Method for information management and big data analysis of power engineering project

Abstract

The application provides a power engineering project informatization management and big data analysis method, relates to the field of information technology, and comprises the following steps: obtaining load time series data of power distribution transformer project management, preprocessing the data, including data cleaning, abnormal value processing and data normalization, and obtaining preprocessed load time series data; according to load transfer characteristics, adopting a clustering algorithm to cluster transformers, dividing transformers with similar load transfer modes into the same category, obtaining transformer clustering results, and inputting the transformer clustering results into a project management system; according to the transformer clustering results in the project management system, combining the change of time, judging the dynamic change of the power distribution network topology structure, and if the change degree of the transformer clustering results at a moment is higher than a threshold value, the topology structure changes at the moment. The application provides support for real-time management of the power distribution network project, and improves the reliability and efficiency of the power distribution network operation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and particularly relates to a method for information management and big data analysis of power engineering projects. BACKGROUND

[0002] In power engineering project management, accurately grasping the dynamic changes of the distribution network topology structure is an important and challenging task. The dynamic changes of the distribution network topology structure bring great challenges to load data analysis. The load time series data contains rich topology change information, but how to accurately extract the topology change features from massive data is a key technical problem. The load mutation phenomenon is closely related to the topology change, but the load mutation may also be caused by other factors, and how to distinguish the mutation caused by topology change is a big challenge. Topology changes will lead to changes in the correlation of loads between different transformers, but it is difficult to accurately describe the change law of correlation in a complex environment. Frequent topology switching operations and fault recovery processes further increase the dynamics and uncertainty of topology changes, bringing more difficulties to feature extraction and topology inference. The massive load data of multiple sources and heterogeneous data bring great pressure to data processing and calculation, and efficient big data analysis methods need to be explored. How to accurately extract the topology change features from complex load data and realize real-time inference of the dynamic topology of the distribution network is a key technical problem to be solved, which is of great significance to power distribution automation and intelligent operation and maintenance. SUMMARY

[0003] To solve the above technical problems, the present application provides a method for information management and big data analysis of power engineering projects, mainly including:

[0004] The load time series data of the power distribution transformer project management is acquired, and the data is preprocessed to obtain preprocessed load time series data; according to the preprocessed load time series data, load mutation characteristics and load correlation change characteristics capable of reflecting load transfer are extracted as load transfer characteristics; according to the load transfer characteristics, a clustering algorithm is used to cluster the transformers to obtain transformer clustering results, and the transformer clustering results are input into the project management system; according to the transformer clustering results in the project management system, the dynamic change of the power distribution network topology is judged in combination with the change of time; if the topology structure changes, the corresponding topology structure change including the load transfer path is inferred according to the dynamic topology structure of the power distribution network in the project management system; according to the historical inferred topology structure change, a decision tree algorithm is used to take the load transfer characteristics and the clustering results of the transformers as inputs to train a topology change inference model, and the current power distribution network topology is inferred according to the real-time collected load data; the topology change inference model is evaluated, the accuracy and generalization ability of the model are evaluated by using a cross-validation method, and the model is optimized and improved according to the evaluation results, including adjusting the model parameters, introducing new features, to obtain a power distribution network dynamic topology inference model.

[0005] The technical scheme provided by the embodiment of the application can include the following beneficial effects:

[0006] The application discloses a method for information management and big data analysis of power engineering projects. BRIEF DESCRIPTION OF DRAWINGS

[0007] Figure 1 The flowchart of the method for information management and big data analysis of power engineering projects. DETAILED DESCRIPTION

[0008] In order for those skilled in the art to better understand the technical solutions in the specification, the technical solutions in the specification will be clearly and completely described in the specification below in combination with the drawings in the specification. Obviously, the described embodiments are only part of the embodiments of the specification, not all. Based on the embodiments in the specification, all other embodiments obtained by those of ordinary skill in the art without creative labor should belong to the scope of protection of the specification.

[0009] As Figure 1 The power engineering project information management and big data analysis method of the embodiment can specifically include:

[0010] In step S101, load time series data of power distribution transformer project management is obtained, and the data is preprocessed to obtain preprocessed load time series data, wherein the preprocessing includes data cleaning, outlier processing and data normalization to obtain preprocessed load time series data.

[0011] The original load time series data in the power distribution transformer management system database is obtained, and the original load time series data is preliminarily screened. According to the load time series data after preliminary screening, a 24-hour (according to the actual situation) window moving average method is used to detect outliers. If a point deviates from the average value by more than 3 times (according to the actual situation) standard deviation, it is determined as an outlier. For the outlier, the median of the data 4 hours (according to the actual situation) before and after the time point is used to replace it. After replacing the outliers, the load sequence is normalized by using the maximum and minimum value normalization method. According to the normalized load sequence, data smoothing is performed by using the exponential smoothing method to reduce short-term fluctuations. For the smoothed data, the time scale is adjusted to 1 hour (second time length) sampling interval from 15 minutes (first time length) sampling interval, and the average load value of each hour is calculated by using the simple average method. The time scale adjusted load time series data is obtained.

