A lithium iron phosphate battery thermal management optimization method and system based on wide-temperature modeling

Abstract

The application discloses a lithium iron phosphate battery thermal management optimization method and system based on wide-temperature modeling, and relates to the technical field of neural networks. The lithium iron phosphate battery thermal management optimization method based on wide-temperature modeling collects current operating state data of a lithium iron phosphate battery, constructs a time sequence input sequence and calculates a wide-temperature state vector; a wide-temperature adaptive temperature field perception network is input, a dynamic network weight and a temperature change sensitive thermal characteristic parameter are output by a conditional parameter generation module, and an attention correction prediction module takes the parameters as parameters, performs correction of fusion of priori and data on the time sequence input sequence, and generates an internal temperature field prediction sequence. The application forms a decision state vector by combining the internal temperature field prediction sequence and the wide-temperature state vector, inputs a reinforced thermal management strategy network, and generates a thermal management optimization control quantity after long-term and short-term safety constraint mapping processing, so that the accuracy of battery internal temperature field estimation under wide-temperature working conditions and the timeliness of thermal management control decision are improved.

Description

Technical Field

[0001] This invention relates to the field of neural network technology, specifically to a thermal management optimization method and system for lithium iron phosphate batteries based on wide-temperature modeling. Background Technology

[0002] Lithium iron phosphate batteries are widely used in electric vehicles and energy storage power stations due to their high safety and cycle life. In actual operation, the battery system needs to adapt to a wide temperature range. At low temperatures, the battery internal resistance increases and the usable capacity decreases. At high temperatures, high-rate charging and discharging will accelerate capacity decay. Therefore, efficient thermal management is a key link to ensure battery performance and safety.

[0003] The goal of battery thermal management is to adjust the heating or cooling power according to the battery's thermal state so that the battery can operate within a suitable temperature range. In order to obtain richer information about the internal thermal state of the battery, a modeling method based on the battery's electrothermal coupling characteristics can be used to deduce the internal temperature distribution of the battery from limited measurement information by establishing a mathematical model of battery heat generation and heat transfer.

[0004] The limitations of existing technologies include at least the following problems: In existing technologies, the internal temperature field of lithium iron phosphate batteries under wide temperature range conditions is difficult to measure directly. It is usually estimated indirectly by relying on the temperature of a limited number of measurement points on the surface. In actual operation, the battery's thermal properties and electrothermal coupling parameters undergo significant nonlinear changes with the operating temperature range, load state, and aging degree. For example, at low temperatures, the increased electrolyte viscosity leads to a sharp increase in internal resistance; at high temperatures and high-rate charge / discharge, the heat generation rate and heat transfer path inside the battery differ from those at room temperature. When the battery frequently switches operating states over a wide temperature range, simplified models using fixed or segmented parameters struggle to accurately track the dynamic changes in the battery's internal temperature field across the entire temperature range. This leads to biases in thermal management control decisions based on the model's estimation results. Such biases can lead to insufficient estimation of the internal temperature difference in low-temperature heating scenarios and lag in response to the core temperature rise rate in high-temperature fast charging scenarios, thus affecting the reliability of the thermal management system across the entire wide temperature range. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a thermal management optimization method and system for lithium iron phosphate batteries based on wide-temperature modeling, which solves the problem that existing technologies struggle to accurately estimate the internal temperature field of batteries over a wide temperature range.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling, comprising the following steps: collecting current operating state data of lithium iron phosphate batteries, constructing a time-series input sequence, and calculating a wide-temperature state vector; inputting the wide-temperature state vector and the input sequence together into a pre-trained wide-temperature adaptive temperature field sensing network, including a conditional parameter generation module and an attention correction prediction module; the conditional parameter generation module takes the wide-temperature state vector as input, extracts temperature region features through its internal shared feature encoding layer, and outputs dynamic network weights and temperature-sensitive thermal characteristic parameters in parallel through a weight generation branch and a thermal characteristic parameter generation branch, respectively; The attention correction prediction module uses dynamic network weights as computational parameters and, guided by temperature-sensitive thermal characteristic parameters, performs correction processing on the time-series input sequence, fusing prior trends and data-driven information, to generate an internal temperature field prediction sequence. Based on the internal temperature field prediction sequence, the internal core temperature and the maximum internal temperature difference are determined, and together with the wide-temperature state vector, they form a decision state vector, which is input into the pre-trained enhanced thermal management strategy network. After long-term and short-term collaborative safety constraint mapping processing, thermal management optimization control quantities are generated.

[0007] Further, the specific steps for calculating the wide-temperature state vector are as follows: extract the temperature of multiple measuring points on the surface from the current operating state data, calculate the average surface temperature, and directly obtain the ambient temperature and current from the current operating state data; calculate the state of charge based on the current using the ampere-hour integration method; obtain historical battery charge-discharge state data from the most recent multiple charge-discharge cycles, calculate the average coulombic efficiency, static internal resistance increment, and cumulative throughput capacity, and concatenate them into an aging feature vector; concatenate the average surface temperature, ambient temperature, state of charge, current, and aging feature vector to obtain the wide-temperature state vector.

[0008] Furthermore, the specific structure of the conditional parameter generation module is as follows: the shared feature encoding layer consists of multiple fully connected layers, with the input being a wide-temperature state vector and the output being a temperature region feature vector; the weight generation branch takes the temperature region feature vector as input, and outputs dynamic network weights through the first fully connected layer, which are used to reconstruct the attention projection matrix and feedforward network weights in the attention correction prediction module; the thermal characteristic parameter generation branch takes the temperature region feature vector as input, and outputs temperature change-sensitive thermal characteristic parameters through the second fully connected layer, which consist of the entropy change coefficient, temperature correction factor, and equivalent convective heat transfer coefficient.

[0009] Furthermore, the specific steps of the attention correction prediction module in performing correction processing that integrates prior trends and data-driven information are as follows: The encoder uses the reconstructed attention projection matrix and feedforward network weights as operational parameters to extract features from the time-series input sequence, generate a hidden state sequence, and map the hidden state sequence to virtual measurements; the thermal equilibrium prior generation layer uses the equivalent convective heat transfer coefficient as the convective term coefficient to perform thermal equilibrium recursion and obtain the thermal equilibrium prior; the prior feature enhancement encoder uses the thermal equilibrium prior and temperature-sensitive thermal characteristic parameters as input, processes them through learnable nonlinear feature mapping, and outputs enhanced prior features; the attention correction layer uses the enhanced prior features as the query vector, the virtual measurements as the key vector and value vector, calculates the correction amount through a scaled dot product attention mechanism; and adds the correction amount to the enhanced prior features to obtain the internal temperature field estimate for the current time step, and recursively generates the internal temperature field prediction sequence starting from this.

[0010] Further, the training steps of the wide-temperature adaptive temperature field sensing network are as follows: A training sample set consisting of multiple sets of samples is obtained, each set containing the true temperature field of each internal node; all parameters of the thermal characteristic parameter generation branch are frozen, and the mean square error between the internal temperature field prediction sequence output by the attention correction prediction module and the true temperature field of each internal node is used as the data fitting loss; after the data fitting loss converges, all parameters of the shared feature encoding layer, weight generation branch, and attention correction prediction module are frozen, and the SGD optimizer is initialized separately for the thermal characteristic parameter generation branch, and the optimizer state is reset, with the sum of energy conservation loss and thermal diffusion residual loss used as the constraint loss for fine-tuning; after the constraint loss converges, all parameters are unfrozen, and the weighted sum of the data fitting loss and constraint loss is used as the total loss, and joint fine-tuning is performed using a learning rate smaller than that of the previous training, resulting in the wide-temperature adaptive temperature field sensing network.

