A power distribution network battery digital dynamic management system based on digital twinning

Abstract

The application relates to the technical field of smart power grids, in particular to a power distribution network battery digital dynamic management system based on digital twinning. The system comprises a data acquisition unit, a digital twinning modeling unit, a dynamic optimization unit and an execution feedback unit. The digital twinning modeling unit is used for constructing a multi-dimensional dynamic twin of a battery-power grid-environment, and realizes virtual-real bidirectional mapping and self-adaptive updating by combining a multi-physical field coupling model and a long short-term memory network time series prediction algorithm. The data acquisition unit is a distributed heterogeneous sensor network, which collects multi-dimensional operation data of battery groups and key nodes of a power distribution network, and generates a high-fidelity data set containing four-dimensional labels of battery states, power grid parameters, time and position by combining space-time feature extraction technology, so that deep fusion of battery full life cycle states and power distribution network global operation data is realized, and comprehensive data support covering the whole world is provided for optimization decision-making.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, and more specifically, to a digital dynamic management system for distribution network batteries based on digital twins. Background Technology

[0002] In power distribution networks, the dynamic management of battery energy storage systems directly affects grid stability and energy utilization efficiency. With the increasing penetration rate of distributed energy, traditional battery management systems, due to limitations in static parameter configuration and local data acquisition, struggle to cope with complex load fluctuations and the collaborative needs of multiple battery banks.

[0003] In existing technologies, some solutions attempt to improve battery management capabilities through algorithm or hardware design optimization. For example, Chinese patent CN202410828794.4 discloses a battery energy storage system allocation method to improve the carrying capacity of unbalanced power distribution networks. It achieves power balancing and voltage regulation of multiple battery packs through a robust optimization module, and its outer loop row and column generation algorithm and inner loop alternating optimization process algorithm effectively address the uncertainties of the power distribution network. However, this solution relies on a pre-set integer resource variable model and does not establish a dynamic mapping relationship between the battery pack and the power distribution network topology and load distribution. This results in a lack of a global perspective in optimization decisions and fails to consider the impact of dynamic factors such as battery aging and environmental temperature changes on system performance. Another example is Chinese patent CN202510373032.4, which discloses a battery pack dynamic management and self-cleaning switching system and a battery pack dynamic management method. It solves the contact reliability problem in battery pack switching through a multi-power selective connection device and achieves real-time monitoring and fault isolation. However, the system focuses on the hardware switching logic inside the battery pack and does not deeply integrate with external data such as power flow calculation and load forecasting of the distribution network, making it difficult to achieve coordinated scheduling and dynamic optimization of multiple battery packs at the grid level.

[0004] While the aforementioned technical solutions possess corresponding design advantages, they also suffer from the following technical shortcomings: First, data interaction is limited to within the battery pack or local networks, failing to fully integrate with external data such as the distribution network topology, load distribution, and renewable energy output, resulting in optimization decisions that cannot adapt to the overall operation requirements of the power grid. Second, the control strategy is based on a fixed parameter model, lacking real-time response capabilities to dynamic factors such as battery aging and environmental temperature changes, thus hindering accurate prediction and adaptive adjustment of battery status. Third, it lacks the ability to assess the health status and predict remaining lifespan throughout the battery's entire lifecycle, making it difficult to support preventative maintenance and optimal resource allocation. Therefore, we propose a digital dynamic management system for distribution network batteries based on digital twins. Summary of the Invention

[0005] The purpose of this invention is to provide a digital dynamic management system for batteries in a power distribution network based on digital twins, so as to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, the present invention aims to provide a digital dynamic management system for batteries in a power distribution network based on digital twins, comprising:

[0007] The data acquisition unit is used to achieve deep perception of the battery pack's full life cycle status and the power distribution network's full-domain operation data, and to build a high-fidelity dataset based on distributed heterogeneous sensor networks and spatiotemporal feature extraction technology.

[0008] The digital twin modeling unit is used to construct a multi-dimensional dynamic twin of battery-grid-environment, and combines a multi-physics coupling model with a long short-term memory network time-series prediction algorithm to achieve bidirectional mapping between virtual and real and adaptive updates.

[0009] The dynamic optimization unit is used to achieve global collaborative optimization of multiple battery packs and power distribution network. It integrates real-time status data of twins with an improved deep reinforcement learning-multi-objective optimization fusion algorithm to generate a globally optimal strategy that dynamically adapts to operating conditions.

[0010] The execution feedback unit is used to achieve accurate execution and dynamic correction of the optimization strategy. It constructs a closed-loop correction link between the twin model and the physical system through instruction parsing technology and real-time status feedback mechanism.

[0011] As a further improvement to this technical solution, the data acquisition unit includes a distributed sensing module, a data transmission module, and a spatiotemporal feature extraction module, wherein:

[0012] The distributed sensing module is deployed at key nodes of the battery pack and power distribution network to collect multi-dimensional operational data.

[0013] Furthermore, the distributed sensing module includes an electrochemical sensor array deployed in the battery pack and an electrical parameter sensor array deployed in the power distribution network. The two arrays are independently networked using different communication protocols; wherein:

[0014] Electrochemical sensor arrays are deployed in key locations within the battery pack, including:

[0015] The voltage monitoring sensor employs a differential amplifier circuit design, and its measurement range covers the normal operating range of the battery.

[0016] The temperature sensor, built on digital temperature measurement technology, supports multi-point temperature measurement;

[0017] The current sensor uses the Hall effect principle and its range covers the battery charging and discharging current range.

[0018] Power parameter sensor arrays are deployed at key nodes in the power distribution network, including:

[0019] Smart meters support standard power communication protocols, and their data acquisition frequency meets the system monitoring requirements.

[0020] The feeder terminal is equipped with fault detection and location functions to monitor the line operating status in real time.

[0021] Load characteristic analyzers are deployed at critical load nodes to analyze load types and trends.

[0022] The data transmission module adopts a hybrid communication architecture to transmit the data collected by the distributed sensing module to the cloud.

[0023] Furthermore, the data transmission module includes an edge computing node and a hybrid communication network. The edge computing node is used to preprocess the raw data, and the hybrid communication network includes wireless transmission links and wired transmission links; wherein:

[0024] Edge computing nodes:

[0025] It adopts an industrial-grade embedded hardware platform and is equipped with a processor, memory, and storage unit;

[0026] Deploy a lightweight operating system to run containerized applications and implement data caching and preprocessing functions;

[0027] It supports parallel processing of multi-sensor data, and its throughput meets the system's real-time requirements.

[0028] Hybrid communication networks:

[0029] The battery-side sensor uses low-power wide-area wireless communication technology, and its operating frequency band meets regional regulatory requirements.

[0030] The grid-side equipment is connected via industrial Ethernet and uses standard power communication protocols to achieve data aggregation.

[0031] Edge nodes and the cloud use a secure communication protocol and are configured with data encryption and retransmission mechanisms.

[0032] The spatiotemporal feature extraction module generates a spatiotemporal index dataset based on a time series alignment algorithm and a spatial association mapping algorithm.

[0033] As a further improvement to this technical solution, the spatiotemporal feature extraction module includes a time series alignment submodule, a spatial correlation mapping submodule, and a spatiotemporal index generation submodule, wherein:

[0034] The time series alignment submodule is used to perform time synchronization processing on the multi-frequency data collected by the distributed sensing module. It adopts a sliding time window algorithm to achieve data alignment, and the window size is dynamically adjusted according to the data fluctuation characteristics.

[0035] The spatial association mapping submodule constructs a graph neural network model based on the power distribution network topology. The input layer of the graph neural network model contains node voltage, current, and power feature vectors, and the number of hidden layer nodes is dynamically configured according to the network size.

[0036] The spatiotemporal index generation submodule is used to fuse time-aligned data with a spatial association model to generate a spatiotemporal index dataset containing four-dimensional labels: battery status, power grid parameters, time, and location.

[0037] As a further improvement to this technical solution, the digital twin modeling unit includes a multiphysics coupling modeling module and an LSTM behavior prediction module, wherein:

[0038] The multiphysics coupling modeling module is used to construct a coupling model of the battery electrochemical field, the power grid electromagnetic field, and the ambient temperature field. The coupling model defines the interaction relationship between different physical fields through the field parameter correlation matrix.

[0039] The LSTM behavior prediction module predicts trends in battery charging and discharging characteristics, power grid flow changes, and environmental parameter evolution based on a long short-term memory network, including the following steps:

[0040] S220.1 Input Layer Feature Processing: Standardize historical operational data and real-time sensing data to generate a sequence of feature vectors. ,in For time steps;

[0041] S220.2, Gating Unit Calculation:

[0042] Calculate the forget gate separately Input gate Candidate cell status Cell state update Output gate and hidden state output ;in:

[0043] Forgotten Gate: Used to filter historical information that needs to be retained;

[0044] Input Gate: This is used to control the proportion of new information included.

[0045] Candidate cell status: This is used to store the feature information of the current input;

[0046] Cell status update: Integrating historical and current information;

[0047] Output gate: , used to control the filtering of output information;

[0048] Hidden output: Generate the feature state of the current time step;

[0049] S220.3, Prediction Output: ,get The predicted value at any given time;

[0050] in, This refers to the weight matrix corresponding to the gating and output layers; For bias terms; Use the Sigmoid activation function; It is the hyperbolic tangent activation function; for The hidden state at any given moment; for The state of a cell at any given moment.

[0051] Furthermore, the weight matrix Initialize the hidden layer to random values ​​following a normal distribution, and initialize the bias term to 0; adjust the hidden layer dimension according to the prediction object (e.g., set 64 dimensions for battery state of charge prediction and 128 dimensions for power grid power flow prediction). Simultaneously, in battery charge and discharge characteristic prediction, a forget gate is used. Historical data that has been idle for a long time is given low weight (such as charge and discharge records that have been idle for more than 24 hours); input gate High weighting is given to real-time temperature changes; cell state Dynamic storage battery capacity degradation trend; output gate The output sensitivity is adjusted based on the current state of charge (e.g., the output gate threshold is automatically increased when the state of charge is close to full). In addition, after the LSTM behavior prediction module outputs the battery temperature prediction sequence for the next hour, the multiphysics coupling modeling module uses this sequence as the input boundary condition of the electrochemical field sub-model, calculates the internal temperature gradient of the battery through the heat conduction equation, and then corrects the charge transfer coefficient in the electrode reaction kinetic equation to achieve a closed-loop linkage of "prediction-simulation".

