A mobile power generation vehicle grid-connected state prediction and cooperative control method

Abstract

The application relates to a mobile power generation vehicle grid-connected state prediction and collaborative control method, steps of which comprise high-precision space-time synchronization, dynamic topology feature extraction, accurate grid-connected state prediction, communication delay prospective compensation, VSG parameter adaptive adjustment and event-triggered sparse communication. The application is based on the grid-connected method of space-time data fusion, machine learning prediction and adaptive collaborative control, and aims to solve the technical problems of insufficient synchronization accuracy, slow collaborative response speed and power oscillation of multiple mobile power generation vehicles during grid connection caused by factors such as clock asynchronization, communication delay and fixed control parameters in a dynamic power grid environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of new power systems, in particular to a mobile power car grid-connected state prediction and collaborative control method. BACKGROUND

[0002] Under the background of accelerating the construction of new power systems, mobile power cars, as an important flexible adjustment resource, play an irreplaceable role in grid emergency power supply, new energy consumption regulation and other scenarios. However, in the prior art, the method based on traditional droop control has inherent frequency and pressure difference problems; while the virtual synchronous generator (VSG) control method based on fixed parameters can provide inertia support, but when multiple machines are connected in parallel and the load suddenly changes, the inertia and damping coefficient cannot dynamically adapt to the grid state, which easily leads to power oscillation and even causes cascading trip-out. In the time and space synchronization level, the traditional time service scheme relying on a single positioning system has nanosecond-level time drift, and the clock difference between grid nodes leads to non-uniform data acquisition time scale, which cannot effectively handle the dynamic time and space correlation of mobile power cars. The traditional polling communication mechanism leads to serious waste of channel resources. Especially in the emergency repair scene, the full-load communication state will greatly increase the transmission delay of control commands, greatly restricting the dynamic response ability of the system. SUMMARY

[0003] The purpose of the embodiments of the present application is to provide a mobile power car grid-connected state prediction and collaborative control method, which can improve the prediction accuracy by fusing space and time features, and optimize the efficiency of multi-car collaborative control.

[0004] To achieve the above purpose, the present application provides the following technical scheme:

[0005] The embodiments of the present application provide a mobile power car grid-connected state prediction and collaborative control method, comprising the following specific steps:

[0006] Step 1: High-precision time and space synchronization

[0007] The fusion time service terminal of each power car simultaneously receives the Beidou satellite signal and the synchronization clock signal distributed through the 5G TSN network, and generates an initial time and space feature matrix;

[0008] Step 2: Dynamic topology feature extraction

[0009] The time and space feature matrix generated in step 1 is obtained, and the real-time electrical quantity collected by the local PMU is spliced to extract the deep space dependence features of the current grid in real time, and output the feature tensor F;

[0010] Step 3: Accurate grid-connected state prediction

[0011] The feature tensor F is immediately sent to the ST-Transformer prediction model to predict the dynamic trajectory of voltage, frequency, phase angle and their change rates in the next K time steps;

[0012] Step 4: Communication delay look-ahead compensation

[0013] Estimate the communication delay τ from the instruction issuance to the current reception, use the prediction result of step 3 to calculate the natural evolution amount ΔSpred of the power grid state in this delay τ time, and execute the instruction s(t) = Sserver + ΔSpred after look-ahead correction;

[0014] Step 5: VSG parameter adaptive adjustment

[0015] According to the instruction s(t) corrected in step 4, accurate power control is performed;

[0016] Step 6: Event-triggered sparse communication

[0017] The real-time measurement value y(t) of the local PMU is continuously compared with the value ylast broadcast successfully last time, and only when the deviation between the two exceeds a dynamic threshold δ(t), a communication is triggered to broadcast the current state to the control center and the adjacent power generation vehicle.

[0018] The specific steps of generating the initial space-time feature matrix are,

[0019] Dynamic data fusion is performed using the Kalman filtering algorithm,

[0020] ,

[0021] ,

[0022] Among them: is defined as [toffset, tdrift], where toffset is the real time offset of the local clock relative to the standard UTC, tdrift is its drift rate, and the observation vector Z k is the time information output by the Beidou timing module and the timestamp information obtained by the 5G network through the IEEE 802.1AS-Rev protocol, H is the observation matrix of the conversion relationship between the physical state and the measurement space, R is the observation noise covariance matrix, when the Beidou satellite signal quality decreases, increase σ²Beidou; When the 5G network congestion causes jitter to increase, increase σ²5G, K k is the Kalman gain, P k∣k−1 is the prediction error covariance;

[0023] Through the prediction and update iteration of Kalman filtering, the optimal time offset is estimated in real time :

[0024] ,

[0025] a simplified characterization of performance;

[0026] spatial synchronization equation:

[0027] ,

[0028] wherein: is the deviation of the actual position of the power generation vehicle from the theoretical grid connection point, is the fusion weight coefficient, PPP-RTK is a precise point positioning technology to solve millimeter-level coordinates, and 5G NR-TDOA is a time difference of arrival compensation signal transmission path error;

[0029] Create a complementary fusion spatio-temporal feature matrix through a spatial coordination mechanism:

[0030] .

