Low-temperature energy-saving temperature control management method for power battery based on working condition self-adaption

Abstract

The application discloses a low-temperature energy-saving temperature control management method for power batteries based on working condition self-adaptation, which comprises the following steps: collecting multi-dimensional battery, vehicle and environment data and performing pretreatment, identifying the current typical working condition of the vehicle and generating a future working condition prediction sequence; quantifying the battery aging attenuation amount per unit time in a low-temperature environment, and constructing a temperature control energy consumption and aging attenuation double-target normalized comprehensive cost function; dynamically allocating double-target weights through a reinforcement learning model and generating an optimization benchmark, and combining the prediction sequence to solve a multi-heat-source collaborative temperature control execution scheme; collecting full-process operation data to evaluate the temperature control effect and iteratively updating all calculation models. The scheme has the characteristics of incorporating real-time aging attenuation into the direct optimization target, dynamically balancing energy saving and battery life, and having working condition self-adaptation and full-life-cycle closed-loop iteration.

Description

Technical Field

[0001] This invention relates to the field of thermal management technology for power batteries in new energy vehicles, specifically a low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions. Background Technology

[0002] The electrochemical reaction characteristics of lithium-ion power batteries are highly sensitive to temperature. In low-temperature environments, increased electrolyte viscosity and decreased lithium-ion diffusion rates can lead to problems such as increased internal resistance, decreased usable capacity, and increased risk of lithium deposition. Active temperature control heating is a core element in ensuring the performance, safety, and lifespan of power batteries. Currently, automotive low-temperature control generally adopts a multi-heat source combination architecture, and the control strategy follows the traditional method of fixed target temperature plus power adjustment, using temperature as the sole direct control target and indirectly ensuring battery life by maintaining the target temperature.

[0003] The existing technology has the following main technical problems: First, the target design has inherent defects. Fixed temperature targets cannot adapt to the differences in battery aging sensitivity under different operating conditions, and cannot solve the problem that energy saving accelerates aging, while maintaining lifespan results in excessive energy consumption. Second, the technology chain is disconnected. Battery aging assessment and temperature control are independent technical branches, and real-time aging degradation is not included in the temperature control closed loop. Third, the adaptability to operating conditions is insufficient. A few multi-objective optimization schemes use manually calibrated fixed weights, which are difficult to adapt to complex and ever-changing actual vehicle use scenarios. Fourth, the adaptability throughout the entire life cycle is poor. The strategy parameters are calibrated based on the initial state of the battery and lack a closed-loop self-learning iteration mechanism. As the battery ages, the optimization effect continues to decay. Summary of the Invention

[0004] The purpose of this invention is to provide a low-temperature energy-saving temperature control management method for power batteries based on operating condition adaptation, so as to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions, comprising: Collect multi-dimensional battery, vehicle and environmental data and perform standardized noise reduction preprocessing. Based on the preprocessed data, identify the vehicle’s current typical operating conditions and generate a future operating condition prediction sequence. By combining the identified current typical operating conditions with the corresponding battery operating status, the battery aging and degradation per unit time under low temperature conditions is quantified. A comprehensive cost function with dimensional normalization is constructed with temperature control energy consumption and the quantified real-time aging and degradation as dual optimization objectives. By combining the current typical operating conditions with the predicted sequence of future operating conditions, a reinforcement learning model is used to dynamically assign weights to the two optimization objectives of the comprehensive cost function and generate a dynamically weighted comprehensive cost optimization benchmark. Using the aforementioned dynamic weighted comprehensive cost optimization benchmark as the optimization objective and combining it with the future operating condition prediction sequence, the multi-heat source collaborative temperature control execution scheme is solved. The temperature control effect is evaluated by collecting full-process operation data and all calculation models are iteratively updated.

[0006] Furthermore, the current typical operating conditions, future operating condition prediction sequence, low temperature level, battery state of charge, estimated remaining runtime of navigation, driving mode, and battery health status are all input into the reinforcement learning model. Multiple discrete weight combinations covering the energy-saving priority interval to the lifespan priority interval are used as the model's optional action set. A reward function is constructed with the benchmark energy consumption and benchmark aging amount corresponding to the fixed temperature control strategy as references. The comprehensive cost corresponding to each weight is evaluated and iteratively optimized to output the optimal weight combination that fits the current scenario. The optimal weights are substituted into the comprehensive cost function to generate the corresponding dynamic weighted comprehensive cost optimization benchmark.

[0007] Furthermore, the battery temperature, charge / discharge current rate, state of charge, health status, ambient temperature, and duration of operation corresponding to the current typical operating conditions are extracted as input features. An aging quantification model that integrates data-driven fitting and electrochemical mechanism constraints is built. The nonlinear mapping relationship between the input features and the battery aging rate is fitted through the backbone of the data-driven fitting network. The electrochemical laws of lithium deposition kinetics and interface film growth kinetics are used as physical constraints to correct the fitting results. The decay contribution of the three aging paths of lithium deposition, interface film growth, and active material loss is decomposed, and the overall battery aging decay value per unit time is output.