[0052] As a further improvement to this technical solution, the multiphysics coupling of the multiphysics coupling modeling module specifically includes:

[0053] Coupling of battery electrochemical field and ambient temperature field: The heat power generated by battery charging and discharging (calculated based on the current prediction value output by the LSTM behavior prediction module) is used as the input of ambient temperature field, and the temperature distribution output by ambient temperature field inversely corrects the reaction rate coefficient of battery electrochemical field.

[0054] Coupling of the power grid electromagnetic field and the ambient temperature field: The power loss of the power grid line (calculated based on line current and resistance) is converted into a heat source of the ambient temperature field. The conductor temperature output by the ambient temperature field is used to correct the line resistance parameters of the power grid electromagnetic field.

[0055] Coupling of battery electrochemical field and grid electromagnetic field: A correlation is established by the balance between battery charging and discharging power and grid node power. Changes in battery output power trigger the recalculation of grid electromagnetic field flow.

[0056] As a further improvement to this technical solution, the digital twin modeling unit combines a multiphysics coupling model with a long short-term memory network time-series prediction algorithm to achieve bidirectional virtual-real mapping and adaptive updating, including the following steps:

[0057] S230.1 Initialization of virtual-real mapping relationship: Based on the high-fidelity dataset output by the data acquisition unit, establish parameter mapping rules between physical entities and virtual twins. The mapping rules include a one-to-one correspondence between battery cell voltage and virtual electrode potential, grid node current and virtual line power flow, and ambient temperature and virtual field temperature, and store them in the mapping rule library.

[0058] S230.2 Prediction Sequence Generation and Input: The LSTM behavior prediction module generates a state prediction sequence for a future preset time period (30-60 minutes) based on historical operation data and real-time sensing data. The prediction sequence includes time-series data of battery charging and discharging current, grid node power and ambient temperature, and is transmitted to the multiphysics coupling modeling module through a standardized interface.

[0059] S230.3 Multiphysics Co-simulation: The multiphysics coupled modeling module takes the predicted sequence as the boundary condition input and starts the co-simulation of the battery electrochemical field, the power grid electromagnetic field and the ambient temperature field. The simulation step size is consistent with the time granularity of the LSTM predicted sequence (e.g., 5-10 minutes / step). The output is the state response results of the virtual twin (including battery state of charge, power grid voltage distribution and field temperature gradient).

[0060] S230.4, Virtual-Real Deviation Calculation: Real-time running data (with the same timestamp as the virtual state response result) is extracted by the distributed sensing module of the data acquisition unit, and the parameter correspondence in the mapping rule base is called to align the data. The root mean square error formula is used to calculate the deviation value between the real-time running data and the virtual state response result. The deviation value corresponds to three types of indicators: battery status, grid parameters, and environmental parameters.

[0061] S230.5 Model Adaptive Update: When the deviation value of any type of index continuously exceeds the preset threshold, the update mechanism is triggered: the gating weight matrix of the LSTM behavior prediction module is adjusted (the input gate and forget gate are updated first), and the coupling coefficient in the field parameter correlation matrix of the multi-physics coupling modeling module is corrected. The update magnitude is positively correlated with the deviation value.

[0062] S230.6 Closed-loop verification and iteration: The updated model repeats steps S230.2-S230.4 until the deviation values ​​of all indicators return to within the threshold, forming a closed-loop iterative process of "prediction-simulation-verification-update".

[0063] As a further improvement to this technical solution, the dynamic optimization unit includes an algorithm fusion module, a strategy generation module, and a working condition adaptation module, wherein:

[0064] The algorithm fusion module integrates the real-time state data of the twin with an improved deep reinforcement learning-multi-objective optimization fusion algorithm. The improved deep reinforcement learning-multi-objective optimization fusion algorithm uses the battery state of charge, grid node voltage, and ambient temperature output by the twin as state inputs.

[0065] The strategy generation module generates charging and discharging scheduling strategies, load allocation strategies, and network topology adjustment strategies based on the calculation results of the algorithm fusion module. The strategies include dual constraints in terms of time dimension (scheduling cycle) and spatial dimension (execution node).

[0066] The operating condition adaptation module is used to monitor changes in the operating conditions of the power distribution network and battery pack (such as load fluctuations and battery aging). When the rate of change in operating conditions exceeds a preset threshold, the parameters of the algorithm fusion module are adaptively adjusted.

[0067] As a further improvement to this technical solution, the execution of the improved deep reinforcement learning-multi-objective optimization fusion algorithm includes the following steps:

[0068] S340.1, State Input and Feature Extraction:

[0069] The algorithm fusion module receives real-time state data from the twin and extracts the battery pack's state of charge sequence. Grid node voltage sequence and ambient temperature Construct state vector The formula is:

[0070] ;

[0071] The vector is then normalized (mapped to the [0,1] interval):

[0072] ;

[0073] ;

[0074] ;

[0075] in, The normalized state of charge. For the first Real-time state of charge of individual battery cells The minimum state of charge for safe battery operation. This represents the maximum state of charge for safe battery operation. The normalized voltage. For the first Real-time voltage of each grid node The minimum voltage required for the safe operation of a power grid node. The maximum voltage for safe operation of a power grid node; The minimum temperature required for safe system operation. The highest temperature for safe system operation. The normalized temperature;

[0076] S340.2, Reward Function Calculation:

[0077] Calculating multi-objective reward values ​​based on state vectors The formula is:

[0078] ;

[0079] in, For cost objectives, In the formula: , These are the grid cost weighting coefficient and the battery cost weighting coefficient, respectively. For power grid operating costs; For battery wear and tear costs; For efficiency goals, In the formula: The amount of energy to be used effectively This represents the total energy consumption of the system. For security purposes, In the formula: This represents the number of nodes in the power distribution network that participate in voltage monitoring. For the first The actual voltage of each grid node This is the minimum allowable voltage at the grid node. The actual temperature of the battery or the environment. The upper limit of the temperature range for safe battery operation; , , These are the weighting coefficients, and The weighting coefficients are dynamically configured by the working condition adaptation module;

[0080] S340.3 Multi-objective optimization solution: An improved NSGA-III algorithm is used to solve the decision variables. ( Battery charging and discharging power, For the switch state), optimization is performed, where Indicates power, Indicates the switch state, objective function for:

[0081] ;

[0082] in, To minimize grid operating costs; To minimize battery wear and tear costs; To maximize energy efficiency;

[0083] Constraints are introduced during the optimization process: charging and discharging power ,in This refers to the battery's rated power; adjacent nodes are open when their switch states are different.

[0084] S340.4 Optimal Solution Screening: From the set of non-dominated solutions obtained through optimization, the Analytic Hierarchy Process (AHP) is used to screen indicators (including cost reduction rate, efficiency improvement rate, and safety constraint satisfaction) and construct a judgment matrix. :

[0085] ;

[0086] And select the globally optimal solution that matches the current working condition;

[0087] S340.5, Strategy Generation and Output: Convert the optimal solution into structured strategy instructions (including execution node ID, power value, and timestamp), and output them to the execution feedback unit through the strategy generation module;

[0088] S340.6 Parameter Iterative Update: After every 100 executions of the strategy, compare the actual reward. With predicted rewards :

[0089] like The crossover probability of NSGA-Ⅲ is dynamically adjusted (from 0.6 to 0.9, with a step size of 0.1).

[0090] The weight coefficients are updated using gradient descent until the policy fit is >90%.

[0091] ;

[0092] ;

[0093] ;

[0094] in, , , These are the updated weighting coefficients; The learning rate; , , All are partial derivatives.

[0095] As a further improvement to this technical solution, the execution feedback unit includes an instruction parsing module, an execution driving module, a status feedback module, and a closed-loop correction module, wherein:

[0096] The instruction parsing module adopts a multi-protocol parsing engine, which is compatible with the power distribution terminal communication protocol and the battery BMS communication protocol. It performs format conversion and legality verification on the optimization strategy instructions output by the dynamic optimization unit to ensure that the instructions can be recognized and executed by physical devices.

[0097] The execution drive module generates physical execution signals based on the parsed instructions, drives the battery charging and discharging device and the power grid switching equipment to operate, and has an instruction timeout retransmission mechanism. When the equipment execution is abnormal, it automatically triggers a degradation execution strategy to maintain the basic operation of the system.

[0098] The status feedback module collects real-time operating status data of the battery charging and discharging device and the power grid switching equipment through the status monitoring interface of the battery charging and discharging device and the power grid switching equipment, including the actual charging and discharging power of the battery and the switching action status. Through the time synchronization mechanism, and based on the twin model status data output by the digital twin modeling unit, the virtual and real status are aligned.

[0099] Furthermore, the status data of the battery charging and discharging device includes actual output power, DC bus voltage, and converter operating status (normal / alarm). The status data of the grid switching equipment includes switch open / close position, number of operations, and operating coil status (energized / de-energized). The acquisition cycle is synchronized with the instruction issuance cycle of the execution drive module.

[0100] The closed-loop correction module is used to calculate the deviation between the real-time operating status data of the battery charging and discharging device and the grid switching equipment and the state of the twin model. When the deviation continues to exceed the set threshold, the twin model parameter correction process is triggered, and the dynamic optimization unit is driven to update the optimization strategy to build a closed-loop correction link between the twin model and the physical system.

[0101] As a further improvement to this technical solution, the closed-loop correction module calculates the deviation between the real-time operating status data of the battery charging and discharging device and the grid switching equipment and the state of the twin model, including the following steps:

[0102] S440.1, Deviation Classification Calculation:

[0103] The physical execution status data (actual output power of the battery charging and discharging device) collected by the status feedback module is received. Opening and closing positions of power grid switchgear ), and the twin model state data (virtual charging and discharging power) output by the digital twin modeling unit. Virtual switch position ), calculate the following two types of deviations respectively:

[0104] Power deviation: ;in, The actual output power of the battery charging and discharging device; Virtual charging and discharging power;

[0105] Switch status deviation: And the fractional part is 1 and the composite part is 0; among them, This refers to the open / closed position of the power grid switching equipment; For virtual switch positions;

[0106] S440.2, Calibration Trigger Judgment:

[0107] When power deviation The power of the battery charging and discharging device exceeded the rated power for three consecutive data acquisition cycles. 5%, or switch state deviation When two consecutive cycles are 1 (i.e., the actual and virtual states are inconsistent), the correction mechanism is triggered, and the duration of the deviation and the cumulative deviation value are recorded at the same time.

