Power grid hierarchical load prediction method, device and equipment

Abstract

The application discloses a power grid hierarchical load prediction method, device and equipment, relates to the technical field of load prediction, and comprises the following steps: a table area load diagram model is constructed, load change analysis is carried out on the nodes of the table area load diagram model, a first load change value is obtained, load superposition disturbance is carried out on the table area load diagram model, a plurality of load superposition paths are obtained, load peak value identification is carried out on the plurality of load superposition paths, and according to the identification result, continuous load superposition time periods are divided, a load superposition segmentation set is obtained, the load superposition paths are screened in combination with the load superposition segmentation set, a first screening path set is obtained, the first screening path set is screened based on the first load change value, a second screening path set is obtained, hierarchical load prediction is carried out based on the second screening path set, and the problem that load superposition leads to table area load peak value burst is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of load prediction, and more particularly, to a power grid hierarchical load prediction method, device and equipment. BACKGROUND

[0002] In the operation and management process of the power system, accurately predicting the dynamic changes of the load of each district in the distribution network has extremely important practical significance. On the one hand, accurate load prediction can guarantee the safe and stable operation of the power grid, identify possible overload, power fluctuation or power supply imbalance in advance, thereby reducing equipment wear and tear and accident risk; on the other hand, it provides a scientific basis for dispatchers, which helps to optimize power generation plans, regulate power flow in the power grid, and improve the flexibility and response speed of power dispatching. In addition, with the development of distributed power supply, renewable energy access and user load diversification, the temporal and spatial fluctuations of district loads present higher complexity and uncertainty, which puts higher requirements on the accuracy and real-time performance of load prediction. Therefore, researching and applying methods that can finely and hierarchically predict district loads have become a key technical link to improve the intelligent management level of the power grid and ensure power supply reliability.

[0003] For example, the power grid load prediction method and device disclosed in the patent No. CN111130107B, comprising: predicting the spatial load of the power grid; predicting the load of various load devices and the input power of various power supply devices to the power grid in the power grid, the load including electric vehicle load and demand side load; the input power including photovoltaic input power, wind turbine input power and energy storage input power; determining the load prediction result of the power grid based on the spatial load. The device comprises: a first prediction module for predicting the spatial load of the power grid; a second prediction module for predicting the electric vehicle load, demand side load, photovoltaic input power, wind turbine input power and energy storage input power in the power grid; a determination module for determining the load prediction result of the power grid. The present application solves the technical problem that the related art cannot meet the requirement of the power grid for the accuracy of load prediction due to the fewer factors considered by the above-mentioned method and device.

[0004] The above-mentioned technical solution has at least the following technical problems:

[0005] In modern distribution networks, with the large-scale access of high-power loads such as electric vehicles, distributed energy and smart homes, the load of the district presents obvious concentrated superposition characteristics. When a large number of loads start or charge at the same time in the same time period, the instantaneous load of the district will have a sudden peak increase, which is much higher than the historical average load level. The rapid superposition of such peak load can cause local voltage drop, current exceeding the rated capacity of the line, and frequent action of protection devices, thereby causing unstable power supply or equipment damage risk in the district. SUMMARY

[0006] In order to overcome the above-mentioned defects of the prior art, embodiments of the present application provide a power grid hierarchical load prediction method, device and equipment, which solves the problem of sudden peak of substation load caused by load superposition through modeling and load superposition analysis of substation load.

[0007] To achieve the above object, the present application provides the following technical scheme:

[0008] The power grid hierarchical load prediction method comprises the following steps: constructing a substation load graph model, performing load change analysis on nodes of the substation load graph model, and obtaining a first load change value; performing load superposition disturbance on the substation load graph model to obtain a plurality of load superposition paths; performing load peak value identification on the plurality of load superposition paths, and dividing continuous load superposition time periods according to the identification results to obtain a load superposition segmentation set; screening the load superposition paths in combination with the load superposition segmentation set to obtain a first screening path set; screening the first screening path set based on the first load change value to obtain a second screening path set, and performing hierarchical load prediction based on the second screening path set.

[0009] In a preferred embodiment, the step of constructing a substation load graph model and performing load change analysis on nodes of the substation load graph model to obtain a first load change value specifically comprises: collecting historical load data of substation load points, and constructing a substation load graph model based on the historical load data; calculating the sliding window variance of the load time series of each node in the substation load graph model to obtain a load fluctuation sequence.

[0010] Performing empirical mode decomposition on the load fluctuation sequence of each node to obtain a plurality of node residual components; calculating the fluctuation amplitude and frequency of the plurality of node residual components to obtain a load fluctuation feature vector; performing cluster analysis on the load fluctuation feature vectors of all nodes, and calculating the maximum load deviation value of the nodes based on the cluster results to obtain the first load change value.

[0011] In a preferred embodiment, the step of performing load superposition disturbance on the substation load graph model to obtain a plurality of load superposition paths specifically comprises: constructing a disturbance priority of nodes in the substation load graph model according to the first load change value; applying random load fluctuation disturbance to the nodes based on the disturbance priority to obtain an initial disturbance scenario; applying a disturbance sequence to the nodes on the substation load graph model under the initial disturbance scenario to obtain a plurality of load superposition paths.

