Digital twin multi-field coupling sodium mobile power station air-cooled thermal management system and control method

Abstract

This invention discloses a digital twin multi-field coupled sodium mobile power station air-cooled thermal management system and control method, belonging to the field of electrochemical energy storage thermal management. Addressing the problems of easy leakage in liquid cooling schemes and the inability of traditional air cooling to adapt to deformation under vibration conditions in mobile power stations, this invention proposes using sodium-ion batteries without configuring liquid cooling pipelines. A reduced-order digital twin model based on intrinsic orthogonal decomposition and Galerkin projection is constructed to reconstruct the flow and temperature fields in real time. Microscopic deformation of the air duct is estimated using an inertial measurement unit and extended Kalman filter, combined with Grassmann manifold interpolation for online compensation of the model. A physically constrained LSTM-XGBoost hybrid model is used to predict hotspots, and model predictive control is employed to collaboratively optimize multi-zone variable frequency fans, active air guide louvers, and dynamic spoilers. This invention achieves efficient temperature control and thermal safety early warning for the air-cooled system under vibration conditions, improves temperature consistency, and reduces system complexity and maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of thermal management and safety control technology for electrochemical energy storage systems. Specifically, it relates to a sodium-ion battery pack air-cooled thermal management system and its control method for mobile power stations or mobile charging vehicles. This invention is particularly applicable to achieving precise temperature control and safety early warning for battery packs without the need for liquid cooling pipelines. It achieves this by constructing a digital twin model and combining it with multi-field coupling analysis, fluid-structure interaction sensing, and model predictive control. This invention belongs to the interdisciplinary technology direction of distributed energy systems, mobile energy storage equipment, and IoT edge intelligent control. Background Technology

[0002] Mobile power stations have become an important infrastructure for emergency power supply, distributed power consumption, and off-grid power supply. Unlike stationary energy storage power stations, mobile power stations need to endure long-term travel on unpaved roads, frequent vibration and shock, and drastic changes in ambient temperature. This places far greater demands on the thermal management and safety reliability of the battery system than on conventional energy storage systems.

[0003] Currently, most mobile energy storage equipment uses high-nickel ternary lithium-ion batteries combined with liquid cooling plates or immersion liquid cooling architectures. However, under high-rate charge-discharge and complex vibration conditions, the welds and joints of the liquid cooling pipelines are prone to fatigue, loosening, and leakage. This not only leads to a decrease in cooling capacity but may also cause a reduction in insulation resistance, short circuits, and even electrical fires. Meanwhile, the upcoming new generation of safety standards (such as GB 38031-2025) imposes more stringent requirements on power batteries, stipulating that the system must remain fire-free and explosion-free for two hours after a single cell experiences thermal runaway, and possess intelligent diagnostic capabilities to issue a warning signal five minutes in advance. This places enormous pressure on the traditional "high-nickel lithium battery + complex liquid cooling" solution in terms of cost, safety, and maintainability.

[0004] Sodium-ion batteries, with an energy density slightly lower than high-nickel ternary lithium batteries and close to that of lithium iron phosphate batteries, offer a wider operating temperature range, milder polarization characteristics, and lower high-rate charge / discharge heat generation rates. Analysis based on the Bernardi heat generation equation and experimental data show that, within the 1C to 4C rate range, sodium-ion batteries exhibit significantly lower polarization heating than lithium-ion batteries of the same capacity, with a smoother heat generation curve and lower peak values. This provides a physical basis for employing optimized active air-cooled thermal management systems in mobile energy storage scenarios.

[0005] However, traditional air-cooled energy storage systems mostly employ static duct structures and fixed wind speed strategies. Cooling air is passively distributed to the battery clusters via a plenum chamber and fixed ducts from the air conditioner. This results in uneven flow field distribution, severe localized "thermal islands," and control logic relying on temperature-dependent feedback lag. More critically, such systems cannot detect or compensate for the minute geometric deformations of the ducts caused by vehicle vibrations. Under the long-term vibration and shock conditions of mobile power stations, relying solely on traditional air-cooling solutions makes it difficult to simultaneously meet temperature safety, consistency, and energy efficiency requirements without introducing a liquid-cooled piping network.

[0006] Therefore, based on the suitability of sodium-ion batteries for air cooling, it is necessary to construct a mobile power station air-cooled thermal management system that combines digital twins, multi-field coupling simulation, fluid-structure interaction sensing, and intelligent predictive control, in order to achieve temperature control and safety margins that are close to or even better than those of traditional liquid-cooled architectures without completely eliminating liquid cooling pipe networks. Summary of the Invention

[0007] To address the technical challenges of existing liquid cooling solutions in mobile energy storage systems posing leakage risks under vibration conditions, and traditional air cooling solutions exhibiting uneven flow field distribution, inability to adapt to vibration deformation, and control lag, making it difficult to simultaneously meet high safety standards, temperature consistency, and low maintenance costs, this invention provides a digital twin multi-field coupled sodium mobile power station air-cooled thermal management system, comprising: The sodium-ion battery cluster, consisting of multiple sodium-ion battery modules, is installed in the energy storage compartment of the mobile power station. The system design is based on the low heat generation characteristics of sodium-ion batteries within the rated rate and ambient temperature range, and determines the thermal management scheme based on air cooling and the boundary of heat source terms in the digital twin model.

