Self-adaptive control method for dynamic voltage restorer

Abstract

The invention discloses a self-adaptive control method for a dynamic voltage restorer, and relates to the technical field of voltage control, and the method comprises the steps: carrying out the voltage fluctuation analysis according to the historical monitoring voltage, and determining the voltage fluctuation trend; reading a power grid monitoring voltage and a voltage fluctuation interference factor, and determining an interference influence degree; inputting the voltage fluctuation trend, the power grid monitoring voltage and the interference influence degree into a voltage prediction channel, and outputting a first prediction voltage; acquiring a voltage prediction deviation, and outputting a voltage compensation coefficient; and performing fixed-point regulation and control on the dynamic voltage restorer. The technical problems that in an existing dynamic voltage restorer control technology, future actions are generally controlled based on the current state, due to control delay and external factor interference, the voltage compensation error is large, and the voltage compensation accuracy and timeliness are poor are solved, and the purposes of improving the voltage compensation accuracy and timeliness and improving the voltage compensation efficiency are achieved. The compensation error of the DVR is effectively reduced; and the voltage sag is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of voltage control, and in particular to a dynamic voltage restorer adaptive control method. Background Art

[0002] With the increasing complexity of power systems and the continuous growth of power loads, it is crucial to ensure the stable operation of the power grid and the quality of power. Voltage fluctuation is an important aspect of power quality issues, which may be caused by a variety of reasons, such as lightning strikes, short circuit faults, start and stop of large equipment, access to non-linear loads, etc. Voltage fluctuations can bring many adverse effects to power users. For various electrical equipment, their normal operation usually requires a stable voltage.

[0003] Existing dynamic voltage restorer control technology usually controls future actions based on the current state. Due to control delays and interference from external factors, the voltage compensation error is large, resulting in technical problems such as poor voltage compensation accuracy and timeliness. Summary of the invention

[0004] The present application provides an adaptive control method for a dynamic voltage restorer, which is used to control the existing dynamic voltage restorer control technology. Usually, future actions are controlled based on the current state. Due to control delays and interference from external factors, the voltage compensation error is large, resulting in technical problems such as poor voltage compensation accuracy and timeliness.

[0005] In view of the above problems, the present application provides a dynamic voltage restorer adaptive control method, the method comprising:

[0006] A voltage fluctuation analysis is performed based on multiple historical monitoring voltages at multiple historical moments to determine a voltage fluctuation trend, wherein the multiple historical moments include moment K and a predetermined number of adjacent moments before moment K; the grid monitoring voltage and voltage fluctuation interference factor at moment K are read, and the influence degree of the voltage fluctuation interference factor is analyzed according to a predetermined fuzzy analysis strategy to determine the interference influence degree; the voltage fluctuation trend, grid monitoring voltage and interference influence degree are input into a voltage prediction channel to perform voltage prediction at moment K+1, and a first predicted voltage is output, wherein the voltage prediction channel includes a simulation prediction unit and an integrated prediction unit; multiple voltage prediction deviations at multiple historical moments are collected, the multiple voltage prediction deviations are merged according to preset dynamic weights, and a voltage compensation coefficient is output; the first predicted voltage is optimized and corrected according to the voltage compensation coefficient, a voltage prediction value at moment K+1 is output, a compensation voltage parameter is determined based on the voltage prediction value, and a dynamic voltage restorer is fixed-point controlled according to the compensation voltage parameter.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] Perform voltage fluctuation analysis based on multiple historical monitoring voltages to determine the voltage fluctuation trend; read the grid monitoring voltage and voltage fluctuation interference factor at time K to determine the interference impact; input the voltage fluctuation trend, grid monitoring voltage and interference impact into the voltage prediction channel to predict the voltage at time K+1, and output the first predicted voltage; collect multiple voltage prediction deviations, fuse the multiple voltage prediction deviations according to the preset dynamic weights, and output the voltage compensation coefficient; optimize and correct the first predicted voltage according to the voltage compensation coefficient, output the voltage prediction value at time K+1, determine the compensation voltage parameters, and perform fixed-point control on the dynamic voltage restorer according to the compensation voltage parameters. The technical effect of improving the accuracy and timeliness of voltage compensation, effectively reducing the compensation error of DVR and solving voltage sag is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0010] Figure 1 A schematic flow chart of a method for adaptively controlling a dynamic voltage restorer provided in an embodiment of the present application;

[0011] Figure 2 A schematic diagram of a flow chart of determining a voltage fluctuation trend in a method for adaptively controlling a dynamic voltage restorer provided in an embodiment of the present application. DETAILED DESCRIPTION

[0012] The present application provides a dynamic voltage restorer adaptive control method for existing dynamic voltage restorer control technology, which usually controls future actions based on the current state. Due to control delays and interference from external factors, the voltage compensation error is large, resulting in technical problems such as poor voltage compensation accuracy and timeliness.

[0013] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0014] Embodiment 1

[0015] like Figure 1As shown, the present application provides a dynamic voltage restorer adaptive control method, the method comprising:

[0016] Step S100: performing voltage fluctuation analysis based on a plurality of historical monitored voltages at a plurality of historical moments to determine a voltage fluctuation trend, wherein the plurality of historical moments include moment K and a predetermined number of adjacent moments before moment K.

