Fuel cell self-learning cold start method based on ice front evolution prediction

Abstract

The application belongs to the technical field of fuel cells, and particularly relates to a fuel cell self-learning cold start method based on ice front evolution prediction. The method first establishes and checks a multi-physical field model, then extracts ice generation, ice expansion, local blockage and transmission degradation and other freezing evolution characteristics based on the model, constructs a start-up accessibility margin, a failure boundary and a cold start sample data set, and finally establishes an ice front evolution prediction model and an ice front risk constraint double-layer distributed soft actor-critic control model on the basis, predicts future freezing expansion trends and local blockage risks, and adjusts the start-up current, gas supply, back pressure, purging and auxiliary heating in combination according to the prediction, and updates the prediction model and the control model in combination with real-time operation data. The application realizes prediction of future freezing risks and adaptive optimization control of the cold start process, and improves the success rate, adaptability and safety of the cold start of the fuel cell in a low-temperature environment.

Description

Technical Field

[0001] This application belongs to the field of fuel cell technology, and specifically relates to a self-learning cold start method for fuel cells based on ice front evolution prediction. Background Technology

[0002] Proton exchange membrane fuel cells (PEMFCs) possess advantages such as high energy conversion efficiency, low operating temperature, fast start-up response, and near-zero carbon emissions during operation, making them promising candidates for applications in new energy vehicles, backup power supplies, and distributed power generation. However, when the ambient temperature drops below zero degrees Celsius, the water produced by the electrochemical reaction during the initial startup phase can easily freeze in areas such as the membrane electrode assembly, gas diffusion layer, catalyst layer, and flow channels. The formation and accumulation of ice can obstruct oxygen transport channels, reduce proton conductivity, and shrink the reactive area, thereby inhibiting the electrochemical reaction. Simultaneously, the freezing process can also induce localized volume expansion, stress concentration, and material structural damage, leading to startup failure, performance degradation, and even decreased durability. Therefore, effectively elucidating the coupling mechanism between water production, heat transfer, and freezing during the early startup phase under low-temperature conditions, and thereby achieving highly reliable and low-damage cold-start control, has become a critical technical problem urgently needing to be solved in this field.

[0003] Existing technical solutions for the low-temperature cold start of proton exchange membrane fuel cells mainly focus on auxiliary heating, purge pretreatment, gradual loading, and rule-based control. Auxiliary heating methods increase the initial temperature of the battery through an external heat source to mitigate the risk of product water freezing; purge pretreatment methods introduce gas to purge before shutdown or startup to minimize residual water content within the system; gradual loading methods adjust the output load in stages, allowing the battery to gradually warm up under low-risk conditions through exothermic reactions; and rule-based control methods switch operating strategies based on preset thresholds or operating condition rules. While these methods can improve low-temperature start-up performance to some extent, they mostly rely on preset parameters, offline calibration results, or fixed program logic, making it difficult to adaptively adjust to changes in ambient temperature, differences in stack aging, and fluctuations in initial water content.

[0004] Meanwhile, the cryogenic cold start-up process of proton exchange membrane fuel cells is essentially a highly nonlinear, strongly coupled, and time-varying dynamic process. During this process, water generation, heat transfer, phase change freezing, and gas-liquid-solid multiphase transport interact with each other, and ice formation and expansion typically exhibit significant spatiotemporal nonuniformity. Existing methods mostly rely on macroscopically measurable signals such as average temperature, stack voltage, single-cell voltage, or outlet water flow for control, which often fails to accurately characterize the formation and evolution of the freezing front in local areas, thus making precise intervention before ice blockage occurs.

[0005] In recent years, with the development of multiphysics modeling, embedded sensing, and artificial intelligence technologies, data-driven methods have begun to be introduced into the field of fuel cell cryogenic management, showing potential in state identification, process prediction, and control optimization. However, existing methods still have significant shortcomings: First, they lack sufficient physical constraints for key internal processes such as freezing phase transition and local water migration, easily leading to prediction results inconsistent with actual mechanisms; second, they lack clear characterization of the freezing initiation location, blockage formation process, and start-up failure boundary, making it difficult to provide clear spatiotemporal basis for control actions; third, under complex conditions such as ambient temperature fluctuations, uncertain residual water, and fuel cell stack aging, the robustness and adaptability of control strategies remain limited. Therefore, constructing a mechanism and data fusion prediction model oriented towards the evolution of the freezing front, and forming a collaborative control method capable of online learning and adaptive correction based on this model, has significant research value and engineering significance for improving the success rate of cryogenic cold start, reducing the risk of local ice blockage, and extending the lifespan of the fuel cell stack.

[0006] For example, Chinese patent application CN201711315725.X, entitled "A Low-Temperature Start-up Method for a Proton Exchange Membrane Fuel Cell," discloses a low-temperature start-up scheme that does not rely on an external heating device. This scheme, after the battery is shut down at room temperature, first uses nitrogen purging to reduce the risk of residual water freezing inside the system; then, during the low-temperature start-up phase, it employs a balanced loading method, allowing the fuel cell to heat up through its own electrochemical reaction and gradually reach the normal start-up temperature. This scheme essentially belongs to a technical route combining residual water pretreatment and fixed-load heating. Its control actions mainly rely on preset programs such as fixed purging and balanced loading, and it has not yet established a clear mapping relationship between the residual water content, local freezing behavior, and start-up results. Therefore, it is difficult to reflect the internal state evolution during the cold start-up process, and it is also difficult to answer key questions such as the location where freezing first occurs, the timing of blockage formation, and the timing of control action switching. Under different ambient temperatures, different residual water levels, and different aging conditions, the adaptability and robustness of this type of empirically-based start-up method remain limited.

[0007] For example, Chinese patent application CN201510989410.8, entitled "A Low-Temperature Start-up Control Method for a Fuel Cell System," discloses a low-temperature start-up scheme that combines post-shutdown state regulation and restart control. During the shutdown phase, this scheme achieves purging and water removal through a large reactant gas supply and maintains membrane wetness through a low intermittent loading current. During the restart phase, it combines the changes and uniformity of single-cell voltage to judge the membrane wetness and dryness state and purging effect, and matches corresponding starting conditions and loading strategies accordingly to improve the success rate of low-temperature start-up. This scheme essentially belongs to a technical route combining shutdown state management and restart matching control. Its judgment of system freezing behavior mainly relies on externally measurable signals such as single-cell voltage and its uniformity, and has not yet established a unified state characterization between local temperature field, water content, ice formation distribution, and blockage risk. Therefore, although this scheme can indirectly reflect the purging effect and membrane wetness and dryness state to a certain extent, it still cannot directly identify the preferential growth region of freezing, nor can it predict the subsequent evolution trend of the freezing front. Overall, this type of method still belongs to the indirect control method based on macroscopic electrical signal feedback, and it is difficult to suppress local ice blockage and local irreversible damage in a timely manner under complex low temperature conditions.

[0008] In summary, existing cryogenic cold-start technologies for proton exchange membrane fuel cells generally suffer from insufficient perception of the internal freezing state, inadequate description of the evolution of the freezing front, and limited adaptive capabilities of the control strategy. These limitations make it difficult to meet the application requirements of high reliability and low-damage cold start under complex operating conditions. Therefore, there is an urgent need to propose a cryogenic cold-start method capable of predicting the evolution of the freezing front and achieving adaptive and coordinated control based on real-time conditions, in order to overcome the shortcomings of existing technologies. Summary of the Invention

[0009] To address the shortcomings of existing fuel cell cold start technologies, which primarily rely on preset programs, fixed control laws, and external macroscopic signals for state assessment during cryogenic cold start, making it difficult to directly characterize ice formation, ice expansion, and local ice blockage evolution, and lacking foresight, adaptability, and robustness under complex cryogenic environments, varying residual water levels, and different aging conditions, this application proposes a self-learning cold start method for fuel cells based on ice front evolution prediction. By constructing an ice front evolution characterization model for the cryogenic cold start process, the method establishes the correlation between freezing expansion, local blockage trends, and start-up accessibility. Furthermore, by combining multi-physics mechanism constraint prediction and risk constraint collaborative control, the method performs self-learning joint adjustment of key variables such as start-up current, gas supply, back pressure, purging, and auxiliary heating. This enables the prediction of future freezing risks and adaptive optimization control of the cold start process, thereby improving the success rate, environmental adaptability, and operational safety of fuel cells in cryogenic environments.

[0010] This application provides a self-learning cold-start method for fuel cells based on ice front evolution prediction, the method comprising:

[0011] Step 1: Construct a low-temperature cold start test platform for proton exchange membrane fuel cells, collect cold start test data under different ambient temperatures, different initial water content states and different start-up conditions, and establish a multiphysics model of fuel cell low-temperature cold start based on the test data, fuel cell structural parameters and low-temperature cold start boundary conditions to characterize the coupled processes of gas transport, liquid water migration, heat transfer, electrochemical reaction and freezing phase change, and verify the multiphysics model using the test data;

[0012] Step 2: Based on the verified multiphysics model, construct a set of frozen state variables for ice formation and ice expansion during the low-temperature cold start process, and extract the characteristics of ice formation, ice expansion, local blockage and transport degradation during the low-temperature cold start process. Establish a set of state representations for ice front evolution, and construct a start-up reachability margin and failure boundary judgment model based on the set of state representations. Combine the ambient temperature, residual water state, gas supply parameters, back pressure parameters, auxiliary heating parameters and start-up path parameters to perform multi-condition simulation and generate frozen evolution data and cold start sample data.

[0013] Step 3: Based on the freezing evolution data and cold start sample data, establish an ice front evolution prediction model and an ice front risk constraint two-layer distributed soft actor-commentator model. Use the ice front evolution prediction model to predict the future freezing expansion trend and local blockage risk, and jointly adjust the start-up current, gas supply, back pressure, purging and auxiliary heating based on the prediction results. During the cold start operation, update the ice front evolution prediction model and the ice front risk constraint two-layer distributed soft actor-commentator model online according to the real-time operation data.

[0014] In a preferred implementation, step 2 further includes:

[0015] Step 2.1: Based on the verified multiphysics model, define the ice front horizontal set function, determine the freezing interface position according to the zero value interface of the ice front horizontal set function, determine the ice front normal unit vector according to the spatial gradient of the ice front horizontal set function, determine the ice front normal advance velocity according to the ratio of the rate of change of the ice front horizontal set function with respect to time to the spatial gradient, define the local ice saturation based on the ratio of the ice volume to the local pore volume in the local control body, and define the local freezing rate based on the derivative of the local ice phase volume fraction with respect to time, so as to construct a set of freezing state variables for ice generation and ice expansion during the low temperature cold start process;

[0016] In step 2.2, the set of frozen state variables output in step 2.1 is used as input to extract features of ice generation, ice expansion, local blockage and transmission degradation during the cold start process, and a set of key frozen features is constructed.

[0017] Step 2.3: Based on the key freezing characteristics in Step 2.2, construct the start-up reachability margin and failure boundary determination model. The start-up reachability margin is obtained by weighted integration of the ratio of the difference between the local temperature and the freezing reference temperature in the key area to the reference temperature rise scale, the ratio of the local current density to the reference current density, the local ice saturation, the local oxygen diffusion blockage coefficient, and the effective reactive area decay rate. The failure boundary determination model is constructed using the start-up reachability margin, the maximum ice saturation in the key area, the maximum oxygen diffusion blockage coefficient in the key area, and the maximum thermo-electric coupling instability index in the key area. The boundary determination of the cold start process is performed based on the sign and duration of the failure boundary function.

[0018] Step 2.4: Based on the key freezing feature set in Step 2.2 and the start-up reachability margin in Step 2.3, construct a freezing feature vector. Construct a multi-condition parameter vector according to the boundary conditions and control parameters affecting the low-temperature cold start freezing evolution. Based on the condition parameter vector, call the verified multiphysics model to automatically solve multiple conditions in batches. Combined with Steps 2.1 to 2.3, extract freezing state variables, extract key freezing features, evaluate start-up reachability, and determine failure boundaries to generate freezing feature trajectories and cold start sample labels for each group of conditions. Construct a cold start sample dataset by combining the condition parameter vector, freezing feature trajectory, and cold start sample labels.

[0019] In a preferred implementation, step 3 further includes:

[0020] Step 3.1: Based on the current operating status, frozen feature vector, and initial operating parameters, construct an ice front evolution prediction model, and perform rolling prediction of the freezing features for the next H steps based on the historical input sequence of length L, to obtain the prediction results of the future freezing features and the probability of successful start-up;

[0021] Step 3.2: Based on the future freezing feature prediction results obtained in Step 3.1, a two-layer distributed soft actor-commentator control model with ice front risk constraints is constructed. The current observation input vector, the future freezing feature prediction results, the cold start failure boundary function, and the action executed at the previous moment are used to form the system state vector. The system state vector is input into the upper-layer policy network, which outputs the stage pattern. The system state vector and the stage pattern are input into the lower-layer policy network, which outputs the original continuous action vector. A comprehensive value function and an immediate reward function are constructed based on the start-up benefit commentator, the freezing risk commentator, and the damage risk commentator.

[0022] Step 3.3: Project the original continuous action vector output in Step 3.2 with safety constraints to obtain the final execution action that satisfies the physical constraints of the equipment, the operating boundary constraints, and the future freeze risk constraints. Then, coordinate the control of the electronic load, the gas supply unit, the back pressure unit, the purging unit, and the auxiliary heating unit based on the final execution action.

