Fuel cell system optimization method based on neural network

Abstract

The invention provides a fuel cell system optimization method based on a neural network, and the method comprises the steps: S1, collecting a three-dimensional temperature field distribution matrix through a nanometer optical fiber sensor array, collecting a pressure gradient tensor through a piezoresistive micro-strain sensor grid, obtaining an eddy current characteristic spectrum through a laser Doppler velocimeter, and collecting an impedance spectrum through a high-frequency impedance analyzer; generating a space-time coupling data set; s2, inputting the space-time coupling data set into a hierarchical attention fusion network, and performing feature processing to generate an enhanced feature vector; s3, inputting the enhanced feature vector into the quantum-brain-like collaborative optimization model, and outputting a control parameter set; s4, adopting a meta-reinforcement learning framework, taking the control parameter set as an action space, adjusting the multi-target reward weight in real time, and generating an optimization strategy; and S5, driving an execution mechanism based on the control parameter set and the optimization strategy. According to the invention, the optimization precision and efficiency are improved through the neural network technology.

Description

Technical Field

[0001] The present invention relates to the technical field of fuel cell system optimization, and in particular to a fuel cell system optimization method based on a neural network. Background Art

[0002] With the rapid development of fuel cell vehicle technology, fuel cell systems, as a normal vehicle power generation device, convert chemical energy into electrical energy to provide power for vehicles. Among fuel cell products, hydrogen fuel cells have attracted widespread attention due to their better cleaning effect and environmental friendliness. As a power output device, the output control system of fuel cells is relatively complex. At the same time, the performance of the fuel cell itself is also affected by many factors, and the various factors are coupled and affect each other.

[0003] Traditional fuel cell optimization methods mainly focus on basic aspects such as material improvement, structural design, and operating parameter adjustment, and gradually improve performance through experimental verification and experience accumulation. However, the accuracy and efficiency of traditional fuel cell optimization methods are often low. Future optimization needs to integrate new materials, intelligent algorithms, and interdisciplinary innovations with traditional experience to promote fuel cell technology towards high efficiency, longevity, and low cost. Summary of the invention

[0004] The present invention provides a fuel cell system optimization method based on a neural network, which improves the optimization accuracy and efficiency through the neural network technology.

[0005] To achieve the above object, the present invention adopts the following technical solutions: The fuel cell system optimization method based on neural network includes: S1: Collecting the three-dimensional temperature field distribution matrix T through the nano-fiber sensor array matrix , the piezoresistive micro-strain sensor grid collects the pressure gradient tensor P gradient , laser Doppler velocimeter obtains eddy current characteristic spectrum V vortex , high frequency impedance analyzer collects impedance spectrum Z spectrum , generate a spatiotemporal coupled dataset D st ={T matrix ,P gradient ,V vortex ,Z spectrum}; S2: D st Input level attention fusion network, generate enhanced feature vector F after feature processing enhanced ; S3: F enhanced Input quantum-brain-like collaborative optimization model, output control parameter set C={Q opt ,P opt ,Topt ,η opt}, where Q opt is the flow setting value, P opt is the pressure setting value, T opt is the temperature target value, η opt is the predicted value of efficiency; S4: Using the meta-reinforcement learning framework, with C as the action space, the multi-objective reward weights w=[w1,w2,w3] are adjusted in real time to generate the optimization strategy π(s); w1 is the power maximization weight, w2 is the temperature stability weight, and w3 is the efficiency change rate weight; S5: Drive actuator based on C and π(s).

[0006] In this specification, in S2, the hierarchical attention fusion network includes the following: Temperature field branch module, used for 3D convolution; input: T matrix ; deal with: 3D convolution: kernel size 3×3×3, output channels 256, generating spatial features F thermal ; Activation function: ReLU nonlinear transformation; Pressure field branch module, used for temporal memory unit processing; input: P gradient ; deal with: Bidirectional GRU: hidden layer dimension 128, handles time series changes, and outputs time-varying features F pressure ; Eddy current branch module, used for frequency domain feature extraction; input: V vortex ; deal with: Multilayer Perceptron: Input layer: 10 dimensions → hidden layer: 64 dimensions, ReLU activation → output layer: 32 dimensions; Output: Eddy current frequency domain characteristics F vortex ; Electrochemical field branch module for impedance spectrum fusion; Input: Z spectrum ; deal with: Complex encoding: separate the real and imaginary parts and concatenate them into a real vector Z real ; Frequency Domain Attention: Attention weight calculation: α i =softmax(W a ⋅Z real ),i=1,…,100; W ais the trainable weight matrix; α i is the attention weight of the i-th frequency point; Weighted output: , is the complex impedance value measured at the i-th frequency point, Physical constraint branch module, used for thermal-electric coupling equation constraints; Input: T matrix ; deal with: Equation embedding: ; is the equivalent thermal conductivity, is the conductivity, is the temperature gradient, is the electric potential gradient, is the heat conduction term, is the Joule heat term; Regularization loss: ; is the 3D grid index, is the grid index in the length direction of the anode channel, i.e., the x-axis; is the grid index in the flow channel width direction, i.e., the y-axis; is the grid index in the flow channel height direction, i.e., the z-axis; N is the total number of grids, is the temperature field Laplace operator; Back propagation generates physical constraint features F phy ; Cross-modal attention fusion module, used for feature fusion; Input: F thermal 、F pressure 、F vortex 、F electro 、F phy ; deal with: Dynamic weight generation: ; ω is the dynamic weight, W ω is the learnable weight matrix, d is the input feature dimension; Weighted splicing: F enhanced =ω1F thermal ⊕ω2F pressure ⊕ω3F vortex ⊕ω4F electro ⊕ω5F phy ; ω1, ω2, ω3, ω4, and ω5 are respectively F thermal 、F pressure 、F vortex 、Felectro 、F phy The corresponding weight coefficient.