[0017] Specifically, multiple historical monitoring voltages at multiple historical moments are arranged in order from large to small time intervals, thereby generating an ordered historical monitoring voltage sequence. The variant slowfast network is used to extract features, and the global voltage fluctuation features are extracted from the historical monitoring voltage sequence at the first extraction frequency. This extraction frequency is low, and the overall and macroscopic features of the voltage fluctuation can be obtained. At the same time, the local voltage fluctuation features are extracted at the second extraction frequency. Since the second extraction frequency is much higher than the first extraction frequency, the details of the voltage fluctuation and the local changes can be captured. The extracted global voltage fluctuation features and local voltage fluctuation features are input into the voltage trend analyzer. The voltage trend analyzer will fuse and comprehensively analyze these two features. Finally, the output is the voltage fluctuation trend, which can more accurately determine the voltage fluctuation trend and improve the accuracy and timeliness of the prediction.

[0018] Step S200: reading the grid monitoring voltage and voltage fluctuation interference factor at time K, performing influence analysis on the voltage fluctuation interference factor according to a predetermined fuzzy analysis strategy, and determining the interference influence degree.

[0019] Specifically, the grid monitoring voltage and voltage fluctuation interference factor at time K are first read. Time K is a specific time point that needs to be analyzed at present, and multiple historical moments include time K and a predetermined number of adjacent moments before time K. These historical data are used for subsequent voltage fluctuation trend analysis. The influence degree of the voltage fluctuation interference factor is analyzed according to the predetermined fuzzy analysis strategy to determine the interference influence degree. The fuzzy analysis strategy is an analysis method that comprehensively considers multiple related factors. It can more comprehensively and meticulously evaluate the impact of interference factors on voltage fluctuations. This strategy will consider the following factors, the type of interference factor: load change value, distributed power supply access characteristics, ambient temperature, ambient humidity and electromagnetic interference characteristics, etc. By combining these factors, the strategy can give a quantitative indicator reflecting the degree of influence of interference factors on voltage fluctuations, that is, the interference influence degree. The role of determining the interference influence degree is that it can help the operator of the power system or related control equipment to better understand and deal with voltage fluctuations. According to the influence degree of different interference factors, corresponding control measures or adjustment strategies are taken to reduce the impact of interference on voltage stability.

[0020] Step S300: inputting the voltage fluctuation trend, the grid monitoring voltage and the interference influence into a voltage prediction channel to perform voltage prediction at time K+1, and outputting a first predicted voltage, wherein the voltage prediction channel includes a simulation prediction unit and an integrated prediction unit.

[0021] Specifically, the three key factors, namely the previously determined voltage fluctuation trend, the grid monitoring voltage at time K, and the calculated interference impact, are input into the voltage prediction channel. This voltage prediction channel consists of a simulation prediction unit and an integrated prediction unit. The simulation prediction unit simulates the operation of the grid based on physical models and mathematical algorithms, and predicts the voltage change trend based on the input information. The integrated prediction unit uses a decision tree algorithm to fuse and analyze the input data to improve the accuracy and reliability of the prediction. Through the collaborative work of these two units, the voltage at time K+1 is predicted, and the first predicted voltage is finally output. It can simulate and predict the grid voltage more accurately, improving the accuracy and timeliness of the prediction.

[0022] Step S400: collecting a plurality of voltage prediction deviations at a plurality of historical moments, fusing the plurality of voltage prediction deviations according to a preset dynamic weight, and outputting a voltage compensation coefficient.

[0023] Specifically, data collection is first performed to obtain voltage prediction deviations at multiple historical moments. These deviations are obtained by calculating the difference between the voltage prediction value and the actual monitoring value at each historical moment. Then, these multiple voltage prediction deviations are fused using a preset dynamic weight (determined by the coefficient of variation method). The coefficient of variation method determines the weight of each deviation based on factors such as the time interval between the historical moment and the current moment. The shorter the time interval, the greater the weight; the longer the time interval, the smaller the weight. During the fusion process, each voltage prediction deviation is multiplied by its corresponding weight, and then these weighted deviations are summed or otherwise calculated in a comprehensive manner. Ultimately, the output result is a voltage compensation coefficient. This coefficient can be used to correct and optimize subsequent voltage predictions to improve the accuracy and reliability of voltage predictions.

[0024] Step S500: optimizing and correcting the first predicted voltage according to the voltage compensation coefficient, outputting the voltage prediction value at time K+1, determining compensation voltage parameters based on the voltage prediction value, and performing fixed-point control on the dynamic voltage restorer according to the compensation voltage parameters.

[0025] Specifically, the voltage compensation coefficient obtained previously is used to adjust and correct the first predicted voltage outputted previously. Through this optimization and correction, the prediction result is made more accurate and reliable, and the voltage prediction value at time K+1 is finally outputted. Based on this voltage prediction value, the specific compensation voltage parameters are further determined. These parameters clarify the specific information such as the voltage size and direction required to compensate for the voltage to reach the ideal state. Then, based on these compensation voltage parameters, the dynamic voltage restorer is targeted and fixed-point controlled, and the dynamic voltage restorer can respond quickly and adjust the voltage to stabilize it within a suitable range.

[0026] In one possible implementation, Figure 2 As shown, step S100 also includes:

[0027] Step S110: Arrange a plurality of historical monitoring voltages at a plurality of historical moments according to time intervals from large to small to generate a historical monitoring voltage sequence.

[0028] Step S120: Based on the variant slowfast network, feature extraction is performed on the historical monitored voltage sequence according to the first extraction frequency and the second extraction frequency to obtain global voltage fluctuation features and local voltage fluctuation features, wherein the first extraction frequency is much smaller than the second extraction frequency.

