Reactive voltage intelligent evaluation and correction method and system based on voltage prediction model

Abstract

The present application relates to a kind of reactive voltage intelligent evaluation and correction method and system based on voltage prediction model, method includes: acquisition grid data and input voltage prediction model, obtain the voltage operation upper and lower limit of each node in future period in whole network;The prediction result of voltage prediction model is input into reactive voltage evaluation optimization model, generates reactive voltage optimization control strategy, and is issued to grid reactive power compensation equipment execution;Wherein, voltage prediction model is based on deep learning network construction, voltage prediction model and reactive voltage evaluation optimization model are based on model performance evaluation index system and carry out self-adapting correction, compare the output result of reactive voltage evaluation optimization model with grid actual measurement data, calculate evaluation index, quantitative evaluation is carried out to voltage prediction model, and based on evaluation result, two models are retrained using online learning algorithm and parameter update.Compared with prior art, the present application realizes whole network voltage accurate regulation, and guarantees the safe and stable operation of power grid.

Description

Technical Field

[0001] This invention relates to the field of power system operation and control technology, and in particular to a method and system for intelligent assessment and correction of reactive power and voltage based on a voltage prediction model. Background Technology

[0002] In recent years, with the continuous expansion of power system scale and the increasing complexity of operating environment, voltage regulation and reactive power optimization of AC / DC power grids have become key technologies for ensuring stable grid operation. Especially in regions like East China, the large-scale integration of distributed energy sources and the complex and variable load characteristics have transformed the power grid into a complex AC / DC hybrid system. Under the influence of complex operating conditions such as multiple energy sources, load fluctuations, and extreme weather, how to accurately regulate reactive power and voltage to ensure the coordination and stability of all aspects of the power grid has become a critical issue that urgently needs to be addressed in power system operation and dispatch. The output of new energy sources such as wind power and photovoltaic power generation exhibits significant randomness and intermittency. Demand from industrial and residential loads fluctuates greatly with production schedules and lifestyles. These drastic fluctuations on both the source and load sides cause the operating state of the AC / DC hybrid system to be constantly changing, with voltage fluctuations frequently exceeding reasonable ranges, seriously affecting the stability of power supply.

[0003] Existing reactive power control methods mainly include the following categories: (1) Traditional optimization methods: including linear programming, nonlinear programming, dynamic programming and mixed integer programming. These methods rely on accurate mathematical models and have good optimization effects under steady-state or small-range fluctuation conditions. However, when faced with the strong randomness and intermittency of new energy power generation, they are difficult to adapt to changes in operating conditions in real time. Moreover, the computational load of solving nonlinear models is large, which makes it difficult to meet the needs of online applications.

[0004] (2) Intelligent optimization algorithms: including heuristic methods such as genetic algorithms, particle swarm optimization algorithms, and simulated annealing algorithms, as well as data-driven methods based on deep reinforcement learning. These methods have advantages in dealing with complex nonlinear optimization problems, but they have problems such as large training data requirements, poor model interpretability, and insufficient robustness under extreme conditions.

[0005] (3) Automatic Voltage Control (AVC): Existing AVC systems mainly realize steady-state voltage control of the main grid, and make switching decisions for reactive power compensation equipment through preset control rules and sensitivity matrices. However, existing AVC systems lack the ability to predict source load fluctuations in advance, and are significantly insufficient in dealing with the strong randomness and intermittency of new energy power generation. They are prone to over-adjustment or under-adjustment, and lack a post-evaluation mechanism for their own control effect, making it difficult to achieve adaptive optimization.

[0006] In summary, the existing technology has the following shortcomings: The lack of voltage prediction models that consider the uncertainty of source load fluctuations makes it impossible to predict the impact of source load fluctuations on reactive power and voltage in advance; the lack of a systematic post-evaluation mechanism for reactive power and voltage assessment and optimization models makes it impossible to quantify the accuracy and reliability of the assessment models; and the lack of adaptive correction capabilities of the models based on the assessment results means that the models cannot be continuously optimized as the grid operating status changes.

[0007] Therefore, there is an urgent need for a reactive voltage correction method that can solve the technical problem that existing reactive voltage assessment and optimization models lack post-evaluation mechanisms and adaptive correction capabilities. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a method and system for intelligent assessment and correction of reactive voltage based on a voltage prediction model, so as to achieve precise voltage regulation of the entire grid and ensure the safe and stable operation of the power grid.

[0009] The objective of this invention can be achieved through the following technical solutions: A method for intelligent assessment and correction of reactive power and voltage based on a voltage prediction model, the method comprising: Collect grid data and input it into a pre-trained voltage prediction model to obtain the upper and lower limits of voltage operation for each node in the entire grid in the future time period; input the prediction results of the voltage prediction model into a pre-constructed reactive voltage evaluation and optimization model to generate a reactive voltage optimization control strategy, and send the control strategy to the grid reactive power compensation equipment for execution. The voltage prediction model is constructed based on a deep learning network, and the power grid data includes historical voltage data, meteorological data, load forecast data, and new energy power generation forecast data. The voltage prediction model and the reactive power voltage assessment optimization model are adaptively corrected based on a model performance evaluation index system. The adaptive correction process includes: comparing the output results of the reactive power voltage assessment optimization model with the actual measurement data of the power grid, calculating evaluation indexes, quantitatively evaluating the accuracy, reliability, and stability of the voltage prediction model, and updating and retraining the parameters of the voltage prediction model and the reactive power voltage assessment optimization model using an online learning algorithm based on the evaluation results.

