An intelligent power supply collaborative control method based on digital twinning and related equipment

Abstract

The application discloses an intelligent power supply cooperative control method based on digital twinning and related equipment, and relates to the technical field of power supply systems. The method comprises the following steps: acquiring current battery time sequence data and a digital twinning model of a power management system; inputting the current battery time sequence data into a pre-trained convolutional neural network to predict a charging current upper limit sequence within a preset future time step; updating a current state of the digital twinning model based on current operation data in the current battery time sequence data to obtain a model update state; performing parallel forward simulation on a pre-generated strategy pool based on the model update state through the updated digital twinning model, and determining and executing an optimal control strategy in combination with the charging current upper limit sequence. The application realizes smooth matching of charging current and load demand when nuclear capacity ends and uniform charging starts, avoids a large drop in bus voltage, guarantees that communication equipment does not restart after power failure, improves system power supply stability, and can be widely applied to the technical field of power systems.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a method and related equipment for intelligent power supply collaborative control based on digital twins. Background Technology

[0002] With the accelerating pace of digitalization and intelligentization in the power industry, communication power supplies, as the core power supply infrastructure for communication network operation, are becoming increasingly important in terms of stable and reliable operation and automated maintenance capabilities. In communication site power supply scenarios, battery banks play a crucial role in emergency backup power and voltage regulation. Relying on intelligent control methods to achieve refined charging and discharging management and remote operation and maintenance control of batteries has become the mainstream direction for the upgrading and development of communication power supply systems. Various intelligent prediction algorithms and digital twin simulation modeling technologies are also gradually being applied to the field of power supply control.

[0003] Current communication power supply systems generally adopt a split architecture to build the operation and maintenance control system. The entire system mainly consists of a high-frequency switching power supply, a battery pack, and various independent functional devices. In practical applications, an independent online battery monitoring device is configured to collect battery status parameters, and an independent remote capacity assessment device is used to perform remote battery capacity testing. Additionally, a protocol conversion device is added to achieve unified adaptation of communication interfaces for various devices, thereby realizing remote operation and maintenance functions such as remote battery capacity assessment, status monitoring, remote start / stop, and remote alarm verification. At the power supply charging and discharging control level, the industry commonly uses model predictive control strategies to optimize charging control, while combining neural networks to predict the remaining battery charge trajectory. Based on the established battery equivalent circuit model, the system's charging and discharging behavior constraints and parameter control are achieved.

[0004] Existing traditional control methods and discrete architectures have significant shortcomings. First, traditional charge and discharge control relies heavily on accurate battery equivalent mechanism models, which are susceptible to model mismatch issues caused by battery aging and ambient temperature, resulting in poor control adaptability and matching with actual operating conditions. Second, traditional solutions lack an integrated collaborative simulation and prediction mechanism, making it impossible to predict changes in operating conditions in advance and match appropriate control strategies. Insufficient equipment coordination can easily lead to practical operational problems such as operational logic conflicts and voltage fluctuations during power supply mode switching. Summary of the Invention

[0005] This invention provides a digital twin-based intelligent power supply collaborative control method and related equipment, which solves the technical problems of model mismatch, weak equipment coordination and voltage fluctuation in existing discrete power supply systems.

[0006] In a first aspect, the present invention provides an intelligent power supply cooperative control method based on digital twins, comprising:

[0007] Acquire the current battery timing data and digital twin model of the power management system;

[0008] The current battery timing data is input into a pre-trained convolutional neural network to predict the upper limit sequence of charging current within a preset future time step.

[0009] The current state of the digital twin model is updated based on the current running data in the current battery time series data to obtain the model update state;

[0010] Based on the updated digital twin model and the updated state of the model, a parallel forward simulation is performed on the pre-generated strategy pool, and the optimal control strategy is determined and executed in conjunction with the charging current upper limit sequence.

[0011] In some embodiments, inputting the current battery timing data into a pre-trained convolutional neural network to predict the upper limit sequence of charging current within a preset future time step includes:

[0012] Convert the current battery timing data into a two-dimensional matrix;

[0013] The two-dimensional matrix is ​​input into a convolutional neural network to perform convolutional feature extraction and fully connected layer prediction, and outputs a sequence of charging current upper limits within a preset future time step.

[0014] In some embodiments, the battery timing data includes timing data of multiple state parameters, and converting the battery timing data into a two-dimensional matrix includes:

[0015] The time-series data of each state parameter is used as a row of a two-dimensional matrix;

[0016] The data corresponding to all state parameters at the same sampling time are arranged in parameter order to form a column of a two-dimensional matrix.

[0017] In some embodiments, the pre-generated policy pool includes multiple candidate control policies. The step of performing parallel forward simulation on the pre-generated policy pool based on the updated state of the updated digital twin model, and determining and executing the optimal control policy in conjunction with the charging current upper limit sequence, includes:

[0018] Each of the candidate control strategies is input into the updated digital twin model. Based on the updated state of the model, the operation of the power management system under each candidate control strategy is simulated, and the system response trajectory data corresponding to each candidate control strategy is output.

[0019] Based on the system response trajectory data, current limit constraint determination is performed in conjunction with the charging current upper limit sequence, and the comprehensive cost of each candidate control strategy is calculated according to the preset comprehensive cost rule.

[0020] The comprehensive costs of all candidate control strategies are ranked, and the candidate control strategy with the minimum comprehensive cost is determined as the optimal control strategy and executed.

[0021] In some embodiments, the digital twin model includes a virtual battery model and a virtual power supply model. The step of inputting each candidate control strategy into the updated digital twin model, simulating the operation of the power management system under each candidate control strategy based on the model update state, and outputting system response trajectory data corresponding to each candidate control strategy includes:

[0022] Analyze each of the candidate control strategies and determine the charging control parameters of each candidate control strategy;

[0023] The charging control parameters and the model update state are input into the virtual battery model to perform battery terminal voltage prediction processing and output the predicted battery terminal voltage values ​​at multiple consecutive simulation times.

[0024] The predicted battery terminal voltage, the model update status, and the charging control parameters are input into the virtual power model to simulate the changes in the operating status of the power management system within a preset time period, and the system operating status data at each simulation time are output.

[0025] The system operation status data at all simulation moments are integrated in time series to generate system response trajectory data.

[0026] In some embodiments, the overall cost is calculated using the following formula:

[0027] ;

[0028] ;

[0029] in, For the first The overall cost-effectiveness of each candidate control strategy For the time-series index of the simulation prediction step size, This represents the predicted total number of steps in a single-round strategy simulation. This is a voltage deviation penalty item. The weighting coefficients, Penalty term for current change The weighting coefficients, The weighting coefficient for the battery health status penalty term. For the future obtained through digital twin model simulation The predicted value of the bus voltage of the step, The preset bus voltage reference target value, For the first The charging current sequence corresponding to each candidate control strategy This is the variance calculation function, characterizing the degree of fluctuation in the current sequence. For the first Among the candidate control strategies, the first one is... The charging current command value for the step. For the future number corresponding to the upper limit sequence of charging current The upper limit of the charging current of the step.

[0030] In some embodiments, the method further includes:

[0031] Obtain the historical state feature vector of the power management system, the current current alarm threshold, and the current maximum charging current limit;

[0032] The current state feature vector is calculated based on the updated state of the model.

[0033] Calculate the historical feature difference vector based on the historical state feature vector and the current state feature vector;

[0034] The historical feature difference vector and the current state feature vector are concatenated to generate a collaborative feature vector;

[0035] The collaborative feature vector is input into a multilayer perceptron neural network, which outputs a first feature parameter and a second feature parameter.

[0036] Based on the first feature parameter and the current maximum charging current, the operating state of the digital twin model is updated and used as the initial model state for the next control simulation cycle; the start time of the next control simulation cycle is the end time of the execution of the optimal control strategy.

[0037] The current current alarm threshold is updated based on the second feature parameter.

[0038] In some embodiments, updating the operating state of the digital twin model based on the first feature parameter and the current maximum charging current includes:

[0039] The first characteristic parameter is inversely transformed to obtain the preliminary charging current value;

[0040] Based on the current maximum charging current value, the amplitude of the initial charging current value is constrained to determine the target charging control current;

[0041] According to the preset timing state iteration update rules, the operating state of the digital twin model is updated based on the target charging control current.

