Regional load prediction method for power distribution network
By using data preprocessing, clustering classification and load prediction methods in the distribution network, the problem of existing technology being difficult to cope with areas with high fluctuations in the power consumption load is achieved, and more accurate load prediction and more scientific power supply planning are achieved.
Patent Information
- Application Number
- CN202411766434.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-04
- Publication Date
- 2025-04-25
AI Technical Summary
The existing distribution network load prediction methods are difficult to effectively deal with specific areas with large fluctuations in electricity loads. The time series method requires high data stability, while the regression prediction algorithm is insufficient in variable analysis and processing.
A distribution network area load prediction method is adopted, including data acquisition and preprocessing, data clustering classification and load prediction based on the cassia group algorithm. The specific steps include: acquiring and preprocessing the original electricity consumption data, dividing the electricity consumption data cluster through clustering processing, and using the sea squirt group algorithm to predict the load.
This method can more accurately predict the load state of the distribution network area, especially in areas with high load fluctuations, which improves the accuracy and effectiveness of the prediction, helps the distribution network plan the power supply plan in advance, avoid oversupply and interrupt supply, and saves power resources.
Smart Images
Figure CN119886855A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of distribution network load management, and in particular relates to a distribution network regional load forecasting method. Background Art
[0002] Distribution network load forecasting is a frequently-involved topic in the daily safe and economic operation and management of power systems. By predicting and analyzing the future regional load status of distribution networks, it is helpful to plan the power supply plan of distribution networks in advance, make the regional load distribution of distribution networks more reasonable and scientific, avoid unnecessary oversupply and outages, save electric energy resources, and improve the operation stability of distribution systems. At present, there are many types of distribution network load forecasting methods, including time series method, regression prediction algorithm, etc. Among them, the time series method has high requirements for the stability of distribution network operation data, and it is difficult to cope with the load analysis work in specific areas with strong load fluctuations. The regression prediction algorithm is easy to implement, but has high requirements for the analysis and processing of variables, and is suitable for mature and effectively verified distribution network models. Summary of the invention
[0003] The object of the present invention is to provide a load forecasting method based on actual needs, which can effectively analyze the load status of the distribution network area in the power supply area with large power load fluctuation.
[0004] To achieve the above purpose, the present invention adopts the following technical solution.
[0005] A method for predicting regional load in a distribution network comprises the following steps:
[0006] Step 1: Acquisition and preprocessing of original power consumption data in the power grid area;
[0007] a1. Collecting raw power consumption data in the distribution network area, the raw power consumption data has a time period of not less than six months, preferably not less than one year; the raw power consumption data includes distribution network user load data, peak voltage data, user power load type data;
[0008] a2. Raw data preprocessing, including identification and correction of a small amount of discrete data and missing data in the raw data, standardization and dimensionality reduction;
[0009] Step 2: Clustering and classification of original data of power grid area;
[0010] For different types of electricity consumption, different regions and even different time periods, there are usually large differences in electricity consumption data in the corresponding regions, which makes the electricity consumption data appear in different similar clusters. The samples between different clusters are very different, and the sample data in the same cluster are highly similar. By independently analyzing and predicting the electricity consumption data of different clusters, it is possible to better obtain the prediction results in the corresponding region and obtain more accurate prediction results in the entire power supply cycle. The specific division of the aforementioned clusters is determined by clustering processing in this application. Specifically:
[0011] s1: Calculate the normalized Laplacian matrix LN.
