Adaptive thermal management system for energy storage containers

Abstract

This invention discloses an adaptive thermal management system for energy storage containers, specifically relating to the field of thermal management technology for energy storage containers. By constructing a joint probabilistic model that integrates multi-source data, introducing Gaussian process regression and multi-precision CFD simulation fusion, and combining chaotic feature extraction and anomaly identification mechanisms, it achieves high-confidence dynamic modeling of complex airflow and temperature fields within the energy storage container and accurate identification of abnormal areas, thereby improving the adaptive control capability of the thermal management system. Furthermore, by collecting temperature and humidity data, constructing a condensation early warning mechanism and micro-airflow intervention strategy, and combining edge computing and reinforcement learning, it achieves real-time identification of condensation risks and optimization of control strategies, effectively solving the defects of traditional solutions such as slow response and control lag in thermal runaway and condensation problems, and improving the safety and reliability of the adaptive thermal management system for energy storage containers under dynamic operating conditions.

Description

Technical Field

[0001] This invention relates to the field of thermal management technology for energy storage containers, and more specifically, to an adaptive thermal management system for energy storage containers. Background Technology

[0002] Within the enclosed operation and maintenance space of an energy storage container, various heat dissipation components such as fans, deflectors, and cooling pipes need to be arranged in a limited space to meet the heat dissipation requirements of high-power battery packs or electronic devices during charging, discharging, and continuous operation. Due to the diverse and dense layout of equipment, airflow channels often form complex three-dimensional turbulent structures. Local eddies and dead zones can easily cause heat spot accumulation, resulting in uneven temperature distribution and reduced heat dissipation efficiency. In addition, frequent environmental conditions (such as load fluctuations and changes in external temperature) and internal layout adjustments exacerbate the dynamics and uncertainties of adaptive thermal management in energy storage containers, placing stringent requirements on the real-time response capability and predictive accuracy of the heat dissipation system.

[0003] Most existing technologies rely on offline CFD simulations or empirical formulas to evaluate airflow and temperature fields under static layouts, and use fixed PID control, preset fan curves, or single-objective optimization strategies to adjust heat dissipation components. These methods are difficult to capture transient eddy current characteristics and local temperature fluctuations in a timely manner when faced with layout changes, new equipment, or sudden operating conditions, resulting in problems such as large prediction errors, control lag, and uncontrollable energy consumption. At the same time, they lack precise quantification of sensor blind zone uncertainties, cannot achieve dynamic trade-offs between energy consumption and temperature balance, and lack a sound online self-learning and verification mechanism, thus making it difficult to effectively prevent the risk of local thermal runaway caused by chaotic airflow disturbances. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides an adaptive thermal management system for energy storage containers. Through multi-fidelity data fusion, adaptive chaotic detection, and multi-objective optimization control, dynamic airflow regulation is achieved to solve the risk of thermal runaway caused by chaotic disturbances mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an adaptive thermal management system for energy storage containers, comprising:

[0006] The data fusion and uncertainty quantification module simultaneously collects multi-source data from the energy storage container when it detects changes in the equipment layout or local temperature anomalies. The multi-source data includes at least real-time airflow data, CFD simulation data, and temperature distribution data. The module then fuses the multi-source data using a Gaussian process regression algorithm and uses a Bayesian method to quantify the interpolation error of the sensor blind zone, outputting a fused flow field map with confidence intervals.

[0007] The chaos intensity analysis module divides the fused flow field map into several grid cells. For grid cells whose confidence intervals meet preset requirements, a temporal convolutional network is used to extract multi-scale chaotic features and classify the chaos intensity levels.

[0008] The abnormal region identification module generates a dynamic temperature threshold by combining the standard deviation of the temperature gradient of the grid cells. Based on the fused flow field map and the dynamic temperature threshold, it constructs an error thermal map of the energy storage container. The error thermal parameter of each grid cell is the difference between the dynamic temperature threshold and the predicted temperature in the fused flow field map. When the chaos intensity level and the error thermal parameter exceed the limit together, it is marked as an abnormal region.

[0009] The environmental control command generation module, when the duration of the abnormal area exceeds the limit, separates the real-time airflow data of the abnormal area into low-frequency steady-state base flow components and high-frequency chaotic disturbance components through variational mode decomposition. It constructs a multi-objective optimization model with the goal of minimizing energy consumption and temperature fluctuation, and uses an improved particle swarm optimization algorithm to obtain the Pareto optimal solution of the deflection angle of the guide vane and the fan speed, generating environmental control commands that balance energy consumption and heat dissipation efficiency.

[0010] The reliability verification module collects the airflow field residual in real time after executing the environmental control command. The airflow field residual represents the difference between the actual measured airflow data and the CFD simulation data. If the airflow field residual exceeds the preset threshold, the closed-loop verification process is triggered.

[0011] The meta-reinforcement learning optimization module builds a digital twin platform for energy storage containers. In the digital twin platform, multi-condition disturbance data (including temperature disturbance, load disturbance, and layout disturbance) is randomly generated. The meta-reinforcement learning model is trained to quickly adapt to multi-condition disturbance scenarios while meeting safety constraints (avoiding the lag of manual debugging). The system searches for environmental control commands within the safety constraint boundaries and outputs a safety-verified meta-reinforcement learning model for subsequent use.

[0012] Preferably, the multi-source data fusion method is as follows: real-time airflow data, temperature distribution data, and CFD simulation data are fused using a Gaussian process regression algorithm to construct a joint probability model;

[0013] The joint probability model is used to complement and fuse CFD simulation data with real-time airflow data and temperature distribution data to generate a fused flow field map of the energy storage container (i.e., a prediction of the joint distribution of the flow field and temperature field).