[0023] Step 3.4: Based on the newly added operating parameters, frozen feature trajectories, action sequences and startup results during the cold start operation, an online experience pool is constructed. Combined with the priority replay mechanism based on failure boundaries, small step incremental update, target network soft update and policy offset constraints, the ice front evolution prediction model and reinforcement learning control model are updated online.

[0024] In the preferred implementation, further, in step 2.2, the key frozen feature set... Including the average advance speed of the ice front Freeze coverage Freeze bias factor (t), oxygen diffusion blocking coefficient field Effective reactive surface area decay rate Thermo-electric coupling instability index field Maximum ice saturation Maximum oxygen diffusion blocking coefficient Maximum thermo-electric coupling instability index and maximum freezing rate ;

[0025] Among them, the frozen coverage rate for:

[0026] ;

[0027] Freeze bias factor (t) is:

[0028] ;

[0029] Effective reactive surface area decay rate for:

[0030] ;

[0031] In the formula: t is the discrete time or the current calculation time; This is a critical area for cold starts; For key areas Volume; In key areas Volume division is performed on top; For the Heaviside function; For a moment Spatial location Local ice phase volume fraction at a given location; It is a spatial position vector; in two dimensions it can be... In three dimensions, it can be ; The threshold for ice phase identification; For key areas The upstream region is divided along the mainstream direction; For key areas The downstream region is divided along the main trend direction; This is the integral value of the ice phase volume fraction in the upstream region; This is the integral value of the ice phase volume fraction in the downstream region; This represents the overall integral value of the ice phase volume fraction within the key region; For a moment The effective reactive surface area; For reference, the initial active area; This is the catalytic layer region; It is a local activity reachability function; To obtain the effective reactive area, the local activity reachability function is integrated over the catalytic layer region.

[0032] In the preferred implementation, further, in step 2.2, the maximum ice saturation... for:

[0033] ;

[0034] Maximum oxygen diffusion blocking coefficient for:

[0035] ;

[0036] Maximum thermo-electric coupling instability index for:

[0037] ;

[0038] Maximum freezing rate for:

[0039] ;

[0040] In the formula: For a moment ,Location Local ice saturation at the location; For a moment ,Location To control the volume of ice within the body; For position Local porosity at the location; To control the overall volume of the body; For a moment ,Location Local oxygen diffusion blocking coefficient at the location; The effective oxygen diffusion coefficient under freezing conditions; The effective oxygen diffusion coefficient under unfrozen conditions; The intrinsic diffusion coefficient of oxygen; For a moment ,Location Local liquid water saturation at the location; m and n are structural correction indices; For a moment ,Location Local thermal-electric coupling instability indicators at the location; For local temperature gradients; The gradient is a local voltage characterization quantity. For local current density gradient; , , These are the weighting coefficients for the temperature gradient, voltage gradient, and current density gradient terms in the thermo-electric coupling instability index, respectively. For a moment ,Location Local freezing rate at the location; This represents the rate of change of the local ice phase volume fraction with respect to time. In key areas The maximum value is obtained for the corresponding local variable.

[0041] In the preferred implementation, further, in step 2.3, the start-up reachability margin at time t... for:

[0042]

[0043] In the formula: This is the reference temperature for freezing. For reference temperature rise scale; Reference current density; ~ These are the weighting coefficients;

[0044] Failure Boundary Function for:

[0045] ;

[0046] In the formula: For a moment ,Location Local current density at the location; This is the reference temperature for freezing. For reference temperature rise scale; For a moment ,Location Local current density at the location; Reference current density; For a moment ,Location Local ice saturation at the location; For a moment ,Location Local oxygen diffusion blocking coefficient at the location; For a moment The effective reactive surface decay rate; It is a volume infinitesimal element; to To activate the reachability weighting coefficient; To enable reachability margin; Maximum ice saturation in the key area; The maximum oxygen diffusion blocking coefficient in the critical area; The maximum thermal-electric coupling instability index in the critical area; to This represents the failure boundary weighting coefficient.

[0047] In the preferred implementation, further, in step 2.4, the feature vector is frozen. for:

[0048] ;

[0049] Construct the operating condition parameter vector p:

[0050] ;

[0051] Cold start sample set D:

[0052] ;

[0053] In the formula: The average advance velocity of the ice front; For frozen coverage; This is the freezing bias factor; To enable reachability margin; This represents the maximum ice saturation within the critical area. The maximum oxygen diffusion blockage coefficient in the critical area; The maximum thermo-electric coupling instability index in the critical area; The effective reactive surface area decay rate; The maximum freezing rate within the critical area; Ambient temperature; This represents the initial water content of the membrane. The residual water level after shutdown; The cathode air excess coefficient; The excess hydrogen coefficient at the anode; This is the cathode back pressure; This is the anode back pressure; To supplement heating power; For the startup current ramp-up rate; These are the parameters for triggering the purge. For the first The vector of working parameters corresponding to the group of working conditions; For the first Freezing feature trajectory corresponding to the group working condition; For the first The sample label corresponding to the working condition group; j is the working condition number; N is the total number of samples.

[0054] In the preferred implementation, further, in step 3.1, the input vector for:

[0055] ;

[0056] The formula for predicting future frozen features is:

[0057] ;

[0058] Success rate of startup for:

[0059] ;

[0060] In the formula: The average reactor temperature; It is an indicator of voltage non-uniformity; and These are the cathode pressure and anode pressure at time t, respectively; and These are the excess air coefficient at the cathode and the excess hydrogen coefficient at the anode, respectively, at time t. These are the excess air coefficient at the cathode and the excess hydrogen coefficient at the anode, respectively, at time t. The auxiliary heating power at time t; These are the initial operating parameters, consisting of ambient temperature, initial membrane water content, and residual water level. Let be the predicted value of the frozen feature at the current time t for the k-th future step; is the mapping function for predicting ice front evolution; L is the length of the historical input sequence; H is the prediction time domain length or the number of future control steps; k is the future prediction step number. ; This is a prediction of the start-up reachability margin corresponding to the predicted end point in the time domain; For Sigmoid mapping functions; To predict the maximum ice saturation value corresponding to the end point in the time domain; The predicted value of the maximum oxygen diffusion blockage coefficient corresponding to the end of the time domain; This is the predicted value of the maximum thermo-electric coupling instability index corresponding to the end point in the time domain; to The weighting coefficients are mapped to the probability of successful startup, with the startup reachability margin taking a positive input and the maximum ice saturation, maximum oxygen diffusion blockage coefficient, and maximum thermo-electric coupling instability index taking a negative input.

[0061] In the preferred implementation, further, in step 3.2, the system state vector for:

[0062] ;

[0063] Discrete-stage patterns are output from the upper-level policy network. for:

[0064] ;

[0065] In the formula, This is the upper-level policy mapping function. These are discrete-stage pattern variables;

[0066] The lower-level policy network outputs the original continuous action vector. for:

[0067] ;

[0068] Comprehensive value function for:

[0069] ;

[0070] Instant reward function for:

[0071] ;

[0072] In the formula: This is the current observation input vector; Predict sequences for future frozen features; This is the boundary function for cold start failure. The action performed in the previous moment; This is the mapping function for the upper-level strategy; This is the mapping function for the lower-level strategy; To initiate revenue commentary output; To freeze risk commentators' output; Output for damage risk commentators; and These are the penalty weights for freezing risk and damage risk, respectively; To initiate the revenue item; To freeze risk items; This is the damage cost term; This is an auxiliary energy consumption item; To initiate the accessibility margin increment; This represents the average reactor temperature increment. to To activate the profit weighting coefficient; To predict the maximum ice saturation in the future time domain; This is the maximum oxygen diffusion blocking coefficient predicted in the future time domain; This is the maximum oxygen diffusion blocking coefficient predicted in the future time domain; This represents the average value of the maximum thermo-electric coupling instability index in the future prediction time domain; The equivalent thermomechanical stress index of the membrane electrode; to To freeze the risk weighting coefficient; to This refers to the damage penalty weighting coefficient; to These are the auxiliary heating energy consumption weight and the purging operation cost weight, respectively.

[0073] In the preferred implementation, further, in step 3.3, the final action to be executed that minimizes the distance to the original continuous action vector is obtained within the feasible domain of safe actions. for:

[0074] ;

[0075] The risk constraints for freezing assets are as follows:

[0076] ;

[0077] In the formula: In order to match the current stage model The corresponding feasible domain for safe actions; It is a norm 2; The predicted value of the failure boundary in the time domain is predicted for the k-th future step. This represents the minimum permissible safety margin.

[0078] The beneficial effects of this application are:

[0079] First, the self-learning cold start method for fuel cells based on ice front evolution prediction proposed in this application focuses on the ice formation, ice expansion, and local blockage evolution mechanism during the low-temperature cold start of proton exchange membrane fuel cells. It first establishes a cold start analysis foundation that characterizes the coupling relationship between gas transport, liquid water migration, heat transfer, electrochemical reactions, and freezing phase change by combining experimental data with a multiphysics model. This overcomes the limitations of existing technologies that mainly rely on average temperature, stack voltage, or empirical rules for indirect judgment. Furthermore, this application constructs a set of frozen state variables, a set of state representations, a start-up reachability margin, and a failure boundary determination model around the ice front evolution process. This enables dynamic identification and early warning of the freezing formation location, expansion trend, blockage risk, and start-up failure boundary. This application enhances the ability to characterize the internal state and failure process of low-temperature cold start. Based on this, it establishes an ice front evolution prediction model and an ice front risk constraint two-layer distributed soft actor-commentator model, directly incorporating future freezing expansion trends and local blockage risks into the control decision-making process. It performs coordinated optimization control of start-up current, gas supply, back pressure, purging, and auxiliary heating, and can continuously update the prediction model and control strategy online based on real-time data during cold start operation. Therefore, it can better adapt to complex operating conditions such as changes in ambient temperature, residual water fluctuations, and stack aging, improve the foresight, adaptability, and robustness of low-temperature cold start control, help improve the success rate of cold start, reduce the risk of local ice blockage and irreversible damage, and extend the service life of fuel cell stacks.

[0080] Secondly, in the preferred implementation, steps 2.1 to 2.4 of this application first construct a set of freezing state variables based on a verified multiphysics model, utilizing the ice front level set function, local ice saturation, and local freezing rate. This enables continuous characterization of the freezing interface location, propagation direction, propagation speed, and local icing intensity during the low-temperature cold start process. This elevates the existing technology from indirect judgment relying primarily on macroscopic signals such as reactor voltage and average temperature to a direct description of the ice formation and expansion process. Furthermore, based on this, freezing coverage, freezing bias factor, oxygen diffusion blockage coefficient, effective reactive area decay rate, thermo-electric coupling instability index, and extreme characteristics of key regions are further extracted. This allows for a comprehensive characterization of cold start risks from multiple dimensions, including freezing expansion, mass transfer degradation, reaction decay, and local instability, thus revealing the location of local ice blockage, freezing evolution trend, and... The degree of transmission degradation has a clearer criterion; secondly, by constructing a start-up reachability margin and failure boundary judgment model, the most unfavorable local freezing state and the overall start-up capability are uniformly coupled into a quantifiable criterion, and the boundary judgment is performed by combining the failure boundary function sign and duration, which helps to reduce misjudgment caused by instantaneous fluctuations and improve the accuracy and stability of low-temperature cold start failure identification; finally, by further constructing the key freezing features and start-up reachability margin into a freezing feature vector, and combining it with multi-condition parameter vectors to carry out automatic batch solution, a cold start sample dataset covering different ambient temperatures, residual water states, gas supply parameters, back pressure parameters, auxiliary heating parameters and start-up paths is generated. This not only provides clear state input for subsequent ice front evolution prediction models and control models, but also enhances the model's coverage of complex conditions, critical conditions and failure conditions.

[0081] Third, in the preferred implementation, steps 3.1 to 3.4 of this application construct an ice front evolution prediction model based on the current operating state, freezing feature vector, and initial operating parameters. A historical input sequence of length L is used to perform rolling predictions of the freezing characteristics for the next H steps. This enables the control system to move from a passive response to the current state to a forward-looking judgment of future freezing expansion trends, local blockage risks, and the probability of successful startup, thereby improving the predictability of low-temperature cold start decisions. Furthermore, based on this, a two-layer distributed soft actor-commentator control model with ice front risk constraints is constructed. The upper-layer policy network outputs stage patterns, and the lower-layer policy network outputs the original continuous action vector. Distributed evaluation is performed using three types of commentators: startup benefits, freezing risks, and damage risks. This ensures that the control decision can match the mechanistic characteristics of different stages of cold start while also considering heating startup efficiency and freezing suppression requirements. Furthermore, by implementing structural safety constraints, the system achieves coordinated optimization of multiple variables, including start-up current, gas supply, back pressure, purging, and auxiliary heating. Secondly, by projecting safety constraints onto the original continuous action vector, the final execution action is obtained while satisfying equipment physical constraints, operational boundary constraints, and future freezing risk constraints. This avoids the problem that theoretically optimal control outputs may be unexecutable or introduce new risks in the actual system, thereby improving the safety of the control strategy. Finally, by constructing an online experience pool and combining a priority replay mechanism based on failure boundaries, small-step incremental updates, target network soft updates, and strategy offset constraints, the ice front evolution prediction model and reinforcement learning control model are updated online. This allows the application to continuously correct its prediction capabilities and control parameters as environmental temperature changes, residual water fluctuations, and stack aging occur, thereby enhancing the self-learning capability of the cryogenic cold start control method. Attached Figure Description

[0082] Figure 1 This is a flowchart of the self-learning cold start method for fuel cells based on ice front evolution prediction according to the present invention.

