CN-122021988-A - Multiple periodic load prediction method based on combined event relation network

CN122021988ACN 122021988 ACN122021988 ACN 122021988ACN-122021988-A

Abstract

The application provides a multi-periodic load prediction method based on a combined event relation network, aiming at the problem of multi-dimensional prediction of service load of a national network cloud platform, comprising the steps of firstly extracting load information from a historical operation log of the national network cloud, extracting multi-dimensional periodic characteristic weights on the basis of a Prophet model, secondly extracting independent events and constructing a correlation diagram structure network thereof, mapping pooled vectors into combined effect coefficients through a GNN analysis diagram structure context relation, and predicting load change when the combined events occur, and thirdly, predicting mutation through residual analysis based on deviation indexes of prediction results, and optimizing the model. Finally, the effect of long-term macroscopic prediction trend and short-term microscopic mutation prevention on the service load of the national network cloud platform is achieved.

Inventors

GONG SHUAI
YUE ZHONGWEN
SONG YUHANG
HONG TAO
Zou Dikai
LIANG BIYAN
WANG YU
WANG XUYU
TAO JUN
WANG CHAO
YIN XIAOYU
LI MING
LIU MUYAO
ZHANG MIN
WANG HAILU
CHEN MEI
CAO WANWAN
YU DONGBO

Assignees

国网安徽省电力有限公司信息通信分公司
东南大学

Dates

Publication Date: 20260512
Application Date: 20251215

Claims (10)