[0014] The internal space of the energy storage container is divided into several grid cells, and the flow field diagram includes the airflow velocity, airflow direction angle and predicted temperature of each grid cell.

[0015] Define the kernel function of the joint probability model and optimize the hyperparameters. The optimization method is as follows: set the optimization objective to maximize the likelihood probability of the joint probability model output fused flow field map and measured data; use the conjugate gradient method to iteratively adjust the kernel function parameters until the joint probability model satisfies the optimization objective.

[0016] Preferably, the chaos intensity analysis module includes:

[0017] Based on the fused flow field map, grid cells that meet the confidence interval requirements are defined as high-confidence regions; gas velocity time series data and temperature time series data are extracted from the high-confidence regions.

[0018] Construct a temporal convolutional network containing 4 dilated convolutional layers with dilation factors of 1, 2, 4, and 8, a kernel size of 3, and 64 channels.

[0019] Input airflow velocity time series data in high confidence regions, output multi-scale mixed feature vectors. The dilated convolution structure of the temporal convolutional network covers time scales from seconds to minutes, adapting to battery charge and discharge cycles (typical cycle 30-60 seconds), ensuring that the feature extraction results reflect the dynamic balance between battery heat generation and dissipation, and avoiding transient noise interference.

[0020] Input the chaotic feature vectors of all grid cells to obtain the chaos intensity level of each region.

[0021] Preferably, when the chaos intensity level of a grid cell exceeds a preset value and the predicted temperature exceeds a dynamic temperature threshold, it is marked as an abnormal grid.

[0022] Morphological closing operations (dilation and erosion operations) are performed on abnormal meshes that meet the triggering conditions to generate a set of coordinates of connected regions, providing spatial positioning input for the targeted control of the environmental control command generation module.

[0023] Preferably, the environmental control instruction generation process includes the following steps:

[0024] When the duration of the abnormal region exceeds the limit, variational mode decomposition technology is used to separate the low-frequency steady-state fundamental current component and the high-frequency steady-state fundamental current component of the airflow in the abnormal region. The low-frequency steady-state fundamental current component represents the average motion of the airflow that maintains heat dissipation; the high-frequency steady-state fundamental current component represents the eddy current pulsation and noise between the battery clusters.

[0025] A multi-objective optimization model is constructed with the objectives of minimizing energy consumption and temperature fluctuation. The safety constraints of the multi-objective optimization model include the safe range of airflow velocity and the upper limit of fan speed.

[0026] An improved particle swarm optimization algorithm was used to find the Pareto optimal solution for the deflection angle of the guide vane and the speed of the fan. Environmental control commands were constructed based on the Pareto optimal solution. A multi-objective optimization model was used to balance energy consumption and heat dissipation efficiency and to suppress the risk of thermal runaway caused by chaotic disturbances.

[0027] Preferably, when the closed-loop verification process is triggered, it includes:

[0028] The residual stability coefficient is calculated based on the ratio between the current airflow field residual and the historical residual mean.

[0029] The historical residual baseline is calculated based on the average value of historical residual data within the sliding time window. The residual stability coefficient is dynamically generated by the ratio of the current residual to the historical residual baseline. The residual stability coefficient is used to quantify the transient degree of the energy storage container environment deviating from the steady state.

[0030] Based on the fluctuation range of the residual stability coefficient, the confidence interval threshold and the policy search direction of the meta-reinforcement learning model are dynamically adjusted.

[0031] When the residual stability coefficient exceeds the preset threshold, the confidence interval threshold is tightened and the weight of the meta-reinforcement learning model for exploring new strategies is increased; when the residual stability coefficient is lower than the preset threshold, the confidence interval threshold is relaxed and validated strategies are called first.

[0032] Preferably, the system further includes:

[0033] The condensation risk disturbance module is used to identify temperature and humidity data of potential condensation risk points inside the energy storage container. When the condensation risk score exceeds the preset value, the sensor data is input into the condensation risk assessment model. Based on the boundary reference conditions generated by CFD simulation, combined with the local dew point proximity rate, airflow disturbance attenuation coefficient and temperature difference change rate, the module outputs the condensation risk score for each condensation risk point and generates a high-risk area identifier. It also generates condensation disturbance instructions for the high-risk area to enhance airflow and suppress condensation formation.

[0034] Preferably, the condensation risk disturbance module includes a temperature and humidity sensing unit, a condensation risk assessment model, a local behavior baseline construction unit, and an anomaly detection model;

[0035] The temperature and humidity sensing unit is used to collect temperature and humidity data in different spatial locations inside the energy storage container in real time.

[0036] The condensation risk assessment model refers to a function model that takes local temperature and humidity data as input, combines simulation boundary conditions, and outputs a condensation risk score to dynamically assess the possibility of condensation forming in a local space.

[0037] The local behavior baseline construction unit is used to construct the local environmental behavior baseline of each monitoring area inside the energy storage container under normal operating conditions, and to serve as a comparison reference for risk scoring.

[0038] The anomaly detection model is used to identify the degree of baseline deviation, identify condensation triggering modes (based on dew point approach behavior and humidity surge characteristics), and assist in determining whether the condensation risk score increases rapidly due to non-periodic fluctuations.