[0012] In a preferred embodiment, the step of performing load peak value identification on the plurality of load superposition paths and dividing continuous load superposition time periods according to the identification results to obtain a load superposition segmentation set specifically comprises:

[0013] In a preferred embodiment, the step of performing load peak value identification on the plurality of load superposition paths and dividing continuous load superposition time periods according to the identification results to obtain a load superposition segmentation set specifically comprises:extracting a load time sequence of each node on each load superimposed path and performing peak detection by using a preset multi-scale peak detection algorithm to obtain a plurality of load peak points; analyzing time distribution of the plurality of load peak points, and aligning peak sequences of different load superimposed paths by using a dynamic time warping algorithm to obtain aligned peak sequences; dividing the load time sequence into continuous time periods based on the aligned peak sequences to obtain a plurality of load superimposed segments; and merging the plurality of load superimposed segments to obtain a load superimposed segment set.

[0014] In a preferred embodiment, the peak detection by using the preset multi-scale peak detection algorithm to obtain the plurality of load peak points specifically includes: setting sliding windows of different time scales, and analyzing load gradient changes of the load time sequence of each node in each sliding window; identifying a plurality of first peak points according to the analysis results and performing multi-scale fusion on the first peak points to obtain the plurality of load peak points.

[0015] In a preferred embodiment, the screening of the load superimposed paths based on the load superimposed segment set to obtain the first screening path set specifically includes: extracting load features of each load superimposed segment, and constructing a segment feature matrix based on the load features; calculating Mahalanobis distances between the load features based on the segment feature matrix, and grouping the load superimposed paths by using a spectral clustering algorithm to obtain grouped load superimposed paths; and dividing the grouped load superimposed paths into different path categories according to the Mahalanobis distances to obtain the first screening path set.

[0016] In a preferred embodiment, the screening of the first screening path set based on the first load change value to obtain the second screening path set specifically includes: extracting load change trend feature values of each load superimposed path in the first screening path set, and analyzing the load change trend feature values with the first load change value; constructing an adaptive screening threshold according to the analysis results, and screening the first screening path set based on the adaptive screening threshold by using a fuzzy comprehensive evaluation method to obtain the second screening path set.

[0017] In a preferred embodiment, the hierarchical load prediction based on the second screening path set specifically includes: extracting load sequences from the second screening path set, and constructing a time series data set based on the load sequences; dividing the time series data set into load grade intervals to obtain the load grade intervals; constructing a spatio-temporal graph neural network model based on the load grade intervals to perform load prediction, and outputting a hierarchical load prediction result.

[0018] The present application can obtain the first load change value by constructing the transformer area load diagram model, accurately analyzing the load change of each node of the transformer area, and providing accurate basic data for subsequent load prediction; on this basis, a plurality of load superposition paths are obtained by applying load superposition disturbance to the transformer area load model, and load peak value identification is performed on these paths, thereby effectively dividing the continuous load superposition period, forming the load superposition segmentation set, and realizing the early identification and analysis of the load peak value burst; further, the paths are screened based on the load superposition segmentation set to obtain the first screening path set, and the second screening path set is obtained based on the first load change value for secondary screening, thereby providing high-precision path basis for hierarchical load prediction. Through the above technical process, the method can not only accurately capture the peak value burst phenomenon of the transformer area load under the superposition effect, but also effectively screen and grade the load prediction path, realize the fine management and early warning of the transformer area load, and significantly improve the accuracy and reliability of the load prediction, thereby providing technical protection for power system operation and dispatch, load optimization distribution and safe operation of the power grid. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The flowchart of the power grid hierarchical load prediction method of the present application is shown.

[0020] Figure 2 The device structure diagram of the power grid hierarchical load prediction method of the present application is shown.

[0021] Figure 3 The electronic device structure diagram of the power grid hierarchical load prediction method of the present application is shown. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0023] Embodiment 1, Figure 1 The power grid hierarchical load prediction method of the present application is given, including the following steps:

[0024] S1, constructing a transformer area load diagram model, and performing load change analysis on the nodes of the transformer area load diagram model to obtain a first load change value;

[0025] In this example, a transformer area load diagram model is constructed, and load change analysis is performed on the nodes of the transformer area load diagram model to obtain a first load change value, specifically:

[0026] Collect historical load data of the transformer area load points and construct a transformer area load map model based on the historical load data;

[0027] Calculate the sliding window variance of the load time series of each node in the transformer area load diagram model to obtain the load fluctuation series;

[0028] Empirical mode decomposition is performed on the load fluctuation sequence of each node to obtain several node residual components;

[0029] Calculate the fluctuation amplitude and frequency of several node residual components to obtain the load fluctuation characteristic vector;

[0030] Cluster analysis is performed on the load fluctuation feature vectors of all nodes, and the maximum deviation of node load is calculated based on the clustering results to obtain the first load change value.

[0031] In this example, the formulas for calculating the fluctuation amplitude and frequency of the nodal residual components are as follows:

[0032] The formula for calculating the fluctuation range is as follows:

[0033]

[0034] in, The fluctuation amplitude of the nodal residual components. The length of the load time series. Let be the Fourier transform amplitude of the residual component of the i-th node at time t. Let be the mean of the residual component of the m-th node over the entire time series. Number the nodes. Number the node residual components.

[0035] The formula for calculating frequency is as follows:

[0036]

[0037] in, The frequency of the nodal residual components, This represents the total number of frequency components in the Fourier transform. For the residual component of the i-th node and the m-th node at frequency The Fourier transform amplitude under the given conditions This is the kth frequency component.