[0008] Furthermore, a rolling time-domain predictive control optimization framework is built using the generated future operating condition prediction sequence as the operating condition input in the prediction time domain. A control time domain with a length matching the prediction time domain is set, and rolling iteration is performed according to a fixed control cycle. In each control cycle, the complete temperature control sequence in the prediction time domain is solved, and only the control quantity of the current cycle is output and executed. The next control cycle is re-solved based on the latest collected operating data. The dynamic weighted comprehensive cost optimization benchmark is loaded as the fitness function of the solution process, with minimizing the comprehensive cost value as the optimization direction. Multi-dimensional constraints are set, including upper and lower limits of battery temperature, individual cell temperature difference, maximum power of each heat source, and liquid flow adjustment range. The optimal control parameters that satisfy all constraints are iteratively solved by a swarm intelligent optimization algorithm with an inertial weight adaptive adjustment mechanism. The output is a multi-heat source collaborative temperature control command sequence including heating power, compressor frequency, valve opening, and liquid pump flow.

[0009] Furthermore, the electrochemical formulas of lithium deposition kinetics and interfacial film growth kinetics are embedded as physical regularization constraints in the loss calculation stage of the aging quantification model. When the model fits the mapping relationship between input features and aging rate, the output results are constrained to conform to the basic laws of electrochemical aging. After each complete charge-discharge cycle, the measured data of internal resistance change and capacity decay corresponding to that cycle are collected as calibration benchmarks. The recursive least squares method with forgetting factor is used to recursively correct the benchmark parameters of the output layer of the aging quantification model online. The forgetting factor balances the weight of historical data and the proportion of newly added measured data.

[0010] Furthermore, a simulation dataset covering multiple operating conditions, multiple levels of low temperature environment, and multiple stages of battery aging is used to pre-train the reinforcement learning model offline until convergence. During actual vehicle operation, after each complete driving or charging cycle, the output layer parameters of the reinforcement learning model are continuously fine-tuned online based on the actual operating data of that cycle. The value boundary range of the weights of the two types of targets is set in the action space of the model to limit the upper and lower limits of the selectable weights. The benefit coefficients corresponding to the energy consumption benefit item and the aging benefit item in the reward mechanism are adjusted according to the driving mode selected by the user.

[0011] Furthermore, using preprocessed multi-dimensional battery, vehicle, and environmental data as input, a fixed-duration sliding time window of data is extracted and input into a lightweight temporal classification network. Through temporal feature extraction and classification calculation, the current typical operating conditions of the vehicle are identified. Based on the statistical data of historical operating condition switching, a basic matrix of transition probabilities between different operating conditions is constructed. The driving features of the road ahead obtained by navigation are introduced to correct the transition probability values ​​of each operating condition in the transition probability matrix. Taking the currently identified typical operating condition as the initial state, the future operating condition prediction sequence is generated step by step in combination with the corrected transition probability matrix. The future operating condition prediction sequence is output to the subsequent multi-heat source temperature control optimization stage as the operating condition input in the prediction time domain.

[0012] Furthermore, the battery pack is divided into multiple independent temperature control zones according to the internal module arrangement. Each zone corresponds to an independent liquid circuit branch and a heating execution unit. Differentiated heating intensity parameters are set according to the real-time aging decay amount per unit time and the current temperature deviation value of each zone. The historical temperature sequence, current heating power, and future operating condition prediction sequence information of each zone are collected and input into the gated loop unit network. The future short-term temperature change trend of each zone is output in real time. When the deviation between the actual collected temperature of the zone and the preset target temperature plus the predicted temperature change exceeds the preset deviation threshold, the heating power and liquid circuit flow control commands of the current cycle are dynamically corrected within the preset proportional adjustment range of the original control command.

[0013] Furthermore, after a single complete vehicle usage cycle, the temperature control effect evaluation value is calculated based on the full-process operation data collected in this cycle. The deviation between the evaluation value and the preset target value is used as the loss basis. The output layer parameters of the aging quantization model and reinforcement learning weight allocation model on the vehicle terminal are fine-tuned and local model gradients are generated. The federated learning framework is used to summarize the model gradient information generated locally by each vehicle terminal. Each terminal only uploads gradient data and does not upload the original operation data. After receiving the gradient data uploaded by all terminals, the cloud uses the federated averaging algorithm to perform gradient aggregation according to the data volume of each terminal, updates the corresponding network parameters of the global basic model, and distributes the updated global basic model to each vehicle terminal through the vehicle network link to complete the iterative update of the local model on the vehicle terminal.

[0014] Furthermore, after a single complete vehicle usage cycle, based on the full-process operation data collected during that cycle, the energy consumption of the heating unit, the power consumption of the liquid circuit circulation, and the power consumption of the refrigeration compressor are extracted as input indicators. The average temperature rise rate of the battery, the amplitude of the single-cell temperature difference control, and the cumulative aging attenuation within the cycle are extracted as output indicators. Based on the evaluation orientation of minimizing input and maximizing output, a data envelopment analysis model is used to calculate the comprehensive relative efficiency value of temperature control input and output for that cycle as the temperature control effect evaluation value. The obtained comprehensive efficiency value is used as the target label. The current operating condition parameters, ambient temperature, battery state of charge, health status, and control parameters of each heat source are substituted into the global feature attribution interpretability algorithm to calculate the contribution value of each input feature to the temperature control effect. The features are sorted from high to low according to the absolute value of the contribution, and features exceeding the preset contribution threshold are selected as core driving factors. The corresponding optimization objects are matched according to the attribute category of the core driving factors. Operating condition and environmental factors correspond to weight allocation models, battery state factors correspond to aging quantification models, and control execution factors correspond to temperature control solution models. The parameter optimization adjustment direction of the corresponding models is output.