[0039] Preferably, the operation process of the condensation risk assessment model includes:

[0040] The local dew point proximity rate, airflow disturbance attenuation coefficient, and temperature difference change rate are obtained for each risk point; these are then mapped to a unified scoring scale using normalization functions. The local dew point proximity rate refers to the degree of proximity between the real-time air temperature and the current dew point temperature of a risk point, used to measure the trend of water vapor in the air approaching saturation. The airflow disturbance attenuation coefficient refers to the attenuation magnitude of the local airflow disturbance intensity per unit time, used to describe the natural weakening of the local air disturbance intensity in spatial dead zones. The temperature difference change rate refers to the rate of change of the difference between the local air temperature and the temperature of adjacent areas per unit time, used to capture the trend of abrupt changes in microenvironmental temperature.

[0041] The relative humidity is combined with the local air temperature, and the local absolute humidity is calculated using the ideal gas law.

[0042] Using weighted parameters derived from experience or training, the standardized local dew point proximity rate, airflow disturbance attenuation coefficient, temperature difference change rate, and local absolute humidity are weighted and summed to obtain the condensation risk score for each risk point.

[0043] Preferably, the execution of the condensation disturbance command is limited by the power consumption limit and the main airflow structure constraints. The disturbance operation needs to enhance the airflow disturbance locally while dynamically compensating through the disturbance amplitude adjustment mechanism to control the overall main airflow field offset within an acceptable range.

[0044] The technical effects and advantages of this invention are as follows:

[0045] (1) The energy storage container adaptive thermal management system provided by the present invention constructs a joint probability model that integrates multi-source data and introduces Gaussian process regression and multi-precision CFD simulation fusion. Combined with chaotic feature extraction and anomaly identification mechanism, it realizes high-confidence dynamic modeling of complex airflow and temperature field in energy storage container and accurate identification of abnormal areas. It effectively solves the problems of slow response to transient disturbances, inaccurate prediction of local thermal runaway and coarse energy consumption control in the prior art, and improves the adaptive regulation capability and operational safety of thermal management system under dynamic working conditions.

[0046] (2) The energy storage container adaptive thermal management system provided by the present invention collects temperature and humidity data of the energy storage container, combines CFD simulation and multi-parameter coupling model to evaluate the risk of local condensation in real time, constructs a high-precision condensation early warning mechanism, and identifies condensation triggering mode based on edge computing, triggers the local disturbance control unit to implement micro-airflow excitation and adjustment, and realizes dynamic intervention on potential cold areas; by introducing reinforcement learning mechanism to optimize control strategy parameters, a data-driven closed-loop adaptive thermal management system is formed, which effectively solves the problem of not being able to identify and intervene in local condensation accumulation in the prior art, improves the response speed and local adaptability of the thermal management system in complex environments, and ensures the safety and reliability of energy storage equipment operation. Attached Figure Description

[0047] Figure 1 This is a block diagram of the adaptive thermal management system for energy storage containers of the present invention.

[0048] Figure 2 This is a flowchart of the adaptive thermal management process for the energy storage container of the present invention.

[0049] Figure 3 This is a block diagram of the condensation risk disturbance module of the present invention. Detailed Implementation

[0050] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0051] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0052] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0053] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0054] Example 1, see Figure 1 Block diagram of the adaptive thermal management system for energy storage containers and Figure 2 The present invention provides a flowchart of the adaptive thermal management process for energy storage containers. Figure 1 The adaptive thermal management system for the energy storage container shown includes:

[0055] The data fusion and uncertainty quantification module simultaneously collects multi-source data from the energy storage container when it detects changes in the equipment layout or local temperature anomalies. The multi-source data includes at least real-time airflow data, CFD simulation data, and temperature distribution data. The module then fuses the multi-source data using a Gaussian process regression algorithm and uses a Bayesian method to quantify the interpolation error of the sensor blind zone, outputting a fused flow field map with confidence intervals.

[0056] The chaos intensity analysis module divides the fused flow field map into several grid cells. For grid cells whose confidence intervals meet preset requirements, a temporal convolutional network is used to extract multi-scale chaotic features and classify the chaos intensity levels.

[0057] The abnormal region identification module generates a dynamic temperature threshold by combining the standard deviation of the temperature gradient of the grid cells. Based on the fused flow field map and the dynamic temperature threshold, it constructs an error thermal map of the energy storage container. The error thermal parameter of each grid cell is the difference between the dynamic temperature threshold and the predicted temperature in the fused flow field map. When the chaos intensity level and the error thermal parameter exceed the limit together, it is marked as an abnormal region.

[0058] The explanation is as follows: The combined exceeding of the limits for both the chaos intensity level and the error thermodynamic parameter refers to:

[0059] After normalizing the difference between the dynamic temperature threshold and the predicted temperature in the fused flow field diagram, the difference is weighted and summed with the chaos intensity level to generate an anomaly risk coefficient. The anomaly risk coefficient exceeds the preset value.

[0060] The environmental control command generation module, when the duration of the abnormal area exceeds the limit, separates the real-time airflow data of the abnormal area into low-frequency steady-state base flow components and high-frequency chaotic disturbance components through variational mode decomposition. It constructs a multi-objective optimization model with the goal of minimizing energy consumption and temperature fluctuation, and uses an improved particle swarm optimization algorithm to obtain the Pareto optimal solution of the deflection angle of the guide vane and the fan speed, generating environmental control commands that balance energy consumption and heat dissipation efficiency.

[0061] The reliability verification module collects the airflow field residual in real time after executing the environmental control command. The airflow field residual represents the difference between the actual measured airflow data and the CFD simulation data, which is used to determine whether the control command is effective. If it is invalid, it is re-verified. If the airflow field residual exceeds the preset threshold, an early warning and closed-loop verification process are triggered. The closed-loop verification process verifies whether the execution of the environmental control command is abnormal. If the execution of the environmental control command is normal, it proceeds to the next module.

