A Thermal Management Method for Energy Storage Systems Based on Intelligent Prediction

Abstract

This application discloses a thermal management method for energy storage systems based on intelligent prediction, including the following steps: S1, through a sensor network deployed in the energy storage system, by scientifically planning the sensor network and using the Gaussian process regression method to reconstruct and supplement state parameters, the actual operating state of the area can be accurately reflected. A hybrid intelligent model using a long short-term memory network and a Gappy POD reduced-order model can effectively extract features such as cell temperature, charge / discharge rate, and environmental parameters. At the same time, it can identify potential thermal runaway risk areas and risk moments in advance, reserving sufficient time for risk handling, effectively curbing the occurrence and spread of thermal runaway accidents, and fully ensuring the safe operation of the energy storage system. Through model predictive control algorithms, differentiated thermal management interventions are implemented for areas with different temperature ranges to avoid problems of insufficient or excessive cooling in certain areas, thereby reducing the overall operating cost of the energy storage system.

Description

Technical Field

[0001] This application relates to the technical field of thermal management, and more particularly to a thermal management method for energy storage systems based on intelligent prediction. Background Technology

[0002] As a core support for mitigating fluctuations in new energy power generation and ensuring stable grid operation, energy storage systems have become the mainstream energy storage medium in current electrochemical energy storage systems. However, their electrochemical performance, cycle life, and operational safety are highly dependent on a stable temperature environment. Batteries continuously generate heat during charging and discharging. If the heat cannot be dissipated in a timely and uniform manner, it will lead to a sharp increase in local temperature and cause irreversible thermal runaway. Therefore, the thermal management system has become a core component that determines the economic efficiency and safety of the entire life cycle of the energy storage system, and its performance directly affects the operational reliability and overall energy efficiency of the energy storage system.

[0003] The sensors in existing energy storage systems mostly rely on experience for placement, resulting in untimely and incomplete parameter monitoring in areas with concentrated heat loads, and obvious monitoring blind spots. For areas where no sensors are deployed, the existing methods for supplementing parameters are relatively simple and cannot accurately reflect the true operating status of the area.

[0004] Secondly, existing thermal management systems cannot effectively predict temperature change trends and heat load distribution in the future. They can only respond to current temperature anomalies and cannot identify potential overheating hazards in advance. In terms of safety warnings, the existing mechanisms have relatively simple judgment criteria and can only trigger warnings after the temperature reaches a set threshold. They cannot detect potential thermal runaway risks in advance, resulting in insufficient time for risk handling and making it difficult to effectively curb the occurrence and spread of thermal runaway accidents, thus failing to fully guarantee the safety of system operation.

[0005] In addition, the cooling control of existing energy storage systems mostly adopts a fixed mode, with the operating parameters of the cooling equipment being preset fixed values. Some systems use a uniform cooling supply method, which cannot be targeted to control according to the temperature differences in different areas. This can easily lead to localized insufficient or excessive cooling. At the same time, the existing control method has not effectively optimized cooling energy consumption, resulting in the cooling system operating in a high-energy-consumption state for a long time, which increases the overall operating cost of the energy storage system.

[0006] Application content

[0007] This application aims to address, at least to some extent, the technical problems in the related art.

[0008] To achieve the above objectives, this application proposes a thermal management method for energy storage systems based on intelligent prediction, comprising the following steps:

[0009] S1. By deploying a sensor network in the energy storage system, the system collects cell temperature, charge / discharge rate, ambient temperature, ambient humidity, and module pressure parameters in real time. Then, it uses the Gaussian process regression method to reconstruct and supplement the state parameters of areas where no sensors are deployed, and builds a high spatiotemporal resolution operating database.

[0010] S2. Based on the operating database, a hybrid intelligent model combining a long short-term memory network and a Gappy POD reduced-order model is used to extract the nonlinear coupling characteristics between multiple physics fields, predict the three-dimensional distribution of heat load and the dynamic evolution trend of cell temperature in the energy storage system in the future, and identify potential thermal runaway risk areas and risk moments in advance based on the prediction results and thermal runaway criteria.

[0011] S3. Based on the predicted heat load distribution, temperature trend and risk identification results, dynamically adapt the coordinated cooling strategy with liquid cooling loop as the main component and air cooling loop as the auxiliary component. Based on the model predictive control algorithm, automatically adjust the flow distribution of each branch pump of the liquid cooling subsystem and the speed coordination of multiple fans in the air cooling subsystem to implement differentiated thermal management intervention for the energy storage system.