[0015] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention incorporates real-time battery aging degradation as a direct optimization target into the temperature control closed loop, and together with temperature control energy consumption, constructs a dual-objective normalized comprehensive cost function, fundamentally resolving the inherent contradiction of "energy saving leads to accelerated aging, while maintaining lifespan results in excessive energy consumption." Based on this, a reinforcement learning model is used to dynamically allocate dual-objective weights based on operating conditions, battery status, and user preferences, replacing the traditional manually calibrated fixed weight mode, which can adaptively match diverse vehicle usage scenarios. Simultaneously, a rolling time-domain predictive control framework is built by combining future operating condition prediction sequences to proactively solve multi-heat source collaborative control schemes, effectively reducing the lag of traditional feedback-based temperature control. Finally, a closed-loop self-learning mechanism is constructed based on a federated learning architecture, enabling continuous iterative optimization of the temperature control strategy throughout the battery's aging process without uploading original driving data and ensuring data privacy, comprehensively improving the energy efficiency, scenario adaptability, and long-term service effectiveness of low-temperature temperature control. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions, provided by an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 The flowchart of a low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions provided in an embodiment of the present invention is as follows: Figure 1 It can be seen that the power battery low-temperature energy-saving temperature control management method based on operating condition adaptation includes: Collect multi-dimensional battery, vehicle and environmental data and perform standardized noise reduction preprocessing. Based on the preprocessed data, identify the vehicle’s current typical operating conditions and generate a future operating condition prediction sequence. By combining the identified current typical operating conditions with the corresponding battery operating status, the battery aging and degradation per unit time under low temperature conditions is quantified. A comprehensive cost function with dimensional normalization is constructed with temperature control energy consumption and the quantified real-time aging and degradation as dual optimization objectives. By combining the current typical operating conditions with the predicted sequence of future operating conditions, a reinforcement learning model is used to dynamically assign weights to the two optimization objectives of the comprehensive cost function and generate a dynamically weighted comprehensive cost optimization benchmark. Using the aforementioned dynamic weighted comprehensive cost optimization benchmark as the optimization objective and combining it with the future operating condition prediction sequence, the multi-heat source collaborative temperature control execution scheme is solved. The temperature control effect is evaluated by collecting full-process operation data and all calculation models are iteratively updated.

[0019] Furthermore, the current typical operating conditions, future operating condition prediction sequence, low temperature level, battery state of charge, estimated remaining runtime of navigation, driving mode, and battery health status are all input into the reinforcement learning model. Multiple discrete weight combinations covering the energy-saving priority interval to the lifespan priority interval are used as the model's optional action set. A reward function is constructed with the benchmark energy consumption and benchmark aging amount corresponding to the fixed temperature control strategy as references. The comprehensive cost corresponding to each weight is evaluated and iteratively optimized to output the optimal weight combination that fits the current scenario. The optimal weights are substituted into the comprehensive cost function to generate the corresponding dynamic weighted comprehensive cost optimization benchmark.

[0020] Preferably, the current typical operating conditions, future operating condition prediction sequence, low temperature level, battery state of charge, estimated remaining navigation runtime, driving mode, and battery health status are numerically encoded and normalized, and then concatenated into a one-dimensional state vector input to the reinforcement learning model. The reinforcement learning model adopts a deep reinforcement learning network architecture based on value functions, with multiple layers of feature extraction networks internally performing nonlinear mapping calculations on the input state vector, outputting the expected cumulative reward value corresponding to each optional action. Multiple sets of discrete bi-objective weight coefficient pairs covering the energy-saving priority interval to the lifespan priority interval are used as the model's optional action set. Each weight coefficient pair includes temperature control energy consumption weight and aging degradation weight, and the sum of the two is 1. The normalization constraint is applied; a reward function is constructed based on the unit time benchmark energy consumption value and unit time benchmark aging decay value corresponding to the same operating conditions and battery state under the fixed temperature control strategy. The unit time actual energy consumption value and unit time actual aging decay value under the temperature control scheme corresponding to the current weight combination are collected, and the energy consumption reduction rate and aging decay reduction rate are calculated respectively. According to the current driving mode, the corresponding energy consumption benefit coefficient and aging benefit coefficient are matched. The two types of reduction rates are multiplied by the corresponding coefficients and then summed to obtain the single-step instant reward value. The single-step instant reward value is used to calculate the temporal difference error of the model to iteratively update the network parameters. After multiple rounds of iterative optimization, the action with the highest expected cumulative benefit value is selected as the optimal weight combination. The two sets of weight coefficients in the optimal weight combination are substituted into the comprehensive cost function to generate the corresponding dynamic weighted comprehensive cost optimization benchmark.

