Reactive power output adjusting method for high-voltage static synchronous compensator

Abstract

The invention discloses a reactive output adjusting method for a high-voltage static synchronous compensator, which comprises a three-layer closed-loop architecture comprising a sensing layer, a control layer and an execution layer, and further comprises a layered multi-target model prediction MPC control unit, the control unit comprises a data preprocessing module, an online parameter identification module, an upper-layer global reactive power distribution module, a lower-layer bridge arm level MPC optimization module, a multi-target weight adaptive module and an AI error compensation module. Belongs to the technical field of high voltage, and can reduce the calculation complexity of the MPC algorithm, reduce the single-cycle control delay to be less than 5ms and meet the real-time control requirement of the high-voltage SVG on the premise of ensuring the reactive power regulation precision.

Description

Technical Field

[0001] This invention relates to the field of high-voltage electrical technology, specifically a method for regulating the reactive power output of a high-voltage static synchronous compensator. Background Technology

[0002] The Static Synchronous Compensator (SVG) is a core component for reactive power compensation and voltage stability control in power grids, and its control performance directly affects the power quality and operational stability of the grid. Model Predictive Control (MPC) has become the mainstream research direction for reactive power regulation of SVG due to its advantages such as strong multi-constraint handling capabilities and excellent nonlinear control performance. However, in practical engineering applications, existing technologies suffer from the following core bottlenecks, making it difficult to meet the high-performance requirements of high-voltage scenarios: 1. Insufficient Real-Time Performance: High-voltage SVG often employs modular multilevel converters (MMCs) and neutral-point clamped NPCs, among other multilevel topologies. MMCs are multilevel converter topologies composed of multiple sub-modules connected in series, offering advantages such as low output voltage harmonics and high withstand voltage ratings. Traditional MPCs, on the other hand, require traversing all switching states; for example, with 20 sub-modules on a single MMC bridge arm, the number of switching states reaches 2... 20 The computational complexity increases exponentially, and the single-cycle control delay usually exceeds 8ms, which cannot meet the millisecond-level (≤5ms) response requirements of high-voltage SVG. 2. Poor Multi-Objective Coordination: Existing solutions often focus solely on reactive power compensation accuracy as a control objective, neglecting key indicators such as switching losses, grid voltage harmonic distortion (THD), and voltage balance of multi-level topology capacitors. Some multi-objective MPC solutions employ a fixed-weighted summation method, where the weights cannot be dynamically adjusted according to operating conditions, resulting in low equipment operating efficiency and unstable grid-side performance. 3. Insufficient robustness: Traditional MPC relies on accurate mathematical models, but high-voltage power grids have complex operating conditions such as impedance fluctuations, load shocks, and voltage dips. The model parameters are prone to deviating from the design values, leading to a decrease in control accuracy or even system instability. At the same time, existing error correction methods are mostly fixed compensation and do not combine data-driven technology to achieve adaptive correction.

[0003] In existing technologies, solutions to real-time problems often employ fixed model reduction or simplification of the switch state set, but this can easily lead to a loss of control accuracy; solutions for multi-objective problems lack dynamic weight adjustment mechanisms; and solutions for robustness problems have not formed a collaborative mechanism that combines parameter identification with error compensation.

[0004] Therefore, it is urgent to develop a method for regulating the reactive power output of a high-voltage static synchronous compensator to solve the problems in the existing technology. Summary of the Invention

[0005] The purpose of this invention is to provide a method for regulating the reactive power output of a high-voltage static synchronous compensator (SVG). This method can reduce the computational complexity of the MPC algorithm and reduce the single-cycle control delay to less than 5ms while ensuring the accuracy of reactive power regulation. It meets the real-time control requirements of high-voltage SVG and has a simple structure and is easy to use, thus solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for regulating the reactive power output of a high-voltage static synchronous compensator (SSC) includes a three-layer closed-loop architecture comprising a sensing layer, a control layer, and an execution layer, and further includes a hierarchical multi-objective model predictive control (MPC) control unit. The control unit includes a data preprocessing module, an online parameter identification module, an upper-layer global reactive power allocation module, a lower-layer bridge arm-level MPC optimization module, a multi-objective weight adaptive module, and an AI error compensation module. The specific implementation includes the following steps: S1: The sensing layer collects power grid-side operation data and SVG device status data, transmits the raw data to the data preprocessing module for processing, and obtains standardized and effective data; S2: The online parameter identification module receives standardized and valid data, identifies key parameters of the power grid and SVG model in real time, and transmits the identified parameters to the corresponding downstream modules. S3: The upper-level global reactive power allocation module synchronously receives the total reactive power demand of the power grid, the status data of each bridge arm of the SVG and the above-mentioned identification parameters from the standardized valid data, allocates the total reactive power demand into reference reactive power values ​​for each bridge arm, and transmits the reference reactive power values ​​to the corresponding downstream modules. S4: The multi-objective weight adaptive module receives the operating parameters and real-time values ​​of each control sub-objective from the standardized valid data, calculates and determines the dynamic weights, and transmits them to the downstream optimization module. S5: The lower-level bridge arm MPC optimization module receives the bridge arm reference reactive power value, key model parameters and dynamic weights, and calculates and obtains the predicted reactive power value. S6: The AI ​​error compensation module synchronously receives the bridge arm reference reactive power value, key model parameters and the predicted reactive power value, outputs the error compensation amount based on the BP neural network model completed offline training, corrects the predicted reactive power value through the compensation amount to obtain the final control signal and transmits it to the execution layer. S7: The execution layer receives the final control signal, drives the SVG main circuit to output the target reactive power, and feeds back the actual reactive power output of the SVG and the equipment operating status to the perception layer to form a complete closed-loop control.