[0042] Secondly, the present invention provides an intelligent power supply cooperative control device based on digital twins, used to execute the aforementioned intelligent power supply cooperative control method based on digital twins, comprising:

[0043] The data acquisition module is used to acquire the current battery timing data and digital twin model of the power management system;

[0044] The current prediction module is used to input the current battery timing data into a pre-trained convolutional neural network to predict the upper limit sequence of charging current within a preset future time step.

[0045] The state update module is used to update the current state of the digital twin model based on the current running data in the current battery time series data, so as to obtain the model update state;

[0046] The strategy control module is used to perform parallel forward simulation of the pre-generated strategy pool based on the updated state of the updated digital twin model, and determine and execute the optimal control strategy in combination with the charging current upper limit sequence.

[0047] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the above-mentioned intelligent power supply cooperative control method based on digital twin.

[0048] As can be seen from the above technical solutions, the present invention has the following advantages: The present invention provides a digital twin-based intelligent power supply collaborative control method and related equipment, which predicts the future charging current upper limit sequence by inputting battery time-series data into a convolutional neural network, and optimizes the control strategy by combining the digital twin model with parallel simulation. It eliminates the need for online identification of time-varying parameters, avoids the control performance degradation caused by model mismatch, and enables the system to dynamically adjust the alarm threshold and charging current. This avoids false alarms triggered by voltage drops during core discharge due to fixed thresholds. By performing parallel forward simulation on the pre-generated strategy pool through the updated digital twin model, the optimal control strategy is selected under the constraint of the charging current upper limit sequence. This achieves a smooth match between the charging current and load demand when the core discharge ends and the equalization charge begins, avoiding a significant drop in bus voltage, ensuring that communication equipment does not power down and restart, and improving the system power supply stability. Through the unified decision-making framework of parallel simulation of the digital twin model and the strategy pool, the alarm threshold adjustment and charging current control are deeply integrated, replacing the complex coordination logic between discrete devices, simplifying the system structure, and reducing hardware costs and debugging complexity. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A flowchart illustrating an intelligent power supply collaborative control method based on digital twins, provided as an embodiment of the present invention;

[0051] Figure 2 A schematic diagram of the structure of an intelligent power supply collaborative control device based on digital twin provided in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0053] This invention provides a digital twin-based intelligent power supply collaborative control method and related equipment to solve the technical problems of model mismatch, weak equipment collaboration, and voltage fluctuation in existing discrete power supply systems.

[0054] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0055] Please see Figure 1 , Figure 1 An optional flowchart of an intelligent power supply collaborative control method based on digital twin provided in an embodiment of the present invention includes steps 101 to 104.

[0056] Step 101: Obtain the current battery timing data and digital twin model of the power management system;

[0057] A power management system (BMS) is an electronic device used to monitor the status of a battery pack, calculate secondary data, report operating conditions, and perform battery balancing management. In this embodiment, the BMS collects at least the voltage, current, and temperature of individual battery cells, as well as the state of charge (SOC) and state of health (SOH) estimated by algorithms.

[0058] The current battery time-series data refers to the real-time, continuously stored time-series operating status data of the battery and communication power system collected by the power management system (BMS). It covers core operating parameters of the power system, specifically including key operating condition data such as bus voltage, load current, battery charge / discharge current, battery state of charge (SOC), and battery state of health (SOH). This time-series data is continuous and can fully characterize the current operating conditions and state change trends of the battery and power system. It serves as the core data foundation for subsequent model prediction, state updates, and strategy simulation. The data acquisition nodes in this step focus on the critical point of the power system's transition from float charging to core capacity operation, accurately capturing the transient characteristics of this transition.

[0059] The digital twin model is a virtual simulation model of the communication power system pre-built and deployed on the local control terminal in this invention. It is used to simulate the dynamic behavior of the real system. Unlike traditional high-precision heavy simulation models, it features lightweight, low latency, and high simulation speed. It can replace traditional online optimization algorithms to achieve rapid forward simulation iteration of control strategies. The model is an integrated collaborative simulation architecture, mainly composed of three coupled parts: a virtual power model, a virtual battery model, and a virtual controller. It can completely replicate the operating characteristics, control logic, and dynamic response laws of the physical communication power system.

[0060] Virtual power supply model refers to a fast power supply dynamic simulator with a pre-trained spatiotemporal graph convolutional network (ST-GCN) as its core. It is used to mine the spatiotemporal correlation characteristics of multiple parameters of the power supply system, simulate the dynamic operating characteristics of the power supply side, and adapt to the power supply state deduction under complex operating conditions.

[0061] A virtual battery model refers to a battery response simulator with a pre-trained two-dimensional convolutional neural network (2D CNN) as its core. It is used to quickly fit the charging and discharging response characteristics and state change patterns of a battery, and accurately replicate the battery's operating condition feedback under different control strategies.

[0062] A virtual controller is a virtual control unit that completely replicates the underlying control logic of a physical controller. It can independently execute and debug various candidate control strategies in a virtual twin space without occupying the operating resources of physical devices, thus providing computing power support for parallel simulation and screening of multiple strategies.

[0063] Step 101 is the basic step for data acquisition and model initialization in the above-mentioned intelligent power supply cooperative control method, providing core input and simulation carrier for subsequent current upper limit prediction, twin model state update, and multi-strategy simulation optimization.

[0064] Traditional communication power control schemes rely on physical device measured feedback and online parameter optimization, which suffers from high computational latency, delayed operational condition prediction, and the inability to perform parallel trial calculations of multiple strategies. This invention pre-constructs an integrated, lightweight digital twin model, creating a simulation environment in virtual space that maps one-to-one with the power management system. Simultaneously, the power management system captures real-time battery time-series operational data at the critical point of the float-to-capacity switching, accurately anchoring the system's transient operating state. Leveraging the dynamic characteristics of the time-series data and the virtual-real mapping capability of the digital twin model, this invention eliminates the reliance on battery mechanism model parameter identification found in traditional schemes. It replaces traditional online optimization calculations with data-driven simulation, laying a data and model foundation for the subsequent rapid completion of forward simulations of multiple control strategies and the selection of the optimal collaborative control strategy.

[0065] For example, in a typical scenario of switching from float charging to remote capacity verification for communication power supply batteries, the data acquisition and model retrieval operations in this step are performed in real time. Specifically, the power system time-series operation data at the current switching critical point is collected in real time through the on-site power management system. The sampling frequency is set to 10Hz, and the time-series data of the current moment and previous historical moments are continuously collected. The final obtained current battery time-series data includes complete operating condition parameters such as real-time bus voltage, real-time load current, battery current at the float charging to capacity verification switching node, real-time battery SOC value, and real-time battery SOH value. At the same time, a locally pre-built lightweight digital twin model of the communication power supply is retrieved. This model has completed offline training and parameter solidification, and internally integrates the ST-GCN virtual power supply model, the 2D CNN virtual battery model, and a virtual controller that replicates the physical control logic, which can realize the full-dimensional virtual replication of the power system dynamic characteristics, battery response characteristics, and equipment control logic. After completing the acquisition of current battery time-series data and the retrieval of the digital twin model, the real-time time-series data is bound and associated with the lightweight twin model to complete the data and model initialization for this round of regulation cycle, providing basic support for subsequent upper limit sequence prediction of charging current, twin model state update, and multi-strategy parallel simulation optimization.

[0066] In some embodiments, it is assumed that an integrated intelligent power supply device is configured in the communication power supply system of a 220kV substation. This device has been offline trained and deployed with an ST-GCN virtual power supply model, a 2D CNN virtual battery model, and the policy pool required for collaborative control. The system sampling period is set to... The current time is recorded as The execution process of step 101 is as follows:

[0067] Reading current battery timing data: The embedded processor reads real-time status data from the high-precision sampling circuit and power management system within the most recent sampling period via its internal data bus. Assume the read data is as follows:

[0068] bus voltage (48V lower than the rated value);

[0069] Battery voltage (At this time, the battery is directly connected in parallel with the bus, so the voltage is the same; it may be different when the battery is discharged with full capacity. In this example, we are at the point of switching from float charging to full capacity charging, so we take the same value for now.)