[0031] Dynamic topology feature extraction is specifically,

[0032] Abstract the power system as a dynamic graph:

[0033] ,

[0034] wherein: V is a node set, including all mobile power generation vehicle grid connection points, important load points and substations, E is an edge set, representing the physical line connection relationship between nodes, from which an adjacency matrix A [i, j] = 1 is constructed to indicate that there is a direct connection between nodes i and j, W is an edge weight matrix, W[i, j] can represent the inverse of line impedance, representing the closeness of the connection;

[0035] The graph convolution network GCN propagation adopts a multi-layer GCN to aggregate node features. The spatio-temporal feature matrix output by step 1 is spliced with the real-time electrical quantities collected by the PMU to form the initial node feature matrix H(0) of the GCN,

[0036] ,

[0037] Here, Ã = A + I is the adjacency matrix with a self-loop, which ensures that the node retains its own information when aggregating neighbor information, D̃ is the degree matrix of Ã, and W (l) is the learnable weight matrix of the l-th layer, For the activation function, the propagation of each layer of GCN is equivalent to diffusing the feature information of each node along the power grid line to its first-order neighbors once. After L layers of GCN, each node obtains the state information of all nodes in its L-order neighborhood, thereby capturing complex spatial dependencies.

[0038] Electric island division:

[0039] ,

[0040] ,

[0041] where λ is the eigenvalue, which identifies the island independence. The number of zero eigenvalues of L is equal to the number of connected components in the graph, i.e. the number of electric islands, which provides global topology awareness for the system, automatically identifies whether a power generation vehicle is in island operation state, and switches to different control strategies, outputting a three-dimensional dynamic topology feature tensor F with dimensions N x T x C, node number x time series length x feature dimension. The feature dimension C not only includes the original electrical quantity, but also includes the spatial dependence strength, topological connectivity and island identification code learned by GCN, and other high-dimensional abstract features.

[0042] Precise grid-connected state prediction specifically refers to taking the dynamic topology feature tensor F output in step 2 as input, applying a self-attention mechanism to the features of all nodes at each time step, and calculating an attention score matrix,

[0043] ,

[0044] where Q, K, and V are query, key, and value matrices, d k is the dimension,

[0045] The static topology of step 2 is converted into a dynamic evolution trend through cascading processing: first, the spatial dependence of the power grid topology is extracted by the adaptive graph convolution layer, then the dynamic time features are learned by the algorithm encoder with spatio-temporal attention gating, and the voltage signal is converted from time domain to frequency domain by the FFT frequency domain branch, and the phase spectrum features are directly extracted as the supplementary input of the machine learning model,

[0046] The input voltage time series v n is converted to the frequency domain by fast Fourier transform to obtain its frequency spectrum V k , and then the phase spectrum is extracted ,

[0047] Frequency domain branch:

[0048] ,

[0049] Phase spectrum extraction:

[0050] ,

[0051] The phase spectrum is spliced with the context vector output by the encoder as a strong prior feature, and is fed into the decoder together, using an autoregressive method to generate a prediction sequence of K steps in the future at each time step, combining the encoder output and frequency domain features. In the training stage, a phase-priority hybrid loss function is used,

[0052] Phase-priority hybrid loss function:

[0053] ,

[0054] Where: L MSE is the mean square error, which ensures the accuracy of the overall prediction of voltage amplitude, frequency, etc. L phase is the phase loss, which uses cosine similarity loss to specifically punish the deviation of the predicted phase angle from the true phase angle,

[0055] ,

[0056] Output high-precision state prediction sequence of K steps in the future, including voltage amplitude, frequency, phase angle and its rate of change.

[0057] The communication delay look-ahead compensation is specifically,

[0058] A star-ground collaborative compensation mechanism is constructed, and the power generation vehicle predicts and timestamps the immediate response to user operation locally and uploads; the collaborative control center backtracks to the accurate time when the operation occurs based on dynamic delay estimation to perform authoritative calculation, and the local controller of the power generation vehicle does not directly execute the server instruction Sserver after receiving it, but performs look-ahead correction,

[0059] Delay compensation equation:

[0060] ,

[0061] Where: is the final state instruction vector that the power generation vehicle should execute after look-ahead correction at local time t, s server is the original instruction vector received from the collaborative control center, and Δspred is a vector that is related to s serverState increment prediction vector with the same dimension, calculated by local machine learning prediction model, to predict the natural evolution of the system state in the tdelay time, which can be obtained by first-order approximation:

[0062] ,

[0063] Where: and are the voltage and phase angle change rate predicted by step 3 model respectively;

[0064] And the results are pushed to the latest state to the generator car; The generator car uses phase jump to correct its role position, and through the topological interpolation algorithm and the adjacency matrix information of step 2, it smooths the transition of the unupdated node state, ensures the alignment of prediction and execution in time and space dimensions, and realizes the real-time synchronization of prediction data and grid-connected action.

[0065] VSG parameter adaptive adjustment is specifically,

[0066] According to the prediction, when it is predicted that there will be a large frequency drop in the future, J is actively increased in advance, and when it is predicted that the system will oscillate, D is actively increased. k1, k2 are determined by offline reinforcement learning training or online fuzzy control rules,

[0067] VSG swing equation:

[0068] ,

[0069] Where θ is the rotor angle, P m is the mechanical power, P e is the electromagnetic power,

[0070] Adaptive parameter adjustment model:

[0071] ,

[0072] ,

[0073] Where k1, k2 are optimization parameters.

[0074] According to the real-time state, an event-triggered communication system is constructed, and in the process of continuously monitoring the physical quantities of the power grid, the dynamic error between the current sampling value and the locally held state is calculated in real time. Only when the deviation between the local measurement value y(t) and the value y last broadcasted last time exceeds a dynamic threshold δ(t), a communication is triggered,

[0075] Trigger condition:

[0076] ,

[0077] ,

[0078] Where y(t) is a physical quantity of the power grid. As the baseline threshold, These are adaptive coefficients.