[0108] S440.3, Twin model parameter correction:

[0109] For power deviation: Correct the reaction rate coefficient of the battery electrochemical field in the multiphysics coupling modeling module. The correction formula is:

[0110] ;

[0111] in, This is a correction factor, with a value ranging from 0.8 to 1.2. This is the original reaction rate coefficient in the battery electrochemical field (corresponding to the reverse-corrected reaction rate coefficient of the ambient temperature field). This is the correction factor after power deviation correction;

[0112] To address switch state deviation: update the adjacency matrix of the graph neural network in the distribution network topology mapping, synchronously correct the coupling coefficient of the field parameter correlation matrix in the multi-physics coupling modeling module (corresponding to the coupling coefficient in the corrected field parameter correlation matrix), and adjust the line resistance parameters of the power grid electromagnetic field based on the conductor temperature output from the ambient temperature field until the virtual switch position is consistent with the actual position.

[0113] S440.4 Optimization Strategy Rolling Update: The corrected multiphysics coupling model parameters are input into the dynamic optimization unit, triggering a rolling window update of the optimization strategy. The window length is set to the current scheduling period plus two prediction steps (consistent with the time granularity of the LSTM prediction sequence, i.e., 5-10 minutes / step). The modified particle swarm optimization algorithm is used to resolve the optimal strategy, where the inertia weight... It decreases linearly as the cumulative deviation increases (initial value 0.9, minimum value 0.4).

[0114] S440.5, Verification of Correction Effect: After the new strategy is issued by the instruction parsing module, the state feedback module continuously collects physical execution state data and calls the co-simulation results of the multi-physics coupling modeling module for alignment verification (comparison with data of the same timestamp). If the conditions are met within two consecutive cycles... ,and If the error is not found, the correction process is terminated; otherwise, steps S400.3-S400.4 are repeated until the deviation returns to the threshold.

[0115] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0116] 1. This invention collects multi-dimensional operational data of battery packs and key nodes of the power distribution network through a distributed heterogeneous sensor network of the data acquisition unit. Combined with spatiotemporal feature extraction technology, it generates a high-fidelity dataset containing four-dimensional labels of battery status, power grid parameters, time, and location. This achieves deep integration of the battery's full life cycle status with the power distribution network's full-domain operational data, providing comprehensive data support covering the entire domain for optimization decisions, making the decisions more adaptable to the overall operation needs of the power grid.

[0117] 2. This invention utilizes a multi-dimensional dynamic twin constructed by a digital twin modeling unit, combined with a multi-physics coupling model and a long short-term memory network time-series prediction algorithm, to achieve real-time perception and trend prediction of battery status, grid operation, and environmental changes. The dynamic optimization unit integrates the real-time status data of the twin and generates a strategy that dynamically adapts to operating conditions through an improved deep reinforcement learning-multi-objective optimization fusion algorithm. The operating condition adaptation module can adjust parameters according to changes such as load fluctuations and battery aging. In conjunction with the closed-loop correction link of the execution feedback unit, it ensures that the control strategy can respond to dynamic factors in real time, achieving accurate prediction and adaptive adjustment of battery status.

[0118] 3. This invention enables continuous and in-depth perception of the battery's state throughout its entire lifecycle through a data acquisition unit, combined with the real-time mapping and evolution simulation of the battery state by a dynamic twin constructed by a digital twin modeling unit. This allows for the full lifecycle monitoring and evaluation of the battery's health status, providing a valid basis for the formulation of preventive maintenance plans and the optimal allocation of resources. Attached Figure Description

[0119] Figure 1 This is a schematic diagram of the system framework of the present invention;

[0120] The meanings of the labels in the diagram are as follows:

[0121] 100. Data Acquisition Unit; 110. Distributed Sensing Module; 120. Data Transmission Module; 130. Spatiotemporal Feature Extraction Module; 131. Time Series Alignment Submodule; 132. Spatial Association Mapping Submodule; 133. Spatiotemporal Index Generation Submodule;

[0122] 200. Digital Twin Modeling Unit; 210. Multiphysics Coupled Modeling Module; 220. LSTM Behavior Prediction Module;

[0123] 300. Dynamic optimization unit; 310. Algorithm fusion module; 320. Strategy generation module; 330. Working condition adaptation module;

[0124] 400. Execution feedback unit; 410. Instruction parsing module; 420. Execution drive module; 430. Status feedback module; 440. Closed-loop correction module. Detailed Implementation

[0125] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0126] like Figure 1As shown, this embodiment provides a digital twin-based dynamic management system for battery digitalization in a power distribution network. Through the organic collaboration of a data acquisition unit 100, a digital twin modeling unit 200, a dynamic optimization unit 300, and an execution feedback unit 400, it achieves digital dynamic management of the battery pack and power distribution network throughout their entire lifecycle. The overall operation process is as follows: The data acquisition unit 100 first collects multi-dimensional data such as voltage, current, and temperature at key nodes of the battery pack and power distribution network via a distributed sensing module 110. After edge preprocessing and hybrid communication by the data transmission module 120, the spatiotemporal feature extraction module 130 completes time alignment and spatial correlation mapping, generating a high-fidelity dataset containing four-dimensional labels: "battery state - grid parameters - time - location". The digital twin modeling unit 200 receives this dataset and constructs a coupled model of the battery electrochemical field, the grid electromagnetic field, and the ambient temperature field through a multi-physics coupling modeling module 210. The LSTM behavior prediction module 220 achieves bidirectional mapping between the virtual and real systems by predicting future states, and maintains synchronization between the twin and the physical system through virtual-real deviation calculation and adaptive parameter updates. The dynamic optimization unit 300 takes the real-time state and prediction data of the twin as input, integrates deep reinforcement learning and multi-objective optimization algorithms through the algorithm fusion module 310 to solve for the optimal solution, converts it into structured instructions by the policy generation module 320, and then dynamically adjusts parameters according to load fluctuations, battery aging, etc., through the operating condition adaptation module 330. The instruction parsing module 410 of the execution feedback unit 400 performs protocol conversion and legality verification on the instructions, generates physical signals through the execution drive module 420 to drive equipment actions, the state feedback module 430 collects the equipment state in real time and aligns it with the twin model, and the closed-loop correction module 440 calculates the deviation. If the deviation exceeds a threshold, it triggers model parameter correction and policy updates by the dynamic optimization unit 300, forming an "execution-feedback-correction" closed loop. Specifically, this includes:

[0127] The data acquisition unit 100 is used to achieve deep perception of the battery pack's full life cycle status and the power distribution network's full domain operation data, and to build a high-fidelity dataset based on distributed heterogeneous sensor networks and spatiotemporal feature extraction technology.

[0128] In this embodiment, the data acquisition unit 100 includes a distributed sensing module 110, a data transmission module 120, and a spatiotemporal feature extraction module 130, wherein:

[0129] The distributed sensing module 110 is deployed at key nodes of the battery pack and power distribution network to collect multi-dimensional operational data;

[0130] Furthermore, the distributed sensing module 110 includes an electrochemical sensor array deployed in the battery pack and an electrical parameter sensor array deployed in the power distribution network. The two arrays are independently networked using different communication protocols; wherein:

[0131] Electrochemical sensor arrays are deployed in key locations within the battery pack, including:

[0132] The voltage monitoring sensor employs a differential amplifier circuit design, and its measurement range covers the normal operating range of the battery.

[0133] The temperature sensor, built on digital temperature measurement technology, supports multi-point temperature measurement;

[0134] The current sensor uses the Hall effect principle and its range covers the battery charging and discharging current range.

[0135] Power parameter sensor arrays are deployed at key nodes in the power distribution network, including:

[0136] Smart meters support standard power communication protocols, and their data acquisition frequency meets the system monitoring requirements.

[0137] The feeder terminal is equipped with fault detection and location functions to monitor the line operating status in real time.

[0138] Load characteristic analyzers are deployed at critical load nodes to analyze load types and trends.

[0139] As a further explanation of this embodiment, the spatiotemporal synchronization mechanism between the electrochemical sensor array and the power parameter sensor array in this embodiment is as follows: a unified time reference is configured for all sensors through a BeiDou or GPS timing module, and the sampling trigger signal adopts a hardware synchronization pulse (the pulse interval is dynamically set according to the acquisition frequency, such as 10ms / time for battery voltage sampling and 1s / time for grid load sampling). When multiple sensors are deployed at the same physical node (such as the access point of a battery energy storage station), a data fusion gateway is used for local data aggregation to avoid signal interference (the gateway is equipped with an anti-electromagnetic interference filter, and the operating frequency band avoids the strong electromagnetic radiation range of the power distribution network).

[0140] As a further explanation of this embodiment, the voltage monitoring sensor in this embodiment compares with a standard voltage source daily. If the deviation exceeds a safety threshold (e.g., ±2%), it automatically records and triggers a calibration command (achieved by adjusting the gain resistor of the differential amplifier circuit). The temperature sensor performs zero-point drift compensation every hour (based on the reference temperature of the ambient constant temperature zone). The feeder terminal performs analog input circuit detection every 12 hours. If the fault detection function determines that the line is abnormal (e.g., the current transformer is disconnected), it actively reports the fault code through the data transmission module.

[0141] The data transmission module 120 adopts a hybrid communication architecture to transmit the data collected by the distributed sensing module 110 to the cloud.

[0142] Furthermore, the data transmission module 120 includes an edge computing node and a hybrid communication network. The edge computing node is used to preprocess the raw data, and the hybrid communication network includes wireless transmission links and wired transmission links; wherein:

[0143] Edge computing nodes:

[0144] It adopts an industrial-grade embedded hardware platform and is equipped with a processor, memory, and storage unit;

[0145] Deploy a lightweight operating system to run containerized applications and implement data caching and preprocessing functions;

[0146] It supports parallel processing of multi-sensor data, and its throughput meets the system's real-time requirements.

[0147] Hybrid communication networks:

[0148] The battery-side sensor uses low-power wide-area wireless communication technology, and its operating frequency band meets regional regulatory requirements.