[0008] The air duct and airflow guiding mechanism includes a main air duct and several branch air ducts arranged around the battery cluster, as well as active airflow guiding louvers set at the flow split nodes of the main air duct and branch air ducts, and dynamic baffles set in some branch air ducts.

[0009] The multi-zone variable frequency fan array divides the energy storage compartment into multiple thermal zones. Each zone is equipped with at least one electronic converter fan independently controlled by a frequency converter to supply cold air to the corresponding zone.

[0010] The sensor matrix includes temperature sensors, wind speed sensors, and differential pressure sensors deployed inside the battery module and at key locations in the air duct, as well as inertial measurement units deployed on the vehicle body structure and air duct frame, used to collect information on battery temperature, wind speed, wind pressure difference, and vehicle vibration.

[0011] The edge computing controller, deployed locally in the vehicle, incorporates a reduced-order aerodynamic and thermodynamic digital twin model built based on intrinsic orthogonal decomposition and Galerkin projection, a Grassmann manifold geometric deformation interpolation module, a fluid-structure interaction displacement estimation module, a hotspot prediction module, and a model predictive control module. It is used to solve the current flow field and temperature field in real time based on sensor data, and to issue control commands to the fan, air guide louvers, and spoilers to realize dynamic reconstruction of the air duct and targeted delivery of cold air.

[0012] The cloud-based optimization platform communicates with the edge computing controller to train and update the digital twin model and control strategy based on historical operating data and full-scale simulation results.

[0013] The corresponding control method includes the following steps: 1. Data on the temperature, wind speed, wind pressure difference, and vehicle vibration of the sodium-ion battery cluster are collected through temperature sensors, wind speed sensors, differential pressure sensors, and inertial measurement units, and then sent to the edge computing controller.

[0014] 2. The edge computing controller abstracts the key air duct and supporting structure into a mass-damping-stiffness system. It uses the extended Kalman filter algorithm to estimate the state of the acceleration signals output by the inertial measurement units installed on the air duct and supporting structure, and calculates the micro-deformation displacement of the air duct relative to the vehicle body. This micro-deformation is then used as a fluid-structure interaction boundary condition input to the digital twin simulation engine based on the reduced-order model.

[0015] 3. Within a given control cycle, the digital twin simulation engine solves for the current flow field and temperature field distribution within the energy storage compartment based on the current boundary conditions and real-time heat generation information. Specifically, the digital twin model uses intrinsic orthogonal decomposition and Galerkin projection to construct reduced-order models of the flow field and temperature field.

[0016] 4. Within the set prediction time window, the hotspot prediction model integrated into the edge computing controller is used to predict the temperature evolution trajectory of each spatial node based on the current operating data and digital twin output results, and to identify the location and intensity of possible hotspots.

[0017] 5. When the prediction results indicate that the future temperature rise of a certain defense zone or a certain channel will exceed the preset safety threshold, the model prediction control module calculates the control quantity according to the objective function and constraint conditions, switches the system from steady-state cruise mode to predictive cold storage mode, increases the speed of the variable frequency fan in the target defense zone and adjusts the opening and angle of attack of the corresponding air guide louvers, and guides more cold air vectors to the target defense zone to achieve feedforward cold storage.

[0018] 6. When the measured temperature, wind speed and differential pressure data indicate that there is severe blockage in the local flow channel and the risk of hot spots is increased, the model predictive control module will switch the system from steady-state cruise mode to emergency dynamic reconfiguration mode, further instruct the variable frequency fan in the target defense zone to run at high speed, control the corresponding air guide louvers to deflect significantly, and control the dynamic baffle to pop out from the duct wall to the set angle of attack, so as to destroy the boundary layer and induce turbulence to enhance local convective heat transfer.

[0019] 7. During the above control process, the edge computing controller updates the digital twin simulation results at a preset frequency, iteratively optimizes the control commands based on the updated flow field and temperature field, and smoothly degrades the system to steady-state cruise mode after the temperature gradient in the target area recovers to a safe range.

[0020] In the digital twin model, the expansions of the transient velocity field and temperature field on the POD orthogonal basis are as follows: (1); (2); in Indicates spatial location and the transient velocity field vector at time t; Indicates spatial location and the transient temperature field at time t; and These are the time-averaged fields (reference fields) of the velocity field and the temperature field, respectively. The modal ordinal number takes values ​​from 1 to 1. ,in The total number of POD modes to be retained; and The velocity field and temperature field are respectively the first The time coefficients corresponding to the first mode; and The first and second fields are the velocity field and the temperature field, respectively. Orthogonal modes (basis functions) in the POD space.