[0070] Load current ;

[0071] Battery current ;

[0072] Battery state of charge ;

[0073] Battery health status ;

[0074] Battery temperature ;

[0075] The above data is organized into a multidimensional state vector in a predefined order. ;

[0076] Loading the digital twin model: The processor reads the stored ST-GCN model parameters (including adjacency matrix, graph convolution kernel weights, temporal convolution kernel weights, etc.) and 2D CNN model parameters (convolutional layer weights, fully connected layer weights, etc.) from the internal Flash memory and initializes a logical copy of the virtual controller. These models are loaded into the dedicated memory of the embedded processor's NPU (Neural Processing Unit) to accelerate subsequent inference.

[0077] Step 102: Input the current battery timing data into a pre-trained convolutional neural network to predict the upper limit sequence of charging current within a preset future time step;

[0078] In this embodiment, the pre-trained convolutional neural network can be configured as a two-dimensional convolutional neural network (2D CNN), which is a deep learning model specifically designed for processing grid-like data (such as images). In this invention, one-dimensional battery time-series data is transformed into a two-dimensional matrix (considered as a single-channel or multi-channel image) through a special arrangement. The convolutional layers of the 2D CNN are used to automatically extract the local correlation features of the battery state in both time and parameter dimensions. Finally, the charging current safety boundary for multiple future time periods is output through a fully connected layer.

[0079] The charging current upper limit sequence refers to an ordered set of the maximum charging current values ​​that a battery can safely withstand under its current health state, temperature, and state of charge within a continuous future time window (preset future time step) (e.g., 5 future sampling periods, each period being 10ms). This sequence is dynamically changing and is updated in real time as the battery ages, temperature changes, and SOC changes.

[0080] Step 102 is the core step of operating condition prediction and constraint generation, realizing data-driven intelligent prediction and constraint generation, replacing the traditional online parameter identification scheme based on mechanistic models. The upper limit of charging current in traditional communication power supplies mostly relies on fixed mechanistic models and static threshold settings. When batteries age, temperature changes, or operating conditions switch, model mismatch and poor threshold adaptability are prone to occur, leading to conservative or over-limit operation of charging control strategies, causing problems such as voltage fluctuations and accelerated battery wear.

[0081] Step 102 uses a pre-trained 2D CNN model to predict the dynamic current upper limit. The core principle consists of three stages: data image reconstruction, 2D convolutional feature extraction, and time-series prediction output. First, the collected multi-dimensional battery time-series operating data is reconstructed into a multi-channel standardized 2D image matrix to meet the input dimension requirements of the 2D CNN model. Second, relying on the multi-layer convolution and pooling operations of the 2D CNN model, the multi-parameter coupling features and time-series change local features of the battery are extracted. Activation functions are then superimposed to complete the nonlinear transformation and purification of features, accurately capturing the dynamic response law of battery charging and discharging. Finally, a fully connected layer is used to perform dimensional mapping on the deep fusion features, and the dynamic charging current upper limit sequence for future multiple time moments is regressed and output.

[0082] The model is trained offline using a large amount of historical data on normal battery charging and discharging conditions, core capacity conditions, and variable temperature aging conditions. It uses the upper limit of the battery charging current at future moments as a supervision signal to complete parameter convergence and solidification. It does not require online parameter identification and can adapt to complex scenarios such as float charging, core capacity, and operating condition switching in real time. It dynamically outputs current constraints that fit the real-time health status of the battery and the system operating conditions, providing accurate and dynamic constraint boundaries for subsequent optimization of collaborative control strategies.

[0083] Optionally, a 2D CNN is used to directly learn the end-to-end mapping from recent battery state images to future charging current upper limit sequences. The technical principle includes:

[0084] Joint extraction of spatiotemporal features: There are complex nonlinear couplings between parameters such as battery voltage, current, temperature, and SOC, as well as between different times. After organizing the time-series data into a two-dimensional matrix, the convolutional kernel of the 2D CNN can slide simultaneously along the time axis (column direction) and the parameter axis (row direction), automatically learning composite features across parameters and across time, such as the decrease in the upper limit of current when the temperature rises, the decrease in the upper limit of current when the SOC rises, and the impact of historical current fluctuations on the current safety boundary, without the need for manual feature engineering design;

[0085] Advantages of graphical representation: Compared to directly inputting time-series data into a one-dimensional CNN or LSTM, the two-dimensional matrix form preserves the spatial relationships between different parameters. For example, by fixing voltage, current, temperature, and SOC as the first four rows of the matrix, a 2D CNN can explicitly model the local correlations between these parameters (similar to the relationship between adjacent pixels in an image) through convolution operations, resulting in more complete information exchange.

[0086] Lightweight Real-Time Inference: The 2D CNN architecture (two convolutional layers + global average pooling + fully connected layers) used in this invention has a small number of parameters and low computational cost. On an embedded NPU, the time for a single forward propagation is much shorter than the sampling period (e.g., only 0.5ms), meeting real-time control requirements. Offline training ensures the model's generalization ability, and online training requires no iterative optimization, resulting in a deterministic output.

[0087] Step 102 is the safety boundary generation stage in the entire cooperative control process. Its output charging current upper limit sequence will serve as an important component of the cost function in step 104 (determining the optimal control strategy): any candidate strategy's charging current command exceeding the upper limit at the corresponding moment will be penalized, thus preventing irreversible damage to the battery (such as lithium plating or overheating) at the source. Simultaneously, this upper limit sequence also provides a benchmark value for the maximum allowable current for cooperative decision-making in the virtual feature space.

[0088] In some embodiments, step 102 may include the following sub-steps:

[0089] Convert the current battery timing data into a two-dimensional matrix;

[0090] The two-dimensional matrix is ​​input into the convolutional neural network to perform convolutional feature extraction and fully connected layer prediction, and outputs a sequence of charging current upper limits within a preset future time step.

[0091] A two-dimensional matrix (graphical representation) refers to the reorganization of multivariate time series data into a two-dimensional array with height (feature dimension) and width (time step). In this invention, the battery data for each sampling period contains multiple state parameters (such as voltage, current, temperature, and SOC), and these parameters are used as rows of the matrix, with the most recent... The data from each sampling period are arranged in chronological order as columns of a matrix, thus forming a... A two-dimensional matrix, where, Given the number of state parameters, this matrix can be viewed as a graph with... One channel, width is The image.

[0092] Specifically, the execution flow of step 102 is as follows:

[0093] Current battery time-series data 2D matrix transformation: Extract multi-dimensional state parameters such as bus voltage, load current, battery current, SOC, and SOH from the current battery time-series data, and use the continuous time-series data of each type of state parameter as single row data of a 2D matrix; at the same time, according to a unified parameter sorting rule, arrange all state parameter data corresponding to the same sampling time vertically to form a single column data of a 2D matrix, and finally generate a standardized 2D feature matrix that adapts to the input of the spatiotemporal graph convolutional network model;

[0094] Spatiotemporal convolutional feature extraction and current upper limit sequence prediction: The constructed two-dimensional feature image matrix is ​​input into the pre-trained 2D CNN model; the model first extracts local deep features of the matrix by traversing through multiple layers of two-dimensional convolutional kernels, and compresses the number of feature parameters and purifies the effective operating condition features by combining max pooling and global average pooling operations. The feature nonlinear transformation and iterative optimization are completed by the ReLU activation function; finally, the dimensionality mapping of the fused deep features is performed through a fully connected layer to output the battery dynamic charging current upper limit sequence at multiple sampling times in the future.

[0095] Optionally, the battery time-series data includes time-series data of multiple state parameters. The battery time-series data is converted into a two-dimensional matrix, specifically including:

[0096] The time series data of each state parameter are used as a row of a two-dimensional matrix;

[0097] The data corresponding to all state parameters at the same sampling time are arranged in parameter order to form a column of a two-dimensional matrix.