[0029] Step S130: A voltage trend analyzer is used to perform a fusion analysis on the global voltage fluctuation characteristics and the local voltage fluctuation characteristics, and output the voltage fluctuation trend.

[0030] Specifically, we first need to identify multiple historical moments and the corresponding voltage values ​​monitored at these moments. The time interval refers to the length of time between two adjacent historical moments. For example, we have the following historical moments and their corresponding monitored voltages: moment A, moment B, moment C, moment D, moment E, and the corresponding voltages are V1, V2, V3, V4, and V5, respectively. Assuming that moment A is the earliest and moment E is the latest, the time intervals between them decrease in sequence. Then, according to the arrangement of time intervals from large to small, the earliest moment A and its corresponding voltage V1 are first placed at the front of the sequence, followed by moment B and voltage V2 with the second shortest time interval, and so on, until the latest moment E and voltage V5 are placed at the end of the sequence. The purpose of this arrangement is to enable subsequent data analysis to more clearly reflect the law of voltage changes over time. Through this orderly arrangement, it is easier to discover long-term voltage trends and possible periodic or staged changes.

[0031] The variant slowfast network is used to process the historical monitoring voltage series generated previously. The first extraction frequency is low, and data is extracted from the historical monitoring voltage series at this frequency for feature extraction. Due to the large extraction interval, the extracted features can reflect the overall change pattern and trend of the voltage over a long time span, which is the global voltage fluctuation feature. The second extraction frequency is much higher, which means that the interval of the extracted data is very small. Based on such a high-frequency extraction to extract features, the obtained features can capture the subtle changes in voltage in a short time and the local rapid fluctuations, which is the local voltage fluctuation feature. It is worth noting that the first extraction frequency is much smaller than the second extraction frequency, which makes the global and local features have obvious differences in time scale and the information reflected. Global features focus on the long-term, macroscopic changes in voltage, while local features focus on short-term, microscopic fluctuation details.

[0032] Build a machine learning model as a voltage trend analyzer. In order to build this analyzer, it is necessary to collect sample data. These sample data include the sample global feature set, which is a series of feature sets that reflect the long-term voltage change pattern; the sample local feature set, which is a feature set that reflects the short-term subtle fluctuations of the voltage; and the sample overall voltage trend set, which is the known correct overall voltage change trend. With abundant sample data, the random decision forest is trained. During the training process, the model will continue to learn and adjust, trying to understand the complex relationship between various combinations of the sample global feature set and the sample local feature set and the sample overall voltage trend set. Through a large amount of data input and repeated optimization and adjustment, the model gradually masters how to accurately predict the overall trend of the voltage based on the input features. When the model training is completed and the voltage trend analyzer is successfully built, the global voltage fluctuation features and local voltage fluctuation features obtained are input into the analyzer. The analyzer will conduct in-depth fusion and comprehensive analysis of these two sets of features, and analyze the long-term trend reflected by the global features and the short-term fluctuation details shown by the local features. Through the internally trained complex decision logic and calculation process, these features are fused and weighed. Finally, after precise calculation and analysis, a result is output that can accurately describe the overall direction and characteristics of voltage fluctuations, namely the voltage fluctuation trend.

[0033] In a possible implementation, step S200 further includes:

[0034] Step S210: reading the voltage fluctuation interference factor at time K, wherein the voltage fluctuation interference factor includes load change value, distributed power supply access characteristics, ambient temperature, ambient humidity and electromagnetic interference characteristics.

[0035] Step S220: taking the target power grid topology as a constraint, retrieving and acquiring a sample fuzzy data set, training a fuzzy analysis strategy based on the sample fuzzy data set, and obtaining a predetermined fuzzy analysis strategy; wherein the fuzzy analysis strategy includes fuzzy reasoning logic, fuzzy rules, and fuzzy membership functions.

[0036] Step S230: performing an influence analysis on the voltage fluctuation interference factor according to the predetermined fuzzy analysis strategy, and outputting an interference influence degree.

[0037] Specifically, we focus on obtaining the voltage fluctuation interference factor at time K. At a specific time point K, we actively collect and read a series of key factors that affect voltage stability. Load change value is one of the important interference factors, which specifically refers to the increase or decrease in the power load in the power grid at time K compared with the previous time. If the load suddenly increases significantly, the current demand will rise sharply, which will cause the voltage to drop; conversely, if the load suddenly decreases significantly, the voltage will rise. The access characteristics of distributed power sources should not be ignored. This covers detailed information such as the number of accesses, access locations, output power size and change rate of distributed power sources (such as solar panels, small wind turbines, etc.) at time K, as well as the way and stability of their connection to the power grid. The intermittent and uncertain output characteristics of distributed power sources may cause fluctuations in the power grid voltage. Ambient temperature is another influencing factor. The ambient temperature value at time K will affect the resistance, inductance and other parameters of electrical equipment, thereby indirectly affecting the current and voltage distribution in the power grid. Higher temperatures lead to increased resistance, increased power loss, and thus affect the voltage. Ambient humidity can also play a role. Too high or too low humidity may affect the insulation performance and heat dissipation effect of electrical equipment at time K, which in turn has a certain impact on voltage stability. The performance of electromagnetic interference characteristics at time K is also crucial. It includes information such as the strength, frequency, and change pattern of the surrounding electromagnetic field. Strong or complex electromagnetic interference may interfere with signal transmission and normal operation of equipment in the power grid, causing voltage fluctuations. By reading the specific values ​​and states of these interference factors, basic data is provided for subsequent voltage fluctuation analysis and countermeasures.