[0021] By applying Galerkin projection to the Navier-Stokes equations, the reduced-order ROM state equations are obtained: (3); in Let the vector be a column vector composed of the time coefficients of each velocity mode, and the points on it be... This represents the first derivative with respect to time. This is the constant term vector, derived from the constant contributions in the Galerkin projection; This is a linear coefficient matrix, corresponding to the linear viscous dissipation term in the flow; For nonlinear convection tensors, corresponding to the nonlinear convection terms in the Navier-Stokes equations, the product... Represents a quadratic form; To control the input matrix; The control vector is a collection of control inputs such as fan speed and louver angle for each protection zone.

[0022] Furthermore, regarding the geometric deformation of the air duct under vibration conditions in mobile power stations, this invention employs the Grassmann manifold interpolation method to generate a "pseudo-POD mode" under the current deformation state: (4); in This indicates the current geometric deformation parameters of the air duct (e.g., the relative displacement of the air duct cross section caused by vibration). and The POD orthogonal basis matrices are pre-computed and stored for two sets of typical geometric deformation parameters (such as before and after deformation), which are regarded as two points on the Grassmann manifold; These are the normalized interpolation coefficients, based on the current deformation. exist and The relative positions between corresponding shape variables are determined; Indicates will Mapped to Logarithmic mapping of the tangent space; Indicates from The tangent space is reverted to an exponential mapping of the Grassmann manifold; These are pseudo-POD modes generated online, used to update the basis functions of the reduced-order model under the current geometric state.

[0023] To estimate the micro-displacement of the air duct caused by vibration, this invention abstracts the key air duct and supporting structure into a mass-damping-stiffness system, whose structural dynamic equations are as follows: (5); in The relative displacement vector of the air duct structure with respect to the vehicle body base (varying with time t); The relative velocity vector ( (first derivative with respect to time) The relative acceleration vector ( (Second derivative with respect to time); M is the mass matrix of the structural system; C is the damping matrix; K is the stiffness matrix; The base excitation acceleration vector is obtained from actual measurements by an IMU mounted on the rigid chassis (non-suspended part).

[0024] An extended Kalman filter (EKF) is used to estimate the state of the system. An augmented state vector is defined. ,in For accelerometer dynamic zero bias, the EKF prediction step is: (6); (7); Where the subscript k represents the discrete time step number; This is the estimate of the prior state at step k, obtained from the posterior state prediction at step k-1; For the state transition function (discretized structural dynamics equations); For predicting the state covariance matrix; Let f be the Jacobian matrix of the state transition function f at step k; Let be the process noise covariance matrix.

[0025] The update steps of EKF (Kalman gain and state update) are as follows: (8); (9); in The Kalman gain matrix; For observation function Jacobian matrix; To measure the noise covariance matrix; This is the measured acceleration vector of the IMU installed on the air duct at step k; The observation function maps the state vector to the observation space. This is the posterior state estimate (filtered result) for the k-th step.

[0026] The relative displacement of the air duct is obtained through the above EKF iteration. The accurate estimation suppresses the low-frequency drift caused by the quadratic integral.

[0027] In the hotspot prediction model, this invention first constructs a physical constraint feature matrix: (10); Where t represents the current time step; The battery's dynamic internal resistance (identified online using an equivalent circuit model); For Joule heat integral, representing the accumulated irreversible heat; and These are the first and second derivatives of the temperature rise extracted from the temperature sequence, respectively. The local wind speed is output from the ROM; For the characteristics of air duct deformation (e.g., by displacement) (derived dimensionless deformation coefficients). To accumulate vibrational energy.

[0028] The expression for the LSTM time-series coding layer is: (11); Where L is the time window length of the LSTM; The input is the feature matrix for L consecutive time steps; The hidden state vector output by the LSTM network represents the dynamic law of heat accumulation and decay.

[0029] The hidden state vector is concatenated with the cell spatial coordinates and other local operating condition features, and then input into the XGBoost regressor to predict the future. Temperature after minutes: (12); in For the predicted future Temperature or hotspot intensity value after minutes; M is the total number of regression trees in the XGBoost model; Let m be the function of the m-th regression tree, belonging to the regression tree function space. ; It represents the spatial coordinates of the battery cell and the characteristic vector of its local operating conditions (such as its relative position to the air duct outlet).

[0030] The prediction results are decomposed into interpretability using SHAP values: (13); in d represents the baseline output (the expected output of the model when all features are averaged); d is the total dimension of the input features. For the j-th input feature, the current predicted value The Shapley contribution value, whose magnitude and sign represent the marginal impact of this feature on hotspot formation.

[0031] Compared with the prior art, the present invention has the following beneficial effects: 1. Significantly improve temperature safety and consistency: By combining a reduced-order digital twin model (ROM), model predictive control (MPC), and dynamic turbulence structure, the maximum temperature and maximum temperature difference can be controlled within the set safety threshold under high-rate operating conditions, effectively suppressing the formation of hot spots and "thermal islands", extending cell life and meeting the requirements of the new generation of safety standards for thermal runaway and early warning.

[0032] 2. Simplify system structure and reduce maintenance costs: Completely eliminate the liquid cooling pipeline network and its supporting pumps, heat exchangers, coolant and other components, fundamentally eliminating the risk of leakage in vibration environments, significantly reducing system weight, structural complexity and maintenance workload and downtime costs under long-term vibration conditions, making it particularly suitable for highly mobile application scenarios such as mobile power stations.

