Power grid simulation method and system based on intelligent control

Abstract

The invention relates to the technical field of crossing of power system automation and artificial intelligence, discloses a power grid simulation method and system based on intelligent control, and aims to solve the problems that a traditional power grid simulator is low in dynamic response precision, weak in multi-working-condition coverage capability and poor in nonlinear disturbance suppression. The method comprises the following steps: collecting and preprocessing multi-dimensional electrical data of key nodes of a power grid; high-dimensional state features are extracted through the CNN-LSTM hybrid network; a topology dynamic reconstruction module is driven to generate a real-time electrical connection map; calling an adaptive load model library to fit optimal parameters; importing an improved Newton-Raphson engine to carry out load flow calculation; four types of control instructions are generated, issued and executed; and triggering online fine tuning of the model based on residual analysis. The system is composed of eight functional subsystems and supports hundred thousand node level dynamic simulation.

Description

Technical Field

[0001] This invention belongs to the field of power system simulation technology, specifically relating to a power grid simulation method and system based on intelligent control. Background Technology

[0002] With the deepening of smart grid construction and the large-scale grid connection of new energy sources, power system operation simulation has placed higher demands on dynamic response accuracy, multi-source collaborative control capabilities, and adaptability to complex operating conditions. Traditional power grid simulation methods mostly rely on static parameter modeling and offline calculation frameworks, and their core logic is based on preset boundary conditions and steady-state assumptions.

[0003] However, modern power grids exhibit strong nonlinearity, high-dimensional coupling, and frequent real-time disturbances: dynamic processes such as distributed power generation output fluctuations, instantaneous load demand jumps, and cascading reactions from protection device actions are intertwined, making it difficult for fixed models to accurately depict the system's true evolution trajectory. Static simulation strategies, lacking online perception and closed-loop feedback mechanisms for power grid operation, are prone to problems such as distorted power flow calculations, misjudgments of transient stability, and delayed control commands, exacerbating the risks of voltage exceeding limits, frequency instability, and even local grid disconnection, severely restricting the timeliness and safety of dispatching decisions.

[0004] Furthermore, in the face of multiple uncertainties such as extreme weather, equipment failure, and market transactions, the rigid architecture of traditional methods cannot achieve adaptive reconfiguration of control strategies and is difficult to support the elastic operation requirements under the coordinated optimization of "source-grid-load-storage".

[0005] Therefore, a power grid simulation method and system based on intelligent control is desired. Summary of the Invention

[0006] This invention provides a power grid simulation method and system based on intelligent control. It achieves high-fidelity real-time simulation of complex power grid operating states by constructing a multi-source heterogeneous power data fusion architecture and combining a dynamic topology reconfiguration mechanism with an adaptive load response model. The system deploys a distributed sensor node array at the physical layer, establishes a hierarchical control structure at the logic layer, and introduces a closed-loop feedback regulation strategy based on state-space prediction at the algorithm layer. This overcomes the technical bottlenecks of traditional power grid simulators in terms of dynamic response accuracy, multi-condition coverage, and nonlinear disturbance suppression.

[0007] On the one hand, a power grid simulation method based on intelligent control is provided, which includes: S1 acquires the original measurement sequences of three-phase voltage instantaneous values, three-phase current instantaneous values, active power, reactive power, frequency deviation and harmonic distortion rate of each key node of the power grid through voltage transformers, current transformers and frequency synchronous acquisition units deployed on transmission lines, substation busbars and distribution terminals. S2, perform timestamp alignment on the original measurement sequence, use cubic spline interpolation to unify the data of different sampling periods to a reference sampling frequency of 100 times per second, and remove abnormal data points whose amplitude exceeds the rated value by ±20% or whose rate of change exceeds 5 volts per ampere per millisecond; S3 inputs the preprocessed multidimensional time series data into a hybrid feature extractor composed of a convolutional neural network and a long short-term memory network. The convolutional neural network is responsible for extracting local transient impact features, while the long short-term memory network is responsible for capturing the steady-state evolution trend across time steps. The outputs of the two are spliced ​​together to form a high-dimensional implicit representation vector of the power grid operation status. S4. Based on the high-dimensional implicit representation vector, drive the power grid topology dynamic reconstruction module. This module automatically adjusts the conduction state of the virtual switch matrix according to the current load distribution density, line impedance change gradient and protection device action signal, and generates an electrical connection relationship map that strictly corresponds to the actual operating conditions. S5. Based on the electrical connection diagram, the adaptive load equivalent model library is called. This model library contains four types of basic load units: constant power, constant current, constant impedance, and composite dynamic. Each type of load unit is equipped with three adjustable parameters: temperature coefficient correction factor, voltage sensitivity index, and frequency response attenuation constant. The optimal values ​​of each parameter are determined by the least squares fitting algorithm so that the mean square error between the equivalent load output characteristic curve and the measured load volt-ampere characteristic curve is less than three per thousand. S6. The reconstructed power grid topology and parameterized load model are jointly imported into the real-time solution engine. The engine uses the improved Newton-Raphson iterative method for power flow calculation. The initial iteration step size is set to 0.01. The convergence criterion is that the maximum value of the node voltage magnitude correction is less than the negative fifth power per unit value. If the convergence condition is not met for 20 consecutive iterations, the Jacobian matrix sparsification reset procedure is started. S7. Generate a control command sequence based on the power flow calculation results. The control command sequence includes four types of action variables: generator excitation adjustment, number of static var compensator switching groups, on-load tap changer tap position offset value, and distributed power output correction coefficient. The sequence is transmitted to the actuator drive layer via the industrial Ethernet bus. S8. After the actuator completes its action, the actual response data of the power grid is collected again and residual analysis is performed on it and the simulation prediction value. If the absolute value of the voltage amplitude deviation of a certain node continues to exceed the rated value by 0.5% for more than three sampling cycles, the online update mechanism of the model is triggered. The weight matrix of the hybrid feature extractor is fine-tuned using the incremental gradient descent method. The initial value of the learning rate is set to 0.001 and gradually reduced according to the exponential decay law.