[0149] The grid-side equipment is connected via industrial Ethernet and uses standard power communication protocols to achieve data aggregation.

[0150] Edge nodes and the cloud use a secure communication protocol and are configured with data encryption and retransmission mechanisms.

[0151] As a further explanation of this embodiment, the preprocessing steps of the edge computing nodes for the original data in this embodiment include:

[0152] Outlier filtering: The 3σ principle is used to identify data that exceeds the normal range (such as battery voltage exceeding 1.5 times the rated voltage range), outliers are marked and the original records are retained (for subsequent fault tracing).

[0153] Data dimensionality reduction: For high-frequency load curves collected by load characteristic analyzers, a piecewise linear fitting algorithm is used to extract feature points (such as peak values, valley values, and rates of change) to reduce data transmission volume;

[0154] Timestamp Correction: For timestamp deviations caused by transmission delays, a linear correction is performed based on the local clock of the edge node. The correction formula is as follows: ,in The mean transmission delay is obtained through historical communication delay statistics.

[0155] As a further explanation of this embodiment, the dynamic adaptation mechanism of the communication link in this embodiment is as follows:

[0156] The battery-side wireless transmission link monitors the signal strength (RSSI) in real time. When the RSSI is below -90dBm for 5 consecutive cycles, it automatically switches to the backup frequency band and triggers the local buffering of the edge node.

[0157] The power grid-side industrial Ethernet adopts a redundant topology design. When the main link is interrupted, it automatically switches to the backup link. During the switching process, the data is temporarily stored in the local storage of the feeder terminal (the storage medium is an industrial-grade SD card, which supports power failure data protection).

[0158] Communication bandwidth scheduling between edge nodes and the cloud: bandwidth is dynamically allocated based on data priority (battery failure data has the highest priority and occupies ≥50% of the bandwidth; normal status data has the next highest priority and bandwidth is allocated on demand).

[0159] The spatiotemporal feature extraction module 130 generates a spatiotemporal index dataset based on the time series alignment algorithm and the spatial association mapping algorithm.

[0160] In this embodiment, the spatiotemporal feature extraction module 130 includes a time series alignment submodule 131, a spatial association mapping submodule 132, and a spatiotemporal index generation submodule 133, wherein:

[0161] The time series alignment submodule 131 is used to perform time synchronization processing on the multi-frequency data collected by the distributed sensing module 110. It uses a sliding time window algorithm to achieve data alignment, and the window size is dynamically adjusted according to the data fluctuation characteristics.

[0162] As a further explanation of this embodiment, the time series alignment submodule 131 in this embodiment achieves time synchronization of multi-frequency data by dynamically adjusting the sliding window, laying a consistent time benchmark for subsequent spatial correlation analysis. Specifically, it includes the following steps:

[0163] First, the time series alignment submodule 131 receives multi-frequency data (such as 100Hz voltage data from the battery side and 10Hz power data from the grid side) collected by the distributed sensing module 110. These data have time axis misalignment due to differences in sampling frequency, requiring fluctuation characteristic analysis to determine the basis for window adjustment. Next, the fluctuation coefficient of the data sequence is calculated. ,in Standard deviation The mean is used to quantify data stability. The larger the value, the more drastic the data fluctuation (such as the current during battery charging and discharging).

[0164] Subsequently, based on Dynamically adjust the size of the sliding time window: when When the value is ≥0.5 (high volatility data), the window is set to 100ms; when 0.1 < When the fluctuation value is <0.5 (medium fluctuation data), the window is set to 500ms; when When the value is ≤0.1 (low fluctuation data), the window is set to 1000ms to adapt to the changing characteristics of different data.

[0165] Then, a baseline time axis is constructed using the adjusted window as the time granularity. ,in, The start time, Define the window size and apply it to each raw data point. Calculate relative to the most recent reference time point offset ;in, No. The timestamp of each raw data point;

[0166] Next, regarding The data is directly incorporated into the corresponding baseline time window; The data was completed using linear interpolation, with the following formula: ;in, For the first The values ​​of the original data points, For the first The values ​​of the original data points, To complete the time point and the first The time difference between the original data points For the first The timestamp of each raw data point For the first Timestamps of each original data point; ensure all data is mapped to a unified timeline;

[0167] Finally, output the time-aligned time series dataset. This provides a time-consistent input for subsequent spatial correlation analysis.

[0168] The spatial association mapping submodule 132 constructs a graph neural network model based on the power distribution network topology. The input layer of the graph neural network model contains node voltage, current, and power feature vectors, and the number of hidden layer nodes is dynamically configured according to the network size.

[0169] As a further explanation of this embodiment, the spatial association mapping submodule 132 in this embodiment mines the spatial association of nodes by constructing a graph neural network model based on time-aligned data, providing a spatial dimension of association basis for the generation of spatiotemporal indexes, specifically including the following steps:

[0170] First, the spatial association mapping submodule 132 receives the output from the time series alignment submodule 131. A graph neural network model is constructed based on the power distribution network topology;

[0171] Subsequently, the input layer features of the graph neural network are defined: the input vector of each node contains voltage, current, and power features (taken from...). (corresponding data within the same time window), ensuring that the input data remains synchronized with the timeline;

[0172] Then, the number of hidden layer nodes is dynamically configured according to the network size: when the number of distribution network nodes... When the hidden layer has two layers, the number of nodes in each layer is 1.5 times that of the input layer; when At that time, there are 3 hidden layers, and the number of nodes in each layer is twice that of the input layer. At the same time, a Dropout layer (dropout rate = 0.2) is introduced to suppress overfitting.

[0173] Next, feature aggregation is performed using a graph neural network: the features of the l-th layer are calculated as follows: ,in To normalize the adjacency matrix (adjacency matrix based on network topology) (processed) For the first Layer weight matrix, The ReLU activation function is used to fuse node and neighbor features. Then, the spatial association weights between nodes are output through a fully connected layer. ,in This represents the total number of floors. The larger the value, the more likely the node is to be larger. and The closer the spatial connections;

[0174] Finally, output the spatial correlation model. This model includes the correlation strength between all nodes, providing a spatial dimension basis for spatiotemporal fusion.

[0175] As a further explanation of this embodiment, the spatiotemporal index generation submodule 133 in this embodiment is used to fuse the time-aligned data with the spatial association model to generate a spatiotemporal index dataset containing four-dimensional labels of battery status, power grid parameters, time, and location.

[0176] First, the spatiotemporal index generation submodule 133 receives the output from the time series alignment submodule 131. Output of spatial association mapping submodule 132 Initiate the data fusion process;

[0177] Then, with timestamps and location number It is a double primary key, and the relationship is... middle and the state data at time position p (such as battery SOC, node voltage) and Middle position Association weight with other nodes This ensures a one-to-one correspondence between time and spatial data;

[0178] Then, a four-dimensional tagging system was constructed: "Battery Status - Grid Parameters - Time - Location" tags were added to each piece of fused data. "Battery Status" includes features such as SOC and temperature; "Grid Parameters" includes indicators such as voltage and power; and "Time" is... The base timestamp, "location" is the unique identifier of the node. ;

[0179] Next, a structured index table is generated: the index table uses "timestamp + location number" as the primary key to store the specific data of the corresponding tag, and supports fast query by time range or location area;

[0180] Finally, a spatiotemporal index dataset is output, which retains the dynamic changes of the time series and integrates the topological relationships of spatial correlation. It is transmitted to the digital twin modeling unit 200 through a standardized interface, providing high-fidelity spatiotemporal data support for building a multi-dimensional dynamic twin of battery-grid-environment, and achieving seamless connection between time and space dimensions.

[0181] The digital twin modeling unit 200 is used to construct a multi-dimensional dynamic twin of battery-grid-environment, and combines a multi-physics coupling model with a long short-term memory network time-series prediction algorithm to achieve virtual-real bidirectional mapping and adaptive updates.

[0182] In this embodiment, the digital twin modeling unit 200 includes a multiphysics coupling modeling module 210 and an LSTM behavior prediction module 220, wherein:

[0183] The multiphysics coupling modeling module 210 is used to construct a coupling model of the battery electrochemical field, the power grid electromagnetic field and the ambient temperature field. The coupling model defines the interaction relationship between different physical fields through the field parameter correlation matrix.

[0184] The LSTM behavior prediction module 220 uses a long short-term memory network to predict trends in battery charging and discharging characteristics, power grid flow changes, and environmental parameter evolution, including the following steps:

[0185] S220.1 Input Layer Feature Processing: Standardize historical operational data and real-time sensing data to generate a sequence of feature vectors. ,in For time steps;

[0186] As a further explanation of this embodiment, in the input layer feature processing of this step, historical operating data and real-time sensing data (such as battery current, grid power, and ambient temperature) are standardized using the following formula: ,in The original data, The feature mean is calculated based on a 24-hour sliding window. The characteristic standard deviation, This is used to avoid a denominator of 0, and the standardized result generates a sequence of feature vectors. ,in The time step is defined. Simultaneously, input features are dynamically selected based on the prediction target, and the Pearson correlation coefficient is used as the reference. Filtering weakly correlated features ( (These features are removed), for example, battery state of charge (SOC) prediction needs to include current, temperature, and historical SOC, while grid power flow prediction needs to include node voltage and load power.

[0187] S220.2, Gating Unit Calculation:

[0188] Calculate the forget gate separately Input gate Candidate cell status Cell state update Output gate and hidden state output ;in:

[0189] Forgotten Gate: Used to filter historical information that needs to be retained;

[0190] Input Gate: This is used to control the proportion of new information included.

[0191] Candidate cell status: This is used to store the feature information of the current input;

[0192] Cell status update: Integrating historical and current information;

[0193] Output gate: , used to control the filtering of output information;

[0194] Hidden output: Generate the feature state of the current time step;

[0195] S220.3, Prediction Output: ,get The predicted value at any given time;

[0196] in, This refers to the weight matrix corresponding to the gating and output layers; For bias terms; Use the Sigmoid activation function; It is the hyperbolic tangent activation function; for The hidden state at any given moment; for The state of a cell at any given moment.

