A real-time scheduling method for large power grid operation based on model predictive control

Abstract

The application discloses a kind of big power grid operation real-time scheduling method based on model prediction control, it is related to intelligent control technical field, including, to power grid historical operation data and meteorological forecast information are extracted features, generate fusion feature data, and generate power grid load prediction data through deep learning prediction model;Based on preprocessed power grid operation data, the voltage phasor of power grid node, line power flow and frequency deviation are calculated, and power grid state estimation result is generated;Using power grid state estimation result, potential abnormal mode is identified through machine learning algorithm, and power grid early warning signal is generated;According to power grid early warning signal and power grid state estimation result, the complexity of scheduling problem and the constraint tension degree of assessment, and generate optimizer selection result;Through fusion multi-source data preprocessing, deep learning load prediction, high-precision state estimation and model prediction control, reliable control instruction can be quickly generated, safety, real-time and adaptability are comprehensively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a real-time scheduling method for large power grid operation based on model predictive control. Background Technology

[0002] With the continuous expansion of modern power systems and the large-scale integration of renewable energy, the operation and dispatch of large power grids face higher requirements for real-time performance, robustness, and optimization performance. Traditional power grid dispatch mainly relies on static dispatch modes based on empirical rules and periodic optimal power flow calculations. Its dispatch cycle is long and it is difficult to adapt to dynamic changes in the system. Model Predictive Control (MPC) has been gradually introduced into the field of power dispatch because it has the ability to explicitly handle multivariable constraints, rolling optimization, and feedforward prediction. MPC solves a finite-time optimal control problem in each control cycle, combines the current state of the system with predicted information such as future load and power generation output, generates optimal dispatch instructions, and re-optimizes based on the updated state at the next time step.

[0003] At the scale of large power grids, existing MPC-based scheduling methods still face challenges in dealing with system complexity, uncertainty, and real-time requirements, especially in the fusion of multi-source heterogeneous data and dynamic operating condition identification. Traditional methods usually rely on pre-defined linearized system models and static feature extraction mechanisms, which make it difficult to fully explore the nonlinear correlation between historical operating data and external environmental factors (such as weather changes), resulting in limited load forecasting accuracy. Existing MPC scheduling frameworks usually use fixed optimization solvers and do not dynamically adjust the optimization strategy according to the complexity and constraint tension of the current power grid operating state. This may lead to low solution efficiency or degraded control performance when the system is in an emergency or critical state. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a real-time scheduling method for large power grid operation based on model predictive control to solve the problem of insufficient load forecasting accuracy caused by deep fusion of multi-source data and optimizer selection mechanism.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a real-time scheduling method for large power grid operation based on model predictive control, comprising:

[0008] The process involves collecting power grid operation data, removing outliers, filling in missing values, and standardizing the format to generate preprocessed power grid operation data. Feature extraction is performed on historical power grid operation data and meteorological forecasts to generate fused feature data. A deep learning prediction model is then used to generate power grid load forecast data. Based on the preprocessed power grid operation data, voltage phasors, line power flows, and frequency deviations at power grid nodes are calculated to generate power grid state estimation results. Using these power grid state estimation results, machine learning algorithms are employed to identify potential anomaly patterns and generate power grid early warning signals. Based on the power grid early warning signals and power grid state estimation results, the complexity and constraint stress of the scheduling problem are assessed, and optimizer selection results are generated. Based on the optimizer selection results, power grid state estimation results, and power grid load forecast data, an MPC optimization problem is constructed and solved to generate standardized scheduling control instructions. These instructions are then used to control power grid nodes and generate a scheduling execution report.

[0009] As a preferred embodiment of the real-time scheduling method for large power grid operation based on model predictive control described in this invention, the power grid operation data includes real-time load, voltage, current, frequency, and new energy output information of power grid nodes.

[0010] As a preferred embodiment of the real-time scheduling method for large power grid operation based on model predictive control described in this invention, the specific steps for generating preprocessed power grid operation data are as follows:

[0011] Statistical tests are performed on power grid operation data to identify and remove abnormal data points with short-term and drastic fluctuations, generating abnormal removal operation data.

[0012] Missing values ​​in the abnormal removal operation data are filled in to generate filled operation data. The filled operation data is then standardized in terms of time resolution, units and coding methods to form preprocessed power grid operation data.

[0013] As a preferred embodiment of the real-time scheduling method for large power grid operation based on model predictive control described in this invention, the specific steps for generating power grid load forecast data are as follows:

[0014] Extract the periodicity of load curves, peak-to-valley difference characteristics, temperature sensitivity characteristics, and historical abnormal operating condition identifiers from historical power grid operation data to generate a set of historical features;

[0015] Time-series trend analysis is performed on meteorological forecast information to generate a set of meteorological features; the historical feature set is then combined with the meteorological feature set to generate fused feature data.

[0016] The fused feature data is input into a deep learning prediction model, which then progressively predicts the grid load curves and renewable energy output curves for multiple future time periods, generating grid load prediction data.

[0017] As a preferred embodiment of the real-time scheduling method for large power grid operation based on model predictive control described in this invention, the specific steps for generating the power grid state estimation results are as follows:

[0018] Based on the power grid topology and line parameters, a nonlinear power flow equation is established with the real and imaginary parts of the grid node voltage as state variables. Voltage amplitude measurement, active and reactive power measurement and current measurement are collected and uniformly expressed as functional relationships of state variables to form a set of state estimation equations.

[0019] By combining preprocessed power grid operation data and state estimation equations, a weighted least squares objective function is constructed. The Newton iteration method is then used to iteratively solve the weighted objective function, gradually correcting the estimated values ​​of voltage phasors at power grid nodes and generating voltage phasor estimation results.

[0020] The voltage phasor estimation results are substituted into the nonlinear power flow equation to calculate the active power and reactive power flow of the power grid lines, generating the line power flow estimation results.

[0021] Frequency measurement data is collected and combined with voltage phasor estimation results to calculate the deviation of the grid node frequency from the nominal frequency, generate frequency deviation estimation results, and integrate the line power flow estimation results and frequency deviation estimation results to generate grid state estimation results.