[0008] On the other hand, a power grid simulation system based on intelligent control is provided, which includes: The multi-source data acquisition subsystem is used to acquire the original measurement sequences of three-phase voltage instantaneous values, three-phase current instantaneous values, active power, reactive power, frequency deviation and harmonic distortion rate of each key node of the power grid through voltage transformers, current transformers and frequency synchronous acquisition units deployed on transmission lines, substation busbars and distribution terminals. The data preprocessing subsystem is used to perform timestamp alignment on the original measurement sequence, and uses cubic spline interpolation to unify the data of different sampling periods to a reference sampling frequency of 100 times per second, and removes abnormal data points whose amplitude exceeds the rated value by ±20% or whose rate of change exceeds 5 volts per ampere per millisecond. The state feature extraction subsystem is used to input the preprocessed multidimensional time series data into a hybrid feature extractor composed of a convolutional neural network and a long short-term memory network. The convolutional neural network is responsible for extracting local transient impact features, while the long short-term memory network is responsible for capturing the steady-state evolution trend across time steps. The outputs of the two are spliced ​​together to form a high-dimensional implicit representation vector of the power grid operation state. The topology dynamic reconfiguration subsystem is used to drive the power grid topology dynamic reconfiguration module based on the high-dimensional implicit representation vector. This module automatically adjusts the conduction state of the virtual switch matrix according to the current load distribution density, line impedance change gradient and protection device action signal, and generates an electrical connection relationship map that strictly corresponds to the actual operating conditions. The load parameter identification subsystem is used to call the adaptive load equivalent model library based on the electrical connection relationship diagram. The model library contains four types of basic load units: constant power, constant current, constant impedance, and composite dynamic. Each type of load unit is equipped with three adjustable parameters: temperature coefficient correction factor, voltage sensitivity index, and frequency response attenuation constant. The optimal values ​​of each parameter are determined by the least squares fitting algorithm so that the mean square error between the equivalent load output characteristic curve and the field measured load volt-ampere characteristic curve is less than three per thousand. The real-time power flow solution subsystem is used to import the reconstructed power grid topology and parameterized load model into the real-time solution engine. The engine uses the improved Newton-Raphson iterative method for power flow calculation. The initial iteration step size is set to 0.01. The convergence criterion is that the maximum value of the node voltage magnitude correction is less than the negative fifth power per unit value. If the convergence condition is not met after 20 consecutive iterations, the Jacobian matrix sparsification reset procedure is started. The control command generation subsystem is used to generate a control command sequence based on the power flow calculation results. The control command sequence includes four types of action variables: generator excitation adjustment, number of static var compensator switching groups, tap position offset of on-load tap changer, and output correction coefficient of distributed power source. It is transmitted to the actuator drive layer through the industrial Ethernet bus. The online model update subsystem is used to collect the actual response data of the power grid again after the actuator completes the action, and perform residual analysis on it with the simulation prediction value. If the absolute value of the voltage amplitude deviation of a certain node continues to exceed 0.5% of the rated value for more than three sampling periods, the online model update mechanism is triggered. The weight matrix of the hybrid feature extractor is fine-tuned using the incremental gradient descent method. The initial value of the learning rate is set to 0.001 and gradually reduced according to the exponential decay law.

[0009] The present invention further explains that the frequency synchronization acquisition unit in the multi-source data acquisition subsystem uses the BeiDou satellite timing signal as the time reference source, and its time synchronization accuracy is better than ten nanoseconds, ensuring that the measurement data of the entire network has a unified time coordinate system.