[0083] Figure 2 This is a flowchart of the IRCB-DSAC algorithm for the self-learning cold start method of fuel cells based on ice front evolution prediction, as described in this invention. Detailed Implementation

[0084] To enable those skilled in the art to better understand the technical solutions of this application, the following will provide a more detailed description of this application in conjunction with the accompanying drawings and embodiments.

[0085] The directional terms such as above, below, left, right, front, and back used in this application are based on the positional relationships shown in the attached drawings. Different attached drawings may result in different positional relationships, therefore they should not be interpreted as limitations on the scope of protection.

[0086] In this application, the terms "installation," "connection," "interlocking," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, an integral connection, a mechanical connection, an electrical connection, or a connection that allows communication between components. They can also refer to a direct connection or an indirect connection through an intermediate medium. They can refer to the internal connection of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in this application can be understood according to the specific circumstances.

[0087] This invention describes a self-learning cold-start method for fuel cells based on ice front evolution prediction. It addresses the shortcomings of existing technologies in proton exchange membrane fuel cells during cryogenic cold start-up, including insufficient perception of the internal freezing state, inadequate description of the formation and expansion of the freezing front, difficulty in precise intervention before local ice blockage forms, and reliance on preset parameters, empirical rules, and offline calibration for control strategies, which are ill-suited to complex conditions such as ambient temperature fluctuations, changes in initial water content, and stack aging. The invention proposes a cold-start technology solution integrating multiphysics mechanism modeling, freezing front state characterization, ice front evolution prediction, and online self-learning collaborative control. By constructing a state characterization system based on the freezing interface position, propulsion speed, local ice saturation, freezing rate, and their risk characteristics, and by adaptively optimizing and adjusting the start-up current, gas supply, back pressure, purging, and auxiliary heating based on real-time operating data, this invention can improve the cold-start success rate of fuel cells in cryogenic environments, reduce the risk of local ice blockage and irreversible damage, and improve the reliability and service life of the fuel cell stack.

[0088] As per the instruction manual Figure 1 This invention provides a self-learning cold-start method for fuel cells based on ice front evolution prediction. This method is applied to the field of low-temperature start-up control of proton exchange membrane fuel cells, particularly to adaptive cold-start control of proton exchange membrane fuel cell systems for new energy vehicles, fuel cell systems for backup power, and fuel cell systems for distributed power generation in low-temperature environments. The method includes:

[0089] Step 1: Construct a low-temperature cold start test platform for proton exchange membrane fuel cells, collect cold start test data under different ambient temperatures, different initial water content states, and different start-up conditions, and establish a multiphysics model of low-temperature cold start of fuel cells based on the test data, fuel cell structural parameters, and low-temperature cold start boundary conditions to characterize the coupled processes of gas transport, liquid water migration, heat transfer, electrochemical reaction, and freezing phase change. The multiphysics model is then verified using the test data.

[0090] Specifically, a low-temperature cold start test platform for proton exchange membrane fuel cells was first constructed. The test platform includes a fuel supply unit, an air supply unit, a back pressure control unit, an auxiliary heating unit, an environmental temperature control unit, an electronic load unit, a multi-source information acquisition unit, and a host computer control unit. The connections are as follows: the fuel supply unit is connected to the anode inlet of the fuel cell stack via the anode inlet pipe; the air supply unit is connected to the cathode inlet of the fuel cell stack via the cathode inlet pipe; the back pressure control unit is located on both the anode and cathode outlet pipes; the electronic load unit is electrically connected to the stack output; the auxiliary heating unit is thermally connected to the stack endplate, shell, and inlet pipes; the environmental temperature control unit is located outside the stack, forming a closed low-temperature test space; and the host computer control unit is communicatively connected to each execution unit and acquisition unit for setting operating conditions and recording data.

[0091] The fuel supply unit, consisting of a hydrogen cylinder, pressure reducing valve, mass flow controller, inlet temperature and humidity control assembly, and solenoid valve, is used to stably supply hydrogen to the anode. The air supply unit, consisting of an air compressor or compressed air source, mass flow controller, inlet temperature and humidity control assembly, and solenoid valve, is used to stably supply gas to the cathode. The back pressure control unit uses a proportional back pressure valve to regulate the anode and cathode exhaust gases separately. The auxiliary heating unit uses an electric heating film, electric heating belt, or electric heating plate to locally heat the end plates or inlet piping. The ambient temperature control unit uses a low-temperature environment chamber to enclose the fuel cell stack and related piping at the target ambient temperature. The electronic load unit is used to output constant current, ramp current, or segmented current loading signals.

[0092] After the platform is set up, integration testing is conducted first. This includes checking the sealing of all gas line connectors to confirm there are no leaks; calibrating the flow meter, pressure sensor, temperature and humidity sensor, voltage acquisition board, and current sampling module; testing the back pressure valve response accuracy; testing the power output stability of the auxiliary heating unit; testing the accuracy of the electronic load; and confirming that the host computer can synchronously control each unit and record all channel data. Formal testing only begins when all units are functioning normally and the time synchronization error is within the preset range.

[0093] After the platform was built, a low-temperature cold start test was conducted. Before the test, the fuel cell stack underwent a shutdown pretreatment to obtain different initial water content states. The initial water content state was achieved by adjusting the shutdown purging time, purging flow rate, purging gas type, pre-shutdown operating current, and post-shutdown settling time. For example, the purging time was set to short, medium, and long, corresponding to high, medium, and low residual water states, respectively. After the shutdown pretreatment, the fuel cell stack was placed in an environmental temperature control unit to cool to the target ambient temperature and maintained for a certain equilibration time to ensure the overall temperature of the fuel cell stack matched the ambient temperature.

[0094] During formal testing, the ambient temperature, initial moisture content, and startup conditions are varied according to a preset test matrix. Multiple low-temperature points can be selected for the ambient temperature, and multiple levels of initial moisture content can be chosen. Startup conditions include at least the startup current loading method, anode and cathode excess coefficients, anode and cathode back pressure, and auxiliary heating power. Each test is executed in the following order: first, the inlet gas supply and back pressure are established; then, the auxiliary heating unit is turned on as needed; subsequently, the electronic load unit loads the startup current and starts timing until startup is successful or the failure criteria are met. The entire testing process is automatically executed and data is stored by the host computer.

[0095] During the low-temperature cold start test, key parameters are collected simultaneously. These key parameters include at least the total voltage, individual cell voltage, current, ambient temperature, stack surface temperature, anode inlet temperature, cathode inlet temperature, anode outlet temperature, cathode outlet temperature, anode inlet humidity, cathode inlet humidity, anode inlet pressure, cathode inlet pressure, anode outlet pressure, cathode outlet pressure, anode flow rate, cathode flow rate, AC impedance, and thermal imaging information of the stack surface.

[0096] Among these parameters, total voltage and individual section voltage reflect the overall output capacity and local imbalance during cold start. Current reflects load conditions. Temperature parameters reflect the heating rate and heat distribution. Humidity parameters reflect water management status. Pressure and flow parameters reflect gas supply capacity and channel blockage risk. AC impedance reflects membrane hydration status, mass transfer resistance, and freeze-induced internal resistance changes. Thermal imaging information is used to identify local cold spots, hot spots, and potential freeze-induced preferential growth areas.

[0097] After completing multiple sets of tests, the validity of the experimental data was verified. Verification included: whether the results of repeated tests under the same test conditions were consistent; whether the voltage recovery curve, temperature rise curve, pressure response curve, and impedance change curve showed a repeatable trend under the same ambient temperature and initial water content; and whether the location of localized low-temperature areas in the thermal image correlated with the location of single-section voltage anomalies. If a set of data exhibited significant sensor drift, sudden communication interruption, or gas path anomalies, that set of data was discarded and not used for model correction.

[0098] After validating the data, parameter corrections are prepared based on the experimental data. It's important to note that this parameter correction doesn't directly modify the model results; instead, it involves working backward from the experiments to deduce the boundary and equivalent parameters needed for the model. For example, the inlet boundary is corrected based on measured inlet flow and pressure data; the range of the external heat transfer coefficient is corrected based on thermal imaging and surface temperature measurements; the initial values ​​of membrane water content parameters are corrected based on the AC impedance variation trend; the residual water level range is estimated based on the startup differences under different shutdown and purging schemes; and the initial value ranges of electrochemical kinetic and thermal parameters are estimated based on the voltage recovery rate and temperature rise rate. This avoids relying entirely on empirically given parameters during subsequent modeling.

[0099] In this embodiment, after the experimental data is verified, a set of basic inputs that can be used for modeling are obtained, including: ambient temperature boundary, cathode and anode inlet flow boundary, cathode and anode inlet pressure boundary, auxiliary heating power boundary, start-up current boundary, initial water content state boundary corresponding to shutdown pretreatment, and temperature rise curve, voltage recovery curve, pressure response curve, impedance change curve and thermal image distribution sequence for model parameter identification.

[0100] After verifying the experimental data and correcting the parameters, a multiphysics model for the low-temperature cold start of the fuel cell was established. The modeling objects include the flow channel, gas diffusion layer, microporous layer, catalyst layer, and proton exchange membrane. The model inputs include the stack geometry, interlayer thickness, porosity, permeability, thermal conductivity, electrical conductivity, specific heat capacity, reaction activity parameters, and the boundary conditions obtained from the aforementioned experiments.

[0101] The multiphysics model includes at least a gas transport sub-model, a liquid water migration sub-model, a heat transfer sub-model, an electrochemical reaction sub-model, and an ice-phase transition sub-model. The gas transport sub-model describes the diffusion and convection transport of hydrogen, oxygen, and water vapor in the flow channel and porous medium; the liquid water migration sub-model describes the seepage, accumulation, and migration of generated water in the catalyst layer, microporous layer, and gas diffusion layer; the heat transfer sub-model describes the heat of reaction, ohmic heat, latent heat of phase change, and heat exchange with the environment; the electrochemical reaction model describes the rates of the anodic hydrogen oxidation reaction and the cathodic oxygen reduction reaction, coupling local temperature, reactant concentration, and water state into the reaction kinetics; and the ice-phase transition model describes the process of liquid water transforming into the ice phase under low-temperature conditions, incorporating the effects of ice on porosity, effective diffusion coefficient, effective thermal conductivity, and local transport resistance into the model.

[0102] In modeling, a geometric domain is first established based on the stack structure, followed by mesh generation. Then, governing equations and boundary conditions are established. Boundary conditions include: ambient temperature boundary, inlet flow rate boundary, inlet pressure boundary, inlet temperature and humidity boundary, outlet back pressure boundary, start-up current loading boundary, and auxiliary heating heat flux boundary. Initial conditions include: initial temperature field, initial membrane water content, and initial residual water distribution. To reflect the characteristics of low-temperature cold start, freezing initiation temperature, ice phase formation rate relationship, and the blocking effect of ice on oxygen diffusion and liquid water migration are set in the phase change model. Through numerical solution, the temperature field, aqueous phase field, ice phase field, current density distribution, oxygen concentration distribution, and local reactivity distribution at each moment during the cold start process are output.

[0103] After the model is established, it is verified using the data obtained from the previous experiments. The specific verification process is as follows: First, a set of known test conditions is input into the model to obtain the simulation output. Then, the simulation output is compared with the test results under the same conditions item by item. If the deviation exceeds the preset threshold, the sensitive parameters are corrected and recalculated until the simulation results are consistent with the test results within the allowable range.

[0104] The comparison indicators should include at least: the average heating rate of the fuel cell stack, the time required to reach the target temperature, the voltage recovery time, the minimum single-cell voltage, the dispersion of single-cell voltage, the anode and cathode pressure fluctuation amplitude, the impedance change trend, and the location of local low-temperature regions in the thermal image. If the model can simultaneously and well reproduce the temperature rise process, voltage recovery process, pressure response process, impedance change process, and the start-up success or failure results under different operating conditions, then the model is considered to have completed the verification.

[0105] In this embodiment, to avoid the model being applicable only to a single operating condition, a cross-verification method using multiple operating conditions can be adopted. That is, a portion of the operating condition data is selected for parameter identification, and another portion of the operating condition data is used for independent verification. Only when the model maintains high consistency under different ambient temperatures, different initial water content states, and different start-up paths is it output as the verified multiphysics model. The verified multiphysics model can provide a reliable foundation for subsequent ice front evolution feature extraction, freezing risk assessment, and self-learning control.

[0106] Step 2: Based on the verified multiphysics model, extract the characteristics of ice generation, ice expansion, local blockage and transport degradation during the low-temperature cold start process, establish a state characterization set for the evolution of the ice front, and construct a start-up reachability evaluation model and a failure boundary judgment model based on the state characterization set. Combine the ambient temperature, residual water state, gas supply parameters, back pressure parameters, auxiliary heating parameters and start-up path to perform multi-condition simulation and generate freezing evolution data and cold start sample data.

[0107] Step 2 includes:

[0108] Step 2.1: Based on the verified multiphysics model, define the ice front horizontal set function, determine the freezing interface position according to the zero-value interface of the ice front horizontal set function, determine the ice front normal unit vector according to the spatial gradient of the ice front horizontal set function, determine the ice front normal advance velocity according to the ratio of the rate of change of the ice front horizontal set function with respect to time to the spatial gradient, define the local ice saturation based on the ratio of the ice volume to the local pore volume within the local control body, and define the local freezing rate based on the derivative of the local ice phase volume fraction with respect to time, so as to construct a set of freezing state variables for ice formation and ice expansion during the low-temperature cold start process.