[0079] Event-triggered dynamic threshold coefficient Method for determining:

[0080] To achieve an adaptive balance between communication overhead and control accuracy, the adaptive coefficients in the dynamic threshold δ(t) are... The threshold should be correlated with the system's operating state and the importance of nodes. During system transients or for critical nodes, a smaller threshold should be used to ensure the sensitivity of information transmission; during steady-state conditions or for non-critical nodes, the threshold can be appropriately relaxed to save communication resources.

[0081] coefficient Determined by the following formula:

[0082] ,

[0083] in: It is a basic adjustment coefficient, C node It is the node centrality, which is obtained from the graph convolutional network (GCN) analysis in step 2, C. node The larger the value of S, the more important the node. state It is a stability factor that characterizes the state of a system.

[0084] Compared with the prior art, the beneficial effects of the present invention are:

[0085] By innovating a multi-protocol adaptive interface, it enables plug-and-play functionality for both new and old equipment. Utilizing high-speed three-phase current sampling and a Lyapunov stability algorithm, it achieves microsecond-level grid connection detection and zero-phase-difference closing control. Furthermore, it employs a triple protection mechanism, including virtual impedance injection and power feedforward compensation, to significantly shorten oscillation suppression time. Its dedicated hardware platform supports long-distance wireless multi-vehicle collaborative control, effectively solving key problems such as inconsistent grid connection protocols in traditional generator vehicles, time-consuming manual operations, and oscillations caused by power backlash. This provides core equipment support for building a new type of power system that is fast-responding, seamlessly switching, and safe and stable, while also promoting the advancement of seamless grid connection technology for emergency generator vehicles. Attached Figure Description

[0086] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be regarded as a limitation on the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0087] Figure 1 Method flowchart of the present application. DETAILED DESCRIPTION

[0088] The technical solutions in the embodiments of the present application will be described below in combination with the drawings in the embodiments of the present application. It should be noted that similar reference numerals and letters represent similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0089] The term "comprising" or "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or device including the element.

[0090] The terms "first", "second", and the like are only used to distinguish one entity or operation from another entity or operation, and cannot be understood as indicating or implying relative importance, and cannot be understood as requiring or implying any such actual relationship or order between the entities or operations.

[0091] The above only describes the embodiments of the present application and does not limit the protection scope of the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0092] As shown in FIG. 1, a mobile power generation vehicle grid-connected state prediction and collaborative control method includes the following specific steps: Figure 1

[0093] Step 1: High-precision space-time synchronization

[0094] The fusion time terminal of each power generation vehicle simultaneously receives the Beidou satellite signal and the synchronous clock signal distributed through the 5G TSN network to generate an initial space-time feature matrix;

[0095] ​Step 2: Dynamic topology feature extraction

[0096] The spatio-temporal feature matrix generated in step 1 is obtained, and the real-time electrical quantity collected by the local PMU is spliced to extract the deep spatial dependence features of the current time grid in real time, and output the feature tensor F;

[0097] Step 3: Precise grid-connected state prediction

[0098] The feature tensor F is immediately sent to the ST-Transformer prediction model to predict the dynamic trajectories of voltage, frequency, and phase angle and their change rates in the next K time steps;

[0099] Step 4: Prospective compensation for communication delay

[0100] The communication delay τ from the issuance of the instruction to the current reception is estimated, and the prediction result of step 3 is used to calculate the natural evolution amount ΔS of the grid state within the delay τ time pred , and the instruction s(t) =S server +ΔS pred after prospective correction is executed;

[0101] Step 5: Adaptive adjustment of VSG parameters

[0102] Accurate power control is performed according to the corrected instruction s(t) of step 4;

[0103] Step 6: Event-triggered sparse communication

[0104] The real-time measurement value y(t) of the local PMU is continuously compared with the value y last broadcast successfully last time, and only when the deviation between the two exceeds a dynamic threshold δ(t), a communication is triggered to broadcast the current state to the control center and the adjacent power car.

[0105] The specific steps for generating the initial spatio-temporal feature matrix are as follows,

[0106] Dynamic data fusion is performed using the Kalman filtering algorithm, as shown in formulas (1) and (2)

[0107] (1),

[0108] (2),

[0109] Wherein: is defined as [toffset, tdrift], where toffset is the real time offset of the local clock relative to the standard UTC, and tdrift is its drift rate, and the observation vector Z kThe time information output by the Beidou timing module and the timestamp information obtained by the 5G network through the IEEE 802.1AS-Rev protocol, H is the observation matrix of the conversion relationship of mapping the physical state to the measurement space, R is the observation noise covariance matrix: diag([σ²Beidou σ²5G]), when the quality of the Beidou satellite signal decreases, increase σ²Beidou; when the 5G network congestion causes the jitter to increase, increase σ²5G, Kk is the Kalman gain, P k∣k−1 is the prediction error covariance;

[0110] Through the prediction and update iteration of Kalman filtering, the optimal time offset is estimated in real time :

[0111] (3),

[0112] is only a simplified representation of performance;

[0113] Space synchronization equation:

[0114] (4),

[0115] wherein: is the deviation of the actual position of the power car from the theoretical grid connection point, is the fusion weight coefficient, PPP-RTK is a precise point positioning technology to solve millimeter-level coordinates, and 5G NR-TDOA is a time difference of arrival compensation signal transmission path error;

[0116] Create a complementary fusion spatio-temporal feature matrix through a space coordination mechanism:

[0117] (5).