[0011] Furthermore, the energy conservation loss is calculated as follows: for each time step in the internal temperature field prediction sequence, the temperature change rate at that time step is calculated; the heat generation power at the current time step is calculated according to the heat generation equation; the heat dissipation power is calculated according to the thermal management control variables and wide-temperature state vector in the time-series input sequence; and the temperature change rate, heat generation power, and heat dissipation power at all time steps are comprehensively processed to obtain the energy conservation loss. ,Right now: ;in, Let be the rate of temperature change at time step t+h. Let be the heat generation power at time step t+h. Let be the heat dissipation power at time step t+h. For the preset cell quality, This is the preset equivalent specific heat capacity.

[0012] Furthermore, after the wide-temperature adaptive temperature field sensing network is trained, an online adaptive correction step is also included: The Fisher information diagonal value of each parameter in the wide-temperature adaptive temperature field sensing network is calculated using the training sample set; when the lithium iron phosphate battery is detected to be in a static state or the current is continuously less than a preset threshold, natural variation data is collected, and gradient updates are performed only on the bias vector of the second fully connected layer of the thermal characteristic parameter generation branch; the loss function optimized by the gradient update includes energy conservation loss and an anti-forgetting penalty term; the anti-forgetting penalty term uses the Fisher information diagonal value as the weight of each parameter, and applies a penalty to the deviation of the bias term from the training convergence. Defined as: ;in, The bias vector of the second fully connected layer of the generation branch only contains thermal characteristic parameters. The set of indexes for the parameters. The current parameter value. These are the parameter values ​​after training convergence.

[0013] Furthermore, the enhanced thermal management strategy network includes a strategy network, a reward function module, and a safety action filter. The long- and short-term collaborative safety constraint mapping is implemented collaboratively by the reward function module and the safety action filter. The specific steps are as follows: In the reward function module, the maximum internal temperature difference is extracted from the internal temperature field prediction sequence. And based on the preset maximum temperature difference penalty function Add a penalty for maximum internal temperature difference. The forms of expression are as follows: ;in , A preset safety limit is set; the safety action filter is connected after the policy network and has a built-in local agent layer; the local agent layer takes the decision state vector and the original action generated by the policy network as input and outputs the predicted maximum internal temperature difference; the safety action filter corrects the original action with the constraint that the predicted maximum internal temperature difference does not exceed the preset safety limit and outputs the thermal management optimization control quantity.

[0014] Further, the training steps for strengthening the thermal management strategy network are as follows: The parameters of the wide-temperature adaptive temperature field sensing network are frozen and used as the environment model. The current decision state vector is input into the strategy network, and the original action is output. The current decision state vector and the original action are simultaneously input into the safety action filter. The local agent layer outputs the predicted maximum internal temperature difference, corrects the original action, obtains the safety action, and inputs it into the environment model to obtain the maximum internal temperature difference. The maximum internal temperature difference extracted after executing the safety action on the environment model is used for online incremental fitting of the local agent layer to update its parameters. The reward value is calculated based on the maximum internal temperature difference. The current decision state vector, the original action, the reward value, and the decision state vector for the next time step are stored in the experience replay buffer. Samples are taken from the experience replay buffer to update the evaluation network parameters by minimizing the Bellman residual of the evaluation network, and to update the strategy network parameters by maximizing the Q-value of the strategy network under the evaluation network.

[0015] A thermal management optimization system for lithium iron phosphate batteries based on wide-temperature modeling includes: a data acquisition unit for acquiring current operating state data of the lithium iron phosphate battery, constructing a time-series input sequence, and calculating a wide-temperature state vector; a wide-temperature modeling unit for inputting the wide-temperature state vector and the input sequence into a pre-trained wide-temperature adaptive temperature field sensing network, including a conditional parameter generation module and an attention correction prediction module; and a conditional parameter generation unit for enabling the conditional parameter generation module to take the wide-temperature state vector as input, extract temperature region features through its internal shared feature encoding layer, and then output dynamic network weights and temperature-sensitive thermal characteristic parameters in parallel through a weight generation branch and a thermal characteristic parameter generation branch, respectively. The attention correction prediction unit enables the attention correction prediction module to use dynamic network weights as computational parameters and, guided by temperature-sensitive thermal characteristic parameters, to perform correction processing on the time-series input sequence, fusing prior trends and data-driven information, to generate an internal temperature field prediction sequence. The thermal management optimization unit determines the internal core temperature and the maximum internal temperature difference based on the internal temperature field prediction sequence, and together with the wide-temperature state vector, forms a decision state vector, which is input into a pre-trained enhanced thermal management strategy network. After long-term and short-term collaborative safety constraint mapping processing, it generates thermal management optimization control variables.

[0016] The present invention has the following beneficial effects: (1) The thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling uses a condition parameter generation module to enable the operation parameters of the attention correction prediction module to be adjusted synchronously with the real-time operating temperature range, load state and aging degree of the battery. The weight generation branch dynamically outputs network parameters according to the wide-temperature state vector. Under low-temperature conditions, it automatically adapts to the changes in heat transfer law caused by the change in operating conditions. Under high-temperature high-rate charge and discharge conditions, it can switch to the operation mode that matches the high heat generation rate and strong convection conditions. The thermal characteristic parameter generation branch outputs the entropy change coefficient temperature correction factor and equivalent convection heat transfer coefficient that change significantly in the wide temperature range, so that the network can maintain accurate reconstruction of the internal temperature field of the battery in the entire temperature range, avoiding estimation deviations at low-temperature or high-temperature ends.

[0017] (2) The thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling adopts a segmented decoupled training strategy in the training stage of the wide-temperature adaptive temperature field sensing network. The thermal characteristic parameter generation branch is frozen, and only the shared feature encoding layer, weight generation branch and attention correction prediction module are trained. This allows the network to first establish a basic temperature field prediction capability without being disturbed by physical constraints. After that, the trained part is frozen, and the optimizer is initialized separately for the thermal characteristic parameter generation branch. It does not inherit any momentum accumulation state and only uses energy conservation loss and thermal diffusion residual loss as targets for fine-tuning. Finally, all parameters are unfrozen, and the overall fine-tuning is completed with a small learning rate. This avoids mutual interference between the data fitting target and the thermodynamic constraint target during the gradient update process, so that the network can ultimately balance the accuracy of temperature field prediction and the rationality of thermal characteristic parameter output.

[0018] (3) The thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling constructs a dual safety constraint mechanism of soft and hard collaboration in the enhanced thermal management strategy network. It adds a penalty term for the maximum internal temperature difference in the reward function. When the internal temperature difference exceeds the preset limit, a penalty signal is generated. This guides the strategy network to gradually learn to avoid control strategies that are prone to causing internal overheating or core temperature rise overshoot during long-term training. At the same time, a safety action filter is deployed at the output end of the strategy network. A local agent layer is built in. The training samples are generated online by the wide-temperature adaptive temperature field sensing network during the strategy training process. The filter aims to minimize the deviation of the action before and after correction and is constrained by the fact that the internal temperature difference predicted by the local agent layer does not exceed the safety limit. After correcting the original action, the filter outputs the actual control quantity executed. This enables the thermal management control decision to have the ability to respond to the potential risks caused by excessive internal temperature difference in real time while pursuing temperature tracking and energy efficiency optimization.