[0010] The present invention further explains that the convolutional neural network in the hybrid feature extractor contains five convolutional layers. The first layer has a kernel size of 3x1, a stride of 1, and the same padding method, with the activation function being a modified linear unit. The kernel sizes of the second to fifth layers are 5x1, 7x1, 9x1, and 11x1 respectively, with the remaining parameters remaining consistent. The long short-term memory network contains two layers of recurrent units, with 256 hidden units in each layer, and the forget gate bias term is initialized to 1.

[0011] The present invention further explains that the virtual switch matrix is ​​stored using a sparse Boolean matrix data structure. The matrix row index corresponds to the power supply side node number, and the column index corresponds to the load side node number. When the element value is one, it indicates that the branch is in a closed conduction state, and when the element value is zero, it indicates that the branch is in a disconnected isolation state. The matrix dimension dynamically expands with the scale of the power grid, and supports a topology description of up to 100,000 nodes.

[0012] The present invention further explains that the composite dynamic load unit in the adaptive load equivalent model library is composed of a mechanical inertial element and an electromagnetic transient element connected in series. The time constant of the mechanical inertial element ranges from 0.1 seconds to 5 seconds, and the time constant of the electromagnetic transient element ranges from 0.001 seconds to 0.1 seconds. The ratio coefficient between the two is determined by inversion of the field waveform data.

[0013] The present invention further explains that the improved Newton-Raphson iterative method introduces a damping factor in each iteration. The initial value of the damping factor is set to 0.8. When the ratio of the correction norm of two adjacent iterations is greater than 1.5, the damping factor is multiplied by 0.9 to decay until the ratio falls back to below 1.2 and then the original value is restored.

[0014] The present invention further explains that the control command generation subsystem has a built-in priority arbitration mechanism. When multiple control objectives conflict, the adjustment resources are allocated according to the principle of voltage stability first, frequency recovery second, and economy last. Specifically, the generator excitation adjustment command has the highest execution priority, followed by the static var compensator command, then the on-load tap changer command, and finally the distributed power source output correction command.

[0015] The present invention further explains that the online model update mechanism only adjusts the weight parameters of the last two fully connected layers in the hybrid feature extractor, freezes the parameters of the remaining layers to avoid catastrophic forgetting, and limits the number of training samples used for each update to the effective data points collected in the most recent 500 sampling periods.

[0016] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: By constructing an end-to-end smart grid simulation architecture, this invention realizes fully automated processing from raw measurement data to control command output, eliminating response delays and operational errors caused by manual intervention. The multi-source heterogeneous data fusion mechanism ensures the integrity and timeliness of input information; the hybrid neural network feature extractor effectively captures the nonlinear dynamic coupling relationships in grid operation; the dynamic topology reconstruction module enables the simulation model to track grid structure changes in real time; and the adaptive load modeling method significantly improves the equivalent accuracy on the load side. The closed-loop feedback regulation and online model update mechanism endow the system with continuous self-optimization capabilities, maintaining stable and reliable simulation performance even in complex scenarios such as large-scale integration of new energy sources and drastic load fluctuations. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention.

[0018] Figure 2 This is a flowchart illustrating the main stages of the process from multi-source data acquisition to closed-loop control command generation in this invention. Detailed Implementation

[0019] The following detailed, non-limiting description of the technical solution of the present invention, in conjunction with preferred embodiments and accompanying drawings, is provided. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 and Figure 2This invention provides a power grid simulation method and system based on intelligent control, constructing a power grid simulation system with high dynamic response capability, strong environmental adaptability, and self-optimizing evolution characteristics. Traditional power grid simulation technologies mostly rely on static parameter modeling and offline calculation processes. When facing complex operating scenarios such as the continuous increase in the penetration rate of new energy sources, the drastic increase in load fluctuations, and frequent changes in topology, they generally suffer from fundamental defects such as insufficient simulation accuracy, significant response delay, and weak model generalization ability.

[0021] Especially with the large-scale integration of distributed power sources, the power grid operation exhibits highly nonlinear, strongly coupled, and time-varying uncertainties. Existing simulation tools are unable to capture transient processes such as voltage surges / dips, frequency oscillations, and harmonic distortions in real time, leading to lag or even failure of control strategies.

[0022] To address the aforementioned issues, this invention proposes an intelligent control architecture that integrates multi-source heterogeneous data acquisition, deep feature extraction, dynamic topology reconstruction, adaptive parameter identification, real-time power flow solution, and closed-loop feedback update. This architecture enables fully automated closed-loop processing from physical quantity perception to control command output, ensuring that simulation results and real power grid behavior maintain a high degree of consistency on a millisecond-level time scale.