[0118] The dynamic topology feature extraction is specifically,

[0119] Abstract the power system as a dynamic graph:

[0120] (6),

[0121] wherein: V is a node set, including all mobile power car grid connection points, important load points and substations, E is an edge set, representing the physical line connection relationship between nodes, from which an adjacency matrix A [i, j] = 1 is constructed, indicating that there is a direct connection between nodes i and j, W is an edge weight matrix, W[i, j] can represent the inverse of line impedance, representing the closeness of the connection;

[0122] The propagation of the graph convolutional network GCN is shown in Equation (7). A multi-layer GCN is used to aggregate node features. The spatiotemporal feature matrix output in step 1 is concatenated with the real-time electrical quantities collected by the PMU to form the initial node feature matrix H(0) of the GCN.

[0123] (7),

[0124] Here, Ã = A + I is the adjacency matrix with self-loops added, ensuring that nodes retain their own information when aggregating neighbor information. D̃ is the degree matrix of Ã, and W(l) is the learnable weight matrix of the l-th layer. As the activation function, the propagation of each layer of GCN is equivalent to spreading the feature information of each node along the power grid line to its first order front and back once. After passing through L layers of GCN, each node obtains the state information of all nodes in its L-order neighborhood, thereby capturing complex spatial dependencies.

[0125] Electric island division:

[0126] (8),

[0127] (9),

[0128] Where λ is the eigenvalue, which identifies the independence of islands. Spectral analysis is performed using the Laplacian matrix L = D - A of the graph. The number of zero eigenvalues ​​of L is exactly equal to the number of connected components in the graph, i.e., the number of electrical islands. This provides the system with global topology awareness, automatically identifying whether a generator is in an islanded operation state, and thus switching to different control strategies. It outputs a three-dimensional dynamic topology feature tensor F with dimensions N×T×C, which is the number of nodes × the length of the time series × the feature dimension. The feature dimension C not only includes the original electrical quantities, but also includes high-dimensional abstract features such as spatial dependency strength, topological connectivity, and island identification codes learned by GCN.

[0129] The precise grid connection status prediction specifically involves taking the dynamic topology feature tensor F output from step 2 as input, and at each time step, applying a self-attention mechanism to the features of all nodes to calculate an attention score matrix.

[0130] (10)

[0131] Where Q, K, and V are the query, key, and value matrices, and d k As a dimension,

[0132] The static topology of step 2 is converted into a dynamic evolution trend through cascading processing: first, the power grid topology spatial dependence is extracted by the adaptive graph convolution layer, then the dynamic time features are learned by the algorithm encoder with spatio-temporal attention gate, and at the same time, the voltage signal is converted from time domain to frequency domain through the FFT frequency domain branch, and the phase spectrum feature is directly extracted as the supplementary input of the machine learning model,

[0133] The input voltage time series v n is converted to the frequency domain by fast Fourier transform to obtain its frequency spectrum V k Then extract its phase spectrum ,

[0134] Frequency domain branch:

[0135] (11),

[0136] Phase spectrum extraction:

[0137] (12),

[0138] The phase spectrum is spliced with the context vector output by the encoder as a strong prior feature, and is sent into the decoder together, which adopts an autoregressive method to generate a K-step prediction sequence step by step, combining the encoder output and the frequency domain feature. In the training stage, a phase-priority hybrid loss function is used,

[0139] Phase-priority hybrid loss function:

[0140] (13),

[0141] Where: L MSE is the mean square error, which ensures the accuracy of the overall prediction of voltage amplitude, frequency, etc. L phase is the phase loss, which uses cosine similarity loss to specifically punish the deviation of the predicted phase angle and the real phase angle,

[0142] (14),

[0143] Output the high-precision state prediction sequence of K steps in the future, including voltage amplitude, frequency, phase angle and its change rate dU / dt, df / dt.

[0144] The communication delay proactive compensation is specifically,

[0145] A star-ground cooperative compensation mechanism is constructed, the power generation vehicle uploads in real time after predicting the user operation locally and marking the timestamp; the cooperative control center performs authoritative calculation based on dynamic delay estimation and backtracking to the precise time of operation, and the local controller of the power generation vehicle does not directly execute the server instruction S server after receiving the server instruction S

[0146] Delay compensation equation:

[0147] (15),

[0148] wherein: is the final state instruction vector that the power generation vehicle should execute at the local t time after the prospective correction, S server is the original instruction vector received from the cooperative control center, Δs pred is a state increment prediction vector with the same dimension as S server , which is calculated by the local machine learning prediction model, and is used to predict the natural evolution of the system state within t delay This time, the first-order approximation of the vector can be obtained:

[0149] (16),

[0150] wherein: and are the voltage and phase angle change rate predicted by the step 3 model, respectively;

[0151] and the results are pushed to the latest state and sent to the power generation vehicle; the power generation vehicle corrects its role position by phase jump, and smoothes the transition of the non-updated node state through the topological interpolation algorithm and the adjacency matrix information of step 2, to ensure that the prediction and execution are aligned in the time and space dimensions, and to realize the real-time synchronization of the prediction data and the grid-connected action.

[0152] VSG parameter adaptive adjustment is specifically,

[0153] According to the prediction, the online adaptive adjustment is performed, when it is predicted that there will be a large frequency drop in the future (|df / dt| is large), J is actively increased in advance, which is equivalent to adding a virtual moment of inertia to the system, which can more effectively resist frequency impact, when it is predicted that the system will have power oscillation, which is manifested as |dU / dt| or power change rate is large, D is actively increased to provide stronger damping and quickly suppress oscillation, k1, k2 can be determined by offline reinforcement learning training or online fuzzy control rules,

[0154] VSG swing equation:

[0155] (17)

[0156] Where θ is the rotor angle, P m For mechanical power, P e Electromagnetic power,

[0157] Adaptive parameter tuning model:

[0158] (18)

[0159] (19)

[0160] Where k1 and k2 are optimization parameters.