[0007] In this manual, in the physical constraint branch module, Z spectrum Fitting to the equivalent circuit model R Ω (R ct C dl ): ; Fit the impedance spectrum to the equivalent circuit model, is the ohmic resistance, is the charge transfer resistance, is the double layer capacitance; j is the imaginary component of impedance, is the angular frequency; The Levenberg-Marquardt algorithm is used to optimize the parameters and minimize the fitting error ‖Z spectrum −Z fit ‖ 2 ; Replace σ with Related functions: ; in is the local charge transfer resistance obtained by inversion of impedance spectroscopy; Equation modification: Replace σ with the spatially dependent : ; Discretization and calculation: Local parameter mapping: Mapping from equivalent circuit models to 3D meshes generates ; Dynamic conductivity calculation: For each grid point ,calculate ; Equation solving: Discretize the equation using the finite difference method and solve T iteratively matrix and Φ; Physical constraint feature generation: Regularization loss term update: ; Feature Backpropagation: Calculation To T matrix and The gradient of F generates the physical-electrochemical coupling feature phy .

[0008] In this manual, the specific process of S2 is as follows: S21. Multimodal feature extraction; S211. 3D convolution processing spatial features: Input data: T matrix and P gradient ; operate: To T matrix and P gradient Apply 3×3×3 three-dimensional convolution kernels to extract local spatial features; To T matrix Perform multi-level convolution to gradually compress the spatial dimension to 25×25×12 and output the feature map F T ∈R 25 ×25×12×C , C is the number of channels; P gradient Upsample to 25×25×25 and then compress to F through convolution p ∈R 25×25×12×C , C is the number of channels, and F T Alignment; S212. Bidirectional gated recurrent unit processing timing characteristics: Input data: V vortex and Z spectrum ; Operation: V vortex and Z spectrum Input bidirectional gated recurrent units separately to capture timing dependencies; Output time series feature vector F V and F Z ; S22. Hierarchical attention fusion; S221. Intra-modal attention: Input: F T 、F p 、F V 、F Z ; Operation: For spatial feature F T and F p , the weight of each position is calculated through the spatial attention module, the key area is highlighted, and the feature and ; For the time series feature F V and F Z , the importance of different time steps is assigned through the time attention module to obtain the feature and ; S222. Cross-modal attention: Input: Features after intra-modal attention , , and ; Operation: , , and Project to uniform dimension R 256 , calculate the cross-modal interaction weights through a multi-head attention mechanism; Fusion formula: ; in, is the modality weight, obtained by normalizing the attention score; To fusion features, is the feature after intra-modal attention; S23. Physical equation constraints; S231. Electric potential field prediction: Input: Fusion features ; operate: Predict the electric potential field matrix Φ through the fully connected layer; Compute the physical constraint residuals: ; in, is the equivalent thermal conductivity, σ is the electrical conductivity, is the electric potential gradient, is the temperature field Laplace operator; S232. Constrained Optimization: Physical Constraint Residuals Add the total loss function as a regularization term to ensure that the network output conforms to the thermal-electric coupling equation and obtain the fusion features after physical constraints. ; S24. Feature enhancement and dimensionality reduction: Input: fusion features after physical constraints ; Operation: Yes The feature dimension is compressed to R through the fully connected layer 576 , generate enhanced feature vector F enhanced .

[0009] In this manual, the specific process of S3 is as follows: S31.F enhanced After standardization, we get the feature vector F with a mean of 0 and a standard deviation of 1. norm , through quantum gate operation on F norm Encoding quantum states: Q state =U(θ q )⋅F norm , where U is the unitary transformation matrix, θ q is the quantum circuit parameter, Q state is the quantum state vector; S32. By simulating the dynamic weight adjustment mechanism of biological synapses, F norm Generate adaptive weight matrix W adaptive ; Spiking neuron model: W adaptive =Sigmoid(LSTM(F norm ))⋅H, where H is the brain-like pulse timing coding matrix, Sigmoid is the activation function, and LSTM is a pulse-driven LSTM, which is used to simulate the dynamic synaptic plasticity of biological neurons; S33. Based on Q state Through the Hamiltonian H Q Calculate the expected energy value E: optimization To minimize E, we can get the solution that minimizes E. ; in, For Take the conjugate transpose as a whole, V is the variational quantum circuit, is the set of adjustable parameters of the line, for Embedded into the structure of the variational quantum circuit V; will minimize E With W adaptive Combined to generate parameter vector C prime =W adaptive ⋅(Q state ⊗ ), where ⊗ is the tensor product operation; S34.To C prime Physical dimension matching is performed and system boundary condition constraints are imposed to obtain the control parameter set C.

[0010] In this specification, in S34, the system boundary condition is: Flow range: Q min ≤Q opt ≤Q max ; Pressure range: P min ≤P opt ≤P max ; Temperature range: T min ≤T opt ≤T max ; Efficiency range: η min ≤η opt ≤η max ; The specific process of physical dimension matching is as follows: C prime=[c1,c2,c3,c4], c1 corresponds to the original value of flow, c2 corresponds to the original value of pressure, c3 corresponds to the original value of temperature, and c4 corresponds to the original value of efficiency; Q opt Generation of: Compress c1 to [0,1] through the Sigmoid function, and then linearly map it to the actual flow range, normalizing the function: Q opt =Sigmoid(c1)⋅(Q max −Q min )+Q min ; Adjust Q according to the real-time current density I opt : ; is the flow dynamic coefficient, is the rated current density; P opt Generation of: The Tanh function is used to limit c2 to [-1,1] and then mapped to the midpoint of the pressure range to avoid extreme values ​​and normalize the function: ; Based on P gradient P opt Make feedback adjustments: ; is the pressure gradient compensation coefficient, is the mean pressure gradient; T opt Generation of: Use the ReLU function to avoid negative temperature values, then linearly map to the actual temperature range, and normalize the function: ; Based on T matrix To T opt To make spatial balance adjustments: ; is the mean value of the temperature field, is the temperature equilibrium coefficient; η opt Generation of: Use the Softplus function to ensure that the efficiency is non-negative, then map it to the actual efficiency range and normalize the function: ; Combined impedance spectroscopy Z spectrum Ohmic loss R ohm For ηopt Make adjustments: ; is the reference impedance.