[0062] The meta-reinforcement learning optimization module builds a digital twin platform for energy storage containers. In the digital twin platform, multi-condition disturbance data (including temperature disturbance, load disturbance, and layout disturbance) is randomly generated. The meta-reinforcement learning model is trained to quickly adapt to multi-condition disturbance scenarios while meeting safety constraints (avoiding the lag of manual debugging). The system searches for environmental control commands within the safety constraint boundaries and outputs a safety-verified meta-reinforcement learning model for subsequent use.

[0063] In this embodiment of the invention, it is necessary to further explain that the multi-source data acquisition method is as follows: pressure sensors are deployed on the top of the energy storage container and in the gaps between the battery clusters; real-time airflow data, including airflow velocity and airflow direction angle, is obtained through the pressure sensor array; a CFD simulation model is invoked to output CFD simulation data; temperature distribution data is collected through an infrared thermal imager to generate a two-dimensional temperature matrix; the coordinates of the pressure sensors are mapped with the CFD simulation data to ensure spatial alignment between the real-time airflow data and the CFD simulation data, and to ensure spatial correspondence during data fusion.

[0064] In one possible embodiment, the CFD simulation model is a multi-precision CFD simulation model, outputting low, medium, and high fidelity CFD simulation data. The hierarchical setting of the multi-precision CFD simulation model, through precision-velocity grading and the complementarity of physical features, provides a cross-scale data foundation for Gaussian process regression, satisfying real-time control requirements while improving prediction reliability through fusion, thus resolving the conflict between dynamic layout and limited computing resources in energy storage container scenarios. If the high-fidelity CFD simulation model takes too long to compute (hours), it cannot participate in online fusion in real time; in this case, the high-fidelity CFD simulation model is only used for offline weight calibration, and online fusion only uses low / medium-fidelity and high-fidelity CFD simulation models. The low, medium, and high-fidelity CFD simulation models are defined according to the following rules:

[0065] Low-fidelity model: The minimum mesh size is 1 / 100 to 1 / 50 of the characteristic length L of the energy storage container. RANS equations and k-∈ turbulence model are used. The single simulation time Tcalc≤10s.

[0066] Medium-fidelity model: minimum mesh size is L / 200 to L / 100, using LES or TRANS equations, single simulation time is 10s. <Tcalc≤300s;

[0067] High-fidelity model: minimum mesh size is L / 500 to L / 200, using DNS or fully analytical LES, single simulation time Tcalc>300s.

[0068] In this embodiment of the invention, it is necessary to further explain that the confidence interval is calculated as follows: using the covariance matrix output by the Gaussian process regression, the prediction standard deviation of each grid cell (reflecting the credibility of the grid cell fusion result) is calculated. Based on the normal distribution assumption of the Gaussian process, the confidence interval boundary is the mean ± 2 times the standard deviation, including the confidence intervals of flow velocity and temperature.

[0069] The explanation explains that by processing only high-confidence regions and ignoring low-confidence regions, error propagation is avoided. If a certain CFD simulation model has a systematic bias in a specific region (such as a high-temperature region), the contribution of high-error CFD simulation models is suppressed by inverse weighting of the Bayesian posterior covariance matrix, thus preventing erroneous predictions from affecting the fusion results.

[0070] In this embodiment of the invention, it is necessary to further explain that real-time airflow data, temperature distribution data, and CFD simulation data are fused using a Gaussian process regression algorithm to construct a joint probability model. This joint probability model is used to complementarily fuse CFD simulation data with real-time airflow and temperature distribution data to generate a fused flow field map of the energy storage container (i.e., a prediction of the joint distribution of the flow field and temperature field). The internal space of the energy storage container is divided into several grid cells, and the fused flow field map includes the airflow velocity, airflow direction angle, and predicted temperature for each grid cell. The kernel function of the joint probability model is defined, and its hyperparameters are optimized. The hyperparameters are optimized by setting the optimization objective to maximize the likelihood probability of the fused flow field map output by the joint probability model compared to the measured data. The kernel function parameters are iteratively adjusted using the conjugate gradient method until the joint probability model satisfies the optimization objective. Gas velocity time-series data and temperature time-series data are used.

[0071] The kernel function is a joint kernel function formed by the Matern kernel function and the linear kernel function. The Matern kernel function is used to capture the smooth variation characteristics of airflow velocity in adjacent regions (e.g., the closer the distance, the stronger the correlation of flow velocity). The linear kernel function is used to characterize the relationship between the prediction bias of CFD simulation models of different accuracies and the change of location. The Matern kernel function is used to force the airflow field to conform to spatial continuity and avoid the anti-physics prediction of pure data-driven methods. The linear kernel function is used to make the advantages of multi-accuracies CFD simulation models complement each other. The low-fidelity CFD simulation model is used to preserve the global trend, and the high-fidelity CFD simulation model is used to extract local accuracy. The joint kernel function meets the time constraints of online control through analytical or fast approximation calculation.

[0072] In this embodiment of the invention, it needs to be further explained that the multi-scale chaotic features refer to the extraction of chaotic characteristic parameters at different time scales through a temporal convolutional network. These multi-scale chaotic features cover time windows from seconds to minutes, adapting to battery charge-discharge cycles and ensuring the capture of transient eddies and periodic fluctuations, including:

[0073] Singular spectral entropy: The entropy value of energy distribution is calculated based on the singular value decomposition of the flow velocity signal. The higher the entropy value, the stronger the energy dispersion of the flow field (the higher the degree of chaos).