[0161] An event-triggered communication system is constructed based on real-time status. During continuous monitoring of power grid physical quantities, the dynamic error between the current sampled value and the locally held state is calculated in real time. This error only occurs when the local measured value y(t) differs from the value y(t) of the last successfully broadcast value. last A communication is only triggered when the deviation between them exceeds a dynamic threshold δ(t).

[0162] Triggering conditions:

[0163] (20)

[0164] (twenty one),

[0165] Where y(t) is a physical quantity of the power grid. As the baseline threshold, These are adaptive coefficients.

[0166] Event-triggered dynamic threshold coefficient Method for determining:

[0167] To achieve an adaptive balance between communication overhead and control accuracy, the adaptive coefficients in the dynamic threshold δ(t) are... The threshold should be correlated with the system's operating state and the importance of nodes. During system transients or for critical nodes, a smaller threshold should be used to ensure the sensitivity of information transmission; during steady-state conditions or for non-critical nodes, the threshold can be appropriately relaxed to save communication resources.

[0168] coefficient It can be determined by the following formula:

[0169] (twenty two),

[0170] in: is a basic regulation coefficient. C node is the centrality of the node, which can be analyzed by the graph convolution network (GCN) in step 2 (for example, the degree centrality or betweenness centrality of the node), C node The larger the value, the more important the node is. S state is a stability factor representing the state of the system. When the predicted system disturbance is small (|df / dt|pred is low), S state The value is larger; when the system is predicted to enter a transient state, S state The value is smaller.

[0171] In steady state (df / dt is small), the threshold δ is small, and the system is very sensitive, ensuring control accuracy. In transient state (df / dt is large), the threshold δ is automatically relaxed, allowing a certain range of normal fluctuations without triggering communication, avoiding confusion during disturbance, and only interacting with key information when the system deviates from the predicted trajectory. The triggered message will be broadcast along the minimum spanning tree topology based on link quality calculation, ensuring that all nodes that need information are covered at the lowest communication cost. This mechanism only communicates when the state of the power grid changes significantly, maintaining a dormant mode under normal conditions, ensuring the spatiotemporal synchronization of critical data and significantly reducing channel noise, ultimately achieving a synergistic optimization of communication efficiency and system response speed.

[0172] Specific embodiments: First stage: system deployment and preparation

[0173] This stage mainly completes hardware deployment, software configuration and initial data collection.

[0174] First, for the mobile power generation truck terminal, we equip each mobile power generation truck in the cluster with a set of integrated time and control terminal. The hardware core of the terminal mainly includes:

[0175] (1) Beidou-3 multi-frequency multi-mode high-precision receiver, which requires the receiver to support PPP-RTK solution and be able to receive differential correction data from national CORS network or self-built base station to achieve centimeter-level positioning and nanosecond-level timing;

[0176] (2) 5G communication module, which needs to support 5G TSN (Time Sensitive Network) protocol stack, especially IEEE802.1AS-Rev clock synchronization protocol, to ensure low latency and high reliability of data transmission;

[0177] (3) High-performance edge computing unit, such as NVIDIA Jetson series or equivalent embedded AI computer, for running local machine learning prediction models, VSG adaptive control algorithms and event-triggered communication logic;

[0178] (4) PMU (Phasor Measurement Unit) and timing module are strictly synchronized to collect electrical data such as voltage, current phasor, frequency and frequency rate of the grid-connected point at a high sampling rate (e.g. 50 times / s);

[0179] (5) The generator controller interface is used to send control instructions to the prime mover governor and excitation system of the power generation vehicle;

[0180] (6) The collaborative control center (cloud or regional server) deploys high-performance servers to gather all power generation vehicle data, monitor the global state, issue collaborative control instructions, and update and maintain the model;

[0181] Secondly, network environment construction and software system configuration:

[0182] 5G private network or slice The purpose of 5G private network or slice is to ensure the determinism, low latency and high reliability of communication, and to provide a dedicated channel for control signaling by deploying a 5G private network or using the slice function of the operator network for the mobile power generation vehicle cluster;

[0183] Beidou ground station The purpose of Beidou ground station is to enhance the network and ensure that the power generation vehicle operating area has stable and reliable differential correction data source coverage;

[0184] Terminal software The purpose of terminal software is to deploy data acquisition and fusion software (implementing Kalman filtering), lightweight machine learning inference framework, VSG control algorithm program and communication agent program on the edge computing unit;

[0185] Center software The purpose of center software is to deploy data aggregation platform, power grid topology database, machine learning training platform (such as PyTorch / TensorFlow) and global situation awareness and decision system;

[0186] Finally, initial data collection and modeling:

[0187] Power grid topology information input The purpose of power grid topology information input is to input the physical connection relationship (line, transformer, load point) of the operating area power grid into the collaborative control center to generate the initial adjacency matrix A and weight matrix W;

[0188] Historical data and simulation data preparation The purpose of historical data and simulation data preparation is to collect historical operation data of the power grid, and use simulation software such as PSASP and DIgSILENT to generate a large amount of power grid dynamic data under different working conditions (such as load mutation, line fault, new energy fluctuation) for subsequent machine learning model training.