[0012] S4. Real-time acquisition of system status feedback data after dynamic regulation, and residual analysis and energy efficiency assessment of the data are performed on the output of the intelligent prediction steps. Online learning is used to continuously update the parameters and structure of the hybrid intelligent prediction model, forming a closed-loop adaptive management architecture from data acquisition, intelligent prediction to dynamic regulation.

[0013] In addition, the application may also include the following additional technical features:

[0014] Specifically, in step S1, the sensor network includes distributed temperature sensors, current sensors, voltage sensors, ambient temperature and humidity sensors, and pressure sensors. The spacing between the distributed temperature sensors does not exceed 1.5 times the diameter of the battery cell, and a dense deployment strategy is adopted in the battery cell tab area, module edge area, and liquid cooling plate inlet and outlet area.

[0015] Specifically, in step S1, the kernel function of the Gaussian process regression method adopts a linear combination of radial basis function and periodic kernel function, determines hyperparameters by maximizing the marginal likelihood function, and sets the spatial correlation length threshold to 0.8-1.2 times the physical size of the cell to ensure the spatial continuity and physical consistency of the reconstructed state parameters.

[0016] Specifically, in step S2, the input layer of the long short-term memory network receives the cell temperature sequence, charge / discharge rate time series data, and environmental parameters. The hidden layer is set with 3-5 LSTM unit layers, each containing 64-128 neurons. The output layer is connected to the three-dimensional heat load distribution prediction module and the temperature dynamic evolution prediction module, respectively. The Gappy POD reduced-order model uses intrinsic orthogonal decomposition to extract the dominant mode and reconstructs the flow field and temperature field of the missing region through the least squares method.

[0017] Specifically, in step S2, the thermal runaway criteria include: when the predicted temperature rise rate of a single cell exceeds 1.5℃ / s and the temperature difference between adjacent cells is greater than 15℃, or when the predicted module pressure increases by more than 20kPa within 30 seconds, it is determined to be a high-risk thermal runaway state, and the spatial coordinates of the risk area and the timestamp of the risk moment are generated.

[0018] Specifically, in step S3, the liquid cooling loop includes a main circulation loop and a multi-level branch loop, each branch loop corresponding to a battery module area. The air cooling path includes a top forced convection channel and a bottom natural convection compensation channel. The optimization objective function of the model predictive control algorithm is: J = a·ΔT + b·P, where a and b are weighting coefficients, ΔT is the deviation between the cell temperature and the set temperature, and P is the total power consumption of the cooling system.

[0019] Specifically, in step S3, the differentiated thermal management intervention includes: for areas where the predicted temperature exceeds 45°C, liquid cooling flow is preferentially allocated, and the flow allocation ratio is proportional to the predicted temperature gradient; for areas where the temperature is between 40-45°C, liquid cooling and air cooling are controlled in a coordinated manner, and the air cooling speed is dynamically adjusted according to the local heat flux density; for areas where the temperature is below 40°C, only basic air cooling is maintained.

[0020] Specifically, in step S4, the residual analysis adopts a sliding window mechanism, with the window length set to 1.5 times the prediction time domain. When the root mean square error of the residuals in three consecutive time windows exceeds a preset threshold, the online update of the hybrid intelligent prediction model is triggered. The energy efficiency evaluation indicators include energy consumption ratio per unit heat load, temperature uniformity index, and risk avoidance success rate. The model weight parameters are dynamically adjusted through a multi-objective optimization algorithm.

[0021] Specifically, in step S4, the online learning adopts a dual update strategy that combines incremental kernel ridge regression and elastic weight consolidation: incremental kernel ridge regression is used to correct the hyperparameters and output bias of the prediction model in real time, and elastic weight consolidation is used to apply decay constraints to the important connection weights of the model to prevent catastrophic forgetting.

[0022] In summary, the beneficial effects of the intelligent prediction-based thermal management method for energy storage systems proposed in this application are as follows:

[0023] 1. By scientifically planning the sensor network, a dense deployment strategy is adopted in areas with concentrated heat loads such as cell tabs, module edges, and liquid cooling plate inlets and outlets. The spacing of distributed temperature sensors is strictly controlled to effectively eliminate monitoring blind spots. For areas where no sensors are deployed, the Gaussian process regression method is used to reconstruct and supplement the state parameters, which can accurately reflect the real operating status of the area.