[0019] (4) The thermal management optimization system for lithium iron phosphate batteries based on wide-temperature modeling organizes the current operating state data into a time-series input sequence and calculates the wide-temperature state vector through the data acquisition unit. The condition parameter generation module dynamically generates the operation parameters required by the attention correction prediction module according to the temperature range and aging information in the wide-temperature state vector. This enables the attention correction prediction module to adaptively match the heat conduction and heat generation characteristics under the current operating conditions during the inference stage. After outputting the internal temperature field prediction sequence, the thermal management optimization unit concatenates the predicted internal core temperature and the maximum internal temperature difference with the observable state into a decision state vector. The optimized thermal management control quantity is output through the enhanced thermal management strategy network, thereby improving the accuracy of the battery internal temperature field estimation and the timeliness of thermal management control decision under wide-temperature operating conditions.

[0020] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0021] Figure 1 This is a flowchart of a thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling, according to the present invention.

[0022] Figure 2 This is a flowchart illustrating the specific steps of the attention correction prediction module in the thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling of the present invention, which performs correction processing that integrates prior trends and data-driven information.

[0023] Figure 3 This is a block diagram of a thermal management optimization system for lithium iron phosphate batteries based on wide-temperature modeling, according to the present invention. Detailed Implementation

[0024] Please see Figure 1 This invention provides a technical solution: a thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling, comprising the following steps: collecting current operating state data of the lithium iron phosphate battery (including thermal management control variables, ambient temperature, current, terminal voltage, and temperature of multiple measuring points on the surface of the lithium iron phosphate battery), constructing a time-series input sequence containing multiple consecutive time steps including the current time step, and calculating a wide-temperature state vector; inputting the wide-temperature state vector and the input sequence together into a pre-trained wide-temperature adaptive temperature field sensing network, which includes a conditional parameter generation module and an attention correction prediction module; The conditional parameter generation module takes a wide-temperature state vector as input, extracts temperature region features through its internal shared feature encoding layer, and outputs dynamic network weights and temperature-sensitive thermal characteristic parameters in parallel through a weight generation branch and a thermal characteristic parameter generation branch, respectively. Among them, the thermal characteristic parameter generation branch only outputs two parameters: entropy change coefficient temperature correction factor and equivalent convective heat transfer coefficient. The battery's equivalent specific heat capacity and thermal conductivity are based on preset calibration constants. The attention correction prediction module uses dynamic network weights as computational parameters and, guided by the temperature-sensitive thermal characteristic parameters, performs correction processing on the time-series input sequence, fusing prior trends and data-driven information, to generate an internal temperature field prediction sequence for multiple future time steps. Based on the internal temperature field prediction sequence, the internal core temperature and the maximum internal temperature difference are determined, and together with each state component in the wide-temperature state vector, they form a decision state vector. This vector is then input into a pre-trained enhanced thermal management strategy network and processed through long-term and short-term collaborative safety constraint mapping to generate thermal management optimization control quantities.

[0025] Specifically, the steps for calculating the wide-temperature state vector are as follows: The temperature of multiple measuring points on the surface of the lithium iron phosphate battery is extracted from the current operating status data, the average surface temperature is calculated, and the ambient temperature and current are directly obtained from the current operating status data. Specifically: Let M temperature measuring points be arranged on the surface of the battery. The temperature of the m-th measuring point at the current time step t is... The average surface temperature Calculate using the following formula: ; Ambient temperature The current is read directly from the ambient temperature sensor located inside the battery pack. The terminal voltage is directly acquired by a current sensor. The voltage is directly collected by a voltage sensor, with positive values ​​taken in the charging direction and negative values ​​taken in the discharging direction. Based on current, the state of charge is calculated using the ampere-hour integration method, specifically as follows: State of charge The update recursive formula is: ; in, For the sampling time step, This is the preset rated capacity of the lithium iron phosphate battery; This is the state of charge value from the previous time step; Obtain historical battery charge / discharge state data (including charging and discharging capacity, terminal voltage, and current) from the most recent multiple charge / discharge cycles. Calculate the average coulombic efficiency, static internal resistance increment, and cumulative throughput capacity. Concatenate these three data points into an aging feature vector, specifically: Recently In one charge-discharge cycle ( The value of is in the range of 5 to 30, preferably 10), and the charging capacity of the j-th charging process is . The discharge capacity of the j-th discharge process is Then the average Coulomb efficiency for: ; Increment of static internal resistance The calculation method is as follows: Extract the most recent moment when the system transitioned from charging or discharging to resting state from historical operating data, and obtain the terminal voltage of the last sampling point before the transition. and current And the terminal voltage of the first sampling point after settling. and current Estimate the ohmic resistance during this resting period using the following formula: ; This estimate The initial ohmic internal resistance specified by the battery at the factory The difference is used as the increment of the static internal resistance. ; Cumulative throughput For near The sum of the discharge capacities in each charge-discharge cycle; The above three quantities are concatenated to form an aging feature vector. ; By concatenating the surface average temperature, ambient temperature, state of charge, current, and aging characteristic vectors, a wide-temperature state vector is obtained. .

[0026] Specifically, the structure of the condition parameter generation module is as follows: The shared feature coding layer consists of multiple fully connected layers, employing the Swish activation function. The input is a wide-temperature state vector, and the output is a temperature-region feature vector. Specifically: The shared feature coding layer consists of two sequentially connected fully connected layers, and the first weight matrix of these two sequentially connected layers is: The first bias vector is The second weight matrix is The second bias vector is ; The activation function used is the Swish function. ; Wide-temperature state vector After forward propagation through the shared feature coding layer, the output temperature region feature vector is obtained. : ; Temperature region eigenvector The comprehensive characteristics of the battery's current temperature range, load intensity, and aging state are encoded for use by the two subsequent branches; The weight generation branch takes the temperature region feature vector as input, outputs dynamic network weights through the first fully connected layer, and is used to reconstruct the attention projection matrix of the encoder and the feedforward network weights in the attention correction prediction module. Specifically: The weight matrix of the first fully connected layer is: The bias vector is This branch outputs a dynamic network weight parameter vector. : ; Will The query projection matrix of each attention layer of the Transformer encoder in the attention correction prediction module is broken down as follows: Key projection matrix Value projection matrix and the weight matrix of the feedforward network and the corresponding bias vector: set up The total dimension is D, and the encoder has a total of Each layer requires attention and reconstruction of the projection matrix. and feedforward weights and the corresponding bias vector, where For the encoder hidden layer dimension, For the attention head dimension, This represents the intermediate layer dimension of the feedforward network.

[0027] Will Sort by floor number from smallest to largest, and within the same floor by... And the corresponding bias order, sequentially extract the elements of the corresponding dimension and assign values ​​until all parameters are assigned, so as to achieve temperature-zone adaptive reconstruction of the encoder feature extraction capability; The thermal characteristic parameter generation branch takes the temperature region feature vector as input and outputs temperature-sensitive thermal characteristic parameters through the second fully connected layer. These parameters consist of the entropy change coefficient, temperature correction factor, and equivalent convective heat transfer coefficient. The equivalent specific heat capacity and thermal conductivity are calibration constants of the same type of lithium iron phosphate battery and are not involved in online identification. Specifically: The weight matrix of the second fully connected layer is: The bias vector is The output of this branch is a vector of temperature-sensitive thermal characteristic parameters. : ; in, This is a temperature correction factor for the entropy change coefficient, used to correct the battery entropy change coefficient. Values ​​at different temperatures; It is the equivalent convective heat transfer coefficient, used to characterize the intensity of the influence of micro-convection inside the electrolyte on the temperature field distribution. Under low temperature conditions, its value automatically approaches zero, degenerating into a pure heat conduction process. Equivalent specific heat capacity With equivalent thermal conductivity Using the same model of lithium iron phosphate battery, the factory calibration values ​​are kept constant during the operation of the wide-temperature adaptive temperature field sensing network.