[0022] As a preferred embodiment of the real-time scheduling method for large power grid operation based on model predictive control described in this invention, the specific steps for generating power grid early warning signals are as follows:

[0023] Standardize the preprocessed power grid operation data to generate a standardized state feature sequence;

[0024] The standardized state feature sequences are input into the parallel anomaly detection process. The isolation tree is constructed and the path length of each sample is calculated through the isolation forest algorithm. The path length is then converted into an isolation forest anomaly score.

[0025] Using a one-dimensional convolutional neural network process, classification calculations are performed using convolution and pooling operations to generate convolutional neural network anomaly scores.

[0026] A weighted fusion decision is made on the anomaly scores of the convolutional neural network and the isolated forest to obtain a comprehensive anomaly score. When the comprehensive anomaly score exceeds the anomaly threshold, an anomaly alarm is triggered, the abnormal device is located, and the anomaly location information is obtained.

[0027] Based on anomaly location information, a clustering algorithm is used to analyze anomaly patterns and generate anomaly classification results. Combined with power grid state estimation results, fuzzy reasoning is used to classify early warnings and generate power grid early warning signals.

[0028] As a preferred embodiment of the real-time scheduling method for large power grid operation based on model predictive control described in this invention, the specific steps for generating the optimizer selection result are as follows:

[0029] By integrating information on anomaly type, severity level, and impact range from power grid early warning signals, and power grid node voltage, line power, and frequency operating parameters from power grid state estimation results, a current power grid state description is formed.

[0030] Based on the current power grid status description, multi-dimensional evaluation indicators are used to generate constraint tension index and problem complexity score; according to the problem complexity score and constraint tension index, the constraint tension index and problem complexity are divided into three levels: low, medium and high, respectively, to obtain the classification results;

[0031] Based on the hierarchical results, optimizer selection decision rules are established, and optimizer selection results are obtained.

[0032] As a preferred embodiment of the real-time scheduling method for large power grid operation based on model predictive control described in this invention, the specific steps for establishing optimizer selection decision rules based on hierarchical results and obtaining optimizer selection results are as follows.

[0033] Based on the results of the hierarchical evaluation, an optimizer selection decision rule is established: when both the constraint tension index and the problem complexity are at a low level, the interior point method is selected.

[0034] When either the constraint tension index or the problem complexity reaches a medium level, the effective set method is selected; when either the constraint tension index or the problem complexity reaches a high level, the alternating direction multiplier method is selected, generating the optimizer selection result.

[0035] As a preferred embodiment of the model predictive control-based real-time dispatching method for large power grid operation described in this invention, the specific steps for generating standardized dispatching control commands are as follows:

[0036] Based on the optimizer selection results, combined with the power grid state estimation results and power grid load forecast data, a model predictive control optimization problem is constructed, and a mathematical expression of the optimization problem is generated.

[0037] An optimizer is used to solve the mathematical expression of the optimization problem, and an optimal scheduling scheme is obtained through iterative calculation.

[0038] The feasibility of the optimized scheduling scheme is verified, a set of scheduling control instructions is generated, and the instructions are sorted and formatted in chronological order to generate standardized scheduling control instructions.

[0039] As a preferred embodiment of the model predictive control-based real-time scheduling method for large power grid operation described in this invention, the specific steps for controlling power grid nodes and generating scheduling execution reports are as follows:

[0040] Scheduling is performed based on standardized scheduling control commands. By collecting feedback signals from the execution equipment and power grid measurement data, the execution results of the standardized scheduling control commands are verified, and command execution verification results are generated.

[0041] Based on the command execution verification results and real-time power grid measurement data, the execution effect of standardized dispatch control commands is evaluated, the deviation between the execution effect and the expected target is calculated, and a dispatch execution report is generated.

[0042] The beneficial effects of this invention are as follows: by integrating multi-source data preprocessing, deep learning load prediction, high-precision state estimation and model predictive control, it achieves early identification and intelligent judgment of potential operational risks, enhances the ability to predict complex faults and proactive response, and can provide early warning of risks and quickly generate reliable control commands when facing complex scenarios such as new energy fluctuations and load changes, thus achieving a comprehensive improvement in safety, real-time performance and adaptability. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of a real-time scheduling method for large power grid operation based on model predictive control.

[0045] Figure 2 A flowchart for generating preprocessed power grid operation data.

[0046] Figure 3 This is a flowchart of load forecasting and feature fusion.

[0047] Figure 4 This is a flowchart for anomaly detection and early warning. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0050] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0051] Reference Figures 1-4 This is one embodiment of the present invention, which provides a real-time scheduling method for large power grid operation based on model predictive control, comprising the following steps:

[0052] S1. Collect power grid operation data, remove outliers, fill in missing values, and format the power grid operation data to generate preprocessed power grid operation data.

[0053] S1.1. Power grid operation data includes real-time load, voltage, current, frequency, and renewable energy output information of power grid nodes.

[0054] S1.2. Perform statistical tests on the power grid operation data, identify and remove abnormal data points with short-term drastic fluctuations, and generate abnormal removal operation data.

[0055] Specifically, when performing statistical tests on power grid operation data, the sliding window quartile method is used for outlier identification and removal. A sliding window of fixed time length is set, for example, 15 sampling periods. Within each sliding window, the 25th percentile and 75th percentile are calculated for each data channel in the power grid operation data, and the interquartile range is calculated. Data points in each data channel that exceed the range from the 25th percentile minus 1.5 times the interquartile range to the 75th percentile plus 1.5 times the interquartile range are identified as outlier data points and removed from the power grid operation data to generate outlier-removed operation data.

[0056] S1.3. Complete the missing values ​​of the abnormal removal operation data to generate filled operation data. Unify the time resolution of the filled operation data and standardize the units and coding methods to form preprocessed power grid operation data.

[0057] Specifically, the original sampling period is determined based on the time interval between adjacent valid sampling points in the abnormal removal operation data, and a complete time index is generated for each measurement sequence of each power grid node. For each missing time point, the missing value is calculated in a linear proportion using the adjacent time series interpolation method when there are valid sampling points before and after it. For the missing time points at the beginning or end, forward or backward filling is used as a supplementary value when there is a nearest neighbor valid sampling value. If the nearest neighbor valid sampling value is unavailable and the length of the missing interval exceeds a certain number of sampling periods, the time band average of the historical operation data of the power grid in the same period is used as an example supplementary value.