[0033] 3. Optimize energy efficiency and noise: By using multi-zone variable frequency fans and dynamic spoilers that pop up as needed, a balance is achieved between zoned strong cooling and global energy saving, avoiding overcooling of the entire cabin and ineffective airflow. While meeting thermal management requirements, fan power consumption and operating noise are reduced, thus improving the overall energy efficiency of the system.

[0034] 4. Enhanced adaptability to complex vibration conditions: The innovative introduction of the IMU+EKF fluid-structure interaction displacement estimation algorithm, combined with the reduced-order model of Grassmann manifold interpolation, enables online compensation for the geometric deformation of the air duct caused by vibration. This allows the air-cooling system to maintain effective cooling and temperature uniformity under complex conditions such as rugged roads and severe bumps, avoiding the problem of traditional static air duct models failing in vibration scenarios.

[0035] 5. Possesses interpretable intelligent predictive control capabilities: Introduces a hybrid hotspot prediction model based on physical constraint features using LSTM+XGBoost, combined with the SHAP interpretability mechanism (in Equation 13). The contribution value enables hotspot prediction and control decisions to have a clear causal chain, which facilitates compliance with automotive-grade functional safety standards and regulatory requirements, and can provide visual evidence in fault analysis and model iteration.

[0036] Other features and advantages of the invention will be set forth in the following description or may be learned by practicing the invention. Attached Figure Description

[0037] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0038] Figure 1 This is a schematic diagram of the overall structure of the mobile sodium-ion energy storage vehicle, including the vehicle body, energy storage compartment, and main functional areas.

[0039] Figure 2 This is a schematic diagram showing the arrangement of the multi-zone air ducts, active air guide louvers, and dynamic spoilers in the battery compartment.

[0040] Figure 3 This is a schematic diagram of the sensor matrix topology, including the locations of temperature, wind speed, differential pressure, and IMUs.

[0041] Figure 4Flowchart for switching between MPC control state machine, steady-state cruise, predictive feedforward cold storage, and emergency dynamic reconfiguration modes. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. Those skilled in the art should understand that these embodiments are only for explaining the invention and are not intended to limit the scope of protection of the invention.

[0043] Example 1, System Hardware Configuration and Installation: This embodiment provides a digital twin multi-field coupled sodium mobile power station air-cooled and heat-cooled management system.

[0044] like Figure 1 As shown, the overall structure of the mobile sodium-ion energy storage vehicle in this embodiment includes components such as the vehicle body and the energy storage compartment. First, a sodium-ion battery cluster is installed in the independent energy storage compartment of the mobile energy storage vehicle. The battery cluster consists of multiple square sodium-ion battery modules with a rated capacity of 150Ah. Based on the characteristic that sodium-ion batteries generate significantly less polarization heat than lithium batteries at 2C to 4C rates, a thermal management scheme primarily based on forced air cooling is determined.

[0045] Secondly, construct the air duct and airflow guiding mechanism. An S-shaped main air duct is constructed around the battery cluster, and the area is divided into four thermal protection zones (A, B, C, and D) according to the battery module layout. The arrangement of the main air duct, branch air ducts, active air guide louvers, and dynamic spoilers is as follows: Figure 2 As shown. At the air inlet branch node of each defense zone, an active air guide louver driven by a servo motor is installed. The blades are made of aluminum alloy and the angle of attack can be adjusted from -30° to +30°. On the inner wall of the cooling channel of each battery module, a retractable dynamic spoiler is arranged. Under normal operating conditions, the spoiler is attached to the wall to reduce wind resistance.

[0046] Next, deploy actuators and sensors. Each zone is independently equipped with a 150W rated power EC fan, achieving 0-100% stepless speed regulation via PWM. The locations of temperature sensors, wind speed sensors, differential pressure sensors, and inertial measurement units are as follows: Figure 3 As indicated in the diagram. The sensor matrix includes: T-type thermocouples arranged at the center of each cell terminal, busbar, and module; wind speed sensors and differential pressure sensors arranged at the main air duct and the characteristic sections of each branch; and a 6-axis inertial measurement unit (IMU) arranged at each of the four corners of the rigid main beam of the vehicle body, the air duct frame, and the key stress points of the battery bracket.

[0047] Finally, the control system is deployed. The local edge computing controller uses an automotive-grade SOC chip, which integrates a cloud-trained and compressed reduced-order digital twin (ROM) model, a Grassmann manifold interpolator, an extended Kalman filter (EKF), an LSTM-XGBoost hybrid prediction model, and a model predictive control (MPC) algorithm module. The cloud optimization platform is a high-performance server used for offline full-scale CFD simulation and structural dynamics simulation, and for constructing POD orthogonal bases.

[0048] Example 2: Digital twin model and online fluid-structure interaction compensation: This embodiment describes in detail how the edge computing controller constructs a digital twin model and achieves online compensation for vibration conditions.