[0012] For the sample data {v1,v2...v n}, and abstractly express it as an undirected weighted network graph G = (V, E), where the elements in the point set V are sample data values, and the elements in the edge set E represent the affinity of the sample data corresponding to the two endpoints of the edge;
[0013] For an undirected weighted network graph G = (V, E), its point v i and v j The edge E between ij The weight w ij =w ji , the weights of each point form the adjacency matrix W; get the degree matrix of each point in
[0014] Further, the unnormalized Laplace matrix L=DW is obtained;
[0015] Construct the normalized Laplacian matrix Where I = H T H, Among them A j refers to the sample data {v1,v2...v n}The jth cluster of cluster classification;
[0016] Based on the spectral clustering algorithm, calculate the unnormalized Laplace matrix L N The eigenvectors f corresponding to the first k eigenvalues of are combined and standardized to obtain the feature matrix F. Each row in the feature matrix F is used as a multidimensional sample and clustered using the traditional clustering algorithm to obtain the final clustering result.
[0017] Step 3: Load forecasting based on salp swarm algorithm;
[0018] The steps include:
[0019] Data sorting: Obtain historical load data, divide it into training data and validation data, and save them in matrix form as training set and data set respectively;
[0020] Set parameters: Set the initial parameters of the salp swarm algorithm;
[0021] The load data stored in the above matrix form is expressed as the position coordinates of the individual, and the closest prediction result value is used as the target position coordinates of the best individual of the algorithm, where each individual can be expressed as the following position matrix Where m is the population size, n is the optimization target dimension, and each column in X represents the position coordinates of an individual; the individual fitness matrix can be expressed as Fitness function
[0022] Define five types of individuals: Type I Leader, Cross-Follower, Type II Leader, Follower, and Mutant;
[0023] One type of leader is used to complete normal process walk, crossover and selection operations;
[0024] Its normal process walk process can be expressed as: Where X1(t) refers to the transition position generated by the platform process in the tth iteration; where X1′(t) represents the coordinates of the type 1 leader at the tth iteration, randn(n,1) represents the generation of n rows and a single column of normally distributed random numbers, ub refers to the upper boundary of the individual search area, and lb refers to the lower boundary of the individual search area; represents the Hadamard product;
[0025] The cross-operation process of a Type 1 leader can be expressed as: Where X r (i) is the i-th dimension coordinate after the intersection of the type I leader body and the transition position; rand is a uniformly distributed random number between 0 and 1, and CR is the crossover probability;
[0026] The selection process of a type 1 leader can be expressed as: X′ refers to the coordinate after the intersection of the type 1 leader’s body and the transition position;
[0027] The crossover follower performs crossover and selection operations on itself and a randomly selected type I leader for iterative updates;
[0028] Type II leaders are individuals generated with the current best position as the center, and the position coordinates can be expressed as:
[0029]
[0030] Where X best is the current best position; R n.1A and B refer to vectors with n rows and 1 column, and the values of the elements in the vector are all 1; C refers to vectors with n rows and 1 column, and the values of the elements in the vector are all 1 or -1; i refers to a uniformly distributed random number between 0 and 1; k1 is a constant that controls the search range, and mod(l,w) refers to the remainder of l divided by w, where l is the number of iterations and w is the change period;
[0031] Follower is the follower in the salp swarm algorithm;
[0032] The mutant mutates iteratively from the initial position of the mutant to the direction of the type II leader based on the DE algorithm, which can be expressed as: X3 = X 2.1 +k3(X 2.2 -X 2.3 ); where X 2.1 ,X 2.2 ,X 2.3 refers to the randomly selected type II leader, and k3 is the scaling factor;
[0033] Based on the aforementioned individuals and iteration methods, sub-individuals are continuously generated, the fitness of the sub-individuals is calculated, the sub-individuals corresponding to the best position are found, and the algorithm parameters are optimized using the training set and data set until the iteration ends and the optimal parameters are output; the measured values are substituted into the algorithm to determine the load forecasting results.