[0074] Recursive graph density: Recursive matrix is ​​generated by reconstructing phase space, and the probability of trajectory repetition is statistically analyzed to quantify determinism (the lower the density, the worse the determinism);

[0075] Permutation entropy: used to quantify the randomness of time-series signals; the higher the value, the stronger the chaos.

[0076] Recursive Quantitative Analysis - Determinism: The proportion of deterministic structures in the recursive matrix; the lower the value, the stronger the chaos.

[0077] In this embodiment of the invention, it should be further explained that the chaos intensity analysis module includes:

[0078] Based on the fused flow field map, the grid cells that meet the confidence interval requirements are defined as high confidence regions; gas velocity time series data and temperature time series data are extracted from the high confidence regions, and in one possible embodiment, the time surface aperture length is set to 60 seconds and the sliding step size is 10 seconds.

[0079] Explanation: The high-confidence region is the core area for airflow chaos detection. By eliminating low-confidence regions (such as container corners or equipment obstruction blind spots), the interference of interpolation errors on subsequent analysis is avoided.

[0080] A temporal convolutional network is constructed, consisting of four dilated convolutional layers with dilation factors of 1, 2, 4, and 8, a kernel size of 3, and 64 channels. It takes as input time-series airflow velocity data from high-confidence regions and outputs multi-scale mixed feature vectors. The dilated convolutional structure of the temporal convolutional network covers time scales from seconds to minutes, adapting to battery charge-discharge cycles (typically 30-60 seconds), ensuring that the feature extraction results reflect the dynamic balance between battery heat generation and dissipation, and avoiding transient noise interference.

[0081] Input the chaotic feature vectors of all grid cells to obtain the chaos intensity level of each region.

[0082] Explanation: Calculate the temperature gradient magnitude in the high-confidence region. The corresponding 90th percentile Q 90 (The 90th percentile represents the temperature gradient of 90% of the grid cells below this value and 10% of the grid cells above this value). The dynamic temperature threshold is defined as the average temperature of the grid region multiplied by twice Q. 90 The sum; when the chaos intensity level of a grid cell exceeds a preset value and the predicted temperature exceeds the dynamic temperature threshold, it is marked as an abnormal grid;

[0083] Morphological closing operations (dilation and erosion operations) are performed on abnormal meshes that meet the triggering conditions to generate a set of coordinates of connected regions, providing spatial positioning input for the targeted control of the environmental control command generation module.

[0084] In this embodiment of the invention, it needs to be further explained that the environmental control instruction generation process includes the following steps:

[0085] When the duration of the abnormal region exceeds the limit, variational mode decomposition technology is used to separate the low-frequency steady-state fundamental current component and the high-frequency steady-state fundamental current component of the airflow in the abnormal region. The low-frequency steady-state fundamental current component represents the average motion of the airflow that maintains heat dissipation; the high-frequency steady-state fundamental current component represents the eddy current pulsation and noise between the battery clusters.

[0086] A multi-objective optimization model is constructed with the objectives of minimizing energy consumption and temperature fluctuation. The safety constraints of the multi-objective optimization model include the safe range of airflow velocity and the upper limit of fan speed.

[0087] The Pareto optimal solution for the deflection angle of the guide vane and the speed of the fan is obtained by improving the particle swarm optimization algorithm, and environmental control commands are constructed based on the Pareto optimal solution. A multi-objective optimization model is used to balance energy consumption and heat dissipation efficiency and suppress the risk of thermal runaway caused by chaotic disturbances.

[0088] In one possible implementation, high-power discharge effectively minimizes temperature fluctuations, while low-power discharge prioritizes reducing total power consumption to meet the heat dissipation requirements of different operating conditions.

[0089] The improved particle swarm optimization algorithm achieves the search for the Pareto optimal solution set by dynamically adjusting the inertia weight parameter, integrating a multi-objective optimization mechanism, and an adaptive constraint handling strategy. Specifically, it includes:

[0090] The inertia weight is dynamically adjusted based on the number of iterations and the particle swarm distribution density. In the early stage, a large weight value is maintained to enhance the global search capability, and in the later stage, it is gradually reduced to improve the local optimization accuracy.

[0091] We employ fast non-dominated sorting and crowding assessment to perform hierarchical screening of the solution set, prevent local optima by maintaining diversity, and introduce an elitist retention strategy to inject historical optimal solutions into the iterative process to accelerate convergence.

[0092] A dual-objective optimization function is constructed with the core of wind turbine energy consumption and battery cluster temperature fluctuation amplitude. The objective weight coefficients are dynamically allocated by combining the energy ratio of the low-frequency steady-state component and the high-frequency transient component of the airflow signal extracted by variational mode decomposition. When the low-frequency component is dominant, the focus is on energy consumption optimization, and when the high-frequency component is significant, the temperature suppression requirement is strengthened.

[0093] A penalty function mechanism is constructed to address the safe range of airflow velocity and the upper limit of fan speed. Constraint violation penalty terms are superimposed on the objective function to ensure that the solution set conforms to physical constraints.

[0094] Based on the normalized comprehensive decision index, the Pareto front solution set is prioritized, and the solution with the highest balance between energy consumption and temperature fluctuation is selected as the global optimal control command for the deflection angle of the guide vane and the speed of the fan, so as to achieve adaptive and coordinated optimization of heat dissipation performance and energy efficiency.