[0189] Second phase: offline model training

[0190] This phase uses the prepared data to train the core machine learning model in the collaborative control center;

[0191] The grid topology feature extraction model (GCN) training takes the grid topology structure (adjacency matrix A) and the node electrical quantities in the simulation data as input; a multi-layer graph convolution network (GCN) is constructed. The training goal is to enable the GCN to learn the electrical coupling relationship between nodes, for example, by learning the weight W(l), so that the output of the model can accurately reflect the degree of influence of the disturbance of a certain node on other nodes. At the same time, through spectral analysis of the graph Laplacian matrix L, the model training identifies the ability of the network connectivity, providing a basis for electrical island judgment;

[0192] The spatio-temporal state prediction model (ST-Transformer) training takes the dynamic topology feature tensor F extracted by GCN as the model input. This tensor contains C features of all N nodes in the past period (e.g. T = 50 sampling points, i.e. 1 second);

[0193] The frequency domain branch is to transform the voltage signal in the input sequence through FFT to extract its phase spectrum φk as a strong prior feature;

[0194] The encoder dynamically learns the importance weights of different nodes and different time points through self-attention mechanism to capture the spatio-temporal dependence;

[0195] The decoder is to fuse the context vector output by the encoder and the phase spectrum feature extracted by the frequency domain branch, and to predict the voltage amplitude, frequency, phase angle and their change rates in the future K time steps (e.g. K = 25, i.e. predicting the future 0.5 seconds) in a self-recurrent manner;

[0196] The loss function is a hybrid loss function L = 0.7L phase + 0.3L MSE Training is carried out to ensure higher accuracy in phase angle prediction, which is crucial for grid connection;

[0197] Training and verification are carried out using a large amount of simulation and historical data sets to train and ensure the model has good generalization ability through cross-validation;

[0198] Model deployment is to optimize and optimize the trained GCN and ST-Transformer model (such as model pruning, quantization), and then deploy it to the edge computing unit of each mobile power generation vehicle.

[0199] Phase 3: Online collaborative operation

[0200] The system enters real-time operation state, and each step forms a high-speed closed-loop collaborative control process.

[0201] Step 1: High-precision spatio-temporal synchronization (continuous)

[0202] Each power generation vehicle's fusion time terminal receives both Beidou satellite signals and synchronized clock signals distributed through the 5G TSN network;

[0203] The Kalman filter algorithm (formula 1, 2) inside the terminal runs in real time, dynamically fusing the two time sources. When the Beidou signal is blocked (σ²Beidou increases), the trust weight of the 5G TSN time is automatically increased; when the 5G network is congested and jittered (σ²5G increases), more trust is placed in the Beidou time;

[0204] At the same time, the PPP-RTK module calculates the centimeter-level vehicle position;

[0205] Finally, each data point collected by the local PMU is stamped with a unified and high-precision space-time stamp, generating an initial space-time feature matrix (formula 5), which fundamentally solves the data alignment problem.

[0206] Step 2: Dynamic Topological Feature Extraction (Real-time)

[0207] The edge computing unit obtains the space-time feature matrix generated in step 1 and splices the real-time electrical quantities (voltage amplitude, phase angle, power) collected by the local PMU;

[0208] The deployed GCN model performs forward propagation calculation on these inputs (formula 7), real-time extracts deep spatial dependence features of the power grid at the current time, and judges whether it is in island operation state (based on the number of zero eigenvalues of Laplacian matrix);

[0209] The output of this step is a three-dimensional feature tensor F that contains rich topological information at the current and past T time points.

[0210] Step 3: Accurate Grid Connection State Prediction (Real-time)

[0211] The feature tensor F is immediately sent to the ST-Transformer prediction model;

[0212] The model performs time and frequency domain analysis in parallel, focuses on key disturbance sources in the power grid through self-attention mechanism, and predicts the dynamic trajectories of voltage, frequency, and phase angle within the next K time steps (e.g. 0.5 seconds) and their change rates (dU / dt, df / dt);

[0213] Step 4: Prospective Compensation for Communication Delay (Executed when receiving instructions)

[0214] When the power generation vehicle receives a control instruction Sserver (such as adjusting power output) from the cooperative control center, it does not execute it immediately;

[0215] The local controller first estimates the communication delay τ from the command issuance to the current reception;

[0216] Then, using the prediction result of step 3, the natural evolution of the grid state ΔSpred in this delay τ time is calculated (formula 16);

[0217] The final executed command is the prospectively corrected command s(t) = Sserver + ΔSpred (formula 15). This ensures that the control action is for the real state of the grid at the execution moment, not for an already outdated state;

[0218] Step 5: Adaptive adjustment of VSG parameters (real-time)

[0219] The VSG controller no longer uses fixed virtual inertia J and damping coefficient D;

[0220] It continuously monitors the future state prediction sequence output by step 3, and each power generation vehicle's VSG controller adjusts in advance according to large deviation values appearing in the prediction sequence;

[0221] Inertia adaptation: If a large frequency drop is predicted in the future (max(|df / dt|pred) exceeds the threshold), each power generation vehicle's controller will proactively increase the virtual inertia J according to formula (18), enhancing the system's anti-disturbance ability, as if adding a virtual flywheel to the system;

[0222] Damping adaptation: If power oscillation is predicted in the future (max(|dU / dt|pred) or power change rate is large), each power generation vehicle's controller will actively increase the damping coefficient D according to formula (19), providing stronger damping to quickly suppress oscillation;

[0223] The adaptively adjusted VSG controller performs accurate power control according to the command s(t) corrected in step 4. That is, when the frequency impact really comes, since each power generation vehicle has far exceeded the virtual inertia at ordinary times, they can instantly and autonomously release active power in proportion (determined by the droop characteristic of VSG) to support the frequency. Since the parameters are adjusted according to unified prediction, the problem of power oscillation between vehicles caused by parameter mismatch is avoided. Ultimately, the actual frequency drop depth is effectively suppressed, and the recovery speed is significantly accelerated.