[0109] Step 2.1 first calls the verified multiphysics model obtained in Step 1. Under given ambient temperature, initial water content, gas supply boundary, back pressure boundary, auxiliary heating boundary, and starting current boundary conditions, a transient solution is performed on the cold start process to obtain the temperature field, liquid water volume fraction field, ice phase volume fraction field, and pore occupancy distribution of the film region, electrode region, gas diffusion layer region, microporous layer region, and flow channel region at each discrete time. The verified multiphysics model can output the temperature field, aqueous phase field, and spatiotemporal evolution results of the reaction state during the low-temperature cold start stage. Step 2.1 defines the frozen state variables based on these solution results.

[0110] In terms of spatial discretization, the membrane, electrodes, gas diffusion layer, microporous layer, and flow channel are divided into multiple control volumes. For each control volume, spatial coordinates x and discrete time series t are established. At each discrete time point, the ice phase volume fraction, liquid water volume fraction, porosity, and control volume within the corresponding control volume are read, and a freezing state calculation grid is established. The freezing state calculation grid corresponds one-to-one with the multiphysics model solution grid, and the control volume index remains consistent in the time direction to ensure direct comparison of the freezing front position and calculation of local freezing rates between adjacent time points.

[0111] During the frozen interface identification phase, the ice front level set function is defined. and will The area is recorded as the unfrozen area. ,Will The area is marked as a frozen area. ,Will The zero-value interface is denoted as the ice front interface. :

[0112]

[0113] In the formula, Let t be a spatial position vector, and t be time. This is an unfrozen area. This is the frozen area. This is the interface of the ice front.

[0114] Ice front level set function The definition applies to two-dimensional or three-dimensional solution domains, where the spatial position vector... When it is a two-dimensional coordinate system, that is This corresponds to two-dimensional frozen interface tracking; when the spatial position vector When it is a three-dimensional coordinate system, that is This corresponds to three-dimensional frozen interface tracking. During the ice front direction calculation phase, the spatial gradient of the level set function is calculated for the control volume near the ice front interface. In two dimensions, The ice front interface is a curve, in three dimensions. The ice front interface is a curved surface. The unit normal vector of the ice front is calculated using a spatial gradient normalization method. That is, the spatial gradient direction is used as the local normal direction of the frozen interface:

[0115]

[0116] During the calculation of the ice front propulsion velocity, the rate of change of the level set function with respect to time is calculated for the interface control volume at each discrete moment. And combined with spatial gradient magnitude Calculate the normal propulsion velocity of the ice front The normal propulsion speed Take the ratio of the time rate of change of the level set function to the magnitude of the spatial gradient:

[0117]

[0118] The rate of change over time is obtained by the difference between the level set function values ​​of two adjacent time points. and When the data is received, the corresponding interface control body's time change rate is... This calculation yields the local propulsion velocity field along the normal direction for each interface control volume. If the normal propulsion velocity... This indicates that the frozen interface is advancing towards the unfrozen area along the normal direction; if the normal advancement speed... This indicates that the partially frozen interface has regressed or thawed.

[0119] During the local ice saturation calculation stage, the ice volume is read for each control volume. Local porosity and total volume of control body The local ice saturation was calculated based on the ratio of ice volume to local pore volume. :

[0120]

[0121] Among them, the local pore volume is composed of and Multiplication yields the local ice saturation, which reflects the proportion of pore space occupied by the ice phase within the current control volume. When the control volume is located in the membrane region, if this region does not participate in the pore medium saturation calculation, it is skipped according to the preset shielding rules; when the control volume is located in the catalyst layer, microporous layer, gas diffusion layer, and flow channel region, local ice saturation calculation is performed. After completing the calculation for all control volumes, the local ice saturation field at the current moment is formed. .

[0122] During the local freezing rate calculation phase, the local ice phase volume fraction is read for each control volume. The local freezing rate was calculated by performing time difference analysis on the data. :

[0123]

[0124] In numerical implementation, the local freezing rate is determined by two adjacent discrete time intervals. and The difference in ice phase volume fraction is obtained by dividing the time step, i.e. When the local freezing rate When the ice phase volume fraction of the control volume increases; when the local freezing rate increases... When this time, it indicates a decrease in the ice phase volume fraction of the control volume. After completing the calculations for all control volumes, the local freezing rate field at the current moment is formed. .

[0125] During the state variable output phase, the ice front interface will be... unit vector of ice front normal normal advance velocity of the ice front Local ice saturation field and local freezing rate field Store the variables in chronological order and create a set of frozen state variables. Among them, the set of frozen state variables. It includes at least interface position data, interface normal data, interface advancement speed data, control volume ice saturation data, and control volume freezing rate data. A set of frozen state variables is generated for each discrete time step, and the set of frozen state variables for consecutive time steps constitutes the original state trajectory of ice formation and ice expansion during the cold start phase.

[0126] In one implementation, to ensure consistency in subsequent feature extraction, the set of frozen state variables is... Unified coordinate mapping and time alignment are performed. Unified coordinate mapping involves mapping the position coordinates of control volumes in different regions to the same fuel cell reference coordinate system. Time alignment involves resampling the simulation outputs under different operating conditions at the same sampling time interval to obtain a unified time step. , , , and After this processing, the frozen state variables output in step 2.1 can be directly used as input data for the average advance velocity of the ice front, frozen coverage, frozen bias factor, and other frozen characteristics in step 2.2.

[0127] Step 2.2: Based on the frozen state variables, extract key freezing features from the verified multiphysics model solution results, including the average advance velocity of the ice front, frozen coverage, frozen bias factor, oxygen diffusion blockage coefficient, effective reactive area decay rate, thermo-electric coupling instability index, maximum ice saturation in key areas, maximum oxygen diffusion blockage coefficient in key areas, maximum thermo-electric coupling instability index in key areas, and maximum freezing rate in key areas.

[0128] In step 2.2, the set of frozen state variables output in step 2.1 is used as input to extract features of ice generation, ice expansion, local blockage and transmission degradation during the cold start process, and a set of key frozen features is constructed.

[0129] The input variables should include at least the ice front interface at each discrete time point. unit vector of ice front normal normal advance velocity of the ice front Local ice saturation field and local freezing rate field Simultaneously, the local ice phase volume fraction output by the verified multiphysics model is invoked. Local liquid water saturation Local temperature T(x,t), local voltage characterization V(x,t), and local current density Local effective oxygen diffusion coefficient under non-freezing conditions and the local effective oxygen diffusion coefficient under freezing conditions .

[0130] Before feature extraction, the key cold start region is first identified. Key areas This includes the catalyst layer region, the microporous layer region, the gas diffusion layer region near the catalyst layer, and the near-wall flow channel region. Subsequently, the key areas... Divided into upstream areas along the main direction and downstream areas A unified index is established for each control volume within the key region, which is used for subsequent calculations of freeze coverage, freeze bias factor, and extreme value characteristics of the key region. After the key region is divided, at each discrete time step... Each freezing characteristic is calculated separately.

[0131] In the calculation of the average advance velocity of the ice front, the time... Ice front interface Normal propulsion speed of each interface control body By performing area averaging or length averaging, the average advance velocity of the ice front at time t can be obtained. For two-dimensional cross-sections, the average is calculated by integrating over the interface length; for three-dimensional structures, the average is calculated by integrating over the interface area. Average advance velocity of the ice front. for:

[0132]

[0133] In the formula, For time t, the area or length of the ice front interface is measured.

[0134] In numerical implementation, the local normal propagation velocity of each interface unit belonging to the ice front interface is read one by one. The velocity is then weighted according to the length or area of ​​each interface unit, summed, and divided by the total length or area of ​​the interface to obtain the time series of the average ice front propagation velocity. This feature is used to characterize the velocity change process of the frozen interface as a whole advancing towards the unfrozen region.

[0135] During the frozen coverage calculation phase, an ice phase identification threshold is pre-set. For key areas Ice phase volume fraction of each control volume Threshold determination is performed. When the ice phase volume fraction of a certain control volume... Ice phase identification threshold When the ice phase volume fraction is... Ice phase identification threshold At time t, this control volume is recorded as an unfrozen control volume. Then, the ratio of the sum of the volumes of the frozen control volumes within the critical region to the total volume of the critical region is calculated to obtain the freeze coverage rate at time t. After completing the calculations at each time point, a time series of frozen area coverage is generated. This feature is used to describe the changing process of the proportion of frozen area in the key region. Frozen Area Coverage for:

[0136]

[0137] In the formula, This is a critical area for cold starts; This represents the volume of the region. For the Heaviside function; The threshold for ice phase identification; This indicates the proportion of the key area that is frozen.

[0138] During the freeze bias factor calculation stage, for the upstream region and downstream areas Calculate the integral value of ice phase volume fraction separately, and take the key region as an example. The integral value of the internal ice phase volume fraction is used as a normalization term to obtain the freezing bias factor at time t. (t):

[0139]

[0140] In the formula, and These are the upstream and downstream regions, respectively, defined along the main flow direction. This quantity characterizes the degree of spatial offset of freezing along the flow direction.

[0141] In numerical implementation, first, respectively... and The ice phase volume fraction of each control volume is multiplied by the volume of the control volume and then summed. The difference between the upstream and downstream integral values ​​is then calculated and finally divided by the overall integral value of the ice phase volume fraction within the critical region. If the freezing bias factor is greater than zero, it indicates that freezing is more concentrated in the upstream region; if the freezing bias factor is less than zero, it indicates that freezing is more concentrated in the downstream region; if the freezing bias factor is close to zero, it indicates that freezing is relatively uniformly distributed along the mainstream direction. This process can transform the spatial distribution of the ice phase into a bias along the mainstream direction.

[0142] In the oxygen diffusion blocking coefficient calculation stage, key areas Each control volume reads the local effective oxygen diffusion coefficient under non-freezing conditions. Local effective oxygen diffusion coefficient under freezing conditions The local oxygen diffusion blockage coefficient is calculated based on the relative difference between the two. The local effective oxygen diffusion coefficient under freezing conditions is also related to the local porosity. Local ice saturation and local liquid water saturation This yields the local oxygen transport attenuation under conditions of coexistence of freezing and liquid water. After completing the calculations for the entire control volume, an oxygen diffusion blockage coefficient field is formed within the critical region. :

[0143]

[0144] In the formula, The effective oxygen diffusion coefficient under unfrozen conditions. The effective oxygen diffusion coefficient under freezing conditions.

[0145] Among them, the local effective oxygen diffusion coefficient under freezing conditions for:

[0146]

[0147] In the formula, The intrinsic diffusion coefficient of oxygen. Let be the local liquid water saturation, and m and n be the structural correction exponents. Clearly, The larger the value, the more severe the obstruction of oxygen transport.

[0148] Subsequently, the average and maximum values ​​of the oxygen diffusion blocking coefficient field in the key region were extracted, with the maximum value denoted as . It is used to characterize the location and intensity of the most severe oxygen transport obstruction in critical regions.

[0149] In the calculation of the effective reactive surface area decay rate, first in the catalyst layer region... Internally defined local activity reachability function This function is jointly determined by the local oxygen supply state, ion conduction state, and freezing state; then, by integrating the local activity reachability function of each control body within the catalyst layer region, the effective reactive area at time t is obtained. Then compare it with the reference initial active area. Compare and calculate the effective reactive surface area decay rate. :

[0150]

[0151] In the formula, The effective active area at time t. This serves as a reference for the initial active area.

[0152] The effective active area at time t is:

[0153]

[0154] In the formula, For the catalytic layer region, It is a local activity reachability function used to characterize the effective reactivity under the combined effects of local oxygen supply, ion conduction, and frozen state.

[0155] In numerical implementation, the local activity reachability function value can be calculated one by one within the control volume of the catalyst layer, multiplied by the control volume volume or control volume area, and then integrated. After completing the calculations at each time point, a time series of the effective reactive area decay rate is obtained. This feature is used to describe the degree to which freezing weakens the effective reactive region of the catalyst layer.

[0156] In the calculation stage of thermo-electric coupling instability index, the key area Each control volume calculates its local temperature gradient separately. Local voltage gradient and local current density gradient Then according to the preset weight , and The local thermo-electric coupling instability index is obtained by weighted summation of the three gradients. Temperature gradient, voltage gradient, and current density gradient were all calculated using the same spatial discretization scheme as the multiphysics model; central difference was used at the internal control volume, and one-sided difference was used at the boundary control volume. After completing the calculations for all control volumes, a thermo-electric coupling instability index field was formed. :

[0157]

[0158] In the formula, For local temperature, It is a characterization of local voltage. For local current density, , and These are the weighting coefficients.

[0159] Furthermore, the maximum thermo-electric coupling instability index within the key region was extracted. This feature is used to describe the degree of imbalance in local temperature, potential, and reaction distribution induced by freezing.

[0160] In the key region extreme value feature extraction stage, the key region Extract the maximum ice saturation for each control volume. Maximum oxygen diffusion blocking coefficient Maximum thermo-electric coupling instability index and maximum freezing rate In numerical implementation, the corresponding variables of each control volume within the key region are compared point by point, the maximum value at that moment is selected, and stored in chronological order as an extreme value feature time series. Among them, the maximum ice saturation is used to mark the local location with the most severe freezing in the key region, the maximum oxygen diffusion blocking coefficient is used to mark the local location with the strongest oxygen transport obstruction, the maximum thermo-electric coupling instability index is used to mark the region with the most significant deviation in local thermoelectric distribution, and the maximum freezing rate is used to mark the local location with the fastest freezing growth.