[0007] By adopting the above technical solution, through the three-layer closed-loop architecture of perception, control, and execution and the coordinated operation of six core modules, the entire process of automatic adjustment from grid and SVG data acquisition and processing to final reactive power output and status feedback is realized, ensuring that the reactive power output of SVG accurately matches the grid demand and forming a complete closed-loop control link.

[0008] As a further aspect of the present invention, the specific implementation process of parameter identification in S2 includes the following: the online parameter identification module receives standardized valid data output by the data preprocessing module, including grid voltage, SVG output current, and bridge arm capacitor voltage, and uses the improved extended Kalman filter (EKF) algorithm to identify key model parameters in real time, including grid impedance, SVG equivalent inductance and equivalent capacitance, and transmits the identified parameters to the upper-level global reactive power allocation module, the lower-level bridge arm-level MPC optimization module and the AI ​​error compensation module respectively.

[0009] By adopting the above technical solutions, the input data, core algorithms and output parameters for parameter identification are clearly defined, providing accurate power grid and SVG key model parameter support for upper-level reactive power allocation, lower-level MPC optimization and AI error compensation, laying the parameter foundation for subsequent regulation processes.

[0010] As a further aspect of the present invention: the improved Extended Kalman Filter (EKF) algorithm of the online parameter identification module in S2, the core improvement of which is the adaptive Kalman gain adjustment strategy, and the specific implementation process includes: S21: Define a convergence criterion, which includes the residual between the observed value and the predicted value obtained based on the prediction model; S22: Set the residual threshold, which is the product of the observed value and the preset scaling factor; S23: Adjust the Kalman gain according to the relationship between the residual and the threshold. When the absolute value of the residual is greater than the threshold, increase the Kalman gain to speed up the convergence of the algorithm; when the absolute value of the residual is less than or equal to the threshold, decrease the Kalman gain to ensure the accuracy of parameter identification. S24: Update the key model parameters of the power grid and SVG according to the preset identification cycle. The key model parameters include the power grid impedance, SVG equivalent inductance and equivalent capacitance.

[0011] By adopting the above technical solution and improving the EKF algorithm through an adaptive Kalman gain adjustment strategy, the convergence speed and accuracy of parameter identification are balanced, ensuring the reliability and stability of key model parameter identification under different working conditions and improving the robustness of the overall adjustment scheme.

[0012] As a further aspect of the present invention, the specific implementation process of reactive power allocation in S3 includes the following: The upper-level global reactive power allocation module receives the status data output by the data preprocessing module, including the total reactive power demand of the power grid, the capacitor voltage of each bridge arm of the SVG, and the temperature of the switching devices. Combining the key model parameters output by S2, based on the dynamic allocation algorithm of capacity weighting combined with state constraints, the capacitor voltage balance weight and temperature weight of each bridge arm are first calculated to obtain the comprehensive weight. Then, the total reactive power demand is allocated as the reference reactive power value of each bridge arm according to the comprehensive weight ratio. The reference reactive power value is transmitted to the lower-level bridge arm-level MPC optimization module and AI error compensation module to realize the reasonable allocation of reactive power demand and equipment status protection.

[0013] By adopting the above technical solution, combining the total reactive power demand of the power grid, SVG equipment status data and identification parameters, a dynamic allocation algorithm is used to achieve a reasonable breakdown of reactive power demand, while also taking into account equipment status protection, and avoiding equipment failures caused by uneven reactive power distribution.

[0014] As a further aspect of the present invention: the capacity-weighted dynamic allocation algorithm of the upper-layer global reactive power allocation module in S3, combined with state constraints, is specifically implemented as follows: S31: Calculate the capacitor voltage balance weight of each bridge arm. This weight is inversely proportional to the capacitor voltage balance of each bridge arm and is calculated by the degree of deviation between the capacitor voltage of a single bridge arm and the average value of the capacitor voltages of all bridge arms. S32: Calculate the temperature weight of each bridge arm. Based on the difference between the real-time temperature of the bridge arm switching device and the safe operating threshold, the weight is determined using a piecewise linear function. The closer the temperature is to the safe threshold, the smaller the weight. When the temperature exceeds the safe threshold, the weight is set to zero. S33: Multiply the capacitor voltage balance weight and temperature weight of the same bridge arm to obtain the comprehensive weight of each bridge arm. S34: Based on the proportion of the comprehensive weight of each bridge arm to the sum of the comprehensive weights of all bridge arms, the total reactive power demand of the power grid is allocated to the corresponding bridge arm to obtain the reference reactive power value of each bridge arm, and the sum of the reference reactive power values ​​of all bridge arms is equal to the total reactive power demand of the power grid.

[0015] By adopting the above technical solution, the implementation logic of the capacity weighted combined with state constraint algorithm is refined, and the comprehensive weight of each bridge arm is accurately calculated. This ensures that the reactive power distribution can balance the capacitor voltage of each bridge arm and avoid damage to the switching devices due to excessive temperature, thereby further improving the scientific nature of the distribution.