[0028] like Figure 2 As shown, the attention correction prediction module includes an encoder, a thermal equilibrium prior generation layer, a prior feature enhancement encoder, and an attention correction layer. The specific steps of the attention correction prediction module in performing correction processing that integrates prior trends and data-driven information are as follows: The encoder uses the reconstructed attention projection matrix and feedforward network weights as operational parameters to extract features from the time-series input sequence, generating a hidden state sequence. This hidden state sequence is then mapped to virtual measurements, which represent data-driven estimations of the temperature at various nodes within the battery. Specifically: Timing input sequence It contains data from the current time step t and the data from the previous N-1 time steps, for a total of N time steps, with the input vector for each time step. , The actual thermal management control quantity executed at time t; The encoder uses the reconstructed attention projection matrix and feedforward network weights as computational parameters to perform multi-head self-attention computation and feedforward network mapping on the temporal input sequence, and outputs a hidden state sequence. ; the hidden state of the last time step It is mapped to a virtual measurement vector through a lightweight projection network consisting of a single fully connected layer. ; The weight matrix of the fully connected layer is: The bias vector is No activation function is used, where Hidden state The dimension is L, which is the number of internal intermediate layer nodes discretized along the battery thickness direction; Each element of the virtual measurement vector corresponds to the data-driven core temperature estimate, the L internal intermediate layer temperature estimates, and the surface temperature estimate, respectively. The encoder performs forward computation only once on the historical time-series input sequence, and outputs a fixed virtual measurement vector. and through Generate a fixed key matrix K and value matrix V; In the attention correction at each subsequent future time step, the fixed key matrix and value matrix are reused and are not calculated again. The length N of the input sequence ranges from 16 to 64, preferably 32; the number of internal intermediate layer nodes... The value ranges from 2 to 5, with 3 being the preferred value; The thermal equilibrium prior generation layer uses the surface multi-point temperature at the end of the time step of the time-series input sequence as the initial temperature distribution, and the equivalent convective heat transfer coefficient in the temperature-sensitive thermal characteristic parameters as the convection term coefficient to perform thermal equilibrium recursion and obtain the thermal equilibrium prior. In the recursive generation process of subsequent time steps, the surface node temperature in the internal temperature field estimation is used instead of the surface multi-point temperature as the starting point for recursion, specifically as follows: During the initial generation, for all nodes (internal nodes) and surface nodes ), its previous time step temperature All are initialized to the measured average surface temperature at the current time step; Combining the simplified thermal model of the battery with discretized nodal equations along the thickness direction, using the equivalent convective heat transfer coefficient To perform a one-step recursion on the convection term coefficients, we obtain an L+2 dimensional physical prior vector. ; The simplified one-dimensional discrete recurrence equation of the thermal model is: For internal nodes: ; For core nodes Adiabatic boundary conditions are adopted, and virtual nodes are introduced. Substituting into the above equation, we get: ; For surface nodes, a recursive equation for convective heat transfer boundary conditions is used, and an active thermal management heat flux term is introduced: ; in For active thermal management heat flux density, a positive value indicates that the thermal management system transfers heat to the battery surface (heating condition), and a negative value indicates that the thermal management system removes heat from the battery surface (heat dissipation condition). Its value is controlled by the thermal management control quantity. Pre-calibrated mapping function Determined, that is ; r is the convective heat transfer coefficient of the battery surface, which is the calibration value of the same type of battery and is kept as a constant during the training and online inference phases; in, For the previous time step Temperature of each node For cell density, For equivalent specific heat capacity, For equivalent thermal conductivity, For discrete node spacing, Let be the heat generation rate per unit volume of the i-th node; Total heat production Based on the current, terminal voltage, and average temperature of each node in the previous time step, the data is calculated using the Bernardi heat generation equation through a preset open-circuit voltage function and entropy change coefficient function, and then calculated according to the cell volume. Distribute evenly to each node, i.e. For all nodes; To address the computational cyclic dependency caused by the coupling between heat generation and temperature in temperature field prediction, at prediction time step t, the internal average temperature required for calculating its heat generation power is approximated using the known predicted value from the previous time step t-1. Due to the time step... The error caused by this lag is very small and negligible; When recursively generating future time steps, the control quantity, current and ambient temperature of each future step are all assumed to be zero-order hold, that is, the measured value of the current time step is taken as the corresponding value of each future step. For the terminal voltage at each future step The following is obtained from online estimation using the current open-circuit voltage and internal resistance voltage drop: ,in , It is obtained by recursion using the zero-order held current through the ampere-hour integration method. Under the zeroth-order hold assumption, it is equal to the current. The prior feature enhancement encoder takes thermal balance priors, thermal management control variables, and temperature-sensitive thermal characteristic parameters as inputs, and processes them through learnable nonlinear feature mapping to output enhanced prior features, specifically: The prior feature enhancement encoder is a small fully connected network whose input is a thermal equilibrium prior vector. Thermal management control quantity and temperature-sensitive thermal characteristic parameters The concatenated vector; After passing through a fully connected layer and the Swish activation function for mapping, the output is an enhanced prior query vector. Its shape is a row vector, with the same dimensions as the internal temperature field vector, and it can be directly used as a physical prior initial estimate of the temperature field: ; in, The weight matrix and bias vector of the encoder are enhanced by prior features; The attention correction layer uses enhanced prior features as the query vector and virtual measurements as the key and value vectors. It calculates the correction amount through a scaled dot product attention mechanism, specifically as follows: In the attention correction layer, the enhanced prior query vector is used as the query, and the virtual measurement vector is used as the source of the key and value. The dynamic projection matrix is ​​generated by the condition parameter generation module. and Map queries and key-value pairs to the same space: Query mapping: ; Key mapping: ; Value mapping: ; Among them, the key vector dimension The value range is 16 to 128, preferably 64; value vector dimension The value ranges from 16 to 128, with 64 being the preferred value; Calculate the scaled dot product attention weight matrix : ; in The dimension is This indicates an enhancement in the matching degree between the prior query vector and the virtual measurement vector; is the dimension of the key vector, used to scale the inner product to prevent gradient saturation of the softmax function; The correction amount is obtained by weighted summation of the attention weight matrix and the value matrix: ; The correction amount is then added to the enhanced prior features to obtain the internal temperature field estimate for the current time step. This estimate is then used as a starting point to recursively generate a sequence of predicted internal temperature fields for multiple future time steps. Specifically: The dimension of the correction amount is By outputting the projection matrix Map it back to the dimension L+2 of the temperature field vector, and then combine it with the enhanced prior features. Adding them together gives the estimated internal temperature field at the current time step. : ; in This is the trainable projection matrix inherent within the attention correction layer, not included in the dynamic network weights output by the conditional parameter generation module. In the training phase, it is optimized along with other parts of the network; Estimate the internal temperature field after correction at the current time step As the starting point of the recursion; for the h-th step in the future ( ),by Starting from the thermal equilibrium prior generation layer, physical priors are recursively obtained. The encoder is enhanced by prior features. Mapped to enhanced prior query vector ; In the attention correction layer, with To query and reuse the fixed key and value matrices generated from the historical sequence at time t, perform a scaled dot product attention calculation to obtain the correction amount and combine it with... Adding them together gives ; Repeat this process to generate a predicted sequence of internal temperature fields for the next H consecutive time steps. ; During the recursive generation process, the control quantity, current and ambient temperature of each future step are assumed to be zero-order hold, and the measured values ​​of the current time step are taken to remain unchanged. The terminal voltage at each future time step is estimated online using the open-circuit voltage and internal resistance voltage drop; the temperature field vector at each time step is determined by the core temperature. The temperatures of L internal intermediate layers equidistantly distributed along the thickness direction of the battery to And it is assembled in sequence according to surface temperature; The prediction time-domain step number H ranges from 5 to 30, preferably 10.