[0161]

[0162] In the feature output stage, the average advance velocity of the ice front is... Freeze coverage Freeze bias factor (t), oxygen diffusion blocking coefficient field Effective reactive surface area decay rate Thermo-electric coupling instability index field Maximum ice saturation Maximum oxygen diffusion blocking coefficient Maximum thermo-electric coupling instability index and maximum freezing rate Stored according to a unified time coordinate to form a set of key frozen features. Key frozen feature set This serves as the input for constructing the startup reachability margin and failure boundary function in step 2.3, and also as the basis for constructing the frozen feature vector in step 2.4.

[0163] In one embodiment, to ensure the comparability of extraction results under different operating conditions, after all frozen features have been calculated, each feature is resampled at a uniform sampling time interval, and the same region partitioning rules and numerical integration format are used for features related to spatial integration. After this processing, the features output in step 2.2 can directly enter the subsequent processes of accessibility evaluation, failure boundary determination, and sample construction.

[0164] Step 2.3: Based on the key freezing characteristics obtained in Step 2.2, a start-up accessibility evaluation model and a failure boundary determination model are constructed. The start-up accessibility evaluation model uses the difference between the local temperature and the freezing reference temperature in the key area, the ratio of the local current density to the reference current density, the local ice saturation, the oxygen diffusion blocking coefficient, and the effective reactive area decay rate to perform weighted integration to obtain the start-up accessibility margin. The failure boundary determination model uses the start-up accessibility margin, the maximum ice saturation in the key area, the maximum oxygen diffusion blocking coefficient in the key area, and the maximum thermo-electric coupling instability index in the key area to construct the failure boundary function, and performs boundary determination on the cold start process according to the sign and duration of the failure boundary function.

[0165] In step 2.3, using the key freezing feature set output from step 2.2 as input, a start-up reachability evaluation model and a failure boundary determination model are constructed. Input variables include at least the local ice saturation within the key region. Local oxygen diffusion blocking coefficient Effective reactive surface area decay rate Local temperature T(x,t), local current density Maximum ice saturation in key areas Maximum oxygen diffusion blocking coefficient in key areas Maximum thermo-electric coupling instability index in key areas and key areas Simultaneously preset the freezing reference temperature. Reference temperature rise scale Reference current density The calculation order for step 2.3 is as follows: first calculate the start-up reachability margin, then calculate the start-up reachability margin. (Note: The original text contains some formatting errors and inconsistencies. A more accurate translation would require the full context.) Based on Calculate the failure boundary function with the extreme value characteristics of the key region. .

[0166] In the initial stage of building the accessibility assessment model, the key areas are first... The local temperature T(x,t) and local current density are read point by point from each control unit. Local ice saturation Local oxygen diffusion blocking coefficient and the effective reactive surface area decay rate Among them, key areas Consistent with step 2.2, this includes the catalyst layer region, the microporous layer region, the gas diffusion layer region near the catalyst layer, and the near-wall flow channel region. For each control volume, the difference between the local temperature and the freezing reference temperature, the normalized term between the local current density and the reference current density, the local ice saturation term, the local oxygen diffusion blockage term, and the active area decay term are calculated and linearly combined according to preset weighting coefficients to form the local contribution value of the control volume's start-up reachability at the current moment.

[0167] After the local contribution value is calculated, the key areas are... Spatial integration of the local contributions of all control volumes yields the start-up reachability margin at time t. :

[0168]

[0169] In the formula, This is the reference temperature for freezing. For reference temperature rise scale, For reference current density, ~ This is a weighting coefficient. This index comprehensively characterizes the system's current heating capacity, responsiveness, and degree of freezing and blockage. The larger the value, the more capable the system is of navigating the frozen sensitive area and maintaining a successful startup.

[0170] In one embodiment, to ensure the comparability of start-up accessibility margins under different operating conditions, the temperature and current terms are normalized using a reference scale. Specifically, the temperature term is calculated by dividing the difference between the local temperature and the freezing reference temperature by a reference temperature rise scale. The current term is represented in the form of local current density and reference current density. The ratios are expressed as follows: local ice saturation, local oxygen diffusion blockage coefficient, and active area decay rate are directly used in the weighted integral in their original dimensionless form. Through this process, the start-up accessibility margin is calculated using a uniform scale under different ambient temperatures, different residual water states, and different start-up paths.

[0171] In the failure boundary determination model construction stage, the maximum ice saturation of the key area output in step 2.2 is read first. Maximum oxygen diffusion blocking coefficient in key areas and the maximum thermo-electric coupling instability index in the key area Combined with the current start-up reachability margin Failure boundary functions are constructed according to preset weight coefficients α1 to α4. In numerical implementation, calculations are performed hourly for each discrete moment:

[0172]

[0173] This yields the time series of the failure boundary function. This function couples the local worst-case freeze state with the overall startup capability into a single criterion.

[0174] During the failure boundary determination stage, the values ​​calculated at each discrete time point are... Perform sign discrimination. If If the current cold start state is recorded as the recoverable start domain, then... Then continue to determine whether the state continues for a preset number of time steps. .when and continue for no less than At each discrete time step, the current operating condition is determined to have entered the cold start failure domain; if during continuous counting... If the value is greater than zero again, the count will be reset to zero and the system will continue to track subsequent time points. This process avoids misjudgments caused by instantaneous fluctuations.

[0175] In the model output phase, the start-up reachability margin at each discrete time step is calculated. Failure Boundary Function The corresponding status judgment markers are stored in a unified time frame to form a startup status evaluation result set. Startup status evaluation result set It should include at least the time series of reachability margin, the time series of failure boundary function, the recoverable start-up domain determination marker, and the failure domain determination marker. This result set serves as the input for constructing the frozen feature vector in step 2.4, and as the basis for risk assessment of the ice front evolution prediction model and control model in the subsequent step 3.

[0176] In one embodiment, to improve the consistency of the evaluation model under multi-condition simulation, the weight coefficients ω1 to ω5 and α1 to α4 are calculated in batches using a fixed parameter set, and the same freezing reference temperature is used for all conditions. Reference temperature rise scale Reference current density and failure duration steps After this processing, the start-up reachability margin and failure boundary function obtained under different operating conditions can be directly used for subsequent sample labeling and state comparison.

[0177] Step 2.4: Based on the key frozen features obtained in Step 2.2 and the start-up reachability margin obtained in Step 2.3, construct a frozen feature vector and a multi-condition simulation parameter vector p. Based on the parameter vector p, call the verified multiphysics model to automatically solve multiple conditions in batches. Combine the frozen state variable extraction, key frozen feature extraction, start-up reachability evaluation and failure boundary determination processes from Steps 2.1 to 2.4 to generate frozen feature trajectories and cold start sample labels for each group of conditions. Construct a cold start sample dataset D based on the results of the automatic batch solution for multiple conditions.

[0178] In step 2.4, using the key freezing feature set obtained in step 2.2 and the start-up state evaluation result set obtained in step 2.3 as inputs, a freezing feature vector oriented towards the evolution of the ice front is constructed, and a multi-condition integrated simulation dataset is established based on this vector. The input variables include at least the average advance velocity of the ice front. Freeze coverage Freeze bias factor (t), Startup reachability margin Maximum ice saturation in key areas Maximum oxygen diffusion blocking coefficient in key areas Maximum thermo-electric coupling instability index in key areas Effective reactive surface area decay rate and the maximum freezing rate in key areas The processing order in step 2.4 is as follows: first construct the frozen feature vector. Then, the working condition parameter vector p is constructed, followed by automatic batch solution of multiple working conditions, and finally a cold start sample set is formed and standardized processing is performed.

[0179] In the frozen feature vector construction stage, the average advance velocity of the ice front, frozen coverage, frozen bias factor, start-up reachability margin, maximum ice saturation in key areas, maximum oxygen diffusion blockage coefficient in key areas, maximum thermo-electric coupling instability index in key areas, effective reactive area decay rate, and maximum freezing rate in key areas at each time point are concatenated into a temporal feature vector according to a unified order. This establishes a set of state representations oriented towards the evolution of the ice front, and the frozen feature vector is constructed. This serves as the state input variable for subsequent ice front evolution prediction models and cold start control models. (Frozen feature vector) for:

[0180]

[0181] In numerical implementation, a corresponding set is generated for each discrete time t. The data are stored as frozen feature trajectories in chronological order. This process compresses the high-dimensional field variables of the multiphysics output into low-dimensional state representations of a fixed dimension.

[0182] In the stage of constructing the operating condition parameter vector, the boundary conditions and control parameters that affect the low-temperature cold start freezing evolution are selected as input variables to construct the operating condition parameter vector p:

[0183]

[0184] in, For ambient temperature, This represents the initial water content of the membrane. The residual water level is at the shutdown level. This is the excess air coefficient at the cathode. The excess hydrogen coefficient at the anode. For cathode back pressure, This is the anode back pressure. To assist in heating power, For the startup current ramp-up rate, These are the purging trigger parameters. In numerical implementation, each simulation condition corresponds to a defined set of condition parameter vectors. , where j represents the operating condition number.

[0185] During the operating condition matrix generation stage, the value range and discrete level of each parameter in the operating condition parameter vector are set, and multiple sets of operating conditions to be solved are generated through parameter combination. Ambient temperature is set according to multiple low-temperature levels; initial membrane water content is set according to multiple hydration levels; shutdown residual water level is set according to multiple shutdown residual water levels; cathode air excess coefficient, anode hydrogen excess coefficient, cathode back pressure, anode back pressure, auxiliary heating power, and start-up current ramp-up rate are each combined according to preset discrete values; purge trigger parameters are set according to different trigger times or trigger thresholds. After completing the parameter combination, a multi-operating condition simulation matrix is ​​formed for subsequent automatic batch solving.

[0186] In the multi-condition automatic batch solution stage, each set of condition parameter vectors is read sequentially. The results are then written into the verified multiphysics model, and transient cold start simulation is automatically executed. After each set of operating conditions is solved, steps 2.1 to 2.3 are invoked to sequentially extract frozen state variables, extract key frozen features, and calculate start-up reachability and failure boundaries, ultimately obtaining the frozen feature trajectory corresponding to that set of operating conditions. Through batch cyclic calculations, a multi-condition simulation result set is generated, covering different ambient temperatures, different residual water states, different gas supply parameters, different back pressure parameters, different auxiliary heating parameters, and different start-up paths.

[0187] During the sample label generation stage, sample labels are further extracted from the solution results of each set of working conditions. The sample labels include at least one of the following: startup success label, startup failure label, startup time, maximum freeze risk, and minimum voltage. During numerical implementation, labels can be generated as follows: if the startup success criterion is met within the preset cold start time, the startup success label is recorded as 1, and the startup failure label as 0; if the startup success criterion is not met, the startup success label is recorded as 0, and the startup failure label as 1; the startup time is the moment corresponding to reaching the target operating state; the maximum freeze risk is the maximum value of the combined risk of key freeze features throughout the simulation; the minimum voltage is the lowest output voltage of the fuel cell stack or the lowest single-cell voltage during the entire cold start process. After processing, each set of operating conditions corresponds to a set of operating condition parameters, a freeze feature trajectory, and a set of sample labels.

[0188] During the cold start sample set construction phase, the operating condition parameter vectors will be... Freezing feature trajectory under corresponding working conditions and sample labels The samples are stored in association according to their working condition numbers to form a cold start sample set D:

[0189]

[0190] Where N is the total number of samples. Each item in the sample set retains the correspondence between operating parameters and freezing evolution trajectory, which can be used for training subsequent ice front evolution prediction and control models.

[0191] During the standardization phase, dimensionless standardization is performed on continuous variables in the operating parameters, frozen feature trajectories, and sample labels of the cold start sample set. For any original variable x, the minimum value of that variable is read from all samples. and maximum value and in accordance with:

[0192]

[0193] Calculate the standardized results. Discrete label variables retain their original encoding form. After standardization, frozen evolution data and cold start sample data are obtained, which are used for training the ice front evolution prediction model and constructing the control model in subsequent step 3.

[0194] In one embodiment, to ensure the sample set covers different freezing evolution paths, the ambient temperature, residual water level, gas supply parameters, back pressure parameters, auxiliary heating parameters, and start-up current ramp-up rate are sampled using a full combination method or stratified sampling method when generating the operating condition matrix. This ensures that the sample set simultaneously includes successful start-up conditions, critical start-up conditions, and failed start-up conditions. After this processing, the dataset output in step 2.4 retains both the operating condition input information and the freezing evolution state trajectory and result label information, and can be directly used as the data foundation for subsequent prediction and control models.

[0195] Step 3: Based on the freezing evolution data and cold start sample data, establish an ice front evolution prediction model and an ice front risk constraint two-layer distributed soft actor-commentator model. Use the ice front evolution prediction model to predict the future freezing expansion trend and local blockage risk, and jointly adjust the start-up current, gas supply, back pressure, purging and auxiliary heating based on the prediction results. During the cold start operation, update the ice front evolution prediction model and the ice front risk constraint two-layer distributed soft actor-commentator model online according to the real-time operation data.