[0021] Furthermore, the battery temperature, charge / discharge current rate, state of charge, health status, ambient temperature, and duration of operation corresponding to the current typical operating conditions are extracted as input features. An aging quantification model that integrates data-driven fitting and electrochemical mechanism constraints is built. The nonlinear mapping relationship between the input features and the battery aging rate is fitted through the backbone of the data-driven fitting network. The electrochemical laws of lithium deposition kinetics and interface film growth kinetics are used as physical constraints to correct the fitting results. The decay contribution of the three aging paths of lithium deposition, interface film growth, and active material loss is decomposed, and the overall battery aging decay value per unit time is output.

[0022] Preferably, the extracted battery temperature, charge / discharge current rate, state of charge, health status, ambient temperature, and operating condition duration input features are standardized and mapped, and then divided into three categories according to attributes: electrochemical state features, lifetime baseline features, and operating condition boundary features before being input into the aging quantification model. The aging quantification model adopts an architecture coupling a multi-task data-driven fitting branch and an electrochemical mechanism constraint branch. The multi-task data-driven fitting branch uses a multi-layer nonlinear transformation network, and after receiving all input features, it extracts aging-related features through feature cross-validation, outputting initial values ​​of the unit time rate corresponding to the three aging paths: lithium deposition decay, interface film growth decay, and active material loss decay. The electrochemical mechanism constraint branch extracts from the input features... Battery temperature, charge / discharge current rate, and state of charge are used as calculation parameters. The theoretical rate boundary of the corresponding aging path is obtained by solving the built-in lithium deposition kinetics equation and interface film growth kinetics equation. This boundary is embedded as a physical regularization term into the loss calculation link of the multi-task data-driven fitting branch. At the same time, the initial rate value is truncated and corrected at the inference output to make the output rate conform to the basic laws of electrochemical reaction. The unit time rate of the three types of aging paths after correction is converted by combining the decay benchmark coefficient corresponding to the lifetime baseline characteristics and the time accumulation coefficient corresponding to the operating condition boundary characteristics to obtain the unit time decay contribution of each of the three types of aging paths. The overall battery aging decay value per unit time is output after summing the three types of decay contributions.

[0023] Furthermore, a rolling time-domain predictive control optimization framework is built using the generated future operating condition prediction sequence as the operating condition input in the prediction time domain. A control time domain with a length matching the prediction time domain is set, and rolling iteration is performed according to a fixed control cycle. In each control cycle, the complete temperature control sequence in the prediction time domain is solved, and only the control quantity of the current cycle is output and executed. The next control cycle is re-solved based on the latest collected operating data. The dynamic weighted comprehensive cost optimization benchmark is loaded as the fitness function of the solution process, with minimizing the comprehensive cost value as the optimization direction. Multi-dimensional constraints are set, including upper and lower limits of battery temperature, individual cell temperature difference, maximum power of each heat source, and liquid flow adjustment range. The optimal control parameters that satisfy all constraints are iteratively solved by a swarm intelligent optimization algorithm with an inertial weight adaptive adjustment mechanism. The output is a multi-heat source collaborative temperature control command sequence including heating power, compressor frequency, valve opening, and liquid pump flow.

[0024] Preferably, a rolling time-domain predictive control optimization framework is built using the generated future operating condition prediction sequence as the operating condition input in the prediction time domain. A control time domain with a length matching the prediction time domain is set, and rolling iteration is performed according to a fixed control cycle. In each control cycle, the time domain calculation boundary is initialized based on the latest acquired operating state. The battery thermal state and energy consumption change trends are synchronously extrapolated at each discrete calculation node in the prediction time domain, and a piecewise fitness calculation function adapted to the time-domain rolling mechanism is constructed. The fitness function uses the set of multi-heat source control variables of each discrete node in the prediction time domain, the extrapolated battery temperature and aging state parameters, and the corresponding node's operating condition characteristic parameters as input variables. The two core components, energy consumption cost and aging cost, are accumulated and calculated node by node along the time domain direction. The energy consumption cost component is obtained by summing the products of the operating power of the corresponding heating unit, refrigeration compressor, and liquid circulation pump at each node and the node's time step. The aging cost component is obtained by summing the products of the battery state parameters at each node. The decay per unit time output by the aging quantization model is obtained by integrating along the prediction time domain. The energy consumption weight coefficient and aging decay weight coefficient corresponding to the current control cycle dynamic weighted comprehensive cost optimization benchmark are substituted into the summation of the two types of total costs to obtain the benchmark target cost value. At the same time, a soft constraint penalty mechanism is adopted to transform the constraints of battery temperature upper and lower limits, single cell temperature difference limits, power limits of each heat source, and liquid flow adjustment range into penalty cost components that are positively correlated with the out-of-bounds amplitude. The larger the out-of-bounds amplitude, the higher the corresponding penalty cost value. The sum of the benchmark target cost value and the penalty cost component is used as the final output value of the fitness function. With minimizing the output value of the fitness function as the optimization direction, the population iterative solution is completed by a swarm intelligent optimization algorithm with an inertial weight adaptive adjustment mechanism. The optimal control parameter set that satisfies all constraints is output, forming a multi-heat source collaborative temperature control command sequence that includes heating power, compressor frequency, valve opening, and liquid pump flow.