[0016] As a further aspect of the present invention: the multi-objective weight adaptive module in S4 includes an AHP fusion entropy weight method strategy, the specific implementation process of which includes: S41: Establish a hierarchical structure including a target layer, a criterion layer, and a scheme layer. The target layer is for multi-objective optimization, the criterion layer is for each control sub-objective, and the scheme layer is for each switching state. A judgment matrix is ​​constructed using a preset scaling method. By calculating the eigenvectors of the judgment matrix and performing a consistency check, the subjective basic weights of each control sub-objective are obtained. S42: Standardize the real-time values ​​of each control sub-target collected under different operating conditions to eliminate the influence of dimensions. Calculate the information entropy of each control sub-target based on the principle of information entropy. The smaller the information entropy, the greater the fluctuation of the target data, the more information it contains, and the greater the corresponding objective weight. S43: A linear weighted fusion method is adopted to combine subjective basic weights and objective weights. The proportion of subjective basic weights and objective weights is allocated by a preset ratio coefficient to obtain the dynamic weights of each control sub-target. S44: Repeat the above steps to update the dynamic weights according to the preset cycle to ensure that the weights can adapt to changes in operating conditions.

[0017] By adopting the above technical solution, the weights of control sub-objectives are dynamically determined through the AHP fusion entropy weight method strategy, so that the weights can adapt to changes in operating conditions in real time, ensuring that the multi-objective optimization effect meets the actual needs under different operating conditions and improving the operating condition adaptability of the adjustment scheme.

[0018] As a further aspect of the present invention: the lower-level bridge arm MPC optimization module in S5 includes a low-complexity strategy of switch state grouping, simplified model, and fast solution, and the specific implementation process includes: S51: The switching states of multiple sub-modules of a single bridge arm in the SVG multilevel topology are divided into three categories: magnetization group, demagnetization group, and holding group. The magnetization group includes key switching states that can increase the output voltage of the bridge arm to increase reactive power output. The demagnetization group includes key switching states that can decrease the output voltage of the bridge arm to reduce reactive power output. The holding group includes key switching states that can stabilize the output voltage of the bridge arm to maintain the reactive power output. Only the key switching states in the above three categories are selected for subsequent prediction. S52: Based on Kirchhoff's voltage law, minor parameters, including submodule parasitic resistance and stray capacitance, are ignored to construct a bridge arm-level low-order simplified model, reducing the computational complexity of the model; S53: The sparse linear programming algorithm is used to solve for the optimal switching state. Based on the dynamic weights output by the multi-objective weight adaptive module, a comprehensive objective function is constructed, including reactive power compensation accuracy and switching loss targets. The optimal switching state of the previous control cycle is used as the initial value to narrow the search range and quickly solve for the predicted reactive power value.

[0019] By adopting the above technical solutions, and through low-complexity strategies such as switching state grouping, simplified modeling, and fast solution, the computational load of lower-level MPC optimization is significantly reduced, the efficiency of predicting reactive power is improved, and the real-time requirements of SVG reactive power regulation are met.

[0020] As a further aspect of the present invention: the training process of the BP neural network model of the AI ​​error compensation module in S6 includes: S601: Build an SVG simulation platform to simulate three typical operating conditions: steady state, transient state, and complex state. Collect input and output parameters under each operating condition. The input parameters include the bridge arm reference reactive power value, SVG equivalent inductance, equivalent capacitance, and grid impedance. The output parameters include the MPC predicted reactive power value and the SVG actual output reactive power value. The error data is obtained by the difference between the two. S602: After denoising and normalizing the collected input parameters and error data, the training set and test set are divided. S603: Construct a BP neural network, including an input layer, hidden layers, and an output layer. Determine the number of nodes and activation functions for each layer. Train the neural network using a preset loss function and optimizer. Verify the model performance using a test set. If the performance does not meet the preset requirements, use preset techniques to suppress overfitting and retrain. After training, solidify the model into the control chip.

[0021] By adopting the above technical solution, a high-precision BP neural network error compensation model is constructed through multi-condition data acquisition, preprocessing, and model training, providing reliable model support for online error compensation and ensuring the accuracy of the compensation effect.

[0022] As a further aspect of the present invention: the online operation process of the BP neural network model of the AI ​​error compensation module in S6 includes: S611: Normalize the key model parameters output by the online parameter identification module and the bridge arm reference reactive power value output by the upper-level global reactive power allocation module, and then input them into the trained BP neural network model to obtain the error compensation amount. S612: Use this error compensation amount to correct the predicted reactive power value output by the lower-level bridge arm MPC optimization module to obtain the final control signal, so that the reactive power compensation error of the SVG meets the preset accuracy requirements.

[0023] By adopting the above technical solution, the online output error compensation amount of the trained BP neural network model is used to correct the predicted reactive power value to obtain a precise control signal, effectively reducing reactive power compensation error and improving the control accuracy of SVG reactive power output.

[0024] As a further aspect of the present invention: the execution layer in S7 includes a switching device driving circuit, an SVG main circuit, and a protection circuit. The driving circuit receives the final control signal and generates a driving pulse to drive the SVG main circuit to output the target reactive power. The protection circuit monitors the overcurrent, overvoltage, and overtemperature status of the SVG main circuit in real time. When an abnormal status is detected, it outputs a protection signal to cut off the driving circuit to protect the equipment safety.

[0025] By adopting the above technical solution, the core components and functions of the execution layer are clearly defined, enabling the effective execution of control signals and the target reactive power output. At the same time, the protection circuit monitors the equipment status in real time, avoiding abnormal risks such as overcurrent, overvoltage, and overtemperature, and ensuring the safe and stable operation of the equipment.

[0026] Compared with the prior art, the beneficial effects of the present invention are: This invention reduces computational load through a low-complexity MPC optimization strategy and accelerates convergence speed by combining adaptive parameter identification, thus solving the problems of computational complexity and response lag in traditional solutions. It adapts to the millisecond-level regulation requirements of high-voltage power grids and significantly improves the real-time performance and operating efficiency of SVG reactive power regulation.

[0027] This invention adapts to different operating conditions through dynamic weight adjustment, combines AI error compensation to correct prediction deviations, and takes into account both equipment condition protection and grid demand matching, thereby achieving high-precision and robust reactive power output regulation and significantly improving reactive power compensation accuracy and operating condition adaptability.