[0038] The target power grid topology plays an important role in constraints, describing the connection mode, location and relationship of each component (such as substation, line, transformer, etc.) in the power grid. This structure determines the flow path and electrical characteristics of electric energy. Based on this clear topology constraint, we start to retrieve and obtain sample fuzzy data sets. This data set contains a large amount of relevant data of the power grid under different circumstances, especially information related to voltage fluctuation interference factors and their impact. Then, we use these rich sample data to train the fuzzy analysis strategy. Fuzzy reasoning logic is a way of thinking that deals with uncertainty and fuzzy information. Unlike traditional precise logic, which only has "yes" or "no" judgments, it can handle fuzzy concepts such as "maybe" and "probably", helping to make reasonable inferences in complex and uncertain situations. Fuzzy rules are a series of criteria based on experience and professional knowledge. For example, when the load change value exceeds a certain range and the ambient temperature is high, it may be concluded that the voltage fluctuation is greatly affected. The fuzzy membership function is used to define the degree to which the input data belongs to different degrees or categories. For example, the load change value is divided into categories such as "small", "medium" and "large", and each specific load change value has a corresponding degree value belonging to different categories. By training and adjusting these elements, the fuzzy analysis strategy can better adapt to the characteristics and actual conditions of the target power grid, and finally obtain a predetermined fuzzy analysis strategy that can accurately analyze the influence of voltage fluctuation interference factors.

[0039] The voltage fluctuation interference factors, such as load change value, distributed power access characteristics, ambient temperature, ambient humidity and electromagnetic interference characteristics, are input into the predetermined fuzzy analysis strategy. The fuzzy reasoning logic in the strategy will analyze and reason these input interference factors according to the set rules and conditions. Fuzzy rules define the corresponding influence judgments within different interference factor combinations or value ranges. The fuzzy membership function is used to determine the degree of belonging of the value of each interference factor in different fuzzy sets, for example, the degree of belonging of a certain load change value in fuzzy sets such as "large", "medium change" and "small". By comprehensively applying these fuzzy analysis methods and tools, the impact of each interference factor on voltage fluctuation is evaluated and measured. Ultimately, the output interference impact is a quantitative result, which indicates the relative importance or influence of each interference factor on voltage fluctuation. Such an output result helps to consider and deal with various interference factors more targetedly in subsequent voltage prediction and regulation, thereby improving the effect and accuracy of voltage management.

[0040] In a possible implementation, step S210 further includes:

[0041] Step S211: taking the K+1 time as a constraint, searching the historical power grid load records to obtain a plurality of sample load data, wherein the sample load data includes a sample starting load and a sample load mean.

[0042] Step S212: Calculate the load deviation between the sample load mean and the sample starting load, determine multiple load deviations, cluster the multiple load deviations according to a predetermined step size, and count the number of deviations in the clustering result as the clustering frequency to obtain multiple clustering frequencies.

[0043] Step S213: retain the clustering results corresponding to the clustering frequencies that meet the predetermined frequency threshold, and perform mean calculation on the remaining clustering results to obtain the standard starting load and the standard load mean.

[0044] Step S214: reading the starting grid load and the grid load mean at time K, calculating a first load deviation between the standard starting load and the starting grid load, compensating the standard load mean according to the first load deviation, and determining a predicted load mean.

[0045] Step S215: Calculate the deviation between the predicted load mean and the grid load mean to obtain the load change value.

[0046] Specifically, in this step, the K+1 moment is set as a key time reference point, and this moment is used as a constraint to search and filter in the database of historical power grid load records. The historical power grid load record is a data set containing a large number of power grid load conditions at different times in the past. By setting the K+1 moment constraint, the system can extract the relevant specific parts from this huge data set. Each of the multiple sample load data obtained contains two key parameters: the sample starting load and the sample load mean. The sample starting load refers to the load borne by the power grid at the beginning of a specific time period, which reflects the initial level of the power grid load at the beginning of the time period. The sample load mean is the average calculation result of the power grid load in this specific time period. It comprehensively considers the changes in the load during the time period and provides a measure of the average load level during the time period.

[0047] For each sample load data obtained, the difference between the sample load mean and the sample starting load is calculated to determine the load deviations one by one. These load deviations reflect the magnitude and direction of the load change of the sample in a specific time period. In order to more effectively analyze and understand these load deviation data, they are clustered according to the pre-set step size and the load deviation values ​​are classified. In the clustering process, load deviations with similar values ​​are classified into the same category. The advantage of this is that the data can be grouped according to a certain rule, making the distribution characteristics of the data clearer and more visible. At the same time, the number of occurrences of the load deviation contained in each clustering result is counted and defined as the clustering frequency. By counting the clustering frequency, the frequency of occurrence of different load deviation ranges in the entire data set can be understood. From the perspective of data screening, it can help us remove some abnormal and isolated load deviation values, because these values ​​may be caused by measurement errors or special circumstances and are not representative. By clustering and counting the frequency, more attention is paid to those load deviation ranges with higher frequency, that is, the mainstream trend of the data, thereby improving the accuracy and reliability of the data.