[0029] The training steps for the wide-temperature adaptive temperature field sensing network are as follows: A training sample set consisting of multiple sets of samples is obtained. Each set of samples contains a wide-temperature state vector synchronously acquired under the full-temperature-range operating conditions of the lithium iron phosphate battery, a time-series input sequence, and the actual temperature field of each internal node obtained by the built-in thermocouple. Specifically: Training sample set The results were obtained through experiments on lithium iron phosphate batteries covering a full temperature range, including low temperature, room temperature, and high temperature. During the experiments, L+2 miniature thermocouples were pre-embedded along the thickness direction inside the battery, corresponding to the surface measuring points, and the core temperature, the temperature of each internal intermediate layer, and the surface temperature were recorded simultaneously as labels for the real temperature field. ; All parameters of the freezing thermal property parameter generation branch are used, with the mean square error between the internal temperature field prediction sequence output by the attention correction prediction module and the actual temperature field of each internal node as the data fitting loss. The shared feature encoding layer, weight generation branch, and attention correction prediction module are trained using the Adam optimizer, as follows: Before this stage begins, the bias vector of the second fully connected layer of the thermal characteristic parameter generation branch is used. It is initialized to a specific value, such that the initial output of this branch is and That is, the initial value of the entropy change coefficient temperature correction factor is 1 (no correction), and the initial value of the equivalent convective heat transfer coefficient is 0 (degenerates into a pure heat conduction model). Under this initialization condition, all parameters of the thermal characteristic parameter generation branch are frozen. In this stage, only the parameters of the shared feature coding layer, the weight generation branch, and the attention correction prediction module are updated. The weights and biases of the thermal characteristic parameter generation branch remain frozen. Data fitting loss Defined as: ; in, This represents the number of samples collected in each training batch. This represents the sum of squared errors of each element; the number of samples in the batch. The value ranges from 16 to 256, with 64 being the preferred value; Initial learning rate of Adam optimizer The range of values ​​is to Preferred The preferred first-order momentum decay coefficient is 0.9, and the preferred second-order momentum decay coefficient is 0.999. After the data fitting loss converges, all parameters of the shared feature encoding layer, weight generation branch, and attention correction prediction module are frozen. A separate SGD optimizer is initialized for the thermal characteristic parameter generation branch, and all optimizer states are reset. Fine-tuning is performed using the sum of energy conservation loss and thermal diffusion residual loss as the constraint loss. Specifically: During this stage, only the parameters of the thermal characteristic parameter generation branch are updated, while the parameters of the shared feature encoding layer, weight generation branch, and attention correction prediction module remain frozen. A separate stochastic gradient descent optimizer (SGD) is initialized for the branch generating thermal characteristic parameters, with momentum coefficients... The value range is from 0.0 to 0.3, preferably 0.1. This SGD optimizer does not inherit any first-order momentum and second-order momentum accumulation states from the first-stage Adam optimizer. Constraint loss Defined as: ; in, Loss due to energy conservation This is the loss due to thermal diffusion residuals; For each time step in the internal temperature field prediction sequence, calculate the finite difference residual of the one-dimensional transient heat conduction equation along the battery thickness direction for the temperature of each internal node. Take the sum of the squares of the residuals of all time steps and all internal nodes as the heat diffusion residual loss. ,Right now: ; in, For the first The predicted temperature at the i-th internal node at time step. Let be the heat generation rate per unit volume of the i-th node in the previous time step; When the equivalent convective heat transfer coefficient When approaching zero, including The convection term automatically degenerates to zero, and the thermal diffusion residual degenerates to a pure thermal conduction constraint; Initial learning rate of SGD optimizer The range of values ​​is to Preferred ; After the constraint loss converges, all parameters are unfrozen, and the total loss is the weighted sum of the data fitting loss and the constraint loss. Joint fine-tuning is then performed using a learning rate smaller than that used in the previous training, resulting in the trained wide-temperature adaptive temperature field sensing network, as follows: After all parameters are unfrozen, the total loss function for: ; in This is the penalty coefficient, with values ​​ranging from 0.01 to 1.0. The preferred value is 0.1. The preferred value is 0.05; The optimizer was switched back to Adam, and the learning rate... The range of values ​​is to Preferred The learning rate is much lower than that in the first stage.

[0030] Specifically, the energy conservation loss is calculated as follows: For each time step in the internal temperature field prediction sequence, the average temperature of all nodes within that time step is calculated as the internal average temperature. The rate of change of this internal average temperature relative to the internal average temperature of the previous time step is then calculated, specifically as follows: internal average temperature The result is obtained by taking the arithmetic mean of the core temperature, L internal intermediate layer temperatures, and surface temperature in the predicted sequence: ; Rate of temperature change It is obtained by dividing the difference between the internal average temperature of the current time step and the previous time step by the time step length; The heat generation power at the current time step is calculated based on the Bernardi heat generation equation, where the open-circuit voltage and entropy change coefficient are obtained through two pre-defined spline functions. and The spline function is obtained by fitting experimental data of the same type of lithium iron phosphate battery over a wide temperature range. It takes the state of charge z as input and outputs a smooth open-circuit voltage value or entropy change coefficient value. The parameters remain unchanged during the training and online operation of the wide-temperature adaptive temperature field sensing network. Specifically: Heat generation power The following is given by Bernardi's heat generation equation: ; in, This is the open-circuit voltage corresponding to the state of charge z; is the entropy change coefficient corresponding to the charged state z; This is the estimated terminal voltage value for this time step, which is obtained from the online estimation formula when predicting future time steps; Internal average temperature required for heat generation power calculation Using the previous time step The predicted values ​​are approximated with lag to avoid circular dependence in temperature field prediction; The heat dissipation power is calculated based on the thermal management control variables in the time-series input sequence and the difference between the average surface temperature and the ambient temperature in the wide-temperature state vector. Specifically: Heat dissipation power Thermal management control quantity And the difference between surface temperature and ambient temperature is determined: ; Where r is the convective heat transfer coefficient of the battery surface, and A is the preset effective heat dissipation area of ​​the battery. For active thermal management heat flux density, a positive value indicates that the thermal management system transfers heat to the battery surface (heating condition), and a negative value indicates that heat is removed from the battery surface (heat dissipation condition). The rate of temperature change, heat generation power, and heat dissipation power are comprehensively processed to obtain the energy conservation residual for the corresponding time step. The energy conservation residuals for all time steps are then summed to obtain the energy conservation loss, i.e.: ; in For cell quality, For equivalent specific heat capacity, the calibration constant of the same type of lithium iron phosphate battery is used.