[0024] 2. The hybrid intelligent model, which employs a long short-term memory network and a Gappy POD reduced-order model, can effectively extract features such as cell temperature, charge / discharge rate, and environmental parameters. It can accurately predict the three-dimensional distribution of heat load and the dynamic evolution trend of cell temperature in the energy storage system in the future, enabling early prediction of temperature changes and heat load distribution. It can promptly identify potential overheating hazards and identify potential thermal runaway risk areas and moments in advance, allowing sufficient time for risk handling, effectively curbing the occurrence and spread of thermal runaway accidents, and fully ensuring the safe operation of the energy storage system.

[0025] 3. Through model predictive control algorithms, the flow distribution of each branch pump in the liquid cooling subsystem and the speed coordination of multiple fans in the air cooling subsystem are automatically adjusted. Differentiated thermal management interventions are implemented for different temperature ranges to avoid local insufficient or excessive cooling. At the same time, the balance between cooling effect and energy consumption is achieved by optimizing the cell temperature deviation and the total power consumption of the cooling system, effectively reducing the operating energy consumption of the cooling system and reducing the overall operating cost of the energy storage system. Attached Figure Description

[0026] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0027] Figure 1 This is a flowchart of a thermal management method for an energy storage system based on intelligent prediction, as described in this application.

[0028] Figure 2 This is a flowchart of the time-series prediction process of a hybrid model for a thermal management method for an energy storage system based on intelligent prediction, as described in this application.

[0029] Figure 3 This is a closed-loop flowchart of thermal management and model optimization for a thermal management method for energy storage systems based on intelligent prediction, as described in this application. Detailed Implementation

[0030] To make the technical means, inventive features, objectives, and effects of this application easier to understand, the application is further described below with reference to specific illustrations. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0031] The present application will now be described in further detail with reference to the accompanying drawings.

[0032] like Figures 1-3 As shown in the figure, an embodiment of this application provides a thermal management method for an energy storage system based on intelligent prediction, which includes the following steps:

[0033] S1. By deploying a sensor network in the energy storage system, the system collects cell temperature, charge / discharge rate, ambient temperature, ambient humidity, and module pressure parameters in real time. Then, it uses the Gaussian process regression method to reconstruct and supplement the state parameters of areas where no sensors are deployed, and builds a high spatiotemporal resolution operating database.

[0034] It should be noted that by deploying a distributed sensor network on the surface of the battery cells, module gaps, cooling circuits, and environmental areas of the energy storage system, the battery cell temperature, charge / discharge rate, ambient temperature, ambient humidity, and module pressure can be captured in real time. Among these, the battery cell temperature directly reflects the core requirements of thermal management, the charge / discharge rate is a key variable affecting the heat generation of the battery cell (the higher the charge / discharge rate, the more intense the heat generation due to the internal resistance of the battery cell), the ambient temperature and humidity determine the basic conditions of the system's heat dissipation environment, and the module pressure can indirectly reflect abnormal states such as battery cell expansion and thermal deformation, thus avoiding heat conduction obstruction caused by local deformation.

[0035] In actual deployment, sensors cannot provide full coverage (such as blind spots inside modules and between battery cells). Gaussian process regression is used to reconstruct and supplement state parameters. This method is based on probability and statistics theory. It uses the collected sensor data as training samples and constructs a kernel function to describe the nonlinear relationship between parameters. It makes probabilistic predictions of parameters in areas where no sensors are deployed. It can output the predicted values ​​of parameters and also provide the prediction uncertainty, thereby filling the data blind spots. Finally, it integrates all collected and reconstructed data to build a high-precision operating database in both time and spatial dimensions.

[0036] S2. Based on the operating database, a hybrid intelligent model using a long short-term memory network and a Gappy POD reduced-order model is adopted to extract the nonlinear coupling characteristics between multiple physics fields, predict the three-dimensional distribution of heat load and the dynamic evolution trend of cell temperature in the energy storage system in the future, and identify potential thermal runaway risk areas and risk moments in advance based on the prediction results and thermal runaway criteria.

[0037] It should be noted that LSTM networks excel at capturing long-term dependencies in time-series data. Based on historical and real-time multi-dimensional parameters in the running database, they can learn the dynamic evolution of cell temperature and heat load over time, solving problems such as gradient vanishing and short-term memory bias in traditional time-series forecasting.