[0038] It should be noted that first, the smart meters, distribution automation terminals and monitoring sensors deployed by the power company in the transformer area collect historical load data of each load point in the transformer area, including active power, reactive power, voltage, current and other multi-dimensional parameters with a sampling interval of minutes or 15 minutes. By time series processing of the historical load data and combining the transformer area topology structure information, the distribution transformers, feeders, branches and user nodes in the transformer area are abstracted as nodes in the graph model, and the electrical connection relationship between the nodes is abstracted as edges in the graph model, so as to obtain the transformer area load graph model. The load graph model can not only reflect the electrical topology relationship of each node, but also carry the load time sequence information corresponding to each node, realize unified modeling of spatial topology and time sequence, and provide a basis for subsequent dynamic load feature analysis.

[0039] Secondly, for the historical load time sequence of each node in the transformer area load graph model, a fixed-length sliding window is used for traversal calculation, and the window length can be set to 24 hours, 48 hours or a week according to actual conditions, so as to balance short-term fluctuations and long-term trends. In each sliding window, the variance value of the load time sequence is calculated to represent the load fluctuation intensity in the time period. As the window slides along the time axis, a series of variance value sequences that change with time, i.e. the load fluctuation sequence of the node, can be obtained. The load fluctuation sequence can effectively describe the load fluctuation characteristics of the node in different time periods, and is the input of subsequent signal decomposition and feature extraction.

[0040] In this embodiment, the empirical mode decomposition algorithm is used to decompose the load fluctuation sequence of the node step by step. First, the upper and lower envelope lines are constructed by local extreme point interpolation, and the mean function is calculated, and the original fluctuation sequence is subtracted from the mean function to obtain the first order intrinsic mode function (IMF); then the above steps are iterated to extract IMF components of different time scales layer by layer until the remaining part is a monotonic function or has no more fluctuation components. Through the process, the load fluctuation sequence of each node is decomposed into several residual components of different scales, and each residual component corresponds to a fluctuation feature of different frequency range, so as to realize multi-scale decomposition of the load fluctuation characteristics.

[0041] In addition, for each node, the fluctuation amplitude and main frequency of the signal of each residual component are extracted. The fluctuation amplitude can be obtained by calculating the standard deviation or peak-to-peak value of the residual component, and the main frequency can be obtained by fast Fourier transform (FFT) or Hilbert spectrum analysis. After obtaining the amplitude and frequency of each residual component, they are spliced in a fixed order to form the load fluctuation feature vector of the node. The feature vector can comprehensively represent the dynamic fluctuation mode of the node at multiple scales, including the sensitivity of short-period disturbance and the stability of long-period load change.

[0042] Finally, the load fluctuation feature vectors of all nodes are collected to form a feature sample set, and an unsupervised clustering algorithm (such as K-means, DBSCAN or spectral clustering) is used for clustering analysis of the nodes. Through the clustering results, the nodes can be divided into several categories according to their load fluctuation characteristics, and each category of nodes has similar fluctuation patterns. Subsequently, the center vector of each category is taken as a reference benchmark to calculate the difference between the load fluctuation feature vector of each node in the category and the category center. Specifically, the Euclidean distance, Mahalanobis distance or cosine similarity can be used to measure the difference. The maximum value of the difference is defined as the maximum deviation value of the node, and the larger the deviation value, the more abnormal the load change of the node. Finally, the maximum deviation values of the nodes are normalized to obtain the first load change value, which is used to represent the load change sensitivity of the node and provide a basis for subsequent disturbance priority and path selection.

[0043] S2, load superposition disturbance is performed on the transformer area load diagram model to obtain a plurality of load superposition paths;

[0044] In this example, load superposition disturbance is performed on the transformer area load diagram model to obtain a plurality of load superposition paths, specifically:

[0045] According to the first load change value, a disturbance priority of the nodes in the transformer area load diagram model is constructed;

[0046] Based on the disturbance priority, random load fluctuation disturbance is applied to the nodes to obtain an initial disturbance scenario;

[0047] Under the initial disturbance scenario, a disturbance sequence is applied to the nodes on the transformer area load diagram model to obtain a plurality of load superposition paths.

[0048] It should be noted that first, the load sensitivity of each node in the transformer area load diagram model is quantified according to the first load change value obtained in the foregoing steps. The size of the first load change value is taken as the main basis, and the position weight of the node in the transformer area topology (for example, the end node of the feeder is more sensitive to overall voltage fluctuation) is combined to construct a disturbance priority sequence. The nodes with higher scores are selected first in the disturbance simulation, thereby establishing the disturbance priority in the transformer area load diagram model. The disturbance priority refers to the order and intensity reference standard of the nodes receiving load fluctuation disturbance in the subsequent disturbance application process, which reflects the dominant degree of the nodes in the overall transformer area load dynamics.

[0049] Further, according to the sorted disturbance priority, high-priority nodes are selected in turn, and random fluctuation disturbances are superimposed on the original load time series thereof. The random disturbance can be generated by using Gaussian noise, Poisson noise or fitting a random sequence based on a historical load deviation distribution, so as to ensure that the disturbance has randomness and conforms to the actual load fluctuation characteristics. The amplitude range of the disturbance can be set according to the first load change value of the node, for example, the upper limit of the disturbance amplitude is a certain percentage of the value. After the above operation, a global load disturbance distribution state, that is, an initial disturbance scenario, is formed in the transformer area load graph model. The initial disturbance scenario is essentially a graph model with random load disturbances, and can simulate the initial unstable state of the power grid under external environmental changes or user load fluctuations, thereby providing a starting point for subsequent path cascade disturbances.