[0197] Furthermore, the weight matrix Initialize the hidden layer to random values ​​following a normal distribution, and initialize the bias term to 0; adjust the hidden layer dimension according to the prediction object (e.g., set 64 dimensions for battery state of charge prediction and 128 dimensions for power grid power flow prediction). Simultaneously, in battery charge and discharge characteristic prediction, a forget gate is used. Historical data that has been idle for a long time is given low weight (such as charge and discharge records that have been idle for more than 24 hours); input gate High weighting is given to real-time temperature changes; cell state Dynamic storage battery capacity degradation trend; output gate The output sensitivity is adjusted according to the current state of charge (e.g., the output gate threshold is automatically increased when the state of charge is close to full charge). In addition, after the LSTM behavior prediction module 220 outputs the battery temperature prediction sequence for the next hour, the multiphysics coupling modeling module uses this sequence as the input boundary condition of the electrochemical field sub-model, calculates the internal temperature gradient of the battery through the heat conduction equation, and then corrects the charge transfer coefficient in the electrode reaction kinetic equation to achieve a closed-loop linkage of "prediction-simulation".

[0198] As a further explanation of this embodiment, the prediction output stage in this step requires physical constraint correction, such as clamping the predicted battery SOC value to the [0,1] interval and limiting the grid voltage to [0,1]. Inside( (Rated voltage). Output temperature prediction sequence for the next 1 hour. ( The predicted temperature at time (the time of the prediction) will be used as the boundary condition for the electrochemical field sub-model, through the heat conduction equation. Calculate the internal temperature gradient of the battery; where, The rate of change of temperature over time. Where is the thermal diffusivity, For thermal power, For volumetric heat capacity, This is the Laplace operator for temperature. This further modifies the charge transfer coefficient in the electrode reaction kinetic equation. This achieves a closed-loop linkage between prediction and simulation; among which, For activation energy, Boltzmann's constant, The reference charge transfer coefficient, This refers to absolute temperature.

[0199] In this embodiment, the multiphysics coupling of the multiphysics coupling modeling module 210 specifically includes:

[0200] Coupling of battery electrochemical field and ambient temperature field: The heat power generated by battery charging and discharging (calculated based on the current prediction value output by LSTM behavior prediction module 220) is used as the input of ambient temperature field, and the temperature distribution output by ambient temperature field inversely corrects the reaction rate coefficient of battery electrochemical field.

[0201] Coupling of the power grid electromagnetic field and the ambient temperature field: The power loss of the power grid line (calculated based on line current and resistance) is converted into a heat source of the ambient temperature field. The conductor temperature output by the ambient temperature field is used to correct the line resistance parameters of the power grid electromagnetic field.

[0202] Coupling of battery electrochemical field and grid electromagnetic field: A correlation is established by the balance between battery charging and discharging power and grid node power. Changes in battery output power trigger the recalculation of grid electromagnetic field flow.

[0203] As a further explanation of this embodiment, the multiphysics coupling modeling module 210 in this embodiment constructs a dynamic coupling model of the battery electrochemical field, the power grid electromagnetic field, and the ambient temperature field by quantifying the interaction relationships between fields, providing physical simulation support for the virtual twin. The specific relationship of the three-field coupling is as follows:

[0204] The heat power generated during battery charging and discharging in the coupling of the battery's electrochemical field and the ambient temperature field. As the input to the ambient temperature field, where This is the charging and discharging current. This refers to the battery's internal resistance. Terminal voltage, The electromotive force is taken from the LSTM prediction sequence; the ambient temperature field output temperature is... Reverse correction of the reaction rate coefficient of the electrochemical field ,in This is the reaction rate coefficient (reflecting the speed of an electrochemical reaction); The reference rate coefficient is E; the activation energy of the reaction is E. It is the gas constant; Absolute temperature;

[0205] In the coupling of the power grid's electromagnetic field and the ambient temperature field, the power loss of the power grid lines... (in For line current, Reference temperature The line resistance below is converted into a heat source for the ambient temperature field. ( (Heat power generated by the line); conductor temperature output from the ambient temperature field. Line resistance parameters used to correct the electromagnetic field of the power grid ,in For temperature The line resistance at that time; Temperature coefficient of resistance (reflects the rate of change of resistance with temperature); Reference temperature;

[0206] The coupling between the battery's electrochemical field and the grid's electromagnetic field is achieved through a power balance relationship: battery output power... ( This refers to the battery terminal voltage. (Battery charging and discharging current) must meet the power balance requirements of the grid nodes. ,in This represents the total output power of all batteries. This refers to the total output power of the power supply. Total load power; This represents the total line loss power. When... When the change exceeds 5% of the rated power, the power flow equations are triggered. To resolve, where It is complex power (including active and reactive power). For node voltage phasors; It is the conjugate of the current phasor.

[0207] Furthermore, the field parameter correlation matrix Used to quantify the three-field coupling strength, element field Opposite field Influence coefficient ( , representing the electrochemical field, electromagnetic field, and temperature field respectively), with initial values ​​calibrated based on historical simulation data to ensure that the coupling relationship is consistent with actual physical laws.

[0208] In this embodiment, the digital twin modeling unit 200 combines a multiphysics coupling model with a long short-term memory network time-series prediction algorithm to achieve bidirectional virtual-real mapping and adaptive updates, including the following steps:

[0209] S230.1 Initialization of virtual-real mapping relationship: Based on the high-fidelity dataset output by the data acquisition unit 100, establish parameter mapping rules between physical entities and virtual twins. The mapping rules include a one-to-one correspondence between battery cell voltage and virtual electrode potential, grid node current and virtual line power flow, and ambient temperature and virtual field temperature, and store them in the mapping rule library.

[0210] As a further explanation of this embodiment, the parameter mapping rules in this embodiment specifically include:

[0211] Battery-side mapping rules: Physical cell voltage and virtual electrode potential are mapped through a linear correlation model: Considering factors such as battery electrode polarization effect and sensor measurement error, the proportional coefficient and offset need to be determined in the experimental calibration stage (such as charge and discharge test) so that the virtual electrode potential can accurately reproduce the voltage characteristics of the physical cell and realize the virtual-real correspondence of the core parameters of the electrochemical field.

[0212] Grid-side mapping rules: The calculation of virtual line power flow (complex power) needs to be combined with the conjugate complex number of virtual node voltage and physical node current: Since there is a phase difference between voltage and current in the power grid, the conjugate operation can separate the active and reactive components of complex power, ensuring that the power flow calculation of the virtual twin is consistent with the electrical characteristics of the physical power grid (such as power flow direction and loss distribution).

[0213] Environmental mapping rules: The deviation between physical temperature and virtual field temperature stems from the spatial difference between the sensor installation location and the virtual model's field reference point (e.g., the sensor is attached to the equipment surface, while the virtual model simulates the center temperature of the field). Compensation values ​​are determined through calibration during the installation phase (e.g., simultaneously acquiring physical and virtual initial temperatures). This corrects measurement deviations caused by spatial location and ensures the consistency of the spatial reference for temperature field simulation.

[0214] S230.2 Prediction Sequence Generation and Input: The LSTM behavior prediction module 220 generates a state prediction sequence for a future preset time period (30-60 minutes) based on historical operation data and real-time sensing data. The prediction sequence includes time-series data of battery charging and discharging current, grid node power and ambient temperature, and is transmitted to the multiphysics coupling modeling module 210 through a standardized interface.

[0215] S230.3 Multi-physics co-simulation: The multi-physics coupling modeling module 210 takes the predicted sequence as the boundary condition input and starts the co-simulation of the battery electrochemical field, the power grid electromagnetic field and the ambient temperature field. The simulation step size is consistent with the time granularity of the LSTM predicted sequence (e.g., 5-10 minutes / step), and outputs the state response results of the virtual twin (including battery state of charge, power grid voltage distribution and field temperature gradient).

[0216] S230.4, Calculation of virtual-real deviation: Real-time running data (with the same timestamp as the virtual state response result) is extracted by the distributed sensing module of the data acquisition unit 100, and the parameter correspondence in the mapping rule base is called to align the data. The root mean square error formula is used to calculate the deviation value between the real-time running data and the virtual state response result. The deviation value corresponds to three types of indicators: battery status, grid parameters, and environmental parameters.

[0217] As a further explanation of this embodiment, the root mean square error formula is used to calculate the three types of deviations in this step, specifically including:

[0218] Battery state deviation ;in This represents the number of sampling points; For the physical entity's first One SOC sample value; For the virtual twin's first One SOC simulation value;

[0219] Power grid parameter deviation ;in The number of nodes; For the first physical node Each voltage sample value; For the virtual node's first One voltage simulation value;

[0220] Environmental parameter deviation ;in This represents the number of temperature sampling points. For the physical environment One temperature sample value; For the virtual field A simulated temperature value.

[0221] S230.5 Model Adaptive Update: When the deviation value of any type of index continuously exceeds the preset threshold, the update mechanism is triggered: the gating weight matrix of the LSTM behavior prediction module 220 is adjusted (the input gate and forget gate are updated first), and the coupling coefficient in the field parameter correlation matrix of the multi-physics coupling modeling module 210 is corrected. The update magnitude is positively correlated with the deviation value.

[0222] S230.6 Closed-loop verification and iteration: The updated model repeats steps S230.2-S230.4 until the deviation values ​​of all indicators return to within the threshold, forming a closed-loop iterative process of "prediction-simulation-verification-update".

[0223] As a further explanation of this embodiment, when the deviation value of any type of indicator continuously exceeds a preset threshold, the model is triggered to adaptively update: the gating weight matrix of the LSTM behavior prediction module is adjusted. ,in This is the updated weight matrix; This is the original weight matrix; This is the proportionality coefficient; (For deviation values, prioritize updating the input gate and forget gate); correct the coupling coefficients in the field parameter correlation matrix of the multiphysics coupling modeling module. ,in The updated coupling coefficients; The original coupling coefficient; This is a correction factor; For the field The deviation value is positively correlated with the update magnitude. The updated model repeats the prediction, simulation, and deviation calculation steps until the deviation values ​​of all indicators return to within the threshold, forming a closed-loop iterative process of "prediction-simulation-verification-update".

[0224] The dynamic optimization unit 300 is used to realize the global collaborative optimization of multiple battery packs and power distribution network. It integrates real-time status data of twins with an improved deep reinforcement learning-multi-objective optimization fusion algorithm to generate a globally optimal strategy that dynamically adapts to the operating conditions.