[0224] Step 6: Event-triggered sparse communication (continuous monitoring)

[0225] To reduce the occupation of the 5G channel, the terminal is in a silent state by default, and only local monitoring-prediction-control closed loop is performed;

[0226] The terminal continuously compares the real-time measurement value y(t) of the local PMU with the last successfully broadcast value y last performs comparison;

[0227] Only when the deviation of the two exceeds a dynamic threshold δ(t) (formula 20), a communication is triggered to broadcast the current state to the control center and neighboring power cars;

[0228] The threshold δ(t) is adaptive (formula 21, 22):

[0229] When the power grid is running smoothly (|df / dt| is small), the threshold is large, saving communication resources;

[0230] When it is predicted that the system will enter a transient state (|df / dt|pred is large), or the node is very important in the network (C node value is large), the threshold δ(t) will automatically decrease and become more sensitive, ensuring that critical information can be transmitted in time and reliably;

[0231] The triggered communication will be along the minimum spanning tree topology calculated based on link quality, achieving the most efficient broadcast. The sparse feedback data is also used to correct the state of the Kalman filter in step 1, forming a global system-level closed loop.

[0232] The above only describes the embodiments of the present application and is not used to limit the protection scope of the present application. For those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting and coordinating the grid connection status of a mobile generator vehicle, characterized in that, The specific steps include the following: Step 1: High-precision spatiotemporal synchronization Each generator vehicle's fusion timing terminal simultaneously receives BeiDou satellite signals and synchronization clock signals distributed through the 5G TSN network to generate an initial spatiotemporal feature matrix. Step 2: Dynamic Topological Feature Extraction Obtain the spatiotemporal feature matrix generated in step 1, and concatenate it with the real-time electrical quantities collected by the local PMU. Extract the deep spatial dependency features of the power grid at the current moment in real time, and output the feature tensor F. Step 3: Precise Grid Connection Status Prediction The feature tensor F is immediately fed into the ST-Transformer prediction model to predict the dynamic trajectory and rate of change of voltage, frequency, and phase angle over the next K time steps. Step 4: Proactive compensation for communication delays Estimate the communication delay τ from the issuance of the command to the current reception, use the prediction results of step 3 to calculate the natural evolution of the power grid state ΔSpred during this delay τ, and execute the forward-corrected command s(t) = Sserver + ΔSpred. Step 5: Adaptive adjustment of VSG parameters Precise power control is performed based on the corrected instruction s(t) from step 4; Step 6: Event-triggered sparse communication The local PMU continuously compares the real-time measurement value y(t) with the value ylast that was successfully broadcast last time. Only when the deviation between the two exceeds a dynamic threshold δ(t) will a communication be triggered to broadcast the current status to the control center and the nearby generator car. The specific details of the communication delay forward compensation are as follows: A space-ground collaborative compensation mechanism is established, whereby the generator vehicle predicts and responds to user operations locally in real time, and uploads the results with timestamps; the collaborative control center performs authoritative calculations based on dynamic latency predictions, backtracking to the precise moment the operation occurred; and the generator vehicle's local controller, upon receiving the server command Sserver, does not execute it directly but instead performs forward-looking corrections. Delay compensation equation: ， in: Let s be the state command vector that the generator car should finally execute at local time t after forward correction. server It is the raw instruction vector received from the collaborative control center, Δspred is a vector related to s server State increment prediction vectors with the same dimensions are calculated by the local machine learning prediction model and used to predict the state increment at time t. delay The natural evolution of the system state during this period, this vector, can be obtained by a first-order approximation: ， in: and These are the voltage and phase angle change rate predicted by the model in step 3, respectively. The results are then updated to the latest state and sent to the generator vehicle. The generator vehicle uses phase jump to correct its own role position and smoothly transitions to the unupdated node state through topological interpolation algorithm and adjacency matrix information from step 2, ensuring that prediction and execution are aligned in the spatiotemporal dimension and realizing real-time synchronization of prediction data and grid connection actions. The VSG parameter adaptive adjustment is specifically as follows: Based on predictions, online adaptive adjustments are made. When a large frequency drop is predicted in the future, J is proactively increased in advance. When power oscillations are predicted, D is proactively increased. k1 and k2 are determined through offline reinforcement learning training or online fuzzy control rules. VSG swing equation: ， Where θ is the rotor angle, P m For mechanical power, P e Electromagnetic power, Adaptive parameter tuning model: ， ， Where k1 and k2 are optimization parameters; An event-triggered communication system is constructed based on real-time status. During continuous monitoring of power grid physical quantities, the dynamic error between the current sampled value and the locally held state is calculated in real time. This error only occurs when the local measured value y(t) differs from the value y(t) of the last successfully broadcast value. last A communication is only triggered when the deviation between them exceeds a dynamic threshold δ(t). Triggering conditions: ， ， Where y(t) is a physical quantity of the power grid. As the baseline threshold, These are adaptive coefficients.