[0028] Other features and advantages of the present invention will be disclosed in detail in the following detailed description and accompanying drawings. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of an overall structure in an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1 In this embodiment of the invention, a method for regulating the reactive power output of a high-voltage static synchronizing compensator is described below. Figure 1As shown, the architecture includes a three-layer structure of perception, control, and execution. At its core is a hierarchical multi-objective MPC control unit, integrating six functional modules: data preprocessing, online parameter identification, upper-layer global reactive power allocation, lower-layer bridge arm-level MPC optimization, multi-objective weight adaptation, and AI error compensation. Its core logic includes: The perception layer collects real-time operating data from the power grid and SVG equipment, and inputs it into the control layer after preprocessing. The online parameter identification module updates model parameters in real time, providing an accurate foundation for MPC optimization; The upper-level global reactive power distribution module dynamically distributes the total reactive power demand to each bridge arm based on grid demand and equipment status. The lower-level bridge arm-level MPC optimization module uses a low-complexity strategy to solve for the optimal switching state; The multi-objective weight adaptive module dynamically adjusts the weights of each objective based on the operating conditions. The AI ​​error compensation module corrects the control output based on the offline trained model; Ultimately, the control signal drives the execution layer to achieve precise reactive power output, while the equipment operating status is fed back to the sensing layer to form a closed loop.

[0032] The sensing layer includes voltage and current sensors, power sensors, temperature sensors, and AD conversion modules; the core of the control layer is a hierarchical multi-objective MPC control unit, which includes the aforementioned six functional modules; the execution layer includes IGBT drive circuits, MMC main circuits, and protection circuits.

[0033] The detailed design of the core modules includes: 1. Low-complexity hierarchical MPC algorithm module This module adopts a layered architecture of "upper-level global reactive power allocation + lower-level bridge arm-level local optimization" to reduce computational complexity from the "spatial dimension". At the same time, it dynamically adjusts the allocation strategy in combination with the device status to avoid the drawbacks of traditional MPC global traversal.

[0034] (1) Upper layer: Global reactive power distribution module Input: The total reactive power demand Q on the grid side collected by the sensing layer, in units of... ; Voltage of each MMC bridge arm capacitor to Its unit is kV; the temperature of each IGBT switching device in the bridge arm to Its unit is ℃; Output: Reference reactive power values ​​for each bridge arm to Its unit is To satisfy all The sum equals Q; Among them, the core algorithm includes: adopting a dynamic allocation algorithm that combines capacity weighting with state constraints, and the specific steps are as follows: Step 1: Calculate the balance weight k of the capacitor voltage of each arm.

[0035] Define the capacitor voltage balance degree ε as the absolute value of the difference between the capacitor voltage V of a single arm and the average value V of the capacitor voltages of 6 arms, and then the ratio to V. k is inversely proportional to and the calculation formula is .

[0036] In this embodiment: If V = 20 kV and V = 21 kV, then , ; Step 2: Calculate the temperature weight k of each arm.

[0037] According to the safe operating threshold of the IGBT device ≤ 100 °C, a piecewise linear function is adopted.

[0038] When T ≤ 80 °C, k = 1. At this time, the device loss is small and it can operate at full load; when 80 °C < T ≤ 100 °C, the calculation formula is , thereby linearly reducing the weight to reduce the loss; when T > 100 °C, k = 0. At this time, the reactive power output of this arm is prohibited to protect the device.

[0039] In this embodiment: T = 85 °C, then k = 1 - 0.01 × 5 = 0.95; Step 3: Calculate the comprehensive weight of each arm. The formula is ; Step 4: Allocate the reference reactive power of each arm. The formula is .

[0040] In this embodiment: , , , then .

[0041] (2) Lower layer: Arm-level low-complexity MPC optimization module Input: The reference reactive power of this arm output by the upper layer ; The equivalent inductance of the SVG output by the online parameter identification module, capacitor ; The weights to output by the multi-objective weight adaptive module; Output: The IGBT switch control signal of this arm; Among them, the core algorithm: Adopt a three-step method of switch state grouping, simplified model, and fast solution. The specific implementation is as follows: Step 1: Switch state grouping.

[0042] Instead of the traditional MPC's traversal of all switching states, the switching states of the 20 sub-modules in a single MMC bridge arm are divided into three categories. The first category is the magnetization group, which increases the bridge arm output voltage, thereby increasing reactive power output. This includes two states: 19 sub-modules on combined with 1 off, and all 20 sub-modules on. The second category is the demagnetization group, which decreases the bridge arm output voltage, thereby reducing reactive power output. This includes two states: 1 sub-module on combined with 19 off, and all 20 sub-modules off. The third category is the holding group, which stabilizes the bridge arm output voltage, thereby maintaining constant reactive power output. This includes one state: 10 sub-modules on combined with 10 off. Predicting only these five key states significantly reduces the optimization dimensionality, unlike the traditional approach. a state; Step 2: Simplify the construction of the prediction model.

[0043] Based on Kirchhoff's voltage law, and ignoring minor parameters such as parasitic resistance and stray capacitance of submodules, a simplified second-order model of the bridge arm level is established.

[0044] The model expression is Where i is the bridge arm current, u is the bridge arm output voltage, and u is the grid voltage. This model reduces its dimensionality from the traditional 6th order to 2nd order, decreasing matrix computation by over 70%. Step 3: Fast solution algorithm. A sparse linear programming algorithm is used to solve for the optimal switching state. This algorithm can shorten the solution time by removing redundant variables.