[0038] The Gappy POD model is used to process high-dimensional multiphysics data (such as the three-dimensional temperature field and heat load distribution within the system). By reducing the order of high-dimensional data, core features are preserved and redundant information is removed, significantly reducing computational complexity and improving real-time prediction performance. The two work together. First, the Gappy POD model is used to extract the nonlinear coupling features between multiphysics fields (thermal field, electric field, force field) (such as the coupling between charge / discharge rate and cell heat generation, and the coupling between ambient temperature and heat dissipation efficiency). Then, the LSTM network uses these features to predict the three-dimensional distribution of heat load and the dynamic evolution trend of cell temperature in the energy storage system within a future period (such as the next 10-30 minutes).

[0039] Meanwhile, by combining preset thermal runaway criteria (such as cell temperature exceeding the critical value, temperature rise rate exceeding the threshold, abnormal sudden change in module pressure, etc.), the predicted results are threshold-judged to identify potential thermal runaway risk areas (such as areas with excessively high cell temperatures) and risk moments in advance, providing early warning basis for subsequent regulation.

[0040] S3. Based on the predicted heat load distribution, temperature trend and risk identification results, dynamically adapt the coordinated cooling strategy with liquid cooling loop as the main component and air cooling loop as the auxiliary component. Based on the model predictive control algorithm, automatically adjust the flow distribution of each branch pump of the liquid cooling subsystem and the speed coordination of multiple fans in the air cooling subsystem to implement differentiated thermal management intervention for the energy storage system.

[0041] It should be noted that the liquid cooling system has the advantages of high temperature control accuracy, high heat dissipation efficiency, and good temperature uniformity. It is used to undertake the heat dissipation task of the main heat load of the system (such as the dense area of ​​battery cells and the high heat generation area). The air cooling system has the characteristics of simple structure, low energy consumption and high flexibility. It is used to assist in adjusting the overall temperature of the system and to supplement the heat dissipation blind spots of the liquid cooling system (such as the gap between modules and the area not covered by the cooling circuit). The two work together to achieve the dual goals of efficient heat dissipation and energy consumption optimization.

[0042] During the regulation process, a model predictive control algorithm is adopted. This algorithm takes the predicted heat load and temperature trend for future periods as input, and solves the optimal control strategy through rolling optimization. It automatically adjusts the flow distribution of each branch pump in the liquid cooling subsystem and the speed coordination of multiple fans in the air cooling subsystem to implement differentiated thermal management intervention for the energy storage system: for high-risk areas and high heat load areas, the liquid cooling flow rate and air cooling speed are increased to enhance heat dissipation; for low load and normal temperature areas, the liquid cooling flow rate and fan speed are appropriately reduced to reduce energy consumption and achieve on-demand regulation.

[0043] S4. Real-time acquisition of system status feedback data after dynamic regulation, and residual analysis and energy efficiency assessment of the data are performed on the output of the intelligent prediction steps. Online learning is used to continuously update the parameters and structure of the hybrid intelligent prediction model, forming a closed-loop adaptive management architecture from data acquisition, intelligent prediction to dynamic regulation.

[0044] It should be noted that after the dynamic adjustment in step S3, the system status feedback data (such as the adjusted cell temperature, heat load distribution, cooling system energy consumption, etc.) is collected in real time through the sensor network. Then, the feedback data and the intelligent prediction output of step S2 are subjected to residual analysis (calculating the deviation between the predicted value and the actual value) to evaluate the accuracy of the prediction model. At the same time, energy efficiency evaluation is performed (analyzing the matching degree between cooling energy consumption and temperature control effect to judge the rationality of the control strategy).

[0045] Finally, by utilizing online learning algorithms, feedback data, residual information, and energy efficiency assessment results are continuously input into the hybrid intelligent prediction model, dynamically updating the model's parameters (such as the weights of LSTM and the reduced-order matrix of Gappy POD) and structure to optimize prediction accuracy. At the same time, the collaborative cooling strategy and MPC algorithm parameters are fine-tuned based on the energy efficiency assessment results, achieving continuous optimization of the model and control strategy, forming a closed-loop iteration, and ensuring that the thermal management system can adapt to changes in the operation of the energy storage system, environmental changes, and aging process, maintaining a high-efficiency thermal management effect in the long term.

[0046] In one embodiment of this application, in step S1, the sensor network includes distributed temperature sensors, current sensors, voltage sensors, ambient temperature and humidity sensors, and pressure sensors. The spacing between the distributed temperature sensors does not exceed 1.5 times the diameter of the battery cell, and a dense deployment strategy is adopted in the battery cell tab area, module edge area, and liquid cooling plate inlet and outlet area.