[0010] Furthermore, the deep learning network includes at least one of recurrent neural networks, long short-term memory networks, and gated recurrent units; the construction process of the voltage prediction model includes: Historical and real-time power grid operation data are collected. The historical operation data includes voltage data, reactive power data, load data, generator output data, real-time wind speed, light intensity, temperature and meteorological data, and new energy power generation data of each node. The collected data is processed by noise reduction, outlier removal and missing value filling to form a standardized dataset. The standardized dataset is divided into a training set and a validation set according to the time series. The voltage prediction model is trained by using multi-dimensional feature vectors within a historical time window as input and the voltage values ​​of all network nodes at the corresponding time as output labels. The hyperparameters are adjusted through cross-validation so that the prediction error of the voltage prediction model on the validation set meets the preset accuracy threshold; the hyperparameters include the number of network layers, the number of hidden layer nodes, and the learning rate. The trained voltage prediction model is deployed as an online inference service to predict the upper and lower limits of voltage operation for all nodes in the network in future periods.

[0011] Furthermore, the evaluation indicators of the model performance evaluation index system include voltage deviation index, reactive power loss index, model convergence speed index, and model prediction accuracy index. The voltage deviation index includes average voltage deviation and maximum voltage deviation; the reactive power loss index includes average reactive power loss of the entire network and average reactive power loss of each node; the model convergence speed index includes the number of optimization iterations and the single calculation time; the model prediction accuracy index includes: root mean square error of voltage prediction and mean absolute error of reactive power prediction.

[0012] Furthermore, the process of comparing the output results of the reactive power voltage assessment optimization model with the actual measurement data of the power grid includes: A fixed-duration sliding time window is constructed, and the actual measurement data of the power grid after the reactive power compensation equipment executes the control strategy is continuously collected within the sliding time window; The output results of the reactive voltage evaluation optimization model within the sliding time window are compared with the actual measurement data of the power grid point by point, and the window statistics of each evaluation index are calculated. The weight coefficients of each evaluation indicator are determined using the analytic hierarchy process (AHP), and the overall evaluation score of the model is calculated. When the window statistical value of any evaluation indicator exceeds the preset first-level warning threshold, a warning message is generated and the current operating condition characteristics are recorded. When the window statistical value of any evaluation indicator exceeds the preset second-level trigger correction threshold or the model's comprehensive evaluation score exceeds the preset comprehensive correction threshold, the model's parameters are updated and retrained.

[0013] Furthermore, the online learning algorithm includes stochastic gradient descent and adaptive moment estimation algorithms; the process of updating and retraining the parameters of the voltage prediction model and the reactive voltage assessment optimization model includes: When the window statistical value of any evaluation index exceeds the preset second-level trigger correction threshold or the model comprehensive evaluation score exceeds the preset comprehensive correction threshold, an online learning algorithm is used to incrementally update the model parameters using the actual power grid measurement data within the latest time window as training samples. Based on the Bayesian optimization method, the optimal hyperparameter combination for the evaluation index is found within the preset hyperparameter search space, and the model structure and training strategy are adjusted.

[0014] Furthermore, during the incremental update of the model parameters, the experience of power system experts is embedded into the model calibration process in the form of rule constraints. These rule constraints include physical action constraints of reactive power compensation equipment, discrete constraints of transformer tap changer adjustment, and voltage stability margin constraints.

[0015] Furthermore, the reactive power and voltage assessment optimization model is established using a quadratic constraint quadratic programming method; the control variables of the reactive power and voltage assessment optimization model include generator terminal voltage amplitude, reactive power compensation equipment switching quantity, and transformer tap position; the reactive power and voltage assessment optimization model aims to minimize grid reactive power loss and voltage deviation; the constraints of the reactive power and voltage assessment optimization model include power flow equation constraints, node voltage upper and lower limit constraints, reactive power compensation equipment capacity constraints, and transformer tap adjustment range constraints; the reactive power and voltage assessment optimization model is solved using the primal-dual interior point method combined with the branch and bound method.

[0016] Furthermore, the method also includes: Before the daily scheduling plan is formulated, the voltage prediction model is used to predict the upper and lower limits of voltage operation of each node in the entire network during a preset period of the next day, so as to form a voltage operation interval plan for the next day. During daily operation, the latest power grid data is collected at preset time intervals. Based on the model performance evaluation index system, the voltage prediction model and reactive voltage evaluation optimization model are adaptively corrected to adjust the voltage prediction results for the remaining time periods.

[0017] Furthermore, the method also includes: after completing the parameter update and retraining of the voltage prediction model and the reactive voltage assessment optimization model, storing the cause of the correction, the assessment index values ​​before and after correction, the characteristics of the deviation samples, and the updated model parameters as correction records in the model knowledge base, wherein the initial parameters or hyperparameter prior knowledge of the online learning algorithm are taken from the model knowledge base.