[0034] Further improvement or specific implementation steps of the above-mentioned distribution network regional load forecasting method, the discrete data identification and repair specifically refers to: statistics the original data separately according to the power supply cycle, for a certain type of original data sequence {l1,l2...l n}, calculate the average value L mid And covariance S, and calculate the data distance corresponding to the original data sequence Analyze and calculate the distance threshold G of the original data sequence L L =m(D i )+k*s(D i );m(·) refers to the mean, s(·) refers to the standard deviation, and k is the adjustment coefficient; compare D i With G L The size of D i >G L The corresponding original data is discrete data;
[0035] The discrete data are extracted respectively. If they are independent non-endpoint discrete data, they are replaced by the average value of the values obtained in the previous and next acquisition cycles. The replaced value can be expressed as l i =0.5(l i+1 +l i-1); if continuous discrete data is found or the discrete data appears at both ends of the periodic data, the weight is used to modify and replace it. The value after replacement can be expressed as Among them l i.t refers to the original data obtained at the i-th time node in the t-th acquisition cycle, where ω i refers to l t The cycle weight of the collection cycle is closer to the current collection cycle, the closer the collection cycle is ... i The larger the
[0036] Further improvement or specific implementation steps of the above-mentioned distribution network regional load forecasting method, the standardized processing specifically refers to: processing the original data in a standardized processing manner, for l i , the normalized value can be expressed as Among them l max represents the maximum value of the sample, l min Indicates the minimum sample value.
[0037] Further improvement or specific implementation steps of the above-mentioned distribution network regional load forecasting method, the dimension reduction processing refers to: for the data sequence L = {l1,l2...l n}, assuming that the corresponding data sequence after dimensionality reduction is C = {c1, c2....c m};but where c k Indicates that the data sequence is the kth element in C. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 It is a schematic diagram of the clustering and classification process of raw data in the power grid area. DETAILED DESCRIPTION
[0039] The present invention is described in detail below in conjunction with specific embodiments.
[0040] The distribution network regional load forecasting method of the present application is mainly used for the forecasting and analysis of the load status of the distribution network region, especially for areas where the load status changes greatly within a cycle, which can obtain better forecasting results and is conducive to improving the accuracy of distribution network load analysis.
[0041] The distribution network regional load forecasting method of the present application mainly includes the following steps:
[0042] Step 1: Acquisition and preprocessing of raw power consumption data in the power grid area
[0043] a1. Collecting raw power consumption data in the distribution network area, the raw power consumption data has a time period of not less than six months, preferably not less than one year; the raw power consumption data includes distribution network user load data, peak voltage data, user power load type data;
[0044] a2. Raw data preprocessing, including identification and correction of a small amount of discrete data and missing data in the raw data, standardization and dimensionality reduction;
[0045] Discrete Data Identification and Repair:
[0046] The original data are counted separately according to the power supply cycle. For a certain type of original data sequence {l1,l2...l n}, calculate the average value L mid And covariance S, and calculate the data distance corresponding to the original data sequence Analyze and calculate the distance threshold G of the original data sequence L L =m(D i )+k*s(D i );m(·) refers to the mean, s(·) refers to the standard deviation, and k is the adjustment coefficient; compare D i With G L The size of D i >G L The corresponding original data is discrete data;
[0047] The discrete data are extracted respectively. If they are independent non-endpoint discrete data, they are replaced by the average value of the values obtained in the previous and next acquisition cycles. The replaced value can be expressed as l i =0.5(l i+1 +l i-1 ); if continuous discrete data is found or the discrete data appears at both ends of the periodic data, the weight is used to modify and replace it. The value after replacement can be expressed as Among them l i.t refers to the original data obtained at the i-th time node in the t-th acquisition cycle, where ω i refers to l t The cycle weight of the collection cycle is closer to the current collection cycle, the closer the collection cycle is ... i The larger the
[0048] Standardization processing: Since the value ranges of various types of raw data vary greatly, in order to facilitate centralized processing and keep each type of data within a relatively valuable range, the raw data is processed using a standardized processing method. i , the normalized value can be expressed as Among them l maxrepresents the maximum value of the sample, l min Indicates the minimum value of the sample;
[0049] Dimensionality reduction processing: Since different types of raw data have different collection frequencies and the frequency differences are large, in the actual analysis process, it is often not necessary to analyze the data collected at any time, but now it is necessary to perform corresponding analysis on the data within a certain section. In order to achieve the above purpose and reduce the demand for data processing, a dimensionality reduction processing method is adopted. For the data sequence L = {l1,l2...l n}, assuming that the corresponding data sequence after dimensionality reduction is C = {c1, c2....c m};but where c k Indicates that the data sequence is the kth element in C;
[0050] Step 2: Clustering and classification of raw data in power grid areas
[0051] For different types of electricity consumption, different regions and even different time periods, there are usually large differences in electricity consumption data in the corresponding regions, which makes the electricity consumption data appear in different similar clusters. The samples between different clusters are very different, and the sample data in the same cluster are highly similar. By independently analyzing and predicting the electricity consumption data of different clusters, it is possible to better obtain the prediction results in the corresponding region and obtain more accurate prediction results in the entire power supply cycle. The specific division of the aforementioned clusters is determined by clustering processing in this application. Specifically:
[0052] s1: Calculate the normalized Laplacian matrix LN.