[0095] In this embodiment of the invention, it needs to be further explained that triggering the closed-loop verification process includes:

[0096] The residual stability coefficient is calculated based on the ratio between the current airflow field residual and the historical residual mean.

[0097] The historical residual baseline is calculated based on the average value of historical residual data within the sliding time window. The residual stability coefficient is dynamically generated by the ratio of the current residual to the historical residual baseline. The residual stability coefficient is used to quantify the transient degree of the energy storage container environment deviating from the steady state.

[0098] Based on the fluctuation range of the residual stability coefficient, the confidence interval threshold and the policy search direction of the meta-reinforcement learning model are dynamically adjusted.

[0099] When the residual stability coefficient exceeds the preset threshold, the confidence interval threshold is tightened and the weight of the meta-reinforcement learning model for exploring new strategies is increased; when the residual stability coefficient is lower than the preset threshold, the confidence interval threshold is relaxed and validated strategies are called first.

[0100] Background: Existing energy storage containers are typically equipped with temperature control devices to regulate the internal temperature and humidity of the container through overall airflow, in order to avoid the risk of cell performance degradation or thermal runaway caused by changes in environmental conditions. However, due to the complex internal structure of the container, especially in local areas such as pipe bends, insulation interfaces, and water collection points, dead air zones can easily form, causing deviations between the local temperature and humidity environment and the overall control environment. Under certain conditions, these small areas of stagnant flow can easily trigger local condensation, leading to the accumulation of condensate. This not only reduces local heat exchange efficiency but may also cause serious hidden dangers such as corrosion of the heat transfer medium and deterioration of electrical insulation performance, threatening the long-term stable operation of the energy storage system.

[0101] To address the aforementioned issues, existing technologies primarily rely on centralized temperature and humidity control equipment on the top of the container for global temperature and humidity regulation. However, they lack the ability to fine-grainedly identify and dynamically intervene in local microenvironments. When condensation accumulates in localized areas, the system cannot detect and respond in real time, leading to the continued spread of the localized cold zone effect. This ultimately impacts the overall thermal management performance and increases the risk of system failure. Furthermore, existing control strategies fail to effectively integrate local climate characteristics with dynamic risk assessments and lack a data-driven adaptive adjustment mechanism. Therefore, Example 2 is designed based on these shortcomings.

[0102] Example 2, see Figure 3 The block diagram of the condensation risk disturbance module in this embodiment of the invention differs from that in Embodiment 1 in that it further includes:

[0103] The condensation risk disturbance module identifies temperature and humidity data at potential condensation risk points within the energy storage container. When the condensation risk score exceeds a preset value, the sensor data is input into the condensation risk assessment model. Based on boundary reference conditions generated by CFD simulation, and combined with local dew point proximity rate, airflow disturbance attenuation coefficient, and temperature difference change rate, the model outputs a condensation risk score for each condensation risk point and generates a high-risk area identifier. It also generates condensation disturbance commands for high-risk areas to enhance airflow and suppress condensation formation. The condensation risk disturbance module includes a temperature and humidity sensing unit, a condensation risk assessment model, a local behavior baseline construction unit, and an anomaly detection model.

[0104] The temperature and humidity sensing unit is used to collect temperature and humidity data in different spatial locations inside the energy storage container in real time.

[0105] The condensation risk assessment model refers to a function model that takes local temperature and humidity data as input, combines simulation boundary conditions, and outputs a condensation risk score to dynamically assess the possibility of condensation forming in a local space.

[0106] The local behavior baseline construction unit is used to construct the local environmental behavior baseline of each monitoring area inside the energy storage container under normal operating conditions, and to serve as a comparison reference for risk scoring.

[0107] The anomaly detection model is used to identify the degree of baseline deviation, identify condensation triggering modes (based on dew point approach behavior and humidity surge characteristics), and assist in determining whether the condensation risk score increases rapidly due to non-periodic fluctuations.

[0108] The explanation is as follows: Boundary reference conditions refer to the preset operating boundary dataset based on computational fluid dynamics simulation of the airflow and heat transfer state inside the energy storage container, which is used as a theoretical benchmark for the analysis of local microclimate monitoring data.

[0109] The explanation is as follows: the local environmental behavior baseline is used to construct the temperature and humidity change benchmark curves of each monitoring area inside the energy storage container under normal operating conditions, and to serve as a comparison reference for condensation risk scores. The construction of the local environmental behavior baseline relies on the multi-scale sliding window method to identify stable patterns and fluctuation boundaries in the time series, forming a behavior mapping of the "normal" climate behavior of the area, which is used for subsequent condensation trigger condition judgment and anomaly detection.

[0110] In this embodiment of the invention, it needs to be further explained that the operation process of the condensation risk assessment model includes:

[0111] The local dew point proximity rate, airflow disturbance attenuation coefficient, and temperature difference change rate are obtained for each risk point; these are then mapped to a unified scoring scale using normalization functions. The local dew point proximity rate refers to the degree of proximity between the real-time air temperature and the current dew point temperature of a risk point, used to measure the trend of water vapor in the air approaching saturation. The airflow disturbance attenuation coefficient refers to the attenuation magnitude of the local airflow disturbance intensity per unit time, used to describe the natural weakening of the local air disturbance intensity in spatial dead zones. The temperature difference change rate refers to the rate of change of the difference between the local air temperature and the temperature of adjacent areas per unit time, used to capture the trend of abrupt changes in microenvironmental temperature.

[0112] The relative humidity is combined with the local air temperature, and the local absolute humidity is calculated using the ideal gas law.