[0047] It should be noted that distributed temperature sensors are mainly used to capture the temperature distribution characteristics of battery cells and modules. Their spacing is strictly controlled to be no more than 1.5 times the diameter of the battery cell. The core purpose of this parameter setting is to ensure the spatial resolution of temperature monitoring, avoid missing local hot spots due to sparse deployment, and at the same time take into account monitoring cost and data transmission efficiency.

[0048] For critical areas such as the cell tab area, module edge area, and liquid cooling plate inlet and outlet areas, a denser deployment strategy is adopted. This is because these areas are where the battery system experiences the most drastic temperature changes and have the highest thermal risks. As a key node for current transmission, the cell tab is prone to localized heating due to contact resistance. The module edge area has poor heat dissipation conditions and is prone to temperature accumulation. The liquid cooling plate inlet and outlet areas directly affect the heat exchange efficiency of the entire cooling system. The denser deployment can accurately capture temperature changes in these critical areas, providing high-quality basic data for subsequent state reconstruction and risk prediction.

[0049] In one embodiment of this application, in step S1, the kernel function of the Gaussian process regression method adopts a linear combination of radial basis function and periodic kernel function, determines the hyperparameter by maximizing the marginal likelihood function, and sets the spatial correlation length threshold to 0.8-1.2 times the physical size of the cell to ensure the spatial continuity and physical consistency of the reconstructed state parameters.

[0050] It should be noted that this combination method has both the global fitting capability of radial basis functions and the ability of periodic kernel functions to capture time-series changes, and can effectively adapt to the characteristics of battery system state parameters (such as temperature and pressure) that have both spatial global distribution patterns and periodic fluctuations over time.

[0051] Hyperparameters were determined by maximizing the marginal likelihood function. Iterative optimization was used to minimize the deviation between the model's predictions and actual monitoring data, ensuring the model's fitting accuracy. Simultaneously, the spatial correlation length threshold was set to 0.8-1.2 times the physical size of the battery cell. This threshold range is based on the physical characteristics of the battery system. The spatial correlation length reflects the correlation of the spatial distribution of parameters. Binding it to the physical size of the battery cell effectively avoids spatial discontinuities or unreasonable fluctuations in the reconstruction results, ensuring the spatial continuity and physical consistency of the reconstructed state parameters, and enabling the reconstruction results to truly reflect the actual operating state of the battery system.

[0052] In one embodiment of this application, in step S2, the input layer of the long short-term memory network receives the cell temperature sequence, charge / discharge rate time series data and environmental parameters. The hidden layer is set with 3-5 LSTM unit layers, each containing 64-128 neurons. The output layer is connected to the three-dimensional heat load distribution prediction module and the temperature dynamic evolution prediction module, respectively. The Gappy POD reduced-order model uses intrinsic orthogonal decomposition to extract the dominant mode and reconstructs the flow field and temperature field of the missing region by least squares method.

[0053] It should be noted that the Long Short-Term Memory Network (LSTM) is used to process the dynamic time-series information of the battery system by capturing its long-term dependence on time-series data. Its input layer specifically receives cell temperature sequence (reflecting the current thermal state of the battery), charge and discharge rate time-series data (affecting the battery's heat generation rate), and environmental parameters (such as ambient temperature and humidity, affecting the battery's heat dissipation efficiency). These three types of input data work together to provide comprehensive input features for the prediction model.

[0054] The hidden layers consist of 3-5 LSTM unit layers, each containing 64-128 neurons. This parameter range was determined through multiple experiments and optimizations to ensure the model's ability to fit complex temporal features while avoiding overfitting and computational inefficiency caused by overly complex network structures. The output layers are connected to the three-dimensional heat load distribution prediction module and the temperature dynamic evolution prediction module, respectively, enabling simultaneous prediction of the spatial distribution of the battery system's heat load and the temperature change trend over time.

[0055] Meanwhile, the Gappy POD reduced-order model is used to solve the problem of efficient reconstruction of flow and temperature fields. It extracts the dominant modes of the flow and temperature fields through intrinsic orthogonal decomposition, retaining the core features of the data, eliminating redundant information, and effectively reducing computational complexity. Then, it uses the least squares method to accurately reconstruct the missing regions of the monitoring data, ensuring the integrity of the flow and temperature field data. Intrinsic orthogonal decomposition is a mathematical method for extracting feature information from discrete data. It describes multidimensional stochastic processes and extracts the essential features of complex systems through low-dimensional approximations. This method determines the basis functions by solving the eigenvalue problem of the covariance matrix, maximizing the energy content of low-order modes and rapidly decreasing the projection.