[0053] For the sample data {v1,v2...v n}, and abstractly express it as an undirected weighted network graph G = (V, E), where the elements in the point set V are sample data values, and the elements in the edge set E represent the affinity of the sample data corresponding to the two endpoints of the edge;
[0054] For an undirected weighted network graph G = (V, E), its point v i and v j The edge E between ij The weight w ij =w ji , the weights of each point form the adjacency matrix W; get the degree matrix of each point in
[0055] Further, the unnormalized Laplace matrix L=DW is obtained;
[0056] Construct the normalized Laplacian matrix Where I = H T H, Among them A j refers to the sample data {v1,v2...v n}The jth cluster of cluster classification;
[0057] Based on the spectral clustering algorithm, calculate the unnormalized Laplace matrix L N The eigenvectors f corresponding to the first k eigenvalues of are combined and standardized to obtain the feature matrix F. Each row in the feature matrix F is used as a multidimensional sample and clustered using the traditional clustering algorithm to obtain the final clustering result.
[0058] Step 3: Load forecasting based on salp swarm algorithm
[0059] The steps include:
[0060] Data sorting: Obtain historical load data, divide it into training data and validation data, and save them in matrix form as training set and data set respectively;
[0061] Set parameters: Set the initial parameters of the salp swarm algorithm;
[0062] The load data stored in the above matrix form is expressed as the position coordinates of the individual, and the closest prediction result value is used as the target position coordinates of the best individual of the algorithm, where each individual can be expressed as the following position matrix Where m is the population size, n is the optimization target dimension, and each column in X represents the position coordinates of an individual;
[0063] The individual fitness matrix can be expressed as Fitness function
[0064]
[0065] Define five types of individuals: Type I Leader, Cross-Follower, Type II Leader, Follower, and Mutant;
[0066] One type of leader is used to complete normal process walk, crossover and selection operations;
[0067] Its normal process walk process can be expressed as: Where X1(t) refers to the transition position generated by the platform process in the tth iteration; where X1′(t) represents the coordinates of the type 1 leader at the tth iteration, randn(n,1) represents the generation of n rows and a single column of normally distributed random numbers, ub refers to the upper boundary of the individual search area, and lb refers to the lower boundary of the individual search area; represents the Hadamard product;
[0068] The cross-operation process of a Type 1 leader can be expressed as: Where Xr (i) is the i-th dimension coordinate after the intersection of the type I leader body and the transition position; rand is a uniformly distributed random number between 0 and 1, and CR is the crossover probability;
[0069] The selection process of a type 1 leader can be expressed as: X′ refers to the coordinate after the intersection of the type 1 leader’s body and the transition position;
[0070] The crossover follower performs crossover and selection operations on itself and a randomly selected type I leader for iterative updates;
[0071] Type II leaders are individuals generated with the current best position as the center, and the position coordinates can be expressed as:
[0072]
[0073] Where X best is the current best position; R n.1 A and B refer to vectors with n rows and 1 column, and the values of the elements in the vector are all 1; C refers to vectors with n rows and 1 column, and the values of the elements in the vector are all 1 or -1; i refers to a uniformly distributed random number between 0 and 1; k1 is a constant that controls the search range, and mod(l,w) refers to the remainder of l divided by w, where l is the number of iterations and w is the change period;
[0074] Follower is the follower in the salp swarm algorithm;
[0075] The mutant mutates iteratively from the initial position of the mutant to the direction of the type II leader based on the DE algorithm, which can be expressed as: X3 = X 2.1 +k3(X 2.2 -X 2.3 ); where X 2.1 ,X 2.2 ,X 2.3 refers to the randomly selected type II leader, and k3 is the scaling factor;
