Intelligent distributed liquid cooling energy storage system thermal management method

Abstract

The application relates to the technical field of energy storage system thermal management, and discloses an intelligent distributed liquid-cooled energy storage system thermal management method. The method acquires a plurality of battery monomer operating state parameters and optimizes timestamp synchronization; temperature data is processed by small-batch time series decomposition, and change trend characteristics and abnormal hotspot positions are extracted. Based on the above information, the non-dominated sorting strategy is used to optimize the heat exchange control rule library parameter set, and the optimized control parameters and priority rules are generated. In combination with the liquid cooling system inlet temperature and flow parameter, the tabu search mechanism is used to determine the cooling liquid flow distribution scheme, and branch control instructions are output. Finally, the approximate gradient projection technology is used to dynamically adjust the flow constraint boundary, the branch and bound strategy is used to discretize the optimization distribution path, the branch flow fine regulation and control are realized, and the energy storage system thermal management precision and reliability are improved. The method improves the accuracy, adaptability and stability of the liquid-cooled energy storage system thermal management.

Description

Technical Field

[0001] This invention relates to the field of thermal management technology for energy storage systems, specifically a thermal management method for an intelligent distributed liquid-cooled energy storage system. Background Technology

[0002] With the rapid development of the new energy industry, energy storage systems, as core equipment for energy storage and dispatch, are constantly expanding in scale and application scenarios. Among them, battery energy storage systems are widely used due to their high energy density and fast response speed. However, batteries generate a large amount of heat during charging and discharging. If this heat cannot be dissipated in time, it will lead to increased battery temperature and uneven temperature distribution, thereby affecting battery performance, cycle life, and even causing safety risks. Therefore, efficient thermal management technology is crucial to ensuring the stable operation of energy storage systems.

[0003] Liquid cooling technology has been widely used in medium and large-scale energy storage systems due to its advantages such as high heat dissipation efficiency and good temperature control uniformity. However, existing thermal management methods for liquid-cooled energy storage systems still have many shortcomings. In terms of parameter acquisition, due to the large number of battery cells in the energy storage cluster, the time synchronization of each acquisition node is poor, resulting in deviations in the acquired operating status parameters over time, which affects the accuracy of subsequent temperature analysis and control decisions.

[0004] In temperature data analysis and processing, traditional methods often employ simple statistical analysis or static models, which struggle to capture dynamic temperature trends in real time. Furthermore, they suffer from lag and insufficient accuracy in identifying localized hotspots. This prevents the system from promptly detecting potential thermal runaway risks, reducing the foresight of thermal management.

[0005] In terms of optimizing heat exchange control rules, existing technologies often rely on empirically set control parameters and lack dynamic optimization mechanisms. As the operating conditions of energy storage systems change, such as fluctuations in charging and discharging power and changes in ambient temperature, fixed control rules are difficult to adapt to the heat dissipation requirements under different scenarios, which can easily lead to insufficient or excessive heat dissipation, resulting in energy waste or poor heat dissipation effect.

[0006] In the coolant flow distribution stage, traditional methods often employ simple proportional distribution or control strategies based on a single index, failing to comprehensively consider real-time inlet temperature, flow parameters, and the temperature status of each battery cell, resulting in inaccurate flow distribution. Furthermore, the lack of effective dynamic control methods for adjusting flow constraint boundaries and optimizing pathways makes it difficult to achieve refined control of the flow in each branch, affecting the uniformity and stability of the overall heat dissipation effect.

[0007] In distributed energy storage systems, the operating states of each energy storage unit cluster differ. Traditional centralized thermal management control models struggle to address the individualized heat dissipation needs of each unit, resulting in low overall system thermal management efficiency. These issues hinder further improvements in the thermal management performance of liquid-cooled energy storage systems, failing to meet the demands of large-scale energy storage systems for high reliability and high-efficiency thermal management. Summary of the Invention

[0008] The purpose of this invention is to provide a thermal management method for an intelligent distributed liquid-cooled energy storage system to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides a thermal management method for an intelligent distributed liquid-cooled energy storage system, the method comprising:

[0010] The system acquires the operating status parameters of multiple battery cells in the energy storage unit cluster and optimizes the data acquisition process by timestamp synchronization.

[0011] The operating status parameters are analyzed in real time, and the temperature data is processed using a small-batch time series decomposition method to obtain temperature change trend characteristics and abnormal hot spot location information.

[0012] Based on the temperature change trend characteristics and abnormal hot spot location information, the parameter set of the heat exchange control rule base is optimized by a non-dominated sorting strategy to generate optimized heat exchange control parameters and priority rules.

[0013] Based on the real-time collected inlet temperature and flow parameters of the liquid cooling system, as well as the optimized heat exchange control parameters and priority rules, a tabu search mechanism is used to calculate the decision of the coolant flow distribution scheme and output the flow control command for each branch.

[0014] In response to the flow control commands of each branch, the flow constraint boundary is dynamically adjusted using approximate gradient projection technology, and the branch delimitation strategy is integrated to discretize and optimize the coolant distribution path, thereby achieving fine-grained control of branch flow.

[0015] Preferably, the operating status parameters of multiple battery cells in the energy storage unit cluster are acquired, and the acquisition process is optimized for timestamp synchronization, including:

[0016] The operating status parameters include temperature distribution, voltage fluctuation, and internal resistance change.

[0017] Based on the rate of change of historical temperature data and system load conditions, the sampling frequency and sampling period of battery cells in different areas are dynamically set.

[0018] Based on the set acquisition frequency and sampling period, the temperature distribution, voltage fluctuation and internal resistance change data of each battery cell are collected synchronously to generate a time-aligned battery cluster state vector.

[0019] Preferably, the operating status parameters are analyzed in real time, and the temperature data is processed using a small-batch time series decomposition method, including:

[0020] The real-time acquired battery cluster temperature data stream is divided into continuous time-segment batches;

[0021] Initialize the window size and iteration step size of the temperature decomposition model;

[0022] In each time segment batch, the periodic component, trend component, and residual component of the temperature data are extracted, and the energy proportion of each component is calculated.

[0023] Based on the spatial distribution density of the residual components, the coordinates of abnormal hot spots exceeding the set threshold are identified, and hot spot risk level markers are generated by combining voltage fluctuation data.

[0024] The output includes analysis results containing temperature change trend characteristics and abnormal hotspot location information marked with risk levels.

[0025] Preferably, based on the temperature change trend characteristics and abnormal hotspot location information, the parameter set of the heat exchange control rule base is optimized using a non-dominated sorting strategy, including:

[0026] Define the objective function for thermal management, including temperature uniformity index, energy consumption index, and hot spot suppression response time;

[0027] Initialize the candidate parameter population for the heat exchange control rule base;

[0028] The parameter combinations in the population are applied to the heat exchange control rule base, and the Pareto front fitness of the parameter combinations is evaluated based on the temperature change trend characteristics and abnormal hot spot location information.