[0048] A predetermined frequency threshold is set, which is used to determine which clustering results are significant and representative. The multiple clustering frequencies obtained are screened, and those clustering results whose clustering frequencies meet the predetermined frequency threshold are retained. These retained clustering results usually reflect the common and important patterns of load deviation. For the remaining clustering results that do not meet the predetermined frequency threshold, their load deviation values ​​are averaged. Through this mean calculation, the standard starting load and the standard load mean are obtained. The standard starting load can be understood as a typical and representative starting load value after comprehensively considering various load conditions. The standard load mean is the average load level corresponding to this standard starting load. The purpose of this is to highlight common load patterns, and by averaging uncommon patterns, more general and representative standard values ​​are obtained, providing a more reliable basis for subsequent analysis and prediction.

[0049] Read the starting grid load and grid load mean at time K. Time K is a specific moment that is currently being paid attention to or needs to be analyzed. Calculate the difference between the standard starting load obtained before and the starting grid load at the current time K. This difference is called the first load deviation. Compensate and adjust the standard load mean according to this first load deviation. The idea of ​​compensation is that if the first load deviation is large, it means that the starting load at the current time K is significantly different from the usual standard starting load, so the adjustment range of the standard load mean will also be correspondingly large; conversely, if the first load deviation is small, the adjustment range will be relatively small. Through this compensation adjustment based on the first load deviation, the predicted load mean is finally determined. This predicted load mean comprehensively considers the standard load mean obtained from historical data and the difference between the starting load at the current time K and the standard starting load, aiming to provide a load mean prediction value that is more in line with the current actual situation.

[0050] The predicted load mean calculated by the previous steps has been obtained, and the actual grid load mean at time K has been obtained. Then, through simple mathematical operations, the predicted load mean is subtracted from the grid load mean, and the difference is the load change value. This load change value reflects the degree of difference between the predicted load mean and the actual grid load mean observed at time K. If the load change value is a positive number, it means that the predicted load mean is higher than the actual grid load mean, and the load has an upward trend; if the load change value is a negative number, it means that the predicted load mean is lower than the actual grid load mean, and the load has a downward trend; and if the load change value is close to zero, it means that the predicted value is close to the actual value, and the load is relatively stable. By calculating the deviation between the predicted load mean and the grid load mean, we can intuitively understand the degree of difference between the predicted result and the actual load, thereby accurately evaluating the accuracy of the load prediction.

[0051] In a possible implementation, step S300 further includes:

[0052] Step S310: Taking the target power grid topology and time interval as constraints, a positive sample search is performed based on the power Internet to obtain a search data set, wherein the search data set includes a sample voltage fluctuation trend, a sample power grid monitoring voltage, a sample interference influence degree and a sample predicted voltage.

[0053] Step S320: Divide the retrieval data set into a first data set and a second data set according to a predetermined ratio, perform simulation based on the target power grid topology structure based on the digital twin, construct a power grid twin model, and use the first data set to train and optimize the power grid twin model to construct a simulation prediction unit.

[0054] Step S330: using the second data set, performing supervised training on the predetermined prediction operator, integrating the convergent prediction operators based on the ensemble learning principle, and generating an integrated prediction unit.

[0055] Step S340: generating the voltage prediction channel based on the simulation prediction unit and the integrated prediction unit.

[0056] Specifically, this step aims to use the topological structure of the target power grid and the specific time interval as the limiting conditions, and use the power Internet to search and obtain positive samples, so as to obtain a retrieval data set. The target power grid topological structure clarifies the connection mode and layout between the various components in the power grid (such as power generation equipment, transmission lines, substations, etc.). The time interval specifies the time period in which the data is retrieved. The retrieval data set contains multiple aspects of information: sample voltage fluctuation trend, which presents the trend of voltage changes over a period of time, such as whether there is periodic rise and fall, whether there is sudden fluctuation, etc.; sample power grid monitoring voltage, which is the specific value of the power grid voltage at each moment measured in practice; sample interference influence, which is used to measure the impact of various interference factors (such as changes in power grid load, external environment, etc.) on the power grid voltage; sample predicted voltage, which is the voltage prediction value obtained by a certain prediction method or model. Through such a retrieval process, various data samples related to the target power grid in a specific time period can be collected, and these sample data are of great significance for subsequent power grid analysis, modeling, prediction and optimization. They can help understand the operating status of the power grid, the patterns of voltage changes, and the impact of interference factors, thereby providing support for more accurate voltage forecasting, power grid optimization and stable operation.

[0057] According to the preset ratio, the retrieved data set is divided into two parts, named as the first data set and the second data set. The preset ratio is determined according to specific needs and experience, for example, it can be divided into 70% and 30%. Based on the digital twin technology, that is, using digital means to create a virtual model that is highly similar to the actual target power grid in structure and characteristics, the simulation is performed according to the topological structure of the target power grid, including the connection mode of each component, equipment parameters and other information. In the simulation process, a power grid twin model is constructed. This model can simulate the operating state and various dynamic changes of the power grid. Next, the constructed power grid twin model is trained and optimized using the first data set. By inputting the sample data in the first data set into the model, the model is allowed to learn and adapt to the operating laws and characteristics of the actual power grid. After continuous training and adjustment of the parameters of the model, the model can more accurately predict the behavior and state of the power grid, and finally construct a unit that can perform simulation prediction, called a simulation prediction unit. For example, if the preset ratio is 70% and 30%, the retrieved data set is divided into the first data set (accounting for 70%) and the second data set (accounting for 30%). Then, according to the topological structure information of the target power grid, including 5 substations and 10 transmission lines, a simulation was performed to build a power grid twin model that can simulate voltage changes, power flows, etc. Then, the model was trained using a large amount of voltage fluctuations, monitoring voltage and other data contained in the first data set. For example, after 1,000 iterations of training, the prediction error of the model reached an acceptable range, thus successfully building a simulation prediction unit.