[0056] In one embodiment of this application, in step S2, the thermal runaway criterion includes: when the predicted temperature rise rate of a single cell exceeds 1.5℃ / s and the temperature difference between adjacent cells is greater than 15℃, or when the predicted module pressure increases by more than 20kPa within 30 seconds, it is determined to be a high-risk thermal runaway state, and the spatial coordinates of the risk area and the timestamp of the risk moment are generated.

[0057] It should be noted that a high-risk thermal runaway state is determined when one of the following two conditions is met: First, the predicted temperature rise rate of a single cell exceeds 1.5℃ / s, and the temperature difference between adjacent cells is greater than 15℃. An excessively rapid temperature rise rate indicates a sharp increase in heat generation inside the cell, while an excessively large temperature difference between adjacent cells indicates that the heat diffusion rate exceeds the safe range. The combination of these two conditions can effectively identify early signs of local thermal runaway.

[0058] Secondly, when the predicted module pressure increases by more than 20 kPa within 30 seconds, the rapid pressure rise originates from the decomposition of the electrolyte and the release of gas inside the cell, which is an important characteristic of thermal runaway entering a severe stage. This threshold setting is based on the battery module's withstand voltage limit and the evolution law of thermal runaway, enabling a rapid response to thermal runaway risks. Once a high-risk state is determined, the system will automatically generate the spatial coordinates of the risk area (precisely locating the specific location of thermal runaway for subsequent targeted intervention) and the timestamp of the risk moment (recording the specific time of the risk occurrence, providing a basis for subsequent fault tracing and model optimization).

[0059] In one embodiment of this application, in step S3, the liquid cooling loop includes a main circulation loop and a multi-level branch loop, each branch loop corresponding to a battery module area. The air cooling path includes a top forced convection channel and a bottom natural convection compensation channel. The optimization objective function of the model predictive control algorithm is: J = a·ΔT + b·P, where a and b are weighting coefficients, ΔT is the deviation between the cell temperature and the set temperature, and P is the total power consumption of the cooling system.

[0060] It should be noted that the main circulation loop is responsible for the coolant circulation drive and temperature control of the entire cooling system, while the multi-level branch loops correspond to different battery module areas. This structural design can realize independent cooling control of each module area, and facilitate differentiated flow distribution according to the heat load differences of different areas, thereby improving cooling efficiency.

[0061] The air-cooled airflow path includes a top forced convection channel and a bottom natural convection compensation channel. The top forced convection channel uses a fan to drive airflow for rapid heat dissipation, while the bottom natural convection compensation channel replenishes cool air, balancing the airflow distribution throughout the battery compartment and avoiding uneven heat dissipation caused by localized dead air zones. To achieve a balance between cooling effect and energy consumption, a model predictive control algorithm is used for optimization control: J = a·ΔT + b·P, where a and b are weighting coefficients that can be dynamically adjusted according to the battery system's operating conditions (such as charge / discharge status and ambient temperature), ΔT is the deviation between the cell temperature and the set temperature, reflecting the temperature control accuracy requirement, and P is the total power consumption of the cooling system, reflecting the energy consumption control requirement. This objective function minimizes the cooling system's energy consumption while ensuring temperature control accuracy, achieving optimal energy efficiency for the thermal management system.

[0062] In one embodiment of this application, in step S3, the differentiated thermal management intervention includes: for areas where the predicted temperature exceeds 45°C, liquid cooling flow is preferentially allocated, and the flow allocation ratio is proportional to the predicted temperature gradient; for areas where the temperature is between 40-45°C, liquid cooling and air cooling are coordinated and controlled, and the air cooling speed is dynamically adjusted according to the local heat flux density; for areas where the temperature is below 40°C, only basic air cooling is maintained.

[0063] It should be noted that for areas where the predicted temperature exceeds 45°C, these areas are high-load heat zones. If cooling is not strengthened in time, thermal runaway risks may occur. Therefore, liquid cooling flow is allocated first, and the flow allocation ratio is proportional to the predicted temperature gradient. The greater the temperature gradient, the more liquid cooling flow is allocated to ensure that high-temperature hot spots can be cooled down quickly.