[0012] For example, raw load time series data is obtained from the power distribution transformer management system database, and preliminary screening is performed on the data to eliminate obviously unreasonable load records such as negative values or values exceeding the rated capacity of the transformer. For missing time point data, linear interpolation is used for supplementation to ensure the continuity and integrity of the time series. Based on the supplemented data, data quality assessment is performed, including integrity, consistency and accuracy checks, and a data quality report is generated. For the preliminary cleaned load sequence, a 24-hour window moving average method is used to detect and mark outliers. The average value and standard deviation within 24 hours before and after each time point are calculated, and if a point deviates from the average value by more than 3 times the standard deviation, it is determined to be an outlier. For outliers, the median of the data 4 hours before and after the time point is used for replacement to ensure that the replacement value is more consistent with the local load change trend. Based on the outlier replacement result, an outlier processing report is generated to record the position of the outlier and the value before and after replacement. The load sequence after removing outliers is normalized, and the maximum and minimum value normalization method is used to map the load value to the range of 0 to 1. The reason for choosing the maximum and minimum value normalization method is to keep the data distribution characteristics unchanged, which is convenient for subsequent analysis and modeling. The normalization formula is X_norm=(X-X_min) / (X_max-X_min), where X is the original load value, X_min and X_max are the minimum and maximum values of the load sequence, and X_norm is the normalized load value. Based on the normalization result, a normalization parameter report is generated to record the maximum value, minimum value and normalization formula. Based on the normalized load sequence, data smoothing is performed to reduce the influence of short-term fluctuations on long-term trends using the exponential smoothing method. The exponential smoothing formula is S_t=αX_t+(1-α)S_t-1, where S_t is the smoothed value at time t, X_t is the actual observation value at time t, and α is the smoothing coefficient, with a value range of 0 to 1. The selection of the smoothing coefficient α is based on the characteristics of the data, usually between 0.2 and 0.3. The smoothed data is adjusted in time scale to adjust the original 15-minute sampling interval to 1-hour interval, and the average load value of each hour is calculated using the simple average method. A data smoothing and time scale adjustment report is generated to record the smoothing parameters and time scale adjustment results. Raw load time series data is obtained from the power distribution transformer management system database, and the data contains timestamps and corresponding load values. During preliminary screening, the reasonable range of load values is set to 0 to 1.2 times the rated capacity of the transformer, and records exceeding this range are marked as invalid. For missing time point data, linear interpolation is used for supplementation. For example, if the load values at 10:00 and 10:30 are 100kW and 120kW respectively, and the data at 10:15 is missing, the load value at 10:15 is calculated to be 110kW by linear interpolation.In the data quality assessment, the integrity check ensures that there is a load value record every 15 minutes; the consistency check verifies that the units of the load values are uniform in kW; the accuracy check compares the load values at adjacent time points, and if the change exceeds 50%, it is marked as suspicious data. The moving average method with a 24-hour window is used to detect outliers. For each time point, the average value and standard deviation within 24 hours are calculated. If the load value of a certain point deviates from the average value by more than 3 times the standard deviation, it is determined to be an outlier. For example, the load value of a certain time point is 500 kW, and the average load within the 24-hour window is 300 kW, and the standard deviation is 50 kW, since 500 kW exceeds 300 kW + 3 * 50 kW = 450 kW, it is determined to be an outlier. The treatment of outliers uses the median of the data within the previous 4 hours to replace. If the load values within 4 hours around the outlier are 280 kW, 290 kW, 310 kW, 320 kW, 330 kW, 340 kW, 350 kW, 360 kW, the median is 325 kW, which replaces the 500 kW of the outlier. The data normalization processing uses the maximum and minimum value normalization method. Assuming that the minimum value in the original load sequence is 100 kW and the maximum value is 1000 kW, then for any original load value X, its normalized value is calculated as (X-100) / (1000-100). For example, the normalized result of the original load value 550 kW is (550-100) / (1000-100)=0.5. This normalization method preserves the relative distribution characteristics of the data, facilitating subsequent analysis and modeling. The data smoothing processing uses the exponential smoothing method, and the smoothing coefficient a is selected as 0.25. For the load value X_t at time t and the smoothing value S_t-1 at the previous time, the smoothing value S_t at time t is calculated as S_t=0.25X_t+0.75S_t-1. For example, if the smoothing value at t-1 is 300 kW and the actual load at t is 400 kW, then the smoothing value at t is 0.25400+0.75300=325 kW. The time scale adjustment converts the 15-minute interval data to 1-hour interval, which is achieved by calculating the average value of the 4 15-minute data points within each hour. For example, the 4 15-minute load values within a certain hour are 280 kW, 300 kW, 320 kW, and 340 kW, then the average load value of the hour is (280+300+320+340) / 4=310 kW.

[0013] In step S102, according to the preprocessed load time series data, the load mutation features and load correlation change features that can reflect the load transfer are extracted as load transfer features.

[0014] Specifically, according to the pre-processed load time series data, the features capable of reflecting load transfer are extracted, including load mutation features and load correlation change features, the first-order difference and the second-order difference of the load data are calculated to obtain the load mutation features, and the Pearson correlation coefficients between different transformer load data are calculated to obtain the load correlation change features, and the load mutation features and the load correlation change features are taken as the load transfer features.

[0015] According to the pre-processed load time series data, the difference between the load values of adjacent time points is calculated to obtain a first-order difference sequence; the difference between adjacent difference values is calculated for the first-order difference sequence to obtain a second-order difference sequence; and the first-order difference sequence and the second-order difference sequence are taken as load mutation features. The Pearson correlation coefficients between transformers are calculated using a 24-hour sliding window; a correlation matrix is constructed, wherein the matrix elements represent the correlation coefficients between two transformers, and if the correlation coefficient changes by more than a preset threshold, it is determined to be a significant change and is marked. By a feature vector splicing method, the load mutation features and the load correlation change features are combined to form a load transfer feature vector; the feature vector includes the first-order difference value, the second-order difference value and the element value of the significant change in the correlation matrix of each time point. Statistics of each dimension of the load transfer feature vector are calculated; principal component analysis is used to reduce the dimension of the feature vector, and principal components whose explained variance ratio reaches a preset value are retained. The K-means clustering algorithm is applied to the reduced feature vector to classify the load transfer features into different categories, and the load transfer modes corresponding to different categories are determined.