[0058] The second data set divided previously is used to conduct supervised training on the predetermined prediction operators. The prediction operator here refers to a tool for prediction built according to a specific algorithm or model. When the prediction operator reaches a convergence state after training, these converged prediction operators are integrated based on the principle of ensemble learning. The basic idea of ​​ensemble learning is to combine multiple prediction operators to form a more powerful integrated prediction unit. In the integration process, a variety of strategies can be adopted, such as averaging, weighted averaging or voting. By combining the advantages and characteristics of multiple prediction operators, the final generated integrated prediction unit can make predictions more comprehensively and accurately. For example, suppose there are several prediction operators based on different algorithms or models, such as decision trees, support vector machines, neural networks, etc. These prediction operators are trained separately using the second data set. After they all converge, the corresponding weights are determined according to the performance of each prediction operator on the validation set or other evaluation indicators. Then, the prediction results of these prediction operators are combined according to the weights by weighted averaging to obtain the final prediction result of the integrated prediction unit. The integrated prediction unit generated in this way can reduce the possible deviation and uncertainty of a single prediction operator and improve the overall prediction performance and generalization ability.

[0059] The voltage prediction channel is created by integrating the simulation prediction unit and the integrated prediction unit constructed previously, and the output results and characteristics of the simulation prediction unit and the integrated prediction unit are clarified. The simulation prediction unit is obtained based on the digital twin model and the first data set training optimization, and it can provide voltage prediction results based on simulation. The integrated prediction unit is generated by integrating multiple converged prediction operators and has the ability to integrate the advantages of multiple prediction methods. Then, the output results of the two units are fused. A variety of fusion methods can be used, such as weighted averaging, that is, the prediction results of the simulation prediction unit and the integrated prediction unit are weighted and summed according to certain weights to obtain the final voltage prediction value; or a more complex fusion strategy can be used to dynamically adjust the fusion method taking into account the prediction accuracy and reliability of different units in different scenarios. Finally, through the above fusion and processing process, a channel that can perform accurate and reliable voltage prediction is successfully generated.

[0060] In a possible implementation, step S330 further includes:

[0061] Step S331: configuring a predetermined prediction operator, wherein the predetermined prediction operator at least includes a BP neural network, a support vector machine and a random decision forest.

[0062] Step S332: Divide the second data set into N equal parts, select N times with replacement, construct the first data set, iterate and select N times to obtain N data sets, where N is the number of prediction operators.

[0063] Step S333: Use the N data sets to perform supervised training and cross-validation on the predetermined prediction operator to obtain N convergence prediction modules, and based on the principle of integrated learning, construct the integrated prediction unit according to the N convergence prediction modules, wherein the output of the integrated prediction unit is the mean of the output results of the N convergence prediction modules.

[0064] Specifically, the required predetermined prediction operators are configured for subsequent prediction tasks. BP neural network is a widely used machine learning algorithm. It learns the complex relationship between input and output by adjusting the connection weights between neurons. BP neural network has strong learning and generalization capabilities and can handle nonlinear problems. Support vector machine is an algorithm based on statistical learning theory. It classifies or regresses data by finding an optimal hyperplane. Support vector machine performs well in processing small samples and high-dimensional data and has good robustness. Random decision forest is an integrated learning algorithm. It consists of multiple decision trees. By randomly selecting features and samples to construct multiple decision trees and combining their prediction results, random decision forest can effectively reduce the risk of overfitting and improve the accuracy and stability of prediction. The reason for selecting at least these three prediction operators is that they have different characteristics and advantages and can analyze and predict data from multiple angles. By combining their results, the reliability and accuracy of the overall prediction can be improved. For example, when predicting power grid load, BP neural network may be able to capture complex nonlinear patterns; support vector machine may have better adaptability to certain types of data distribution; and random decision forest can provide more robust and diverse prediction results.

[0065] The second data set is divided into N equal parts, where N is the number of prediction operators (the total number of prediction operators such as BP neural network, support vector machine and random decision forest configured in step S331). Then, selection with replacement is performed N times, and selection with replacement means that after each selection, the selected data is put back into the original data set, so that the next time the selection is made, there is still a chance to select the previously selected data. Through this iterative selection method, a total of N data sets are obtained. The purpose of doing this is that each data set can be used to train and test different prediction operators separately in the future. Each prediction operator can obtain an independent data set for learning and verification, so that the performance of different prediction operators on different data subsets can be compared, their prediction effects can be comprehensively evaluated, and the accuracy and reliability of the overall prediction can be improved.

[0066] There are N data sets, each of which will be used for supervised training and cross-validation of the corresponding predetermined prediction operator. For each predetermined prediction operator (such as BP neural network, support vector machine and random decision forest), the corresponding data set is used for training. During the training process, by continuously adjusting the parameters of the model, the error between the predicted result and the actual result is gradually reduced until the model reaches a convergence state, thereby obtaining N converged prediction modules. Next, based on the principle of integrated learning, these N converged prediction modules are integrated to construct an integrated prediction unit. In this integrated prediction unit, the final output result is the average value of the output results of these N converged prediction modules. When a prediction is required, the input data is input into these N converged prediction modules respectively to obtain N prediction results. Then, these N prediction results are added and divided by N. The average value obtained is the final output of the integrated prediction unit, which combines the advantages of multiple prediction modules, reduces the deviation and uncertainty of a single prediction module, and improves the accuracy and stability of the overall prediction.