[0225] In this embodiment, the dynamic optimization unit 300 includes an algorithm fusion module 310, a strategy generation module 320, and a working condition adaptation module 330, wherein:

[0226] The algorithm fusion module 310 integrates the real-time state data of the twin with an improved deep reinforcement learning-multi-objective optimization fusion algorithm. The improved deep reinforcement learning-multi-objective optimization fusion algorithm uses the battery state of charge, grid node voltage, and ambient temperature output by the twin as state inputs.

[0227] The strategy generation module 320 generates charging and discharging scheduling strategies, load allocation strategies, and network topology adjustment strategies based on the calculation results of the algorithm fusion module 310. The strategies include dual constraints in the time dimension (scheduling cycle) and the spatial dimension (execution node).

[0228] As a further explanation of this embodiment, the policy generation module 320 converts the optimal solution into structured policy instructions, including execution timestamps, battery charging and discharging scheduling, load allocation, and network topology adjustments. The instructions are transmitted to the execution feedback unit via an HTTPS encrypted channel, with a CRC checksum in the header. Upon successful verification, the receiving end returns an acknowledgment message. If no acknowledgment is received within a timeout period, a retransmission is triggered (up to 3 times) to ensure accurate policy delivery.

[0229] The operating condition adaptation module 330 is used to monitor changes in the operating conditions of the power distribution network and battery pack (such as load fluctuations and battery aging). When the rate of change in operating conditions exceeds a preset threshold, the parameters of the algorithm fusion module 310 are automatically adjusted.

[0230] As a further explanation of this embodiment, the operating condition adaptation module 330 in this embodiment monitors the changes in the operating conditions of the power distribution network and the battery pack in real time, and judges the stability of the operating conditions by load fluctuation rate (|current load - load 1 hour ago| / load 1 hour ago × 100%) and battery aging degree ((initial capacity - current capacity) / initial capacity × 100%), with thresholds set at 20% and 10% respectively;

[0231] When the load fluctuation rate exceeds the threshold, the algorithm fusion module is triggered to increase the NSGA-Ⅲ crossover probability (step size 0.1) and increase the efficiency target weight; when the battery aging degree exceeds the threshold, the upper limit of battery charging and discharging power is reduced and the safety target weight is increased, so as to realize the dynamic adaptation of the algorithm to the operating conditions.

[0232] In this embodiment, the execution of the improved deep reinforcement learning-multi-objective optimization fusion algorithm includes the following steps:

[0233] S340.1, State Input and Feature Extraction:

[0234] The algorithm fusion module 310 receives real-time state data from the twin and extracts the battery pack state of charge sequence. Grid node voltage sequence and ambient temperature Construct state vector The formula is:

[0235] ;

[0236] The vector is then normalized (mapped to the [0,1] interval):

[0237] ;

[0238] ;

[0239] ;

[0240] in, The normalized state of charge. For the first Real-time state of charge of individual battery cells The minimum state of charge for safe battery operation. This represents the maximum state of charge for safe battery operation. The normalized voltage. For the first Real-time voltage of each grid node The minimum voltage required for the safe operation of a power grid node. The maximum voltage for safe operation of a power grid node; The minimum temperature required for safe system operation. The highest temperature for safe system operation. The normalized temperature;

[0241] S340.2, Reward Function Calculation:

[0242] Calculating multi-objective reward values ​​based on state vectors The formula is:

[0243] ;

[0244] in, For cost objectives, In the formula: , These are the grid cost weighting coefficient and the battery cost weighting coefficient, respectively. For power grid operating costs; For battery wear and tear costs; For efficiency goals, In the formula: The amount of energy to be used effectively This represents the total energy consumption of the system. For security purposes, In the formula: This represents the number of nodes in the power distribution network that participate in voltage monitoring. For the first The actual voltage of each grid node This is the minimum allowable voltage at the grid node. The actual temperature of the battery or the environment. The upper limit of the temperature range for safe battery operation; , , These are the weighting coefficients, and The weighting coefficients are dynamically configured by the working condition adaptation module 330;

[0245] S340.3 Multi-objective optimization solution: An improved NSGA-III algorithm is used to solve the decision variables. ( Battery charging and discharging power, For the switch state), optimization is performed, where Indicates power, Indicates the switch state, objective function for:

[0246] ;

[0247] in, To minimize grid operating costs; To minimize battery wear and tear costs; To maximize energy efficiency;

[0248] Constraints are introduced during the optimization process: charging and discharging power ,in This refers to the battery's rated power; adjacent nodes are open when their switch states are different.

[0249] S340.4 Optimal Solution Screening: From the set of non-dominated solutions obtained through optimization, the Analytic Hierarchy Process (AHP) is used to screen indicators (including cost reduction rate, efficiency improvement rate, and safety constraint satisfaction) and construct a judgment matrix. :

[0250] ;

[0251] And select the globally optimal solution that matches the current working condition;

[0252] S340.5, Strategy Generation and Output: The optimal solution is converted into structured strategy instructions (including execution node ID, power value, and timestamp), and output to the execution feedback unit 400 through the strategy generation module 320;

[0253] S340.6 Parameter Iterative Update: After every 100 executions of the strategy, compare the actual reward. With predicted rewards :

[0254] like The crossover probability of NSGA-Ⅲ is dynamically adjusted (from 0.6 to 0.9, with a step size of 0.1).

[0255] The weight coefficients are updated using gradient descent until the policy fit is >90%.

[0256] ;

[0257] ;

[0258] ;

[0259] in, , , These are the updated weighting coefficients; The learning rate; , , All are partial derivatives.

[0260] As a further explanation of this embodiment, the real-time status data output by the digital twin modeling unit 200 is received through a standardized interface (such as a real-time data bus based on the MQTT protocol). This data includes the battery pack state of charge (SOC) sequence, grid node voltage sequence, and ambient temperature. The battery pack SOC sequence is uploaded in real-time by the battery management system (BMS) of each individual battery cell, with a sampling frequency consistent with the twin simulation step size (e.g., 5-10 minutes / time). The grid node voltage sequence covers the line voltage data of all load nodes and power supply nodes in the distribution network and is collected by the feeder terminal units (FTUs). The ambient temperature is a 5-minute moving average collected by temperature sensors in the battery compartment and line corridor. After receiving the data, the algorithm fusion module 310 normalizes the state vector, mapping SOC, voltage, and temperature to the [0,1] interval. The safe range for SOC is set to 10%-90% based on the battery type (e.g., lithium iron phosphate battery), the safe range for grid node voltage is set to 90%-110% of the rated voltage based on the operating procedures, and the safe range for ambient temperature is determined by combining the battery operating temperature (-20℃-55℃) and the line current carrying capacity limit. When the number of battery packs or grid nodes changes, the state vector is expanded using zero-padding to ensure algorithm compatibility.

[0261] As a further explanation of this embodiment, in the reward function calculation, the algorithm fusion module 310 generates a multi-objective reward value through a weighted fusion of cost, efficiency, and security objectives, wherein:

[0262] The cost target includes grid operation cost and battery loss cost. Grid operation cost includes electricity purchase cost (calculated based on time-of-use price and electricity purchase volume) and line loss cost (calculated based on loss power and electricity price). Battery loss cost is estimated based on charge-discharge depth and cycle count. The two are weighted by a weighting coefficient.

[0263] The efficiency target is the ratio of the effective energy utilization (the sum of clean energy consumed by the load and the battery discharge, minus line losses and battery charging and discharging losses) to the total energy of the system (the sum of the absolute values ​​of grid-purchased electricity, clean energy generation, and battery charging and discharging).

[0264] Safety objectives are achieved through voltage over-limit penalties (accumulated based on the absolute value of the deviation when the node voltage exceeds the safe range) and temperature over-limit penalties (calculated at twice the excess value when the temperature exceeds the safe upper limit). The weighting coefficients of the three objectives are dynamically adjusted by the operating condition adaptation module based on real-time operating conditions, such as the weighting of safety objectives during high-temperature periods. It can be increased to 0.5.

[0265] It should be added that the multi-objective optimization solution in this embodiment adopts the improved NSGA-Ⅲ algorithm. The decision variables include battery charging and discharging power (real number encoding, negative values ​​for charging and positive values ​​for discharging) and switch state (binary encoding, 1 for open and 0 for closed). The algorithm improves the quality of the solution by using an initial population (containing 30% of the historical best solutions) and adopts simulated binary crossover (SBX) and a polynomial mutation operator. The crossover probability is dynamically adjusted according to load fluctuations (0.9 when the fluctuation is large and 0.6 when the load is stable).

[0266] In terms of constraint handling, solutions that exceed the power range are truncated and corrected, and solutions where adjacent nodes are simultaneously disconnected are repaired by randomly flipping the state.

[0267] Meanwhile, the optimal solution selection adopts the analytic hierarchy process (AHP), constructing a judgment matrix based on cost reduction rate, efficiency improvement rate, and safety constraint satisfaction. The rationality of the matrix is ​​ensured through consistency checks, and the globally optimal solution is selected according to the principle of the highest weighted score. After every 100 executions of the strategy, the algorithm updates the weight coefficients using gradient descent. The learning rate (initially 0.01) is dynamically adjusted according to the deviation between the actual and predicted rewards (increasing to 0.05 when the deviation is >15%, and decreasing to 0.005 when the deviation is <5%), until the strategy fit (actual reward / predicted reward × 100%) stabilizes above 90%.

[0268] The execution feedback unit 400 is used to achieve accurate execution and dynamic correction of the optimization strategy. It constructs a closed-loop correction link between the twin model and the physical system through instruction parsing technology and real-time status feedback mechanism.

[0269] In this embodiment, the execution feedback unit 400 includes an instruction parsing module 410, an execution driving module 420, a status feedback module 430, and a closed-loop correction module 440, wherein:

[0270] The instruction parsing module 410 adopts a multi-protocol parsing engine, which is compatible with the power distribution terminal communication protocol and the battery BMS communication protocol. It performs format conversion and legality verification on the optimization strategy instructions output by the dynamic optimization unit 300 to ensure that the instructions can be recognized and executed by physical devices.