[0050] Finally, the initial disturbance scenario is taken as a reference state, and then disturbance sequences are applied to nodes from high to low according to the disturbance priority order. The construction rules of the disturbance sequence include: (1) the disturbance form can be selected as a step type, a pulse type or a periodic type fluctuation to simulate different load change scenarios; (2) the disturbance amplitude is randomly selected according to the first load change value range of the node to ensure that the disturbance intensity is consistent with the node sensitivity; and (3) the disturbance duration can be set as a fixed time window or subject to an exponential distribution to simulate the uncertainty of the actual load fluctuation duration. During the application process, when a node is disturbed, the load change caused thereby is transmitted to adjacent nodes through the transformer area topology, thereby gradually forming a complete load superposition path. By repeating the above process, a path set formed by the spatial and temporal superposition of multiple disturbances, that is, a plurality of load superposition paths, can be obtained. These paths intuitively reflect the dynamic evolution process of the disturbance propagation from high-priority nodes to other nodes, thereby laying a data foundation for subsequent peak value identification and segmentation analysis.

[0051] S3, load peak values are identified for the plurality of load superposition paths, and continuous load superposition time periods are divided according to the identification results, to obtain a load superposition segmentation set;

[0052] In this example, load peak values are identified for the plurality of load superposition paths, and continuous load superposition time periods are divided according to the identification results, to obtain a load superposition segmentation set, which specifically includes:

[0053] Load time series of each node on each load superposition path are extracted, and a preset multi-scale peak detection algorithm is used for peak detection, to obtain a plurality of load peak points;

[0054] Time distribution of the plurality of load peak points is analyzed, and a dynamic time warping algorithm is used to align the peak value sequences of different load superposition paths, to obtain aligned peak value sequences;

[0055] Divide the load time sequence into continuous time periods based on the aligned peak sequence, to obtain a plurality of load superposition segments;

[0056] Merge the plurality of load superposition segments to obtain a load superposition segment set.

[0057] It should be noted that firstly, a plurality of load superposition paths formed by perturbation are obtained in the transformer area load diagram model, and the corresponding load time sequence is extracted node by node on each path. The load time sequence of each node is composed of its historical load data and the load change value after superposition perturbation, and is usually formed into a discrete sequence at a fixed sampling interval (such as 15 minutes or 1 hour). When extracting, first determine the time span of the path to ensure that all the node load data involved are aligned, and then add the original load data and the perturbation value of each node in the period point by point to obtain the complete load time sequence of the node. In this way, a multi-dimensional time sequence matrix covering all nodes on the path can be obtained, providing a basis for subsequent peak detection and alignment processing.

[0058] In addition, firstly, a preset multi-scale peak detection algorithm is used to process the load time sequence of each node, identify a plurality of peak points, and record the specific time and amplitude of the occurrence. Subsequently, statistical analysis is performed on the time distribution of all peak points, and the concentration degree, time interval distribution and cross-node synchronization of different peaks are calculated to judge the timing characteristics of load fluctuation. Since different load superposition paths may have misalignment or rhythm difference on the time axis, a dynamic time warping (DTW) algorithm is further used to align the peak sequences between the paths. The specific method is: calculate the cumulative distance matrix between any two peak sequences, find the optimal matching path through dynamic programming, so as to realize the adaptive alignment of different sequences in the time scale. After alignment, the peak sequence under the unified time reference is obtained, providing a standardized input for subsequent continuous period division.

[0059] Further, the load time sequence is divided into a plurality of continuous time periods based on the aligned peak sequence. The specific method is: taking the time interval between every two adjacent aligned peak points as the boundary, the load time sequence is divided into independent segments, each segment completely contains the rising, peak and falling process of a load peak. If no obvious peak point is detected in a certain period of time, the time period can be kept as a stable segment. In this way, a plurality of load superposition segments with clear structure and clear boundary can be obtained, each of which can reflect a relatively complete load fluctuation process.

[0060] Finally, in order to avoid too many segments and high complexity of subsequent analysis, the preliminary obtained load superposition segments need to be merged. The specific steps are as follows: first, the load characteristics of each segment are calculated, such as average load level, peak-valley difference, duration, etc., and the similarity of the characteristics between adjacent segments is calculated; if the difference between adjacent segments is lower than the preset threshold, it is determined that they belong to the same load fluctuation process, and they are merged into a longer segment. The merging operation can be iteratively performed until the difference between all segments is greater than the threshold. The final obtained load superposition segment set is a set of continuous time period sets that can represent the main load fluctuation mode, which not only retains the fluctuation characteristics but also reduces the redundant information, and lays a foundation for subsequent path screening and classification analysis.

[0061] In the present example, a preset multi-scale peak detection algorithm is used for peak detection to obtain a plurality of load peak points, specifically:

[0062] Set different time scale sliding windows, and calculate the load gradient change of the load time series of each node in each sliding window;

[0063] According to the analysis result, a plurality of first peak points are identified and multi-scale fusion is performed on the first peak points to obtain a plurality of load peak points.

[0064] It should be noted that in the specific implementation process, in order to capture the load fluctuation characteristics at different time scales, a plurality of sliding windows are first set on the load time series. The window size can be determined according to the operation characteristics of the power system and the sampling interval of the load data. For example, when the sampling interval is 15 minutes, short time scale windows (such as 2 hours, 6 hours), medium time scale windows (such as 24 hours, 48 hours), and long time scale windows (such as one week) can be set. Each window slides along the time axis with a fixed step (such as half the window length or one sampling interval) to fully cover the time series. Through the setting of such multi-scale sliding windows, both local rapid fluctuations and long-term trend fluctuations can be captured, providing a multi-level data perspective for subsequent gradient change analysis and peak detection.