[0076] Based on the aforementioned individuals and iteration methods, sub-individuals are continuously generated, the fitness of the sub-individuals is calculated, the sub-individuals corresponding to the best position are found, and the algorithm parameters are optimized using the training set and data set until the iteration ends and the optimal parameters are output; the measured values are substituted into the algorithm to determine the load forecasting results.
[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.
Claims
1. A method for predicting regional load in a distribution network, characterized in that: The steps include: Step 1: Acquisition and preprocessing of raw power consumption data in the power grid area a1. Collecting raw power consumption data in the distribution network area, the raw power consumption data has a time period of not less than six months, preferably not less than one year; the raw power consumption data includes distribution network user load data, peak voltage data, user power load type data; a2. Raw data preprocessing, including identification and correction of a small amount of discrete data and missing data in the raw data, standardization and dimensionality reduction; Step 2: Clustering and classification of raw data in power grid areas For different types of electricity consumption, different regions and even different time periods, there are usually large differences in electricity consumption data in the corresponding regions, which makes the electricity consumption data appear in different similar clusters. The samples between different clusters are very different, and the sample data in the same cluster are highly similar. By independently analyzing and predicting the electricity consumption data of different clusters, it is possible to better obtain the prediction results in the corresponding region and obtain more accurate prediction results in the entire power supply cycle. The specific division of the aforementioned clusters is determined by clustering processing in this application. Specifically: s1: Calculate the normalized Laplacian matrix; For the sample data {v1,v2...v n }, and abstractly express it as an undirected weighted network graph G = (V, E), where the elements in the point set V are sample data values, and the elements in the edge set E represent the affinity of the sample data corresponding to the two endpoints of the edge; For an undirected weighted network graph G = (V, E), its point v i and v j The edge E between ij The weight w ij =w ji , the weights of each point form the adjacency matrix W; get the degree matrix of each point in Further, the unnormalized Laplace matrix L=DW is obtained; Construct the normalized Laplacian matrix Where I = H T H, Among them A j refers to the sample data {v1,v2...v n }The jth cluster of cluster classification; Based on the spectral clustering algorithm, calculate the unnormalized Laplace matrix L N The eigenvectors f corresponding to the first k eigenvalues of are combined and standardized to obtain the feature matrix F. Each row in the feature matrix F is used as a multidimensional sample and clustered using the traditional clustering algorithm to obtain the final clustering result. Step 3: Load forecasting based on salp swarm algorithm The steps include: Data sorting: Obtain historical load data, divide it into training data and validation data, and save them in matrix form as training set and data set respectively; Set parameters: Set the initial parameters of the salp swarm algorithm; The load data stored in the above matrix form is expressed as the position coordinates of the individual, and the closest prediction result value is used as the target position coordinates of the best individual of the algorithm, where each individual can be expressed as the following position matrix Where m is the population size, n is the optimization target dimension, and each column in X represents the position coordinates of an individual; the individual fitness matrix can be expressed as Fitness function Define five types of individuals: Type I Leader, Cross-Follower, Type II Leader, Follower, and Mutant; One type of leader is used to complete normal process walk, crossover and selection operations; Its normal process walk process can be expressed as: Where X1(t) refers to the transition position generated by the platform process in the tth iteration; where X1′(t) represents the coordinates of the type 1 leader at the tth iteration, randn(n,1) represents the generation of n rows and a single column