[0011] In this specification, S5 also includes: inputting real-time data into the multi-scale digital twin, when the electrochemical deviation δ echem =‖Z act −Z sim ‖ / ‖Z sim ‖>10% or thermodynamic deviation δ thermo =|T act −T opt ∣>5K, trigger S3 control parameter re-optimization; Z act is the actual impedance spectrum, Z sim To simulate the impedance spectrum, T act is the real-time temperature measurement value, T opt is the temperature target value.

[0012] In this specification, in S4, the pre-training stage of meta-reinforcement learning: the meta-policy network is trained using the historical failure dataset, and the loss function is L meta =0.8L task +0.2L robust , where L task is the control error, L robust To resist interference loss; The action space of meta-reinforcement learning is C={Q opt ,P opt ,T opt}, the state space is s=[F enhanced ,δ echem ,δ thermo ], the multi-objective reward function R is defined as: ; is the real-time pressure measurement value, w1, w2, and w3 are dynamically updated through Pareto front analysis, with the constraints that the sum of the weights is 1 and each weight is greater than 0.1; According to the state space s, the control parameters are dynamically adjusted, and the strategy π(s) maps the state to the action space, that is: ; in is the incremental adjustment value of the control parameter, Q opt The incremental adjustment value of P opt The incremental adjustment value of T opt The incremental adjustment value of C and Combined, the control instructions are generated, that is, the strategy π(s) is: ; Q final is the flow target value, P final is the pressure target value, T final is the temperature target value.

[0013] In this specification, multi-scale digital twins include: Microscale: Molecular dynamics simulations were performed using the ReaxFF reaction force field with a time step of 0.1 fs and input Q opt , output the simulated impedance spectrum Z of the catalyst surface reaction sim ; Macro scale: Establish a multi-physics field coupling model using finite element software, with a mesh size of ≤0.1 mm and input T opt , output temperature field distribution T sim ; when When , S3 control parameters are re-optimized; When thermodynamic deviation , adjust w2=w2+0.1.

[0014] In this specification, the actuator includes a hydrogen circulation pump, a coolant flow valve and a voltage regulator.

[0015] In summary, the present invention has at least the following beneficial effects: The present invention improves the optimization efficiency and effect through multiple dimensions (pressure, temperature, eddy current characteristics, impedance) using neural networks, quantum-brain-like, and meta-reinforcement learning technologies, thereby greatly improving the optimization efficiency and effect of the fuel cell system, and providing a higher foundation for the development of fuel cell system optimization technology, which is conducive to the development and further promotion of this technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0017] Figure 1 It is a schematic diagram of the neural network-based fuel cell system optimization method involved in the present invention.

[0018] Figure 2 It is a schematic diagram of the steps of S2 involved in the present invention.

[0019] Figure 3 It is a schematic diagram of the steps of S21 involved in the present invention.

[0020] Figure 4 It is a schematic diagram of the steps of S22 involved in the present invention.

[0021] Figure 5 It is a schematic diagram of the step S23 involved in the present invention. DETAILED DESCRIPTION

[0022] In the following, only some exemplary embodiments are briefly described. As those skilled in the art will appreciate, the described embodiments may be modified in various ways without departing from the spirit or scope of the embodiments of the present invention. Therefore, the drawings and descriptions are considered to be exemplary and non-restrictive in nature.

[0023] The disclosure below provides many different embodiments or examples to implement different structures of the embodiments of the present invention. In order to simplify the disclosure of the embodiments of the present invention, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the embodiments of the present invention. In addition, the embodiments of the present invention can repeat reference numbers and / or reference letters in different examples, and this repetition is for the purpose of simplification and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed.

[0024] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0025] like Figure 1 As shown, this embodiment provides a fuel cell system optimization method based on a neural network, including: S1: Collecting the three-dimensional temperature field distribution matrix T through the nano-fiber sensor array matrix , the piezoresistive micro-strain sensor grid collects the pressure gradient tensor P gradient , laser Doppler velocimeter obtains eddy current characteristic spectrum V vortex , high frequency impedance analyzer collects impedance spectrum Z spectrum , generate a spatiotemporal coupled dataset D st ={T matrix ,P gradient ,V vortex ,Z spectrum}; S2: D st Input level attention fusion network, generate enhanced feature vector F after feature processing enhanced ; S3: F enhanced Input quantum-brain-like collaborative optimization model, output control parameter set C={Q opt ,Popt ,T opt ,η opt}, where Q opt is the flow setting value, P opt is the pressure setting value, T opt is the temperature target value, η opt is the predicted value of efficiency; S4: Using the meta-reinforcement learning framework, with C as the action space, the multi-objective reward weights w=[w1,w2,w3] are adjusted in real time to generate the optimization strategy π(s); w1 is the power maximization weight, w2 is the temperature stability weight, and w3 is the efficiency change rate weight; S5: Drive actuator based on C and π(s).