[0045] The objective function is It includes core and secondary objectives. The optimal switching state from the previous control cycle is used as the initial value to narrow the search range. Traditional global traversal requires searching all states, while this scheme searches only 5 grouped states, and the initial value is close to the optimal solution, reducing the optimization time for a single bridge arm from 8ms to less than 1.5ms. 2. Multi-objective weight adaptive module This module employs a strategy that combines the Analytic Hierarchy Process (AHP) with the entropy weight method to integrate subjective needs with objective working conditions. It dynamically adjusts the weights of multiple objective functions, addressing the poor adaptability of traditional fixed-weight methods. AHP is a multi-criteria decision-making method that combines qualitative and quantitative analysis, while the entropy weight method is an objective weighting method based on data information entropy to determine weights. Input: Operating parameters collected by the sensing layer, including load type, grid voltage fluctuation amplitude, and IGBT temperature; real-time values ​​of each sub-target, including... to ; Output: Dynamic weights to ,satisfy , , , The sum equals 1; The core algorithm includes the following steps: Step 1: Determine the subjective weights of AHP.

[0046] A hierarchical structure is established, divided into an objective layer, a criterion layer, and a solution layer. The objective layer involves multi-objective optimization, and the criterion layer... to The scheme layer represents the states of each switch. A 1-9 scaling method is used to construct the judgment matrix. The basic weights are obtained by calculating the eigenvectors and performing a consistency check (CR < 0.1). =0.5 corresponds , =0.25 corresponds to , =0.15 corresponds to , =0.1 corresponds ; Step 2: Objective weight correction using the entropy weight method.

[0047] First, data standardization is performed on the data collected under different operating conditions. to The data is normalized to eliminate the influence of dimensions; secondly, the information entropy is calculated using the formula: ,in , For the sample size, For the first The first sample The proportion of each objective; finally, calculate the entropy weight, using the formula: In this embodiment: under the condition of sudden load change, Information entropy This indicates that it fluctuates greatly and contains a large amount of information; at this time, the entropy weight... ; Step 3: Calculate the fusion weights.

[0048] A linear weighted fusion method is used, and the formula is as follows: Where α = 0.6, representing a subjective weighting of 60% and an objective weighting of 40%. In this embodiment: under the condition of sudden load change, In practical applications, it is necessary to combine sample data to increase the weight of J to 0.6~0.7; Step 4: Dynamic weight update.

[0049] Repeat steps 1-3 every 10 control cycles (10 × 5ms = 50ms), updating according to the current operating conditions. .

[0050] 3. Online parameter identification module This module employs an improved Extended Kalman Filter (EKF) algorithm to identify key parameters of the power grid and SVG in real time. EKF is a parameter estimation method suitable for nonlinear systems. However, the traditional EKF suffers from slow convergence speed in nonlinear systems; this solution addresses this issue. Input: Grid voltage collected by the sensing layer SVG output current Bridge arm capacitor voltage ; Output: Mains impedance SVG equivalent inductance Equivalent capacitance Among them, the power grid impedance Includes resistors With inductance Two parts; Among them, the core improvement is that the traditional EKF uses a fixed Kalman gain K, which has a slow convergence speed of about 50ms.

[0051] This scheme adopts an adaptive Kalman gain adjustment strategy: Define the convergence criterion, i.e., the residual. .in For the observed values, For the observation matrix, This is a predicted value; Set threshold ,when When the Kalman gain K is increased by a factor of 1.2, the convergence speed is accelerated; when At this time, the Kalman gain K is reduced by a factor of 0.8 to ensure the estimation accuracy; The identification period was set to 10ms, with parameters updated once per period to ensure that the model parameters matched the actual working conditions. Testing showed that the improved EKF convergence time was reduced from 50ms to 20ms, and the identification error was ≤2%.

[0052] 4. AI Error Compensation Module This module uses a backpropagation (BP) neural network to build an offline error compensation library. During online operation, it corrects the control signal by interpolation and model output to solve the MPC prediction error problem. The BP neural network is a multi-layer feedforward neural network trained using the backpropagation algorithm, possessing strong function approximation capabilities. This module belongs to the data-driven AI model category, and its training process is as follows: Step 1: Training data collection.

[0053] A 110kV MMC-SVG simulation platform was built using MATLAB / Simulink tools; three typical operating conditions were simulated: steady-state, transient, and complex.

[0054] The steady-state operating condition requires load power fluctuation ≤5%; the transient operating condition requires load change amplitude between 20% and 50%; the complex operating condition includes two situations: voltage sag of 10%-20% and grid impedance fluctuation of 10%-30%.

[0055] Under each type of operating condition, input and output parameters are collected. Input parameters include... , , , The output parameters include MPC predicted reactive power. With actual no power The error was obtained through calculation. A total of 10,000 samples were collected; they were divided into training and testing sets according to the ratio, with the training set containing 8,000 samples and the testing set containing 2,000 samples. Step 2: Data preprocessing.

[0056] First, noise reduction is performed using wavelet transform, specifically the db4 wavelet, decomposed into three layers to remove high-frequency noise from the collected data. Second, normalization is performed, normalizing the input parameters and error e to the [0,1] interval. The normalization formula is as follows: , where x and x are sample extreme values; finally, the data is split, and the training set is divided into a training subset and a validation subset in a 7:3 ratio. The training subset is used for model training, and the validation subset is used for overfitting suppression. Step 3: BP neural network construction.

[0057] The number of input layer nodes is set to 4, corresponding to respectively , , , Four input parameters; the number of hidden layer nodes was determined to be 12 through trial and error, which minimizes the validation set error; the number of output layer nodes was set to 1, corresponding to the error compensation amount. Regarding the choice of activation function, the hidden layer uses the Sigmoid function, the function expression of which is: The output layer uses a linear function, and the function expression is as follows: ; Step 4: Model training.