[0029] Perform tournament selection, simulated binary crossover, and polynomial mutation operations on the parameter population to update the candidate solution set;

[0030] The next generation of populations is selected based on the non-dominated sorting level and crowding distance until the convergence condition is met, and the optimized heat exchange control parameters and priority rules are output.

[0031] Preferably, based on the real-time collected inlet temperature and flow parameters of the liquid cooling system, as well as the optimized heat exchange control parameters and priority rules, a tabu search mechanism is used to calculate the coolant flow distribution scheme, including:

[0032] Construct a thermal resistance network model with branch flow as the decision variable;

[0033] Initialize the tabu list length and neighborhood search step size;

[0034] An initial flow allocation scheme is generated based on the optimized heat exchange control parameters and priority rules.

[0035] In each iteration, a neighborhood solution set is generated based on the tabu list state, and the heat balance deviation and pump power consumption corresponding to each solution are calculated.

[0036] The pardon criterion is used to accept taboo solutions that are better than the historical best, and the current optimal traffic allocation scheme is updated.

[0037] When the maximum number of iterations is met, output the flow control instructions for each branch.

[0038] Preferably, in response to the flow control commands of each branch, the flow constraint boundary is dynamically adjusted using approximate gradient projection technology, and a branch-bounding strategy is integrated to discretize and optimize the coolant distribution path, including:

[0039] Analyze the target flow values ​​in the flow control commands for each branch;

[0040] Dynamic flow constraint boundaries are set based on pipeline topology characteristics and pump performance curves;

[0041] Calculate the direction of the absolute error gradient between the current traffic allocation and the target traffic;

[0042] Projecting along the gradient descent direction onto the feasible region boundary yields a continuous flow solution that satisfies the constraints.

[0043] The continuous flow is discretized into candidate valve opening combinations, and the optimal valve opening command is determined by pruning the invalid search space through a branch and bound strategy.

[0044] Preferably, the flow constraint boundary is dynamically adjusted using an approximate gradient projection technique, including:

[0045] Set the relaxation factor and convergence threshold for the flow constraint boundary;

[0046] Calculate the heat transfer efficiency loss of each branch under the current flow distribution scheme;

[0047] Update the relaxation coefficient of the flow constraint boundary according to the direction of heat conduction efficiency loss;

[0048] The updated constraint boundary is applied to the gradient projection calculation to generate a flow adjustment scheme that satisfies the new boundary.

[0049] Preferably, the method further includes:

[0050] The temperature feedback data of the battery cells after regulation is collected by the adaptive Kalman estimation algorithm. The processed temperature feedback data is then transmitted to the heat exchange control rule base, time series decomposition module and optimization module to update the thermal management strategy.

[0051] The method of acquiring temperature feedback data of individual battery cells after regulation using an adaptive Kalman estimation algorithm includes:

[0052] Collect multi-point temperature monitoring data of individual battery cells within a set time window;

[0053] Construct a battery thermal state transition model that considers the coolant time delay effect;

[0054] The Kalman gain coefficient is adaptively adjusted based on the noise statistics of the temperature monitoring data.

[0055] Outputs an estimated battery body temperature after removing measurement noise and time delay interference.

[0056] Preferably, the processed temperature feedback data is transmitted to the heat exchange control rule base, the time series decomposition module, and the optimization module, including:

[0057] Extract the spatial gradient distribution and temporal differential characteristics of the temperature estimate;

[0058] Input the spatial gradient distribution characteristics into the heat exchange control rule base and update the hotspot suppression rule weights;

[0059] Input the time differential features into the time series decomposition module and adjust the window parameters for trend component extraction;

[0060] Inputting the complete temperature feedback dataset into the optimization module triggers a re-optimization process of the parameter set in the heat exchange control rule base.

[0061] Preferably, the method further includes:

[0062] Based on the cumulative distribution characteristics of temperature feedback data, a multidimensional thermodynamic feature vector is generated;

[0063] Extracting the core pattern components of feature vectors using tensor decomposition techniques;

[0064] Map the core mode components to the system aging state space and update the battery health state compensation coefficient.

[0065] The updated battery health state compensation coefficients are fed back into the residual analysis step in the time series decomposition method.

[0066] Compared with the prior art, the beneficial effects of the present invention are:

[0067] This intelligent distributed liquid-cooled energy storage system thermal management method effectively improves the accuracy and adaptability of thermal management through multi-stage collaborative optimization. During the operational status parameter acquisition phase, timestamp synchronization optimization is performed on the acquisition process to ensure the consistency of operational status parameters of different battery cells across the time dimension. This avoids data analysis errors caused by time deviations and provides a reliable data foundation for subsequent temperature analysis and control decisions. This synchronization optimization mechanism enables the system to more accurately reflect the real-time operational status of each battery cell, laying a data-level guarantee for precise thermal management.

[0068] In temperature data analysis and processing, a small-batch time series decomposition method is employed to extract temperature change trend characteristics and abnormal hotspot location information in real time. Compared to traditional static analysis methods, this method can dynamically capture subtle temperature changes over time, promptly identify local abnormal hotspots, and allow the system to detect potential thermal risks earlier. By accurately grasping temperature change trends, the thermal management system can adjust its heat dissipation strategy in advance, preventing the continued expansion of abnormal hotspots and ensuring that the battery operates within a safe temperature range.

[0069] To optimize the heat exchange control rule base, a non-dominated sorting strategy is introduced to optimize the parameter set. The resulting optimized heat exchange control parameters and priority rules are better able to adapt to the complex and variable operating conditions of energy storage systems. The non-dominated sorting strategy considers multiple optimization objectives, enabling the control rules to meet heat dissipation requirements while also taking into account factors such as energy consumption. This avoids the problems of insufficient or excessive heat dissipation that occur with traditional fixed rules when operating conditions change, thus improving the flexibility and rationality of thermal management control.

[0070] In the coolant flow allocation decision-making stage, a tabu search mechanism is used in conjunction with real-time collected data on the liquid cooling system inlet temperature, flow parameters, and optimized control rules to calculate the flow allocation scheme. The tabu search mechanism possesses powerful global optimization capabilities, quickly finding a better solution among numerous possible flow allocation schemes, ensuring that the flow allocation of each branch matches the current heat dissipation requirements of each battery cell. This dynamic flow allocation decision allows heat dissipation resources to be allocated on demand, improving heat dissipation efficiency.

[0071] When responding to flow control commands, the approximate gradient projection technique is used to dynamically adjust the flow constraint boundary, and a branch-bound strategy is integrated to discretize and optimize the coolant distribution path, achieving fine-grained control of branch flow. The approximate gradient projection technique can flexibly adjust the constraint boundary according to real-time conditions, ensuring the feasibility and accuracy of flow control; the branch-bound strategy can efficiently optimize in the discrete path space, ensuring the rationality of the distribution path. The combination of these two techniques allows the flow of each branch to accurately match the corresponding battery cell's heat dissipation requirements, effectively improving the problem of uneven temperature distribution, enhancing the uniformity and stability of the entire energy storage system's thermal management, helping to extend battery life, and ensuring the safe and reliable operation of the energy storage system. Attached Figure Description

[0072] Figure 1 This is a schematic diagram illustrating the working principle of the thermal management method for the intelligent distributed liquid-cooled energy storage system described in this invention.