[0025] Furthermore, the electrochemical formulas of lithium deposition kinetics and interfacial film growth kinetics are embedded as physical regularization constraints in the loss calculation stage of the aging quantification model. When the model fits the mapping relationship between input features and aging rate, the output results are constrained to conform to the basic laws of electrochemical aging. After each complete charge-discharge cycle, the measured data of internal resistance change and capacity decay corresponding to that cycle are collected as calibration benchmarks. The recursive least squares method with forgetting factor is used to recursively correct the benchmark parameters of the output layer of the aging quantification model online. The forgetting factor balances the weight of historical data and the proportion of newly added measured data.

[0026] Preferably, the aging quantification model incorporates an electrochemical formula for lithium deposition kinetics, using the negative electrode surface overpotential, battery operating temperature, and charge / discharge current density as input parameters. This formula calculates the critical rate threshold and theoretical rate range for lithium deposition under the corresponding operating conditions. The in-house electrochemical formula for interface film growth kinetics uses battery temperature, state of charge, and cycle cumulative equivalent as input parameters. This formula calculates the theoretical rate range for interface film thickness growth. These two theoretical rate ranges are then transformed into interval-constrained physical regularization terms, which are weighted and summed with the fitting loss function of the data-driven fitting branch to form the total loss function for model training. When the corresponding aging path rate output by the model exceeds the theoretical rate range, a penalty loss is added based on the magnitude of the exceedance. The model fitting process always conforms to the reaction kinetics of electrochemical aging. After each complete charge-discharge cycle, the internal resistance difference and capacity decay at the beginning and end of the cycle are collected as measured observation vectors. The baseline offset parameter vector of the aging quantification model output layer is determined as the object to be corrected. A linear mapping relationship between the parameter to be corrected and the measured observation vector is established. The parameters are iteratively updated using the recursive least squares method with a forgetting factor. The historical covariance matrix is ​​attenuated by the forgetting factor to reduce the weight of historical data. The recursive gain matrix is ​​calculated based on the measured observation values ​​of the current cycle, and the corrected output layer baseline parameter vector is updated to achieve online calibration of the model output accuracy as the battery aging process progresses.

[0027] Furthermore, a simulation dataset covering multiple operating conditions, multiple levels of low temperature environment, and multiple stages of battery aging is used to pre-train the reinforcement learning model offline until convergence. During actual vehicle operation, after each complete driving or charging cycle, the output layer parameters of the reinforcement learning model are continuously fine-tuned online based on the actual operating data of that cycle. The value boundary range of the weights of the two types of targets is set in the action space of the model to limit the upper and lower limits of the selectable weights. The benefit coefficients corresponding to the energy consumption benefit item and the aging benefit item in the reward mechanism are adjusted according to the driving mode selected by the user.

[0028] Preferably, a power battery temperature control simulation sample library is constructed, covering multiple operating conditions, multiple levels of low-temperature environments, and multiple stages of battery aging states. State feature samples, weight action samples, and corresponding reward values ​​from the sample library are stored in an experience replay pool. The reinforcement learning model randomly selects samples from the experience replay pool for batch training, updating the network's full parameters with the goal of minimizing temporal difference error. Offline pre-training is completed when the cumulative reward fluctuation amplitude of multiple consecutive training rounds is lower than a preset convergence threshold, enabling the model to have weight decision-making capabilities in basic scenarios. During actual vehicle operation, after each complete driving or charging cycle, the actual temperature control energy consumption, actual aging degradation, and corresponding weight combinations for each time step within that cycle are extracted, and the calculations for each time step are performed. The actual reward value is used to freeze the backbone feature extraction network parameters of the reinforcement learning model. Gradient descent is only performed on the policy mapping parameters of the output layer to achieve continuous online fine-tuning of the model. Within the model's action space, the upper and lower limits of energy consumption weights and aging weights are determined according to the minimum temperature control performance requirements and minimum lifespan guarantee requirements in low-temperature scenarios, respectively. Boundary constraint rules for the action space are constructed. When the model outputs an action, weight combinations exceeding the boundary interval are truncated to limit the range of selectable weight values. According to the user's selected driving mode, the corresponding benefit coefficient mapping rules are matched, and the weighting coefficients of energy consumption benefit items and aging benefit items in the reward calculation are adjusted respectively. Different coefficient ratios correspond to different driving modes, thereby correcting the model's optimization orientation.

[0029] Furthermore, using preprocessed multi-dimensional battery, vehicle, and environmental data as input, a fixed-duration sliding time window of data is extracted and input into a lightweight temporal classification network. Through temporal feature extraction and classification calculation, the current typical operating conditions of the vehicle are identified. Based on the statistical data of historical operating condition switching, a basic matrix of transition probabilities between different operating conditions is constructed. The driving features of the road ahead obtained by navigation are introduced to correct the transition probability values ​​of each operating condition in the transition probability matrix. Taking the currently identified typical operating condition as the initial state, the future operating condition prediction sequence is generated step by step in combination with the corrected transition probability matrix. The future operating condition prediction sequence is output to the subsequent multi-heat source temperature control optimization stage as the operating condition input in the prediction time domain.