[0016] For example, according to the preprocessed load time series data, the first-order difference value of each time point is directly calculated, that is, the difference between the load values of adjacent two time points, to obtain a first-order difference sequence. For the first-order difference sequence, the difference between adjacent two first-order difference values is calculated again to obtain a second-order difference sequence. The first-order difference sequence and the second-order difference sequence are taken as load mutation features to reflect the speed and acceleration of load change. The load data of different transformers is extracted from the preprocessed load time series data, and the Pearson correlation coefficient between the transformers is calculated. A 24-hour sliding window is used, and the correlation coefficient is calculated every 6 hours to obtain a sequence of load correlation between transformers changing over time. A correlation matrix is constructed, and each element in the matrix represents the correlation coefficient between two transformers. The correlation coefficient change threshold is set to 0.3, and when the correlation coefficient change exceeds this threshold, it is determined to be a significant change and is marked. The first-order difference sequence and the second-order difference sequence calculated are taken as load mutation features, and the load correlation change sequence between transformers is taken as load correlation change features. Using a feature vector splicing method, these sequences are combined to form a load transfer feature vector. For each time point, the feature vector contains the first-order difference value, the second-order difference value at the time point, and the element value of the significant change in the correlation matrix. The extracted load transfer features are preliminarily analyzed, and the mean, variance, skewness, kurtosis and other statistical quantities of each dimension of the feature vector are calculated. Principal component analysis is used to reduce the dimension of the feature vector, and the principal components whose explanation variance ratio reaches 95% are retained. The K-means clustering algorithm is applied to the reduced feature vector to classify the load transfer features into different categories, and different types of load transfer patterns are preliminarily identified. The preprocessed load time series data contains the load value every 15 minutes within 24 hours. When calculating the first-order difference, for time point t, the first-order difference value is load value(t)-load value(t-1). For example, if the load at time t is 100 kW and the load at time t-1 is 95 kW, then the first-order difference value at time t is 5 kW. The second-order difference is calculated similarly, which is the difference between adjacent two first-order difference values. If the first-order difference at time t is 5 kW and the first-order difference at time t-1 is 3 kW, then the second-order difference at time t is 2 kW. These difference values reflect the speed and acceleration of load change. For the load data of three transformers A, B and C, a 24-hour sliding window is used, and the Pearson correlation coefficient is calculated every 6 hours. Assuming that in a certain time window, the correlation coefficient between A and B is 0.8, the correlation coefficient between B and C is 0.6, and the correlation coefficient between A and C is 0.5, then a correlation matrix [[1, 0.8, 0.5], [0.8, 1, 0.6], [0.5, 0.6, 1]] is constructed. The correlation coefficient change threshold is set to 0.3, and if the correlation coefficient between A and B in the next time window changes to 0.4, which exceeds the threshold, it is marked as a significant change. When splicing the feature vector, for each time point t, the first-order difference value, the second-order difference value at time t, and the element value of the significant change in the correlation matrix are included.For example, the feature vector at time t can be [5kW, 2kW, 0.4]. Statistical analysis is performed on the extracted load transfer features to calculate the mean, variance, skewness and kurtosis. Assuming that the mean of the first-order difference is 2kW, the variance is , the skewness is 0.5, and the kurtosis is 3.2. Dimensionality reduction is performed using principal component analysis. If the proportion of explained variance of the first two principal components reaches 95%, the two principal components are retained. K-means clustering algorithm is applied to the dimensionally reduced feature vector. Set K = 3 to divide the load transfer features into 3 categories, and preliminarily identify load transfer patterns such as stable, slow growth and sharp fluctuation.

[0017] Step S103, according to the load transfer feature, a clustering algorithm is used to cluster the transformers to obtain a transformer clustering result, and input into the project management system.

[0018] Specifically, according to the load transfer feature, a clustering algorithm is used to cluster the transformers, dividing transformers with similar load transfer patterns into the same category to obtain a transformer clustering result, which is input into the project management system.

[0019] The K-means clustering algorithm is applied to the load transfer feature. By calculating the within-cluster sum of squares under different clustering numbers, a curve graph is drawn. According to the inflection point of the curve graph, the number of initial clustering centers is determined. The Euclidean distance is used as the distance measurement standard, and the transformers are divided into different categories through iterative calculation to obtain a preliminary transformer clustering result. The average distance of the sample and other samples in the same category and the average distance to the nearest neighbor class sample are obtained. According to the average distance, the silhouette coefficient is calculated. The transformer clustering result is converted into structured data in JSON format. The structured data includes transformer number, category, and distance to clustering center. The structured data is imported into the project management system through RESTful API to update the transformer grouping information. According to the transformer grouping information, a differentiated load management strategy is developed for each category of transformers.