[0058] The loss function used is the mean squared error (MSE), and the calculation formula is as follows: , where e is the model prediction error, e is the actual error, and n is the number of samples; the optimizer is the adaptive momentum estimation Adam, with a learning rate of 0.001, momentum coefficient β=0.9, and momentum coefficient β=0.999; training termination conditions are set, and the training will terminate when any condition is met, namely, the number of training rounds reaches 1000, and the validation set MSE≤0.001; Step 5: Model validation and optimization.

[0059] The model performance was validated using a test set. If the MSE on the test set was greater than 0.001, dropout was used to suppress overfitting with a dropout rate of 0.2, and then the model was retrained. The final trained model had an MSE of 0.0008 on the test set and an error compensation accuracy of ≥98%. Step 6: Model Deployment and Updates.

[0060] The trained BP neural network model is embedded into the DSP control chip to form an offline error compensation library. Every 3 months, new operating condition data, about 2000 sets, are collected to fine-tune the model. The learning rate is reduced to 0.0001, and the model is trained for 50 rounds to ensure compensation accuracy.

[0061] The online operation process in this embodiment includes: Input: Output of the online parameter identification module , , The upper level allocation ; Output: Error compensation amount ; Core logic: Normalize the input parameters and then input them into the BP neural network model to obtain the compensation amount. ;use The output of the lower-level MPC optimization module is corrected; the corrected formula is as follows: Where Q is the MPC predicted reactive power and Q is the final control reactive power. Testing showed that this module can further reduce the reactive power compensation error from ±2% to within ±1%. The specific flow of the input / output coherence logic between modules is as follows: Sensing layer: Voltage and current sensors acquire three-phase voltages on the grid side. Current The power sensor collects the SVG output power P and Q; the temperature sensor collects the IGBT temperature. to Capacitive voltage sensor data acquisition to All data is converted into digital signals by a high-speed AD conversion module. The sampling frequency of the AD conversion module is set to 20kHz, and the converted data is input to the control layer. The data preprocessing module performs three processes on the input digital signal: denoising, filtering, and synchronization alignment. Denoising uses wavelet transform, filtering uses mean filtering, and the processed output is clean, standardized data, including... , Q to , to ; Online parameter identification module: receives preprocessed parameters. , , Output the identified , , These identification parameters are respectively input into the upper-level global reactive power allocation module, the lower-level bridge arm-level MPC optimization module, and the AI ​​error compensation module; Upper-layer global reactive power allocation module: Receives the pre-processed total reactive power demand Q. , Output Input the lower-level bridge arm-level MPC optimization module and AI error compensation module; Multi-objective weight adaptive module: Receives preprocessed load data and real-time values ​​of each sub-objective. Load data includes load type, Fluctuation amplitude, T; real-time values ​​of each sub-target including to The module outputs dynamic weights. to Input to the lower-level bridge arm-level MPC optimization module; Lower-level bridge arm MPC optimization module: receiving , , , to Output the predicted reactive power Qpred and input it into the AI ​​error compensation module; AI error compensation module: receiving , , , , Output compensation amount .pass Correction get Then It is converted into an IGBT switching control signal, i.e., a PWM pulse, and input to the execution layer; Execution layer: The IGBT driver circuit receives the PWM signal and drives the MMC main circuit to output. The protection circuit monitors the equipment status in real time, including overcurrent, overvoltage, and overtemperature. If any abnormality occurs, a protection signal is output to cut off the drive circuit. Simultaneously, the execution layer will output the actual data. The device status is fed back to the sensing layer to form a closed-loop control; The construction of the experimental platform includes: This invention is verified using a 110kV MMC-SVG experimental prototype, with the following specific parameters: MMC topology: 6-arm structure, 20 sub-modules per arm, rated capacity 100Mvar, rated voltage 110kV; Control core: It adopts a DSP combined with FPGA architecture. The DSP is responsible for algorithm solving, including hierarchical MPC, weight adaptation, parameter identification, and error compensation; the FPGA is responsible for high-speed switching drive, specifically PWM signal generation, with the generation frequency set to 2kHz. Sensors: Voltage and current sensors with an accuracy of 0.2 class, model ABB PQ3000; power sensors with an accuracy of 0.1 class, model YOKOGAWA WT3000; temperature sensors with an accuracy of 0.5 class, model PT100. Software platform: MATLAB / Simulink was used for simulation modeling; CCS9.0 was used for DSP algorithm programming; LabVIEW was used for data acquisition and analysis. The implementation steps include the following: Step 1: Algorithm modeling and simulation.

[0062] Build an MMC-SVG simulation model in MATLAB / Simulink and write algorithm code for hierarchical MPC, improved EKF, and BP neural network error compensation.

[0063] Three types of operating conditions were set up: steady state, transient state, and complex state. Simulation tests were completed, and algorithm parameters were optimized, including the AHP judgment matrix and the number of hidden layer nodes in the BP neural network.

[0064] Step 2: Hardware platform development.

[0065] Assemble a 110kV MMC-SVG prototype and install sensors and AD conversion modules; develop a DSP combined with FPGA control board, port the simulation-verified algorithm code to the DSP, and write it in C language; the FPGA driver code is written in Verilog language.

[0066] Step 3: Train the BP neural network offline.

[0067] 10,000 sets of sample data were collected through a simulation platform to complete the training and validation of the BP neural network, and the model was then embedded into the DSP.

[0068] Step 4: Joint debugging and testing.