[0065] Secondly, for the load time series of each node, the corresponding subsequence is extracted in each sliding window range, and then the gradient change of the subsequence is calculated. The specific steps include: first, the first-order difference of the load subsequence is obtained, and the load change rate sequence is obtained; second, the sign analysis is performed on the load change rate sequence, and the turning points from negative to positive or from positive to negative are marked; third, the change amplitude before and after the turning point is calculated, that is, the absolute value of the difference value, which is used to represent the rate strength of the load rise or fall; fourth, the turning points with larger change amplitude are screened out by setting a threshold, and the occurrence time and change trend (rise or fall) are recorded. This process is the load gradient change analysis, which can reveal the mutation characteristics and fluctuation trend of the load sequence at different time scales.

[0066] Further, the turning points obtained by the above gradient change analysis are used as candidate points, and then combined with the actual value of the load time series, the first peak point is further screened out. The specific method is: if the load value at a certain turning point is in the local maximum or minimum in the sliding window, and the gradient change amplitude exceeds the preset threshold, then the point is determined as the effective first peak point; for multiple adjacent candidate points, the extreme point with higher or lower load value is selected as the first peak point. In this way, a batch of representative first peak points can be identified from the windows of different time scales, which can reflect the significant fluctuation characteristics of the load at each scale.

[0067] Finally, since the sliding windows of different time scales may identify first peak points that overlap with each other or are close in time, multi-scale fusion processing is needed. The specific method is: first, arrange all the first peak points at different time scales in chronological order; second, calculate the interval between time adjacent points, if less than a preset fusion threshold (such as 1 hour), it is determined as the same peak event, and merged into one peak point; in the merging process, the weighted average method can be used to determine the final peak amplitude and time, and the weight can be determined by the window size or the gradient amplitude. After the above fusion process, redundant repeated detection points can be eliminated, and a set of independent and representative load peak points can be obtained. These load peak points as the final result can fully represent the fluctuation peak characteristics of the load at different time scales, and provide reliable input for subsequent segmentation and path screening.

[0068] S4, screening the load superposition path based on the load superposition segmentation set to obtain a first screening path set;

[0069] In this example, the load superposition path is screened based on the load superposition segmentation set to obtain a first screening path set, specifically:

[0070] Extract the load features of each load superposition segmentation, and construct a segmentation feature matrix based on the load features;

[0071] The distance metric between the load features is calculated based on the segmented feature matrix, and a spectral clustering algorithm is used to group the load superposition paths to obtain the grouped load superposition paths; the distance metric includes calculating Mahalanobis distance, Euclidean distance or cosine distance, and the Mahalanobis distance is preferred in this embodiment.

[0072] The grouped load superposition paths are divided into different path categories according to the Mahalanobis distance to obtain a first screening path set.

[0073] In this example, the calculation formula of the Mahalanobis distance between the load features based on the segmented feature matrix is as follows:

[0074]

[0075] wherein, is the Mahalanobis distance between the a-th load superposition segment and the b-th load superposition segment, and are the a-th and b-th load superposition segments in the segmented feature matrix, and are the load feature vector of the a-th segment and the load feature vector of the b-th segment, is the inverse of the covariance matrix, is the transpose symbol.

[0076] It should be noted that for each divided load superposition segment, the load time series thereof is statistically analyzed and pattern analyzed to extract an index capable of representing the characteristics of the segment, i.e., a load feature. The load features usually include but are not limited to the following categories: load level features (such as average load, maximum load, minimum load), fluctuation features (such as standard deviation, peak-valley difference, fluctuation frequency), morphological features (such as rising rate, falling rate, symmetry index), and time series features (such as peak value occurrence time, duration, etc.). These features comprehensively reflect the overall trend, fluctuation degree and time series characteristics of the load in the segment, and are the basic input for subsequent segment clustering and path screening.

[0077] Secondly, the load features of all load superposition segments are quantified into numerical vectors to ensure that each segment corresponds to a feature vector. Then these feature vectors are arranged one by one according to the segment number to form a two-dimensional matrix, i.e., a segmented feature matrix. The rows of the matrix represent different load superposition segments, the columns represent various load features (such as average load, standard deviation, peak time, etc.), and the matrix elements are the numerical values of the segments under the corresponding features. The segmented feature matrix can collectively express the feature distribution of all segments, and is the core data structure for distance measurement and clustering analysis.

[0078] Further, the segmented feature matrix is first normalized to eliminate the influence of different feature dimension scales. Subsequently, the Mahalanobis distance formula is used to calculate the similarity between each segment. This distance measure takes into account the correlation between features and is more capable of reflecting the differences between segments than the Euclidean distance. By calculating the Mahalanobis distance between any two segments, a symmetric distance matrix can be obtained, which provides input for subsequent spectral clustering.

[0079] Subsequently, the Mahalanobis distance matrix described above is used as input. First, it is converted into a similarity matrix (for example, by using a Gaussian kernel function for similarity mapping), and then a graph Laplacian matrix is constructed. Next, the Laplacian matrix is subjected to eigenvalue decomposition, and the eigenvectors corresponding to the first several smallest eigenvalues are extracted and used as the new embedding space representation. Finally, K-means or other clustering methods are used in this embedding space to cluster the load superposition segments. Since each load superposition path is composed of multiple segments, the clustering results can be mapped back to the path level, thereby dividing all load superposition paths into several groups, each group of paths having similar segment feature patterns, i.e., obtaining the grouped load superposition paths.