[0018] A reactive power voltage intelligent assessment and correction system based on a voltage prediction model, the system comprising: The data acquisition module is used to collect historical and real-time operating data of the power grid and perform preprocessing. The voltage prediction module is used to build and run voltage prediction models based on deep learning networks. The reactive power optimization module is used to build and solve the reactive power voltage evaluation optimization model and generate reactive power compensation control strategies. The evaluation module is used to establish a model performance evaluation index system to quantitatively evaluate the accuracy, reliability and stability of the reactive voltage evaluation optimization model. The correction module is used to adaptively correct the voltage prediction model and the reactive voltage assessment optimization model based on the evaluation results of the evaluation module using an online learning algorithm. The scheduling interface module is used to send the reactive voltage optimization control strategy generated by the reactive power optimization module to the power grid reactive power compensation equipment, and to receive the execution feedback from the power grid reactive power compensation equipment. The modules interact with each other through standardized data interfaces, and the system supports coordinated operation of three time scales: daily forecasting, intraday rolling forecasting, and real-time optimization.

[0019] Compared with the prior art, the beneficial effects of the present invention include: 1. This invention integrates the voltage prediction model and the reactive voltage evaluation optimization model as an organic whole, constructing a complete closed loop from prediction to decision-making, execution, evaluation, and correction. The evaluation index system in this invention not only focuses on prediction accuracy but also extends to the execution effect of the control strategy, which can accurately locate performance bottlenecks in the entire link, achieve synergistic correction of the two models, and improve the overall control effect.

[0020] 2. This invention, through real-time monitoring via a sliding time window and a multi-level threshold triggering mechanism, can automatically detect model performance degradation and immediately initiate the correction process. Employing an online learning algorithm, the model can quickly learn and iteratively optimize from real-time deviation samples without manual intervention, thereby improving the model's agility and robustness in responding to changes in power grid conditions and extreme operating conditions.

[0021] 3. This invention constructs a comprehensive model performance evaluation index system, including voltage deviation index, reactive power loss index, model convergence speed index, and model prediction accuracy index. These indexes enable quantitative evaluation of the model's accuracy, reliability, and stability.

[0022] 4. This invention introduces a model knowledge base to store historical correction records, associating the operating condition characteristics of each correction with the parameter update scheme. When similar operating conditions reappear, the system can quickly retrieve historical experience as prior knowledge for correction, achieving intelligent evolution and accelerating the correction convergence speed.

[0023] 5. This invention enhances the interpretability of the model and the credibility of decisions by embedding expert rules from the power system field into the model calibration process. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0025] 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, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0026] Example 1 This embodiment discloses a method for intelligent assessment and correction of reactive power and voltage based on a voltage prediction model. The method is as follows: Figure 1 As shown, steps S1-S6 are included, and each step is described in detail below: Step S1, Data Acquisition and Preprocessing: Collect historical and real-time power grid operation data and preprocess them to form a standardized dataset.

[0027] Historical operating data specifically includes voltage data, reactive power data, load data, generator output data, real-time wind speed, solar intensity, temperature and meteorological data, and new energy power generation data for each node.

[0028] The preprocessing of the collected data includes noise reduction, outlier removal, and missing value imputation to form a standardized dataset.

[0029] Step S2, Voltage Prediction Model Construction: Based on a standardized dataset, a voltage prediction model is constructed using a deep learning network. The voltage prediction model takes historical voltage data, meteorological data, load prediction data, and new energy power generation prediction data as inputs and outputs the upper and lower limits of voltage operation of each node in the entire network for the next day.

[0030] Deep learning networks include at least one of recurrent neural networks, long short-term memory networks, and gated recurrent units.

[0031] The process of constructing a voltage prediction model includes: The standardized dataset is divided into a training set and a validation set according to the time series. The voltage prediction model is trained by using multi-dimensional feature vectors within a historical time window as input and the voltage values ​​of all network nodes at the corresponding time as output labels. The hyperparameters are adjusted through cross-validation so that the prediction error of the voltage prediction model on the validation set meets the preset accuracy threshold; the hyperparameters include the number of network layers, the number of hidden layer nodes, and the learning rate. The trained voltage prediction model is deployed as an online inference service to predict the upper and lower limits of voltage operation for all nodes in the network in future periods.

[0032] Specifically, the output of the voltage prediction model is the upper and lower limits of voltage operation for all nodes in the entire network at a preset time resolution for the next day, such as 15 minutes / 1 hour. These limits are directly used as the node voltage constraint boundaries for the subsequent reactive power and voltage optimization model.

[0033] The labels of the standardized training dataset are based on historical measured voltage data and are quantile-calibrated: the upper limit corresponds to the 97.5% quantile value of the voltage under the same operating conditions during the same period in history, and the lower limit corresponds to the 2.5% quantile value, forming a voltage operating range with a 95% confidence level. This ensures coverage of the vast majority of operating scenarios while avoiding an excessively wide range that would lose its constraining significance.

[0034] Specifically, a sliding time window method is used to generate training sample pairs: the historical input window length is set, and the prediction window length is 24 hours of the next day. The standardized dataset is continuously truncated over time steps to generate sample pairs of historical multi-source feature sequences and future voltage range labels.