[0026] In some embodiments, in S2, the hierarchical attention fusion network includes the following: Temperature field branch module, used for 3D convolution; input: T matrix ; deal with: 3D convolution: kernel size 3×3×3, output channels 256, generating spatial features F thermal ; Activation function: ReLU nonlinear transformation; Pressure field branch module, used for temporal memory unit processing; input: P gradient ; deal with: Bidirectional GRU: hidden layer dimension 128, handles time series changes, and outputs time-varying features F pressure ; Eddy current branch module, used for frequency domain feature extraction; input: V vortex ; deal with: Multilayer Perceptron: Input layer: 10 dimensions → hidden layer: 64 dimensions, ReLU activation → output layer: 32 dimensions; Output: Eddy current frequency domain characteristics F vortex ; Electrochemical field branch module for impedance spectrum fusion; Input: Z spectrum ; deal with: Complex encoding: separate the real and imaginary parts and concatenate them into a real vector Z real ; Frequency Domain Attention: Attention weight calculation: α i =softmax(W a ⋅Z real ),i=1,…,100; Wa is the trainable weight matrix; α i is the attention weight of the i-th frequency point; Weighted output: , is the complex impedance value measured at the i-th frequency point, Physical constraint branch module, used for thermal-electric coupling equation constraints; Input: T matrix ; deal with: Equation embedding: ; is the equivalent thermal conductivity, is the conductivity, is the temperature gradient, is the electric potential gradient, is the heat conduction term, is the Joule heat term; Regularization loss: ; is the 3D grid index, is the grid index in the length direction of the anode channel, i.e., the x-axis; is the grid index in the flow channel width direction, i.e., the y-axis; is the grid index in the flow channel height direction, i.e., the z-axis; N is the total number of grids, is the temperature field Laplace operator; Back propagation generates physical constraint features F phy ; Cross-modal attention fusion module, used for feature fusion; Input: F thermal 、F pressure 、F vortex 、F electro 、F phy ; deal with: Dynamic weight generation: ; ω is the dynamic weight, W ω is the learnable weight matrix, d is the input feature dimension; Weighted splicing: F enhanced =ω1F thermal ⊕ω2F pressure ⊕ω3F vortex ⊕ω4F electro ⊕ω5F phy ; ω1, ω2, ω3, ω4, and ω5 are respectively F thermal 、F pressure 、Fvortex 、F electro 、F phy The corresponding weight coefficient.

[0027] In some embodiments, in the physical constraint branch module, Z spectrum Fitting to the equivalent circuit model R Ω (R ct C dl ): ; Fit the impedance spectrum to the equivalent circuit model, is the ohmic resistance, is the charge transfer resistance, is the double layer capacitance; j is the imaginary component of impedance, is the angular frequency; The Levenberg-Marquardt algorithm is used to optimize the parameters and minimize the fitting error ‖Z spectrum −Z fit ‖ 2 ; Replace σ with Related functions: ; in is the local charge transfer resistance obtained by inversion of impedance spectroscopy; Equation modification: Replace σ with the spatially dependent : ; Discretization and calculation: Local parameter mapping: Mapping from equivalent circuit models to 3D meshes generates ; Dynamic conductivity calculation: For each grid point ,calculate ; Equation solving: Discretize the equation using the finite difference method and solve T iteratively matrix and Φ; Physical constraint feature generation: Regularization loss term update: ; Feature Backpropagation: Calculation To T matrix and The gradient of F generates the physical-electrochemical coupling feature phy .

[0028] In some embodiments, Figure 2 , Figure 3, Figure 4 and Figure 5 As shown, the specific process of S2 is as follows: S21. Multimodal feature extraction; S211. 3D convolution processing spatial features: Input data: T matrix and P gradient ; operate: To T matrix and P gradient Apply 3×3×3 three-dimensional convolution kernels to extract local spatial features; To T matrix Perform multi-level convolution to gradually compress the spatial dimension to 25×25×12 and output the feature map F T ∈R 25 ×25×12×C , C is the number of channels; P gradient Upsample to 25×25×25 and then compress to F through convolution p ∈R 25×25×12×C , C is the number of channels, and F T Alignment; S212. Bidirectional gated recurrent unit processing timing characteristics: Input data: V vortex and Z spectrum ; Operation: V vortex and Z spectrum Input bidirectional gated recurrent units separately to capture timing dependencies; Output time series feature vector F V and F Z ; S22. Hierarchical attention fusion; S221. Intra-modal attention: Input: F T 、F p 、F V 、F Z ; Operation: For spatial feature F T and F p , the weight of each position is calculated through the spatial attention module, the key area is highlighted, and the feature and ; For the time series feature F V and F Z , the importance of different time steps is assigned through the time attention module to obtain the feature and ; S222. Cross-modal attention: Input: Features after intra-modal attention , , and ; Operation: , , and Project to uniform dimension R 256 , calculate the cross-modal interaction weights through a multi-head attention mechanism; Fusion formula: ; in, is the modality weight, obtained by normalizing the attention score; To fusion features, is the feature after intra-modal attention; S23. Physical equation constraints; S231. Electric potential field prediction: Input: Fusion features ; operate: Predict the electric potential field matrix Φ through the fully connected layer; Compute the physical constraint residuals: ; in, is the equivalent thermal conductivity, σ is the electrical conductivity, is the electric potential gradient, is the temperature field Laplace operator; S232. Constrained Optimization: Physical Constraint Residuals Add the total loss function as a regularization term to ensure that the network output conforms to the thermal-electric coupling equation and obtain the fusion features after physical constraints. ; S24. Feature enhancement and dimensionality reduction: Input: fusion features after physical constraints ; Operation: Yes The feature dimension is compressed to R through the fully connected layer 576 , generate enhanced feature vector F enhanced .