[0023] A power grid simulation method based on intelligent control includes the following steps: S1 acquires raw measurement sequences of three-phase voltage instantaneous values, three-phase current instantaneous values, active power, reactive power, frequency deviation, and harmonic distortion rate covering the entire network by deploying voltage transformers, current transformers, and frequency synchronization acquisition units at key nodes of transmission lines, busbars of substations, and user sides of distribution terminals.

[0024] Among them, the voltage transformer uses the principle of electromagnetic induction to convert the primary high-voltage signal into the secondary low-voltage standard signal proportionally. The measurement range covers the level of 10 kV to 500 kV, and the accuracy level is not lower than zero point level 2.

[0025] Current transformers achieve high-current isolation sampling based on Rogowski coils or iron core structures, with a rated input current range of 100 amperes to 5,000 amperes and an angle difference controlled within one degree.

[0026] The frequency synchronization acquisition unit uses the BeiDou satellite timing signal as the global time reference source. Its internal crystal oscillator is locked to nanosecond-level time pulses to ensure that the timestamp error of all measuring devices is less than ten nanoseconds, thereby establishing a unified time coordinate system for the entire network and eliminating phase distortion and power calculation deviation caused by asynchronous sampling.

[0027] All raw measurement data are stored in a local cache in floating-point format and are accompanied by a 16-bit cyclic redundancy check code for transmission integrity verification.

[0028] S2, perform timestamp alignment on the original measurement sequence, and use cubic spline interpolation to unify the data of different sampling periods to a reference sampling frequency of 100 times per second.

[0029] Specifically, for the sampling sequence of fifty times per second output from the voltage transformer and the sampling sequence of two hundred times per second output from the current transformer, the absolute time value of each timestamp is first extracted to construct a time axis grid in microseconds. Then, cubic spline function fitting is performed on the missing sampling points on the time axis, and a piecewise cubic polynomial is constructed using four adjacent known data points to ensure that the interpolation curve has continuous first and second derivatives at the nodes, thereby preserving the high-frequency details of the original signal to the greatest extent.

[0030] After interpolation is completed, outlier removal is performed on all data points. If the amplitude of a sampling point exceeds the range of ±20% of the corresponding electrical quantity rating, or the rate of change between two adjacent points exceeds the threshold of 5 volts per ampere per millisecond, it is determined to be invalid data and is removed. The positions of the removed data are filled by the linear interpolation results of the valid points before and after, so as to avoid introducing false mutations.

[0031] The preprocessed multidimensional time series data forms a structured tensor. Its first dimension represents the time step, the second dimension represents the measurement channel number, and the third dimension represents the specific electrical quantity type. The total data size increases linearly with the number of power grid nodes. In typical regional power grid scenarios, it can achieve a throughput of 100,000 feature vectors per second.

[0032] S3 inputs the preprocessed multidimensional time series data into a hybrid feature extractor composed of a convolutional neural network and a long short-term memory network to generate a high-dimensional implicit representation vector of the power grid operation status.

[0033] This hybrid feature extractor employs a dual-branch parallel architecture, with the left branch consisting of a five-layer one-dimensional convolutional neural network and the right branch consisting of a two-layer stacked long short-term memory network.

[0034] The first layer of the convolutional neural network is configured with a 3x1 kernel, a fixed stride of 1, and the same padding strategy. The activation function is a modified linear unit, which is used to extract local transient impact features in voltage / current waveforms, such as spike pulses caused by short-circuit faults or step jumps caused by switching actions. The kernel sizes of the second to fifth layers are successively expanded to 5x1, 7x1, 9x1, and 11x1, while the other parameters remain unchanged. By increasing the receptive field layer by layer, multi-scale feature capture from microscopic perturbations to macroscopic trends is achieved.

[0035] Each layer of the Long Short-Term Memory Network contains 256 hidden units. The forget gate bias term is initialized to one to alleviate the gradient vanishing problem. Its core function is to model the steady-state evolution across time steps, such as the slow load ramp-up or the generator inertial response process.

[0036] The output feature maps of the two branches are concatenated along the channel dimension to form a joint representation vector of length 1024. The joint representation vector contains both local transient characteristics and long-term memory dependencies, which form the basic input for subsequent topology reconstruction and load identification.

[0037] The weight parameters of the hybrid feature extractor are pre-trained offline using historical power grid operation datasets in the initial stage. The loss function is defined as the mean square error between the predicted voltage trajectory and the measured trajectory. The optimization algorithm adopts the stochastic gradient descent method with a driving term. The batch size is set to 128, and the training rounds are no less than 500 until the validation set error converges.

[0038] S4. Based on the high-dimensional implicit representation vector, drive the power grid topology dynamic reconstruction module to generate an electrical connection relationship map that strictly corresponds to the measured operating conditions.