[0067] In a possible implementation, step S300 further includes:

[0068] Step S350: Obtain a first prediction convergence degree of the simulation prediction unit and a second prediction convergence degree of the integrated prediction unit.

[0069] Step S360: constructing a weighted fusion layer based on the first prediction convergence degree and the second prediction convergence degree, and generating the voltage prediction channel by combining the simulation prediction unit and the integrated prediction unit.

[0070] Specifically, two key indicators are obtained, the first prediction convergence of the simulation prediction unit and the second prediction convergence of the integrated prediction unit. The prediction error of the simulation prediction unit on the training set and the validation set is monitored in a series of training iterations. As the training progresses, if the prediction error gradually decreases and tends to be stable, it means that the model is converging. Calculate the error indicators of different training stages, such as mean square error (MSE), mean absolute error (MAE), etc., and observe the changing trends of these error indicators. If the error no longer decreases significantly after a certain number of training rounds, or the error on the validation set can fluctuate within a small range, it can be considered that the simulation prediction unit has reached a certain degree of convergence. For the second prediction convergence of the integrated prediction unit, the acquisition method is similar, but the performance of each sub-model in the integrated model needs to be considered. Evaluate the error changes of each sub-model during the training and validation process, as well as the consistency and complementarity between the sub-models. If the overall prediction error of the sub-models in the integrated model is stable and performs well after training, and the sub-models can cooperate with each other to improve the overall prediction performance, then it can be considered that the integrated prediction unit has reached the corresponding degree of convergence. Finally, through the above evaluation and calculation, the values ​​of the first prediction convergence degree and the second prediction convergence degree that can quantify the convergence conditions of the simulation prediction unit and the integrated prediction unit are obtained.

[0071] The weighted fusion layer is constructed based on the first prediction convergence of the simulation prediction unit and the second prediction convergence of the integrated prediction unit obtained previously. The numerical values ​​and relative relationship of the first prediction convergence and the second prediction convergence are analyzed. If the first prediction convergence is high, it means that the prediction result of the simulation prediction unit is more stable and reliable. In this case, the weight assigned to the simulation prediction unit in the weighted fusion layer will be relatively large; conversely, if the second prediction convergence is high, the integrated prediction unit will be given a higher weight when fused. The specific weight allocation can be achieved by simply setting the weight to be proportional to the convergence. After determining the weight, the weighted fusion layer multiplies the output result of the simulation prediction unit by its corresponding weight and the output result of the integrated prediction unit by its corresponding weight. Then, the two weighted results are added together to obtain the final fused prediction value. This fused prediction value constitutes the output of the voltage prediction channel. In this way, the respective advantages of the simulation prediction unit and the integrated prediction unit are fully utilized, and their prediction capabilities and stability are comprehensively considered to obtain more accurate and reliable voltage prediction results.

[0072] In a possible implementation, step S400 further includes:

[0073] Step S410: Calculate the time intervals between multiple historical moments and moment K to obtain multiple time intervals.

[0074] Step S420: Based on the coefficient of variation method, a preset dynamic weight is configured according to the multiple time intervals, wherein the weight is negatively correlated with the time interval.

[0075] Specifically, first, it is necessary to identify the specific time points of multiple historical moments. These historical moments may be recorded according to a certain pattern, such as every hour, day, or week. Then, each historical moment is compared with the K moment. The K moment is usually the time point that is currently being paid attention to or used as a benchmark. The time interval can be calculated in a unified time unit, such as seconds, minutes, hours, etc. Through a simple subtraction operation, the time value of each historical moment is subtracted from the time value of the K moment to obtain the corresponding time interval.

[0076] The coefficient of variation method is used to configure a preset dynamic weight for each historical moment based on multiple time intervals calculated previously. The coefficient of variation method is a statistical method for measuring the degree of data dispersion. Here, it is used to determine the weight corresponding to the time interval. Since the weight is negatively correlated with the time interval, it means that the longer the time interval, the lower the credibility of the data, and the smaller the weight assigned. The specific calculation process is as follows: First, the mean and standard deviation of multiple time intervals are calculated, and then the weight corresponding to each time interval is calculated by the formula (weight = 1 / (time interval × standard deviation + mean)). Such weight configuration reflects the impact of time interval on data reliability. The more recent historical moments are assigned higher weights because their data are more valuable due to the short time interval; while the more distant historical moments are assigned smaller weights because their data are relatively less reliable due to the long time interval. Combining these two steps, historical data can be used more effectively for analysis and prediction, improving the accuracy and reliability of the results, while adapting to different time patterns and data characteristics.

[0077] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of this specification are described. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0078] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

[0079] This specification and the drawings are merely exemplary illustrations of the present application and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, a person skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these modifications and variations.