[0031] After the wide-temperature adaptive temperature field sensing network is trained, an online adaptive correction step is also included: After the training of the wide-temperature adaptive temperature field sensing network converges, the Fisher information diagonal value of each parameter in the network is calculated using the training sample set and saved. Specifically: For the nth parameter in the network Its Fisher information diagonal value Calculate using the following formula: ; in, The total number of samples in the training sample set. For the total loss function with respect to parameters Gradient; Fisher information diagonal value Parameters were measured The importance of the old task (wide-range temperature field prediction) The larger the value, the more important the parameter. The diagonal values ​​of Fisher information for all parameters are saved and deployed to the battery management system along with the trained wide-temperature adaptive temperature field sensing network. During the operation of the lithium iron phosphate battery, when the battery is detected to be in a static state or the current is continuously less than a preset threshold, natural temperature change data at multiple measuring points on the surface are collected, and bias vectors for the second fully connected layer of the branch are generated only for thermal characteristic parameters. Gradient updates are performed, specifically as follows: The condition for determining a static state is: Among them, the preset current threshold The value range is from 0.02C to 0.1C, preferably the current value corresponding to 0.05C; When the time during which the conditions are met continuously exceeds a preset time window, the natural temperature variation sequence of M measuring points on the surface within that window is collected as online fine-tuning data. ; Only the bias vector of the second fully connected layer of the branch is generated for the thermal characteristic parameters. Perform gradient updates and freeze all other parameters; The loss function optimized by gradient update consists of the surface temperature prediction error (i.e., the mean square error between the predicted and measured surface temperatures), energy conservation loss, a time-series smoothing regularization term composed of the square of the maximum temperature difference within adjacent time steps, and an anti-forgetting penalty term. Specifically: Online loss function Defined as: ; in, The mean square error between the predicted and measured surface temperature values. Loss due to energy conservation The time-series smoothing regularization term is formed by the square of the maximum temperature difference within adjacent time steps. This is a penalty item to combat forgetting; This is an adjustment factor for online energy conservation losses, with a value ranging from 0.01 to 1.0, preferably 0.1; This is the adjustment coefficient for the time-series smoothing regularization term, with a value ranging from 0.1 to 10, preferably 0.1; The adjustment coefficient for the anti-forgetting penalty term ranges from 0.1 to 10, preferably 1.0; Since the open-circuit voltage and entropy change coefficient required for energy conservation losses are provided by a preset fixed spline function, the online fine-tuning stage only involves adjusting the bias vector. The systematic bias in the temperature field prediction is corrected without updating the spline function itself, so as to maintain the basic electrochemical properties unchanged; The anti-forgetting penalty term uses the diagonal value of the Fisher information as the weight of each parameter, and applies a penalty to the deviation of the bias term from training convergence. Specifically: Anti-forgetting penalty Defined as: ; in, The bias vector of the second fully connected layer of the generation branch only contains thermal characteristic parameters. The set of indexes for the parameters. The current parameter value. The parameter values ​​after training convergence; The anti-forgetting penalty term applies stronger pull-back constraints to parameters with large Fisher information values, enabling the model to maintain its learned wide-temperature-range temperature field reconstruction capability while adapting to aging drift.

[0032] Specifically, the enhanced thermal management strategy network includes a strategy network, a reward function module, and a safety action filter. The long-short-term collaborative safety constraint mapping is implemented collaboratively by the reward function module and the safety action filter, with the following specific steps: The reward function module extracts the maximum internal temperature difference from the predicted sequence of the internal temperature field, and adds a penalty term for the maximum internal temperature difference based on the preset maximum temperature difference penalty function. Specifically: Maximum internal temperature difference Extracted from the internal temperature field prediction sequence output by the wide-temperature adaptive temperature field sensing network, the value is the maximum absolute value of the difference between the core temperature and the temperature of each internal intermediate layer: ; Define the maximum temperature difference penalty function A piecewise nonlinear approach is adopted to adaptively increase the penalty intensity according to the degree of exceeding the limit: ; in , To pre-determine safety limits; The reward function R is composed of a weighted sum of temperature deviation penalty, core over-limit penalty, internal temperature difference penalty, and energy consumption penalty, all weighted by normalized weighting coefficients. ; in, The target operating temperature ranges from 20°C to 35°C, with 25°C being preferred. The core temperature safety upper limit is set between 50°C and 60°C, preferably 55°C. To establish a preset safety limit, the value range is 5°C to 15°C, preferably 10°C; The total power consumption of the thermal management system under control is expressed as a function of the active thermal management heat flux density. The mapping relationships were all pre-calibrated by the thermal management system bench experiment; The normalized baseline value for temperature-related penalty items is taken from a preset safety limit. ; The normalized reference value for the power consumption penalty term is the rated power of the thermal management system. These are normalized weighting coefficients that sum to 1. Their values ​​are allocated according to the importance of each item. Preferably, While ensuring temperature tracking accuracy, priority is given to ensuring core safety and internal temperature difference safety. The maximum internal temperature difference penalty term generates a penalty when the internal temperature difference exceeds the preset safety limit, guiding the policy network to actively avoid dangerous control strategies that may generate hidden hot spots or core overshoot in long-term decision-making. The security action filter is connected after the policy network and has a built-in local proxy layer, which is as follows: The security action filter is located at the output of the policy network and intercepts and corrects the original actions generated by the policy network in real time. The local agent layer is a fully connected network with two hidden layers, each containing 64 neurons and the activation function being the Swish function. The output layer is a single neuron with no activation function, taking the decision state vector and any candidate action as input and outputting the prediction of the maximum internal temperature difference. The parameters are updated online synchronously during the training phase of the enhanced thermal management strategy network. The local agent layer takes the decision state vector and the original actions generated by the policy network as input, and outputs a prediction of the maximum internal temperature difference, specifically: Decision state vector The internal core temperature (the temperature field vector at each time step in the internal temperature field prediction sequence is formed by splicing the core temperature, the temperatures of multiple internal intermediate layers that are equidistantly distributed along the battery thickness direction, and the surface temperature in a spatial order from the inside to the outside, and the internal core temperature is directly read from the position of the first element of the vector) and the internal maximum temperature difference in sequence. The policy network is a fully connected network with two hidden layers, each containing 256 neurons. The activation function is the Swish function. The output layer contains three neurons, corresponding to heating power, coolant pump speed, and compressor speed, respectively. The activation function is the Tanh function to limit the output to […]. Within the range of 1, 1]; Policy networks with As input, output the original action vector. These correspond to heating power, coolant pump speed, and compressor speed, respectively; The local proxy layer is a fully connected network with two hidden layers, each containing 64 neurons and the activation function being the Swish function. The output layer is a single neuron with no activation function. Local proxy layer with and The concatenated vector is used as input, and the output is the predicted maximum internal temperature difference. ; The safety action filter, constrained by the prediction that the maximum internal temperature difference will not exceed a preset safety limit, corrects the original action and outputs a thermal management optimization control quantity, which is as follows: The corrected problem is constructed as the following constrained optimization problem: ; Since the local proxy layer is a nonlinear fully connected network, the constraint function... It is an action The nonlinear function is solved using the iterative gradient correction method to address this constrained optimization problem. With original action For the initial solution In the k-th iteration, if Then output the current action. For safety reasons; Otherwise, correct the action along the gradient direction of the constraint function with respect to the action: ; in The step size coefficient is adaptively adjusted according to the degree of constraint violation; the above iterations are repeated until the constraint is satisfied or the preset maximum number of iterations is reached, and the final action is projected back into the legal range of the action space to output the final safe action. This serves as the control quantity for thermal management optimization at the next moment.