[0030] Preferably, the preprocessed multi-dimensional battery, vehicle, and environmental data undergoes temporal alignment and dimensional normalization. A temporal feature matrix is ​​constructed by sliding and truncating continuous time series segments with a fixed step size, and input into a lightweight temporal classification network. This lightweight temporal classification network employs a multi-layer, depthwise separable temporal convolutional module stacking architecture. It extracts temporal dimension features through a combination of channel-wise and pointwise convolutions. After global average pooling, it outputs the category probability values ​​corresponding to each preset typical operating condition. The operating condition category with the highest probability value is selected as the current typical operating condition of the vehicle. The frequency of operating condition switching within each fixed time step in historical operating data is statistically analyzed. A frequency statistics matrix is ​​constructed with the current operating condition category as the row and the next operating condition category as the column. Normalization is performed on the frequency of each row to obtain the basic matrix of transition probabilities between operating conditions. Each element in the matrix corresponds to the probability of switching from the current operating condition to the next operating condition, and the sum of the elements in each row is 1. Extraction... The navigation system obtains parameters such as road slope, speed limit, intersection distribution, and congestion level. These parameters are mapped to generate transition probability correction coefficients for each operating condition type. The elements in the corresponding rows of the transition probability base matrix are then weighted and adjusted using these correction coefficients. After adjustment, each row is normalized again to obtain the corrected operating condition transition probability matrix. Using the currently identified typical operating condition as the initial state node, the system iterates sequentially along the prediction time domain at fixed time steps. At each iteration step, the next operating condition category with the highest transition probability is selected from the corrected transition probability matrix based on the current operating condition category. This process is repeated to iterate through all time steps within the prediction time domain, generating a continuous future operating condition prediction sequence. The time step of this future operating condition prediction sequence corresponds one-to-one with the discrete calculation nodes of the subsequent rolling time domain predictive control optimization framework, serving as the operating condition input / output to multi-heat source temperature control optimization stage within the prediction time domain.

[0031] Furthermore, the battery pack is divided into multiple independent temperature control zones according to the internal module arrangement. Each zone corresponds to an independent liquid circuit branch and a heating execution unit. Differentiated heating intensity parameters are set according to the real-time aging decay amount per unit time and the current temperature deviation value of each zone. The historical temperature sequence, current heating power, and future operating condition prediction sequence information of each zone are collected and input into the gated loop unit network. The future short-term temperature change trend of each zone is output in real time. When the deviation between the actual collected temperature of the zone and the preset target temperature plus the predicted temperature change exceeds the preset deviation threshold, the heating power and liquid circuit flow control commands of the current cycle are dynamically corrected within the preset proportional adjustment range of the original control command.

[0032] Preferably, the battery pack is divided into multiple independent temperature control zones according to the physical arrangement of the internal modules and the branching of the liquid circuit. Each zone corresponds to an independent coolant branch and an independently controllable heating execution unit. For each zone, parameters are adjusted based on the difference between the current temperature and the preset target temperature, and parameters are corrected by weighting the real-time aging decay amount per unit time of the corresponding zone. Differentiated initial heating intensity parameters are set for each zone. Historical temperature sampling sequences, real-time heating power values, and corresponding operating condition characteristic parameters for a fixed duration are collected for each zone. After time-series alignment and dimensional normalization, these are used to construct an input feature sequence input gated loop unit network. The network filters the time-series dimension through reset gates and update gates. The system identifies the correlation characteristics of temperature and captures the thermal inertia during temperature changes. It outputs the hourly temperature prediction values ​​for each zone within a preset time period. Based on the difference between the predicted value and the current temperature, it calculates the short-term temperature change trend of the zone. It also calculates the real-time deviation between the current measured temperature and the preset target temperature of the zone. The real-time deviation value is then superimposed with the short-term temperature change trend to obtain the comprehensive deviation. When the comprehensive deviation exceeds the preset deviation threshold, the system calculates the control correction amount according to the direction and amplitude of the comprehensive deviation, within the preset proportional adjustment range corresponding to the original control command, based on the linear correspondence between the deviation amplitude and the adjustment amount. This allows for the same-direction adjustment of the heating power command and the liquid flow command for the current cycle, ensuring that the adjusted operating parameters do not exceed the hardware rated operating range of the corresponding zone.

[0033] Furthermore, after a single complete vehicle usage cycle, the temperature control effect evaluation value is calculated based on the full-process operation data collected in this cycle. The deviation between the evaluation value and the preset target value is used as the loss basis. The output layer parameters of the aging quantization model and reinforcement learning weight allocation model on the vehicle terminal are fine-tuned and local model gradients are generated. The federated learning framework is used to summarize the model gradient information generated locally by each vehicle terminal. Each terminal only uploads gradient data and not the original operation data. After receiving the gradient data uploaded by all terminals, the cloud uses the federated averaging algorithm to perform gradient aggregation according to the data volume of each terminal, updates the corresponding network parameters of the global basic model, and distributes the updated global basic model to each vehicle terminal through the vehicle network link to complete the iterative update of the local model on the vehicle terminal.