[0196] Step 3 includes:

[0197] Step 3.1: Based on the current operating status, frozen feature vector, and initial operating parameters, construct an ice front evolution prediction model, and perform rolling prediction of the freezing features for the next H steps based on the historical input sequence of length L, to obtain the prediction results of the future freezing features and the probability of successful start-up.

[0198] In step 3.1, after generating the freezing evolution data and cold start sample data in step 2, an ice front evolution prediction model is constructed based on the current operating state, freezing feature vector, and initial operating parameters. This model is used to predict the freezing evolution trend over several future control steps. The prediction object here is not a single temperature or voltage quantity, but rather the freezing feature vector constructed in step 2. The freezing feature vector consists of the average advance velocity of the ice front, freezing coverage, freezing bias factor, initiation reachability margin, maximum ice saturation, maximum oxygen diffusion blockage coefficient, maximum thermo-electric coupling instability index, active area decay rate, and maximum freezing rate. Therefore, the output of the ice front evolution prediction model simultaneously covers three dimensions: freezing expansion, transport degradation, and initiation reachability.

[0199] In this embodiment, the input vector at time t is first defined. Input vector It consists of the current running state variables, the current frozen feature vector, and the initial operating parameters:

[0200]

[0201] In the formula: The average reactor temperature; It is an indicator of voltage non-uniformity; and These are the cathode and anode pressures, respectively. and These are the excess air coefficients at the cathode and the excess hydrogen coefficients at the anode, respectively. For the starting current, To supplement heating power; Freeze the feature vector; The initial operating parameters consist of ambient temperature, initial membrane water content, and residual water level.

[0202] To enable the prediction model to utilize the temporal evolution information during the cold start process, instead of directly using single-time-point inputs, a historical input sequence of length L is constructed. Specifically, at the current prediction time t, input vectors from time t−L+1 to t (a total of L discrete times) are sequentially read from the historical database or real-time cache to form the historical input sequence. This historical input sequence preserves the time-varying trajectories of temperature rise, voltage dispersion, gas supply status, back pressure status, heating status, and freezing characteristics during the cold start process, serving as the temporal input for the ice front evolution prediction model. If the original sampling time step for a certain operating condition differs from the model's set time step, resampling is performed at a uniform time interval before constructing the historical input sequence.

[0203] In the model building phase, a mapping function for predicting ice front evolution is established. Mapping function Taking a historical input sequence of length L as input, and outputting the predicted frozen features within the next H control steps, the prediction result is denoted as... Where k = 1, 2, ..., H. Correspondingly, the future H-step frozen feature prediction process can be expressed as:

[0204]

[0205] In practical implementation, the mapping function This can be achieved using a prediction network trained on time-series samples, while maintaining consistency between the input dimension, output dimension, and frozen feature vector dimension. The model training samples come from the cold-start sample set constructed in step 2, where each set of samples contains operating parameters, frozen feature trajectories, and start-up results, which can be directly used to establish a mapping relationship from historical states to future frozen features.

[0206] During the prediction execution phase, for the current time t, the historical input sequence is first fed into the ice front evolution prediction model to obtain the predicted freezing characteristics for the next H steps. Then, the latest observations are added to the input buffer, and a new historical input sequence of length L is constructed at the next control time. This prediction process is repeated, forming a rolling prediction mechanism. Through rolling prediction, the model can continuously output the changing trends of future freezing front advancement, freezing cover expansion, oxygen diffusion obstruction, and thermo-electric coupling instability throughout the entire cold start process, without relying solely on single-step judgments based on current measurements.

[0207] In one implementation, the ice front evolution prediction model outputs not only future freezing characteristic trajectories, but also maps them to a probability of successful initiation. The specific approach is as follows: read the start-up reachability margin corresponding to the predicted time-domain endpoint t+H. Maximum ice saturation Maximum oxygen diffusion blocking coefficient and maximum thermo-electric coupling instability index and according to the preset weighting coefficient to After combining the results, input the Sigmoid mapping function to obtain the probability of successful startup at the current time. :

[0208]

[0209] Among them, the start-up reachability margin is taken as a positive input, while the maximum ice saturation, maximum oxygen diffusion blockage coefficient and maximum thermo-electric coupling instability index are taken as negative inputs, so as to ensure that the probability of successful start-up decreases accordingly when freezing is enhanced, transmission is blocked and thermo-electric instability increases.

[0210] To improve the model's applicability under different cold-start conditions, the multi-condition cold-start sample set formed in step 2 is used during the training phase. Specifically, the historical input sequence and future frozen feature trajectory are extracted for each sample group according to the condition number. The historical input sequence is used as the training input, and the future H-step frozen feature trajectory is used as the supervised output. The prediction model parameters are updated through iterative optimization. The initial condition parameters in the input vector are maintained during training. The operating conditions are consistent with the corresponding freezing characteristic trajectory, thus enabling the model to distinguish freezing evolution paths under different ambient temperatures, different initial membrane water content, and different residual water levels.

[0211] During the model output phase, a prediction result set is generated for each control time step. The prediction result set must include at least the frozen feature prediction sequence for the next H steps. and the probability of successful startup at the current moment. In this process, the predicted sequence of future frozen features is used as a component of the reinforcement learning state vector in step 3.2, and the success probability of initiation is used as the input to the immediate reward function of the subsequent control model. Through this processing, step 3.1 further transforms the frozen evolution state representation constructed offline in step 2 into feedforward prediction information for control decision-making.

[0212] Step 3.2: Based on the future freezing feature prediction results obtained in Step 3.1, a two-layer distributed soft actor-commentator control model with ice front risk constraints is constructed. The current observation input vector, the future freezing feature prediction results, the cold start failure boundary function, and the action executed in the previous time step are combined to form the reinforcement learning state vector. The upper-layer policy network outputs the stage pattern, and the lower-layer policy network outputs the original continuous action vector. A comprehensive value function and an immediate reward function are constructed based on the start-up reward commentator, the freezing risk commentator, and the damage risk commentator.

[0213] The phased modes include at least the conservative preheating mode, the controlled heat generation mode, the ice suppression crossing mode, and the stable climbing mode; the original continuous action vector includes at least the starting current correction, the cathode air excess coefficient correction, the anode hydrogen excess coefficient correction, the cathode back pressure correction, the anode back pressure correction, the purging trigger control, and the auxiliary heating power correction.

[0214] In step 3.2, after obtaining the predicted freezing features within several future control steps in step 3.1, a two-layer distributed soft actor-critic control model with ice front risk constraints, denoted as IRCB-DSAC, is constructed for the low-temperature cold start scenario. In IRCB, I represents the ice front, R represents risk, C represents constrained constraints, B represents bilevel, and DSAC stands for Distributed Soft Actor-Critic. DSAC emphasizes modeling the reward distribution and is commonly used in risk-sensitive reinforcement learning. This model is used to generate multivariate collaborative control actions based on the current operating state, future freezing risk, and stage pattern. The control model employs a two-layer policy structure and a distributed multi-critic structure: the two-layer policy structure separates the stage switching decisions and continuous control quantity generation during the cold start process, while the distributed multi-critic structure evaluates the start-up benefits, freezing risk, and potential damage risk separately, thus forming a control decision framework adapted to the low-temperature cold start process.

[0215] In this embodiment, the system state vector at reinforcement learning time t is first defined. System state vector It consists of four parts: the current observation input vector Future frozen feature prediction sequence Cold start failure boundary function and the action performed at the previous moment Among them, the current observation input vector The input vector should be consistent with that of the ice front evolution prediction model in step 3.1, and should include at least the average reactor temperature, voltage inhomogeneity index, cathode pressure, anode pressure, cathode air excess coefficient, anode hydrogen excess coefficient, start-up current, auxiliary heating power, freezing feature vector, and initial operating parameters; the future freezing feature prediction sequence should use the output of step 3.1. The action performed at the previous moment This represents the actual execution action obtained after projecting the safety constraints in step 3.3. Therefore, the system state vector... for:

[0216]

[0217] This state vector further incorporates future freeze risk information and safety boundary information on the basis of the current measurement, so that the subsequent policy output does not depend on the feedback quantity at a single moment.

[0218] In the phased pattern decision-making phase, the system state vector is... Input upper-layer policy network The discrete-stage pattern is output by the upper-level policy network. :

[0219]

[0220] In the formula, This is the upper-level policy mapping function. These are discrete-stage pattern variables.

[0221] In this embodiment, based on the characteristics of the low-temperature cold start process, the stage modes are divided into four categories: conservative preheating mode, controlled heat generation mode, ice suppression and crossing mode, and stable ramp-up mode. The conservative preheating mode corresponds to the initial stage of cold start where heat accumulation is insufficient and the risk of freezing is high; the controlled heat generation mode corresponds to the stage where reaction exotherm is established through controlled loading; the ice suppression and crossing mode corresponds to the stage where freezing expansion is significant and ice blockage and oxygen diffusion blockage need to be prioritized; and the stable ramp-up mode corresponds to the later stage of cold start where temperature and reaction state gradually stabilize and output capacity continuously increases. The upper-level strategy network switches between each stage mode based on the predicted trend of future freezing characteristics, the current failure boundary level, and the actions performed at the previous moment.

[0222] During the continuous action generation phase, the system state vector is... and phase mode Common input lower-level policy network The lower-level policy network outputs the original continuous action vector. :

[0223]

[0224] The original continuous action vector includes at least: the starting current correction amount. Correction amount for excess cathode air coefficient Correction amount for excess hydrogen at the anode Cathode back pressure correction amount Anode back pressure correction amount , purge trigger control quantity And auxiliary heating power correction amount Therefore, the original continuous action vector is:

[0225]

[0226] This action vector covers five types of actuators: electronic load, gas supply, back pressure, purging, and auxiliary heating, so that the control output is no longer limited to a single current regulation or a single heating regulation.

[0227] During the commentator structure building phase, separate start-up revenue commentators were set up. Freezing risk commentators and damage risk commentators :

[0228]

[0229] All three types of commentators used the system state vector. Actual actions performed and phased mode The inputs are the same, but the output objectives differ: the startup benefit commentator assesses the contribution of the current action to the establishment of temperature rise and maintenance of startup capability; the freeze risk commentator assesses the impact of the current action on future freeze propagation and oxygen diffusion blockage; and the damage risk commentator assesses the impact of the current action on thermo-electric coupling instability and structural stress. To unify the outputs of the three commentators, a comprehensive value function is further constructed in this embodiment. :

[0230]

[0231] In the formula, and These are the penalty weights for freezing risk and damage risk, respectively. This comprehensive value function embodies the core idea of ​​this application, namely: cryogenic cold start control does not simply pursue rapid start-up, but rather achieves a dynamic balance between start-up benefits, freezing risk, and structural safety.

[0232] In the immediate reward function construction phase, to enable the control model to directly utilize the future freeze prediction information output in step 3.1 during training, an immediate reward function is constructed consisting of a startup benefit term, a freeze risk term, a damage cost term, and an auxiliary energy consumption term. :

[0233]

[0234] The startup benefit item is defined as follows:

[0235]

[0236] In the formula, To initiate the accessibility margin increment; This represents the average reactor temperature increment. To predict the probability of a successful startup; ~ This is the profit weighting coefficient.

[0237] By using an instant reward function, the average amount of frozen risk in the future prediction time domain will be directly included in the reward calculation during policy training.

[0238] The definition of a freeze risk item is:

[0239]

[0240] In the formula,

[0241] These represent the average predicted values ​​of maximum ice saturation, maximum oxygen diffusion blocking coefficient, and freeze coverage within the future prediction time domain, respectively. ~ This is for freezing risk weighting coefficients.

[0242] The damage cost term is defined as:

[0243]

[0244] In the formula, This represents the average value of the maximum thermo-electric coupling instability index in the future prediction time domain; This is an indicator of current voltage non-uniformity. The equivalent thermomechanical stress index of the membrane electrode; ~ This represents the damage penalty weighting coefficient. This process integrates the local thermal imbalance, electrical imbalance, and structural stress risk induced by freezing into the strategy training process.

[0245] The auxiliary energy consumption item is defined as:

[0246]

[0247] in, and These are the auxiliary heating energy consumption weight and the purging operation cost weight, respectively.

[0248] During the policy training phase, the lower-level policy network is optimized using a soft actor-critic framework. Specifically, the state vector, stage pattern, action, and reward are written into the experience replay pool, and then optimized based on the comprehensive value function. and instant reward function The parameters of the actor network and the three types of critic networks are iteratively updated. The optimization objective for the actor network is expressed as the difference between the entropy regularization term and the comprehensive value term, ensuring that the strategy retains some exploratory capability while pursuing a high comprehensive value. The optimization objective for the actor network is:

[0249]

[0250] In the formula, For experience replay pool, Let be the entropy temperature coefficient. This objective function indicates that the policy network maintains a moderate exploratory capability while pursuing high overall value, in order to improve its adaptability to complex low-temperature operating conditions.

[0251] The commentator network loss function is constructed using the deviation between the target value and the current commentator's output, allowing the three types of commentators to approximate the target functions corresponding to the initial gain, freeze risk, and damage risk, respectively. The loss function for each commentator network is uniformly expressed as:

[0252]

[0253] in, For the first Target value for critics:

[0254]

[0255] In the formula, As a discount factor, For the corresponding target critic network.