[0058] After completing the interpolation and imputation of adjacent time series for all missing time points, the interpolation results are merged with the valid sampling points and sorted by time to generate imputed operation data. The imputed operation data is resampled to unify the time resolution and linear interpolation is used to adjust the samples to the target sampling period, for example, unifying the time resolution to the minute level. The resampled imputed operation data is standardized in terms of units and coding, unifying active power and reactive power to kilowatts (kW), voltage to volts (V), current to amperes (A), and frequency to hertz (Hz). The grid node identifier is standardized to a unified identifier code and the timestamp format is standardized to ISO-8601. The preprocessed grid operation data is then output.

[0059] S2. Extract features from historical power grid operation data and meteorological forecast information to generate fused feature data, and generate power grid load forecast data through a deep learning prediction model.

[0060] S2.1. Extract the periodicity characteristics of the load curve, peak-valley difference characteristics, temperature sensitivity characteristics, and historical abnormal operating condition identifiers from the historical power grid operation data to generate a historical feature set.

[0061] Specifically, based on historical power grid operation data, the Fast Fourier Transform is used to extract periodic characteristic components such as daily and weekly cycles from the load curve. The daily peak-to-valley difference characteristic is obtained by calculating the difference and ratio between the daily maximum load and minimum load. A linear regression relationship between load value and temperature is established to extract the regression coefficient and determination coefficient in the temperature sensitivity characteristic. At the same time, historical abnormal operating conditions are marked according to the fault waveform and protection action event in the historical records. After being aligned and integrated according to the time series, a historical feature set is generated.

[0062] It should be noted that the periodic feature extraction uses Fast Fourier Transform to obtain the main frequency components; the peak-valley difference feature is obtained through daily load extreme values; the temperature sensitivity feature is fitted by linear regression using the least squares method; and the historical abnormal operating condition identifiers are directly taken from the power grid event log file.

[0063] Historical power grid operation data includes load data (including historical active and reactive power records for each node and the overall system), generation data (including historical output, start-up and shutdown status, and operating limits of various generator sets), power grid measurement data (including SCADA and PMU measurement records of historical voltage amplitude, current amplitude, active power, and reactive power), environmental data (including meteorological information such as historical temperature, humidity, wind speed, and solar radiation intensity), and event record data (including event sequences such as historical fault waveform files, protection action signals, and switch change records).

[0064] S2.2. Perform time-series trend analysis on meteorological forecast information to generate a set of meteorological features; combine the historical feature set with the meteorological feature set through feature splicing to generate fused feature data.

[0065] Specifically, meteorological forecast information includes temperature, humidity, and wind speed data;

[0066] The moving average method is used to smooth the temperature, humidity and wind speed data. The temporal trend slope and curvature characteristics of each meteorological element are calculated by multinomial fitting. Principal component analysis is used to reduce the dimensionality and retain the principal components whose variance contribution rate exceeds 95% of the example values, thus generating a set of meteorological features.

[0067] The meteorological feature set and the historical feature set are aligned by timestamp, and vertical fusion is performed by feature vector concatenation. Weights are assigned to load features and meteorological features, and weighted concatenation is performed to generate fused feature data.

[0068] It should be noted that assigning weights to load characteristics and meteorological characteristics is done by testing different weight combinations (such as load characteristics 0.5-0.7 and meteorological characteristics 0.3-0.5) using a grid search method, and selecting the weight combination that minimizes the prediction error.

[0069] The trend slope of the time series represents the average rate and direction of change of meteorological elements (temperature, humidity, wind speed) over time. A positive slope indicates that the meteorological element is trending upward over time, while a negative slope indicates a downward trend. The absolute value reflects the speed of change. The first-order polynomial (linear fitting) of the meteorological data series after moving average smoothing is performed by least squares method, and the coefficients of the first-order terms are the trend slope characteristics.

[0070] Curvature features characterize the acceleration of the rate of change of meteorological elements, that is, the rate of change of the trend slope itself. Positive curvature indicates that the rate of change is accelerating (accelerating upward or accelerating downward), negative curvature indicates that the rate of change is slowing down (the upward or downward trend is weakening), and zero curvature indicates uniform change. The curvature feature is obtained by fitting a second-order polynomial (quadratic curve fitting) to the same meteorological data sequence using the least squares method. Twice the coefficients of the second-order terms are used to fit the curvature feature.

[0071] S2.3. Input the fused feature data into the deep learning prediction model. The deep learning prediction model will gradually predict the power grid load curve and the new energy output curve for multiple future time periods to generate power grid load prediction data.

[0072] Specifically, the fused feature data is input into the deep learning prediction model. The deep learning prediction model adopts an encoder-decoder architecture. The encoder uses a bidirectional long short-term memory network to extract the spatiotemporal features of the input sequence. The decoder uses an attention mechanism to fuse feature information and gradually predicts the grid load curve and renewable energy output curve for multiple future time periods (e.g., the next 24 time periods) through an autoregressive approach. The output of each time step in the prediction process is used as the input of the next time step, and the grid load prediction data is output.

[0073] In the deep learning prediction model processing, the encoder uses a bidirectional long short-term memory network to process the temporal dependencies in the fused feature data; the decoder uses an attention mechanism to calculate feature weights and generates multi-step prediction results through an autoregressive loop; and converts the predicted values ​​back to the original dimensions through inverse normalization, and Kalman filtering is used to smooth the fluctuations of the prediction curve.

[0074] It should be noted that the training process of the deep learning prediction model is as follows: the load curve and new energy output curve of the corresponding period in the fused feature data and the historical power grid operation data are used to construct the input sequence and label sequence. The input sequence is input into the deep learning prediction model and the prediction output is calculated through forward propagation. The error between the prediction output and the label sequence is used as the loss function value. The parameters of the deep learning prediction model are iteratively updated using an optimization algorithm based on gradient descent to obtain the trained deep learning prediction model.

[0075] S3. Based on the preprocessed power grid operation data, calculate the voltage phasors, line power flow and frequency deviation of the power grid nodes, and generate power grid state estimation results.