[0098] For example, following the example scenario in step 101, the system sampling period is set to The current time is recorded as Predicting the length of the time domain (That is, predicting the next 50ms), the 2D CNN model has been trained offline using the historical capacity data of the substation over the past year and deployed in the NPU of the embedded processor. The execution process of step 102 is as follows:

[0099] Extracting battery timing data from the most recent M cycles: Reading the most recent 5 sampling cycles from the historical data buffer ( The data per cycle includes battery-related parameters. Each cycle contains four key state parameters: battery voltage. Battery current Battery temperature and battery state of charge Assume the collected data is as shown in Table 1 below:

[0100] Table 1 Battery Timing Data Table

[0101]

[0102] The battery timing data is converted into a two-dimensional matrix: the matrix is ​​organized with parameters as rows and time as columns, with a total of 4 state parameters and a time step of 5, generating a... Two-dimensional matrix , As shown below:

[0103] ;

[0104] The matrix It is considered as a grayscale image with 4 channels and a width of 5, adapted to the standard input format of 2D CNN;

[0105] Inputting the above matrix into a 2D CNN model for forward propagation: The input is a 2D CNN model already loaded into the NPU. The calculations for each layer of the model are as follows:

[0106] Input layer: Receiver matrix ;

[0107] The first convolutional layer consists of 32 3×3 convolutional kernels with a stride of 1 and padding of 0. The calculated output feature map size is 2×3×32.

[0108] ReLU activation: introduces nonlinearity;

[0109] Max pooling: 2×2 pooling, output 1×1×32;

[0110] The second convolutional layer consists of 64 3×3 convolutional kernels, which further extract higher-order features.

[0111] ReLU+ global average pooling: compresses each feature map into a single value, resulting in a 64-dimensional vector;

[0112] Fully connected layer (128) + Dropout: Dimensionality reduction and prevention of overfitting;

[0113] Output layer: 5 neurons, each corresponding to one of the next 5 time points ( to 5) The predicted upper limit of the charging current;

[0114] After numerical calculations (actually matrix multiplication and nonlinear functions), the model outputs a vector of length 5:

[0115] ;

[0116] This is a sequence of upper limits for charging current over the next five moments. Its physical meaning is: based on the battery's state evolution over the past 50ms, the model infers that the battery can safely withstand the maximum charging current within the next 10ms. The charging current; the upper limit of the safe current within the second 10ms is The third safe current limit within 10ms is The upper limit of safe current within the fourth 10ms is The fifth safe current limit within 10ms is: As the state of charge (SOC) gradually increases or the temperature may rise, the upper limit of the safe current tends to decrease.

[0117] Output and temporarily store the prediction results: The predicted upper limit sequence of charging current will be output. Stored in a designated memory region of the embedded processor for use by the cost function in subsequent step 104 (determining the optimal control strategy); simultaneously, the currently constructed two-dimensional matrix... It is also retained in the history buffer for constructing a new image in the next cycle (using a sliding window mechanism, discarding the oldest column and adding the newest column each time).

[0118] Step 103: Update the current state of the digital twin model based on the current running data in the current battery time series data to obtain the model update state;

[0119] Current operating data refers to the latest set of state parameters collected in real time from the power management system, after analog-to-digital conversion and filtering. It is the data vector corresponding to the latest sampling time in the current battery time-series data, and includes at least key operating condition data such as bus voltage, load current, battery charge / discharge current, battery state of charge (SOC), battery state of health (SOH), and battery temperature.

[0120] The current state of a digital twin model refers to a set of state variables maintained by each sub-model within the digital twin, used to characterize the operating status of the virtual power system at a certain moment. Specifically, it includes:

[0121] The state of the virtual power source model (ST-GCN) typically includes the feature matrices of each node within the historical window, the hidden states of each layer of the graph convolutional network (e.g., node embedding vectors retained in a graph neural network), and internal memory units for recursive prediction (if a recursive structure is used). In the offline training and online deployment of this invention, ST-GCN is designed to receive data from a fixed-length historical window and output future predictions; therefore, its state can be simplified to a sequence of system features from the most recent τ time steps, which is continuously updated with new data.

[0122] The state of the virtual battery model (2D CNN): Since 2D CNN is a feedforward network, it has no internal recursive states. Its state mainly refers to the historical data buffer (battery data from the most recent M cycles) needed to construct the input image and the model parameters themselves. Updating the state means updating this buffer;

[0123] The state of the virtual controller includes the internal time counter, the last output control command, the integral error (if any), and the intermediate buffer of the policy pool simulation.

[0124] The model update state refers to the new virtual state reached by the digital twin after the current running data has been synchronized and updated. This state reflects all known information about the physical system up to the current moment and can serve as the starting point for subsequent parallel forward simulations. The model update state is an accurate mirror image of physical reality in the virtual world.

[0125] Optionally, the ST-GCN model requires the system feature matrix of the most recent τ time steps as input, while the 2D CNN model requires battery data from the most recent M cycles to construct the image. Step 103, while injecting the currently running data, discards the oldest time step data from the historical buffer and adds the new data to the end, achieving a rolling update of the sliding window. This operation ensures that the model input continuously contains the latest system dynamics, balancing temporal locality and computational efficiency.

[0126] In some embodiments, following the examples of steps 101 and 102, the system sampling period is set to... The current time is recorded as The digital twin has loaded the ST-GCN model parameters and 2D CNN model parameters from Flash, and the initial state of the virtual controller has been reset. Therefore, the execution flow of step 103 is as follows:

[0127] Obtain current running data: the current time read from step 101. The currently running data is extracted from the real-time state vector. Based on the previous example, this data includes:

[0128] bus voltage (48V lower than the rated value);

[0129] Battery voltage ;

[0130] Load current ;

[0131] Battery current ;

[0132] Battery state of charge ;

[0133] Battery health status ;

[0134] Battery temperature ;

[0135] Update the state of the virtual power model (ST-GCN): The virtual power model internally maintains a state of length... A historical feature sequence buffer (corresponding to 50ms, sampling period 10ms) stores one feature vector at each time step. This includes the rectifier module output current, individual battery cell voltages, load current, and bus voltage; Previously, the buffer already contained The feature vectors at six time points, where, This represents the corrected or predicted value at the end of the previous period; then, eigenvectors Write to the last position of the buffer, while discarding the oldest. eigenvectors The updated buffer contents are: , , , , , This buffer constitutes the standard input historical data window of the ST-GCN model. (in ST-GCN employs a non-recursive, purely spatiotemporal convolutional structure, which does not rely on gated recurrent units (GRU / LSTM), thus eliminating the need for additional network hidden state resets. After the historical feature sequence buffer is updated, the model update state is uniquely determined by the latest historical data window, as shown below: ;

[0136] Updating the state of the virtual battery model (2D CNN): The virtual battery model internally maintains a... A two-dimensional matrix buffer, wherein, , This buffer stores battery voltage, current, temperature, and SOC data for the most recent 5 cycles. Previously, the buffer already contained to The four columns of data will display the battery data at the current moment. As a new column, it is added to the end of the buffer (the rightmost column), while the leftmost column is discarded. Column data; updated The matrix is ​​equivalent to a two-dimensional image of battery data, denoted as The state of the virtual battery model is the updated state. The updated matrix's internal image data fully reflects the actual evolution of the physical battery over the past five cycles.

[0137] Update the state of the virtual controller, including:

[0138] The virtual controller has a simulated step counter inside. This counter is used to record the number of steps executed in the virtual simulation. At the start of each physical sampling period, this counter is reset to 0, indicating that a new round is about to begin. Forward simulation;

[0139] The virtual controller also caches the charging current command that was last applied in the previous cycle. The cache is updated to reflect the instructions actually being executed by the physical system before the start of the current cycle (i.e., the last instruction output in the previous cycle, which can be obtained from feedback from the physical driver). Assuming the charging current was [value missing] at the end of the previous cycle... (In the floating charge state), the virtual controller updates its internal variables as follows: ;

[0140] After the above three-step sub-model state update operation is completed, all sub-models in the digital twin (virtual power model, virtual battery model, and virtual controller) have been updated using the physical system. The current running data is synchronized at any given time. The state of the digital twin at this point is called the model update state, denoted as:

[0141] ;

[0142] in, For virtual controllers in The internal state at any given moment.

[0143] Step 104: Based on the updated model update state, perform parallel forward simulation on the pre-generated strategy pool using the updated digital twin model, and determine and execute the optimal control strategy by combining the charging current upper limit sequence.

[0144] The pre-generated strategy pool refers to the set of candidate charging current sequences pre-constructed during the offline phase. Each candidate strategy is a sequence of length [length missing]. Discrete current sequences (in the prediction time domain) represent the future. The charging current command to be applied within each sampling period. The policy pool can contain various types, such as uniform charging policy, exponential decay charging policy, and piecewise constant policy, and can also contain no-charging policy (all-zero sequence). The policy pool is generated and stored offline, and does not need to be recalculated when called online.