[0064] For areas with temperatures between 40-45℃, which are considered medium heat load areas, a combined liquid cooling and air cooling control method is adopted. The liquid cooling system maintains a basic cooling flow rate, while the air cooling speed is dynamically adjusted according to the local heat flux density. The higher the heat flux density, the higher the air cooling speed, thus achieving a balance between cooling effect and energy consumption.

[0065] For areas with temperatures below 40°C, the heat load is low, and basic air cooling is sufficient to meet temperature control requirements. There is no need to start the liquid cooling system, which effectively reduces the overall power consumption of the cooling system and extends the battery life.

[0066] In one embodiment of this application, in step S4, the residual analysis adopts a sliding window mechanism, with the window length set to 1.5 times the prediction time domain. When the root mean square error of the residuals of three consecutive time windows exceeds a preset threshold, the online update of the hybrid intelligent prediction model is triggered. The energy efficiency evaluation indicators include the energy consumption ratio per unit heat load, the temperature uniformity index, and the risk avoidance success rate. The model weight parameters are dynamically adjusted through a multi-objective optimization algorithm.

[0067] It should be noted that the residual analysis employs a sliding window mechanism, with the window length set to 1.5 times the prediction time domain. This setting ensures that the window covers the entire prediction period, effectively capturing the changing trend of prediction deviations and avoiding misjudgments due to excessively short windows. When the root mean square error of the residuals exceeds a preset threshold for three consecutive time windows, it indicates that the model's prediction accuracy can no longer meet actual needs. At this point, the online update of the hybrid intelligent prediction model is triggered to promptly correct model deviations. Simultaneously, a multi-dimensional energy efficiency evaluation index system is constructed, and the model weight parameters are dynamically adjusted through a multi-objective optimization algorithm, enabling the model to adapt to complex scenarios such as battery system aging and changes in operating conditions, ensuring the long-term efficient operation of the thermal management system.

[0068] In one embodiment of this application, in step S4, online learning employs a dual update strategy combining incremental kernel ridge regression and elastic weight consolidation: incremental kernel ridge regression is used to correct the hyperparameters and output bias of the prediction model in real time, while elastic weight consolidation is used to apply decay constraints to the important connection weights of the model to prevent catastrophic forgetting.

[0069] It should be noted that incremental kernel ridge regression is used to correct the hyperparameters and output bias of the prediction model in real time. This algorithm can use newly collected monitoring data to incrementally update the model parameters without retraining the entire model, effectively reducing computational complexity, improving model update efficiency, and ensuring that the model can quickly adapt to the dynamic changes of the battery system.

[0070] Elastic weight consolidation addresses the catastrophic forgetting problem during online model updates. By imposing decay constraints on the weights of important connections, it preserves key features and historical knowledge learned by the model when updating parameters, preventing the model from forgetting previously useful information due to the introduction of new data. This ensures the continuity and stability of the model update process. The synergistic effect of these dual update strategies enables the model to respond quickly to changes in new data while maintaining long-term prediction accuracy, meeting the long-term operational requirements of the battery thermal management system.

[0071] It should be noted that, in this document, 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.

[0072] The present application and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present application. The actual structure is not limited to this. In conclusion, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present application, such design should fall within the protection scope of the present application.

Claims

1. A thermal management method for an energy storage system based on intelligent prediction, characterized in that, Includes the following steps: S1. Through the sensor network deployed in the energy storage system, the cell temperature, charge / discharge rate, ambient temperature, ambient humidity and module pressure parameters are collected in real time. The Gaussian process regression method is used to reconstruct and supplement the state parameters of the areas where no sensors are deployed, and to build a high spatiotemporal resolution operation database. S2. Based on the operating database, a hybrid intelligent model combining a long short-term memory network and a Gappy POD reduced-order model is used to extract the nonlinear coupling characteristics between multiple physical fields, predict the three-dimensional distribution of heat load and the dynamic evolution trend of cell temperature in the energy storage system in the future period, and identify potential thermal runaway risk areas and risk moments in advance based on the prediction results and thermal runaway criteria. S3. Based on the predicted heat load distribution, temperature trend and risk identification results, dynamically adapt the coordinated cooling strategy with liquid cooling loop as the main component and air cooling loop as the auxiliary component. Based on the model predictive control algorithm, automatically adjust the flow distribution of each branch pump of the liquid cooling subsystem and the speed coordination of multiple fans in the air cooling subsystem to implement differentiated thermal management intervention for the energy storage system. S4. Real-time acquisition of system status feedback data after dynamic regulation, and residual analysis and energy efficiency assessment of the data are performed on the output of the intelligent prediction steps. Online learning is used to continuously update the parameters and structure of the hybrid intelligent prediction model, forming a closed-loop adaptive management architecture from data acquisition, intelligent prediction to dynamic regulation.