[0256] This completes the construction of a two-layer distributed reinforcement learning control model for low-temperature cold start scenarios. The IRCB-DSAC algorithm, through its overall structure of upper-layer stage pattern scheduling, lower-layer continuous action optimization, distributed risk commentators, and future freeze risk feedforward rewards, achieves safe, robust, and adaptive cooperative control of the low-temperature cold start process. In the model output phase, step 3.2 outputs the stage pattern at the current control moment. Original continuous action vector Three types of critic output values ​​and the comprehensive value function Among them, the original continuous action vector As input to the safety constraint projection module in step 3.3, the phase mode This also serves as a conditional input for determining the feasible domain of safe actions in step 3.3. In this way, a clear data transfer relationship is established between steps 3.2 and 3.3.

[0257] Step 3.3: Project the original continuous motion vector output in Step 3.2 with safety constraints to obtain the final execution action that satisfies the physical constraints of the equipment, the operating boundary constraints, and the future freeze risk constraints. Then, coordinate the control of the electronic load, the gas supply unit, the back pressure unit, the purging unit, and the auxiliary heating unit based on the final execution action.

[0258] In step 3.2, output the original continuous motion vector. Subsequently, instead of directly sending the original continuous motion vector to the electronic load, air supply unit, back pressure unit, purging unit, and auxiliary heating unit for execution, it first enters the safety constraint projection module. The safety constraint projection module operates in the current stage mode. Original continuous action vector Using the current operating status, device allowable boundaries, and the future freeze risk prediction results output in step 3.1 as inputs, construct the feasible domain of safe actions corresponding to the current stage mode. Within the feasible region of this safe action, the final action that minimizes the distance to the original continuous action vector is obtained. :

[0259]

[0260] In the formula, In order to match the current stage model The corresponding feasible domain for safe actions. This feasible domain simultaneously considers equipment physical constraints, operational boundary constraints, and freeze risk constraints. The main constraints include current change rate limits, upper and lower limits for air and hydrogen excess coefficients, upper and lower limits for anode and cathode back pressures, upper limit for auxiliary heating power, and future failure boundary margin constraints.

[0261] In this embodiment, the original continuous action vector The original continuous action vector, output from the lower-level policy network in step 3.2, includes at least: the starting current correction. Correction amount for excess cathode air coefficient Correction amount for excess hydrogen at the anode Cathode back pressure correction amount Anode back pressure correction amount , purge trigger control quantity And auxiliary heating power correction amount After receiving the original continuous motion vector, the safety constraint projection module first determines the current stage mode. Select the corresponding set of constraint parameters. The permissible range of action variation differs depending on the stage mode: in conservative preheating mode, the current correction range is small and the upper limit of auxiliary heating is high; in controlled heat generation mode, the current and gas supply can be gradually increased; in ice suppression crossing mode, freezing-related risk constraints are tightened first; in stable ramp-up mode, the output level can be increased while ensuring safety margins are met. Therefore, the feasible domain of safe actions is not a fixed set, but rather a set of parameterized constraints that dynamically switches with the stage mode.

[0262] During the safe action feasible domain construction phase, equipment physical constraints are first established. These constraints include at least: limits on the rate of change of starting current, the range of variation for the cathode air excess coefficient, the range of variation for the anode hydrogen excess coefficient, the range of variation for the cathode back pressure, the range of variation for the anode back pressure, the range of values ​​for the purging trigger quantity, and the upper limit for auxiliary heating power. Specifically, maximum rising and falling step sizes are set for the starting current correction quantity to ensure that the current change within adjacent control cycles does not exceed the allowable response range of the electronic load; upper and lower limits are set for the cathode air excess coefficient and the anode hydrogen excess coefficient to ensure that the gas supply command does not exceed the adjustable range of the gas supply unit; upper and lower limits are set for the cathode back pressure and the anode back pressure to ensure that the back pressure command does not exceed the allowable operating range of the back pressure valve; a maximum allowable power is set for the auxiliary heating power to ensure that the heating unit output does not exceed the rated power; and binary or interval constraints are set for the purging trigger control quantity to ensure that the purging command is consistent with the execution mode of the purging unit.

[0263] During the operation boundary constraint construction phase, state-related constraints are further established according to the current operating conditions of the fuel cell system. The operation boundary constraints at least include: the maximum allowable current correction amplitude under the current stack temperature condition, the maximum allowable backpressure correction amplitude under the current supply pressure condition, the minimum supply safety margin under the current gas excess coefficient condition, and the maximum allowable additional heating power under the current heating state. When specifically implemented, first read the current average stack temperature 、the current cathode pressure 、the current anode pressure 、the current cathode air excess coefficient 、the current anode hydrogen excess coefficient and the current auxiliary heating power , and then, based on the preset operation boundary table or constraint function, the adjustable range of each action component is further shrunk to make the finally executed action meet the boundary requirements of the device capacity and the current operating state at the same time.

[0264] During the freeze risk constraint construction phase, the predicted failure boundary values within the future k-step prediction time domain output in step 3.1 are introduced into the safe action feasible region. Specifically, set the allowable minimum safety margin , and gradually impose constraints on the prediction results from the 1st step to the Hth step in the future, requiring:

[0265]

[0266] where is further calculated from the predicted freeze characteristic results under the action of the current candidate action. If a certain candidate action causes the predicted failure boundary value at any future step to be lower than the minimum safety margin, then this candidate action does not belong to the safe action feasible region under the current stage mode. Through this processing, the future freeze failure risk directly participates in the action screening, rather than being indirectly reflected only in the policy reward.

[0267] After the constraint projection, each execution variable is updated according to the corresponding correction amount, which is used to drive the coordinated actions of the electronic load, the gas supply unit, the backpressure unit, the purge unit and the auxiliary heating unit. Thus, the present invention realizes the multi-variable linkage adjustment under the condition that the future freeze risk is controlled.

[0268] Step 3.4: Based on the newly added operating condition parameters, freeze characteristic trajectories, action sequences and start-up results during the cold start operation process, an online experience pool is constructed, and combined with the priority replay mechanism based on the failure boundary construction, small-step increment update, target network soft update and policy deviation constraint, the ice front evolution prediction model and the reinforcement learning control model are updated online.

[0269] After each low-temperature cold start process, the operational data generated throughout the entire cold start process is processed, and the ice front evolution prediction model and reinforcement learning control model are updated online based on the processed operational data. This online self-update is not offline retraining detached from the operational process, but rather a gradual correction of the prediction and control model parameters while retaining the current model parameters and incorporating newly acquired operational samples. The input data for the online self-update includes at least the operating parameters, freezing feature trajectory, action sequence, and start-up result corresponding to this cold start. The operating parameters include at least ambient temperature, initial membrane water content, residual water level, gas supply parameters, back pressure parameters, auxiliary heating parameters, and start-up path parameters; the freezing feature trajectory includes the freezing feature vector sequence extracted and predicted in steps 2 and 3; the action sequence includes the original continuous action vectors and the final executed action within each control cycle; and the start-up result includes at least one of the following: start-up success label, start-up failure label, start-up time, minimum voltage, and maximum freezing risk.

[0270] During the online experience pool construction phase, the state sequence, action sequence, reward sequence, predicted freeze feature sequence, and final startup result of this cold start process are first read in the order of the control cycle, and then organized into experience samples and written into the online experience pool. Each experience sample includes at least the current system state vector. Current original continuous action vector Current final action Current instant rewards The system state vector at the next moment Current stage mode Current failure boundary function value The results of this cold start are also labeled. For the ice front evolution prediction model, the correspondence between historical input sequences and the true trajectories of future freezing features is also saved simultaneously, enabling the prediction model to be updated under supervision based on new samples. After processing, the newly added samples from this cold start are added to the online experience pool, forming a continuously expanding set of running samples.

[0271] During the sample priority calculation phase, sampling priorities are assigned to each sample in the online experience pool based on the failure boundary and the intensity of frozen evolution. Specifically, the failure boundary function value corresponding to each sample is first read. The system includes setting the freeze characteristic change rate and startup result markers. For samples with failure boundary function values ​​close to zero, startup failure samples, and samples with large freeze characteristic change amplitudes, their priority is increased. For samples with failure boundary function values ​​far from the failure boundary, gradual freeze changes, and stable startup processes, their priority is decreased. In numerical implementation, a priority scoring function can be used to score samples. The scoring function should include at least a failure boundary distance term, a failure penalty term, and a freeze change amplitude term. After scoring, probability sampling is performed according to sample priority, so that critical samples, failed samples, and rapidly frozen samples are prioritized during online updates. Through this process, newly emerging dangerous and difficult samples can participate in parameter updates at a higher frequency.

[0272] During the update phase of the ice front evolution prediction model, a batch of supervised update samples is extracted from the online experience pool according to priority. Each supervised update sample includes at least a historical input sequence of length L and the corresponding true trajectory of the freezing feature in the next H steps. Based on this batch of samples, the prediction error between the current prediction model output and the true freezing feature trajectory is recalculated, and the parameters of the ice front evolution prediction model are updated incrementally with small steps according to the prediction error. The meaning of small step incremental update is that each update only allows the model parameters to be corrected by a preset small learning rate based on the current parameters, so as to ensure that the model can adapt to new working conditions, new residual water states, and new aging states without causing drastic drift in prediction results due to a single batch of samples. After completing one parameter correction, the updated prediction model is reused for rolling prediction in the subsequent cold start process.

[0273] During the reinforcement learning control model update phase, a batch of reinforcement learning training samples is extracted from the online experience pool according to priority, and the upper-layer policy network, lower-layer policy network, and distributed multi-commentator network are updated based on these samples. For the lower-layer policy network, the soft actor-commentator framework is still used for optimization, that is, the policy output is adjusted according to the comprehensive value function and the immediate reward function. For the starting reward commentator, the frozen risk commentator, and the damage risk commentator, the loss function is calculated based on the deviation between the current commentator output and the target commentator output, and the network parameters are updated accordingly. During the update, the state, action, reward, next state, and stage pattern information in the experience replay pool are used, so that the control model can use the newly acquired operating experience to correct the control policy.

[0274] During the target network soft update phase, a soft update operation is performed on the target network parameters in the reinforcement learning control model. Specifically, the current network parameters are denoted as θ, and the target network parameters are denoted as θ. Let the soft update coefficient be denoted as τ, then the updated target network parameters are calculated according to... The calculation is performed. This update method involves slowly shifting the target network parameters along the direction of the current network parameters, rather than directly replacing them with the current network parameters. This reduces abrupt changes in the critic's target value during online learning, maintaining the continuity of the policy optimization process.

[0275] During the policy offset constraint phase, to prevent abrupt changes in the output of the lower-level policy network due to online updates, the lower-level policy before the update is modified after the parameter update is completed. and updated lower-level strategies Perform a consistency check. Specifically, select a batch of representative state samples and check the consistency of state vectors. and phase mode Simultaneously input the two sets of lower-level policy networks before and after the update, and calculate the corresponding output action difference; if the square norm of the output difference exceeds the preset maximum policy offset threshold... If the error occurs, the update step size is reduced or the current parameter update is rolled back; if the output difference does not exceed the threshold, the update result is accepted. The policy offset constraint is:

[0276]

[0277] In the formula, and These are the lower-level strategies before and after the update. This represents the maximum allowable policy offset. The above design ensures that the control policy remains sufficiently smooth and executable while adapting to new operating conditions and aging states.

[0278] In one implementation, the online update process follows a sequence of "updating the prediction model first, then the control model." Specifically, the ice front evolution prediction model is first updated incrementally with small steps based on new samples. Then, the updated prediction model is used to regenerate the future freezing feature prediction sequence, and the updated prediction results are used as one of the inputs during the training of the reinforcement learning control model. This ensures that the feedforward risk information upon which the control model relies remains consistent with the latest operating state, avoiding policy adjustments based on outdated predictor parameters.

[0279] During the update triggering mechanism phase, either a fixed-period update mode or an event-triggered update mode can be set. The fixed-period update mode performs an online update once after each preset number of cold start tests or operations. The event-triggered update mode triggers an online update immediately when a start-up failure occurs, the minimum voltage falls below a preset threshold, the failure boundary function remains close to zero for an extended period, or the rate of change of the frozen characteristic exceeds a preset threshold. By setting the update triggering mechanism, the model update frequency can be matched to the degree of change in the system's operating state.

[0280] During the online experience pool maintenance phase, to prevent the experience pool from growing indefinitely, a sample retention window and sample elimination rules can be set. Preferably, all samples from the most recent cold start processes are retained, along with a portion of historical failure samples and critical samples. For stable samples that are far from the failure boundary and have high repeatability, the retention ratio can be reduced. After this maintenance, the online experience pool retains both recent samples representative of the current operating conditions and historical samples valuable for risk boundary identification. This maintenance method, used in conjunction with a failure boundary-based priority replay mechanism, can improve the effectiveness of online sample updates.

[0281] During the update output phase, the updated parameters of the ice front evolution prediction model, the updated parameters of the upper-layer policy network, the lower-layer policy network, the critic network, and the target network are written back to the model parameter library. The sample IDs, priority distributions, learning rates, soft update coefficients, and policy offset verification results used in this update round are also recorded. After the write-back is complete, the updated model parameters are directly put into operation during the next cold start process.

[0282] Thus, step 3.4 completes the entire online self-updating process from new sample collection, experience pool construction, priority playback, incremental update of the prediction model, soft update of the control model to policy offset verification, enabling the ice front evolution prediction model in step 3.1 and the reinforcement learning control model in step 3.2 to be continuously corrected as the working conditions change.