[0076] S3.1. Based on the power grid topology and line parameters, establish nonlinear power flow equations with the real and imaginary parts of the grid node voltage as state variables. Collect voltage amplitude measurements, active and reactive power measurements, and current measurements, and uniformly express them as functional relationships of state variables to form a set of state estimation equations.

[0077] Specifically, based on the node connection relationships in the power grid topology and the resistance and reactance values ​​in the line parameters, a nonlinear power flow equation is established with the real and imaginary parts of the grid node voltage as state variables. The collected voltage amplitude measurement, active power measurement, reactive power measurement, and current measurement are expressed as functional relationships between the real and imaginary parts of the grid node voltage, respectively. The voltage amplitude measurement is expressed as the square root of the sum of the squares of the real and imaginary parts of the grid node voltage. The active power measurement and reactive power measurement are expressed as complex operation functions of the real and imaginary parts of the grid node voltage and the line admittance matrix. The current measurement is expressed as a complex operation function of the voltage difference and the line impedance, forming a set of state estimation equations.

[0078] It should be noted that by collecting the open / closed status of circuit breakers and disconnectors in the substation in real time through Supervisory Control and Data Acquisition (SCADA), the actual operation wiring mode of the power grid is determined, and the power grid topology is obtained;

[0079] Line parameters are obtained through on-site measurements (e.g., using a line parameter tester);

[0080] The line admittance matrix is ​​a mathematical tool that describes the topology of a power grid (who is connected to whom) and the physical parameters of all lines (the conductivity of each line). The line admittance matrix is ​​obtained by using the power grid topology and line parameters, according to Kirchhoff's laws and Ohm's law in circuit theory, to obtain the self-admittance of all nodes and the mutual admittance between nodes, and then arranging them into a square matrix according to the node numbers.

[0081] S3.2. Combine the preprocessed power grid operation data and the state estimation equations to construct a weighted least squares objective function, and use the Newton-Raphson iteration method to iteratively solve the weighted objective function, gradually correcting the estimated values ​​of the voltage phasors of the power grid nodes, and generating voltage phasor estimation results.

[0082] Specifically, based on the measured values ​​and measurement error statistics in the preprocessed power grid operation data, the inverse matrix of the measurement error covariance matrix is ​​constructed as the weight matrix. The functional relationship between the measurement equations and state variables in the state estimation equation set is substituted into the weighted least squares objective function expression to form the weighted least squares objective function.

[0083] The Newton-Raphson iteration method is used to iteratively solve the weighted least squares objective function. The initial point is a flat start or historical estimate. In each iteration, the gradient vector and Hessian matrix of the objective function are calculated. The linear equation system is solved to obtain the state variable correction and update the estimate. The converged estimate of the real part and imaginary part of the grid node voltage is output to generate the voltage phasor estimation result.

[0084] S3.3. Substitute the voltage phasor estimation results into the nonlinear power flow equation to calculate the active power and reactive power flow of the power grid lines, and generate the line power flow estimation results.

[0085] Specifically, based on the estimated real and imaginary parts of the grid node voltage in the voltage phasor estimation results, combined with the node connection relationships in the grid topology and the resistance and reactance values ​​in the line parameters, the active and reactive power flows at the beginning and end of each line are calculated using the power calculation formula in the nonlinear power flow equation. The power flow results of all lines are verified to ensure that the power balance condition is met, and the line power flow estimation results are generated.

[0086] It should be noted that the expression for calculating the active power flow of the line is:

[0087] ;

[0088] in, Indicates from the first end of the power grid node Flow to end grid nodes The estimated active power, Indicates the first-end power grid node. Indicates the end-point grid node. Indicates the first-end power grid node Flow to end grid nodes electrical conductivity, Indicates the first-end power grid node Flow to end grid nodes susceptivity, Indicates the first-end power grid node The estimated real part of the voltage. Indicates the end grid node The estimated real part of the voltage. Indicates the first-end power grid node The estimated value of the imaginary part of the voltage. Indicates the end grid node The estimated value of the imaginary part of the voltage;

[0089] The expression for calculating the reactive power flow of a line is:

[0090] ;

[0091] in, Indicates the first-end power grid node Flow to end grid nodes The estimated reactive power.

[0092] S3.4. Collect frequency measurement data and combine it with voltage phasor estimation results to calculate the deviation of the grid node frequency from the nominal frequency, generate frequency deviation estimation results, and integrate the line power flow estimation results and frequency deviation estimation results to generate grid state estimation results.

[0093] Specifically, frequency measurement data is collected, and the node voltage phase angle information in the voltage phasor estimation results is combined with the weighted average method to calculate the deviation of the grid node frequency from the 50Hz nominal frequency: frequency measurement data of each node is obtained from the synchronous phasor measurement device, the rate of change of voltage phase angle difference between adjacent nodes is calculated based on the voltage phasor estimation results, the direct frequency measurement and the frequency value derived from the phase angle difference are weighted and fused, the deviation value relative to the 50Hz nominal frequency is calculated, and the frequency deviation estimation result is generated.

[0094] The line power flow estimation results, frequency deviation estimation results, and voltage phasor estimation results are integrated with the data fusion method based on confidence level to generate power grid state estimation results.

[0095] S4. Using the power grid state estimation results, identify potential abnormal patterns through machine learning algorithms and generate power grid early warning signals.

[0096] S4.1. Standardize the preprocessed power grid operation data to generate a standardized state feature sequence.

[0097] Specifically, the sliding window Z-score standardization method is used to standardize the preprocessed power grid operation data. A fixed-length time window (e.g., a window length of 60 sampling periods) is used to calculate the arithmetic mean and standard deviation of each feature parameter in the preprocessed power grid operation data. The mean effect is eliminated by centering, and then unit variance is achieved by scaling, so that each feature parameter is converted into a standard normal distribution with a mean of zero and a standard deviation of one, generating a standardized state feature sequence.

[0098] S4.2. Input the standardized state feature sequences into the parallel anomaly detection process, construct the isolation tree through the isolation forest algorithm, calculate the path length of each sample, and convert the path length into an isolation forest anomaly score.

[0099] Specifically, the standardized state feature sequence is input into the isolated forest algorithm process, multiple isolation trees are constructed using a random partitioning strategy (e.g., 100 isolation trees are constructed), and recursive random partitioning is performed. The path length of the multidimensional feature vector corresponding to each sampling time in the standardized state feature sequence from the root node to the terminal node is recorded, and the path length is calculated to obtain the isolated forest anomaly score through a mathematical transformation formula.