[0073] Figure 2 A flowchart for collecting and synchronizing operational status parameters;

[0074] Figure 3 Flowchart for optimizing parameters of the heat exchange control rule base;

[0075] Figure 4 A flowchart for dynamic adjustment and discretization optimization of branch flow. Detailed Implementation

[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] Please see Figure 1 This invention provides a thermal management method for an intelligent distributed liquid-cooled energy storage system, the method comprising:

[0078] The system achieves precise temperature control of the battery cluster through multi-module collaboration. During system initialization, the communication protocol between the data acquisition module and the liquid cooling control module is configured, and an initial parameter set for the heat exchange control rule base is established. During operation, temperature, voltage, and internal resistance data of each battery cell within the energy storage unit cluster are synchronously collected and, after timestamp alignment, a battery cluster state vector is generated. The temperature data stream is divided into time-segment batches using a sliding window, and a seasonal decomposition algorithm is used to extract trend and residual components. Abnormal hotspot coordinates are marked using voltage fluctuation data. Parameter optimization of the heat exchange control rule base employs a multi-objective genetic algorithm, generating a Pareto optimal solution set based on temperature uniformity, energy consumption, and response time indices. The coolant flow distribution scheme is solved in the thermal resistance network model using a tabu search mechanism, and the valve opening is discretized and optimized using gradient projection and branch-and-bound strategies. The system receives temperature feedback data in real time, eliminates measurement noise through Kalman filtering, and updates the parameters of each module, forming a closed-loop control.

[0079] Example 1: See Figure 2 The data acquisition system adopts a distributed architecture, with the main controller connecting to the monitoring units of each battery module via Ethernet. Each monitoring unit is equipped with a high-precision temperature sensor, voltage sampling circuit, and AC impedance measurement module. The temperature detection point is located at the center of the battery terminal and sidewall. The voltage sampling frequency is 20kHz, and the internal resistance measurement uses the 1kHz AC injection method. The acquisition strategy is dynamically configured based on the battery's historical temperature rise, dividing the battery cluster into an 8×8 grid area. Each grid area records the maximum temperature rise rate of the last 10 samples. Areas with a temperature rise rate exceeding 2℃ / min are marked as red high-activity areas, with a sampling period of 500ms; areas with a temperature rise rate between 0.5-2℃ / min are marked as yellow medium-activity areas, with a sampling period of 1s; and blue low-activity areas with a temperature rise rate below 0.5℃ / min have a sampling period extended to 2s. The main controller broadcasts a time synchronization signal via the PTP precision clock protocol, and each monitoring unit performs local sampling upon receiving the synchronization pulse. The collected data is appended with millisecond-level timestamps. A temperature-voltage-internal resistance state matrix is ​​constructed using the battery cluster ID, timestamp, and three-dimensional coordinates (module number / row number / column number) as indexes, with time alignment accuracy controlled within ±0.8ms.

[0080] Temperature data processing employs an improved STL decomposition process. The data stream is segmented by time windows, with the window length initialized to 15 sampling points. After the temperature sequence is input into the decomposition module, the first iteration uses locally weighted regression to fit the trend component. A quadratic polynomial regression model is used, with adjacent point weights decaying according to a Gaussian distribution, and the bandwidth parameter is set to 40% of the window length. The fitted residual sequence is input into the periodic component extraction module, where the main frequency components are identified through Fast Fourier Transform, and frequency components with an energy share exceeding 5% are retained to reconstruct the periodic curve. In the second iteration, the periodic component is subtracted from the original data to obtain the deperiodic sequence, and weighted regression is performed again to fit the new trend component. After three iterations, the final trend component, periodic component, and residual component are output. The energy share of each component is calculated as follows: the trend component is the integral of the square of the linear fitting slope, the periodic component is the sum of the squares of the Fourier coefficient moduli, and the residual component is the square of the standard deviation. When the residual component exceeds 3.2 times the standard deviation of the historical mean, the battery cell is marked as an anomaly candidate.

[0081] Spatial localization of anomalous hotspots is based on the physical coordinate layout of the battery clusters. Each battery cell is treated as a node on a two-dimensional plane, and a spatial mesh is constructed using Delaunay triangulation. The residual values ​​are used as height values ​​input to the interpolation algorithm to generate a continuous two-dimensional temperature anomaly surface. Surface gradient analysis identifies local extrema; areas within a 30cm radius of these extrema with a residual increase exceeding 15% are identified as anomalous hotspots. Voltage fluctuation data are used to calculate the standard deviation over a time window; cells with a standard deviation greater than 5mV are marked as high-risk. Combining spatial localization results with voltage fluctuation values, three risk levels are defined: Level 1 (residual > 3σ and voltage fluctuation < 3mV), Level 2 (residual > 4σ or voltage fluctuation > 5mV), and Level 3 (residual > 4σ and voltage fluctuation > 5mV). The final output is a hotspot distribution map with coordinate locations and risk levels, and the first derivative of the trend component is generated as a temperature change rate feature.

[0082] The time window parameters are dynamically adjusted based on data characteristics. After processing five time windows, the average rate of change of the trend component and the coefficient of variation of the residual component are calculated. If the rate of change increases by more than 20%, the window length is shortened by two sampling points; if the coefficient of variation decreases by more than 15%, the window length is increased by one sampling point. The structured output of the analysis results includes four dimensions: temperature trend feature set (rate of change, mean acceleration), periodic feature vector (dominant frequency amplitude, phase shift), hotspot location list (3D coordinates, risk level), and residual distribution matrix (gridized residual values). This structured data is passed to the subsequent optimization module via shared memory.

[0083] The 3D coordinate mapping employs a spatial indexing mechanism. The physical layout of the battery cluster is modeled using a Cartesian coordinate system, with the origin placed at the geometric center of the cluster. The X-axis has one unit every 10cm along the battery row direction, the Y-axis has one unit every 8cm along the battery column direction, and the Z-axis corresponds to the support layer number. Each battery cell is assigned a unique spatial identifier, composed of X coordinate / 10, Y coordinate / 8, and the layer number. Temperature data interpolation in the spatial grid uses the inverse distance weighting method, with a fixed number of four nearest neighbors, and the weight is inversely proportional to the square of the distance. After the residual surface is generated, the gradient field is calculated using the Sobel operator; regions with gradient magnitudes exceeding a set threshold are identified as anomalous boundaries. Within these boundaries, a flooding-fill algorithm marks connected regions. The centroid coordinates of a single hotspot are calculated as the weighted average of the coordinates of all cells within the region, with the weight proportional to the residual magnitude.