[0049] 2.1 Offline Construction and Online Update of Reduced-Order Digital Twin Models: In the offline phase, full-scale CFD simulations were performed for various parameter combinations, including fan speed, louver angle, ambient temperature, sodium ion-generated hot water, and several typical deformation states that the duct might undergo under vibration (e.g., duct cross-section compression or tension of 0.5mm, 1mm, 2mm, etc.), to obtain snapshots of the flow and temperature fields. Intrinsic orthogonal decomposition (POD) was then used to extract a set of spatially orthogonal modes ordered by energy. and The transient velocity and temperature fields are expanded on the POD basis, as shown in equations (1) and (2). For the velocity field, For the average velocity field, The time coefficients of the i-th velocity mode are... Let be the spatial basis functions of the i-th velocity mode; similarly, For the temperature field, For the average temperature field, and These represent the time coefficient and spatial basis functions of the temperature field, respectively. The number of modes retained. Determined based on energy percentage (e.g., 99% energy).

[0050] By applying Galerkin projection to the Navier-Stokes equations, the reduced-order ROM state equations are obtained, as shown in equation (3). In the equation... It is a time coefficient vector The derivative with respect to time; , , and The constant term, linear matrix, nonlinear tensor, and input matrix are respectively obtained offline from the Galerkin projection. The control vector includes the setpoints for the fan speed and the angle of the air guide louvers for each zone. The edge computing controller only needs to perform numerical integration on the ordinary differential equation system (3) within the control cycle (with a cycle of approximately 5 milliseconds) to obtain the time coefficient in real time. Substituting these equations into equations (1) and (2) reconstructs the flow field and temperature field of the entire cabin, reducing the computational burden far from directly solving the original partial differential equations.

[0051] 2.2 Fluid-structure interaction displacement estimation based on IMU+EKF: While the vehicle is in motion, the chassis IMU continuously collects the base excitation acceleration from the road surface. Meanwhile, the IMU installed on the air duct frame collects apparent acceleration. The edge computing controller abstracts the air duct and supporting structure into a mass-damping-stiffness system, whose dynamic equation is shown in equation (5). Where... The relative displacement of the air duct structure is represented by M, C, and K, which are the mass, damping, and stiffness matrices, respectively (pre-calibrated through finite element modal analysis). (Right side) This represents the inertial force of the base.

[0052] To accurately estimate displacement from noisy acceleration signals The controller runs an extended Kalman filter (EKF). An augmented state vector is defined. ,in For speed, The accelerometer is dynamically zero biased. The prediction step of EKF is shown in equations (6) and (7): It is based on the posterior state of the previous time step. and base incentive Prior state estimates obtained through recursion using the state transition function f; For prior covariance, For f, Jacobi The process noise covariance is then calculated. Then, IMU observations installed on the air duct are used. Update: First, calculate the Kalman gain. As in equation (8), where Let h be the Jacobian of the observation function. To measure the noise covariance; then, the posterior state estimate is obtained as shown in equation (9). ,in These are predicted observations. This EKF simultaneously estimates displacement and accelerometer bias, effectively suppressing low-frequency drift caused by direct quadratic integration of acceleration.

[0053] 2.3 Grassmann manifold interpolation and geometric deformation processing: The displacement vector obtained by EKF estimation Mapped to the geometric deformation parameters of the current air duct (For example, the percentage change in the height of the duct cross-section). For different typical deformation parameters, corresponding POD orthogonal bases have been constructed in the cloud. (correspond =0) and (correspond =1), they are considered as two points on the Grassmann manifold. The edge controller utilizes the log-exponential mapping on the Grassmann manifold, as shown in Equation (4), to generate the "pseudo-POD mode" in the current deformation state. In the formula: To be based on the current Interpolation coefficients calculated online; It is Mapped to Logarithmic mapping of the tangent space (offline computation and storage); For the exponential mapping from the tangent space back to the manifold (computed online), the pseudo-POD mode is... Substitute the original bases in equations (1) and (2) and This allows for the acquisition of a reduced-order model that matches the current vibration deformation, enabling online compensation of fluid-structure interaction.

[0054] Example 3: Hotspot prediction and MPC collaborative control: This embodiment describes the specific control process under the 4C high-rate charging condition of the vehicle.

[0055] 3.1 Physically Constrained Hotspot Prediction: The controller detects an increase in charging current, causing the temperatures of various features to rise. The hotspot prediction module first calculates the dynamic internal resistance online based on Bernardi's heat generation equation and the equivalent circuit model. Joule heat integral Entropy change thermal fluctuation coefficient, etc., and extract the first derivative of temperature rise from the temperature series. and second derivative Simultaneously read the local average wind speed output from the ROM. Characteristics of air duct deformation (by displacement) (Exported) and accumulated vibrational energy The feature matrix that constitutes each node As shown in equation (10). The meanings of each symbol have been defined in detail in the invention description.

[0056] Long Short-Term Memory (LSTM) network features a sequence of features from the past L=20 time steps (60 seconds in total). Encode the hidden state vector as shown in Equation (11). .here The dynamic laws governing heat accumulation and decay are encoded. Then the hidden state vector is... With cell space coordinates The model is concatenated and input into an extreme gradient boosting tree (XGBoost) regression model, as shown in Equation (12), to predict the future. =Temperature after 10 minutes In the formula, M represents the total number of trees, and each regression tree represents a regression tree. Features The scores are accumulated. The contribution of each feature to the predicted value is obtained through SHAP decomposition according to equation (13). ,in As the baseline, The sign and size indicate whether the j-th feature (such as local wind speed) inhibits or promotes hotspot formation.