[0080] Finally, for the clustered and grouped load superposition paths, the intra-class compactness and inter-class difference are further calculated based on the Mahalanobis distance. If the average Mahalanobis distance of a group of paths is small, it means that the load behavior within the group is highly consistent and can be divided into the same path category. If the average Mahalanobis distance between different groups is large, it means that they represent different load behavior patterns. Ultimately, according to the clustering results and the Mahalanobis distance measurement, all paths are divided into several path categories, and the paths in each category represent a typical load evolution pattern. Collecting the typical paths in these categories that meet the research requirements, we obtain the first screening path set.

[0081] S5, screening the first screening path set based on the first load change value to obtain a second screening path set, and performing hierarchical load prediction based on the second screening path set.

[0082] In this example, the first screening path set is screened based on the first load change value to obtain a second screening path set, specifically:

[0083] Extract the load change trend feature value of each load superposition path in the first screening path set and analyze it with the first load change value;

[0084] According to the analysis result, an adaptive screening threshold is constructed, and a fuzzy comprehensive evaluation method is used to screen the first screening path set based on the adaptive screening threshold to obtain the second screening path set.

[0085] It should be noted that for each load superimposed path in the first screening path set, the corresponding time series is extracted, and the load change trend characteristic value is calculated by the trend analysis method. The specific method includes: using the least square method to fit the overall trend of the path load time series, and extracting the slope as the load change rate characteristic; the overall change amplitude characteristic is obtained by calculating the mean difference or cumulative change of the sequence; at the same time, the main fluctuation direction of the sequence (such as continuous rise, continuous decline or fluctuation type) is counted, and the trend direction characteristic is formed. Finally, the above rate, amplitude and directionality indicators are combined into the load change trend characteristic value of the path. The characteristic value can comprehensively reflect the dynamic change characteristics of the path load evolution with time, and is the key parameter for comparison with the first load change value.

[0086] Secondly, the load change trend characteristic value of each load superimposed path is associated with the first load change value calculated in advance in the feeder load graph model. The analysis method includes: on the one hand, the consistency of the change rate of the trend characteristic value and the maximum deviation value of the node is compared, which is used to judge whether the path reflects the main load change characteristics of the sensitive node; on the other hand, the correlation coefficient or similarity measure between the trend characteristic value and the first load change value is calculated, and the matching degree between the path load change and the overall sensitive load change of the feeder is evaluated.

[0087] Further, according to the matching degree of the path characteristic value and the first load change value, an adaptive screening threshold is dynamically constructed. The specific method is: first, the correlation coefficient or similarity distribution of all paths and the first load change value is counted; then, based on the mean and standard deviation of the distribution, a dynamic threshold function is constructed, for example, threshold = mean + λ × standard deviation, where λ is an adjustable parameter. In this way, the adaptive threshold can be automatically adjusted according to different data sets without fixed standards, so as to ensure the flexibility and adaptability of the screening conditions in different feeders or different running scenes.

[0088] In addition, with the adaptive screening threshold as the reference, a fuzzy comprehensive evaluation method is used to screen the first screening path set. The specific steps include: first, define multiple evaluation indexes, such as the degree of consistency of the load change rate with the first load change value, the degree of closeness of the load change amplitude with the maximum deviation value of the node, and the consistency of the overall trend direction; then, establish a fuzzy membership function for each index, and convert the quantitative index into a fuzzy evaluation grade (such as high, medium and low); then, assign weights to different indexes, and use weighted average method for fuzzy comprehensive operation to get the comprehensive membership degree of each path; finally, compare the comprehensive membership degree with the adaptive threshold, and the paths with membership degree greater than the threshold are retained, and the paths with membership degree less than the threshold are removed. Through the fuzzy comprehensive evaluation method, the path set that meets the screening conditions can be obtained, that is, the second screening path set.

[0089] Finally, the first screening path set is a candidate path set obtained based on the load superposition segmentation set and the path feature similarity preliminary screening, which has the characteristics of containing most typical path categories, but there may still be some paths that are not completely related to the key load change of the transformer area or have strong interference. The second screening path set is a refined set obtained by further combining the first load change value and performing secondary screening through adaptive threshold and fuzzy comprehensive evaluation based on the first screening path set. Compared with the first screening path set, the second screening path set is more representative and targeted, as it only retains the paths that are highly related to the sensitive node load change and can reflect the overall load evolution law of the transformer area, and thus is the core data input for hierarchical load forecasting.

[0090] In this example, hierarchical load forecasting is performed based on the second screening path set, specifically as follows:

[0091] Load sequences are extracted from the second screening path set, and a time series data set is constructed based on the load sequences;

[0092] The time series data set is divided into load grade intervals to obtain the load grade intervals;

[0093] A spatio-temporal graph neural network model is constructed based on the load grade intervals to perform load forecasting, and the hierarchical load forecasting result is output.

[0094] It should be noted that first, the node load data contained in each load superposition path is selected from the second screening path set, and the load sequence of each node is extracted in chronological order. In order to ensure data consistency, all sequences are aligned using a uniform sampling interval (such as 15 minutes or 1 hour), and missing values are interpolated and completed, and abnormal values are smoothed and corrected. Subsequently, these load sequences are combined according to the path dimension and the time dimension to construct a multi-dimensional time series data set, where the rows represent time steps, the columns represent load values of paths or nodes, and the matrix elements are load data at corresponding time points. This time series data set not only retains the time dynamic characteristics of the load, but also reflects the spatial correlation between paths, providing a data basis for the construction of subsequent spatio-temporal models.