[0029] In some embodiments, the specific process of S3 is as follows: S31.F enhanced After standardization, we get the feature vector F with a mean of 0 and a standard deviation of 1. norm , through quantum gate operation on F norm Encoding quantum states: Q state =U(θ q)⋅F norm , where U is the unitary transformation matrix, θ q is the quantum circuit parameter, Q state is the quantum state vector; S32. By simulating the dynamic weight adjustment mechanism of biological synapses, F norm Generate adaptive weight matrix W adaptive ; Spiking neuron model: W adaptive =Sigmoid(LSTM(F norm ))⋅H, where H is the brain-like pulse timing coding matrix, Sigmoid is the activation function, and LSTM is a pulse-driven LSTM, which is used to simulate the dynamic synaptic plasticity of biological neurons; S33. Based on Q state Through the Hamiltonian H Q Calculate the expected energy value E: ,optimization To minimize E, we can get the solution that minimizes E. ; in, For Take the conjugate transpose as a whole, V is the variational quantum circuit, is the set of adjustable parameters of the line, for Embedded into the structure of the variational quantum circuit V; will minimize E With W adaptive Combined to generate parameter vector C prime =W adaptive ⋅(Q state ⊗ ), where ⊗ is the tensor product operation; S34.To C prime Physical dimension matching is performed and system boundary condition constraints are imposed to obtain the control parameter set C.

[0030] In some embodiments, in S34, the system boundary condition is: Flow range: Q min ≤Q opt ≤Q max ; Pressure range: P min ≤P opt ≤P max ; Temperature range: T min ≤T opt ≤T max ; Efficiency range: η min ≤ηopt ≤η max ; The specific process of physical dimension matching is as follows: C prime =[c1,c2,c3,c4], c1 corresponds to the original value of flow, c2 corresponds to the original value of pressure, c3 corresponds to the original value of temperature, and c4 corresponds to the original value of efficiency; Q opt Generation of: Compress c1 to [0,1] through the Sigmoid function, and then linearly map it to the actual flow range, normalizing the function: Q opt =Sigmoid(c1)⋅(Q max −Q min )+Q min ; Adjust Q according to the real-time current density I opt : ; is the flow dynamic coefficient, is the rated current density; P opt Generation of: The Tanh function is used to limit c2 to [-1,1] and then mapped to the midpoint of the pressure range to avoid extreme values ​​and normalize the function: ; Based on P gradient P opt Make feedback adjustments: ; is the pressure gradient compensation coefficient, is the mean pressure gradient; T opt Generation of: Use the ReLU function to avoid negative temperature values, then linearly map to the actual temperature range, and normalize the function: ; Based on T matrix To T opt To make spatial balance adjustments: ; is the mean value of the temperature field, is the temperature equilibrium coefficient; η opt Generation of: Use the Softplus function to ensure that the efficiency is non-negative, then map it to the actual efficiency range and normalize the function: ; Combined impedance spectroscopy Z spectrum Ohmic loss R ohm For η opt Make adjustments: ; is the reference impedance.

[0031] In some embodiments, S5 further includes: inputting real-time data into the multi-scale digital twin, when the electrochemical deviation δ echem =‖Z act −Z sim ‖ / ‖Z sim ‖>10% or thermodynamic deviation δ thermo =|T act −T opt ∣>5K, trigger S3 control parameter re-optimization; Z act is the actual impedance spectrum, Z sim To simulate the impedance spectrum, T act is the real-time temperature measurement value, T opt is the temperature target value.

[0032] In some embodiments, in S4, the pre-training phase of meta-reinforcement learning: the meta-strategy network is trained using the historical failure dataset, and the loss function is L meta =0.8L task +0.2L robust , where L task is the control error, L robust To resist interference loss; The action space of meta-reinforcement learning is C={Q opt ,P opt ,T opt}, the state space is s=[F enhanced ,δ echem ,δ thermo ], the multi-objective reward function R is defined as: ; is the real-time pressure measurement value, w1, w2, and w3 are dynamically updated through Pareto front analysis, with the constraints that the sum of the weights is 1 and each weight is greater than 0.1; According to the state space s, the control parameters are dynamically adjusted, and the strategy π(s) maps the state to the action space, that is: ; in is the incremental adjustment value of the control parameter, Q opt The incremental adjustment value of P opt The incremental adjustment value of T opt The incremental adjustment value of C and Combined, the control instructions are generated, that is, the strategy π(s) is: ; For example: If Q opt =120L / min, , then Q final =125L / min; Q final is the flow target value, P final is the pressure target value, T final is the temperature target value.

[0033] Control instruction conversion: Hydrogen circulation pump: speed ω is determined by the flow setting value Q final calculate: ω=ξ⋅ω base (ξ=Q final / Q rated ,ω base =3000rpm) For example: If Q final =125L / min, rated flow rate Q rated =150L / min, then ξ=0.83, ω=2490rpm.

[0034] Coolant flow valve: The opening coefficient ψ is determined by the temperature deviation δthermo: ψ=ψ base +k p ⋅δ thermo (ψ base =0.5,k p =0.1 / ℃) For example: If δ thermo =3℃, then ψ=0.5+0.3=0.8.

[0035] Voltage regulator: Sets the voltage V set According to the efficiency prediction value η opt Adjustment: V set =V nominal ⋅(1+0.02⋅(η opt −60%) For example: If η opt =62%, then Vset =1.02⋅V nominal . The actuator performs the action and the command is issued: The control instructions are sent to the actuator controller via the real-time operating system (RTOS) at a fixed period (e.g. 10ms): Hydrogen circulation pump: receives PWM signal and adjusts the motor speed to the target ω; Coolant valve: receives 4-20mA analog signal and adjusts the opening to ψ; Motion Compensation:

[0036] If the actual flow rate Q act With Q final If the deviation exceeds 5%, PID feedback control is triggered: where k p =0.5, k i =0.1, k d =0.2. Closed-loop feedback and re-optimization Sensor Feedback: Real-time collection of post-execution data:

[0037] Flow meter measures Q act ; The pressure sensor measures P act ; Thermocouple measurement T act .

[0038] Digital Twin Validation: The actual data (Q act ,P act ,T act ) Input multi-scale digital twins to generate predicted values ​​(Q sim ,P sim ,T sim ), calculate the deviation: ; If any deviation is >8%, re-optimization in step S3 is triggered.