[0084] The correlation analysis between voltage fluctuations and thermal risk is performed on a separate coprocessor. Voltage data streams are filtered by a Butterworth low-pass filter, and the standard deviation sequence within a 10-second window is calculated. When a sudden change in standard deviation exceeds 50% of the historical mean, a time alignment check with the temperature residual is triggered, with the alignment time window set to ±200ms. The risk level criterion is a fuzzy logic with a dead zone: when the temperature residual is within 3.0-3.5 times the standard deviation and the voltage fluctuation is within 3-4mV, a risk level of 1.5 is output; when the residual is within 3.5-4.0 times the standard deviation and the voltage fluctuation is within 4-5mV, a risk level of 2.5 is output. Linear interpolation is used for the critical values ​​to avoid abrupt changes in risk level. The final generated hotspot report table contains sixteen attribute fields: timestamp, spatial identifier, risk level, temperature residual, voltage standard deviation, temperature change rate, and distance to adjacent hotspots, which are retrieved by the rule base optimization module.

[0085] Example 2: See Figure 3 During system initialization, a parameter search space is established in a dedicated optimization coprocessor equipped with an independent memory pool and parallel computing units. The parameter solution vector is set to 16 dimensions, containing three types of parameters: proportional-integral-derivative control coefficients (6 dimensions), branch flow allocation weights (8 dimensions), and anomaly response priority coefficients (2 dimensions). The parameter boundaries are preset according to physical constraints, with proportional coefficients ranging from 0.1 to 5.0, integral time constants from 0.01 to 10 seconds, and derivative coefficients from 0 to 2.0; the flow weight range is [0,1] and the sum of the weights of each branch is 1; the priority coefficients are integers from 1 to 5. The initial population generation uses Latin hypercube sampling, uniformly sampling 100 candidate solution vectors within a 100-dimensional hypercube.

[0086] Each candidate solution vector is converted into an executable control rule by a decoder. Proportional-integral-derivative (PID) parameters are input to the temperature closed-loop control module, flow weights are imported into the flow allocation algorithm, and priority coefficients are written to the thermal event response queue. The evaluation process loads the temperature change trend characteristics and abnormal hotspot location information output from Example 1, and performs a 300-second virtual run in the battery thermal model simulation environment. The model uses a three-dimensional non-uniform thermal network, containing 512 battery nodes and 24 cooling branches with thermal capacity and thermal resistance parameters. The solution step size for the heat conduction equation is 50 milliseconds. The objective function calculation outputs three scalars based on the simulation results: the temperature uniformity index is the ratio of the maximum temperature difference to the average temperature difference of the battery cluster during the stable phase; the energy consumption index is the integral value of the pump power over time, with power converted according to the cubic relationship of flow rate; the hotspot suppression response time starts from the abnormal marker and ends when the hotspot temperature drops below the threshold, with the longest response time taken for multiple hotspots.

[0087] The genetic optimization engine employs a modified NSGA-II architecture. The selection operation implements a binary tournament, randomly selecting two individuals each time, prioritizing those with lower Pareto rank, and choosing those with larger crowding distances when ranks are the same. The crossover operation uses a simulated binary crossover operator, operating independently on each dimension of the selected parent solution vector. The crossover probability is set to 0.92, and the distribution exponent is adjusted to 30. When generating offspring solutions, samples are taken from the vicinity of the parent solutions according to the probability distribution. The mutation phase implements multinomial mutation, with a mutation probability set to 0.15 and a mutation distribution exponent of 35, perturbing the random dimensions of the solution vector. The newly generated offspring individuals are merged with the original population to form 200 solutions, which are then ranked using a fast non-dominated sort. During the sorting process, the dominance count for each solution is calculated. Non-dominated solutions are assigned to the first frontier layer. Solutions in this layer are removed, and the dominance count for the remaining solutions is recalculated to form the second layer, until all solutions are ranked.

[0088] Crowding distance calculation is performed separately for each Pareto front layer. For individuals within the same layer, they are sorted according to the values ​​of the three objective functions, and the Euclidean distances of adjacent solutions in each objective space are calculated. The crowding distance of each individual is the sum of its adjacent distances in each objective dimension. When selecting the next generation of the population, individuals from lower-level front layers are preferentially retained. When a front layer cannot accommodate all individuals, they are selected from largest to smallest crowding distance. The elite retention strategy is set to retain the 10 historically best solutions in each generation and directly add them to the next generation of the population. During the iteration process, the rate of change of the Pareto front is monitored, and convergence is determined when the front position shift is less than 0.5% for five consecutive generations.

[0089] The parameter weighting module is activated after each generation of evolution. It records the extreme values ​​of each objective function in the current population. The weight for temperature uniformity is set to the reciprocal of the current maximum value, the weight for energy consumption is the reciprocal of the minimum value, and the weight for response time is fixed at 1. A weighted sum of three terms for each candidate solution is calculated as a supplementary evaluation metric. When the number of Pareto front solutions exceeds 20, the five solutions with the smallest weighted sum are additionally selected and added to the elite set. A dynamic parameter adjustment mechanism monitors the optimization process. If no new Pareto solutions are generated for three consecutive generations, the crossover probability is increased by 0.05; if population diversity is below a threshold (mean crowding distance decreases by 30%), the mutation probability is increased by 0.1.

[0090] After optimization, the Pareto front solution set is output and presented in a three-dimensional target space distribution through an interactive decision-making interface. After the operator selects a compromise solution, the system automatically parses the parameter vector: the proportional-integral-derivative (PID) parameter set is written to the temperature controller register; the flow weight coefficients are updated to the weight matrix of the allocation algorithm; and the priority coefficients are converted into the sorting rules of the event response queue. Before the parameters take effect, an integrity check is performed to verify that the flow weight sum is 1 and that the PID parameters are within the hardware's allowable range. The final generated optimization parameter set is marked with a timestamp and optimization batch number and stored in an encrypted format in the rule base version management system.

[0091] The parameter update engine is activated during system idle periods, backing up the current rule base before loading new parameters. The update process consists of three phases: the first phase updates only priority coefficients and observes the response time to thermal events within 12 hours; the second phase injects flow allocation weights and monitors the matching degree of flow sensors in each branch; the third phase fully enables proportional-integral-derivative parameters and evaluates the temperature control accuracy within 72 hours. Each phase must be passed before proceeding to the next, and an exception rollback mechanism automatically restores the previous version when critical parameter deviations exceed limits. Throughout the optimization cycle, performance baseline data is recorded, including computation time, memory usage, and convergence algebra, for subsequent hyperparameter tuning of the optimization algorithm.

[0092] Example 3: During the system initialization phase, a thermal resistance network model is constructed, abstracting the battery clusters as a directed weighted graph structure. Nodes in the graph represent battery modules, with N nodes corresponding to the actual physical layout; edges represent coolant flow paths, including M main branches and K bypass branches. The edge weights are defined as the thermal resistance coefficients.

[0093] The calculation method is as follows:

[0094]

[0095] in: Indicates the length of the branch pipe. The thermal conductivity of the coolant, It is the cross-sectional area of ​​the branch road. This is the convective heat transfer coefficient. The model is stored in the graph database Neo4j. Node attributes include three-dimensional coordinates, rated heat capacity, and current temperature, while edge attributes record real-time flow rate, thermal resistance, and valve opening status.