[0095] Further, first, statistical analysis is performed on all load values in the time series data set to calculate the global maximum value, minimum value and distribution characteristics. Then, a hierarchical division method (such as equidistant division, equal frequency division or adaptive division based on clustering) is used to divide the load value interval into several levels, for example, it can be divided into "low load", "medium load", "high load" and "ultra-high load" four levels. Each level corresponds to a load interval, for example: 0-100kW is low load, 100-300kW is medium load, 300-600kW is high load, and 600kW or more is ultra-high load. Each load value in the time series data set is marked according to the interval it belongs to, and the labeled results of the load level interval are obtained to provide label data for subsequent hierarchical prediction.

[0096] Finally, based on the time series data set and the load level interval, a spatio-temporal graph neural network (ST-GNN) model is constructed. The specific method is as follows: first, the transformer model is used as the graph structure input, wherein the nodes represent the load points and the edges represent the electrical connection relationship; second, the historical load time series of each node is used as the input feature, the spatial dependence relationship between nodes is captured through the graph convolution network (GCN), and then the dynamic characteristics of the time series are captured through the time convolution or the gated recurrent unit (GRU / LSTM); finally, a classification module is set in the output layer to map the predicted load value to the corresponding load level interval. Cross-entropy loss or weighted loss function is used in the model training process to improve the prediction accuracy of high-level load. After training, the model can input historical load data to predict the load level in the future period. For example, the prediction result of a certain transformer station at 8am is "medium load (about 250kW)", the prediction result at 12pm is "high load (about 480kW)", and the prediction result at 7pm is "ultra-high load (about 710kW)". The prediction result can provide early warning and hierarchical management basis for power grid dispatching.

[0097] As shown in Figure 2 Fig. 1 is a schematic diagram of the device structure of the power grid hierarchical load prediction method. According to the functions realized, the power grid hierarchical load prediction device can include a model construction module, a disturbance superposition module, a peak value identification module, a path screening module, and a load prediction module. The modules of the present application can also be referred to as units, which refer to a series of computer program segments that can be executed by the processor of the personnel performance data processing device and can complete a fixed function. They are stored in the memory of the personnel performance data processing device.

[0098] In this embodiment, the functions of each module are as follows:

[0099] The model construction module is used to construct the transformer station load graph model and analyze the load change of the nodes of the transformer station load graph model to obtain the first load change value.

[0100] The disturbance superposition module is configured to perform load superposition disturbance on the transformer area load diagram model to obtain a plurality of load superposition paths.

[0101] The peak identification module is configured to perform load peak identification on the plurality of load superposition paths, and divide continuous load superposition time periods according to the identification result to obtain a load superposition segmentation set.

[0102] The path screening module is configured to perform first-level screening on the load superposition paths based on the segmentation set to obtain a first screening path set.

[0103] The load prediction module is configured to screen the first screening path set based on the first load change value to obtain a second screening path set, and perform hierarchical load prediction based on the second screening path set.

[0104] In detail, each module in the power grid hierarchical load prediction device in the embodiment of the present application adopts the same technical means as the power grid hierarchical load prediction method described in the accompanying drawings when in use, and can produce the same technical effects, which will not be described here.

[0105] As shown in Figure 3 , it is an electronic device structure schematic diagram of the power grid hierarchical load prediction method of the present application.

[0106] The electronic device can include a processor, a memory, a communication bus, and a communication interface, and can further include a computer program stored in the memory and executable on the processor, such as a power grid hierarchical load prediction program.

[0107] The processor is the control core (Control Unit) of the electronic device, which connects all components of the electronic device through various interfaces and lines, and executes or runs the programs or modules stored in the memory, and calls the data stored in the memory, to perform various functions and process data of the electronic device.

[0108] The memory includes at least one type of readable storage medium, and in some embodiments, the memory can be an internal storage unit of the electronic device, such as a mobile hard disk of the electronic device. The memory can be used to store application software and various data installed in the electronic device.

[0109] The communication bus is configured to realize the connection and communication between the memory and at least one processor.

[0110] The communication interface is used for communication between the above-mentioned electronic device and other devices, including a network interface and a user interface.

[0111] The electronic device with components is shown in the figure, and those skilled in the art can understand that the structure shown in the figure does not constitute a limitation on the electronic device, and can include fewer or more components than the figure, or combine certain components, or different component arrangements.

[0112] It should be understood that the embodiments are only for illustration, and the scope of the patent application is not limited by the structure.

[0113] The power grid hierarchical load prediction program stored in the memory in the electronic device is a combination of a plurality of instructions, which, when executed in the processor, can implement the steps in the power grid hierarchical load prediction method.

[0114] Specifically, the specific implementation system of the processor for the above instructions can refer to the description of the related steps in the corresponding embodiment of the accompanying drawings, which will not be described here.

[0115] The above formulas are all dimensionless numerical calculations, and the formulas are obtained by collecting a large amount of data to simulate the most recent real situation. The preset parameters in the formula are set by those skilled in the art according to the actual situation.

[0116] The above embodiments can be realized wholly or partially by software, hardware, firmware or any combination thereof. When realized by software, the above embodiments can be realized in the form of a computer program product, wholly or partially.