[0283] The above descriptions are merely embodiments of this application, and common knowledge regarding specific structures and characteristics in the solutions is not described in detail here. It will be apparent to those skilled in the art that this application is not limited to the details of the above exemplary embodiments, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Therefore, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A self-learning cold-start method for fuel cells based on ice front evolution prediction, characterized in that, The method includes: Step 1: Construct a low-temperature cold start test platform for proton exchange membrane fuel cells, collect cold start test data under different ambient temperatures, different initial water content states and different start-up conditions, and establish a multiphysics model of fuel cell low-temperature cold start based on the test data, fuel cell structural parameters and low-temperature cold start boundary conditions to characterize the coupled processes of gas transport, liquid water migration, heat transfer, electrochemical reaction and freezing phase change, and verify the multiphysics model using the test data; Step 2: Based on the verified multiphysics model, construct a set of frozen state variables for ice formation and ice expansion during the low-temperature cold start process, and extract the characteristics of ice formation, ice expansion, local blockage and transport degradation during the low-temperature cold start process. Establish a set of state representations for ice front evolution, and construct a start-up reachability margin and failure boundary judgment model based on the set of state representations. Combine the ambient temperature, residual water state, gas supply parameters, back pressure parameters, auxiliary heating parameters and start-up path parameters to perform multi-condition simulation and generate frozen evolution data and cold start sample data. Step 3: Based on the freezing evolution data and cold start sample data, establish an ice front evolution prediction model and an ice front risk constraint two-layer distributed soft actor-commentator model. Use the ice front evolution prediction model to predict the future freezing expansion trend and local blockage risk, and jointly adjust the start-up current, gas supply, back pressure, purging and auxiliary heating based on the prediction results. During the cold start operation, update the ice front evolution prediction model and the ice front risk constraint two-layer distributed soft actor-commentator model online according to the real-time operation data.

2. The fuel cell self-learning cold start method based on ice front evolution prediction according to claim 1, characterized in that, Step 2 includes: Step 2.1: Based on the verified multiphysics model, define the ice front horizontal set function, determine the freezing interface position according to the zero value interface of the ice front horizontal set function, determine the ice front normal unit vector according to the spatial gradient of the ice front horizontal set function, determine the ice front normal advance velocity according to the ratio of the rate of change of the ice front horizontal set function with respect to time to the spatial gradient, define the local ice saturation based on the ratio of the ice volume to the local pore volume in the local control body, and define the local freezing rate based on the derivative of the local ice phase volume fraction with respect to time, so as to construct a set of freezing state variables for ice generation and ice expansion during the low temperature cold start process; In step 2.2, the set of frozen state variables output in step 2.1 is used as input to extract features of ice generation, ice expansion, local blockage and transmission degradation during the cold start process, and a set of key frozen features is constructed. Step 2.3: Based on the key freezing characteristics in Step 2.2, construct the start-up reachability margin and failure boundary determination model. The start-up reachability margin is obtained by weighted integration of the ratio of the difference between the local temperature and the freezing reference temperature in the key area to the reference temperature rise scale, the ratio of the local current density to the reference current density, the local ice saturation, the local oxygen diffusion blockage coefficient, and the effective reactive area decay rate. The failure boundary determination model is constructed using the start-up reachability margin, the maximum ice saturation in the key area, the maximum oxygen diffusion blockage coefficient in the key area, and the maximum thermo-electric coupling instability index in the key area. The boundary determination of the cold start process is performed based on the sign and duration of the failure boundary function. Step 2.4: Based on the key freezing feature set in Step 2.2 and the start-up reachability margin in Step 2.3, construct a freezing feature vector. Construct a multi-condition parameter vector according to the boundary conditions and control parameters affecting the low-temperature cold start freezing evolution. Based on the condition parameter vector, call the verified multiphysics model to automatically solve multiple conditions in batches. Combined with Steps 2.1 to 2.3, extract freezing state variables, extract key freezing features, evaluate start-up reachability, and determine failure boundaries to generate freezing feature trajectories and cold start sample labels for each group of conditions. Construct a cold start sample dataset by combining the condition parameter vector, freezing feature trajectory, and cold start sample labels.

3. The fuel cell self-learning cold start method based on ice front evolution prediction according to claim 1, characterized in that, Step 3 includes: Step 3.1: Based on the current operating status, frozen feature vector, and initial operating parameters, construct an ice front evolution prediction model, and perform rolling prediction of the freezing features for the next H steps based on the historical input sequence of length L, to obtain the prediction results of the future freezing features and the probability of successful start-up; Step 3.2: Based on the future freezing feature prediction results obtained in Step 3.1, a two-layer distributed soft actor-commentator control model with ice front risk constraints is constructed. The current observation input vector, the future freezing feature prediction results, the cold start failure boundary function, and the action executed at the previous moment are used to form the system state vector. The system state vector is input into the upper-layer policy network, which outputs the stage pattern. The system state vector and the stage pattern are input into the lower-layer policy network, which outputs the original continuous action vector. A comprehensive value function and an immediate reward function are constructed based on the start-up benefit commentator, the freezing risk commentator, and the damage risk commentator. Step 3.3: Project the original continuous action vector output in Step 3.2 with safety constraints to obtain the final execution action that satisfies the physical constraints of the equipment, the operating boundary constraints, and the future freeze risk constraints. Then, coordinate the control of the electronic load, the gas supply unit, the back pressure unit, the purging unit, and the auxiliary heating unit based on the final execution action. Step 3.4: Based on the newly added operating parameters, frozen feature trajectories, action sequences and startup results during the cold start operation, an online experience pool is constructed. Combined with the priority replay mechanism based on failure boundaries, small step incremental update, target network soft update and policy offset constraints, the ice front evolution prediction model and reinforcement learning control model are updated online.

4. The fuel cell self-learning cold start method based on ice front evolution prediction according to claim 2, characterized in that, In step 2.2, the key frozen feature set Including the average advance speed of the ice front Freeze coverage Freeze bias factor (t), oxygen diffusion blocking coefficient field Effective reactive surface area decay rate Thermo-electric coupling instability index field Maximum ice saturation Maximum oxygen diffusion blocking coefficient Maximum thermo-electric coupling instability index and maximum freezing rate ; Among them, the frozen coverage rate for: ； Freeze bias factor (t) is: ； Effective reactive surface area decay rate for: ； In the formula: t is the discrete time or the current calculation time; This is a critical area for cold starts; For key areas Volume; In key areas Volume division is performed on top; For the Heaviside function; For a moment Spatial location Local ice phase volume fraction at a given location; It is a spatial position vector; in two dimensions it can be... In three dimensions, it can be ; The threshold for ice phase identification; For key areas The upstream region is divided along the mainstream direction; For key areas The downstream region is divided along the main trend direction; This is the integral value of the ice phase volume fraction in the upstream region; This is the integral value of the ice phase volume fraction in the downstream region; This represents the overall integral value of the ice phase volume fraction within the key region; For a moment The effective reactive surface area; For reference, the initial active area; This is the catalytic layer region; It is a local activity reachability function; To obtain the effective reactive area, the local activity reachability function is integrated over the catalytic layer region.

5. The fuel cell self-learning cold start method based on ice front evolution prediction according to claim 4, characterized in that, In step 2.2, the maximum ice saturation for: ； Maximum oxygen diffusion blocking coefficient for: ； Maximum thermo-electric coupling instability index for: ； Maximum freezing rate for: ； In the formula: For a moment ,Location Local ice saturation at the location; For a moment ,Location To control the volume of ice within the body; For position Local porosity at the location; To control the overall volume of the body; For a moment ,Location Local oxygen diffusion blocking coefficient at the location; The effective oxygen diffusion coefficient under freezing conditions; The effective oxygen diffusion coefficient under unfrozen conditions; The intrinsic diffusion coefficient of oxygen; For a moment ,Location Local liquid water saturation at the location; m and n are structural correction indices; For a moment ,Location Local thermal-electric coupling instability indicators at the location; For local temperature gradients; The gradient is a local voltage characterization quantity. For local current density gradient; , , These are the weighting coefficients for the temperature gradient, voltage gradient, and current density gradient terms in the thermo-electric coupling instability index, respectively. For a moment ,Location Local freezing rate at the location; This represents the rate of change of the local ice phase volume fraction with respect to time. In key areas The maximum value is obtained for the corresponding local variable.

6. The fuel cell self-learning cold start method based on ice front evolution prediction according to claim 4, characterized in that, In step 2.3, the start-up reachability margin at time t. for: ； In the formula: This is the reference temperature for freezing. For reference temperature rise scale; Reference current density; ~ These are the weighting coefficients; Failure Boundary Function for: ； In the formula: For a moment ,Location Local current density at the location; This is the reference temperature for freezing. For reference temperature rise scale; For a moment ,Location Local current density at the location; Reference current density; For a moment ,Location Local ice saturation at the location; For a moment ,Location Local oxygen diffusion blocking coefficient at the location; For a moment The effective reactive surface decay rate; It is a volume infinitesimal element; to To activate the reachability weighting coefficient; To enable reachability margin; Maximum ice saturation in the key area; The maximum oxygen diffusion blocking coefficient in the critical area; The maximum thermal-electric coupling instability index in the critical area; to This represents the failure boundary weighting coefficient.

7. The fuel cell self-learning cold start method based on ice front evolution prediction according to claim 5, characterized in that, In step 2.4, the feature vector is frozen. for: ； Construct the operating condition parameter vector p: ； Cold start sample set D: ； In the formula: The average advance velocity of the ice front; For frozen coverage; This is the freezing bias factor; To enable reachability margin; This represents the maximum ice saturation within the critical area. The maximum oxygen diffusion blockage coefficient in the critical area; The maximum thermo-electric coupling instability index in the critical area; The effective reactive surface area decay rate; The maximum freezing rate within the critical area; Ambient temperature; This represents the initial water content of the membrane. The residual water level after shutdown; The cathode air excess coefficient; The excess hydrogen coefficient at the anode; This is the cathode back pressure; This is the anode back pressure; To supplement heating power; For the startup current ramp-up rate; These are the parameters for triggering the purge. For the first The vector of working parameters corresponding to the group of working conditions; For the first Freezing feature trajectory corresponding to the group working condition; For the first The sample label corresponding to the working condition group; j is the working condition number; N is the total number of samples.

8. The fuel cell self-learning cold start method based on ice front evolution prediction according to claim 3, characterized in that, In step 3.1, the input vector for: ； The formula for predicting future frozen features is: ； Success rate of startup for: ； In the formula: The average reactor temperature; It is an indicator of voltage non-uniformity; and These are the cathode pressure and anode pressure at time t, respectively; and These are the excess air coefficient at the cathode and the excess hydrogen coefficient at the anode, respectively, at time t. These are the excess air coefficient at the cathode and the excess hydrogen coefficient at the anode, respectively, at time t. The auxiliary heating power at time t; These are the initial operating parameters, consisting of ambient temperature, initial membrane water content, and residual water level. Let be the predicted value of the frozen feature at the current time t for the k-th future step; is the mapping function for predicting ice front evolution; L is the length of the historical input sequence; H is the prediction time domain length or the number of future control steps; k is the future prediction step number. ; This is a prediction of the start-up reachability margin corresponding to the predicted end point in the time domain; For Sigmoid mapping functions; To predict the maximum ice saturation value corresponding to the end point in the time domain; The predicted value of the maximum oxygen diffusion blockage coefficient corresponding to the end of the time domain; This is the predicted value of the maximum thermo-electric coupling instability index corresponding to the end point in the time domain; to The weighting coefficients are mapped to the probability of successful startup, with the startup reachability margin taking a positive input and the maximum ice saturation, maximum oxygen diffusion blockage coefficient, and maximum thermo-electric coupling instability index taking a negative input.

9. The fuel cell self-learning cold start method based on ice front evolution prediction according to claim 8, characterized in that, In step 3.2, the system state vector for: ； Discrete-stage patterns are output from the upper-level policy network. for: ； In the formula, This is the upper-level policy mapping function. These are discrete-stage pattern variables; The lower-level policy network outputs the original continuous action vector. for: ； Comprehensive value function for: ； Instant reward function for: ； In the formula: This is the current observation input vector; Predict sequences for future frozen features; This is the boundary function for cold start failure. The action performed in the previous moment; This is the mapping function for the upper-level strategy; This is the mapping function for the lower-level strategy; To initiate revenue commentary output; To freeze risk commentators' output; Output for damage risk commentators; and These are the penalty weights for freezing risk and damage risk, respectively; To initiate the revenue item; To freeze risk items; This is the damage cost term; This is an auxiliary energy consumption item; To initiate the accessibility margin increment; This represents the average reactor temperature increment. to To activate the profit weighting coefficient; To predict the maximum ice saturation in the future time domain; This is the maximum oxygen diffusion blocking coefficient predicted in the future time domain; This is the maximum oxygen diffusion blocking coefficient predicted in the future time domain; This represents the average value of the maximum thermo-electric coupling instability index in the future prediction time domain; The equivalent thermomechanical stress index of the membrane electrode; to To freeze the risk weighting coefficient; to This refers to the damage penalty weighting coefficient; to These are the auxiliary heating energy consumption weight and the purging operation cost weight, respectively.

10. The fuel cell self-learning cold start method based on ice front evolution prediction according to claim 9, characterized in that, In step 3.3, the final action that minimizes the distance to the original continuous action vector is selected within the feasible region of safe actions. for: ； The risk constraints for freezing assets are as follows: ； In the formula: In order to match the current stage model The corresponding feasible domain for safe actions; It is a norm 2; The predicted value of the failure boundary in the time domain is predicted for the k-th future step. This represents the minimum permissible safety margin.

Images