Claims

1. A dynamic voltage restorer adaptive control method, characterized in that: include: Perform voltage fluctuation analysis based on multiple historical monitored voltages at multiple historical moments to determine the voltage fluctuation trend, wherein the multiple historical moments include moment K and a predetermined number of adjacent moments before moment K; Read the power grid monitoring voltage and voltage fluctuation interference factor at time K, perform influence analysis on the voltage fluctuation interference factor according to a predetermined fuzzy analysis strategy, and determine the interference influence degree; The voltage fluctuation trend, the grid monitoring voltage and the interference influence are input into a voltage prediction channel to perform voltage prediction at time K+1, and a first predicted voltage is output, wherein the voltage prediction channel includes a simulation prediction unit and an integrated prediction unit; Collect multiple voltage prediction deviations at multiple historical moments, fuse the multiple voltage prediction deviations according to preset dynamic weights, and output a voltage compensation coefficient; The first predicted voltage is optimized and corrected according to the voltage compensation coefficient, the voltage prediction value at time K+1 is output, the compensation voltage parameter is determined based on the voltage prediction value, and the dynamic voltage restorer is fixed-point controlled according to the compensation voltage parameter.

2. The method for adaptively controlling a dynamic voltage restorer according to claim 1, characterized in that: Determine voltage fluctuation trends, including: Arrange multiple historical monitoring voltages at multiple historical moments according to time intervals from large to small to generate a historical monitoring voltage sequence; Based on the variant slowfast network, feature extraction is performed on the historical monitoring voltage sequence according to the first extraction frequency and the second extraction frequency respectively, so as to obtain global voltage fluctuation feature and local voltage fluctuation feature, wherein the first extraction frequency is much smaller than the second extraction frequency; The global voltage fluctuation characteristics and the local voltage fluctuation characteristics are integrated and analyzed by a voltage trend analyzer to output the voltage fluctuation trend.

3. The method for adaptively controlling a dynamic voltage restorer according to claim 1, characterized in that: Determine the impact of interference, including: Read the voltage fluctuation interference factor at time K, where the voltage fluctuation interference factor includes load change value, distributed power supply access characteristics, ambient temperature, ambient humidity and electromagnetic interference characteristics; Taking the target power grid topology as a constraint, a sample fuzzy data set is retrieved and obtained, and a fuzzy analysis strategy is trained based on the sample fuzzy data set to obtain a predetermined fuzzy analysis strategy; wherein the fuzzy analysis strategy includes fuzzy reasoning logic, fuzzy rules and fuzzy membership function; The voltage fluctuation interference factor is subjected to influence analysis according to the predetermined fuzzy analysis strategy, and the interference influence degree is output.

4. The method for adaptively controlling a dynamic voltage restorer according to claim 3, characterized in that: Get load change values, including: Taking the K+1 time as a constraint, retrieve the historical power grid load records to obtain multiple sample load data, where the sample load data includes the sample starting load and the sample load mean; Calculate the load deviation between the sample load mean and the sample starting load, determine multiple load deviations, cluster the multiple load deviations according to a predetermined step length, and count the number of deviations in the clustering result as the clustering frequency to obtain multiple clustering frequencies; The clustering results corresponding to the clustering frequencies that meet the predetermined frequency threshold are retained, and the mean of the remaining clustering results is calculated to obtain the standard starting load and the standard load mean; Reading the starting grid load and the grid load mean at time K, calculating a first load deviation between the standard starting load and the starting grid load, compensating the standard load mean according to the first load deviation, and determining a predicted load mean; The predicted load mean and the power grid load mean are calculated to obtain the load change value.

5. The method for adaptively controlling a dynamic voltage restorer according to claim 1, characterized in that: Construct a voltage prediction channel, including: Taking the target power grid topology and time interval as constraints, a positive sample search is performed based on the power Internet to obtain a search data set, where the search data set includes sample voltage fluctuation trend, sample power grid monitoring voltage, sample interference impact and sample predicted voltage; Dividing the search data set into a first data set and a second data set according to a predetermined ratio, based on the digital twin, simulating according to the target power grid topology structure, constructing a power grid twin model, and using the first data set to train and optimize the power grid twin model to construct a simulation prediction unit; Using the second data set, supervised training is performed on the predetermined prediction operator, and the convergent prediction operators are integrated based on the ensemble learning principle to generate an integrated prediction unit; The voltage prediction channel is generated based on the simulation prediction unit and the integrated prediction unit.

6. A dynamic voltage restorer adaptive control method according to claim 5, characterized in that: Generate an integrated prediction unit, including: Configuring a predetermined prediction operator, wherein the predetermined prediction operator includes at least a BP neural network, a support vector machine, and a random decision forest; Divide the second data set into N equal parts, select N parts with replacement, construct the first data set, iterate and select N times, and obtain N data sets, where N is the number of prediction operators; The predetermined prediction operator is supervisedly trained and cross-validated using the N data sets to obtain N convergence prediction modules. Based on the principle of integrated learning, the integrated prediction unit is constructed according to the N convergence prediction modules, wherein the output of the integrated prediction unit is the mean of the output results of the N convergence prediction modules.

7. A dynamic voltage restorer adaptive control method according to claim 5, characterized in that: Generating the voltage prediction channel includes: Acquire a first prediction convergence degree of the simulation prediction unit and a second prediction convergence degree of the integrated prediction unit; A weighted fusion layer is constructed based on the first prediction convergence degree and the second prediction convergence degree, and the voltage prediction channel is generated by combining the simulation prediction unit and the integrated prediction unit.

8. The method for adaptively controlling a dynamic voltage restorer according to claim 1, characterized in that: Configure preset dynamic weights, including: Calculate the time intervals between multiple historical moments and moment K to obtain multiple time intervals; Based on the coefficient of variation method, preset dynamic weights are configured according to the multiple time intervals, wherein the weights are negatively correlated with the time intervals.

Images