[0020] For example, according to the load transfer feature, the transformer load data is preprocessed, and first, the abnormal values and missing values are detected and processed. The quartile method is used to identify abnormal values, and data less than Q1-1.5IQR or greater than Q3+1.5IQR is marked as abnormal, where Q1 is the first quartile, Q3 is the third quartile, and IQR is the interquartile range. For abnormal values and missing values, the median of the feature dimension is used for filling. Then the preprocessed data is applied to the Z-score standardization method, and the numerical value of each feature dimension is converted to a standard normal distribution with a mean of 0 and a standard deviation of 1, to obtain the standardized load transfer feature data. The K-means clustering algorithm is applied to the standardized load transfer feature data. The elbow rule is used to determine the initial number of clustering centers, the within-cluster sum of squares under different clustering numbers is calculated, a curve graph is drawn, and the clustering number corresponding to the curve inflection point is selected as the initial value. The Euclidean distance is used as the distance measurement standard, and the transformers are divided into different categories through iterative calculation to obtain the preliminary transformer clustering result. The silhouette coefficient is used to evaluate the clustering effect, and the silhouette coefficient of each transformer sample is calculated. For sample i, the average distance a(i) of other samples in the same class and the average distance b(i) of the nearest neighbor class sample are calculated, and the silhouette coefficient s(i)=(b(i)-a(i)) / max(a(i),b(i)). The average value of the silhouette coefficients of all samples is calculated, and if the average silhouette coefficient is less than 0.5, the number of clustering centers is increased and the K-means clustering is re-executed until the average silhouette coefficient is greater than or equal to 0.5 or the number of clustering centers reaches the preset upper limit. The final number of clustering centers and the corresponding transformer classification result are recorded. The final transformer clustering result is converted into structured data in JSON format, including transformer number, class, distance to clustering center, etc. The clustering result is imported into the project management system using RESTful API (Representational State Transfer style interface) to update the transformer grouping information. In the project management system, differentiated load management strategies are developed for each class of transformers based on the clustering result, such as arranging more frequent maintenance plans for transformers in the high load category. A scatter plot of the clustering result is generated, with the X-axis and Y-axis representing two main load features, different colors representing different clusters, and the transformer grouping situation is visually displayed. When processing the transformer load transfer feature data, first, the original data is subjected to abnormal value detection and processing. For example, for a certain feature dimension, Q1=100kW, Q3=300kW, and IQR=200kW are calculated, and data less than -200kW or greater than 600kW is marked as abnormal. An abnormal value of 700kW is replaced by the median of the feature dimension, 200kW. The processed data is subjected to Z-score standardization, such as a feature with a mean of 250kW and a standard deviation of 50kW, and the original value of 300kW is standardized to 1. When applying the K-means clustering algorithm, the elbow rule is used to determine the initial number of clusters.The sum of squares within groups of 1 to 10 clusters is calculated, and it is found that a clear inflection point appears in the curve when the number of clusters is 4, so 4 is selected as the initial number of clusters. The K-means iterative process is performed, and it is assumed that convergence is achieved after 50 iterations, obtaining 4 cluster centers. When calculating the silhouette coefficient, a sample is selected, and the average distance a(i) of this sample from other samples in the same class is 0.2, and the average distance b(i) of this sample from the nearest neighbor class sample is 0.5, so the silhouette coefficient s(i) of this sample is (0.5-0.2) / 0.5=0.6. The average silhouette coefficient of all samples is calculated, and if it is 0.55, which is greater than the threshold value 0.5, the current clustering result is accepted. The clustering result is converted into JSON format, and the data is imported into the project management system through the RESTful API. In the project management system, the transformer group with a clustering result of 2 is arranged to have a monthly maintenance plan, while the transformer groups with clustering results of 1, 3, and 4 are arranged to have a maintenance plan every two months. Finally, a scatter plot is generated, with the X-axis representing the average load and the Y-axis representing the load fluctuation rate, and four colors representing the four clusters, respectively, which intuitively shows the grouping of the transformers and the load characteristics of each group.

[0021] In step S104, according to the transformer clustering result in the project management system, the dynamic change of the power distribution network topology structure is judged in combination with the change of time.

[0022] Specifically, according to the transformer clustering result in the project management system, the dynamic change of the power distribution network topology structure is judged in combination with the change of time, and if the degree of change of the transformer clustering result at a certain time is higher than a threshold value, it is determined that a topology structure change occurs at the certain time.

[0023] The transformer clustering results in the project management system in consecutive time periods are obtained, and a time series data structure is constructed, which records the clustering labels of each transformer at each time point. For the time series matrix, the clustering change degree between adjacent time points is calculated, and the Jaccard similarity is used to quantify the change of the clustering labels. If the clustering change degree exceeds a preset threshold value, the time point is marked as a potential topology structure change point. The potential topology structure change point is verified, the transformer load data and the line connection information are combined, and the graph edit distance algorithm is used to calculate the topology similarity between the before and after time points. According to the topology similarity, it is determined whether a topology structure change occurs, and if the topology similarity is lower than a preset threshold value, it is determined that a topology structure change occurs at the certain time. A topology structure change report is generated, which includes the change time, the involved transformers and lines, and the visual display of the network graph before and after the change.

[0024] For example, the transformer clustering results in a continuous time period are obtained from the project management system, a time series data structure is constructed, and the clustering labels of each transformer at each time point are recorded to form a time series matrix of transformer clustering results. In the matrix, the rows represent different time points, the columns represent different transformers, and the matrix elements are the clustering labels of the transformers at the corresponding time points. For the time series matrix, the clustering change degree between adjacent time points is calculated, and the Jaccard similarity is used to quantify the change of the clustering labels. For two adjacent time points t1 and t2, the Jaccard similarity of all transformer clustering labels is calculated, J(t1, t2) = |A∩B| / |A∪B|, where A and B are the sets of transformers with the same clustering labels at t1 and t2, respectively. The time series of clustering change degrees is obtained, and the change degree is defined as 1-J(t1, t2). A threshold of clustering change degree is set, and the time series of clustering change degrees is scanned by the sliding window method. The window size is set to 3 time points, and the threshold is determined by statistical analysis of historical data, taking the 95% quantile of the change degree distribution as the threshold. If the change degree in the window exceeds the preset threshold, the time point is marked as a potential topology change point. Preliminary analysis is performed on the marked potential change point, and the difference in clustering distribution before and after the change is recorded. The marked potential topology change point is verified, and the topology similarity of the two time points is calculated by combining the transformer load data and the line connection information, using the graph edit distance algorithm. The distribution network topology is represented as a graph G(V, E), where V is the set of transformers and E is the set of lines. The edit distance of two graphs G1 and G2 is calculated, including the minimum number of node (transformer) insertion, deletion and replacement, and edge insertion, deletion and replacement. The normalized edit distance gives the topology similarity, and if the similarity is lower than the set threshold 0.9, it is confirmed that the topology structure has changed at that time. A topology change report is generated, including the change time, the involved transformers and lines, and the network graph visualization before and after the change. In the implementation of the dynamic change detection of the distribution network topology, the transformer clustering results in the last 24 hours are first obtained from the project management system, with an interval of 15 minutes, a total of 96 time points. A 96-row and 100-column time series matrix is constructed, with rows representing time points and columns representing 100 transformers, and matrix elements being integers from 1 to 5 representing the clustering of transformers. For example, matrix element M[50, 30] = 3 indicates that the 30th transformer at the 50th time point (12:15) belongs to the 3rd cluster. The Jaccard similarity of adjacent time points is calculated, such as t1 clustering label set {1, 2, 3, 4, 5} and t2 clustering label set {1, 2, 3, 4, 6}, then J(t1, t2) = 4 / 6 = 0.667, and the change degree is 1-0.667 = 0.333. Statistical analysis of the 95 change degree values gives a 95% quantile of 0.2, which is set as the threshold.The change degree sequence is scanned using a sliding window of size 3, and if the maximum value in the window exceeds 0.2, the center time point is marked as a potential change point. For the marked potential change point, such as 13:30, the clustering distribution before and after the change is recorded: before the change {1:20, 2:25, 3:30, 4:15, 5:10}, and after the change {1:15, 2:20, 3:40, 4:15, 5:10}. Combined with the transformer load data and line connection information, a graph G(V, E) is constructed, V is 100 transformer nodes, and E is 150 connection lines. The edit distance of the graph G1 and G2 at two time points 13:15 and 13:45 is calculated, including 5 transformer label changes (cluster from 2 to 3) and 2 edge additions, and the normalized topological similarity is 0.88, which is lower than the threshold 0.9, confirming that the topological structure changes at 13:30. A change report is generated, including the change time 13:30, the 5 transformer numbers involved, and the 2 newly added lines, and the network structure before and after the change is visualized by force-directed graph algorithm, which intuitively reflects the adjustment of local network connection.