[0039] The power grid topology dynamic reconfiguration module maintains a virtual switch matrix in sparse Boolean form. The matrix row index corresponds to the node number on the power source side, and the column index corresponds to the node number on the load side. When the element value is one, it indicates that the branch is in a closed conducting state, and when the element value is zero, it indicates that the branch is in an open isolation state.

[0040] The matrix dimension expands dynamically according to the current power grid scale, supporting topology description capabilities up to 100,000 nodes.

[0041] The refactoring process is divided into three sub-phases: First, the load distribution density information encoded in the high-dimensional implicit representation vector is analyzed to identify the spatial clustering patterns of high-load and light-load areas, and the tie-line switching state is adjusted accordingly to balance the power flow between areas. Secondly, the gradient component of the line impedance change is analyzed. When the equivalent impedance growth rate of a certain section of the transmission line exceeds the preset threshold of 0.05 ohms per kilometer, it is determined to be parameter drift caused by line aging or icing. The line is automatically disconnected and a backup path is activated. Finally, the monitoring protection device action signal input port detects a relay protection output command such as overcurrent tripping or distance protection action, and immediately sets the corresponding branch to zero in the virtual switch matrix to simulate the actual circuit breaker tripping behavior.

[0042] After reconstruction, the system calls the graph theory algorithm library to perform connectivity checks and loop detection on the newly generated electrical connection relationship graph to ensure that there are no isolated nodes or illegal closed loop structures. Finally, it outputs a standardized adjacency matrix and node attribute list for downstream modules to use.

[0043] S5. Based on the electrical connection diagram, call the adaptive load equivalent model library to determine the optimal parameter values ​​for various load units.

[0044] The adaptive load equivalent model library includes four types of basic load units: constant power, constant current, constant impedance, and composite dynamic. Each type of load unit is equipped with three adjustable parameters: temperature coefficient correction factor, voltage sensitivity index, and frequency response attenuation constant.

[0045] The parameter identification process uses the least squares fitting algorithm. The objective function is defined as the mean square error between the equivalent load output characteristic curve and the measured load current-voltage characteristic curve. The constraint condition requires that the error value must be less than three per thousand.

[0046] In practice, the system first determines the load type label of each load node based on the electrical connection relationship diagram. For example, residential areas are classified as constant impedance type, industrial areas as composite dynamic type, and commercial areas as constant power type. Then, for each load node, voltage-current sampling pairs within the last ten complete power frequency cycles are collected to construct an observation matrix. Next, the observation matrix is ​​substituted into the mathematical model of the corresponding load type, and the optimal solution of parameters is solved using the Gauss-Newton iteration method. The iteration termination condition is that the parameter update norm is less than 10 to the power of negative 6 or the maximum number of iterations reaches 100.

[0047] For the composite dynamic load unit, its internal structure consists of a mechanical inertial element and an electromagnetic transient element connected in series. The time constant of the mechanical inertial element is limited to the range of 0.1 to 5 seconds, while the time constant of the electromagnetic transient element is limited to the range of 0.001 to 0.1 seconds. The ratio coefficient between the two is determined by inversion from the field waveform data. The inversion process uses a particle swarm optimization algorithm to search for the global optimum, with a population size of fifty and a maximum number of iterations of two hundred. The parameter identification results are stored in a distributed in-memory database in key-value pairs, supporting millisecond-level read and write access.

[0048] S6 imports the reconstructed power grid topology and parameterized load model into the real-time solver engine and executes the improved Newton-Raphson iterative method for power flow calculation.

[0049] When the real-time solver engine is initialized, the latest version of the node admittance matrix and load parameter table is loaded, the initial iteration step size is set to 0.01, and the convergence criterion is that the maximum value of the voltage amplitude correction of any node is less than the negative fifth power per unit value.

[0050] In each iteration, the system first constructs a Jacobian matrix, whose elements consist of the partial derivatives of the node injected power with respect to the voltage magnitude and phase angle. Then, a damping factor mechanism is introduced. The initial value of the damping factor is set to 0.8. When the ratio of the correction norm between two adjacent iterations is greater than 1.5, the damping factor is multiplied by 0.9 to decay until the ratio falls back to below 1.2 and then the original value is restored. This measure effectively suppresses the oscillation and divergence phenomenon in the solution of the nonlinear equation system.

[0051] If the convergence condition is not met after 20 consecutive iterations, the Jacobian matrix sparsification reset procedure is triggered. The system scans the elements in the matrix whose absolute value is less than 10 to the power of -8, forces them to be set to zero, and recalculates the sparsity mode to reduce the matrix condition number and improve numerical stability.

[0052] The power flow calculation results include key indicators such as the voltage amplitude and phase angle of each node, the active and reactive power of each branch, and the total network loss. These data are encapsulated into structured messages and passed to the control command generation module through a shared memory buffer.