[0145] Parallel forward simulation refers to the independent and synchronous (computationally accelerated using multi-core or vector instruction methods) simulation of a length of [length missing] for each candidate control policy in the policy pool within a digital twin model. The simulation is performed using a step-by-step recursive prediction method. In each step, the virtual battery model predicts the battery terminal voltage based on the current battery state and candidate charging current, while the virtual power supply model (ST-GCN) predicts the system state (bus voltage, load current, SOC, etc.) at the next moment based on the current system state, charging current command, and battery voltage. The simulations of all strategies start from the same model update state (i.e., the virtual-real synchronization state output in step 103) and do not interfere with each other.

[0146] Optionally, step 104 involves performing parallel forward simulation on each candidate control strategy in the pre-generated strategy pool based on the updated digital twin model state, generating the system response trajectory corresponding to each strategy, completing the safety constraint verification by combining the upper limit sequence of charging current, and selecting the optimal control strategy according to the comprehensive cost rule, and issuing its current instruction to the physical system for execution.

[0147] In some embodiments, the pre-generated policy pool includes multiple candidate control policies, and step 104 may include the following sub-steps:

[0148] Each candidate control strategy is input into the updated digital twin model. Based on the model update state, the operation process of the power management system under each candidate control strategy is simulated, and the system response trajectory data corresponding to each candidate control strategy is output.

[0149] Based on system response trajectory data, current limit constraint determination is performed by combining the charging current upper limit sequence, and the comprehensive cost of each candidate control strategy is calculated according to the preset comprehensive cost rule.

[0150] The comprehensive costs of all candidate control strategies are ranked, and the candidate control strategy with the minimum comprehensive cost is determined as the optimal control strategy and executed.

[0151] The system response trajectory refers to the completion of a candidate policy. After forward simulation, the system's key state variables are recorded at various future time points. These include at least the bus voltage sequence, battery SOC sequence, and optional other state variables (such as load current). This trajectory is used to evaluate the performance of candidate control strategies.

[0152] The overall cost is a quantitative indicator used to evaluate the merits of a candidate strategy; a lower overall cost indicates a better strategy. The overall cost function defined in this invention includes the following three penalty terms:

[0153] Voltage deviation penalty: The absolute value of the maximum deviation of the predicted bus voltage from the reference value (e.g., 48V), reflecting the voltage holding capability;

[0154] Current variation penalty term: The variance of the charging current sequence is penalized for drastic current fluctuations, thus smoothing the control.

[0155] SOH Safety Penalty: Based on the charging current upper limit sequence predicted in step 102, check whether there are current commands in the candidate control strategies that exceed the safety upper limit. If so, apply a positive penalty to protect battery health.

[0156] In some embodiments, the overall cost is calculated using the following formula:

[0157] ;

[0158] ;

[0159] in, For the first The overall cost-effectiveness of each candidate control strategy For the time-series index of the simulation prediction step size, This represents the predicted total number of steps in a single-round strategy simulation. This is a voltage deviation penalty item. The weighting coefficients, Penalty term for current change The weighting coefficients, The weighting coefficient for the battery health status penalty term. For the future obtained through digital twin model simulation The predicted value of the bus voltage of the step, The preset bus voltage reference target value, For the first The charging current sequence corresponding to each candidate control strategy This is the variance calculation function, characterizing the degree of fluctuation in the current sequence. For the first Among the candidate control strategies, the first one is... The charging current command value for the step. For the future number corresponding to the upper limit sequence of charging current The upper limit of the charging current of the step.

[0160] In some embodiments, the digital twin model includes a virtual battery model and a virtual power supply model. The aforementioned input of each candidate control strategy into the updated digital twin model, simulation of the power management system operation under each candidate control strategy based on the model update state, and output of system response trajectory data corresponding to each candidate control strategy, specifically includes:

[0161] Analyze each candidate control strategy and determine the charging control parameters for each candidate control strategy;

[0162] The charging control parameters and model update status are input into the virtual battery model to perform battery terminal voltage prediction processing and output the predicted battery terminal voltage values ​​at multiple consecutive simulation times.

[0163] The predicted battery terminal voltage, model update status, and charging control parameters are input into the virtual power model to simulate the changes in the operating status of the power management system within a preset time period, and the system operating status data at each simulation time are output.

[0164] The system operation status data at all simulation moments are integrated in time series to generate system response trajectory data.

[0165] A virtual battery model refers to a sub-model in a digital twin used to simulate the charging and discharging response characteristics of a battery. This invention employs a battery terminal voltage response model built on a 2D CNN (distinct from the 2D CNN model used in step 102 to predict the upper limit of the charging current). This model receives the current battery state (including SOC, SOH, and temperature) and the proposed charging current command, and outputs the corresponding predicted value of the battery terminal voltage. This model also uses a two-dimensional convolutional neural network architecture, but its training objective is to learn the dynamic mapping of current command → terminal voltage response, rather than learning the upper limit of the charging current.

[0166] In some embodiments, a spatiotemporal graph convolutional network (ST-GCN) is used as a virtual power supply model. This virtual power supply model models the dynamics of the communication power supply system. It captures the electrical topology coupling relationship through spatial convolution and captures the temporal evolution law of the operating conditions through temporal convolution, thereby achieving accurate prediction of the future state of the power supply system and replacing the traditional mechanism model based on differential equations.

[0167] Specifically, the communication power supply system is abstracted as a spatiotemporal graph structure. The node set covers all key equipment units, the edge set replicates the actual electrical connection relationships, and the node feature matrix integrates multi-dimensional operating parameters, realizing a digital representation of the power supply system's physical topology and operating status. This provides a structured data foundation for subsequent spatiotemporal convolution operations. The structure is as follows:

[0168] Node definition: The rectifier module, each battery cell, each load branch, and the bus in the communication power system are defined as nodes in the space-time diagram, and a node set is constructed. ;

[0169] Edge relationship definition: Define edge sets based on actual electrical connection relationships. The bus node is connected to all other nodes, and each battery cell node is connected in series in sequence, completely replicating the electrical topology of the power system.

[0170] Node feature matrix construction: Collect operational data such as voltage, current, temperature, SOC, and SOH of each node to construct the node feature matrix. ,in, For the number of nodes, The feature dimensions for each node.

[0171] ST-GCN gradually fuses the global spatiotemporal information of the system through multi-layer spatiotemporal convolution operations. The multi-layer spatiotemporal convolution operation is specifically represented by the following formula:

[0172] ;

[0173] in, For the first The hidden state matrix of the nodes in the first layer is derived from the first layer. The features of one layer are obtained after undergoing spatiotemporal convolution operations and become the input features for the next layer. For the first The hidden state matrix of the nodes in the layer has a dimension of , For the first The feature dimension of a layer contains the feature information of each node in the power system at the current layer. It is an adjacency matrix (including self-loops) with dimension . This is used to represent the electrical connections between nodes. Diagonal elements are set to 1 to preserve the node's own characteristics, while off-diagonal elements are set to 1 to indicate an electrical connection between nodes, and 0 to indicate no connection. It is a degree matrix with dimension 1. A diagonal matrix, where the diagonal elements are the connectivity degree of the corresponding node (i.e., the number of connections between the node and other nodes plus the number of self-loops), is used to normalize the adjacency matrix. The inverse square root of the degree matrix is ​​used to achieve symmetric normalization of the adjacency matrix, eliminating the influence of node degree differences on convolution operations. For the first The spatial convolutional weight matrix of the layer is a learnable parameter with dimension . Used to capture the spatial electrical coupling characteristics between nodes. For the first The temporal convolutional weight matrix of the layer is a learnable parameter with dimension . ( (where the time convolution kernel size is used), and through one-dimensional convolution operations along the time dimension, the temporal variation trend of operating condition data is captured. As the activation function, this embodiment uses the ReLU function to introduce nonlinear transformation and enhance the model's ability to fit the complex dynamic characteristics of the power system.