[0035] The dataset is divided as follows: Training set: accounting for 70%~80% of the total data, selecting historical data from earlier periods for iterative updates of model parameters; Validation set: accounting for 10%~15% of the total data, selected from continuous time periods, used for accuracy monitoring, overfitting judgment and initial selection of hyperparameters during the training process; Test set: accounting for 10%~15% of the total data, selecting the latest independent time period data (not used for training and validation), used for the final generalization performance verification of the model.

[0036] The Z-score standardization method is used for all input features to map each dimension of features to a distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of different units on model training. The standardization parameters (mean and standard deviation) are obtained only from the training set statistics, and the validation set and test set reuse these parameters to ensure that the data processing logic is consistent with the actual online inference.

[0037] In this embodiment, the model training loss function is tailored to the task characteristics of voltage range prediction. The core loss function is quantile loss, which weights the quantile prediction errors of the upper and lower voltage limits to ensure range coverage and boundary accuracy. Simultaneously, the mean square error (MSE) loss of the voltage mean is added to constrain the prediction accuracy of the range center. The total loss function is the weighted sum of these two losses.

[0038] In this embodiment, the model training optimizer uses the Adaptive Moment Estimation (Adam) optimizer to iteratively update parameters, ensuring that the optimization logic is consistent between the training phase and the subsequent online calibration phase; the initial learning rate is a configurable hyperparameter, and a learning rate decay strategy is added in the later stage of training to improve convergence stability.

[0039] After training is completed, the core accuracy metrics are calculated on the validation set: root mean square error of voltage prediction (RMSE), mean absolute error (MAE), and voltage range coverage. The metrics are confirmed to meet the preset accuracy threshold requirements. If they do not meet the requirements, the hyperparameters are adjusted and the training is repeated.

[0040] Once the accuracy target is met, the model is solidified and packaged into a standardized online service, which is then connected to the power grid dispatch system to form a closed loop with subsequent optimization and correction processes. Before the daily dispatch plan is formulated, the latest historical operating data, the next day's load forecast, the new energy power generation forecast, and the weather forecast data are automatically loaded. The upper and lower limits of voltage operation for each node during the entire next day are output to form the voltage operation range plan for the next day, which is then sent to the reactive power and voltage assessment and optimization model.

[0041] Step S3, Reactive Voltage Assessment and Optimization Model Operation: Collect grid data and input it into the voltage prediction model to obtain the upper and lower limits of voltage operation for each node in the entire grid in the future period; input the prediction results of the voltage prediction model into the pre-built reactive voltage assessment and optimization model to generate a reactive voltage optimization control strategy, and send the control strategy to the grid reactive power compensation equipment for execution.

[0042] The reactive voltage assessment optimization model is established using the quadratic constrained quadratic programming (QCQP) method.

[0043] The control variables of the reactive power voltage assessment and optimization model include generator terminal voltage amplitude, reactive power compensation equipment switching quantity, and transformer tap position. Its objective is to simultaneously minimize grid reactive power losses and voltage deviation.

[0044] The constraints of the reactive power voltage assessment optimization model include power flow equation constraints, node voltage upper and lower limit constraints, reactive power compensation equipment capacity constraints, and transformer tap adjustment range constraints.

[0045] The reactive voltage assessment optimization model is solved by combining the original dual interior point method with the branch and bound method. The original dual interior point method is used to solve continuous variable optimization problems, while the branch and bound method is used to handle discrete integer variables such as transformer tap position and the number of capacitor switching groups.

[0046] Specifically, in this embodiment, the mathematical expression of the reactive power voltage assessment optimization model is: in, The control variable vector includes generator terminal voltage amplitude, reactive power compensation equipment switching quantity, and transformer tap position. The coefficient matrix of the quadratic term of the objective function represents the quadratic coupling relationship between the control variables, corresponding to the quadratic terms of power grid reactive power loss and voltage deviation. This is the coefficient vector of the first-order term of the objective function, corresponding to the first-order components of the power grid reactive power loss and voltage deviation. Let be the coefficient matrix of the quadratic term of the k-th inequality constraint. Let be the coefficient vector of the linear term of the k-th inequality constraint. Let k be the constant term of the k-th inequality constraint. Let be the coefficient matrix of the quadratic term of the j-th equality constraint. Let be the coefficient vector of the linear term of the j-th equality constraint. Let be the constant term of the j-th equality constraint.

[0047] In another embodiment, the solution of the reactive power voltage assessment optimization model is accelerated using sparse matrix techniques, specifically: For the coefficient matrix , , Using a compressed storage format, in the form of triple arrays Row number, column number, value Describing sparse matrices as arrays of binary tuples Line number, value Describes sparse vectors; In the iterative solution process of the original dual interior point method, the structural characteristics of sparse matrices are used to accelerate the solution of linear equations and reduce memory usage and computational load. For mixed integer optimization problems containing discrete variables (transformer tap position, number of capacitor banks switched on and off), the branch and bound method is used to perform sensitivity analysis at the optimal solution of the relaxation problem. The discrete variables with the greatest impact on the objective function are selected for branching to improve the solution efficiency.