[0113] Using weighted parameters derived from experience or training, the standardized local dew point proximity rate, airflow disturbance attenuation coefficient, temperature difference change rate, and local absolute humidity are weighted and summed to obtain the condensation risk score for each risk point.

[0114] In this embodiment of the invention, it is necessary to further explain that when the condensation risk score of the condensation risk point exceeds the preset requirement and the anomaly detection model identifies the condensation triggering mode, a condensation disturbance command is generated. The condensation disturbance command includes wind speed, wind direction and disturbance duration parameters, which are used to control the local guide vanes or air supply unit to implement airflow disturbance operation, and continuously track the risk score change trend after the disturbance is executed. The disturbance parameters are optimized in a closed loop through a reinforcement learning mechanism to improve the intervention effect and energy efficiency ratio.

[0115] The explanation is that the execution of the condensation disturbance command is limited by the power consumption limit and the main airflow structure constraints. The disturbance operation needs to enhance the airflow disturbance locally while dynamically compensating through the disturbance amplitude adjustment mechanism to control the overall main airflow field offset within an acceptable range.

[0116] In this embodiment of the invention, it is necessary to further explain that environmental data of the corresponding area after the execution of the condensation disturbance command is collected to evaluate the disturbance response quality coefficient and update the condensation risk assessment model; the changing trend of the condensation risk score in the high-risk area is tracked. If the condensation risk score decreases, it is marked as the strategy is effective and the current disturbance control parameters are recorded; otherwise, the strategy is invalid. If the strategy is invalid, the control parameter weights, including the air supply direction, speed and action time, are iteratively updated through reinforcement learning. The reinforcement learning process is based on a constrained exploration mechanism, that is, under safety constraints, the disturbance amplitude and time window are limited, and the control parameters are iteratively updated gradually within a controllable range.

[0117] In summary, this invention collects temperature and humidity data from energy storage containers, combines CFD simulation with a multi-parameter coupled model, to assess local condensation risk in real time, constructs a high-precision condensation early warning mechanism, and identifies condensation triggering modes based on edge computing, triggering the local disturbance control unit to implement micro-airflow excitation and adjustment, thereby achieving dynamic intervention in potential cold zones. By introducing a reinforcement learning mechanism to optimize control strategy parameters, a data-driven closed-loop adaptive thermal management system is formed, effectively solving the problem of not being able to identify and intervene in local condensation accumulation in existing technologies. This improves the response speed and local adaptability of the thermal management system in complex environments, ensuring the safety and reliability of energy storage equipment operation.

[0118] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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. An adaptive thermal management system for energy storage containers, characterized in that, include: The data fusion and uncertainty quantification module simultaneously collects multi-source data from the energy storage container when it detects changes in the equipment layout or local temperature anomalies. The multi-source data includes at least real-time airflow data, CFD simulation data, and temperature distribution data. The module then fuses the multi-source data using a Gaussian process regression algorithm and uses a Bayesian method to quantify the interpolation error of the sensor blind zone, outputting a fused flow field map with confidence intervals. The chaos intensity analysis module divides the fused flow field map into several grid cells. For grid cells whose confidence intervals meet preset requirements, a temporal convolutional network is used to extract multi-scale chaotic features and classify the chaos intensity levels. The abnormal region identification module generates a dynamic temperature threshold by combining the standard deviation of the temperature gradient of the grid cells. Based on the fused flow field map and the dynamic temperature threshold, it constructs an error thermal map of the energy storage container. The error thermal parameter of each grid cell is the difference between the dynamic temperature threshold and the predicted temperature in the fused flow field map. When the chaos intensity level and the error thermal parameter exceed the limit together, it is marked as an abnormal region. The environmental control command generation module, when the duration of the abnormal area exceeds the limit, separates the real-time airflow data of the abnormal area into low-frequency steady-state base flow components and high-frequency chaotic disturbance components through variational mode decomposition. It then constructs a multi-objective optimization model with the goal of minimizing energy consumption and temperature fluctuation, and uses an improved particle swarm optimization algorithm to obtain the Pareto optimal solution of the deflection angle of the guide vane and the fan speed, thereby generating environmental control commands that balance energy consumption and heat dissipation efficiency.

2. The adaptive thermal management system for energy storage containers according to claim 1, characterized in that, The multi-source data fusion method is as follows: real-time airflow data, temperature distribution data and CFD simulation data are fused using a Gaussian process regression algorithm to construct a joint probability model; The joint probability model is used to complement and fuse CFD simulation data with real-time airflow data and temperature distribution data to generate a fused flow field map of the energy storage container. The internal space of the energy storage container is divided into several grid cells, and the flow field diagram includes the airflow velocity, airflow direction angle and predicted temperature of each grid cell. Define the kernel function of the joint probability model and optimize the hyperparameters. The optimization method is as follows: set the optimization objective to maximize the likelihood probability of the joint probability model output fused flow field map and measured data; use the conjugate gradient method to iteratively adjust the kernel function parameters until the joint probability model satisfies the optimization objective.

3. The adaptive thermal management system for energy storage containers according to claim 2, characterized in that, When the chaos intensity level of a grid cell exceeds a preset value and the predicted temperature exceeds a dynamic temperature threshold, it is marked as an abnormal grid. Morphological closing operations are performed on abnormal meshes that meet the triggering conditions to generate a set of coordinates of connected regions, providing spatial positioning input for the targeted control of the environmental control command generation module.