[0100] It should be noted that the isolated forest algorithm constructs isolation trees based on the multidimensional feature parameters in the standardized state feature sequence; the path length represents the ease with which the multidimensional feature vector corresponding to each sampling time in the standardized state feature sequence is isolated.

[0101] It should be noted that the expression for calculating the anomaly score of an isolated forest is as follows:

[0102] ;

[0103] in, Indicates sample Isolated forest anomaly scores, Indicates the sampling time index. Indicates the subsample size. Indicates sample Path length in the isolation tree This represents the normalization factor.

[0104] It should also be noted that, This refers to the number of samples randomly drawn when constructing each isolation tree. It is a comprehensive assessment of the entire forest, making it more stable than the results of a single tree. A small value indicates that the sample point is easily isolated in the entire forest and is very likely an outlier. A large value indicates that the sample point requires many segments to be isolated in the entire forest, suggesting that it is in a data-dense area and is very likely a normal point.

[0105] S4.3. Through a one-dimensional convolutional neural network process, the standardized state feature sequence is extracted using convolution and pooling operations and classified to generate a convolutional neural network anomaly score.

[0106] Specifically, the standardized state feature sequence is input into a one-dimensional convolutional neural network process, and temporal feature extraction is performed using multiple convolutional kernels (e.g., using 32 convolutional kernels of length 3). Nonlinear transformation is achieved through the ReLU activation function. After dimensionality reduction through max pooling, the output feature map is passed to a fully connected layer. The anomaly probability score at each time point is calculated through the Sigmoid function to generate the convolutional neural network anomaly score.

[0107] In a one-dimensional convolutional neural network, the convolutional kernel slides to extract local temporal patterns in the normalized state feature sequence; the max pooling operation preserves significant features (e.g., the active and reactive power of a certain line changes in the same trend under normal conditions, but reverses under abnormal conditions) and reduces the number of parameters; the fully connected layer integrates global feature information; the sigmoid function maps the output to anomaly probabilities between 0 and 1.

[0108] It should be noted that historical normal state data and abnormal state data are used as training sets, cross-entropy loss function is used as optimization objective, backpropagation algorithm is used to update network weight parameters, and gradient descent method is used to minimize prediction error to obtain a one-dimensional convolutional neural network after training.

[0109] S4.4. Perform a weighted fusion decision on the anomaly scores of the convolutional neural network and the isolated forest to obtain a comprehensive anomaly score. When the comprehensive anomaly score exceeds the anomaly threshold, trigger an anomaly alarm, locate the abnormal device, and obtain anomaly location information.

[0110] Specifically, when making a weighted fusion decision on the anomaly scores of convolutional neural networks and isolated forests, a fixed weight allocation method is adopted (e.g., the weight of the anomaly score of the convolutional neural network is 0.6, and the weight of the anomaly score of the isolated forest is 0.4). The comprehensive anomaly score is calculated by weighted summation. When the comprehensive anomaly score exceeds the anomaly threshold (an exemplary value range is 0.75 to 0.85, derived from the statistical analysis of the distribution of anomaly scores in historical normal operation data and the balance of operational risk control objectives), an anomaly alarm is triggered. The abnormal equipment is located by analyzing the feature contribution and anomaly location information is generated.

[0111] Furthermore, by analyzing the contribution of features to locate abnormal devices, specifically by calculating the gradient response value of the operating parameters of each monitored power device or node in the standardized state feature sequence in the convolutional neural network and combining it with the importance of the segmentation path features of the isolated forest, the device with the highest response value and the greatest feature importance is identified as the source of the anomaly.

[0112] It should be noted that the importance of the splitting path features in the isolated forest is obtained by statistically analyzing the frequency and depth at which these features are selected as splitting points during the construction of the isolation trees. In the process of constructing multiple isolation trees, the number of times a node is used for splitting is recorded, and the average depth at which the split occurs (i.e., the distance from the root node) is considered to obtain the importance of the splitting path features in the isolated forest.

[0113] The anomaly threshold is obtained by analyzing historical normal operation data. The anomaly scores of the convolutional neural network and the isolated forest are collected over a sufficiently long period of time (such as 3 to 6 months) of normal power grid operation. The statistical distribution of the comprehensive anomaly score corresponding to the normal data is calculated, and the high quantile of the statistical distribution (such as the 99th percentile or 99.5th percentile) is selected as the anomaly threshold.

[0114] S4.5. Based on anomaly location information, clustering algorithm is used to analyze anomaly patterns, generate anomaly classification results, and combine them with power grid state estimation results to perform early warning classification through fuzzy reasoning, generating power grid early warning signals.

[0115] Specifically, based on anomaly location information and power grid state estimation results, the anomaly location information and anomaly location scores are aligned and normalized with the power grid node voltage deviation, line power margin, and frequency deviation in the power grid state estimation results at the corresponding time points in the time and node dimensions to form an anomaly sample feature vector set. The density-based spatial clustering algorithm DBSCAN is applied to the anomaly sample feature vector set for clustering. By setting the neighborhood radius and minimum number of samples (e.g., example neighborhood radius = 0.5, minimum number of samples = 5), several anomaly pattern clusters are divided, and the statistics and spatial distribution characteristics of each anomaly pattern cluster are calculated.

[0116] Based on the statistical and spatial distribution characteristics of the abnormal pattern clusters, the abnormal pattern clusters are mapped to the abnormal classification results. When the abnormal pattern clusters are concentrated in a single or a few power grid nodes and are accompanied by a continuous drop in power grid node voltage and local line active power anomalies, they are identified as equipment failure type. When the abnormal pattern clusters cover a wide range of power grid nodes and there is a sudden large active power shift that is synchronous in time, they are identified as network attack type. When the abnormal pattern clusters and meteorological characteristics appear simultaneously and show a gradual change in amplitude, they are identified as natural disturbance type.

[0117] Using the anomaly classification results and the corresponding comprehensive anomaly scores as input, fuzzy reasoning is used to implicitly map the severity of anomalies, power grid vulnerability indicators, and fault propagation risks to generate power grid early warning signals.