[0117] Those of ordinary skill in the art can realize that the modules and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0118] In addition, the functional modules in each embodiment of the present application can be integrated in one processing module, or each module can exist physically alone, or two or more modules can be integrated in one module.

[0119] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0120] Finally: the above only for the preferred embodiments of the present application, and not for limiting the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application, should be included in the scope of protection of the present application.

Claims

1. A method for hierarchical load forecasting of an electrical grid, characterized in that, The method comprises the following steps: constructing a load diagram model of a transformer area and analyzing load variation characteristics of nodes of the load diagram model to obtain a first load variation value; applying a load superimposition disturbance to the load diagram model of the transformer area to generate a plurality of load superimposition paths: determining a disturbance priority of a node in the load diagram model of the transformer area according to the first load variation value; applying a random load fluctuation disturbance to the corresponding node based on the disturbance priority to generate an initial disturbance scenario; under the initial disturbance scenario, applying a disturbance sequence to the load diagram model of the transformer area according to a preset rule to generate a plurality of load superimposition paths; identifying load peak value characteristics of the superimposition paths, and dividing continuous time periods based on the peak value characteristics to form a load superimposition segmentation set; performing primary screening on the load superimposition paths based on the segmentation set to obtain a first screening path set; performing secondary screening on the first screening path set based on the first load variation value to obtain a second screening path set, and performing hierarchical load prediction based on the second screening path set.

2. The power grid hierarchical load forecasting method according to claim 1, characterized in that, The method of constructing a load diagram model of a transformer area and analyzing load variation characteristics of nodes of the load diagram model to obtain a first load variation value comprises the following steps: collecting historical load data of load points in a transformer area to construct a load diagram model of the transformer area; calculating the sliding window variance of the load time series of each node in the load diagram model to obtain a load fluctuation sequence; performing modal decomposition on the load fluctuation sequence to obtain a plurality of node residual components; constructing a load fluctuation feature vector based on the fluctuation amplitude and frequency of the node residual components; performing clustering analysis on the load fluctuation feature vectors of all nodes, calculating the maximum load deviation value of the nodes based on the clustering results, and obtaining the first load variation value.

3. The power grid hierarchical load forecasting method according to claim 2, characterized in that, The method of identifying load peak value characteristics of the superimposition paths and dividing continuous time periods based on the peak value characteristics to form a load superimposition segmentation set comprises the following steps: performing multi-scale peak value detection on the load time series of the nodes on each load superimposition path to obtain load peak points; analyzing the time distribution of the load peak points, and aligning the peak value sequences of different load superimposition paths using a dynamic time warping algorithm to obtain aligned peak value sequences; dividing the load time series into continuous time periods based on the aligned peak value sequences to obtain load superimposition segments; merging the load superimposition segments to form a load superimposition segmentation set.

4. The power grid hierarchical load forecasting method according to claim 3, characterized in that, The method of performing multi-scale peak value detection on the load time series of the nodes on each load superimposition path to obtain load peak points comprises the following steps: setting sliding windows of different time scales, and calculating the load gradient change of the load time series of each node in each sliding window; identifying a plurality of first peak points based on the analysis results, and performing multi-scale fusion on the first peak points to obtain a plurality of load peak points.

5. The power grid hierarchical load forecasting method of claim 4, wherein, The method of performing primary screening on the load superimposition paths based on the segmentation set to obtain a first screening path set comprises the following steps: extracting load features of each load superimposition segment, and constructing a segment feature matrix based on the load features; calculating the distance measure between the features based on the segment feature matrix, and grouping the load superimposition paths using a spectral clustering algorithm; dividing the grouped load superimposition paths into different path categories based on the distance measure to obtain the first screening path set.

6. The power grid hierarchical load forecasting method according to claim 5, characterized in that, The method of performing secondary screening on the first screening path set based on the first load variation value to obtain a second screening path set comprises the following steps: Extract the load change trend characteristic value of each load superposition path in the first screening path set, and analyze it with the first load change value; According to the analysis result, an adaptive screening threshold is constructed, and a fuzzy comprehensive evaluation method is used based on the adaptive screening threshold to screen the first screening path set, to obtain a second screening path set.

7. The power grid hierarchical load forecasting method according to claim 6, characterized in that, The hierarchical load prediction based on the second screening path set is specifically: Extract the load sequence from the second screening path set, and construct a time series data set based on the load sequence; Divide the time series data set into load grade intervals; Based on the load grade interval, a spatio-temporal graph neural network model is constructed to predict the load, and the hierarchical load prediction result is output.

8. An apparatus for using the grid staging load forecasting method according to any one of claims 1 to 7, characterized by, The method comprises a model construction module, a disturbance superposition module, a peak value identification module, a path screening module, and a load prediction module: The model construction module is used to construct a substation load graph model and analyze the load change characteristics of its nodes to obtain a first load change value; The disturbance superposition module is used to apply load superposition disturbance to the load graph model to generate a plurality of load superposition paths; The peak value identification module identifies the load peak value characteristics of the superposition paths, and divides the continuous time periods based on the peak value characteristics to form a load superposition segmentation set; The path screening module is used to perform primary screening on the load superposition paths based on the segmentation set to obtain a first screening path set; The load prediction module is used to perform secondary screening on the first screening path set based on the first load change value to obtain a second screening path set, and perform hierarchical load prediction based on the second screening path set.

9. An electronic device, comprising: The electronic device comprises: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program is executed by the at least one processor, so that the at least one processor can execute the power grid hierarchical load prediction method of any one of claims 1 to 8.

Images