average method, the weighting coefficients... The relationship between sensor location and fuel cell heat flux density is determined by the following formula: ； In the formula, β is the distance sensitivity coefficient; Let n be the distance between sensor i and the hotspot, n be the number of temperature measurement points, and j be the summation index variable.

6. The active intelligent thermal management method for proton exchange membrane fuel cells as described in claim 5, characterized in that, Based on the current state and the reward function, a preliminary control action is generated using a reinforcement learning algorithm. During the process, the current state includes the actual stack temperature at the current time t. Coolant flow rate Phase change heat storage ; The initial control action Including fan speed Liquid cooling pump flow rate regulation and valve opening degree v; The reward function The calculation method is as follows: for the maximum temperature difference inside the fuel cell stack Average temperature deviation of fuel cell stack Cooling system energy consumption After weighted summation and taking the negative value, the calculation formula is as follows: ； In the formula, Let be the reward value at time t; Energy consumption of the cooling system; The maximum temperature difference weighting coefficient; The average temperature difference weighting coefficient; t represents the energy consumption weighting coefficient; t represents the time step index.

7. The active intelligent thermal management method for proton exchange membrane fuel cells as described in claim 6, characterized in that, In the process of optimizing cooling parameters using model predictive control, the optimal control parameter at is obtained by constructing and solving an optimization problem. The optimization problem aims to minimize the temperature deviation and control input variation within the future prediction time domain, and the calculation formula is as follows: ； And satisfy the following constraints: Predicted stack temperature at the k-th step Must be within the allowable temperature range to between; Fan speed Must be within the rated range to between; Phase change heat storage Within capacity range to between; In the formula: J is the objective function value; α is the control smoothing coefficient; Δa(k) is the change in control input at step k; To predict the number of time-domain steps; To control the number of time-domain steps; k is the prediction step index; and These are the minimum and maximum allowable temperatures for the fuel cell stack, respectively. and These are the lower and upper limits of the fan speed, respectively, in r / min; and These are the lower and upper limits for phase change heat storage, respectively.

8. The active intelligent thermal management method for proton exchange membrane fuel cells as described in claim 7, characterized in that, During the process of controlling the storage or release of phase change heat in the phase change heat storage manifold, the phase change heat storage consists of both the sensible heat and latent heat of the phase change material. The calculation steps are as follows: Determine the mass of phase change materials ; Calculate the sensible heat of a phase change material in the solid state using its specific heat capacity. With solid-state temperature change The product is obtained; Calculate the sensible heat of a phase change material in its liquid state using the specific heat capacity of the liquid. With the change in liquid temperature Δ The product is obtained; Calculate the latent heat of the phase change process by considering the mass of the phase change material. Obtained by multiplying with the latent heat of phase transition λ; The total heat storage Qp is obtained by adding the solid sensible heat, liquid sensible heat, and latent heat of phase change. The calculation formula is: 。 9. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the active intelligent thermal management method for proton exchange membrane fuel cells according to any one of claims 3 to 8.

10. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the active intelligent thermal management method for a proton exchange membrane fuel cell according to any one of claims 3 to 8.

Images