[0053] S7. Generate a control command sequence based on the power flow calculation results. The control command sequence includes four types of action variables: generator excitation adjustment, number of static var compensator switching groups, tap position offset of on-load tap changer, and output correction coefficient of distributed power source.

[0054] The instruction generation process has a built-in priority arbitration mechanism. When multiple control objectives conflict, regulation resources are allocated according to the principle of prioritizing voltage stability, followed by frequency recovery, and lastly, economy.

[0055] Specifically, the generator excitation regulation command has the highest execution priority, and its regulation amount is directly related to the generator terminal voltage amplitude control. The calculation formula is as follows: ; in For proportional gain, For integral gain, For the target voltage, This is the measured voltage.

[0056] The Static Var Compensator (SVC) command is next, with the number of switching groups determined by the reactive power deficit. Each switching group of capacitors can provide 50 Mvars of reactive power support. The On-Load Tap Changer (OTCT) command is next, with the tap position offset value calculated based on the central point voltage deviation. Each adjustment can change the voltage by 1.25 percent. The Distributed Power Generation (DPG) Output Correction command is the lowest, and is only activated when the system frequency deviation exceeds 0.05 Hz and other adjustment methods have reached their limits.

[0057] All control commands are formatted and encoded, and then transmitted to the actuator drive layer via the industrial Ethernet bus. The communication protocol adopts the IEC 61850 standard, and the message transmission delay is controlled within five milliseconds.

[0058] S8: After the actuator completes its action, the actual response data of the power grid is collected again, and residual analysis is performed between the data and the simulation prediction value to trigger the online model update mechanism.

[0059] The system continuously monitors the absolute value of voltage amplitude deviation at each key node. If the deviation of a node exceeds 0.5% of the rated value for more than three sampling cycles, it is determined to be a model mismatch event, and the incremental gradient descent method is started to fine-tune the weight matrix of the hybrid feature extractor.

[0060] The update process only adjusts the parameters of the last two fully connected layers, freezing the parameters of the remaining layers to avoid catastrophic amnesia.

[0061] The initial learning rate is set to 0.001 and gradually decreased according to the exponential decay law, with a decay coefficient of 0.95.

[0062] The number of training samples used for each update is limited to valid data points collected within the last 500 sampling periods. The sample set is dynamically refreshed through a sliding window mechanism to ensure that the model always focuses on the latest operating conditions.

[0063] After the weights are updated, the system automatically saves a snapshot of the new model and records the version number. At the same time, it sends a model iteration notification log to the operations and maintenance personnel, which includes information such as the data period used for this update, the magnitude of the loss function decrease, and statistics on the changes in the main parameters.

[0064] The system's hardware implementation relies on a distributed computing cluster architecture. The main control server is equipped with a multi-core central processing unit and a graphics processing unit accelerator card, which is responsible for running the hybrid feature extractor and the power flow solving engine. The data acquisition front end is deployed with an industrial-grade embedded controller, which has a built-in high-precision analog-to-digital converter chip and a Beidou time synchronization module. The communication network adopts a dual-ring redundant fiber optic Ethernet with a bandwidth of no less than gigabits per second, and supports the IEEE 1588 precision time protocol to achieve nanosecond-level clock synchronization.

[0065] At the software level, a microservice design pattern is adopted, with each functional subsystem deployed independently in a containerized environment. Asynchronous communication and load balancing are achieved through message queues. The system supports hot-swappable module replacement and online firmware upgrades, ensuring that algorithm iteration and performance optimization can be completed without downtime.

[0066] The technical effects of this invention are reflected in multiple dimensions: Firstly, by using a multi-source heterogeneous data fusion mechanism, the information loss caused by the asynchronous nature of sensor sampling and the inconsistency of dimensions is eliminated, ensuring the integrity and timeliness of the input data; Secondly, the hybrid neural network feature extractor effectively captures the nonlinear dynamic coupling relationship in power grid operation, overcoming the high dependence of traditional feature engineering methods on expert experience. Third, the dynamic topology reconfiguration module enables the simulation model to track changes in the power grid structure in real time, solving the model distortion problem caused by fixed topology assumptions; Fourth, the adaptive load modeling method significantly improves the equivalent accuracy of the load side, especially showing excellent generalization ability in complex dynamic load scenarios; Fifth, the closed-loop feedback regulation and online model update mechanism endow the system with continuous self-optimization capabilities, and can still maintain stable and reliable simulation performance when facing complex scenarios such as large-scale access of new energy sources and drastic load fluctuations.

[0067] The aforementioned technical features work together to construct a high-fidelity real-time simulation platform for future smart grids, providing strong technical support for grid planning, operation scheduling, fault diagnosis, and control strategy verification.