[0069] Connect the perception layer, control layer, and execution layer, and conduct three sets of comparative experiments.

[0070] The three experimental groups were: Control Group 1: Traditional single-target MPC; Control Group 2: Traditional multi-target MPC; Experimental Group: Scheme of this invention.

[0071] Collect key data under various operating conditions, including reactive power compensation error, control delay, switching loss, THD, etc.

[0072] Step 5: Optimization and finalization.

[0073] By analyzing test data and addressing error fluctuations under complex operating conditions, the BP neural network compensation parameters and EKF convergence threshold were fine-tuned; the finalized solution meets all core objective requirements.

[0074] The test results are shown in the table below: This invention provides a method for regulating the reactive power output of a high-voltage static synchronous compensator (SVG). It can reduce the computational complexity of the MPC algorithm while ensuring the accuracy of reactive power regulation, reduce the single-cycle control delay to less than 5ms, meet the real-time control requirements of high-voltage SVG, and has high reliability.

[0075] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0076] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for regulating the reactive power output of a high-voltage static synchronous compensator, characterized in that, The architecture includes a three-layer closed-loop structure comprising a perception layer, a control layer, and an execution layer, and also includes a hierarchical multi-objective model predictive control (MPC) control unit. The control unit includes a data preprocessing module, an online parameter identification module, an upper-layer global reactive power allocation module, a lower-layer bridge arm-level MPC optimization module, a multi-objective weight adaptive module, and an AI error compensation module. The specific implementation includes the following steps: S1: The sensing layer collects power grid-side operation data and SVG device status data, transmits the raw data to the data preprocessing module for processing, and obtains standardized and effective data; S2: The online parameter identification module receives standardized and valid data, identifies key parameters of the power grid and SVG model in real time, and transmits the identified parameters to the corresponding downstream modules. S3: The upper-level global reactive power allocation module synchronously receives the total reactive power demand of the power grid, the status data of each bridge arm of the SVG and the above-mentioned identification parameters from the standardized valid data, allocates the total reactive power demand into reference reactive power values ​​for each bridge arm, and transmits the reference reactive power values ​​to the corresponding downstream modules. S4: The multi-objective weight adaptive module receives the operating parameters and real-time values ​​of each control sub-objective from the standardized valid data, calculates and determines the dynamic weights, and transmits them to the downstream optimization module. S5: The lower-level bridge arm MPC optimization module receives the bridge arm reference reactive power value, key model parameters and dynamic weights, and calculates and obtains the predicted reactive power value. S6: The AI ​​error compensation module synchronously receives the bridge arm reference reactive power value, key model parameters and the predicted reactive power value, outputs the error compensation amount based on the BP neural network model completed offline training, corrects the predicted reactive power value through the compensation amount to obtain the final control signal and transmits it to the execution layer. S7: The execution layer receives the final control signal, drives the SVG main circuit to output the target reactive power, and feeds back the actual reactive power output of the SVG and the equipment operating status to the perception layer to form a complete closed-loop control.

2. The method for regulating the reactive power output of a high-voltage static synchronous compensator according to claim 1, characterized in that, The specific implementation process of parameter identification in S2 includes the following: The online parameter identification module receives standardized and valid data output by the data preprocessing module, including grid voltage, SVG output current, and bridge arm capacitor voltage. It uses the improved extended Kalman filter (EKF) algorithm to identify key model parameters in real time, including grid impedance, SVG equivalent inductance and equivalent capacitance. The identified parameters are then transmitted to the upper-level global reactive power allocation module, the lower-level bridge arm-level MPC optimization module, and the AI ​​error compensation module, respectively.

3. The method for adjusting the reactive power output of a high-voltage static synchronous compensator according to claim 2, characterized in that, The improved Extended Kalman Filter (EKF) algorithm in the online parameter identification module of S2 has a core improvement in its adaptive Kalman gain adjustment strategy. The specific implementation process includes: S21: Define a convergence criterion, which includes the residual between the observed value and the predicted value obtained based on the prediction model; S22: Set the residual threshold, which is the product of the observed value and the preset scaling factor; S23: Adjust the Kalman gain according to the relationship between the residual and the threshold. When the absolute value of the residual is greater than the threshold, increase the Kalman gain to speed up the convergence of the algorithm; when the absolute value of the residual is less than or equal to the threshold, decrease the Kalman gain to ensure the accuracy of parameter identification. S24: Update the key model parameters of the power grid and SVG according to the preset identification cycle. The key model parameters include the power grid impedance, SVG equivalent inductance and equivalent capacitance.

4. The method for regulating the reactive power output of a high-voltage static synchronizing compensator according to claim 1, characterized in that, The specific implementation process of reactive power allocation in S3 includes the following: The upper-level global reactive power allocation module receives the status data output by the data preprocessing module, including the total reactive power demand of the power grid, the capacitor voltage of each bridge arm of the SVG, and the temperature of the switching devices. Combining the key model parameters output by S2, based on the dynamic allocation algorithm of capacity weighting combined with state constraints, the capacitor voltage balance weight and temperature weight of each bridge arm are first calculated to obtain the comprehensive weight. Then, the total reactive power demand is allocated to the reference reactive power value of each bridge arm according to the comprehensive weight ratio. The reference reactive power value is transmitted to the lower-level bridge arm-level MPC optimization module and AI error compensation module to realize the reasonable allocation of reactive power demand and equipment status protection.