4. The adaptive thermal management system for energy storage containers according to claim 1, characterized in that, The process of generating environmental control instructions includes the following steps: When the duration of the abnormal region exceeds the limit, variational mode decomposition technology is used to separate the low-frequency steady-state fundamental current component and the high-frequency steady-state fundamental current component of the airflow in the abnormal region. The low-frequency steady-state fundamental current component represents the average motion of the airflow that maintains heat dissipation; the high-frequency steady-state fundamental current component represents the eddy current pulsation and noise between the battery clusters. A multi-objective optimization model is constructed with the objectives of minimizing energy consumption and temperature fluctuation. The safety constraints of the multi-objective optimization model include the safe range of airflow velocity and the upper limit of fan speed. An improved particle swarm optimization algorithm was used to find the Pareto optimal solution for the deflection angle of the guide vane and the speed of the fan. Environmental control commands were constructed based on the Pareto optimal solution. A multi-objective optimization model was used to balance energy consumption and heat dissipation efficiency and to suppress the risk of thermal runaway caused by chaotic disturbances.

5. The adaptive thermal management system for energy storage containers according to claim 1, characterized in that, Also includes: The reliability verification module collects the airflow field residual in real time after executing the environmental control command. If the airflow field residual exceeds the preset threshold, the closed-loop verification process is triggered. The meta-reinforcement learning optimization module builds a digital twin platform for energy storage containers. Multi-condition disturbance data is randomly generated in the digital twin platform, and the meta-reinforcement learning model is trained to quickly adapt to multi-condition disturbance scenarios while meeting safety constraints. The model searches for environmental control commands within the safety constraint boundaries and outputs a safety-verified meta-reinforcement learning model for subsequent use.

6. The adaptive thermal management system for energy storage containers according to claim 5, characterized in that, When the closed-loop verification process is triggered, it includes: The residual stability coefficient is calculated based on the ratio between the current airflow field residual and the historical residual mean. The historical residual baseline is calculated based on the average value of historical residual data within the sliding time window. The residual stability coefficient is dynamically generated by the ratio of the current residual to the historical residual baseline. The residual stability coefficient is used to quantify the transient degree of the energy storage container environment deviating from the steady state. Based on the fluctuation range of the residual stability coefficient, the confidence interval threshold and the policy search direction of the meta-reinforcement learning model are dynamically adjusted. When the residual stability coefficient exceeds the preset threshold, the confidence interval threshold is tightened and the weight of the meta-reinforcement learning model for exploring new strategies is increased. When the residual stability coefficient is lower than the preset threshold, the confidence interval threshold is relaxed and the validated strategy is called first.

7. The adaptive thermal management system for energy storage containers according to claim 1 or 6, characterized in that, The system also includes: The condensation risk disturbance module is used to identify temperature and humidity data of potential condensation risk points inside the energy storage container. When the condensation risk score exceeds the preset value, the sensor data is input into the condensation risk assessment model. Based on the boundary reference conditions generated by CFD simulation, combined with the local dew point proximity rate, airflow disturbance attenuation coefficient and temperature difference change rate, the module outputs the condensation risk score for each condensation risk point and generates a high-risk area identifier. It also generates condensation disturbance instructions for the high-risk area to enhance airflow and suppress condensation formation.

8. The adaptive thermal management system for energy storage containers according to claim 7, characterized in that, The condensation risk disturbance module includes a temperature and humidity sensing unit, a condensation risk assessment model, a local behavior baseline construction unit, and an anomaly detection model. The temperature and humidity sensing unit is used to collect temperature and humidity data in different spatial locations inside the energy storage container in real time. The condensation risk assessment model refers to a function model that takes local temperature and humidity data as input, combines simulation boundary conditions, and outputs a condensation risk score to dynamically assess the possibility of condensation forming in a local space. The local behavior baseline construction unit is used to construct the local environmental behavior baseline of each monitoring area inside the energy storage container under normal operating conditions, and to serve as a comparison reference for risk scoring. The anomaly detection model is used to identify the degree of baseline deviation, identify condensation triggering patterns, and assist in determining whether the condensation risk score increases rapidly due to non-periodic fluctuations.

9. The adaptive thermal management system for energy storage containers according to claim 8, characterized in that, The operation process of the condensation risk assessment model includes: The local dew point proximity rate, airflow disturbance attenuation coefficient, and temperature difference change rate are obtained for each risk point; these are then mapped to a unified scoring scale using normalization functions. The local dew point proximity rate refers to the degree of proximity between the real-time air temperature and the current dew point temperature of a risk point, used to measure the trend of water vapor in the air approaching saturation. The airflow disturbance attenuation coefficient refers to the attenuation magnitude of the local airflow disturbance intensity per unit time, used to describe the natural weakening of the local air disturbance intensity in spatial dead zones. The temperature difference change rate refers to the rate of change of the difference between the local air temperature and the temperature of adjacent areas per unit time, used to capture the trend of abrupt changes in microenvironmental temperature. The relative humidity is combined with the local air temperature, and the local absolute humidity is calculated using the ideal gas law. Using weighted parameters derived from experience or training, the standardized local dew point proximity rate, airflow disturbance attenuation coefficient, temperature difference change rate, and local absolute humidity are weighted and summed to obtain the condensation risk score for each risk point.

10. The adaptive thermal management system for energy storage containers according to claim 9, characterized in that, The execution of the condensation disturbance command is limited by the power consumption limit and the main airflow structure. The disturbance operation needs to enhance the airflow disturbance locally while dynamically compensating through the disturbance amplitude adjustment mechanism to control the overall main airflow field offset within an acceptable range.

Images