2. The thermal management method for an energy storage system based on intelligent prediction according to claim 1, characterized in that, In step S1, the sensor network includes distributed temperature sensors, current sensors, voltage sensors, ambient temperature and humidity sensors, and pressure sensors. The spacing between the distributed temperature sensors does not exceed 1.5 times the diameter of the battery cell, and a dense deployment strategy is adopted in the battery cell tab area, module edge area, and liquid cooling plate inlet and outlet area.

3. The thermal management method for an energy storage system based on intelligent prediction according to claim 1, characterized in that, In step S1, the kernel function of the Gaussian process regression method is a linear combination of radial basis function and periodic kernel function. The hyperparameters are determined by maximizing the marginal likelihood function, and the spatial correlation length threshold is set to 0.8-1.2 times the physical size of the cell to ensure the spatial continuity and physical consistency of the reconstructed state parameters.

4. The thermal management method for an energy storage system based on intelligent prediction according to claim 1, characterized in that, In step S2, the input layer of the long short-term memory network receives the cell temperature sequence, charge / discharge rate time series data and environmental parameters. The hidden layer is set with 3-5 LSTM unit layers, each containing 64-128 neurons. The output layer is connected to the three-dimensional heat load distribution prediction module and the temperature dynamic evolution prediction module, respectively. The Gappy POD reduced-order model uses intrinsic orthogonal decomposition to extract the dominant mode and reconstructs the flow field and temperature field of the missing region through the least squares method.

5. The thermal management method for an energy storage system based on intelligent prediction according to claim 1, characterized in that, In step S2, the thermal runaway criteria include: when the predicted temperature rise rate of a single cell exceeds 1.5℃ / s and the temperature difference between adjacent cells is greater than 15℃, or when the predicted module pressure increases by more than 20kPa within 30 seconds, it is determined to be a high-risk thermal runaway state, and the spatial coordinates of the risk area and the timestamp of the risk moment are generated.

6. The thermal management method for an energy storage system based on intelligent prediction according to claim 1, characterized in that, In step S3, the liquid cooling loop includes a main circulation loop and a multi-level branch loop, with each branch loop corresponding to a battery module area. The air cooling path includes a top forced convection channel and a bottom natural convection compensation channel. The optimization objective function of the model predictive control algorithm is: J = a·ΔT + b·P, where a and b are weighting coefficients, ΔT is the deviation between the cell temperature and the set temperature, and P is the total power consumption of the cooling system.

7. The thermal management method for an energy storage system based on intelligent prediction according to claim 1, characterized in that, In step S3, the differentiated thermal management intervention includes: for areas where the predicted temperature exceeds 45°C, liquid cooling flow is preferentially allocated, and the flow allocation ratio is proportional to the predicted temperature gradient; for areas where the temperature is between 40-45°C, liquid cooling and air cooling are controlled in a coordinated manner, and the air cooling speed is dynamically adjusted according to the local heat flux density; for areas where the temperature is below 40°C, only basic air cooling is maintained.

8. The thermal management method for an energy storage system based on intelligent prediction according to claim 1, characterized in that, In step S4, the residual analysis adopts a sliding window mechanism with a window length set to 1.5 times the prediction time domain. When the root mean square error of the residuals in three consecutive time windows exceeds a preset threshold, the online update of the hybrid intelligent prediction model is triggered. The energy efficiency evaluation indicators include energy consumption ratio per unit heat load, temperature uniformity index, and risk avoidance success rate. The model weight parameters are dynamically adjusted through a multi-objective optimization algorithm.

9. The thermal management method for an energy storage system based on intelligent prediction according to claim 1, characterized in that, In step S4, the online learning adopts a dual update strategy that combines incremental kernel ridge regression and elastic weight consolidation: incremental kernel ridge regression is used to correct the hyperparameters and output bias of the prediction model in real time, and elastic weight consolidation is used to apply decay constraints to the important connection weights of the model to prevent catastrophic forgetting.

Images