[0068] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this invention, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. 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, and these 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 power grid simulation method based on intelligent control, characterized in that, include: The raw measurement sequences of instantaneous three-phase voltage, instantaneous three-phase current, active power, reactive power, frequency deviation, and harmonic distortion rate are collected by sensors deployed at key nodes of the power grid. The original measurement sequence is preprocessed, including timestamp alignment, interpolation to unify the sampling frequency, and removal of abnormal data with amplitude exceeding the limit or rate of change exceeding the threshold. The preprocessed multidimensional time-series data is input into the hybrid feature extractor to generate a high-dimensional latent representation vector of the power grid operating status. Based on the high-dimensional implicit representation vector, the power grid topology dynamic reconfiguration module is driven to generate an electrical connection relationship map that strictly corresponds to the measured operating conditions; Based on the electrical connection diagram, the load parameters are identified by calling the adaptive load equivalent model library. The reconstructed power grid topology and the parameterized load model are jointly imported into the real-time solver engine for power flow calculation. Based on the power flow calculation results, a sequence of control commands is generated and transmitted to the actuator drive layer via the industrial Ethernet bus. After the actuator completes its action, the actual response data of the power grid is collected again, and residual analysis is performed on the data and the simulation prediction values ​​to trigger the online model update mechanism.

2. The power grid simulation method based on intelligent control according to claim 1, characterized in that, The interpolation method is cubic spline interpolation, which unifies data from different sampling periods to a reference sampling frequency of one hundred times per second, including: Extract the absolute time values ​​of the timestamps from the 50 sampling sequences per second output by the voltage transformer and the 200 sampling sequences per second output by the current transformer, and construct a time axis grid in microseconds. On the time axis grid, a cubic spline function is fitted to the missing sampling points. A piecewise cubic polynomial is constructed using four adjacent known data points to ensure that the interpolation curve has continuous first and second derivatives at the nodes. After interpolation is completed, outlier removal is performed on all data points. If the amplitude of a sampling point exceeds the range of ±20% of the corresponding electrical quantity rating, or the rate of change between two adjacent points exceeds the threshold of 5 volts per ampere per millisecond, it is determined to be invalid data and is removed. The positions of the removed data are filled by the linear interpolation results of the valid points before and after.

3. The power grid simulation method based on intelligent control according to claim 1, characterized in that, The preprocessed multidimensional time series data is input into a hybrid feature extractor consisting of a convolutional neural network and a long short-term memory network. The convolutional neural network is responsible for extracting local transient impact features, while the long short-term memory network is responsible for capturing the steady-state evolution trend across time steps. The outputs of the two are concatenated to form the high-dimensional latent representation vector. The multidimensional time series data is input into the five-layer one-dimensional convolutional neural network of the left branch. The first layer is configured with a 3x1 convolutional kernel, the stride is fixed at 1, the padding method adopts the same padding strategy, and the activation function is the modified linear unit. The kernel size of the second to fifth layers is expanded to 5x1, 7x1, 9x1, and 11x1 respectively, while the other parameters remain unchanged. The multidimensional time-series data is simultaneously input into the two-layer stacked long short-term memory network on the right branch. Each layer contains 256 hidden units, and the forget gate bias term is initialized to one. The feature maps output from the two branches are concatenated along the channel dimension to form a joint representation vector of length 1024, which serves as the high-dimensional latent representation vector.

4. The power grid simulation method based on intelligent control according to claim 1, characterized in that, When the topology dynamic reconstruction module generates the electrical connection relationship map, it includes: The load distribution density information encoded in the high-dimensional implicit representation vector is analyzed to identify the spatial clustering patterns of high-load and light-load areas, and the tie line switching state is adjusted accordingly to balance the power flow between areas. Analyze the gradient components of line impedance change. When the equivalent impedance growth rate of a certain section of transmission line exceeds the preset threshold of 0.05 ohms per kilometer, it is determined to be parameter drift caused by line aging or icing. The line is automatically disconnected and a backup path is activated. The monitoring protection device's action signal input port detects a relay protection output command and immediately sets the corresponding branch to zero in the virtual switch matrix to simulate the actual circuit breaker's tripping behavior. The graph theory algorithm library is used to perform connectivity checks and loop detection on the newly generated electrical connection graph to ensure that there are no isolated nodes or illegal closed loop structures.

5. The power grid simulation method based on intelligent control according to claim 1, characterized in that, The adaptive load equivalent model library contains four types of basic load units: constant power, constant current, constant impedance, and composite dynamic. Each type of load unit is equipped with three adjustable parameters: temperature coefficient correction factor, voltage sensitivity index, and frequency response attenuation constant. The optimal values ​​of each parameter are determined by the least squares fitting algorithm, so that the mean square error between the equivalent load output characteristic curve and the field measured load volt-ampere characteristic curve is less than three per thousand. Based on the electrical connection relationship diagram, determine the load type label of each load node. For each load node, collect voltage-current sampling pairs within the last ten complete power frequency cycles to construct an observation matrix. Substitute the observation matrix into the mathematical model of the corresponding load type and use the Gauss-Newton iteration method to solve the optimal solution of the parameters. The iteration termination condition is that the parameter update norm is less than 10 to the power of negative 6 or the maximum number of iterations reaches 100. For the composite dynamic load unit, its internal structure consists of a mechanical inertial element and an electromagnetic transient element connected in series. The time constant of the mechanical inertial element is limited to the range of 0.1 to 5 seconds, and the time constant of the electromagnetic transient element is limited to the range of 0.001 to 0.1 seconds. The ratio coefficient between the two is determined by inversion of the field recorded waveform data.