[0271] As a further explanation of this embodiment, the instruction parsing module 410 in this embodiment adopts a multi-protocol parsing engine, compatible with the communication protocols of the power distribution terminal and the battery BMS, ensuring that the optimization strategy instructions can be recognized and executed by the physical devices. Specifically, the power distribution terminal communication supports DL / T645 (electricity meter protocol) and IEC61850 (substation standard) for parsing remote signaling and telemetry instructions from the power grid switching equipment; the battery BMS communication is compatible with CANopen (vehicle bus) and Modbus-RTU (industrial bus) for parsing battery charging and discharging power instructions and status query frames. This module converts the structured instructions output by the dynamic optimization unit 300 into the device's native frame format, retaining key parameters such as power values ​​and execution time during the conversion process. Simultaneously, the legality of the instructions is ensured through format verification (such as Modbus-RTU CRC check), parameter verification, and permission verification (comparing the device's unique identifier with the pre-stored list).

[0272] The execution drive module 420 generates physical execution signals based on the parsed instructions, drives the battery charging and discharging device and the power grid switching equipment to operate, and has an instruction timeout retransmission mechanism. When the equipment execution is abnormal, it automatically triggers the degradation execution strategy to maintain the basic operation of the system.

[0273] As a further explanation of this embodiment, the execution drive module 420 in this embodiment generates physical execution signals according to the parsed instructions, drives the equipment to operate, and handles abnormal situations. For the battery charging and discharging device, it outputs a PWM (Pulse Width Modulation) signal to adjust the inverter duty cycle (e.g., 50% duty cycle for 50kW), and simultaneously outputs a DC24V relay control signal to switch the charging and discharging mode; for the grid switching equipment, it outputs an AC220V opening and closing coil drive signal (duration matches the equipment operation requirements, e.g., 100ms for a vacuum circuit breaker), and a matching DC110V status holding signal to ensure the switch is stable in the target position. To ensure reliability, a 5-second timer is started after the instruction is issued. If no equipment confirmation frame (e.g., Modbus ACK response) is received, the instruction is retransmitted, up to a maximum of 3 times; if the equipment execution is abnormal (e.g., battery inverter alarm), a degradation strategy is automatically triggered: the battery device switches to "constant voltage mode" to maintain the DC bus voltage stability (e.g., 380V±5%), the switching equipment triggers the backup switch to ensure power supply through topology reconfiguration, and an abnormal code is reported at the same time.

[0274] The status feedback module 430 collects real-time operating status data of the battery charging and discharging device and the power grid switching equipment through the status monitoring interface of the battery charging and discharging device and the power grid switching equipment, including the actual charging and discharging power of the battery and the switching action status. Through the time synchronization mechanism, and based on the twin model status data output by the digital twin modeling unit 200, the virtual and real status are aligned.

[0275] As a further explanation of this embodiment, the status feedback module 430 in this embodiment collects real-time operating status data through the device's built-in interface and achieves spatiotemporal alignment with the digital twin model. Data from the battery charging and discharging device includes: actual output power, DC bus voltage, and converter status word collected via the RS485 interface; data from the power grid switching equipment includes: switch open / closed position (open = 1, closed = 0), number of actions (cumulative counter), and coil status (energized = 1, de-energized = 0) collected via the remote signaling interface. The acquisition cycle is synchronized with the command issuance cycle (5 minutes / time), and high-frequency acquisition at the 1-second level is triggered in emergency situations (such as fault alarms). Simultaneously, spatiotemporal alignment is achieved by calibrating the timestamp using the NTP protocol to ensure that the deviation from the simulation time axis of the digital twin modeling unit 200 is ≤100ms; for physical and virtual data at the same timestamp, small time deviations within 50ms are corrected through linear interpolation to ensure consistent comparison benchmarks.

[0276] Furthermore, the status data of the battery charging and discharging device includes actual output power, DC bus voltage, and converter operating status (normal / alarm). The status data of the grid switching equipment includes switch open / close position, number of operations, and operating coil status (energized / de-energized). The acquisition cycle is synchronized with the instruction issuance cycle of the execution drive module 420.

[0277] The closed-loop correction module 440 is used to calculate the deviation between the real-time operating status data of the battery charging and discharging device and the grid switching equipment and the status of the twin model. When the deviation continues to exceed the set threshold, the twin model parameter correction process is triggered, and the dynamic optimization unit 300 is driven to update the optimization strategy to build a closed-loop correction link between the twin model and the physical system.

[0278] In this embodiment, the closed-loop correction module 440 calculates the deviation between the real-time operating status data of the battery charging and discharging device and the grid switching equipment and the state of the twin model, including the following steps:

[0279] S440.1, Deviation Classification Calculation:

[0280] The physical execution status data (actual output power of the battery charging and discharging device) collected by the status feedback module 430 is received. Opening and closing positions of power grid switchgear ), and the twin model state data (virtual charging and discharging power) output by the digital twin modeling unit 200. Virtual switch position ), calculate the following two types of deviations respectively:

[0281] Power deviation: ;in, The actual output power of the battery charging and discharging device; Virtual charging and discharging power;

[0282] Switch status deviation: And the fractional part is 1 and the composite part is 0; among them, This refers to the open / closed position of the power grid switching equipment; For virtual switch positions;

[0283] S440.2, Calibration Trigger Judgment:

[0284] When power deviation The power of the battery charging and discharging device exceeded the rated power for three consecutive data acquisition cycles. 5%, or switch state deviation When two consecutive cycles are 1 (i.e., the actual and virtual states are inconsistent), the correction mechanism is triggered, and the duration of the deviation and the cumulative deviation value are recorded at the same time.

[0285] S440.3, Twin model parameter correction:

[0286] Regarding power deviation: Correct the reaction rate coefficient of the battery electrochemical field in the multiphysics coupling modeling module 210. The correction formula is:

[0287] ;

[0288] in, This is a correction factor, with a value ranging from 0.8 to 1.2. This is the original reaction rate coefficient in the battery electrochemical field (corresponding to the reverse-corrected reaction rate coefficient of the ambient temperature field). This is the correction factor after power deviation correction;

[0289] To address the switch state deviation: update the adjacency matrix of the graph neural network for the distribution network topology mapping, synchronously correct the coupling coefficient of the field parameter correlation matrix in the multi-physics coupling modeling module 210 (corresponding to the coupling coefficient in the corrected field parameter correlation matrix), and adjust the line resistance parameters of the power grid electromagnetic field based on the conductor temperature output from the ambient temperature field until the virtual switch position is consistent with the actual position.

[0290] S440.4 Optimization Strategy Rolling Update: The corrected multiphysics coupling model parameters are input into the dynamic optimization unit 300, triggering a rolling window update of the optimization strategy: the window length is set to the current scheduling period plus two prediction steps (consistent with the time granularity of the LSTM prediction sequence, i.e., 5-10 minutes / step). The modified particle swarm optimization algorithm is used to resolve the optimal strategy, where the inertia weight... It decreases linearly as the cumulative deviation increases (initial value 0.9, minimum value 0.4).

[0291] S440.5, Verification of Correction Effect: After the new strategy is issued by the instruction parsing module 410, the state feedback module 430 continuously collects physical execution state data and calls the co-simulation results of the multi-physics coupling modeling module 210 for alignment verification (comparison with data of the same timestamp). If the conditions are met within two consecutive cycles... ,and If the error is not found, the correction process is terminated; otherwise, steps S400.3-S400.4 are repeated until the deviation returns to the threshold.

[0292] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0293] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A digital dynamic management system for batteries in a power distribution network based on digital twins, characterized in that, include: The data acquisition unit (100) is used to realize the deep perception of the battery pack's full life cycle status and the power distribution network's full domain operation data. It constructs a high-fidelity dataset based on distributed heterogeneous sensor network and spatiotemporal feature extraction technology, and transmits the high-fidelity dataset to the digital twin modeling unit (200). The digital twin modeling unit (200) is used to construct a multi-dimensional dynamic twin of battery-grid-environment. The digital twin modeling unit (200) receives the high-fidelity dataset output by the data acquisition unit (100), combines the multi-physics coupling model and the long short-term memory network time-series prediction algorithm to realize virtual-real bidirectional mapping and adaptive update, and outputs the real-time status of the twin and the prediction data to the dynamic optimization unit (300). The digital twin modeling unit (200) combines a multiphysics coupling model with a long short-term memory network time-series prediction algorithm to achieve bidirectional mapping between virtual and real worlds and adaptive updates, including the following steps: S230.1 Initialization of virtual-real mapping relationship: Based on the high-fidelity dataset output by the data acquisition unit (100), establish parameter mapping rules between physical entities and virtual twins, and store them in the mapping rule library; S230.2, Prediction Sequence Generation and Input: The LSTM behavior prediction module (220) generates a state prediction sequence for a future preset period based on historical operation data and real-time sensing data, and transmits it to the multi-physics coupling modeling module (210) through a standardized interface. S230.3 Multi-physics co-simulation: The multi-physics coupling modeling module (210) takes the prediction sequence as the boundary condition input and starts the co-simulation of the battery electrochemical field, the power grid electromagnetic field and the ambient temperature field. The simulation step size is consistent with the time granularity of the LSTM prediction sequence, and outputs the state response results of the virtual twin. Among them, the prediction sequence includes the time series data of battery charging and discharging current, power grid node power and ambient temperature. S230.4, Calculation of virtual-real deviation: Real-time running data is extracted by the distributed sensing module of the data acquisition unit (100), and the parameter correspondence in the mapping rule base is called to align the data. The root mean square error formula is used to calculate the deviation between the real-time running data and the virtual state response result. S230.5, Model Adaptive Update: When the deviation value of any type of index continuously exceeds the preset threshold, the update mechanism is triggered: the gating weight matrix of the LSTM behavior prediction module (220) is adjusted, and the coupling coefficient in the field parameter correlation matrix of the multi-physics coupling modeling module (210) is corrected. The update magnitude is positively correlated with the deviation value. S230.6 Closed-loop verification and iteration: The updated model repeats steps S230.2-S230.4 until the deviation values ​​of all indicators return to within the threshold, forming a closed-loop iterative process of "prediction-simulation-verification-update". The dynamic optimization unit (300) is used to realize the global collaborative optimization of multiple battery packs and power distribution network. The dynamic optimization unit (300) receives the real-time status and prediction data of the twin output by the digital twin modeling unit (200), integrates the real-time status data of the twin with the improved deep reinforcement learning-multi-objective optimization fusion algorithm, generates a globally optimal strategy that dynamically adapts to the working conditions, and converts the globally optimal strategy into a structured strategy instruction and transmits it to the execution feedback unit (400). The dynamic optimization unit (300) includes an algorithm fusion module (310), a strategy generation module (320), and a working condition adaptation module (330), wherein: The algorithm fusion module (310) integrates the real-time state data of the twin with the improved deep reinforcement learning-multi-objective optimization fusion algorithm. The improved deep reinforcement learning-multi-objective optimization fusion algorithm uses the battery state of charge, grid node voltage, and ambient temperature output by the twin as state inputs. The strategy generation module (320) generates a charging and discharging scheduling strategy, a load allocation strategy, and a network topology adjustment strategy based on the calculation results of the algorithm fusion module (310). The strategy includes dual constraints in terms of time and space dimensions. The operating condition adaptation module (330) is used to monitor the operating condition changes of the power distribution network and battery pack. When the operating condition change rate exceeds the preset threshold, the parameter adaptive adjustment of the algorithm fusion module (310) is triggered. The execution of the improved deep reinforcement learning-multi-objective optimization fusion algorithm includes the following steps: S340.1, State Input and Feature Extraction: The algorithm fusion module (310) receives real-time state data from the twin and extracts the battery pack state of charge sequence. Grid node voltage sequence and ambient temperature Construct state vector ; Then normalize the vector: ； ； ； in, The normalized state of charge. For the first Real-time state of charge of individual battery cells The minimum state of charge for safe battery operation. This represents the maximum state of charge for safe battery operation. The normalized voltage. For the first Real-time voltage of each grid node The minimum voltage required for the safe operation of a power grid node. The maximum voltage for safe operation of a power grid node; The minimum temperature required for safe system operation. The highest temperature for safe system operation. The normalized temperature; S340.2, Reward Function Calculation: Calculating multi-objective reward values ​​based on state vectors ; S340.3 Multi-objective optimization solution: An improved NSGA-III algorithm is used to solve the decision variables. Optimize, in the formula Indicates the battery's charging and discharging power. Indicates the switch state, objective function for: ； in, To minimize grid operating costs; To minimize battery wear and tear costs; To maximize energy efficiency; Constraints are introduced during the optimization process: charging and discharging power ,in This refers to the battery's rated power; adjacent nodes are open when their switch states are different. S340.4 Optimal Solution Screening: From the set of non-dominated solutions obtained through optimization, the analytic hierarchy process (AHP) is used to screen indicators and construct a judgment matrix. : ； And select the globally optimal solution that matches the current working condition; S340.5, Strategy generation and output: Convert the optimal solution into structured strategy instructions and output them to the execution feedback unit (400) through the strategy generation module (320). S340.6 Parameter Iterative Update: After every 100 executions of the strategy, compare the actual reward. With predicted rewards : like Dynamically adjust the NSGA-Ⅲ crossover probability; The weight coefficients are updated using gradient descent until the policy fit is greater than 90%. The execution feedback unit (400) is used to realize the accurate execution and dynamic correction of the optimization strategy. It receives the structured strategy instructions output by the dynamic optimization unit (300) and constructs a closed-loop correction link between the twin model and the physical system through instruction parsing technology and real-time status feedback mechanism.