[0048] Step S4, Model Performance Evaluation: Establish a model performance evaluation index system. By comparing the output results of the reactive voltage evaluation optimization model with the actual measurement data of the power grid, calculate the evaluation index, and quantitatively evaluate the accuracy, reliability, and stability of the reactive voltage evaluation optimization model.

[0049] The evaluation indicators of the model performance evaluation index system include voltage deviation index, reactive power loss index, model convergence speed index, and model prediction accuracy index.

[0050] Voltage deviation indicators include average voltage deviation and maximum voltage deviation Reactive power loss indicators include the average reactive power loss of the entire network. and average reactive power loss at each node; the model convergence speed index includes the number of optimization iterations. and single calculation time The model prediction accuracy metrics include: root mean square error (RMSE) for voltage prediction and mean absolute error (MAE) for reactive power prediction.

[0051] Specifically, in this embodiment, the average voltage deviation The calculation expression is: In this embodiment, the maximum voltage deviation The calculation expression is: in, The total number of nodes. To assess the number of time periods, For nodes exist Predicted voltage at time [time] This corresponds to the measured voltage.

[0052] Specifically, in this embodiment, the calculation expression for the root mean square error (RMSE) of voltage prediction is as follows: .

[0053] The process of comparing the output results of the reactive power voltage assessment optimization model with the actual measurement data of the power grid includes: Construct a fixed-duration sliding time window and continuously collect actual power grid measurement data after the reactive power compensation equipment executes the control strategy within the sliding time window; The output results of the reactive voltage evaluation optimization model within the sliding time window are compared with the actual measurement data of the power grid point by point, and the window statistics of each evaluation index are calculated. The weight coefficients of each evaluation indicator are determined using the analytic hierarchy process (AHP), and the overall evaluation score of the model is calculated. When the window statistical value of any evaluation indicator exceeds the preset first-level warning threshold, a warning message is generated and the current operating condition characteristics are recorded. When the window statistical value of any evaluation indicator exceeds the preset second-level trigger correction threshold or the model's comprehensive evaluation score exceeds the preset comprehensive correction threshold, the model's parameters are updated and retrained.

[0054] Online learning algorithms include stochastic gradient descent and adaptive moment estimation.

[0055] The process of updating and retraining the parameters of the voltage prediction model and the reactive voltage assessment optimization model includes: When the window statistical value of any evaluation index exceeds the preset second-level trigger correction threshold or the model comprehensive evaluation score exceeds the preset comprehensive correction threshold, an online learning algorithm is used to incrementally update the model parameters using the actual power grid measurement data within the latest time window as training samples. Based on the Bayesian optimization method, the optimal hyperparameter combination for the evaluation index is found within the preset hyperparameter search space, and the model structure and training strategy are adjusted.

[0056] Specifically, in this embodiment, the update formula for incremental updates is: in, for Model parameters at time 10:00 For learning rate, For loss function, This is the latest time window dataset.

[0057] During the incremental update of model parameters, the experience of power system experts is embedded into the model calibration process in the form of rule constraints. These rule constraints include physical action constraints of reactive power compensation equipment, discrete constraints of transformer tap changer adjustment, and voltage stability margin constraints.

[0058] Step S5, Intelligent Model Correction: Based on the evaluation results of step S4, an online learning algorithm is used to update and retrain the parameters of the voltage prediction model and the reactive voltage evaluation optimization model to achieve adaptive correction of the model.

[0059] Step S6, Rolling Update: Steps S1 to S5 are executed cyclically according to the set time interval to achieve real-time tracking and dynamic adjustment of the reactive power and voltage operation status of the power grid.

[0060] Specifically, rolling updates include: Before the daily scheduling plan is formulated, the voltage prediction model is used to predict the upper and lower limits of voltage operation of each node in the entire network during the preset time period of the next day, so as to form the voltage operation interval plan for the next day. During the daytime operation, the latest power grid data is collected at preset time intervals. Based on the model performance evaluation index system, the voltage prediction model and reactive voltage evaluation optimization model are adaptively corrected to adjust the voltage prediction results for the remaining time periods.

[0061] After completing the parameter update and retraining of the voltage prediction model and the reactive voltage assessment optimization model, the cause of the correction, the assessment index values ​​before and after the correction, the characteristics of the deviation samples, and the updated model parameters are stored in the model knowledge base as correction records. The initial parameters or hyperparameter prior knowledge of the online learning algorithm are taken from the model knowledge base.

[0062] For the extreme scenario of a sharp drop in offshore wind power output under severe convective weather in summer, the following robustness verification was conducted: Scenario setting: On a summer day from 14:00 to 15:00, affected by severe convective weather, the output of an offshore wind farm suddenly dropped from 80% to 10% of its rated power within 30 minutes. The DC transmission power changed rapidly accordingly, and the receiving AC grid faced a serious reactive power deficit.

[0063] Verification steps: (1) Complete the basic parameter configuration. Retrieve historical operating data under extreme weather scenarios from the pre-built typical scenario library to construct a source-load fluctuation simulation scenario, or directly collect historical and real-time operating data of the power grid. After retrieval or collection, preprocess the data to construct a standardized dataset.

[0064] Collect electrical quantity data: voltage amplitude at each node, active power, reactive power, generator active / reactive output, transformer tap position, capacitor / reactor switching status, and reactive power output of SVC / STATCOM.