[0057] 3.2 Model Predictive Control (MPC) and Mode Switching: When the predictive model indicates that the temperature of a module in the center of Zone B will exceed the safe threshold of 55°C (currently 48°C) in 10 minutes, the system immediately switches from steady-state cruise mode to predictive cooling mode (e.g., Figure 4 As shown), the MPC feedforward cold storage is started.

[0058] The MPC controller uses the discrete temperature field of the entire cabin as the state vector, and the fan speed, louver angle, and spoiler displacement of the four zones as the control vectors. Within a 10-minute prediction time domain, the state evolution is deduced using the ROM state equation of equation (3) and the hotspot prediction model of equation (12), and the optimal control sequence is obtained by solving a quadratic optimization problem. The objective function includes three terms: the sum of squares of the deviations of each temperature node from the reference safe temperature; the sum of squares of the control quantity increments between adjacent control cycles (to suppress frequent large-amplitude movements of the actuator); and the weighted sum of the fan and servo power (to suppress auxiliary energy consumption). The constraints include: the predicted temperature of all temperature nodes does not exceed 60℃, the temperature difference between any two nodes does not exceed 8℃, and the fan speed, the angle of the air guide louvers, and the displacement and rate of change of the baffle do not exceed the physical allowable range.

[0059] The first step of the optimal control sequence obtained is to increase the fan speed in zone B from 60% to 95% and adjust the angle of attack of the air guide louvers in zone B to +25° (pressurized air supply), while simultaneously reducing the fan speed in the other zones to 45%. After the command is issued, a large amount of cold air is vectored to zone B to perform feedforward cold storage.

[0060] However, due to localized blockage of the air ducts in Zone B caused by vibration, the measured temperature did not decrease effectively, and the wind speed and differential pressure data indicated severe blockage in the local flow channels, further increasing the risk of hotspots. Based on the conditions triggered by the measured data, the MPC controller automatically switched from "steady-state cruise" mode to "emergency dynamic reconfiguration" mode. Figure 4 As shown, the state machine triggers mode switching here based on the hotspot prediction risk value and temperature rise slope. The controller issues a command: the B zone fan accelerates to 100%, the corresponding air guide louver angle is adjusted to +30°, and an activation command is sent to the dynamic spoiler drive mechanism in the B zone battery module. The micro push rod immediately pops the spoiler to a 15° angle of attack with the airflow direction, artificially inducing eddy shedding and secondary flow, thus disrupting the boundary layer. This significantly improves the local convective heat transfer coefficient, and the module hotspot temperature drops rapidly. After the temperature recovers to below 48°C, the controller retracts the spoiler, the fan speeds smoothly decrease, and the system returns to steady-state cruise mode.

[0061] During this process, SHAP analysis (Equation 13) shows that "local wind speed" and "cumulative vibrational energy" These are the main contributing features that led to this hot topic (corresponding to) (If the absolute value is large), this information is recorded for subsequent model optimization and functional safety auditing.

[0062] Example 4, Fail-Safe Mode: When the edge computing controller detects that the digital twin simulation engine has crashed for unknown reasons or that the algorithm is not converging, the system automatically executes a safe return mode. In this mode, all advanced control functions (such as MPC, hotspot prediction, etc.) are suspended, and the controller performs the following operations: 1. Set all zone EC fans to the preset safe speed (e.g., 80% of the rated speed). 2. Return all active air guide louvers to 0° (even airflow distribution state); 3. Forcefully retract all dynamic spoilers to their wall-attached positions.

[0063] This mode ensures that even in the event of a control system malfunction, the system can still maintain basic heat dissipation and thermal safety at the maximum capacity of traditional air cooling, avoiding the risk of thermal runaway due to control failure, thereby meeting the requirements of automotive-grade functional safety standards (such as ISO26262).