[0025] In step S105, if the topological structure changes, the corresponding topological structure change including the load transfer path is inferred according to the dynamic topological structure of the distribution network in the project management system.

[0026] Specifically, if the topological structure changes, it is judged whether the topological change occurs in the process of frequent switching operation or in the process of fault recovery according to the dynamic topological structure of the distribution network in the project management system. If so, the specific change of the topological structure is judged by analyzing the change mode of the transformer clustering result. If the clustering result of the transformer appears a specified change mode, the corresponding topological structure change including the load transfer path is inferred.

[0027] The graph structure information before and after the topological change is obtained according to the dynamic topological structure of the distribution network in the project management system, and the graph structure information includes node (transformer) connection relationship and edge weight. The difference graph of the topological change is represented by an adjacency matrix, and the element value of the adjacency matrix is 1, indicating an added edge, -1, indicating a deleted edge, and other non-zero values, indicating a weight change amount. The time interval of continuous change and the affected load are determined by analyzing the time characteristics and range characteristics of the topological change. If the time interval is less than a first preset time threshold and the affected load is less than a first proportion threshold of the total load, it is determined as a frequent switching operation. The time series data of the transformer clustering result is obtained, and the change matrix of the clustering labels of adjacent time points is calculated according to the time series data. The change matrix is subjected to sequence pattern mining by the SPADE algorithm, and a specified change mode is obtained. According to the matching result of the change mode and the predefined topological structure change template and the difference graph, the specific topological structure change is judged, and the specific topological structure change includes the load transfer path.

[0028] For example, according to the dynamic topology of the power distribution network in the project management system, the graph structure information before and after the topology change is extracted, including the connection relationship of nodes (transformers) and the edge weight. A difference graph of topology change is constructed, represented by an adjacency matrix, and the matrix element value is 1, indicating a new edge, -1 indicating a deleted edge, and other non-zero values indicating the weight change. Record the number of changed edges and the total weight change. Analyze the time characteristics and range characteristics of the topology change, calculate the time interval of continuous change and the affected load. If the time interval is less than 30 minutes and the affected load is less than 10% of the total load, it is determined to be a frequent switching operation; if the time interval is greater than 2 hours or the affected load is greater than 30% of the total load, it is determined to be a fault recovery process. Get the type label of the topology change, and record the change frequency of the frequent switching operation and the recovery duration of the fault recovery process. Extract the time series data of the transformer clustering result, calculate the change matrix of the clustering label at adjacent time points. Apply the SPADE algorithm for sequence pattern mining to identify specific change patterns. Predefine the topology structure change template, including typical scenarios such as load transfer, line switching, and transformer commissioning and decommissioning, obtained through expert experience and historical data statistics. Match the identified patterns with the predefined templates and calculate the matching score. Based on the matching result and the topology difference graph, infer the specific topology structure change, including the load transfer path. Use Dijkstra's algorithm to find the shortest path between the changed transformers in the difference graph to determine the specific line and direction of the load transfer. Calculate the edge weight change on the path to estimate the transferred load, and generate a time sequence visualization graph of the topology change process, with the X-axis representing time and the Y-axis representing the affected transformers and lines, marked with different colors to represent the change type. In the implementation of the power distribution network topology structure change analysis, first extract the graph structure information before and after the topology change from the project management system, and construct a 100x100 adjacency matrix difference graph. For example, matrix element M[25,30]=1 indicates that a new edge is added between transformers 25 and 30, M[40,45]=-1 indicates that an edge is deleted between transformers 40 and 45, and M[60,65]=0.5 indicates that the edge weight between transformers 60 and 65 increases by 0.5. The statistical results show that the number of changed edges is 15 and the total weight change is 8.5. Then analyze the time and range characteristics of the topology change, find that the time interval of continuous change is 20 minutes and the affected load is 8% of the total load, and determine it as a frequent switching operation, recording the change frequency as 3 times / hour. Extract the time series data of the transformer clustering result to generate a 100x96 matrix, with rows representing 100 transformers and columns representing 96 15-minute time points. Apply the SPADE algorithm for sequence pattern mining, set the minimum support to 0.05, and get the frequent sequence pattern {1->2->1}, which represents the pattern of changing the clustering label from 1 to 2 and then back to 1. Match this pattern with the predefined "line switching" template and calculate the matching score as 0.85.Based on the matching result and the topology difference graph, the Dijkstra algorithm is used to find the shortest path between the transformer 25 and the transformer 30 with the cluster change, and the path [25, 28, 29, 30] is obtained, and the transferred load is estimated to be 2MW. Finally, the time sequence visualization graph of the topology change process is generated, the X axis represents the 24-hour time period, the Y axis lists the 10 affected transformers and the 15 lines, the newly added connection is marked in red, the deleted connection is marked in blue, and the weight change is marked in yellow, which intuitively shows the frequent switching operation process during 20:00 to 20:20.