[0033] The training steps for the enhanced thermal management strategy network are as follows: The parameters of the wide-temperature adaptive temperature field sensing network are frozen and used as the environment model. The current decision state vector is input into the policy network, and the original action is output, specifically as follows: Initialize the experience replay buffer Set its maximum capacity to 100 empirical tuples The range of values ​​is to Preferred ; Before training the enhanced thermal management strategy network, the wide-temperature adaptive temperature field sensing network has been trained, and all its parameters are frozen as an environment model that cannot be updated. Each time step t of the input time series of the environmental model includes the measured surface temperature, terminal voltage, current, ambient temperature, and the safety control quantity finally executed at time t. ; For the current time step t, the control measures the safety actions resulting from this decision. The future step control variable adopts the zero-order hold assumption; At each step of training, the current decision state vector is input into the policy network, and the policy network outputs the original action vector. The current decision state vector and the original action are simultaneously input into the safety action filter. The local agent layer outputs the predicted maximum internal temperature difference. The safety action filter uses an iterative gradient correction method to correct the original action, obtaining the safety action, which is then input into the environmental model to obtain the predicted internal temperature field sequence for the next time step. The maximum internal temperature difference is extracted from this sequence. Specifically: Will Simultaneously input the security action filter, and output the local proxy layer. The safety action filter is solved using the iterative gradient correction method to obtain the safety action. ; Will Input a frozen, wide-temperature adaptive temperature field sensing network environment model. The environment model outputs the predicted internal temperature field sequence for the next time step, from which the maximum internal temperature difference is extracted. ; After performing safety actions on the environmental model, the maximum internal temperature difference is extracted. Online incremental fitting is then performed on the local proxy layer to update its parameters. Specifically: Using the actual safety actions performed and the current decision state vector as inputs to the local agent layer, and the maximum internal temperature difference extracted from the environmental model output as the supervision label, the mean squared error loss is calculated: ; The loss is used to perform a one-step gradient update on the parameters of the local agent layer, enabling it to gradually learn the mapping relationship between safety actions and internal temperature differences. The learned mapping relationship is then extended to the unseen action space through the generalization ability of the fully connected network, achieving online incremental fitting. The reward function module calculates the reward value based on the maximum internal temperature difference, and stores the current decision state vector, the original action, the reward value, and the decision state vector for the next time step into the experience replay buffer. Specifically: The reward function module calculates the reward value for the current step based on the maximum internal temperature difference and core temperature output by the environment model. , experience tuple Store in the experience replay buffer; When the number of experience tuples in the experience playback buffer exceeds its maximum capacity At that time, the oldest experience tuple is removed in the order of its storage time; The evaluation network parameters are updated by sampling from the experience replay buffer to minimize the Bellman residual of the evaluation network, and the policy network parameters are updated by maximizing the Q-value of the policy network under the evaluation network. Specifically: A batch of experience tuples is randomly sampled from the experience replay buffer. For each experience tuple, the decision state vector for the next time step is generated. The input policy network obtains the raw action for the next time step. ,Will The input safety action filter yields the safety action for the next time step. Calculate the target Q value: ; in The discount factor has a value ranging from 0.95 to 0.99, preferably 0.99. For the target evaluation network, the smaller value of the outputs of the two target evaluation networks is selected to mitigate the overestimation bias of the Q value; the loss function of the evaluation network is: ; Minimize this loss to update the evaluation network parameters Policy Network To optimize the parameter update by maximizing the Q value of the safe action: first, update the original action output by the policy network. Input the safety action filter to obtain the safety action. Then calculate the loss: ; To enable backpropagation of the gradient, a pass-through estimator is used to approximate the gradient of the safety action filter. That is, a complete iterative gradient correction process is performed during forward propagation to obtain the safety action. During backpropagation, the gradient of the safety action filter with respect to the input action is approximated as an identity matrix, i.e. This allows the gradient of the policy network to be backpropagated via this approximation; The target evaluation network is updated as follows: every C steps, the network parameters are updated using a soft update method. ; in C is the soft update coefficient, ranging from 0.001 to 0.01, preferably 0.005; C is the preset update interval step number, ranging from 1 to 10, preferably 2. Through the aforementioned alternating updates, the policy network gradually learns to output a thermal management control strategy that balances temperature equilibrium, internal safety, and optimal energy consumption, guided by the reward function and within the constraints of the safety action filter.

[0034] Please see Figure 3 This invention provides a technical solution: a thermal management optimization system for lithium iron phosphate batteries based on wide-temperature modeling, comprising: a data acquisition unit for acquiring current operating state data of lithium iron phosphate batteries, constructing a time-series input sequence, and calculating a wide-temperature state vector; a wide-temperature modeling unit for inputting the wide-temperature state vector and the input sequence into a pre-trained wide-temperature adaptive temperature field sensing network, including a conditional parameter generation module and an attention correction prediction module; and a conditional parameter generation unit for enabling the conditional parameter generation module to take the wide-temperature state vector as input, extract temperature region features through its internal shared feature encoding layer, and then output dynamic network weights and temperature-sensitive thermal characteristic parameters in parallel through a weight generation branch and a thermal characteristic parameter generation branch, respectively. The attention correction prediction unit enables the attention correction prediction module to use dynamic network weights as computational parameters and, guided by temperature-sensitive thermal characteristic parameters, to perform correction processing on the time-series input sequence, fusing prior trends and data-driven information, to generate an internal temperature field prediction sequence. The thermal management optimization unit determines the internal core temperature and the maximum internal temperature difference based on the internal temperature field prediction sequence, and together with the wide-temperature state vector, forms a decision state vector, which is input into a pre-trained enhanced thermal management strategy network. After long-term and short-term collaborative safety constraint mapping processing, it generates thermal management optimization control variables.

[0035] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0036] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling, characterized in that, Includes the following steps: Collect the current operating status data of lithium iron phosphate batteries, construct a time-series input sequence, and calculate a wide-temperature state vector; The wide-temperature state vector and the input sequence are fed into a pre-trained wide-temperature adaptive temperature field sensing network, which includes a conditional parameter generation module and an attention correction prediction module. The condition parameter generation module takes a wide-temperature state vector as input, extracts temperature region features through its internal shared feature encoding layer, and then outputs dynamic network weights and temperature-sensitive thermal characteristic parameters in parallel through the weight generation branch and the thermal characteristic parameter generation branch, respectively. The attention correction prediction module uses dynamic network weights as computational parameters and, guided by temperature-sensitive thermal characteristic parameters, performs correction processing on the time-series input sequence, fusing prior trends and data-driven information, to generate an internal temperature field prediction sequence. Based on the internal temperature field prediction sequence, the internal core temperature and the maximum internal temperature difference are determined, and together with the wide-temperature state vector, they form a decision state vector. This vector is then input into a pre-trained enhanced thermal management strategy network and processed through long-term and short-term collaborative safety constraint mapping to generate thermal management optimization control variables.

2. The thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling according to claim 1, characterized in that, The specific steps for calculating the wide-temperature state vector are as follows: The temperature of multiple measuring points on the surface is extracted from the current operating status data, the average surface temperature is calculated, and the ambient temperature and current are directly obtained from the current operating status data. The state of charge is calculated based on the current using the ampere-hour integration method. Obtain historical battery charge-discharge state data from the most recent multiple charge-discharge cycles, calculate the average coulombic efficiency, static internal resistance increment, and cumulative throughput capacity, and concatenate them into an aging feature vector; By concatenating the average surface temperature, ambient temperature, state of charge, current, and aging characteristic vectors, a wide-temperature state vector is obtained.