6. The power grid simulation method based on intelligent control according to claim 1, characterized in that, The real-time solution engine uses the improved Newton-Raphson iterative method for power flow calculation. The initial iteration step size is set to 0.01, and the convergence criterion is that the maximum value of the node voltage amplitude correction is less than 10 to the power of negative 5 per unit. In each iteration, a Jacobian matrix is ​​constructed, and a damping factor mechanism is introduced. The initial value of the damping factor is set to 0.

8. When the ratio of the correction norm between two adjacent iterations is greater than 1.5, the damping factor is multiplied by 0.9 to decay until the ratio falls back to below 1.2 and then the original value is restored. If the convergence condition is not met after 20 consecutive iterations, the Jacobian matrix sparsification reset procedure is triggered, which scans the elements in the matrix whose absolute value is less than 10 to the power of -8, forces them to be set to zero, and recalculates the sparsity mode.

7. The power grid simulation method based on intelligent control according to claim 1, characterized in that, The control command sequence includes four types of action variables: generator excitation regulation, number of static var compensator switching groups, tap position offset of on-load tap changer, and output correction coefficient of distributed power source. These variables are transmitted to the actuator drive layer via an industrial Ethernet bus. The command generation process has a built-in priority arbitration mechanism. When multiple control objectives conflict, regulation resources are allocated according to the principle of voltage stability first, frequency recovery second, and economy last. Generator excitation regulation commands have the highest execution priority, followed by static var compensator commands, then on-load tap-changing transformer commands, and distributed generation output correction commands have the lowest priority.

8. The power grid simulation method based on intelligent control according to claim 1, characterized in that, The online model update mechanism includes: Continuously monitor the absolute value of voltage amplitude deviation at each key node. If the deviation of a node continues to exceed 0.5% of the rated value for more than three sampling cycles, it is determined to be a model mismatch event. Initiate incremental gradient descent to fine-tune the weight matrices of the last two fully connected layers in the hybrid feature extractor, and freeze the parameters of the remaining layers. The initial learning rate is set to 0.001 and gradually decreased according to the exponential decay law, with a decay coefficient of 0.

95. The number of training samples used for each update is limited to valid data points collected within the last 500 sampling periods, and the sample set is dynamically refreshed through a sliding window mechanism.

9. The power grid simulation method based on intelligent control according to claim 1, characterized in that, The sensor includes voltage transformers, current transformers, and frequency synchronization acquisition units deployed on transmission lines, substation busbars, and distribution terminals. The frequency synchronization acquisition unit uses the BeiDou satellite timing signal as the time reference source, and its time synchronization accuracy is better than ten nanoseconds, ensuring that the measurement data of the entire network has a unified time coordinate system.

10. A power grid simulation system based on intelligent control, characterized in that, include: The multi-source data acquisition subsystem acquires raw measurement sequences of instantaneous three-phase voltage, instantaneous three-phase current, active power, reactive power, frequency deviation, and harmonic distortion rate through sensors deployed at key nodes of the power grid. The data preprocessing subsystem is used to preprocess the original measurement sequence, including timestamp alignment, interpolation to unify the sampling frequency, and removing abnormal data with amplitude exceeding the limit or rate of change exceeding the threshold. The state feature extraction subsystem is used to input preprocessed multidimensional time series data into the hybrid feature extractor to generate a high-dimensional implicit representation vector of the power grid operating state. The topology dynamic reconfiguration subsystem is used to drive the power grid topology dynamic reconfiguration module based on the high-dimensional implicit representation vector to generate an electrical connection relationship map that strictly corresponds to the measured operating conditions. The load parameter identification subsystem is used to identify load parameters based on the electrical connection relationship diagram by calling the adaptive load equivalent model library. The real-time power flow solution subsystem is used to import the reconstructed power grid topology and parameterized load model into the real-time solution engine for power flow calculation. The control command generation subsystem is used to generate a sequence of control commands based on the power flow calculation results and transmit them to the actuator drive layer via the industrial Ethernet bus. The online model update subsystem is used to collect the actual response data of the power grid again after the actuator completes its action, perform residual analysis on the data and simulation prediction values, and trigger the online model update mechanism.

Images