[0029] In step S106, according to the historical inferred topology structure change, a decision tree algorithm is used to take the load transfer characteristics of the transformer and the clustering result as input, train a topology change inference model, and infer the current power distribution network topology according to the real-time collected load data, so as to realize real-time monitoring of the power distribution network topology.

[0030] The topology structure change record in the historical database is obtained, and the topology structure change record includes the transformer load transfer characteristics, the clustering result and the corresponding topology change type; the topology structure change record is data cleaned, the continuous type characteristics are processed by using the Z-score standardization method, and the category type characteristics are one-hot encoded to obtain a standardized data set; key features are extracted according to the standardized data set, the key features include the load change rate, the clustering label conversion frequency and the load transfer direction, wherein the load transfer direction is quantified by calculating the load difference change between adjacent transformers; a C4.5 decision tree algorithm is used to construct a topology change inference model, the maximum tree depth is set to a preset value, the minimum leaf node sample number is set to a preset value, and the information gain ratio is used as the node splitting criterion; the load data and the transformer clustering result of the current power distribution network are obtained from the real-time data acquisition system, the same preprocessing and feature extraction as the training data are performed, the processed data are input into the topology change inference model, and the current power distribution network topology inference result is output; it is judged whether the difference between the power distribution network topology inference result and the topology structure of the last moment exceeds a preset threshold, if the preset threshold is exceeded, an artificial review process is triggered; the power distribution network topology inference result is visualized as a dynamic topology graph, the nodes in the dynamic topology graph represent transformers, the edges represent line connections, and the color depth represents the load size.

[0031] Exemplarily, the inferred topological structure change records, including transformer load transfer features, clustering results and corresponding topological change types, are extracted from the historical database. The data is cleaned to remove outliers and missing data, and the Z-score standardization method is used to process continuous features. The category type features are one-hot encoded to obtain a standardized dataset for model training. Key features are extracted from the standardized dataset, including load change rate, clustering label conversion frequency, load transfer direction, etc. The load transfer direction is quantified by calculating the load difference change between adjacent transformers. Principal component analysis is applied to continuous features to reduce dimensionality, and principal components with cumulative contribution rate exceeding 95% are selected. The original category type features are retained and combined with the reduced continuous features as input features for the decision tree algorithm. The C4.5 decision tree algorithm is used to construct the topological change inference model, with a maximum tree depth of 10 and a minimum leaf node sample size of 5. The information gain ratio is used as the node splitting criterion. The optimal parameters are selected by 5-fold cross-validation to obtain the trained decision tree model. The accuracy and F1 score of the model on the validation set are calculated, and if the accuracy is less than 90% or the F1 score is less than 0.85, the parameters are adjusted and retrained. The current load data and transformer clustering results of the distribution network are obtained from the real-time data acquisition system, and after the same preprocessing and feature extraction steps as the training data, the processed data is input into the trained decision tree model to output the current distribution network topological structure inference result. An anomaly detection mechanism is set up, and if the difference between the model output and the previous time topological structure exceeds the preset threshold, the manual review process is triggered. The inference result is visualized as a dynamic topology graph, with nodes representing transformers and edges representing line connections, and color depth representing load size. The model is updated incrementally every 24 hours by adding new historical data to adapt to long-term changes in the distribution network. In the implementation of real-time monitoring of the distribution network topological structure, 1000 topological structure change records are first extracted from the historical database, each record containing load transfer features of 20 transformers, 5 clustering results and 1 topological change type. The data is cleaned to remove 50 outliers and 30 missing data, and the Z-score standardization method is used to process 15 continuous features, adjusting the mean to 0 and the standard deviation to 1. The 5 category type features are one-hot encoded to expand to 25 binary features. Key features such as load change rate, 15-minute percentage change, clustering label conversion frequency, and load transfer direction are extracted. Principal component analysis is applied to reduce the dimensionality of continuous features, and the first 8 principal components are selected, with a cumulative contribution rate of 96.5%. The 8 reduced continuous features and 25 binary features are combined to obtain 33 input features. The C4.5 decision tree algorithm is used to construct the model, with a maximum tree depth of 10 and a minimum leaf node sample size of 5. The information gain ratio is used as the splitting criterion.The optimal parameters are selected by 5-fold cross-validation, and the trained decision tree model reaches an accuracy of 92.5% and an F1 score of 0.88 on the validation set. The current power distribution network data is obtained from the real-time data acquisition system every 5 minutes, preprocessed and input into the model, and the topology inference result is output. Set the anomaly detection threshold to 30%, if the difference between the model output and the topology structure at the last time exceeds 30%, trigger manual review. The inference result is visualized as a dynamic topology graph, containing 100 transformer nodes and 150 line edges, and the node color from green to red represents the load from light to heavy. The model is updated every 24 hours, adding 200 new historical data, retraining the decision tree model to ensure that the model adapts to the long-term changes of the power distribution network.

[0032] Step S107, evaluate the topology change inference model, use cross-validation method to evaluate the accuracy and generalization ability of the model, according to the evaluation result, optimize and improve the model, including adjusting the model parameters, introducing new features, get the dynamic topology inference model of power distribution network, provide support for real-time management of power distribution network project.