3. The thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling according to claim 1, characterized in that, The specific structure of the condition parameter generation module is as follows: The shared feature coding layer consists of multiple fully connected layers, with the input being a wide-temperature state vector and the output being a temperature region feature vector; The weight generation branch takes the temperature region feature vector as input, outputs dynamic network weights through the first fully connected layer, and is used to reconstruct the attention projection matrix and feedforward network weights in the attention correction prediction module. The thermal characteristic parameter generation branch takes the temperature region feature vector as input and outputs temperature change sensitive thermal characteristic parameters through the second fully connected layer. The temperature change sensitive thermal characteristic parameters are composed of entropy change coefficient, temperature correction factor and equivalent convective heat transfer coefficient.

4. The thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling according to claim 3, characterized in that, The specific steps of the attention correction prediction module in performing correction processing that integrates prior trends and data-driven information are as follows: The encoder uses the reconstructed attention projection matrix and feedforward network weights as operational parameters to extract features from the temporal input sequence, generate a hidden state sequence, and map the hidden state sequence to a virtual measurement. The thermal equilibrium prior generation layer uses the equivalent convective heat transfer coefficient as the convective term coefficient to perform thermal equilibrium recursion and obtain the thermal equilibrium prior. The prior feature enhancement encoder takes thermal equilibrium priors and temperature-sensitive thermal characteristic parameters as inputs, processes them through learnable nonlinear feature mapping, and outputs enhanced prior features. The attention correction layer uses enhanced prior features as query vectors and virtual measurements as key and value vectors, and calculates the correction amount through a scaled dot product attention mechanism. The correction amount is added to the enhanced prior features to obtain the internal temperature field estimate for the current time step, and the internal temperature field prediction sequence is recursively generated from this estimate.

5. The thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling according to claim 4, characterized in that, The training steps for the wide-temperature adaptive temperature field sensing network are as follows: Obtain a training sample set consisting of multiple sets of samples, each set of samples containing the real temperature field of each internal node; All parameters of the freezing thermal property parameter generation branch are used, and the mean square error between the internal temperature field prediction sequence output by the attention correction prediction module and the actual temperature field of each internal node is used as the data fitting loss. After the data fitting loss converges, freeze all parameters of the shared feature encoding layer, weight generation branch, and attention correction prediction module, initialize the SGD optimizer separately for the thermal characteristic parameter generation branch, and reset the optimizer state. Fine-tune the optimization using the sum of energy conservation loss and thermal diffusion residual loss as the constraint loss. After the constraint loss converges, all parameters are unfrozen, and the total loss is the weighted sum of the data fitting loss and the constraint loss. Joint fine-tuning is then performed using a learning rate lower than that of the previous training to obtain a wide-temperature adaptive temperature field sensing network.

6. The thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling according to claim 5, characterized in that, The method for calculating energy conservation losses is as follows: For each time step in the internal temperature field prediction sequence, calculate the rate of temperature change at that time step; Calculate the heat generation power at the current time step based on the heat generation equation; Calculate the heat dissipation power based on the thermal management control variables and wide-temperature state vector in the time-series input sequence; By comprehensively processing the temperature change rate, heat generation power, and heat dissipation power at all time steps, the energy conservation loss is obtained. ,Right now: ； in, Let be the rate of temperature change at time step t+h. Let be the heat generation power at time step t+h. Let be the heat dissipation power at time step t+h. For the preset cell quality, This is the preset equivalent specific heat capacity.

7. The thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling according to claim 5, characterized in that, After the wide-temperature adaptive temperature field sensing network is trained, an online adaptive correction step is also included: The Fisher information diagonal value for each parameter in the wide-temperature adaptive temperature field sensing network is calculated using the training sample set; When the lithium iron phosphate battery is detected to be in a static state or the current is continuously less than the preset threshold, natural variation data is collected, and gradient updates are performed only on the bias vector of the second fully connected layer of the thermal characteristic parameter generation branch. The loss function optimized by gradient update includes energy conservation loss and anti-forgetting penalty term; The anti-forgetting penalty term uses the diagonal value of the Fisher information as the weight of each parameter, and applies a penalty to the deviation of the bias term from training convergence. The anti-forgetting penalty term... Defined as: ； in, The bias vector of the second fully connected layer of the generation branch only contains thermal characteristic parameters. The set of indexes for the parameters. The current parameter value. These are the parameter values ​​after training convergence.

8. The thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling according to claim 1, characterized in that, The enhanced thermal management strategy network includes a strategy network, a reward function module, and a safety action filter. The long-short-term collaborative safety constraint mapping is implemented collaboratively by the reward function module and the safety action filter, with the following specific steps: The reward function module extracts the maximum internal temperature difference from the internal temperature field prediction sequence. And based on the preset maximum temperature difference penalty function Add a penalty for maximum internal temperature difference. The forms of expression are as follows: ； in , To pre-determine safety limits; The security action filter is connected after the policy network and has a built-in local proxy layer. The local agent layer takes the decision state vector and the original actions generated by the policy network as input, and outputs the prediction of the maximum internal temperature difference. The safety action filter, constrained by the prediction that the maximum internal temperature difference does not exceed the preset safety limit, corrects the original action and outputs a thermal management optimization control quantity.

9. The thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling according to claim 8, characterized in that, The training steps for the enhanced thermal management strategy network are as follows: The parameters of the wide-temperature adaptive temperature field sensing network are frozen and used as the environment model. The current decision state vector is input into the policy network and the original action is output. The current decision state vector and the original action are simultaneously input into the safety action filter. The local agent layer outputs the predicted maximum internal temperature difference, which corrects the original action to obtain the safety action. This action is then input into the environmental model to obtain the maximum internal temperature difference. After performing safety actions on the environmental model, the maximum internal temperature difference is extracted, and online incremental fitting is performed on the local proxy layer to update the parameters of the local proxy layer. The reward value is calculated based on the maximum internal temperature difference, and the current decision state vector, original action, reward value, and decision state vector of the next time step are stored in the experience replay buffer. Sample from the experience replay buffer to update the evaluation network parameters by minimizing the Bellman residual of the evaluation network, and update the policy network parameters by maximizing the Q-value of the policy network under the evaluation network.

10. A thermal management optimization system for lithium iron phosphate batteries based on wide-temperature modeling, employing the thermal management optimization method for lithium iron phosphate batteries based on wide-temperature modeling as described in any one of claims 1-9, characterized in that, include: The data acquisition unit is used to collect the current operating status data of the lithium iron phosphate battery, construct the time-series input sequence, and calculate the wide-temperature state vector; The wide-temperature modeling unit is used to input the wide-temperature state vector and the input sequence into the pre-trained wide-temperature adaptive temperature field sensing network, including a conditional parameter generation module and an attention correction prediction module. The condition parameter generation unit is used to enable the condition parameter generation module to take a wide temperature state vector as input, extract temperature region features through its internal shared feature encoding layer, and then output dynamic network weights and temperature change sensitive thermal characteristic parameters in parallel through the weight generation branch and the thermal characteristic parameter generation branch, respectively. The attention correction prediction unit is used to enable the attention correction prediction module to use dynamic network weights as computational parameters and, guided by temperature-sensitive thermal characteristic parameters, to perform correction processing on the time-series input sequence, fusing prior trends and data-driven information, to generate an internal temperature field prediction sequence. The thermal management optimization unit is used to determine the internal core temperature and the maximum internal temperature difference based on the internal temperature field prediction sequence. It is then combined with the wide-temperature state vector to form a decision state vector, which is input into the pre-trained enhanced thermal management strategy network. After long-term and short-term collaborative safety constraint mapping, the thermal management optimization control quantity is generated.

Images