[0034] Furthermore, after a single complete vehicle usage cycle, based on the full-process operation data collected during that cycle, the energy consumption of the heating unit, the power consumption of the liquid circuit circulation, and the power consumption of the refrigeration compressor are extracted as input indicators. The average temperature rise rate of the battery, the amplitude of the single-cell temperature difference control, and the cumulative aging attenuation within the cycle are extracted as output indicators. Based on the evaluation orientation of minimizing input and maximizing output, a data envelopment analysis model is used to calculate the comprehensive relative efficiency value of temperature control input and output for that cycle as the temperature control effect evaluation value. The obtained comprehensive efficiency value is used as the target label. The current operating condition parameters, ambient temperature, battery state of charge, health status, and control parameters of each heat source are substituted into the global feature attribution interpretability algorithm to calculate the contribution value of each input feature to the temperature control effect. The features are sorted from high to low according to the absolute value of the contribution. Features exceeding the preset contribution threshold are selected as core driving factors. According to the attribute category of the core driving factors, the corresponding optimization objects are matched respectively: operating condition and environmental factors correspond to weight allocation models, battery state factors correspond to aging quantification models, and control execution factors correspond to temperature control solution models. The parameter optimization adjustment direction of the corresponding models is output.

[0035] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0036] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions, characterized in that, include: Collect multi-dimensional battery, vehicle and environmental data and perform standardized noise reduction preprocessing. Based on the preprocessed data, identify the vehicle’s current typical operating conditions and generate a future operating condition prediction sequence. By combining the identified current typical operating conditions with the corresponding battery operating status, the battery aging and degradation per unit time under low temperature conditions is quantified. A comprehensive cost function with dimensional normalization is constructed with temperature control energy consumption and the quantified real-time aging and degradation as dual optimization objectives. By combining the current typical operating conditions with the predicted sequence of future operating conditions, a reinforcement learning model is used to dynamically assign weights to the two optimization objectives of the comprehensive cost function and generate a dynamically weighted comprehensive cost optimization benchmark. Using the aforementioned dynamic weighted comprehensive cost optimization benchmark as the optimization objective and combining it with the future operating condition prediction sequence, the multi-heat source collaborative temperature control execution scheme is solved. The temperature control effect is evaluated by collecting full-process operation data and all calculation models are iteratively updated.

2. The low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions according to claim 1, characterized in that, The process of combining current typical operating conditions and future operating condition prediction sequences, using a reinforcement learning model to dynamically assign weights to the two optimization objectives of the comprehensive cost function, and generating a dynamically weighted comprehensive cost optimization benchmark includes: The current typical operating conditions, future operating condition prediction sequence, low temperature level, battery state of charge, estimated remaining runtime of navigation, driving mode, and battery health status are all input into the reinforcement learning model. Multiple discrete weight combinations covering the energy-saving priority interval to the lifespan priority interval are used as the model's optional action set. A reward function is constructed with the benchmark energy consumption and benchmark aging amount corresponding to the fixed temperature control strategy as references. The comprehensive cost corresponding to each weight is evaluated and iteratively optimized to output the optimal weight combination that fits the current scenario. The optimal weights are substituted into the comprehensive cost function to generate the corresponding dynamic weighted comprehensive cost optimization benchmark.

3. The low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions according to claim 1, characterized in that, The quantitative measurement of battery aging degradation per unit time under low-temperature conditions, obtained by combining the current typical operating conditions and corresponding battery operating states, includes: The battery temperature, charge / discharge current rate, state of charge, health status, ambient temperature, and duration of operation under the current typical operating conditions are extracted as input features. An aging quantification model that integrates data-driven fitting and electrochemical mechanism constraints is built. The nonlinear mapping relationship between the input features and the battery aging rate is fitted by the backbone of the data-driven fitting network. The electrochemical laws of lithium deposition kinetics and interface film growth kinetics are used as physical constraints to correct the fitting results. The decay contribution of the three aging paths of lithium deposition, interface film growth, and active material loss is decomposed, and the overall battery aging decay value per unit time is output.

4. The low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions according to claim 1, characterized in that, The method of solving the multi-heat-source collaborative temperature control execution scheme by using the dynamic weighted comprehensive cost optimization benchmark as the optimization objective and combining it with the future operating condition prediction sequence includes: A rolling time-domain predictive control optimization framework is built using the generated future operating condition prediction sequence as the operating condition input in the prediction time domain. A control time domain with a length matching the prediction time domain is set, and rolling iteration is performed according to a fixed control cycle. In each control cycle, the complete temperature control sequence in the prediction time domain is solved, and only the control quantity of the current cycle is output and executed. The next control cycle is re-solved based on the latest collected operating data. The dynamic weighted comprehensive cost optimization benchmark is loaded as the fitness function of the solution process, with minimizing the comprehensive cost value as the optimization direction. Multi-dimensional constraints such as upper and lower limits of battery temperature, single cell temperature difference, maximum power of each heat source, and liquid flow adjustment range are set. The optimal control parameters that satisfy all constraints are iteratively solved by a swarm intelligent optimization algorithm with an inertial weight adaptive adjustment mechanism. The output is a multi-heat source collaborative temperature control command sequence including heating power, compressor frequency, valve opening, and liquid pump flow.