[0174] Using future bus voltage and battery status as monitoring signals, historical operating data of the communication power supply system under normal, full-capacity, and fault conditions are collected to train the ST-GCN model offline. The model's forward propagation function is:

[0175] ;

[0176] in, It is the predicted sequence output by the model, representing the predicted system state within a future time window. It is a sequence of lengths... A collection of time series data, containing data from time 1000 to 1000. arrive In this embodiment, the predicted feature values ​​for all key nodes include at least the bus voltage sequence and the battery. sequence, The current time refers to the last moment in the input data window, which is also the baseline time for the model to make predictions. For the prediction time domain (prediction steps), it represents the length of the future time window that the model needs to predict, assuming This means predicting the state over the next 5 sampling periods. is the forward propagation function of the spatiotemporal graph convolutional network model, representing the entire inference process of the neural network. It receives historical spatiotemporal graph data as input, performs multiple layers of spatiotemporal convolution operations, and finally outputs a prediction of the future state. The historical spatiotemporal graph data window input to the model contains past operating state information of the power system and is a window of length [missing information]. Time series data, containing data from time point up to the current moment The set of feature matrices of all nodes The historical window length (backtracking steps) represents the number of past sampling periods included in the input data. Assuming... This corresponds to 5 sampling periods (total duration 50ms), meaning the model uses data from the past 5 periods to predict the future state.

[0177] During the ST-GCN model training process, the model parameters are optimized through the backpropagation algorithm, enabling the model to accurately predict the operating status of future key nodes based on historical spatiotemporal graph data. After training, the model can be used as a data-driven prediction model to replace the traditional differential equation mechanism model, realizing rapid simulation and prediction of power system dynamics.

[0178] In some embodiments, taking the scenario of switching from floating charging to core capacity charging of communication power supply batteries as an example, the parallel simulation and optimal strategy selection process in step 104 is executed as follows:

[0179] Parallel forward simulation initialization: Based on the updated digital twin model state from step 103 To simulate the initial state, K groups of candidate charging current sequences (each group having a length of...) will be pre-generated from the policy pool. The instruction sequence of each step (i.e., the candidate control strategy) is input into the digital twin model to start parallel simulation;

[0180] Single-step timing simulation derivation (taking a single candidate control strategy as an example) includes:

[0181] Initialize simulation state: Let Reset the simulation step counter ;

[0182] Extract current step instruction: Obtain candidate control strategy The One charging current command ;

[0183] Virtual battery model response prediction: Using the current battery state (battery voltage, current, temperature, SOC data) and the voltage response at the charging current command input terminal 2D CNN, predict the battery terminal voltage. ;

[0184] Virtual power model system prediction: predicting the current simulation state Charging current command Battery terminal voltage Input an ST-GCN model, perform spatiotemporal convolution operations, and output the system state vector at the next time step. , This includes the predicted bus voltage at the next moment. Load current Key node state variables such as battery SOC;

[0185] Record simulation trajectory: Store data such as bus voltage, battery terminal voltage, and battery SOC of the current step into the trajectory set. ;

[0186] Update simulation state: Let , Repeat the above steps until completion. Step-by-step simulation;

[0187] Comprehensive cost calculation and constraint verification, including:

[0188] After the simulation is completed, perform the following operations for each candidate control strategy:

[0189] Current limit constraint determination: Based on the upper limit sequence of charging current predicted in step 102, the strategy current command is checked step by step, and the SOH penalty term (battery health state penalty term) is calculated.

[0190] Multi-objective cost calculation: Based on simulation trajectory data, voltage deviation penalty term and current change penalty term are calculated, and combined with preset weighting coefficients, the comprehensive cost value of each candidate control strategy is obtained;

[0191] Finally, the comprehensive cost of all candidate control strategies is sorted, the index of the strategy with the lowest cost is determined, the current command of the strategy at the current moment is extracted, and it is sent to the physical power controller for execution. Subsequent commands are determined by resimulating the latest operating conditions in the next cycle.

[0192] In some embodiments, the above-described intelligent power supply cooperative control method based on digital twins may further include the following steps:

[0193] Obtain the historical state feature vector of the power management system, the current current alarm threshold, and the current maximum charging current limit;

[0194] Calculate the current state feature vector based on the model update state;

[0195] Calculate the historical feature difference vector based on the historical state feature vector and the current state feature vector;

[0196] By concatenating the historical feature difference vector and the current state feature vector, a collaborative feature vector is generated;

[0197] The collaborative feature vector is input into the multilayer perceptron neural network, which outputs the first feature parameter and the second feature parameter.

[0198] Based on the first characteristic parameter and the current maximum charging current, the operating state of the digital twin model is updated and used as the initial model state for the next control simulation cycle; the start time of the next control simulation cycle is the end time of the execution of the optimal control strategy.

[0199] The current current alarm threshold is updated based on the second feature parameter.

[0200] Historical state feature vectors refer to normalized feature vectors stored in the local historical feature buffer and calculated from the state update of the model in the previous control cycle. The dimension is consistent with the current state feature vector. It integrates operating condition information such as the power system bus voltage, battery current, state of charge, and health status in the previous cycle, and is the basic data for calculating the trend of operating condition changes.

[0201] The current alarm threshold refers to the current anomaly judgment threshold preset by the power management system, which is used to identify faults such as excessive battery charging and discharging current and circuit abnormalities in real time. Traditional solutions use a fixed threshold, while this embodiment dynamically adjusts it according to operating conditions to avoid false alarms in scenarios such as capacity limits and float charging switching.

[0202] The current maximum charging current limit refers to the maximum allowable charging current of the battery determined by combining the battery health status and the system bus operating conditions. It is the specific value corresponding to the current moment in the upper limit sequence of charging current predicted by the two-dimensional convolutional neural network model in step 102, or it is preset based on human experience. It is used to constrain the safe amplitude of the charging current command to prevent the battery from overcharging.

[0203] The current state feature vector refers to the standardized feature vector extracted based on the updated state of the digital twin model. Through nonlinear transformation (such as tanh, σ function), state parameters such as bus voltage deviation, battery current, SOC, and SOH are mapped to the interval [-1,1] or [0,1], providing normalized input for collaborative decision-making.

[0204] Historical feature difference vectors refer to the sequence of differences between the current state feature vector and the historical feature vectors. They represent the dynamic change trend of the working conditions, provide temporal context information for multilayer perceptron neural networks (MLP), and enhance the robustness of decision-making.

[0205] The collaborative feature vector is a feature vector formed by concatenating the current state feature vector and the historical feature difference vector. It integrates the current working condition and historical change trends and serves as the input of the collaborative MLP model.

[0206] Multilayer perceptron (MLP) refers to a lightweight feedforward neural network used for nonlinear mapping in a virtual feature space. In this embodiment, the cooperative MLP takes a cooperative feature vector as input and outputs two types of feature parameters: a first feature parameter (used for fine-tuning the charging current) and a second feature parameter (used for dynamic adjustment of the alarm threshold), thereby achieving cooperative optimization of control strategy and safety logic.

[0207] The first characteristic parameter is the parameter output by the MLP model used to correct the charging current. After inverse transformation and amplitude constraint, the final target charging control current is obtained, realizing fine adjustment of the optimal control strategy.

[0208] The second feature parameter is a parameter output by the MLP model used to dynamically adjust the inner loop alarm threshold. It generates an alarm threshold that adapts to the current operating conditions through a sigmoid function mapping, thus eliminating the false alarm problem of traditional fixed thresholds.

[0209] After selecting the optimal control strategy and executing the charging current command at the current moment, the above steps fine-tune the charging current command and dynamically update the alarm threshold through a virtual feature space collaborative decision-making mechanism. At the same time, the current control effect is fed back to the digital twin to provide an accurate initial state for the simulation of the next cycle.

[0210] In some embodiments, updating the operating state of the digital twin model based on the first characteristic parameter and the current maximum charging current specifically includes the following sub-steps:

[0211] The first characteristic parameter is inversely transformed to obtain the preliminary charging current value;

[0212] Based on the current maximum charging current limit, the initial charging current value is constrained to determine the target charging control current.

[0213] According to the preset timing state iteration update rules, the operating state of the digital twin model is updated based on the target charging control current.

[0214] The timing state iteration update rule refers to the state deduction logic preset by the digital twin model. It uses the target charging control current as the input command to drive the virtual power supply model, virtual battery model, and virtual controller to run synchronously, complete the single-step working condition iteration, and obtain the updated model running state.