[0039] Policy Update: The actual state s′=[F enhanced ′,δ echem ′,δ thermo ′] Input meta-reinforcement learning framework and update strategy π(s): ; Where α=0.001 is the learning rate and γ=0.9 is the discount factor. Exception handling and safety mechanisms Over limit protection: >T opt +10℃, immediately cut off the hydrogen supply and start emergency cooling; Failure recovery: When an actuator failure is detected (such as a stuck valve), it switches to the backup controller. Example: Execution process under cold start condition Initial state: =25℃, target T opt =70℃,δ thermo =45℃. Strategy driven:

[0040] π(s) output increment ; Final Command: Q final =140L / min, P final =115kPa.

[0041] Actuator Action: The speed of the hydrogen pump is increased to ω=2800rpm; The cooling valve is fully opened (ψ=1.0) to accelerate the temperature rise.

[0042] Feedback Adjustment: 5 minutes later T act =65℃,δ thermo =5℃, π(s) decreases To +5L / min, avoid overshoot.

[0043] In some embodiments, the multi-scale digital twin includes: Microscale: Molecular dynamics simulations were performed using the ReaxFF reaction force field with a time step of 0.1 fs and input Q opt , output the simulated impedance spectrum Z of the catalyst surface reaction sim ; Macro scale: Establish a multi-physics field coupling model using finite element software, with a mesh size of ≤0.1 mm and input T opt , output temperature field distribution T sim ; When δ echem =‖Z act −Z sim ‖ / ‖Z sim ‖>10%, triggering S3 control parameter re-optimization; When the thermodynamic deviation δ thermo =|T act −T sim ∣>5K, adjust w2=w2+0.1.

[0044] In some embodiments, the actuator includes a hydrogen circulation pump, a coolant flow valve, and a voltage regulator.

[0045] The above-described embodiments are used to illustrate the present invention, not to limit the present invention, so changes in the exemplified values ​​or replacement of equivalent elements should still fall within the scope of the present invention.

[0046] From the above detailed description, it can be understood by those skilled in the art that the present invention can indeed achieve the aforementioned objectives and is in compliance with the provisions of the Patent Law.

[0047] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic creative concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. It should be pointed out that any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0048] It should be noted that the above description of the relevant process is only for example and explanation, and does not limit the scope of application of this specification. For those skilled in the art, various modifications and changes can be made to the process under the guidance of this specification. However, these modifications and changes are still within the scope of this specification.

[0049] The basic concepts have been described above. Obviously, for those of ordinary skill in the art who have read this application, the above invention disclosure is only for example and does not constitute a limitation of this application. Although not explicitly stated here, those of ordinary skill in the art may make various modifications, improvements and amendments to this application. Such modifications, improvements and amendments are suggested in this application, so such modifications, improvements and amendments still belong to the spirit and scope of the exemplary embodiments of this application.

[0050] At the same time, the present application uses specific words to describe the embodiments of the present application. For example, "one embodiment", "an embodiment", and / or "some embodiments" refer to a certain feature, structure or characteristic related to at least one embodiment of the present application. Therefore, it should be emphasized and noted that "one embodiment" or "an embodiment" or "an alternative embodiment" mentioned twice or more in different positions in this specification does not necessarily refer to the same embodiment. In addition, some features, structures or characteristics in one or more embodiments of the present application can be appropriately combined.

[0051] In addition, it will be understood by those skilled in the art that various aspects of the present application may be illustrated and described by a number of patentable categories or situations, including any new and useful process, machine, product or combination of substances, or any new and useful improvement thereof. Therefore, various aspects of the present application may be implemented entirely by hardware, entirely by software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. The above hardware or software may be referred to as a "unit", "module" or "system". In addition, various aspects of the present application may take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0052] The computer program code required for the operation of each part of the present application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python, etc., conventional procedural programming languages ​​such as C programming language, VisualBasic, Fortran2103, Perl, COBOL2102, PHP, ABAP, dynamic programming languages ​​such as Python, Ruby and Groovy, or other programming languages. The program code can be run entirely on the user's computer, or run on the user's computer as an independent software package, or run partially on the user's computer and partially on the remote computer, or run entirely on the remote computer or server. In the latter case, the remote computer can be connected to the user's computer through any network form, such as a local area network (LAN) or a wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as software as a service (SaaS).

[0053] In addition, unless explicitly stated in the claims, the order of the processing elements and sequences described in this application, the use of alphanumeric characters, or the use of other names are not intended to limit the order of the processes and methods of this application. Although the above disclosure discusses some invention embodiments that are currently considered useful through various examples, it should be understood that such details are only for illustrative purposes, and the attached claims are not limited to the disclosed embodiments. On the contrary, the claims are intended to cover all modifications and equivalent combinations that are consistent with the essence and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a pure software solution, for example, installation on an existing server or mobile device.

[0054] Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the above description of the embodiments of the application, multiple features are sometimes combined into one embodiment, figure or description thereof. However, this method of the application should not be interpreted as reflecting the intention that the claimed object requires more features than those explicitly stated in each claim. On the contrary, the subject of the invention should have fewer features than the above single embodiment.

Claims

1. A fuel cell system optimization method based on a neural network, characterized in that: include: S1: Collecting the three-dimensional temperature field distribution matrix T through the nano-fiber sensor array matrix , the piezoresistive micro-strain sensor grid collects the pressure gradient tensor P gradient , laser Doppler velocimeter obtains eddy current characteristic spectrum V vortex , high frequency impedance analyzer collects impedance spectrum Z spectrum , generate a spatiotemporal coupled dataset D st ={T matrix ,P gradient ,V vortex ,Z spectrum }; S2: D st Input level attention fusion network, generate enhanced feature vector F after feature processing enhanced ; S3: F enhanced Input quantum-brain-like collaborative optimization model, output control parameter set C={Q opt ,P opt ,T opt ,η opt }, where Q opt is the flow setting value, P opt is the pressure setting value, T opt is the temperature target value, η opt is the predicted value of efficiency; S4: Using the meta-reinforcement learning framework, with C as the action space, the multi-objective reward weights w=[w1,w2,w3] are adjusted in real time to generate the optimization strategy π(s); w1 is the power maximization weight, w2 is the temperature stability weight, and w3 is the efficiency change rate weight; S5: Drive actuator based on C and π(s).