[0065] Collect meteorological data: real-time wind speed, light intensity, ambient temperature, and relative humidity.

[0066] Two levels of evaluation thresholds are preset: a first-level warning threshold, a second-level correction trigger threshold, and a model comprehensive evaluation score correction threshold.

[0067] Preset constraints include power flow equation constraints, node voltage upper and lower limits constraints, reactive power compensation equipment capacity constraints, and transformer tap adjustment range constraints.

[0068] Preprocess the collected data: Denoising: Wavelet transform is used to remove high-frequency noise from voltage and current signals.

[0069] Outlier removal: The 3σ criterion is used to identify and remove outlier data that deviates significantly from the normal range.

[0070] Missing value imputation: Linear interpolation is used to impute short-term missing data, and similar day matching is used to impute long-term missing data.

[0071] Standardization: The Z-score standardization method is used to transform data of different dimensions into standardized data with a mean of 0 and a variance of 1. The formula is as follows: in, Here are the standardized parameters, and x represents the original data. The mean of this feature. This represents the standard deviation of the feature.

[0072] (2) Using the Monte Carlo sampling method, ±20% random disturbance is superimposed on the basis of the sudden drop in wind power output to generate 100 sets of source-load fluctuation scenarios.

[0073] This example uses a hybrid deep learning voltage prediction model that combines LSTM and GRU: Input layer: Receives multi-dimensional feature vectors from the past 24 hours, including voltage, reactive power, load, renewable energy output, meteorological data, etc. of each node. The input dimension is (96, N), where 96 is the number of 24-hour data points with a 15-minute sampling interval, and N is the number of features. Hidden layers: The first layer is an LSTM layer, which extracts long-term dependency features of the time series; the second layer is a GRU layer, which further extracts short-term fluctuation features and reduces computational cost; the third layer is a fully connected layer, which performs feature mapping. Output layer: Outputs the upper and lower limits of voltage operation for all nodes in the entire network at 15-minute intervals for the next 24 hours. The output dimension is (96, 2M), where M is the number of nodes.

[0074] The model training process is as follows: The standardized dataset was divided into a training set and a validation set in an 8:2 ratio. The root mean square error (RMSE) is used as the loss function, as shown in the following formula: in, This is the actual voltage value. To predict the voltage value, n is the number of samples; The Adam optimizer was used for model training; hyperparameters were adjusted using 5-fold cross-validation, and the final network layer count was determined to be 3 layers, with 128 hidden nodes in the LSTM layer and 64 hidden nodes in the GRU layer. When the validation set RMSE is below 0.5kV, training is stopped and the model parameters are saved.

[0075] The trained model is deployed as an online inference service, receiving the latest power grid data in real time and outputting the upper and lower limits of voltage operation at each node for the next 24 hours.

[0076] (3) Input 100 scenarios into the whole network voltage prediction model, calculate the voltage prediction RMSE under each scenario, and evaluate the prediction accuracy of the model under extreme conditions.

[0077] Rolling Predictive Response Test: Throughout the disturbance process, the voltage prediction input is updated at preset rolling intervals, and the upper and lower limits of voltage operation at each node in the future time period are output. The predicted voltage curve is compared with the measured voltage curve, and the voltage prediction RMSE, mean absolute error (MAE), and maximum voltage deviation are calculated point by point. .

[0078] Verification of the effectiveness of over-limit prediction: Statistical model predicts the time lead of voltage over-limit at key nodes in advance, and verifies whether the predicted upper and lower voltage limits can cover the actual voltage fluctuation range in time; statistics are compiled on the number and proportion of missed and false predictions.

[0079] Prediction bias trend observation: Record the magnitude of change in prediction error during the development of disturbance to verify the prediction stability of the model under the scenario of rapid reactive power fluctuation; mark the moment when the error first exceeds the first-level warning threshold and the second-level correction threshold to provide time nodes for subsequent self-correction verification.

[0080] Control strategy execution tracking: Record the control commands issued by the optimization model one by one, including the generator terminal voltage adjustment, the number of reactive power compensation equipment switching groups, and the transformer tap change; verify that all control quantities meet the equipment capacity constraints, adjustment range constraints, and expert rule constraints.

[0081] Quantitative verification of voltage regulation effect: The number of voltage over-limits, maximum over-limit amplitude, and total duration of over-limits of all network nodes during the entire disturbance process are statistically analyzed and compared with the control group; The two core indicators, average voltage deviation and total network reactive power loss, were calculated to verify whether the optimization strategy could control the voltage within the acceptable range while minimizing reactive power loss under severe reactive power deficit.

[0082] After the correction is triggered, the verification system automatically uses the measured data within the latest sliding time window as samples to perform incremental parameter updates on the voltage prediction model through the stochastic gradient descent / adaptive moment estimation algorithm.

[0083] The Bayesian optimization module is launched simultaneously to search for the optimal number of network layers, learning rate, and other parameters in the preset hyperparameter space, and to fine-tune the model structure.

[0084] Record the total time taken from triggering the correction to completing the parameter update, and verify whether the correction response speed can meet the real-time operation requirements of the power grid.