[0215] The above steps drive the digital twin model with the target charging control current, update the model state according to the preset timing deduction rules, update the model state to the end time of the optimal strategy execution, so that the initial state of the next simulation cycle is completely synchronized with the real-time state of the physical system, avoid model state drift, ensure the accuracy of subsequent simulation optimization, and realize the timing closed loop of the virtual and real system.

[0216] In some embodiments, taking the scenario of switching from float charging to full capacity charging of a communication power supply battery as an example, the virtual feature space collaborative decision-making and state update process is executed, specifically including:

[0217] Read the historical state feature vector retained from the previous control cycle from the local storage unit of the power management system. The system's current alarm threshold and the maximum current limit predicted in real time by a two-dimensional convolutional neural network;

[0218] Based on the digital twin model update in step 103, the state parameters of each node in the model are extracted, and the current state feature vector is calculated through feature mapping. The specific formula is as follows:

[0219] ;

[0220] in, For feature mapping function, For the Sigmoid function, To update the state of the model, This represents the measured current bus voltage value of the power management system. The preset bus voltage reference target value, is the voltage deviation scaling factor, a learnable hyperparameter used to adjust the voltage deviation within... The scaling factor of the function input interval controls the sensitivity of the feature mapping. This is a hyperbolic tangent activation function used to map the input to the [-1,1] interval, thereby normalizing the voltage deviation and current parameters. Its positive or negative value reflects the direction of the deviation. This represents the measured current of the physical battery. is the current scaling factor, a learnable hyperparameter used to adjust the current at... The scaling factor of the function input interval controls the mapping sensitivity of the current characteristics. This represents the current state of charge (SOC) of the physical battery, i.e., the percentage of remaining battery charge. The value ranges from [0,1], with a value of 1 indicating a fully charged battery. for The scaling factor is a learnable hyperparameter used to adjust... exist The scaling factor of the function input range. This represents the current health status of the physical battery, i.e., the degree of battery aging. Its value ranges from [0,1], with a value of 1 indicating a brand new battery. This is a scaling factor for aging, a learnable hyperparameter used to adjust the mapping weights of aging metrics. Indicates vector transpose;

[0221] Read the historical state feature vector With the current state feature vector Perform element-wise interpolation to obtain the historical feature difference vector. The calculation formula is as follows:

[0222] ;

[0223] Following the order of vector dimensions, the historical feature difference vector is concatenated with the current state feature vector to integrate static working condition information and dynamic trend information, generating a collaborative feature vector. This vector has a dimension twice that of the original feature vector, which meets the input dimension requirements of the multilayer perceptron neural network.

[0224] The collaborative feature vector is input into a multilayer perceptron (MLP) that has been trained offline and whose parameters have been fixed. After the network goes through fully connected layer operations and nonlinear activation, it outputs two types of output quantities in parallel: the first feature parameter and the second feature parameter.

[0225] The digital twin model's running state is updated based on the first feature parameter, including:

[0226] Perform an inverse normalization transformation on the first characteristic parameter to restore the dimensionless characteristic value output by the network to physical dimensions, and obtain the preliminary charging current value.

[0227] The initial charging current value is compared with the current maximum charging current value. If the initial charging current value exceeds the limit range, the current maximum charging current value is used as the upper limit for truncation constraint, and the target charging control current is finally determined.

[0228] The system invokes its preset timing state iteration update rules, inputting the target charging control current as a control command into the digital twin model. The virtual battery model, virtual power supply model, and virtual controller within the model respond synchronously, completing a single-step operating condition simulation to obtain the iterative model operating state. This state serves as the initial model state for the next control simulation cycle, and the start time of the next control cycle is consistent with the end time of the current optimal control strategy, achieving seamless timing transitions.

[0229] Using the second feature parameter output by the multilayer perceptron neural network, the current current alarm threshold is adaptively corrected to generate a new current alarm threshold that is adapted to the current core capacity switching condition. The new threshold is written into the alarm configuration unit of the power management system to replace the original fixed threshold for current anomaly determination in the next cycle.

[0230] The current state feature vector calculated in this cycle is stored in the historical feature buffer, overwriting the data from the previous cycle, and used as the historical state feature vector required for the next round of process calculation, thus completing the entire process closed loop.

[0231] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0232] The following describes the intelligent power supply cooperative control device based on digital twins provided in the embodiments of this application. The intelligent power supply cooperative control device based on digital twins described below can be referred to in correspondence with the intelligent power supply cooperative control method based on digital twins described above.

[0233] Reference Figure 2 , Figure 2 This is an optional structural diagram of an intelligent power supply cooperative control device based on digital twins provided in an embodiment of the present invention. This device is used to implement the aforementioned intelligent power supply cooperative control method based on digital twins, and may include:

[0234] The data acquisition module 200 is used to acquire the current battery timing data and digital twin model of the power management system;

[0235] The current prediction module 300 is used to input the current battery timing data into a pre-trained convolutional neural network to predict the upper limit sequence of charging current within a preset future time step.

[0236] The state update module 400 is used to update the current state of the digital twin model based on the current running data in the current battery time series data, so as to obtain the model update state;

[0237] The strategy control module 500 is used to perform parallel forward simulation of the pre-generated strategy pool based on the updated digital twin model and the model update state, and to determine and execute the optimal control strategy in combination with the charging current upper limit sequence.

[0238] In some embodiments, the battery timing data includes timing data of multiple state parameters, and the current prediction module 300 is configured as follows:

[0239] Convert the current battery timing data into a two-dimensional matrix;

[0240] The two-dimensional matrix is ​​input into the convolutional neural network to perform convolutional feature extraction and fully connected layer prediction, and outputs a sequence of charging current upper limits within a preset future time step.

[0241] Convert battery time-series data into a two-dimensional matrix, including:

[0242] The time series data of each state parameter are used as a row of a two-dimensional matrix;

[0243] The data corresponding to all state parameters at the same sampling time are arranged in parameter order to form a column of a two-dimensional matrix.

[0244] In some embodiments, the digital twin model includes a virtual battery model and a virtual power supply model, and the policy control module 500 is configured as follows:

[0245] Each candidate control strategy is input into the updated digital twin model. Based on the model update state, the operation process of the power management system under each candidate control strategy is simulated, and the system response trajectory data corresponding to each candidate control strategy is output.

[0246] Based on system response trajectory data, current limit constraint determination is performed by combining the charging current upper limit sequence, and the comprehensive cost of each candidate control strategy is calculated according to the preset comprehensive cost rule.

[0247] The comprehensive costs of all candidate control strategies are ranked, and the candidate control strategy with the minimum comprehensive cost is determined as the optimal control strategy and executed.

[0248] Each candidate control strategy is input into the updated digital twin model. Based on the model update state, the operation process of the power management system under each candidate control strategy is simulated, and the system response trajectory data corresponding to each candidate control strategy is output, including:

[0249] Analyze each candidate control strategy and determine the charging control parameters for each candidate control strategy;

[0250] The charging control parameters and model update status are input into the virtual battery model to perform battery terminal voltage prediction processing and output the predicted battery terminal voltage values ​​at multiple consecutive simulation times.

[0251] The predicted battery terminal voltage, model update status, and charging control parameters are input into the virtual power model to simulate the changes in the operating status of the power management system within a preset time period, and the system operating status data at each simulation time are output.

[0252] The system operation status data at all simulation moments are integrated in time series to generate system response trajectory data.

[0253] In some embodiments, the digital twin-based intelligent power supply cooperative control device may further include a model iteration module, which is configured as follows:

[0254] Obtain the historical state feature vector of the power management system, the current current alarm threshold, and the current maximum charging current limit;

[0255] Calculate the current state feature vector based on the model update state;

[0256] Calculate the historical feature difference vector based on the historical state feature vector and the current state feature vector;

[0257] By concatenating the historical feature difference vector and the current state feature vector, a collaborative feature vector is generated;

[0258] The collaborative feature vector is input into the multilayer perceptron neural network, which outputs the first feature parameter and the second feature parameter.

[0259] Based on the first characteristic parameter and the current maximum charging current, the operating state of the digital twin model is updated and used as the initial model state for the next control simulation cycle; the start time of the next control simulation cycle is the end time of the execution of the optimal control strategy.