[0033] Obtain the historical data set carrying the topology change type identifier, the historical data set is generated by the power distribution network monitoring system; according to the historical data set, use the rolling window method to divide the time series, get the training window data set and the test window data set; calculate the model performance indicators for the test window data set, the performance indicators include accuracy, precision, recall and F1 score; use the weighted average method to comprehensively evaluate the performance indicators, get the model evaluation score; if the model evaluation score is lower than the preset threshold, adjust the model parameters using the Bayesian optimization method; obtain new feature data, the new feature data includes load time series autocorrelation coefficient, adjacent transformer load cross-correlation coefficient and load fluctuation spectrum analysis result; retrain the model according to the new feature data and the optimized model parameters, get the dynamic topology inference model of power distribution network; receive real-time power distribution network monitoring data, the real-time power distribution network monitoring data is collected by the power distribution network monitoring system every hour; use the dynamic topology inference model of power distribution network to process the real-time power distribution network monitoring data, get the real-time topology inference result;

[0034] For example, the historical data set of the topology change inference model is divided into time series, and cross-validation is performed using the rolling window method. The initial training window is set to 6 months, and the test window is 1 month. Each time, it is rolled forward by 1 month, and a total of 12 validations are performed. Each validation uses the data in the training window to train the model and evaluates the model performance in the test window. For each validation, the accuracy, precision, recall, and F1 score of the model on the test window are calculated, and the performance of the model on different topology change types is recorded. The weighted average method is used to evaluate the overall performance of the model, with weights set as: accuracy 0.3, precision 0.2, recall 0.2, and F1 score 0.3. The identification ability of the model on specific types of topology changes is analyzed to find the performance deficiencies of the change types. According to the cross-validation results, the model parameters are adjusted using Bayesian optimization. The optimization objective function is set to the weighted average performance score, and the optimization parameters include the maximum depth of the decision tree (range 5-20), the minimum leaf node sample size (range 1-10), and the information gain threshold (range 0.01-0.1). Gaussian process regression is used as the surrogate model, and the optimal parameter combination is found through 100 iterations. New features are introduced, including the time series autocorrelation coefficients of transformer load (lag 1, 6, 24 hours), the cross-correlation coefficients of adjacent transformer loads (lag 0, 1 hour), and the spectral analysis results of load fluctuations, extracting the amplitude of the top 3 main frequency components. These features are combined with the original features to train the dynamic topology inference model of the distribution network. The SHAP (SHapley Additive ex Planations) value is used to analyze the importance of the features, and the top 90% of the features are retained. In the actual project, the dynamic topology inference model of the distribution network is applied, and the topology inference results are updated every hour. The real-time accuracy is calculated by comparing it with the manual verification results. When the real-time accuracy is less than 85%, the dynamic topology inference model of the distribution network is triggered to retrain. In the implementation of the dynamic topology inference model optimization process of the distribution network, the historical data of 18 months is first divided into time series, and cross-validation is performed using the rolling window method. The initial training window is set to the previous 6 months, from January to June 2023, and the test window is the following month, July 2023. Each time, it is rolled forward by 1 month, and a total of 12 validations are performed. In the first validation, the performance indicators of the model on the test window are accuracy 88%, precision 85%, recall 87%, and F1 score 86%. The identification ability of the 5 main topology change types is line switching 93%, transformer overload switching 82%, fault isolation 89%, load transfer 85%, and new device grid connection 78%. The weighted average method is used to calculate the overall performance score as 86.9%. Through Bayesian optimization, the model parameters are adjusted, and the optimization objective function is set to the weighted average performance score. After 100 iterations, the optimal parameter combination is obtained: maximum depth 15, minimum leaf node sample size 3, and information gain threshold 0.03.After introducing the new features, the 1-hour, 6-hour and 24-hour lag autocorrelation coefficients of the transformer load are 0.85, 0.72 and 0.65, respectively, the 0-hour and 1-hour lag cross-correlation coefficients of the adjacent transformer loads are 0.62 and 0.58, respectively, and the amplitude of the three main frequency components extracted by the load fluctuation spectrum analysis is 0.45, 0.32 and 0.24. The SHAP value analysis shows that the newly introduced 6-hour lag autocorrelation coefficient and 0-hour lag cross-correlation coefficient enter the top 10 important features. In the actual application of the distribution network dynamic topology inference model in the last month, June 2024, the topology inference result is updated every hour, and the real-time accuracy compared with manual verification is 92%, which is improved by 4 percentage points compared with before optimization.

[0035] The specific embodiments of the application are described above. It should be understood that the application is not limited to the specific embodiments described above, and various changes or modifications can be made by those skilled in the art within the scope of the claims, which do not affect the essential content of the application. The embodiments of the present application and the features in the embodiments can be arbitrarily combined with each other without conflict.

Claims

1. A method for information management and big data analysis of power engineering projects, characterized in that, The method includes: Obtain load time series data from the distribution transformer project management, and preprocess the data to obtain preprocessed load time series data; Based on the preprocessed load time series data, load abrupt change characteristics and load correlation change characteristics that can reflect load transfer are extracted as load transfer characteristics; Based on load transfer characteristics, a clustering algorithm is used to cluster the transformers, and the resulting clustering results are then entered into the project management system. Based on the transformer clustering results in the project management system and the changes over time, the dynamic changes in the distribution network topology are determined. If the topology changes, the corresponding topology changes, including load transfer paths, can be inferred based on the dynamic topology of the distribution network in the project management system. Based on historical topology changes, a decision tree algorithm is used to train a topology change inference model, taking transformer load transfer characteristics and clustering results as input. The model then infers the current distribution network topology based on real-time collected load data. The topology change inference model is evaluated using cross-validation to assess its accuracy and generalization ability. Based on the evaluation results, the model is optimized and improved, including adjusting model parameters and introducing new features, to obtain the dynamic topology inference model of the distribution network. The method for inferring topology changes based on historical data employs a decision tree algorithm. Using transformer load transfer characteristics and clustering results as input, a topology change inference model is trained. Based on real-time collected load data, the current distribution network topology is inferred, including: Retrieve topology change records from the historical database. These records include transformer load transfer characteristics, clustering results, and corresponding topology change types. The data records of topological changes are cleaned, continuous features are processed using a standardization method, and categorical features are one-hot encoded to obtain a standardized dataset. Key features are extracted from a standardized dataset, including load change rate, cluster label switching frequency, and load transfer direction, wherein the load transfer direction is quantified by calculating the load difference change between adjacent transformers. A topology change inference model is constructed using a decision tree algorithm. The maximum tree depth and the minimum number of leaf node samples are set as preset values, and the information gain ratio is used as the node splitting criterion. The load data and transformer clustering results of the current distribution network are obtained from the real-time data acquisition system. After the same preprocessing and feature extraction as the training data, the processed data is input into the topology change inference model and the current distribution network topology inference results are output. The topology change inference model is evaluated using cross-validation to assess its accuracy and generalization ability. Based on the evaluation results, the model is optimized and improved, including adjusting model parameters and introducing new features, to obtain a dynamic topology inference model for the distribution network, including: Obtain a historical dataset carrying a topology change type identifier, the historical dataset being recorded and generated by the distribution network monitoring system; The time series data is divided using the rolling window method based on the historical dataset to obtain the training window dataset and the test window dataset. Calculate model performance metrics for the test window dataset. Performance metrics include accuracy, precision, recall, and F1 score. A weighted average method is used to comprehensively evaluate the performance indicators and obtain the model evaluation score; If the model evaluation score is lower than a preset threshold, the model parameters are adjusted using a Bayesian optimization method. Acquire new feature data, including load time series autocorrelation coefficient, cross-correlation coefficient of adjacent transformer loads, and load fluctuation spectrum analysis results; The model is retrained based on the new feature data and optimized model parameters to obtain the dynamic topology inference model of the distribution network.