2. The method for optimizing a fuel cell system based on a neural network according to claim 1, characterized in that: In S2, the hierarchical attention fusion network includes the following: Temperature field branch module, used for 3D convolution; input: T matrix ; deal with: 3D convolution: kernel size 3×3×3, output channels 256, generating spatial features F thermal ; Activation function: ReLU nonlinear transformation; Pressure field branch module, used for temporal memory unit processing; input: P gradient ; deal with: Bidirectional GRU: hidden layer dimension 128, handles time series changes, and outputs time-varying features F pressure ; Eddy current branch module, used for frequency domain feature extraction; input: V vortex ; deal with: Multilayer Perceptron: Input layer: 10 dimensions → hidden layer: 64 dimensions, ReLU activation → output layer: 32 dimensions; Output: Eddy current frequency domain characteristics F vortex ; Electrochemical field branch module for impedance spectrum fusion; Input: Z spectrum ; deal with: Complex encoding: separate the real and imaginary parts and concatenate them into a real vector Z real ; Frequency Domain Attention: Attention weight calculation: α i =softmax(W a ⋅Z real ),i=1,…,100; W a is the trainable weight matrix; α i is the attention weight of the i-th frequency point; Weighted output: , is the complex impedance value measured at the i-th frequency point, Physical constraint branch module, used for thermal-electric coupling equation constraints; Input: T matrix ; deal with: Equation embedding: ; is the equivalent thermal conductivity, is the conductivity, is the temperature gradient, is the electric potential gradient, is the heat conduction term, is the Joule heat term; Regularization loss: ； is the 3D grid index, is the grid index in the length direction of the anode channel, i.e., the x-axis; is the grid index in the flow channel width direction, i.e., the y-axis; is the grid index in the flow channel height direction, i.e., the z-axis; N is the total number of grids, is the temperature field Laplace operator; Back propagation generates physical constraint features F phy ; Cross-modal attention fusion module, used for feature fusion; Input: F thermal 、F pressure 、F vortex 、F electro 、F phy ; deal with: Dynamic weight generation: ; ω is the dynamic weight, W ω is the learnable weight matrix, d is the input feature dimension; Weighted splicing: F enhanced =ω1F thermal ⊕ω2F pressure ⊕ω3F vortex ⊕ω4F electro ⊕ω5F phy ; ω1, ω2, ω3, ω4, and ω5 are respectively F thermal 、F pressure 、F vortex 、F electro 、F phy The corresponding weight coefficient.

3. The method for optimizing a fuel cell system based on a neural network according to claim 2, characterized in that: In the physical constraint branch module, set Z spectrum Fitting to the equivalent circuit model R Ω (R ct C dl ): ； Fit the impedance spectrum to the equivalent circuit model, is the ohmic resistance, is the charge transfer resistance, is the double layer capacitance; j is the imaginary component of impedance, is the angular frequency; The Levenberg-Marquardt algorithm is used to optimize the parameters and minimize the fitting error ‖Z spectrum −Z fit ‖ 2 ; Replace σ with Related functions: ； in is the local charge transfer resistance obtained by inversion of impedance spectroscopy; Equation modification: Replace σ with the spatially dependent : ； Discretization and calculation: Local parameter mapping: Mapping from equivalent circuit models to 3D meshes generates ; Dynamic conductivity calculation: For each grid point ,calculate ; Equation solving: Discretize the equation using the finite difference method and solve T iteratively matrix and Φ; Physical constraint feature generation: Regularization loss term update: ; feature Back Propagation: Calculation To T matrix and The gradient of F generates the physical-electrochemical coupling feature phy .

4. The method for optimizing a fuel cell system based on a neural network according to claim 1, characterized in that: The specific process of S2 is as follows: S21. Multimodal feature extraction; S211. 3D convolution processing spatial features: Input data: T matrix and P gradient ; operate: To T matrix and P gradient Apply 3×3×3 three-dimensional convolution kernels to extract local spatial features; right T matrix Perform multi-level convolution to gradually compress the spatial dimension to 25×25×12 and output the feature map F T ∈R 25×25×12×C , C is the number of channels; P gradient Upsample to 25×25×25 and then compress to F through convolution p ∈R 25×25×12×C , C is the number of channels, and F T Alignment; S212. Bidirectional gated recurrent unit processing timing characteristics: enter Data: V vortex and Z spectrum ; Operation: V vortex and Z spectrum Input bidirectional gated recurrent units separately to capture timing dependencies; Output time series feature vector F V and F Z ; S22. Hierarchical attention fusion; S221. Intra-modal attention: Input: F T 、F p 、F V 、F Z ; Operation: For spatial feature F T and F p , the weight of each position is calculated through the spatial attention module, the key area is highlighted, and the feature and ; For the time series feature F V and F Z , the importance of different time steps is assigned through the time attention module to obtain the feature and ; S222. Cross-modal attention: Input: Features after intra-modal attention , , and ; Operation: , , and Project to uniform dimension R 256 , calculate the cross-modal interaction weights through a multi-head attention mechanism; Fusion formula: ; in, is the modality weight, obtained by normalizing the attention score; To fusion features, is the feature after intra-modal attention; S23. Physical equation constraints; S231. Electric potential field prediction: Input: Fusion features ; operate: Predict the electric potential field matrix Φ through the fully connected layer; Compute the physical constraint residuals: ; in, is the equivalent thermal conductivity, σ is the electrical conductivity, is the electric potential gradient, is the temperature field Laplace operator; S232. Constrained Optimization: Physical Constraint Residuals Add the total loss function as a regularization term to ensure that the network output conforms to the thermal-electric coupling equation and obtain the fusion features after physical constraints. ; S24. Feature enhancement and dimensionality reduction: Input: fusion features after physical constraints ; Operation: Yes The feature dimension is compressed to R through the fully connected layer 576 , generate enhanced feature vector F enhanced .