[0096] The tabu search algorithm is executed in a dedicated computing unit equipped with a cache to store the current solution state. Initial solution generation employs a uniform distribution strategy, with traffic allocated to each branch. Set to 1 / M of the total flow. The tabu list is designed as a circular buffer structure, recording the branch numbers and flow adjustment directions of the most recent H moves. Neighborhood solution generation uses a three-branch perturbation strategy: three branches are randomly selected, and the flow of the first branch is increased. Article 2 reduces Article 3 reduces Maintain total flow conservation. Perturbation step size. The algorithm is dynamically adjusted, with an initial value of 5% of the design flow rate. If no improvement is found after three consecutive iterations, the step size is reduced to 80% of the original value. Objective function. Calculations include thermal balance deviation Pump power consumption Two items:

[0097]

[0098] in: and The weighting coefficients were obtained from the optimization results of Example 2. Let the temperature be the temperature of the j-th node. It is the average temperature of the cluster; It refers to the density of the coolant. It is the acceleration due to gravity. Let represent the head of the i-th branch. When calculating the heat balance term, the node temperature distribution is first solved using a thermal resistance network model, and the Jacobi iteration method is used until the residual is less than 0.1℃. : indicates that the coolant is at the first The volumetric flow rate in each branch is calculated by interpolation based on the pipeline characteristic curve, and the characteristic curve data is stored in a two-dimensional lookup table.

[0099] The pardon criterion employs a two-level judgment: If the objective function value of the neighborhood solution is better than the historical best solution, the taboo state is immediately lifted; if the quality of the solution is better than the current solution and satisfies the condition... The taboo is lifted after a 5-generation delay. An elite retention strategy is implemented during the search process, storing the top 3 optimal solutions from each generation in an external archive. These archived solutions are used for a restart mechanism; when the search gets stuck in a local optimum (no improvement for 10 consecutive generations), a solution is randomly selected from the archive as a new starting point to continue the search.

[0100] During the flow control command generation phase, continuous flow values ​​are converted into discrete valve openings. The valve characteristic curve uses a piecewise linear approximation, with a control point set every 5% of opening. The final command consists of three parts: a list of target flow rates for the main branch, a bitmap of bypass opening / closing status, and an execution time window. The command is sent to each actuator via the CAN bus. The actuators provide feedback on the deviation between the actual opening and the target value; if the deviation exceeds 5%, a command retransmission mechanism is triggered.

[0101] The dynamic parameter adjustment module monitors the search process status. The tabu list length H is initially set to 10 and automatically adjusted based on search efficiency: when the acceptance rate is below 20%, H decreases by 2; when the number of repeated solution visits exceeds 5, H increases by 3. Weighting coefficients... and Recalibration is performed every 20 iterations, based on the proportional distribution of the current Pareto front solutions. A cooling mechanism is activated in the later stages of the search; when the number of iterations exceeds 70% of the maximum, the perturbation step size is adjusted. It decays exponentially, with the decay factor set to 0.95.

[0102] The exception handling mechanism includes three scenarios: when flow is infeasible, the gradient projection algorithm is invoked to pull the solution back to the feasible region; when the thermal resistance model fails to converge, the simplified lumped parameter model is switched to perform calculations; and in the event of hardware feedback timeout, the valid instructions from the previous cycle are used. All exception events are recorded in the system log for subsequent algorithm improvement analysis. Key parameters throughout the search process, including the tabu list status, objective function value, and neighborhood acceptance rate, are pushed to the monitoring interface in real time via WebSocket, supporting manual intervention to adjust the search direction.

[0103] Convergence is determined based on a combination of conditions: a maximum of 100 iterations, or an improvement of less than 0.1% in the optimal solution over 15 consecutive iterations, or a computation time exceeding 300ms. The final output flow allocation scheme must pass a verification process, including thermal resistance network simulation verification, branch flow conservation checks, and valve actuator capability verification. Verified schemes are marked with timestamps and version numbers and stored in a scheme database for historical querying and rollback. The scheme database employs a differential storage strategy, recording only changes relative to the previous version, saving storage space and improving retrieval efficiency.

[0104] Example 4: See Figure 4 During system initialization, pipeline topology parameters are loaded, including physical characteristic data for 24 main branches and 8 bypasses, stored in a structure array. Each branch attribute includes design flow rate, maximum allowable pressure drop, pipe diameter, equivalent length, and its corresponding zone number. The flow constraint boundary is initially set to 70%-130% of the design value, forming the initial feasible region. The dynamic adjustment module updates the boundary based on real-time monitoring data, with the data sampling period synchronized with the control system clock, refreshing the constraints every 5 seconds.

[0105] The gradient projection calculation employs an iterative approximation method. After parsing the current flow allocation scheme, a 24-dimensional vector is generated, and the target flow command is converted into a reference vector of the same dimension. Error calculation is performed on a branch-by-branch basis, taking the percentage of absolute error to the design flow. The projection direction is determined based on the error gradient, and the gradient field is calculated using the finite difference method. A small perturbation of ±0.5% in the flow of each branch is observed to monitor the change in the objective function value. An adaptive strategy is used for the projection step size, with an initial step size of 2% of the design flow, dynamically adjusted based on convergence. When the error reduction rate is less than 5% for three consecutive iterations, the step size is reduced to 50% of its original value.

[0106] The discretization process maps continuous flow rates to combinations of valve openings. Valve characteristic data is stored in a lookup table, see Table 1.

[0107] Table 1: Valve opening characteristic data lookup table.

[0108] Percentage of opening (%) Flow coefficient (Kv) Equivalent cross-sectional area (cm²) 0 0.00 0.00 20 8.35 5.28 40 16.70 10.56 60 25.05 15.84 80 33.40 21.12 100 41.75 26.40

[0109] When implementing the branch-and-bound strategy, a depth-first search tree is constructed. The first level of the tree corresponds to the first main branch, and each node's branch has five candidate opening values ​​(20%, 40%, 60%, 80%, 100%). Lower-level nodes inherit the opening value selection from their parent nodes and expand the branch for the next branch. The bound function calculation consists of two parts: the actual error of the determined branches and the optimistic estimate of the undetermined branches. The optimistic estimate takes the remaining error tolerance of the current optimal solution; when the cumulative error of a partial solution exceeds the bound, the branch is pruned. During the search process, a priority queue is maintained, and the node with the smallest error is expanded each time.

[0110] Dynamic constraint adjustments are implemented based on branch heat transfer efficiency. The efficiency loss calculation formula compares the current thermal resistance with the ideal thermal resistance; when the loss exceeds 15%, boundary relaxation is triggered. Constraint updates are divided into three levels: Level 1 relaxation (loss 15-25%) expands the boundary by 5%, Level 2 relaxation (loss 25-35%) expands by 10%, and Level 3 relaxation (loss >35%) expands by 15%. Each relaxation operation is recorded in the constraint revision log, including a timestamp, branch number, original boundary value, new boundary value, and triggering condition. The relaxed boundary must undergo hydrodynamic verification to ensure it does not exceed the pipeline's pressure limit.