2. The method according to claim 1, characterized in that, The process of acquiring load time series data from distribution transformer project management and preprocessing the data to obtain preprocessed load time series data includes: Obtain raw load time series data from the distribution transformer management system database, and perform preliminary screening on the raw load time series data; Based on the initially screened load time series data, outliers are detected using the moving average method with a preset time window to obtain outlier points; For outliers, the median of the data before and after the outlier for a specified time period is used for replacement; The load sequence after replacing outliers is normalized. Data smoothing is performed based on the normalized load sequence; The time scale of the smoothed data is adjusted by changing the first time period sampling interval to the second time period sampling interval, and the average load value per hour is calculated by the simple averaging method. We obtain the time-scale adjusted load time series data.

3. The method according to claim 1, characterized in that, The step of extracting load abrupt change features and load correlation change features that can reflect load transfer from the preprocessed load time series data as load transfer features includes: The difference between load values ​​at adjacent time points is calculated based on the preprocessed load time series data to obtain a first-order difference sequence; The difference between adjacent difference values ​​is calculated for the first-order difference sequence to obtain the second-order difference sequence; First-order difference sequences and second-order difference sequences are used as characteristics of load mutations; The Pearson correlation coefficient of the inter-transformer load was calculated using a 24-hour sliding window. Construct a correlation matrix to obtain the load correlation change characteristics, where the matrix elements represent the correlation coefficient between two transformers. If the correlation coefficient changes beyond a preset threshold, it is judged as a significant change and marked. By using the feature vector concatenation method, load abrupt change features and load correlation change features are combined to form load transfer features.

4. The method according to claim 1, characterized in that, The process of clustering transformers based on load transfer characteristics, obtaining transformer clustering results, and inputting them into the project management system includes: The K-means clustering algorithm was applied to the load transfer characteristics, and the sum of squares within groups under different numbers of clusters was plotted. The initial number of cluster centers is determined based on the inflection point of the curve. Using Euclidean distance as the distance metric, transformers are classified into different categories through iterative calculations, yielding preliminary transformer clustering results. Obtain the average distance between a sample and other samples of the same class, and the average distance between a sample and its nearest neighbor sample. Calculate the profile coefficient based on the average distance; The transformer clustering results are converted into structured data in JSON format, which includes the transformer number, its category, and its distance from the cluster center. The structured data is imported into the project management system via a RESTful API to update the transformer grouping information.

5. The method according to claim 1, characterized in that, The step of determining the dynamic changes in the distribution network topology based on transformer clustering results in the project management system and changes over time includes: Obtain the transformer clustering results for consecutive time periods in the project management system, construct a time series data structure, and record the clustering labels of each transformer at each time point; Calculate the degree of cluster change between adjacent time points and use a similarity measure to quantify the change in cluster labels; If the degree of cluster change exceeds a preset threshold, then that time point is marked as a potential topological change point; The potential topology change points are verified, and the topological similarity between the previous and next time points is calculated by combining transformer load data and line connection information using a graph editing distance algorithm. The topological similarity is used to determine whether a topological change has occurred. If the topological similarity is lower than a preset threshold, then a topological change is determined to have occurred at that moment.

6. The method according to claim 1, characterized in that, If the topology changes, the corresponding topology changes, including load transfer paths, are inferred based on the dynamic topology of the distribution network in the project management system, including: The graph structure information before and after the topology change is obtained from the dynamic topology of the distribution network in the project management system. The graph structure information includes node connection relationships and edge weights. The difference graph of topological changes is represented using an adjacency matrix; The temporal and extent characteristics of topology changes are analyzed to determine the time intervals of continuous changes and the load impact. If the time interval is less than the preset first time threshold and the amount of load affected is less than the first proportion threshold of the total load, it is determined to be a frequent switching operation. Obtain time series data of transformer clustering results, and calculate the change matrix of cluster labels between adjacent time points based on the time series data; Perform sequence pattern mining on the change matrix to obtain the specified change pattern; Based on the matching results and difference diagrams between the change pattern and the predefined topology change template, the specific topology change is determined, including the load transfer path.

7. The method according to claim 6, characterized in that, The adjacency matrix has an element value of 1 indicating a new edge, -1 indicating a deleted edge, and other non-zero values ​​indicating changes in weight.

8. The method according to claim 1, characterized in that, After outputting the current distribution network topology inference result, the following is included: Determine whether the difference between the inferred distribution network topology and the topology at the previous time exceeds a preset threshold. If it exceeds the preset threshold, trigger the manual review process. The inferred distribution network topology is visualized as a dynamic topology graph, where nodes represent transformers, edges represent line connections, and color intensity represents load size.

Images