5. The method for optimizing a fuel cell system based on a neural network according to claim 1, characterized in that: The specific process of S3 is as follows: S31.F enhanced After standardization, we get the feature vector F with a mean of 0 and a standard deviation of 1. norm , through quantum gate operation on F norm Encoding quantum states: Q state =U(θ q )⋅F norm , where U is the unitary transformation matrix, θ q is the quantum circuit parameter, Q state is the quantum state vector; S32. By simulating the dynamic weight adjustment mechanism of biological synapses, F norm Generate adaptive weight matrix W adaptive ; Spiking neuron model: W adaptive =Sigmoid(LSTM(F norm ))⋅H, where H is the brain-like pulse timing coding matrix, Sigmoid is the activation function, and LSTM is a pulse-driven LSTM, which is used to simulate the dynamic synaptic plasticity of biological neurons; S33. Based on Q state Through the Hamiltonian H Q Calculate the expected energy value E: optimization To minimize E, we can get the solution that minimizes E. ; in, For Take the conjugate transpose as a whole, V is the variational quantum circuit, is the set of adjustable parameters of the line, for Embedded into the structure of the variational quantum circuit V; will minimize E With W adaptive Combined to generate parameter vector C prime =W adaptive ⋅(Q state ⊗ ), where ⊗ is the tensor product operation; S34.To C prime Physical dimension matching is performed and system boundary condition constraints are imposed to obtain the control parameter set C.

6. The method for optimizing a fuel cell system based on a neural network according to claim 5, characterized in that: In S34, the system boundary conditions are: Flow range: Q min ≤Q opt ≤Q max ; Pressure range: P min ≤P opt ≤P max ; Temperature range: T min ≤T opt ≤T max ; Efficiency range: η min ≤η opt ≤η max ; The specific process of physical dimension matching is as follows: C prime =[c1,c2,c3,c4], c1 corresponds to the original value of flow, c2 corresponds to the original value of pressure, c3 corresponds to the original value of temperature, and c4 corresponds to the original value of efficiency; Q opt Generation of: Compress c1 to [0,1] through the Sigmoid function, and then linearly map it to the actual flow range, normalizing the function: Q opt =Sigmoid(c1)⋅(Q max −Q min )+Q min ; Adjust Q according to the real-time current density I opt : ； is the flow dynamic coefficient, is the rated current density; P opt Generation of: The Tanh function is used to limit c2 to [-1,1] and then mapped to the midpoint of the pressure range to avoid extreme values ​​and normalize the function: ； Based on P gradient P opt Make feedback adjustments: ； is the pressure gradient compensation coefficient, is the mean pressure gradient; T opt Generation of: Use the ReLU function to avoid negative temperature values, then linearly map to the actual temperature range, and normalize the function: ； Based on T matrix To T opt To make spatial balance adjustments: ； is the mean value of the temperature field, is the temperature equilibrium coefficient; η opt Generation of: Use the Softplus function to ensure that the efficiency is non-negative, then map it to the actual efficiency range and normalize the function: ； Combined impedance spectroscopy Z spectrum Ohmic loss R ohm For η opt Make adjustments: ； is the reference impedance.

7. The method for optimizing a fuel cell system based on a neural network according to claim 1, characterized in that: S5 also includes: inputting real-time data into the multi-scale digital twin, when the electrochemical deviation δ echem =‖Z act −Z sim ‖ / ‖Z sim ‖>10% or thermodynamic deviation δ thermo =|T act −T opt ∣>5K, trigger S3 control parameter re-optimization; Z act is the actual impedance spectrum, Z sim To simulate the impedance spectrum, T act is the real-time temperature measurement value, T opt is the temperature target value.

8. The method for optimizing a fuel cell system based on a neural network according to claim 7, characterized in that: In S4, the pre-training phase of meta-reinforcement learning: the meta-policy network is trained using the historical failure dataset, and the loss function is L meta =0.8L task +0.2L robust , where L task is the control error, L robust To resist interference loss; The action space of meta-reinforcement learning is C={Q opt ,P opt ,T opt }, the state space is s=[F enhanced ,δ echem ,δ thermo ], the multi-objective reward function R is defined as: ； is the real-time pressure measurement value, w1, w2, and w3 are dynamically updated through Pareto front analysis, with the constraints that the sum of the weights is 1 and each weight is greater than 0.1; According to the state space s, the control parameters are dynamically adjusted, and the strategy π(s) maps the state to the action space, that is: ； in is the incremental adjustment value of the control parameter, Q opt The incremental adjustment value of P opt The incremental adjustment value of T opt The incremental adjustment value of C and Combined, the control instructions are generated, that is, the strategy π(s) is: ； Q final is the flow target value, P final is the pressure target value, T final is the temperature target value.

9. The method for optimizing a fuel cell system based on a neural network according to claim 7, characterized in that: Multi-scale digital twins include: Microscale: Molecular dynamics simulations were performed using the ReaxFF reaction force field with a time step of 0.1 fs and input Q opt , output the simulated impedance spectrum Z of the catalyst surface reaction sim ; Macro scale: Establish a multi-physics field coupling model using finite element software, with a mesh size of ≤0.1 mm and input T opt , output temperature field distribution T sim ; when When , S3 control parameters are re-optimized; When thermodynamic deviation , adjust w2=w2+0.

1.

10. The method for optimizing a fuel cell system based on a neural network according to claim 1, characterized in that: The actuator includes a hydrogen circulation pump, a coolant flow valve and a voltage regulator.

Images