[0118] S5. Based on the power grid early warning signals and power grid state estimation results, assess the complexity and constraint tension of the scheduling problem, and generate optimizer selection results.

[0119] S5.1. Integrate the information on anomaly type, severity level and impact range in the power grid early warning signal, and the power grid node voltage, line power and frequency operating parameters in the power grid state estimation results to form a description of the current power grid state.

[0120] Specifically, based on the information on anomaly type, severity level, and impact range in the power grid early warning signal, and combined with the power grid node voltage amplitude, voltage phase angle, active power, reactive power, and system frequency deviation operating parameters in the power grid state estimation results, a multi-dimensional state description method is used to integrate and form a complete description of the current power grid state. The complete description of the current power grid state reflects the attribute characteristics of the abnormal event while maintaining the accurate numerical characteristics of the power grid operating parameters.

[0121] S5.2. Based on the current power grid status description, multi-dimensional evaluation indicators are used to generate constraint tension index and problem complexity score; according to the problem complexity score and constraint tension index, the constraint tension index and problem complexity are divided into three levels: low, medium and high, respectively, to obtain the classification results.

[0122] Specifically, based on the current power grid status description, multi-dimensional evaluation indicators are used to calculate the constraint tension index and problem complexity score. The constraint tension index is obtained by weighted summation of node voltage deviation rate, line power load rate and frequency deviation value.

[0123] The problem complexity score is obtained by linearly combining the anomaly type weights (0.3 for equipment failure, 0.5 for network attack, and 0.2 for natural disturbance), the anomaly impact range coefficient, and the prediction time domain length coefficient.

[0124] The constraint tension is divided into three levels: low, medium, and high (the first 30% is low, the middle 40% is medium, and the last 30% is high) using the percentile division method, and the problem complexity is divided into three levels to generate a graded result.

[0125] It should be noted that the weights of node voltage deviation rate, line power load rate, and frequency deviation value are allocated through statistical analysis of power grid fault events over the years, based on the proportion of cascading faults caused by line overload, voltage anomaly, and frequency anomaly.

[0126] Network attack anomalies are characterized by their suddenness, rapid propagation, and potential for chain reactions, making them the most urgent to respond to, and therefore they are assigned a higher weight (Example 0.5). Equipment failure anomalies usually have obvious warning signs and their impact is relatively controllable, so they are of secondary importance (Example 0.3). Natural disturbance anomalies are often predictable and have a relatively slow development process, so they are of the lowest importance (Example 0.2).

[0127] S5.3. Based on the hierarchical results, establish optimizer selection decision rules and obtain optimizer selection results.

[0128] S5.3.1 Based on the results of the hierarchical evaluation, establish the optimizer selection decision rule. When both the constraint tension index and the problem complexity are at a low level, the interior point method is selected.

[0129] Specifically, based on the combination of constraint tension index level and problem complexity level in the graded evaluation results, an optimizer selection decision rule is established. When the constraint tension index level is low and the problem complexity level is low, the interior point method is selected as the optimization solution method and instantiated as an interior point optimizer. The interior point optimizer is suitable for problems with low constraint tension and low complexity, and has the advantages of high computational efficiency and fast convergence speed. At the same time, parameters such as convergence tolerance, maximum number of iterations and step size are set (for example, convergence tolerance example 10^-6, maximum number of iterations example 100, step size example 0.1).

[0130] S5.3.2 When either the constraint tension index or the problem complexity reaches a medium level, the effective set method is selected; when either the constraint tension index or the problem complexity reaches a high level, the alternating direction multiplier method is selected, and the optimizer selection result is generated.

[0131] Specifically, when either the constraint tension index level or the problem complexity level reaches the medium level, the effective set method is selected as the optimization solution method and instantiated as an effective set optimizer. The effective set optimizer is suitable for problems with medium constraint tension or medium complexity and can maintain good numerical stability when the constraints are relatively tight. At the same time, parameters such as the active set strategy and the maximum number of iterations are set (for example, the maximum number of iterations is 200).

[0132] When either the constraint tension index level or the problem complexity level reaches a high level, the alternating direction multiplier method is selected as the optimization solution method and instantiated as the alternating direction multiplier optimizer. The alternating direction multiplier optimizer is suitable for high constraint tension or high complexity problems. It adopts a distributed solution framework and sets parameters such as penalty parameters and adaptive step size.

[0133] Based on the constraint tension index and the level of problem complexity, different optimizers are selected, and the optimizer selection results are obtained.

[0134] It should be noted that, in order to classify the constraint tension index, all sample data of the constraint tension index under historical normal working conditions are collected, a statistical distribution curve is plotted, and specific quantiles are selected as the level division according to management requirements. Usually, the 30th percentile is taken as the dividing threshold between low / medium level and the 70th percentile is taken as the dividing threshold between medium / high level.

[0135] To classify the complexity of a problem, collect complexity score samples corresponding to all historical events (including normal operations and various abnormal events) to form a statistical distribution of the score. Based on scheduling experience and risk control requirements, select specific quantiles as the level division. Typically, the 30th percentile is used as the boundary threshold between low and medium levels, and the 70th percentile is used as the boundary threshold between medium and high levels.

[0136] S6. Based on the optimizer selection results, power grid state estimation results, and power grid load forecast data, construct and solve the MPC optimization problem, generate standardized dispatch control instructions, control the power grid nodes, and generate a dispatch execution report.

[0137] S6.1. Based on the optimizer selection results, combined with the power grid state estimation results and power grid load forecast data, construct the model predictive control optimization problem and generate the mathematical expression of the optimization problem.

[0138] Specifically, based on the optimizer type and parameter configuration determined in the optimizer selection results, combined with the real-time data of node voltage amplitude, node voltage phase angle, line active power, line reactive power and frequency deviation provided by the power grid state estimation results, and the load demand and new energy power generation forecast values ​​for multiple future periods provided by the power grid load forecast data, a mathematical expression of the model predictive control rolling optimization problem is constructed.

[0139] By weighting and combining the three requirements of minimizing power generation fuel costs, minimizing the sum of squared voltage deviations, and frequency stability requirements, and using the power balance equations of power grid nodes, the upper and lower limits of generator active and reactive power output, the thermal stability limit of line transmission power, and the operating range of node voltage as constraints, a mathematical expression for the optimization problem is generated.