5. The method for adjusting the reactive power output of a high-voltage static synchronous compensator according to claim 4, characterized in that, The capacity-weighted dynamic allocation algorithm combined with state constraints of the upper-layer global reactive power allocation module in S3 is specifically implemented as follows: S31: Calculate the capacitor voltage balance weight of each bridge arm. This weight is inversely proportional to the capacitor voltage balance of each bridge arm and is calculated by the degree of deviation between the capacitor voltage of a single bridge arm and the average value of the capacitor voltages of all bridge arms. S32: Calculate the temperature weight of each bridge arm. Based on the difference between the real-time temperature of the bridge arm switching device and the safe operating threshold, the weight is determined using a piecewise linear function. The closer the temperature is to the safe threshold, the smaller the weight. When the temperature exceeds the safe threshold, the weight is set to zero. S33: Multiply the capacitor voltage balance weight and temperature weight of the same bridge arm to obtain the comprehensive weight of each bridge arm. S34: Based on the proportion of the comprehensive weight of each bridge arm to the sum of the comprehensive weights of all bridge arms, the total reactive power demand of the power grid is allocated to the corresponding bridge arm to obtain the reference reactive power value of each bridge arm, and the sum of the reference reactive power values ​​of all bridge arms is equal to the total reactive power demand of the power grid.

6. The method for regulating the reactive power output of a high-voltage static synchronous compensator according to claim 1, characterized in that, The multi-objective weight adaptive module in S4 includes an AHP fusion entropy weight method strategy, and its specific implementation process includes: S41: Establish a hierarchical structure including a target layer, a criterion layer, and a scheme layer. The target layer is for multi-objective optimization, the criterion layer is for each control sub-objective, and the scheme layer is for each switching state. A judgment matrix is ​​constructed using a preset scaling method. By calculating the eigenvectors of the judgment matrix and performing a consistency check, the subjective basic weights of each control sub-objective are obtained. S42: Standardize the real-time values ​​of each control sub-target collected under different operating conditions to eliminate the influence of dimensions. Calculate the information entropy of each control sub-target based on the principle of information entropy. The smaller the information entropy, the greater the fluctuation of the target data, the more information it contains, and the greater the corresponding objective weight. S43: A linear weighted fusion method is adopted to combine subjective basic weights and objective weights. The proportion of subjective basic weights and objective weights is allocated by a preset ratio coefficient to obtain the dynamic weights of each control sub-target. S44: Repeat the above steps to update the dynamic weights according to the preset cycle to ensure that the weights can adapt to changes in operating conditions.

7. The method for adjusting the reactive power output of a high-voltage static synchronous compensator according to claim 1, characterized in that, The lower-level bridge arm MPC optimization module in S5 includes low-complexity strategies such as switch state grouping, simplified model, and fast solution. The specific implementation process includes: S51: The switching states of multiple sub-modules of a single bridge arm in the SVG multilevel topology are divided into three categories: magnetization group, demagnetization group, and holding group. The magnetization group includes key switching states that can increase the output voltage of the bridge arm to increase reactive power output. The demagnetization group includes key switching states that can decrease the output voltage of the bridge arm to reduce reactive power output. The holding group includes key switching states that can stabilize the output voltage of the bridge arm to maintain the reactive power output. Only the key switching states in the above three categories are selected for subsequent prediction. S52: Based on Kirchhoff's voltage law, minor parameters, including submodule parasitic resistance and stray capacitance, are ignored to construct a bridge arm-level low-order simplified model, reducing the computational complexity of the model; S53: The sparse linear programming algorithm is used to solve for the optimal switching state. Based on the dynamic weights output by the multi-objective weight adaptive module, a comprehensive objective function is constructed, including reactive power compensation accuracy and switching loss targets. The optimal switching state of the previous control cycle is used as the initial value to narrow the search range and quickly solve for the predicted reactive power value.

8. The method for regulating the reactive power output of a high-voltage static synchronous compensator according to claim 1, characterized in that, The training process of the BP neural network model of the AI ​​error compensation module in S6 includes: S601: Build an SVG simulation platform to simulate three typical operating conditions: steady state, transient state, and complex state. Collect input and output parameters under each operating condition. The input parameters include the bridge arm reference reactive power value, SVG equivalent inductance, equivalent capacitance, and grid impedance. The output parameters include the MPC predicted reactive power value and the SVG actual output reactive power value. The error data is obtained by the difference between the two. S602: After denoising and normalizing the collected input parameters and error data, the training set and test set are divided. S603: Construct a BP neural network, including an input layer, hidden layers, and an output layer. Determine the number of nodes and activation functions for each layer. Train the neural network using a preset loss function and optimizer. Verify the model performance using a test set. If the performance does not meet the preset requirements, use preset techniques to suppress overfitting and retrain. After training, solidify the model into the control chip.

9. The method for regulating the reactive power output of a high-voltage static synchronizing compensator according to claim 1, characterized in that, The online operation process of the BP neural network model of the AI ​​error compensation module in S6 includes: S611: Normalize the key model parameters output by the online parameter identification module and the bridge arm reference reactive power value output by the upper-level global reactive power allocation module, and then input them into the trained BP neural network model to obtain the error compensation amount. S612: Use this error compensation amount to correct the predicted reactive power value output by the lower-level bridge arm MPC optimization module to obtain the final control signal, so that the reactive power compensation error of the SVG meets the preset accuracy requirements.

10. A method for regulating the reactive power output of a high-voltage static synchronizing compensator according to claim 1, characterized in that, The execution layer in S7 includes a switching device drive circuit, an SVG main circuit, and a protection circuit. The drive circuit receives the final control signal and generates a drive pulse to drive the SVG main circuit to output the target reactive power. The protection circuit monitors the overcurrent, overvoltage, and overtemperature status of the SVG main circuit in real time. When an abnormal status is detected, it outputs a protection signal to cut off the drive circuit to protect the equipment.

Images