[0260] Update the current alarm threshold based on the second feature parameter;

[0261] Based on the first characteristic parameter and the current maximum charging current, update the operating state of the digital twin model, including:

[0262] The first characteristic parameter is inversely transformed to obtain the preliminary charging current value;

[0263] Based on the current maximum charging current limit, the initial charging current value is constrained to determine the target charging control current.

[0264] According to the preset timing state iteration update rules, the operating state of the digital twin model is updated based on the target charging control current.

[0265] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0266] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0267] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0268] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0269] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0270] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned intelligent power supply collaborative control method based on digital twins. This electronic device can be any smart terminal, including tablet computers.

[0271] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0272] Please see Figure 3 , Figure 3 The hardware structure of an electronic device according to another embodiment is illustrated, including:

[0273] The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention.

[0274] The memory 902 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 902 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and is called and executed by the processor 901 to implement the intelligent power cooperative control method based on digital twins of the embodiments of this invention.

[0275] The 903 input / output interface is used to implement information input and output.

[0276] The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0277] Bus 905 transmits information between various components of the device (e.g., processor 901, memory 902, input / output interface 903, and communication interface 904);

[0278] The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0279] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned intelligent power supply collaborative control method based on digital twins.

[0280] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0281] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0282] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent power supply collaborative control based on digital twins, characterized in that, include: Acquire the current battery timing data and digital twin model of the power management system; The current battery timing data is input into a pre-trained convolutional neural network to predict the upper limit sequence of charging current within a preset future time step. The current state of the digital twin model is updated based on the current running data in the current battery time series data to obtain the model update state; the digital twin model includes a virtual battery model and a virtual power supply model; The virtual power supply model is a spatiotemporal graph convolutional neural network, used to model the dynamics of communication power supply systems. It captures electrical topology coupling relationships through spatial convolution and captures the time-series evolution of operating conditions through temporal convolution, predicting the future state of the power supply system. Based on the updated digital twin model and the updated state of the model, a parallel forward simulation is performed on the pre-generated strategy pool, and the optimal control strategy is determined and executed in combination with the charging current upper limit sequence. The pre-generated policy pool includes multiple candidate control policies. The process involves performing parallel forward simulations on the pre-generated policy pool based on the updated digital twin model's state update, and determining and executing the optimal control policy in conjunction with the charging current upper limit sequence. This includes: Each of the candidate control strategies is input into the updated digital twin model. Based on the updated state of the model, the operation process of the power management system under each candidate control strategy is simulated, and the system response trajectory data corresponding to each candidate control strategy is output. Based on the system response trajectory data, current limit constraint determination is performed in conjunction with the charging current upper limit sequence, and the comprehensive cost of each candidate control strategy is calculated according to the preset comprehensive cost rule. The comprehensive costs of all candidate control strategies are ranked, and the candidate control strategy with the minimum comprehensive cost is determined as the optimal control strategy and executed.

2. The intelligent power supply cooperative control method based on digital twin according to claim 1, characterized in that, The step of inputting the current battery timing data into a pre-trained convolutional neural network to predict the upper limit sequence of charging current within a preset future time step includes: Convert the current battery timing data into a two-dimensional matrix; The two-dimensional matrix is ​​input into a convolutional neural network to perform convolutional feature extraction and fully connected layer prediction, and outputs a sequence of charging current upper limits within a preset future time step.

3. The intelligent power supply collaborative control method based on digital twins according to claim 2, characterized in that, The battery time-series data includes time-series data of multiple state parameters. Converting the battery time-series data into a two-dimensional matrix includes: The time-series data of each state parameter is used as a row of a two-dimensional matrix; The data corresponding to all state parameters at the same sampling time are arranged in parameter order to form a column of a two-dimensional matrix.

4. The intelligent power supply cooperative control method based on digital twin according to claim 1, characterized in that, The process of inputting each of the candidate control strategies into the updated digital twin model, simulating the operation of the power management system under each candidate control strategy based on the updated state of the model, and outputting system response trajectory data corresponding to each candidate control strategy includes: Analyze each of the candidate control strategies and determine the charging control parameters of each candidate control strategy; The charging control parameters and the model update state are input into the virtual battery model to perform battery terminal voltage prediction processing and output the predicted battery terminal voltage values ​​at multiple consecutive simulation times. The predicted battery terminal voltage, the model update status, and the charging control parameters are input into the virtual power model to simulate the changes in the operating status of the power management system within a preset time period, and the system operating status data at each simulation time are output. The system operation status data at all simulation moments are integrated in time series to generate system response trajectory data.

5. The intelligent power supply cooperative control method based on digital twin according to claim 1, characterized in that, The overall cost is calculated using the following formula: ； ； in, For the first The overall cost-effectiveness of each candidate control strategy For the time-series index of the simulation prediction step size, This represents the predicted total number of steps in a single-round strategy simulation. This is a voltage deviation penalty item. The weighting coefficients, Penalty term for current change The weighting coefficients, The weighting coefficient for the battery health status penalty term. For the future obtained through digital twin model simulation The predicted value of the bus voltage of the step, The preset bus voltage reference target value, For the first The charging current sequence corresponding to each candidate control strategy This is the variance calculation function, characterizing the degree of fluctuation in the current sequence. For the first Among the candidate control strategies, the first one is... The charging current command value for the step. For the future number corresponding to the upper limit sequence of charging current The upper limit of the charging current of the step.

6. The intelligent power supply cooperative control method based on digital twin according to claim 1, characterized in that, Also includes: Obtain the historical state feature vector of the power management system, the current current alarm threshold, and the current maximum charging current limit; The current state feature vector is calculated based on the updated state of the model. Calculate the historical feature difference vector based on the historical state feature vector and the current state feature vector; The historical feature difference vector and the current state feature vector are concatenated to generate a collaborative feature vector; The collaborative feature vector is input into a multilayer perceptron neural network, which outputs a first feature parameter and a second feature parameter. Based on the first feature parameter and the current maximum charging current, the operating state of the digital twin model is updated and used as the initial model state for the next control simulation cycle; the start time of the next control simulation cycle is the end time of the execution of the optimal control strategy. The current current alarm threshold is updated based on the second feature parameter.

7. The intelligent power supply cooperative control method based on digital twin according to claim 6, characterized in that, The step of updating the operating state of the digital twin model based on the first feature parameter and the current maximum charging current includes: The first characteristic parameter is inversely transformed to obtain the preliminary charging current value; Based on the current maximum charging current value, the amplitude of the initial charging current value is constrained to determine the target charging control current; According to the preset timing state iteration update rules, the operating state of the digital twin model is updated based on the target charging control current.

8. A smart power supply collaborative control device based on digital twins, characterized in that, The method for performing the intelligent power supply cooperative control method based on digital twins as described in any one of claims 1 to 7 includes: The data acquisition module is used to acquire the current battery timing data and digital twin model of the power management system; The current prediction module is used to input the current battery timing data into a pre-trained convolutional neural network to predict the upper limit sequence of charging current within a preset future time step. The state update module is used to update the current state of the digital twin model based on the current running data in the current battery time series data to obtain the model update state; the digital twin model includes a virtual battery model and a virtual power supply model; the virtual power supply model is a spatiotemporal graph convolutional neural network, used to model the dynamics of the communication power supply system, and captures the electrical topology coupling relationship through spatial convolution and the evolution law of the operating condition time series through temporal convolution to predict the future state of the power supply system; The strategy control module is used to perform parallel forward simulation on the pre-generated strategy pool based on the updated state of the updated digital twin model, and determine and execute the optimal control strategy in combination with the charging current upper limit sequence; the pre-generated strategy pool includes multiple candidate control strategies. The policy control module is configured as follows: Each of the candidate control strategies is input into the updated digital twin model. Based on the updated state of the model, the operation process of the power management system under each candidate control strategy is simulated, and the system response trajectory data corresponding to each candidate control strategy is output. Based on the system response trajectory data, current limit constraint determination is performed in conjunction with the charging current upper limit sequence, and the comprehensive cost of each candidate control strategy is calculated according to the preset comprehensive cost rule. The comprehensive costs of all candidate control strategies are ranked, and the candidate control strategy with the minimum comprehensive cost is determined as the optimal control strategy and executed.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the intelligent power supply cooperative control method based on digital twins as described in any one of claims 1 to 7.

Images