[0140] S6.2. An optimizer is used to solve the mathematical expression of the optimization problem. The corresponding solution algorithm is called according to the optimizer type, and the optimized scheduling scheme is obtained through iterative calculation.

[0141] Specifically, the corresponding solution algorithm is called according to the optimizer type in the optimizer selection result. When the interior point method is selected, the original dual interior point algorithm is used to solve the problem and the convergence tolerance is set. When the effective set method is selected, the active set strategy is used to solve the problem and the maximum number of iterations is set to the example value of 200. When the alternating direction multiplier method is selected, the distributed optimization framework is used to solve the problem and the penalty parameter is set to an adaptive adjustment strategy.

[0142] The decision variables are gradually corrected through iterative calculations. In each iteration, the gradient vector and Hessian matrix of the objective function are calculated, the linear equation system is solved to obtain the state variable correction amount and update the estimated value. The calculation stops when the change in the objective function value is less than the convergence tolerance or the maximum number of iterations is reached. The optimized scheduling schemes for generator active and reactive power output setpoints, reactive power compensation equipment switching status and load adjustment schemes for multiple future time periods are obtained.

[0143] S6.3. Perform feasibility verification on the optimized scheduling scheme, generate a set of scheduling control instructions, and sort and format them in chronological order to generate standardized scheduling control instructions.

[0144] Specifically, the feasibility of the optimized scheduling scheme is verified. The power flow method is used to verify whether the optimized scheduling scheme meets the power balance constraints of all nodes. The over-limit check method is used to confirm whether the line power exceeds the thermal stability limit. The voltage stability assessment method is used to determine whether the node voltage is within the allowable operating range (e.g., 0.95 ~ 1.05 per-unit value (pu)). For over-limits or conflicts found during verification, conservative processing and safety constraint priority strategies are used for correction. For over-limits or conflicts found during verification, the execution of the optimized scheme with safety risks is immediately terminated, and the system is rolled back to the known safe scheduling state of the previous cycle. Based on the principle of absolute safety, the control commands are recalculated. Priority is given to reducing the power output of power sources that cause overload of critical equipment (such as lines and transformers), and reserve capacity is unconditionally put into operation to ensure power balance and voltage stability until the power flow calculation verifies that all safety constraints are met, and a feasible scheduling scheme that ensures the safe operation of the power grid is generated.

[0145] Based on the generator active and reactive power output setpoints, reactive power compensation equipment switching status, and load adjustment scheme in the feasible scheduling scheme, generator active power setpoint commands, generator reactive power setpoint commands, capacitor bank switching commands, reactor bank switching commands, and interruptible load control commands are generated through command mapping. The commands are sorted in chronological order, and data is encapsulated using a standard communication protocol format to generate standardized scheduling control commands.

[0146] It should be noted that power balance constraint is the most basic and core physical law that must be followed in power operation. Thermal stability limit is the maximum current (unit: Ampere A) or maximum apparent power (unit: Megavolt-ampere MVA) that a transmission line can continuously carry under the premise of ensuring safety.

[0147] S6.4. The execution equipment of the power grid node performs scheduling based on standardized scheduling control commands. By collecting feedback signals from the execution equipment and power grid measurement data, the execution results of the standardized scheduling control commands are verified, and command execution verification results are generated.

[0148] Specifically, the execution equipment at the power grid nodes (including generator excitation regulators, reactive power compensation device controllers, load control terminals, etc.) receives and executes standardized dispatch control commands. At the same time, the data acquisition and monitoring devices collect feedback signals from the execution equipment (such as actual generator output and actual capacitor switching status) and power grid measurement data (such as node voltage and line power) in real time. The feedback signals and measurement data are compared and analyzed with the expected values ​​of the standardized dispatch control commands to calculate the command execution deviation rate (for example, a generator output deviation rate of no more than 2% is considered qualified). For commands with deviations exceeding the allowable range, a retransmission mechanism is initiated to generate command execution verification results.

[0149] S6.5. Based on the command execution verification results and real-time power grid measurement data, evaluate the execution effect of standardized dispatch control commands, calculate the degree of deviation between the execution effect and the expected target, and generate a dispatch execution report.

[0150] Specifically, based on the command execution success rate, deviation statistics, and execution timeliness data in the command execution verification results, combined with the node voltage, line power, and frequency measurements in the real-time power grid measurement data, the execution effect of standardized dispatch control commands is evaluated. The difference calculation method is used to calculate the degree of deviation between each execution effect and the expected target, the percentage deviation of voltage qualification rate from the target value, the degree of deviation of frequency deviation from the allowable range, the difference between line load rate and thermal stability limit, and the error rate of generator output from the set value. These are then compared and analyzed with historical data from the same period to generate a dispatch execution report.

[0151] It should be noted that the expression for calculating the deviation between the execution effect and the expected goal is as follows:

[0152] ;

[0153] in, This indicates the degree of deviation between the performance and the expected goals. This represents the actual measured value. This represents the expected target value. This indicates the reference base value.

[0154] This embodiment also provides a computer device applicable to the real-time scheduling method for large power grid operation based on model predictive control, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the real-time scheduling method for large power grid operation based on model predictive control as proposed in the above embodiment.

[0155] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0156] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the real-time scheduling method for large power grid operation based on model predictive control as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0157] In summary, this invention achieves early identification and intelligent judgment of potential operational risks by integrating multi-source data preprocessing, deep learning load prediction, high-precision state estimation, and model predictive control. It enhances the ability to predict and respond proactively to complex faults. When facing complex scenarios such as new energy fluctuations and sudden load changes, it can not only provide early warnings of risks but also quickly generate reliable control commands, thus achieving a comprehensive improvement in safety, real-time performance, and adaptability.