[0111] The flow distribution scheme verification phase involves three checks: branch flow conservation verification, which accumulates the deviation between the flow rate of each branch and the total input flow rate, with an allowable error range of ±1.5%; valve response capability testing, which simulates the transient process during each opening degree switch to check for branches with overshoot exceeding 10%; and thermal balance verification, which calculates the node temperature distribution through a simplified model to ensure that the maximum temperature difference does not exceed a set threshold. Schemes that fail verification trigger an adjustment mechanism, prioritizing the adjustment of the three branches with the largest errors, and re-executing the gradient projection and discretization process.

[0112] The exception handling module monitors the entire execution process. In case of hardware communication failure, it uses the valid opening combination verified in the previous cycle; in case of real-time calculation timeout, it activates a fast approximation algorithm to generate a suboptimal solution; when boundary conflict detection identifies an infeasible solution, it automatically switches to a safety net strategy, setting all branches to a uniform opening of 60%. Execution results are sent to field devices via the control bus and simultaneously written to the operation database for archiving. Database records include the solution number, generation time, execution status, key parameters, and exception event codes, supporting retrieval and statistical analysis by time range.

[0113] The system maintenance interface provides manual intervention capabilities. Operators can manually adjust constraint boundary weights, overriding automatically generated parameters; forcibly specify the opening value of specific branches to lock their participation in the optimization process; and import historical successful solutions as initial solutions to accelerate the convergence process. All manual intervention operations generate audit logs, recording the operator ID, modified content, timestamp, and system status snapshot. Maintenance operations require a double confirmation process to prevent accidental operations from causing system instability.

[0114] The performance optimization module runs in the background, analyzing pattern characteristics in historical data. Branch flow distribution schemes are categorized by temperature distribution patterns, establishing a scheme feature library; constraint boundary adjustment records are statistically analyzed to identify high-frequency relaxation regions, marking them as system weaknesses; valve actuation frequency analysis identifies actuators with rapid wear, prompting preventative maintenance. These analysis results are used to optimize algorithm parameters, such as adjusting gradient projection step size coefficients and revising the search depth for branch and bound. Optimization parameters are automatically updated every 24 hours, and their effectiveness is verified through offline simulation before each update.

[0115] Example 5: The temperature monitoring system is configured with a multi-sensor array. Each battery cell is equipped with two platinum resistance temperature probes, located at the terminal connection and the middle of the casing, respectively. The data acquisition card polls and reads the probe data at a frequency of 4Hz. The raw readings are transmitted to the preprocessing unit via shielded twisted-pair cables. The preprocessing process includes three operations: probe cross-verification (alarm triggered when the difference exceeds 1.5℃), thermocouple cold junction compensation, and sensor nonlinearity correction. The corrected data is organized into a three-dimensional temperature matrix according to the spatial structure of the battery cluster, and the timestamp is aligned with the control system clock to an accuracy of ±5ms.

[0116] The adaptive Kalman filter employs a dual-state-space model. The main state vector is defined as the actual temperature of the battery cell, while the auxiliary state vector represents the coolant's thermal conduction delay effect. The process noise covariance matrix is ​​initially set as a diagonal matrix with diagonal elements of 0.25℃ and off-diagonal elements of zero. The measurement noise covariance is initialized according to the probe's accuracy level: 0.1℃ for Class A probes and 0.3℃ for Class B probes. The filtering prediction stage is based on thermodynamic equations, and the state transition matrix includes the battery material's specific heat capacity, the thermal coupling coefficient between adjacent cells, and the coolant flow rate parameters. The time lag effect is modeled as a first-order inertial element, with the time constant calculated inversely proportional to the current branch flow rate.

[0117] The Kalman gain dynamic adjustment mechanism implements a three-level calibration: the base gain is calculated using the standard covariance update formula; short-term calibration monitors the most recent 10 predicted residuals, scaling the gain by a factor of 0.8 when the residual standard deviation exceeds a set threshold; long-term calibration refits the noise statistical characteristics every 24 hours, updating the covariance matrix element values. The filtering output includes two data streams: the volumetric temperature estimate sequence is refreshed at 0.2-second intervals, and the hysteresis temperature field spatial distribution map is updated every 5 seconds. The anomaly detection module monitors abrupt gradient changes in the estimated values, generating a data quality warning when the absolute value of the gradient exceeds 0.8℃ / s for more than 3 seconds.

[0118] The feature extraction engine separates two types of features from the filtered data: spatial gradient features, which calculate the absolute value of the temperature difference between adjacent cells and cluster them into eight feature groups by region. Each group contains three statistics: maximum gradient value, average gradient value, and the percentage of gradient points exceeding 0.5℃. Temporal differential features use a fifth-order Savitzky-Golay filter to calculate the rate of temperature change with a window length of 15 seconds and a polynomial order of 3. The differential results are used to calculate the mean, variance, and zero-crossing frequency for each channel, forming a dynamic temporal feature vector.

[0119] After receiving spatial gradient features, the heat exchange control rule base update module adjusts parameters within a preset rule framework: when the average gradient of a region increases by 10%, the flow weight of that region increases by 0.05; if a single gradient point continuously exceeds the limit, the corresponding local control coefficient is increased by two levels. Rule weight updates are subject to priority constraints, with the maximum adjustment range for high-priority regions being three times that of low-priority regions. Update operations are executed in batches, with a maximum of three region parameters adjusted per control cycle (30 seconds) to prevent system oscillation.

[0120] The parameter adaptation process of the time series decomposition module is based on time differential features. The window length adjustment rule is as follows: if the variance of the rate of change increases by 20%, the window is shortened by 2 seconds; if the zero-crossing frequency decreases by 15%, the window is extended by 3 seconds. The switching threshold for the trend component extraction algorithm is set as follows: when the mean of the rate of change exceeds 0.3℃ / min, quadratic polynomial fitting is enabled; otherwise, linear fitting is maintained. The residual analysis standard is dynamically revised, and the residual threshold scales linearly with the standard deviation of the differential features, with a scaling factor of 0.6 times the standard deviation offset.

[0121] The full-cycle feedback process triggers optimization every 300 complete temperature datasets. The datasets are divided into three parts chronologically: the first 100 sets are used for parameter initialization, the middle 100 sets are used to train the rule base, and the last 100 sets are used to verify the optimization effect. The optimization instruction queue has two levels of buffering: a primary buffer stores unprocessed feedback data, and a deep buffer stores preprocessed optimization tasks. The task scheduler allocates computing resources based on system load, initiating full optimization during idle periods and performing only fine-tuning of key parameters during peak periods. Data transmission uses a shared memory area with verification, and each temperature data packet is appended with a CRC-16 checksum. When the block transmission error rate exceeds 0.1%, it switches to redundant transmission mode.

[0122] The closed-loop system stability monitoring implements triple protection: a parameter change rate limiter prevents single adjustments from exceeding the allowable range, such as limiting sudden changes in flow weight to within ±0.02; control output range constraints ensure that command values ​​do not exceed the physical limits of the actuator; and a historical trajectory backtracking module compares the temperature standard deviation under new and old parameters, automatically reverting to the previous stable version when the new parameter causes a 5% increase in the standard deviation. All parameter changes are recorded in the version control system, marking the operation time, change content, and system state snapshot, supporting the restoration of any historical configuration at a specific point in time.