2. The method for predicting and coordinating the grid connection status of a mobile generator vehicle according to claim 1, characterized in that, The specific steps for generating the initial spatiotemporal feature matrix are as follows: Dynamic data fusion is performed using the Kalman filter algorithm. ， ， in: Defined as [toffset, tdrift], where tofoffset is the true time offset of the local clock relative to standard UTC, tdrift is its drift rate, and the observation vector Z... k The time information output by the BeiDou timing module and the timestamp information acquired by the 5G network are given. H is the observation matrix that maps the physical state to the measurement space, and R is the observation noise covariance matrix. When the BeiDou satellite signal quality deteriorates, σ²BeiDou is increased; when 5G network congestion leads to increased jitter, σ²5G is increased. K k For Kalman gain, P k∣k−1 For the prediction error covariance; The optimal time offset is estimated in real time through Kalman filtering prediction and update iteration. : ， A simplified characterization of performance; Space synchronization equation: ， in: This is the deviation between the actual location of the generator vehicle and the theoretical grid connection point. To integrate the weighting coefficients, To solve millimeter-level coordinates for precise single-point positioning technology, To utilize the 5G NR-TDOA time difference of arrival to compensate for signal transmission path errors; A complementary and integrated spatiotemporal feature matrix is ​​created through a spatial collaboration mechanism.

3. The method for predicting and coordinating the grid connection status of a mobile generator vehicle according to claim 1, characterized in that, Dynamic topological feature extraction specifically involves, Abstracting the power system as a dynamic graph: ， Where: V is the set of nodes, including all mobile generator grid connection points, important load points and substations; E is the set of edges, representing the physical line connection relationship between nodes; thus, the adjacency matrix A[i, j] = 1 is constructed to indicate that there is a direct connection between nodes i and j; W is the edge weight matrix, W[i, j] represents the reciprocal of the line impedance, representing the tightness of the connection; The graph convolutional network (GCN) propagation employs a multi-layer GCN to aggregate node features. The spatiotemporal feature matrix output in step 1 is concatenated with real-time electrical quantities acquired by the PMU to form the initial node feature matrix H(0) of the GCN. ， Here, Ã = A + I is the adjacency matrix with self-loops added, ensuring that nodes retain their own information when aggregating neighbor information, D̃ is the degree matrix of Ã, and W (l) It is the learnable weight matrix of the l-th layer. As the activation function, the propagation of each layer of GCN is equivalent to spreading the feature information of each node along the power grid line to its first order front and back once. After passing through L layers of GCN, each node obtains the state information of all nodes in its L-order neighborhood, thereby capturing complex spatial dependencies. Electric island division: ， ， Where λ is the eigenvalue, which identifies the independence of islands. Spectral analysis is performed using the Laplacian matrix L = D - A of the graph. The number of zero eigenvalues ​​of L is exactly equal to the number of connected components in the graph, i.e., the number of electrical islands. This provides the system with global topology awareness, automatically identifying whether the generator is in an islanded operation state, and thus switching to different control strategies. It outputs a three-dimensional dynamic topology feature tensor F with dimensions N×T×C, which is the number of nodes × the length of the time series × the feature dimension. The feature dimension C not only contains the original electrical quantities, but also the high-dimensional abstract features of spatial dependency strength, topological connectivity, and island identification codes learned by GCN.

4. The method for predicting and coordinating the grid connection status of a mobile generator vehicle according to claim 1, characterized in that, The precise grid connection status prediction specifically involves taking the dynamic topology feature tensor F output from step 2 as input, and at each time step, applying a self-attention mechanism to the features of all nodes to calculate an attention score matrix. ， Where Q, K, and V are the query, key, and value matrices, and d k As a dimension, The static topology from step 2 is transformed into a dynamic evolution trend through cascaded processing: First, the spatial dependency of the power grid topology is extracted by an adaptive graph convolutional layer. Then, the dynamic time features are learned by an algorithm encoder with spatiotemporal attention gating. Simultaneously, the voltage signal is transformed from the time domain to the frequency domain through FFT frequency domain branching analysis, and the phase spectrum features are directly extracted as supplementary input to the machine learning model. The input voltage time series v n Its spectrum V is obtained by converting it to the frequency domain using Fast Fourier Transform. k Then extract its phase spectrum. , Frequency domain branch: ， Phase spectrum extraction: ， The phase spectrum As a strong prior feature, it is concatenated with the context vector output by the encoder and fed into the decoder. Using an autoregressive approach, combining the encoder output and frequency domain features, it generates the predicted sequence for the next K steps step by step. During the training phase, a phase-first hybrid loss function is employed. Phase-priority hybrid loss function: ， Where: L MSE To ensure the accuracy of overall predictions of voltage amplitude and frequency, L is used to calculate the mean square error. phase For phase loss, cosine similarity loss is used to specifically penalize the deviation between the predicted phase angle and the true phase angle. ， Output a high-precision state prediction sequence for the next K steps, including voltage amplitude, frequency, phase angle and its rate of change.

5. The method for predicting and coordinating the grid connection status of a mobile generator vehicle according to claim 1, characterized in that, Event-triggered dynamic threshold adaptive coefficient Method for determining: To achieve an adaptive balance between communication overhead and control accuracy, the adaptive coefficients in the dynamic threshold δ(t) are... The threshold should be correlated with the system's operating state and the importance of nodes. During system transients or for critical nodes, a smaller threshold should be used to ensure the sensitivity of information transmission; during steady-state conditions or for non-critical nodes, the threshold should be relaxed to save communication resources. Adaptive coefficients Determined by the following formula: ， in: It is a basic adjustment coefficient, C node It is the node centrality, which is obtained from the graph convolutional network (GCN) analysis in step 2, C. node The larger the value of S, the more important the node. state It is a stability factor that characterizes the state of a system.

Images