[0158] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A real-time scheduling method for large power grid operation based on model predictive control, characterized in that: include, Collect power grid operation data, remove outliers, fill in missing values, and format the data to generate preprocessed power grid operation data; Feature extraction is performed on historical power grid operation data and meteorological forecast information to generate fused feature data, and power grid load forecast data is generated through a deep learning prediction model. Based on preprocessed power grid operation data, the voltage phasors, line power flow and frequency deviation of power grid nodes are calculated to generate power grid state estimation results; Using the power grid state estimation results, potential abnormal patterns are identified through machine learning algorithms, and power grid early warning signals are generated. Based on power grid early warning signals and power grid state estimation results, the complexity and constraint stress of the scheduling problem are assessed, and optimizer selection results are generated. The specific steps are as follows. By integrating information on anomaly type, severity level, and impact range from power grid early warning signals, and power grid node voltage, line power, and frequency operating parameters from power grid state estimation results, a current power grid state description is formed. Based on the current power grid status description, multi-dimensional evaluation indicators are used to generate constraint tension index and problem complexity score; according to the problem complexity score and constraint tension index, the constraint tension index and problem complexity are divided into three levels: low, medium and high, respectively, to obtain the classification results; Based on the hierarchical results, optimizer selection decision rules are established, and optimizer selection results are obtained. Based on the hierarchical results, an optimizer selection decision rule is established to obtain the optimizer selection result. The specific steps are as follows: Based on the classification results, an optimizer selection decision rule is established: when both the constraint tension index and the problem complexity are low, the interior point method is selected. When either the constraint tension index or the problem complexity reaches a medium level, the effective set method is selected; when either the constraint tension index or the problem complexity reaches a high level, the alternating direction multiplier method is selected, and the optimizer selection result is generated. Based on the optimizer selection results, power grid state estimation results, and power grid load forecast data, the MPC optimization problem is constructed and solved, standardized dispatch control instructions are generated, power grid nodes are controlled, and a dispatch execution report is generated.

2. The real-time dispatching method for large power grid operation based on model predictive control as described in claim 1, characterized in that: The power grid operation data includes real-time load, voltage, current, frequency, and renewable energy output information of power grid nodes.

3. The real-time scheduling method for large power grid operation based on model predictive control as described in claim 1, characterized in that: The specific steps for generating preprocessed power grid operation data are as follows. Statistical tests are performed on power grid operation data to identify and remove abnormal data points with short-term and drastic fluctuations, generating abnormal removal operation data. Missing values ​​in the abnormal removal operation data are filled in to generate filled operation data. The filled operation data is then standardized in terms of time resolution, units and coding methods to form preprocessed power grid operation data.

4. The real-time dispatching method for large power grid operation based on model predictive control as described in claim 1, characterized in that: The specific steps for generating power grid load forecast data are as follows: Extract the periodicity of load curves, peak-to-valley difference characteristics, temperature sensitivity characteristics, and historical abnormal operating condition identifiers from historical power grid operation data to generate a set of historical features; Perform time-series trend analysis on meteorological forecast information to generate a set of meteorological features; By splicing historical feature sets with meteorological feature sets, fused feature data is generated. The fused feature data is input into a deep learning prediction model, which then progressively predicts the grid load curves and renewable energy output curves for multiple future time periods, generating grid load prediction data.

5. The real-time dispatching method for large power grid operation based on model predictive control as described in claim 1, characterized in that: The specific steps for generating the power grid state estimation results are as follows: Based on the power grid topology and line parameters, a nonlinear power flow equation is established with the real and imaginary parts of the grid node voltage as state variables. Voltage amplitude measurement, active and reactive power measurement and current measurement are collected and uniformly expressed as functional relationships of state variables to form a set of state estimation equations. By combining preprocessed power grid operation data and state estimation equations, a weighted least squares objective function is constructed. The Newton iteration method is then used to iteratively solve the weighted objective function, gradually correcting the estimated values ​​of voltage phasors at power grid nodes and generating voltage phasor estimation results. The voltage phasor estimation results are substituted into the nonlinear power flow equation to calculate the active power and reactive power flow of the power grid lines, generating the line power flow estimation results. Frequency measurement data is collected and combined with voltage phasor estimation results to calculate the deviation of the grid node frequency from the nominal frequency, generate frequency deviation estimation results, and integrate the line power flow estimation results and frequency deviation estimation results to generate grid state estimation results.

6. The real-time scheduling method for large power grid operation based on model predictive control as described in claim 1, characterized in that: The specific steps for generating the power grid early warning signal are as follows: Standardize the preprocessed power grid operation data to generate a standardized state feature sequence; The standardized state feature sequences are input into the parallel anomaly detection process. The isolation tree is constructed and the path length of each sample is calculated through the isolation forest algorithm. The path length is then converted into an isolation forest anomaly score. Using a one-dimensional convolutional neural network process, classification calculations are performed using convolution and pooling operations to generate convolutional neural network anomaly scores. A weighted fusion decision is made on the anomaly scores of the convolutional neural network and the isolated forest to obtain a comprehensive anomaly score. When the comprehensive anomaly score exceeds the anomaly threshold, an anomaly alarm is triggered, the abnormal device is located, and the anomaly location information is obtained. Based on anomaly location information, a clustering algorithm is used to analyze anomaly patterns and generate anomaly classification results. Combined with power grid state estimation results, fuzzy reasoning is used to classify early warnings and generate power grid early warning signals.

7. The real-time dispatching method for large power grid operation based on model predictive control as described in claim 1, characterized in that: The specific steps for generating standardized scheduling and control instructions are as follows: Based on the optimizer selection results, combined with the power grid state estimation results and power grid load forecast data, a model predictive control optimization problem is constructed, and a mathematical expression of the optimization problem is generated. An optimizer is used to solve the mathematical expression of the optimization problem, and an optimal scheduling scheme is obtained through iterative calculation. The feasibility of the optimized scheduling scheme is verified, a set of scheduling control instructions is generated, and the instructions are sorted and formatted in chronological order to generate standardized scheduling control instructions.

8. The real-time dispatching method for large power grid operation based on model predictive control as described in claim 1, characterized in that: The specific steps for controlling the power grid nodes and generating scheduling execution reports are as follows. Scheduling is performed based on standardized scheduling control commands. By collecting feedback signals from the execution equipment and power grid measurement data, the execution results of the standardized scheduling control commands are verified, and command execution verification results are generated. Based on the command execution verification results and real-time power grid measurement data, the execution effect of standardized dispatch control commands is evaluated, the deviation between the execution effect and the expected target is calculated, and a dispatch execution report is generated.

Images