[0085] Verification of the effect after correction: After correction, continue running for 15-30 minutes to recalculate indicators such as prediction error and voltage deviation to verify whether the model accuracy has recovered to the acceptable range; compare the difference in indicators before and after correction to quantify the performance recovery effect of the self-correction mechanism.

[0086] Rule constraint embedding verification: The optimized strategy after full verification and correction still meets the expert rules such as reactive equipment operation interval, tap changer adjustment times, and voltage stability margin, and there is no situation where the control strategy after correction violates physical constraints.

[0087] (4) For scenarios where the predicted RMSE exceeds the threshold, analyze the key features that lead to the increase in prediction error, and include these scenarios in the model retraining dataset to improve the robustness of the model under extreme conditions through incremental learning.

[0088] Verification results: After three rounds of iterative correction, the voltage prediction RMSE of the model under extreme weather scenarios decreased from 0.0089 pu to 0.0041 pu, a reduction of 54%, which meets the voltage control accuracy requirements under extreme conditions.

[0089] Example 2 This embodiment, based on Embodiment 1 above, discloses a reactive power voltage intelligent assessment and correction system based on a voltage prediction model. The system includes: The data acquisition module is used to collect historical and real-time operating data of the power grid and perform preprocessing.

[0090] The voltage prediction module is used to build and run voltage prediction models based on deep learning networks.

[0091] The reactive power optimization module is used to build and solve the reactive power voltage evaluation optimization model and generate reactive power compensation control strategies.

[0092] The evaluation module is used to establish a model performance evaluation index system to quantitatively evaluate the accuracy, reliability, and stability of the reactive voltage evaluation optimization model.

[0093] The correction module is used to adaptively correct the voltage prediction model and the reactive voltage assessment optimization model based on the evaluation results of the evaluation module using an online learning algorithm.

[0094] The scheduling interface module is used to send the reactive voltage optimization control strategy generated by the reactive power optimization module to the power grid reactive power compensation equipment, and to receive the execution feedback from the power grid reactive power compensation equipment.

[0095] The modules interact with each other through standardized data interfaces, and the system supports collaborative operation across three time scales: daily forecasting, intraday rolling forecasting, and real-time optimization.

[0096] For details regarding the above modules, please refer to the relevant descriptions and effects in Example 1 for further understanding.

[0097] Example 3 Based on Embodiment 2, this embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the aforementioned reactive voltage intelligent assessment and correction method based on a voltage prediction model.

[0098] At the hardware level, the electronic device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the aforementioned intelligent reactive voltage assessment and correction method based on a voltage prediction model. Of course, in addition to software implementation, this invention does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0099] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0100] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0101] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent assessment and correction of reactive power and voltage based on a voltage prediction model, characterized in that, The method includes: Collect grid data and input it into a pre-trained voltage prediction model to obtain the upper and lower limits of voltage operation for each node in the entire grid in the future time period; input the prediction results of the voltage prediction model into a pre-constructed reactive voltage evaluation and optimization model to generate a reactive voltage optimization control strategy, and send the control strategy to the grid reactive power compensation equipment for execution. The voltage prediction model is constructed based on a deep learning network, and the power grid data includes historical voltage data, meteorological data, load forecast data, and new energy power generation forecast data. The voltage prediction model and the reactive power voltage assessment optimization model are adaptively corrected based on a model performance evaluation index system. The adaptive correction process includes: comparing the output results of the reactive power voltage assessment optimization model with the actual measurement data of the power grid, calculating evaluation indexes, quantitatively evaluating the accuracy, reliability, and stability of the voltage prediction model, and updating and retraining the parameters of the voltage prediction model and the reactive power voltage assessment optimization model using an online learning algorithm based on the evaluation results.

2. The method for intelligent assessment and correction of reactive power and voltage based on a voltage prediction model according to claim 1, characterized in that, The deep learning network includes at least one of recurrent neural networks, long short-term memory networks, and gated recurrent units; the construction process of the voltage prediction model includes: Historical and real-time power grid operation data are collected. The historical operation data includes voltage data, reactive power data, load data, generator output data, real-time wind speed, light intensity, temperature and meteorological data, and new energy power generation data of each node. The collected data is processed by noise reduction, outlier removal and missing value filling to form a standardized dataset. The standardized dataset is divided into a training set and a validation set according to the time series. The voltage prediction model is trained by using multi-dimensional feature vectors within a historical time window as input and the voltage values ​​of all network nodes at the corresponding time as output labels. The hyperparameters are adjusted through cross-validation so that the prediction error of the voltage prediction model on the validation set meets the preset accuracy threshold; the hyperparameters include the number of network layers, the number of hidden layer nodes, and the learning rate. The trained voltage prediction model is deployed as an online inference service to predict the upper and lower limits of voltage operation for all nodes in the network in future periods.

3. The method for intelligent assessment and correction of reactive power and voltage based on a voltage prediction model according to claim 1, characterized in that, The evaluation indicators of the model performance evaluation index system include voltage deviation index, reactive power loss index, model convergence speed index, and model prediction accuracy index. The voltage deviation index includes the average voltage deviation and the maximum voltage deviation; the reactive power loss index includes the average reactive power loss of the entire network and the average reactive power loss of each node. The model convergence speed metrics include the number of optimization iterations and the time per computation; the model prediction accuracy metrics include the root mean square error of voltage prediction and the mean absolute error of reactive power prediction.