[0123] Redundancy and backup mechanisms are implemented on heterogeneous hardware. The primary processor executes real-time control tasks, while the backup processor synchronously runs a simplified estimation algorithm. If the primary processor's data stream is interrupted for more than 3 seconds, the backup system takes over control and simplifies the feature processing flow: it only calculates spatial gradient features to update the weights of key branches, pausing updates to temporal differential features. The primary and backup systems synchronize their states hourly, triggering a manual diagnostic request when the difference exceeds a safety threshold. The system maintenance interface supports importing external calibration data to correct thermal model parameter deviations. The calibration data must contain at least 48 hours of continuous operating condition records and be validated through quadruple cross-validation.

[0124] Example 6: Upon system startup, the data acquisition buffer is initialized, and contiguous storage space is allocated to store the raw temperature time series. Temperature monitoring data for each battery cell is continuously input at a sampling rate of 4Hz. The data preprocessing module implements moving average filtering, with a fixed window width of 7 sampling points. Temperature series segments are extracted according to time windows, with the base time window set to 24 hours. After windowing, the segments are divided into 1440 1-minute segments. Three types of time-domain features are extracted from each segment: the arithmetic mean of the temperature mean, the variance using an unbiased estimation formula, and the skewness coefficient using the normalized moments method. The frequency domain feature extraction process includes four steps: applying the Hanning window function to suppress spectral leakage, performing a 512-point fast Fourier transform, normalizing the amplitude spectrum to the 0-1 range, and filtering the harmonic component index values ​​and corresponding amplitudes of the top five largest amplitude values. The feature normalization module scales all features to the same dimension, the mean feature is linearly mapped to the [0,1] interval, the variance feature is scaled after taking the natural logarithm, and the skewness feature retains the original value but is limited to ±10.

[0125] The feature tensor is constructed using a four-dimensional data structure. The first dimension corresponds to the spatial number of the 128 cells within the battery cluster, arranged in a zigzag pattern. The second dimension represents the time window index, with the indices of the most recent 30 time windows stored in a rolling manner. The third dimension allocates 16 feature slots: the first 8 slots store time-domain features (mean, variance, and skewness each occupy 1 slot, with the remaining 5 slots reserved), and the last 8 slots store frequency-domain features (fundamental amplitude + 4 harmonic amplitudes). The fourth dimension serves as an extended reserved space, currently filled with zero values. The tensor data is stored in a sparse format, and elements with a change in feature value less than 0.01 are marked as empty to reduce storage overhead.

[0126] Tensor decomposition employs Tucker model optimization. The core tensor is initialized to 8×8×8×8 dimensions, and the factor matrices are randomly initialized. The decomposition algorithm uses alternating least squares optimization, updating the four factor matrices sequentially in each iteration. The matrix update steps are as follows: fix the other three matrices, solve for the least squares solution of the current matrix, apply non-negativity constraints, and normalize the row vectors. Convergence is determined using a dual criterion: the rate of change of the objective function is less than 0.005% for three consecutive iterations, or the residual norm drops to 1% of its initial value. The core pattern component extraction strategy retains elements with an energy percentage exceeding 5% in the core tensor, and outputs the top 12 component indices and energy values ​​sorted by pattern importance.

[0127] The aging state mapping module comprises two layers. The first layer, a pattern filter, selects the top three core components by energy percentage as input features, with the cumulative cycle count of the battery as an auxiliary input. The second layer, a radial basis function network, is configured with 3 nodes in the input layer and 8 nodes in the hidden layer, employing a Gaussian kernel function, with the kernel width adaptively adjusted according to the range of input features. The network is trained using a historical aging dataset containing 200 battery samples with different degradation levels. During the online prediction phase, the network outputs a health state compensation coefficient, with a numerical range limited to 0.8-1.2. The coefficient correction logic implements a three-stage processing: when the coefficient is in the range of 0.92-1.08, the original value is output directly; below 0.92, a logarithmically smoothed value is applied; above 1.08, an upper limit truncation is applied.

[0128] The residual analysis compensation mechanism is implemented through dynamic thresholds. The basic anomaly threshold is set to 3.0 times the standard deviation of the historical residual mean. After receiving the health status coefficient, the compensation module applies an adjustment factor of 0.9 to the threshold. The coefficient adjustment range is limited to ±20%. When the coefficient is below 0.95, a sensitive mode is activated, and the threshold is reduced by 10% simultaneously. Maintenance alarm generation is set with two trigger conditions: a yellow alert is generated when the health coefficient remains below 0.85 for 3 consecutive hours; a red alert is generated when the coefficient remains below 0.80 for two consecutive days. The alarm information includes three parts: individual unit identifier code, health coefficient trajectory curve, and recommended maintenance level (observation / inspection / replacement).

[0129] The aging model update process is initiated during the monthly maintenance cycle. It loads the accumulated feature tensor data for the current month and re-executes tensor decomposition to generate new core patterns. Canonical correlation analysis is used to compare the old and new patterns, calculating the spatial correlation coefficient of the patterns. If the correlation of the main patterns is below 0.7, radial basis function network retraining is triggered: the network structure is preserved, and the hidden layer weights are updated using an incremental learning algorithm. Retraining validation uses the retained dataset for evaluation; if the prediction error exceeds 5%, it reverts to the previous model version. All model versions are stored with labeled data time intervals and key statistical indicators, supporting historical state backtracking analysis.

[0130] The system self-test module performs daily integrity checks. The signal-to-noise ratio (SNR) of the measured signals in the feature acquisition channels is monitored; channels with a SNR below 20dB are marked with a quality warning. Tensor decomposition convergence monitoring records the iteration distribution; abnormal divergence triggers algorithm parameter resets. The compensation coefficient output range verification setting includes an over-limit alarm; if the range is continuously exceeded, the output is frozen and switched to safe mode. The maintenance interface provides manual calibration functionality, allowing the import of offline detection health coefficient overlay automatic calculation results; all overlay operations record the operator's identity and the calibration reference file number. Data archiving employs a hierarchical storage strategy: raw temperature data is retained for 7 days, feature tensors are stored for 30 days, and core mode fractionation is simplified into a summary vector for permanent storage.

[0131] The operational status visualization interface renders a 3D heatmap. The basic view overlays real-time temperature distribution and a health coefficient mapping, using a color coding scheme: blue for health coefficients ≥1.0, cyan for the 0.95-0.99 range, yellow for the 0.90-0.94 range, and red for the ≤0.89 range. The historical trend view supports selecting individual health coefficient change curves for 12 months, automatically marking important inflection point events. The early warning management interface lists currently active alarms, supporting filtering by region and priority sorting. All views support data export to standard industrial report formats, with report templates conforming to the ISO 13379 condition monitoring document standard.