5. The low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions according to claim 3, characterized in that, The construction of the aging quantification model that integrates data-driven fitting and electrochemical mechanism constraints also includes: Electrochemical formulas for lithium deposition kinetics and interfacial film growth kinetics are embedded as physical canonical constraints in the loss calculation stage of the aging quantification model. When the model fits the mapping relationship between input features and aging rate, the output results are constrained to conform to the basic laws of electrochemical aging. After each complete charge-discharge cycle, the measured data of internal resistance change and capacity decay corresponding to that cycle are collected as calibration benchmarks. The recursive least squares method with forgetting factor is used to recursively correct the benchmark parameters of the output layer of the aging quantification model online. The forgetting factor balances the weight of historical data and the proportion of newly added measured data.

6. The low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions according to claim 2, characterized in that, The step of dynamically assigning weights to the two optimization objectives of the comprehensive cost function through a reinforcement learning model further includes: The reinforcement learning model is pre-trained offline until convergence using a simulation dataset covering multiple operating conditions, multiple levels of low temperature environment, and multiple stages of battery aging. During actual vehicle operation, after each complete driving or charging cycle, the output layer parameters of the reinforcement learning model are continuously fine-tuned online based on the actual operating data of that cycle. The boundary range of the weights of the two types of targets is set in the action space of the model to limit the upper and lower limits of the selectable weights. The benefit coefficients corresponding to the energy consumption benefit item and the aging benefit item in the reward mechanism are adjusted according to the driving mode selected by the user.

7. The low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions according to claim 1, characterized in that, The process of identifying typical current operating conditions of the vehicle based on preprocessed data and generating a prediction sequence for future operating conditions includes: Using preprocessed multi-dimensional battery, vehicle, and environmental data as input, a lightweight temporal classification network is input with data from a fixed-duration sliding time window. Through temporal feature extraction and classification calculation, the current typical operating conditions of the vehicle are identified. Based on the statistical data of historical operating condition switching, a basic matrix of transition probabilities between different operating conditions is constructed. The driving features of the road ahead obtained by navigation are introduced to correct the transition probability values ​​of each operating condition in the transition probability matrix. Taking the currently identified typical operating condition as the initial state, the future operating condition prediction sequence is generated step by step in combination with the corrected transition probability matrix. The future operating condition prediction sequence is output to the subsequent multi-heat source temperature control optimization stage as the operating condition input in the prediction time domain.

8. The low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions according to claim 4, characterized in that, The solution for the multi-heat-source coordinated temperature control execution scheme also includes: The battery pack is divided into multiple independent temperature control zones according to the internal module layout. Each zone corresponds to an independent liquid circuit branch and heating execution unit. Differentiated heating intensity parameters are set according to the real-time aging decay amount per unit time and the current temperature deviation value of each zone. The historical temperature sequence, current heating power, and future operating condition prediction sequence information of each zone are collected and input into the gated loop unit network. The future short-term temperature change trend of each zone is output in real time. When the deviation between the actual collected temperature of the zone and the preset target temperature plus the predicted temperature change exceeds the preset deviation threshold, the heating power and liquid circuit flow control commands of the current cycle are dynamically corrected within the preset proportional adjustment range of the original control command.

9. The low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions according to claim 1, characterized in that, The process of collecting data throughout the entire operation to evaluate the temperature control effect and iteratively update all calculation models includes: After a single complete vehicle usage cycle, the temperature control effect evaluation value is calculated based on the full-process operation data collected in this cycle. The deviation between the evaluation value and the preset target value is used as the loss basis. The output layer parameters of the aging quantization model and reinforcement learning weight allocation model on the vehicle terminal are fine-tuned and local model gradients are generated. The federated learning framework is used to summarize the model gradient information generated locally by each vehicle terminal. Each terminal only uploads gradient data and not the original operation data. After receiving the gradient data uploaded by all terminals, the cloud uses the federated averaging algorithm to perform gradient aggregation according to the data volume of each terminal, updates the corresponding network parameters of the global basic model, and distributes the updated global basic model to each vehicle terminal through the vehicle network link to complete the iterative update of the local model on the vehicle terminal.

10. The low-temperature energy-saving temperature control management method for power batteries based on adaptive operating conditions according to claim 9, characterized in that, The evaluation of temperature control effectiveness by collecting data from the entire process includes: After a single complete vehicle usage cycle, based on the full-process operation data collected during that cycle, the energy consumption of the heating unit, the power consumption of the liquid circuit circulation, and the power consumption of the refrigeration compressor are extracted as input indicators. The average temperature rise rate of the battery, the amplitude of the single-cell temperature difference control, and the cumulative aging attenuation within the cycle are extracted as output indicators. Based on the evaluation orientation of minimizing input and maximizing output, a data envelopment analysis model is used to calculate the comprehensive relative efficiency value of temperature control input and output for that cycle as the temperature control effect evaluation value. The obtained comprehensive efficiency value is used as the target label. The current operating condition parameters, ambient temperature, battery state of charge, health status, and control parameters of each heat source are substituted into the global feature attribution interpretability algorithm to calculate the contribution value of each input feature to the temperature control effect. The features are sorted from high to low according to the absolute value of the contribution and the features exceeding the preset contribution threshold are selected as core driving factors. According to the attribute category of the core driving factors, the corresponding optimization objects are matched respectively. Operating condition and environmental factors correspond to weight allocation models, battery state factors correspond to aging quantification models, and control execution factors correspond to temperature control solution models. The parameter optimization adjustment direction of the corresponding models is output.

Images