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art can make various improvements and modifications without departing from the spirit and principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A digital twin multi-field coupled sodium mobile power station air-cooled and heat-managed system, characterized in that, include: The sodium-ion battery cluster, consisting of multiple sodium-ion battery modules, is installed in the energy storage compartment of the mobile power station. The system design is based on the low heat generation characteristics of sodium-ion batteries within the rated rate and ambient temperature range, and a thermal management scheme based on air cooling is determined. The air duct and airflow guiding mechanism includes a main air duct and several branch air ducts arranged around the sodium-ion battery cluster, as well as active airflow guiding louvers set at the flow split nodes of the main air duct and the branch air ducts, and dynamic baffles set in some of the branch air ducts. The multi-zone variable frequency fan array divides the energy storage compartment into multiple thermal zones. Each zone is equipped with at least one electronic converter fan independently controlled by a frequency converter to supply cold air to the corresponding zone. The sensor matrix includes temperature sensors, wind speed sensors, and differential pressure sensors deployed inside the battery module and at key locations in the air duct, as well as inertial measurement units deployed on the vehicle body structure and air duct frame, used to collect battery temperature, wind speed, wind pressure difference and vehicle vibration information. The edge computing controller, deployed locally in the vehicle, incorporates a reduced-order aerodynamic and thermodynamic digital twin model based on intrinsic orthogonal decomposition and Galerkin projection, a Grassmann manifold geometric deformation interpolation module, a fluid-structure interaction displacement estimation module, a hotspot prediction module, and a model predictive control module. It is used to solve the current flow field and temperature field in real time based on the data collected by the sensor matrix, and to send control commands to the multi-zone variable frequency fan array, active air guide louvers, and dynamic spoilers to realize dynamic reconstruction of the air duct and targeted delivery of cold air. The cloud-based optimization platform communicates with the edge computing controller and is used to train and update the digital twin model and control strategy based on historical operating data and full-scale simulation results. Without configuring a liquid cooling network, the system achieves consistent temperature control and hotspot prediction and intervention for sodium-ion batteries in the energy storage compartment through the coordinated operation of the above-mentioned components.

2. A control method for a digital twin multi-field coupled sodium mobile power station wind-cooled and heat-managed system as described in claim 1, characterized in that, Includes the following steps: S1 collects data on sodium-ion battery cluster temperature, wind speed, wind pressure difference, and vehicle vibration through temperature sensors, wind speed sensors, differential pressure sensors, and inertial measurement units, and sends the data to the edge computing controller. S2, the edge computing controller abstracts the key air duct and support structure into a mass-damping-stiffness system, uses the extended Kalman filter algorithm to estimate the state of the acceleration signal output by the inertial measurement unit installed on the air duct and support structure, calculates the micro deformation displacement of the air duct relative to the vehicle body, and inputs the micro deformation as a fluid-structure interaction boundary condition into the digital twin simulation engine based on the reduced-order model. S3, within a given control cycle, the digital twin simulation engine solves for the current flow field and temperature field distribution inside the energy storage compartment based on the current boundary conditions and real-time heat generation information; S4. Within the set prediction time window, the hotspot prediction model integrated into the edge computing controller is used to predict the temperature evolution trajectory of each spatial node based on the current operating data and the digital twin output results, and to identify the location and intensity of possible hotspots. S5, when the prediction results indicate that the future temperature rise of a certain defense zone or a certain channel will exceed the preset safety threshold, the model prediction control module calculates the control quantity according to the objective function and constraint conditions, increases the speed of the variable frequency fan in the target defense zone and adjusts the opening and angle of attack of the corresponding air guide louvers, and guides more cold air vectors to the target defense zone to achieve feedforward cold storage; S6, when the measured temperature, wind speed and differential pressure data indicate that there is severe blockage in the local flow channel and the risk of hot spots is increased, the model prediction control module further instructs the variable frequency fan in the target defense zone to run at high speed, controls the corresponding air guide louvers to deflect significantly, and controls the dynamic baffle to pop out from the duct wall to the set angle of attack, so as to destroy the boundary layer and induce turbulence to enhance local convective heat transfer. S7. During the above control process, the edge computing controller updates the digital twin simulation results at a preset frequency, iteratively optimizes the control commands based on the updated flow field and temperature field, and smoothly degrades the system to steady-state cruise mode after the temperature gradient of the target area is restored to a safe range.

3. The system according to claim 1, characterized in that, The digital twin model uses intrinsic orthogonal decomposition and Galerkin projection to construct reduced-order models of the flow and temperature fields. The transient velocity and temperature fields are expanded on the POD orthogonal basis as follows: ； ； in Indicates spatial location and the transient velocity field vector at time t; Indicates spatial location and the transient temperature field at time t; and These are the time-averaged fields of the velocity field and the temperature field, respectively. The modal ordinal number takes values ​​from 1 to 1. ,in The total number of POD modes to be retained; and The velocity field and temperature field are respectively the first The time coefficients corresponding to the first mode; and The first and second fields are the velocity field and the temperature field, respectively. Orthogonal modes in the POD space; By applying Galerkin projection to the Navier-Stokes equations, the reduced-order ROM state equations are obtained: ; in Let the vector be a column vector composed of the time coefficients of each velocity mode, and the points on it be... This represents the first derivative with respect to time. This is the constant term vector, derived from the constant contributions in the Galerkin projection; This is a linear coefficient matrix, corresponding to the linear viscous dissipation term in the flow; For nonlinear convection tensors, corresponding to the nonlinear convection terms in the Navier-Stokes equations, the product... Represents a quadratic form; To control the input matrix; The control vector is a collection of control inputs such as fan speed and louver angle for each protection zone.

4. The system according to claim 3, characterized in that, The reduced-order model constructs multiple sets of POD orthogonal bases offline for different duct geometric deformation parameters, and generates pseudo-POD modes under the current deformation state through Grassmann manifold subspace interpolation mechanism: ； in These are the current geometric deformation parameters of the air duct; and The POD orthogonal basis matrices are pre-calculated and stored for two sets of typical geometric deformation parameters, which are considered as two points on the Grassmann manifold; These are the normalized interpolation coefficients, based on the current deformation. exist and The relative positions between corresponding shape variables are determined; Indicates will Mapped to Logarithmic mapping of the tangent space; Indicates from The tangent space is returned to the exponential mapping of the Grassmann manifold; These are pseudo-POD modes generated online, used to update the basis functions of the reduced-order model under the current geometric state.