2. The digital twin-based dynamic management system for power distribution network batteries according to claim 1, characterized in that, The data acquisition unit (100) includes a distributed sensing module (110), a data transmission module (120), and a spatiotemporal feature extraction module (130), wherein: The distributed sensing module (110) is deployed at key nodes of the battery pack and power distribution network to collect multi-dimensional operating data; The data transmission module (120) adopts a hybrid communication architecture to transmit the data collected by the distributed sensing module (110) to the cloud. The spatiotemporal feature extraction module (130) generates a spatiotemporal index dataset based on a time series alignment algorithm and a spatial association mapping algorithm.

3. The digital twin-based dynamic management system for power distribution network batteries according to claim 2, characterized in that, The spatiotemporal feature extraction module (130) includes a time series alignment submodule (131), a spatial association mapping submodule (132), and a spatiotemporal index generation submodule (133), wherein: The time series alignment submodule (131) is used to perform time synchronization processing on the multi-frequency data collected by the distributed sensing module (110), and adopts a sliding time window algorithm to achieve data alignment. The window size is dynamically adjusted according to the data fluctuation characteristics. The spatial association mapping submodule (132) constructs a graph neural network model based on the power distribution network topology. The input layer of the graph neural network model contains node voltage, current, and power feature vectors, and the number of hidden layer nodes is dynamically configured according to the network size. The spatiotemporal index generation submodule (133) is used to fuse the time-aligned data with the spatial association model to generate a spatiotemporal index dataset containing four-dimensional labels of battery status, power grid parameters, time, and location.

4. The digital twin-based dynamic management system for distribution network batteries according to claim 1, characterized in that, The digital twin modeling unit (200) includes a multiphysics coupling modeling module (210) and an LSTM behavior prediction module (220), wherein: The multiphysics coupling modeling module (210) is used to construct a coupling model of the battery electrochemical field, the power grid electromagnetic field and the ambient temperature field. The coupling model defines the interaction relationship between different physical fields through the field parameter correlation matrix. The LSTM behavior prediction module (220) predicts trends in battery charging and discharging characteristics, power grid flow changes, and environmental parameter evolution based on a long short-term memory network, including the following steps: S220.1 Input Layer Feature Processing: Standardize historical operational data and real-time sensing data to generate a sequence of feature vectors. ,in For time steps; S220.2, Gating Unit Calculation: Calculate the forget gate separately Input gate Candidate cell status Cell state update Output gate and hidden state output ; S220.3, Prediction Output: ,get The predicted value at time; where, This is the weight matrix of the output layer. For bias terms, This is the Sigmoid activation function.

5. The digital twin-based dynamic management system for distribution network batteries according to claim 4, characterized in that, The multiphysics coupling of the multiphysics coupling modeling module (210) specifically includes: Coupling of battery electrochemical field and ambient temperature field: The heat power generated by battery charging and discharging is used as the input of ambient temperature field, and the temperature distribution output by ambient temperature field inversely corrects the reaction rate coefficient of battery electrochemical field. Coupling of the power grid electromagnetic field and the ambient temperature field: The power loss of the power grid line is converted into a heat source of the ambient temperature field. The conductor temperature output by the ambient temperature field is used to correct the line resistance parameters of the power grid electromagnetic field. Coupling of the battery electrochemical field and the power grid electromagnetic field: A correlation is established through the balance relationship between the battery charging and discharging power and the power of the power grid nodes. Changes in the battery output power trigger the recalculation of the power flow of the power grid electromagnetic field.

6. The digital twin-based dynamic management system for distribution network batteries according to claim 1, characterized in that, The execution feedback unit (400) includes an instruction parsing module (410), an execution drive module (420), a status feedback module (430), and a closed-loop correction module (440), wherein: The instruction parsing module (410) adopts a multi-protocol parsing engine, which is compatible with the power distribution terminal communication protocol and the battery BMS communication protocol. It performs format conversion and legality verification on the optimization strategy instructions output by the dynamic optimization unit (300) to ensure that the instructions can be recognized and executed by physical devices. The execution drive module (420) generates physical execution signals according to the parsed instructions, drives the battery charging and discharging device and the power grid switching equipment to operate, and has an instruction timeout retransmission mechanism. When the equipment execution is abnormal, it automatically triggers the degradation execution strategy to maintain the basic operation of the system. The status feedback module (430) collects real-time operating status data of the battery charging and discharging device and the power grid switching equipment through the status monitoring interface of the battery charging and discharging device and the power grid switching equipment. Through the time synchronization mechanism, and based on the twin model status data output by the digital twin modeling unit (200), the virtual and real status are aligned. The closed-loop correction module (440) is used to calculate the deviation between the real-time operating status data of the battery charging and discharging device and the grid switching equipment and the status of the twin model. When the deviation continues to exceed the set threshold, the twin model parameter correction process is triggered, and the dynamic optimization unit (300) is driven to update the optimization strategy to build a closed-loop correction link between the twin model and the physical system.

7. The digital twin-based dynamic management system for power distribution network batteries according to claim 6, characterized in that, The closed-loop correction module (440) calculates the deviation between the real-time operating status data of the battery charging and discharging device and the grid switching equipment and the state of the twin model, including the following steps: S440.1, Deviation Classification Calculation: The physical execution status data collected by the receiving status feedback module (430) and the twin model status data output by the digital twin modeling unit (200) are compared with the following two types of deviations: Power deviation : ;in, The actual output power of the battery charging and discharging device; Virtual charging and discharging power; Switching status deviation : And the fractional part is 1 and the composite part is 0; among them, This refers to the open / closed position of the power grid switching equipment; For virtual switch positions; S440.2, Calibration Trigger Judgment: When power deviation The power of the battery charging and discharging device exceeded the rated power for three consecutive data acquisition cycles. 5%, or switch state deviation When the value is 1 for two consecutive cycles, the correction mechanism is triggered, and the duration of the deviation and the cumulative deviation value are recorded simultaneously. S440.3, Twin model parameter correction: For power deviation: Correct the reaction rate coefficient of the battery electrochemical field in the multiphysics coupling modeling module (210). The correction formula is: ； in, For correction factors, The primary reaction rate coefficient in the battery electrochemical field; This is the correction factor after power deviation correction; For the switch state deviation: update the graph neural network adjacency matrix of the distribution network topology mapping, synchronously correct the coupling coefficient of the field parameter correlation matrix in the multi-physics field coupling modeling module (210), and adjust the line resistance parameters of the power grid electromagnetic field based on the conductor temperature output by the ambient temperature field until the virtual switch position is consistent with the actual position. S440.4 Optimization Strategy Rolling Update: Input the corrected multiphysics coupling model parameters into the dynamic optimization unit (300) to trigger the rolling window update of the optimization strategy: The window length is set to the current scheduling period plus 2 prediction steps. The modified particle swarm optimization algorithm is used to resolve the optimal strategy, where the inertia weight... It decreases linearly as the cumulative deviation increases; S440.5, Verification of Correction Effect: After the new strategy is issued by the instruction parsing module (410), the state feedback module (430) continuously collects physical execution state data and calls the co-simulation results of the multi-physics coupling modeling module (210) for alignment verification. If the alignment is satisfied within two consecutive cycles... ,and If the error is not found, the correction process is terminated; otherwise, steps S400.3-S400.4 are repeated until the deviation returns to the threshold.

Images