of normally distributed random numbers, ub refers to the upper boundary of the individual search area, and lb refers to the lower boundary of the individual search area; represents the Hadamard product; The cross-operation process of a Type 1 leader can be expressed as: Where X r (i) is the i-th dimension coordinate after the intersection of the type I leader body and the transition position; rand is a uniformly distributed random number between 0 and 1, and CR is the crossover probability; The selection process of a type 1 leader can be expressed as: X′ refers to the coordinate after the intersection of the type 1 leader’s body and the transition position; The crossover follower performs crossover and selection operations on itself and a randomly selected type I leader for iterative updates; Type II leaders are individuals generated with the current best position as the center, and the position coordinates can be expressed as: Where X best is the current best position; R n.1 A and B refer to vectors with n rows and 1 column, and the values of the elements in the vector are all 1; C refers to vectors with n rows and 1 column, and the values of the elements in the vector are all 1 or -1; i refers to a uniformly distributed random number between 0 and 1; k1 is a constant that controls the search range, and mod(l,w) refers to the remainder of l divided by w, where l is the number of iterations and w is the change period; Follower is the follower in the salp swarm algorithm; The mutant mutates iteratively from the initial position of the mutant to the direction of the type II leader based on the DE algorithm, which can be expressed as: X3 = X 2.1 +k3(X 2.2 -X 2.3 ); where X 2.1 ,X 2.2 ,X 2.3 refers to the randomly selected type II leader, and k3 is the scaling factor; Based on the aforementioned individuals and iteration methods, sub-individuals are continuously generated, the fitness of the sub-individuals is calculated, the sub-individuals corresponding to the best position are found, and the algorithm parameters are optimized using the training set and data set until the iteration ends and the optimal parameters are output; the measured values are substituted into the algorithm to determine the load forecasting results.
2. The method for regional load forecasting of distribution network according to claim 1, characterized in that: The discrete data identification and repair specifically refers to: performing statistics on the original data according to the power supply cycle, for a certain type of original data sequence {l1,l2...l n }, calculate the average value L mid And covariance S, and calculate the data distance corresponding to the original data sequence Analyze and calculate the distance threshold G of the original data sequence L L =m(D i )+k*s(D i );in m(·) refers to the mean, s(·) refers to the standard deviation, and k is the adjustment coefficient; compare D i With G L The size of D i >G L The corresponding original data is discrete data; The discrete data are extracted respectively. If they are independent non-endpoint discrete data, they are replaced by the average value of the values obtained in the previous and next acquisition cycles. The replaced value can be expressed as l i =0.5(l i+1 +l i-1 ); if continuous discrete data is found or the discrete data appears at both ends of the periodic data, the weight is used to modify and replace it. The value after replacement can be expressed as Among them l i.t refers to the original data obtained at the i-th time node in the t-th acquisition cycle, where ω i refers to l t The cycle weight of the collection cycle is closer to the current collection cycle, the closer the collection cycle is ... i The larger the 3. The method for regional load forecasting of distribution network according to claim 2, characterized in that: The standardization processing specifically refers to: processing the original data in a standardized way, i , the normalized value can be expressed as Among them l max represents the maximum value of the sample, l min Indicates the minimum sample value.
4. The method for regional load forecasting of distribution network according to claim 3, characterized in that: The dimensionality reduction process refers to: for the data sequence L = {l1,l2...l n }, assuming that the corresponding data sequence after dimensionality reduction is C = {c1, c2....c m };but where c k Indicates that the data sequence is the kth element in C.