5. The system according to claim 1, characterized in that, The fluid-structure interaction displacement estimation module constructs the structural dynamics state-space equations based on the mass-damping-stiffness model: ； in This is the relative displacement vector of the air duct structure relative to the vehicle body base; It is a relative velocity vector; M is the relative acceleration vector; C is the mass matrix of the structural system; K is the damping matrix; and K is the stiffness matrix. The base excitation acceleration vector is obtained from measurements taken by an IMU mounted on a rigid chassis; the augmented state vector is defined. ,in To achieve dynamic zero bias of the accelerometer, displacement estimates are obtained through prediction and update steps. The estimated displacement vector is then mapped to the duct geometric deformation parameters and input into a reduced-order digital twin model to achieve online compensation for vibration-induced fluid-structure interaction effects.

6. The control method according to claim 2, characterized in that, The hotspot prediction model employs a hybrid model combining a physically constrained long short-term memory network and an extreme gradient boosting tree, wherein: Dynamic internal resistance, Joule heat integral, and entropy change thermal fluctuation coefficient were calculated using the Bernardi heat generation equation and equivalent circuit model. The first and second derivatives of temperature rise were extracted from the temperature sequence. Simultaneously, local wind speed, duct deformation characteristics, and accumulated vibration energy were read from the reduced-order model output to construct a physical constraint feature matrix. ； Where t represents the current time step; This refers to the battery's dynamic internal resistance. For Joule heat integral, representing the accumulated irreversible heat; and These are the first and second derivatives of the temperature rise extracted from the temperature sequence, respectively. The local wind speed is output from the ROM; This refers to the deformation characteristics of the air duct; To accumulate vibrational energy; Time-series encoding of the feature matrix is ​​performed using a Long Short-Term Memory (LSTM) network. ； Where L is the time window length of the LSTM; The input is the feature matrix for L consecutive time steps; The hidden state vector output by the LSTM network represents the dynamic law of heat accumulation and decay. The hidden state vector is concatenated with the cell's spatial coordinates and local operating condition features. An extreme gradient boosting tree regression model is then used to predict the temperature of each spatial node within a preset time window. ； in For the predicted future Temperature or hotspot intensity value after minutes; M is the total number of regression trees in the XGBoost model; Let m be the function of the m-th regression tree, belonging to the regression tree function space. ; This represents the spatial coordinates of the battery cell and the characteristic vector of its local operating conditions. The prediction results are then decomposed for interpretability using the SHAP value: ； in The baseline output is d; d represents the total dimension of the input features. For the j-th input feature, the current predicted value The Shapley contribution value, whose magnitude and sign represent the marginal impact of this feature on hotspot formation.

7. The system according to claim 1, characterized in that, The multi-zone variable frequency fan array uses EC fans, which achieve stepless speed regulation from 0 to 100% using pulse width modulation. Each fan in each zone independently receives the speed setting value sent by the edge computing controller, achieving decoupled control between different zones and allocating different cooling capacities according to their respective heat load states.

8. The system according to claim 1, characterized in that, The active air guide louvers are made of high-strength metal or composite materials and are driven by a servo motor to adjust the angle of attack. The angle of attack adjustment range includes positive angle of attack, zero angle of attack and negative angle of attack, so as to realize pressurized air supply, uniform distribution or flow restriction for a certain protection zone according to different working conditions. The servo motor has a built-in current detection module, which is used to issue an alarm signal and trigger the flow field compensation strategy when the air guide louvers are mechanically stuck.

9. The system or method according to claim 1 or 2, characterized in that, The dynamic spoiler is attached to the duct wall under normal operating conditions to reduce wind resistance. When the controller issues an activation command, it pops out to a set angle of attack through a micro push rod or shape memory alloy drive mechanism, inducing eddy current shedding and secondary flow in the battery module cooling channel to improve the local convective heat transfer coefficient. After the danger is cleared, it automatically retracts to the initial position to reduce wind resistance and noise.

10. The system or method according to claim 1 or 2, characterized in that, The model predictive control module sets three control modes during execution: steady-state cruise, predictive feedforward cold storage, and emergency dynamic reconfiguration. When the predicted risk value of hotspots or the temperature rise slope exceeds a preset threshold, it automatically switches from steady-state cruise mode to predictive cold storage mode. When measured temperature, wind speed, and differential pressure indicate severe blockage of local flow channels and increased hotspot risk, it automatically switches from steady-state cruise mode to emergency dynamic reconfiguration mode. When the edge computing controller detects an anomaly in the digital twin simulation engine or upper-layer software, it automatically enters safe return mode, running the fans in each defense zone to a preset safe speed and resetting the air guide louvers to an evenly distributed state to ensure that basic heat dissipation capacity and safety margin can still be maintained in the event of a control system failure.

Images