4. The method for intelligent assessment and correction of reactive power and voltage based on a voltage prediction model according to claim 1, characterized in that, The process of comparing the output results of the reactive power voltage assessment optimization model with the actual measurement data of the power grid includes: A fixed-duration sliding time window is constructed, and the actual measurement data of the power grid after the reactive power compensation equipment executes the control strategy is continuously collected within the sliding time window; The output results of the reactive voltage evaluation optimization model within the sliding time window are compared with the actual measurement data of the power grid point by point, and the window statistics of each evaluation index are calculated. The weight coefficients of each evaluation indicator are determined using the analytic hierarchy process (AHP), and the overall evaluation score of the model is calculated. When the window statistical value of any evaluation indicator exceeds the preset first-level warning threshold, a warning message is generated and the current operating condition characteristics are recorded. When the window statistical value of any evaluation indicator exceeds the preset second-level trigger correction threshold or the model's comprehensive evaluation score exceeds the preset comprehensive correction threshold, the model's parameters are updated and retrained.

5. The method for intelligent assessment and correction of reactive power and voltage based on a voltage prediction model according to claim 4, characterized in that, The online learning algorithms include stochastic gradient descent and adaptive moment estimation algorithms; The process of updating and retraining the parameters of the voltage prediction model and the reactive voltage assessment optimization model includes: When the window statistical value of any evaluation index exceeds the preset second-level trigger correction threshold or the model comprehensive evaluation score exceeds the preset comprehensive correction threshold, an online learning algorithm is used to incrementally update the model parameters using the actual power grid measurement data within the latest time window as training samples. Based on the Bayesian optimization method, the optimal combination of hyperparameters for the evaluation index is found within a pre-defined hyperparameter search space, and the model structure and training strategy are adjusted accordingly.

6. The method for intelligent assessment and correction of reactive power and voltage based on a voltage prediction model according to claim 5, characterized in that, During the incremental update of model parameters, the experience of power system experts is embedded into the model calibration process in the form of rule constraints. These rule constraints include physical action constraints of reactive power compensation equipment, discrete constraints of transformer tap changer adjustment, and voltage stability margin constraints.

7. The method for intelligent assessment and correction of reactive power and voltage based on a voltage prediction model according to claim 1, characterized in that, The reactive power and voltage assessment optimization model is established using a quadratic constraint quadratic programming method. The control variables of the reactive power and voltage assessment optimization model include generator terminal voltage amplitude, reactive power compensation equipment switching quantity, and transformer tap position. The reactive power and voltage assessment optimization model aims to minimize grid reactive power loss and voltage deviation. The constraints of the reactive power and voltage assessment optimization model include power flow equation constraints, node voltage upper and lower limit constraints, reactive power compensation equipment capacity constraints, and transformer tap adjustment range constraints. The reactive power and voltage assessment optimization model is solved using the primal-dual interior point method combined with the branch and bound method.

8. The method for intelligent assessment and correction of reactive power and voltage based on a voltage prediction model according to claim 1, characterized in that, The method further includes: Before the daily scheduling plan is formulated, the voltage prediction model is used to predict the upper and lower limits of voltage operation of each node in the entire network during a preset period of the next day, so as to form a voltage operation interval plan for the next day. During daily operation, the latest power grid data is collected at preset time intervals. Based on the model performance evaluation index system, the voltage prediction model and reactive voltage evaluation optimization model are adaptively corrected to adjust the voltage prediction results for the remaining time periods.

9. The method for intelligent assessment and correction of reactive power and voltage based on a voltage prediction model according to claim 1, characterized in that, The method further includes: after completing the parameter update and retraining of the voltage prediction model and the reactive voltage assessment optimization model, storing the cause of the correction, the assessment index values ​​before and after correction, the characteristics of the deviation samples, and the updated model parameters as correction records in the model knowledge base, and taking the initial parameters or hyperparameter prior knowledge of the online learning algorithm from the model knowledge base.

10. A reactive power and voltage intelligent assessment and correction system based on a voltage prediction model, characterized in that, The system includes: The data acquisition module is used to collect historical and real-time operating data of the power grid and perform preprocessing. The voltage prediction module is used to build and run voltage prediction models based on deep learning networks. The reactive power optimization module is used to build and solve the reactive power voltage evaluation optimization model and generate reactive power compensation control strategies. The evaluation module is used to establish a model performance evaluation index system to quantitatively evaluate the accuracy, reliability and stability of the reactive voltage evaluation optimization model. The correction module is used to adaptively correct the voltage prediction model and the reactive voltage assessment optimization model based on the evaluation results of the evaluation module using an online learning algorithm. The scheduling interface module is used to send the reactive voltage optimization control strategy generated by the reactive power optimization module to the power grid reactive power compensation equipment, and to receive the execution feedback from the power grid reactive power compensation equipment. The modules interact with each other through standardized data interfaces, and the system supports coordinated operation of three time scales: daily forecasting, intraday rolling forecasting, and real-time optimization.

Images