[0132] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0133] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for thermal management of an intelligent distributed liquid-cooled energy storage system, characterized in that, The application comprises: acquiring operating state parameters of a plurality of battery monomers in a cluster of energy storage units, and optimizing timestamp synchronization in the acquisition process; performing real-time analysis on the operating state parameters, processing temperature data using a small-batch time series decomposition method to obtain temperature variation trend characteristics and abnormal hot spot position information; optimizing a parameter set of a heat exchange control rule library through a non-dominated sorting strategy according to the temperature variation trend characteristics and the abnormal hot spot position information, generating optimized heat exchange control parameters and priority rules; based on the real-time collected liquid cooling system inlet temperature, flow parameter, and the optimized heat exchange control parameters and priority rules, using a tabu search mechanism to perform decision calculation on the cooling liquid flow distribution scheme, and outputting each branch flow control instruction; in response to the each branch flow control instruction, using an approximate gradient projection technology to dynamically adjust the flow constraint boundary, and combining a branch and bound strategy to discretize and optimize the cooling liquid distribution path, to realize fine regulation and control of the branch flow; based on the real-time collected liquid cooling system inlet temperature, flow parameter, and the optimized heat exchange control parameters and priority rules, using a tabu search mechanism to perform decision calculation on the cooling liquid flow distribution scheme, comprising: constructing a thermal resistance network model with branch flow as the decision variable; initializing the length of the tabu table and the neighborhood search step length; generating an initial flow distribution scheme according to the optimized heat exchange control parameters and priority rules; in each iteration, generating a neighborhood solution set based on the tabu table state, calculating the thermal balance deviation and pump power consumption corresponding to each solution; using a pardon criterion to receive a tabu solution better than the historical optimum, and updating the current optimal flow distribution scheme; when the maximum number of iterations is met, outputting each branch flow control instruction.

2. The intelligent distributed liquid-cooled energy storage system thermal management method of claim 1, wherein, acquiring operating state parameters of a plurality of battery monomers in a cluster of energy storage units, and optimizing timestamp synchronization in the acquisition process, comprising: the operating state parameters include temperature distribution, voltage fluctuation, and internal resistance change; dynamically setting the acquisition frequency and sampling period of battery monomers in different regions according to the historical temperature data change rate and system load working condition; synchronously collecting the temperature distribution, voltage fluctuation, and internal resistance change data of each battery monomer according to the set acquisition frequency and sampling period, to generate a time-aligned battery cluster state vector.

3. The intelligent distributed liquid-cooled energy storage system thermal management method of claim 1, wherein, performing real-time analysis on the operating state parameters, processing temperature data using a small-batch time series decomposition method, comprising: dividing the real-time acquired battery cluster temperature data stream into continuous time segment batches; initializing the window size and iteration step length of the temperature decomposition model; in each time segment batch, extracting the periodic component, trend component, and residual component of the temperature data, and calculating the energy proportion of each component; according to the spatial distribution density of the residual component, identifying abnormal hot spot coordinates that exceed a set threshold, and combining voltage fluctuation data to generate a hot spot risk level label; outputting analysis results containing temperature variation trend characteristics and abnormal hot spot position information with risk level labels.

4. The intelligent distributed liquid-cooled energy storage system thermal management method of claim 1, wherein, according to the temperature variation trend characteristics and the abnormal hot spot position information, optimizing a parameter set of a heat exchange control rule library through a non-dominated sorting strategy, comprising: Setting a thermal management objective function, including temperature uniformity index, energy consumption index and hotspot suppression response time; Initializing the candidate parameter population of the heat exchange control rule base; Applying the parameter combinations in the population to the heat exchange control rule base, evaluating the parameter combinations based on the temperature change trend characteristics and abnormal hotspot position information to obtain the Pareto front fitness of the parameter combinations; Performing tournament selection, simulated binary crossover and polynomial mutation operations on the parameter population to update the candidate solution set; Filtering the new generation population according to the non-dominated sorting rank and crowding distance until the convergence condition is met, and outputting the optimized heat exchange control parameters and priority rules.

5. The intelligent distributed liquid-cooled energy storage system thermal management method of claim 1, wherein, In response to the branch flow control instructions, the approximate gradient projection technique is used to dynamically adjust the flow constraint boundary, and the branch and bound strategy is used to discretize and optimize the cooling liquid distribution path, including: Analyzing the target flow value in the branch flow control instruction; Setting the dynamic flow constraint boundary according to the pipeline topological characteristics and pump performance curve; Calculating the absolute error gradient direction of the current flow distribution and the target flow; Projecting to the feasible region boundary along the gradient descent direction to obtain a continuous flow solution that satisfies the constraint; Discretizing the continuous flow solution into candidate valve opening combinations, pruning the invalid search space through the branch and bound strategy, and determining the optimal valve opening instruction.

6. The intelligent distributed liquid-cooled energy storage system thermal management method of claim 5, wherein, Using the approximate gradient projection technique to dynamically adjust the flow constraint boundary, including: Setting the relaxation factor and convergence threshold of the flow constraint boundary; Calculating the heat conduction efficiency loss of each branch under the current flow distribution scheme; Updating the relaxation factor of the flow constraint boundary according to the heat conduction efficiency loss direction; Applying the updated constraint boundary to the gradient projection calculation to generate a flow adjustment scheme that satisfies the new boundary.

7. The intelligent distributed liquid-cooled energy storage system thermal management method of claim 1, wherein, Further comprising: Collecting the temperature feedback data of the battery monomer after regulation and control through the adaptive Kalman estimation algorithm, and transmitting the processed temperature feedback data to the heat exchange control rule base, the time series decomposition module and the optimization module for updating the thermal management strategy; The adaptive Kalman estimation algorithm for collecting the temperature feedback data of the battery monomer after regulation and control, including: Collecting multi-point temperature monitoring data of the battery monomer within a set time window; Building a battery thermal state transfer model considering the time delay effect of the cooling liquid; Adaptively adjusting the Kalman gain coefficient according to the noise statistical characteristics of the temperature monitoring data; Outputting the battery body temperature estimate value after removing the measurement noise and time delay interference.

8. The intelligent distributed liquid-cooled energy storage system thermal management method of claim 7, wherein, Transmitting the processed temperature feedback data to the heat exchange control rule base, the time series decomposition module and the optimization module, including: Extracting the spatial gradient distribution and time differential characteristics of the temperature estimate value; Inputting the spatial gradient distribution characteristics into the heat exchange control rule base to update the hotspot suppression rule weight; Inputting the time differential characteristics into the time series decomposition module to adjust the window parameter of the trend component extraction; Inputting the complete temperature feedback data set into the optimization module to trigger the parameter set re-optimization process of the heat exchange control rule base.

9. The intelligent distributed liquid-cooled energy storage system thermal management method of claim 1, wherein, Further comprising: Generating a multi-dimensional thermodynamic feature vector according to the cumulative distribution characteristics of the temperature feedback data; Extracting the core mode component of the feature vector through tensor decomposition technology; Mapping the core mode component to the system aging state space to update the battery health state compensation coefficient; The updated battery health state compensation factor is fed back to the residual analysis step in the time series decomposition method. The updated battery health state compensation factor is fed back to the residual analysis step in the time series decomposition method.

Images