1. A method for multiple periodic load prediction based on a combined event relationship network, the method comprising the steps of: Firstly, extracting load information from a national network cloud history operation log, and extracting a multi-dimensional periodic characteristic weight on the basis of a Prophet model; Secondly, extracting independent events, constructing a correlation graph structure network, mapping the pooled vectors into a combined effect coefficient through MLP (maximum likelihood) by analyzing the context relation of the graph structure through GNN, and predicting the load change when the combined events occur; and thirdly, predicting mutation through residual analysis based on the deviation index of the prediction result, and optimizing the model.
2. The method for multiple periodic load prediction based on a combined event relationship network according to claim 1, wherein in the first step, time series data in a historical running load log of a national network cloud is extracted, and the first 80% of the time series data is passed through based on a propset model The change points Equally spaced into linear segments for each point in time Creating a length of Vector of (3) Determining the time of each element representation according to the formula Whether or not to tag the time period after each change point, after the corresponding change point: Wherein the method comprises the steps of As candidate change points , Is that The j-th component of (b).
3. The method for predicting multiple periodic loads based on the combined event relation network according to claim 2, wherein the method is characterized in that data of various long-term trends (business increase and seasonal capacity expansion) and abrupt trends (temporary task scheduling and abrupt faults) exist in the load change of the national network cloud, trend items are calculated in a segmented mode according to a formula, and grid load trend turning is fitted: Wherein the method comprises the steps of Is an initial slope, the trend term is a combination of a number of linear segments, the change points are the junctions of these segments, the slope of the trend allows an adjustment at each change point, Is at the point of change A vector of slope adjustment values at; Is the intercept; Is a vector of offset adjustment amounts for ensuring that the trend term function is continuous at the change point, in order for the function to be at the change point Where continuous, i.e. The values approximated from the left and right sides are equal and must be matched Constraint at the point of change Where the slope suddenly changes To compensate for this jump and to make the curve continuous, the intercept must be adjusted I.e. Will be Substituting the original formula to obtain a form which is more convenient to calculate: According to the formula, the slope change quantity Applying the laplace a priori distribution avoids excessive sensitivity to noise data when fitting grid load changes: Wherein the method comprises the steps of Is a scale parameter, if The larger the a priori distribution is, the flatter, allowing Take larger value and change trend more flexibly, if Smaller, the prior distribution is more concentrated and forced Towards 0, the trend is smoother.
4. A method of multiple periodic load prediction based on a combined event relationship network according to claim 3, wherein a fourier series is used to approximate load fluctuations at different periods in the actual change of grid load, for a given period according to the formula For each point in time Generating The method is characterized in that: The periodic effect is modeled according to a formula as a linear combination of its eigenvectors: According to the formula pair Performing exponential conversion to ensure that the value is always positive, and obtaining multiplication factors through conversion: Wherein the method comprises the steps of Is a weight coefficient vector which the Prophet model needs to learn, then minimizes an objective function comprising a loss function and a regularization term according to a formula, trains all unknown parameters : Wherein the method comprises the steps of Is at the point of time Is used for the data of the real data of (a), Is the model at the time point Based on parameters The predictions that are made, Is the regularization strength.
5. The method for predicting multiple periodic loads based on a combined event relation network according to claim 1, wherein in the second step, events which affect load change and lack periodicity are analyzed according to historical logs of a national network cloud platform, including temporary task scheduling, local faults and burst accesses, the events are defined as independent events, and independent event relevance combined relations are analyzed to construct a graph structure, namely, if event a and event b are independent events, a and b are respectively used as a node of the graph structure, if event a and event b can occur simultaneously (namely have combined relations) and affect loads, an edge for connecting the node a with the node b is added in the graph structure, the GNN graph neural network is divided into three layers, the first layer serves as an input layer, the independent events serve as inputs of the GNN graph neural network, the second layer and the third layer serve as hidden layers, and information of class 1 and class 2 neighbor nodes is respectively aggregated through a message passing mechanism according to a formula: firstly, according to a formula, centering a node Each of the neighboring nodes of (a) Generates a strip to be sent to Is a message of: Wherein the method comprises the steps of Representing neighbor nodes To the central node Is used for the message of (a), Is a learnable message function , Representing neighbor nodes In the first place The representation vector of the layer is a function of the layer, Representing a central node In the first place The representation vector of the layer is a function of the layer, Representing nodes And Feature vectors of the edges between them, and then aggregating all the slave neighbors according to the formula Sent message : Wherein the method comprises the steps of Representative node Is a set of all neighbor nodes; The method is a permutation-invariant aggregation function, namely an aggregation result is not changed along with the change of a neighbor input sequence, each independent event is generated into a unique corresponding embedded vector through a GNN network, and if the relation of a plurality of independent events with edges exists, the embedded vectors of the independent events with the combination relation are subjected to average pooling to generate a combined event embedded vector.
6. The method for predicting multiple periodic loads based on a combined event relation network according to claim 5, wherein the combined effect intensity coefficient is defined, the coefficient is a floating point number and is used for representing the influence intensity of the occurrence of the combined event on the power grid load, the combined effect intensity coefficient is obtained through MLP, the MLP network is divided into three layers, the first layer is an input layer, the combined event embedded vector and the training data set are used as the input of the MLP network, the second layer is a hidden layer, the data are subjected to nonlinear transformation through an activation function, the third layer is an output layer, and the combined effect intensity coefficient with a mapping relation with the combined event embedded vector is obtained through linear transformation.
7. The method for multiple periodic load prediction based on a combined event relationship network according to claim 6, wherein a part of combined effect intensity coefficients are manually labeled according to a formula as a training data set: Wherein the method comprises the steps of Representing the actual load value in the log, Is a trend term of the trend of the web, Is a multiplication factor which is a function of the multiplication factor, Is the intensity coefficient of the combined effect to be marked, defines the mean square error loss function Training a model according to a formula : Wherein the method comprises the steps of Is the number of training samples to be used, Is the first The true load value of the individual samples, Is the propset model at the time point The predicted trend term(s), Is the periodic multiplicative factor of propset model predictions, Obtaining embedded vector representation of independent event based on message transmission and aggregation mechanism of GNN, nonlinear transformation and linear mapping of the embedded vector via MLP network, and final output as combined effect intensity coefficient 。
8. The method for multiple periodic load prediction based on a combined event relationship network according to claim 7, wherein the load is predicted according to a formula Is a variation of (a): Wherein the method comprises the steps of Is a trend term of the trend of the web, Is a periodic multiplication factor that is a function of the periodic multiplication factor, Is the combined effect intensity coefficient; setting a desired threshold Calculating residual error of load prediction and actual result according to formula : Detecting periodic variation by using an autocorrelation function (ACF) on the residual, if the residual exceeds a threshold value and the autocorrelation function diagram has no obvious autocorrelation (periodicity), indicating that the model omits unknown and aperiodic independent events or event combinations in the power grid system, the diagram structure needs to be updated, the GNN and MLP networks are retrained, if the autocorrelation function diagram of the residual shows a certain period The related significant peak value shows that the existing periodic component of the Prophet model is incomplete, and one period is omitted In the Prophet model, at this time, needs to be added to And (3) recalculating multiplication factors as influence factors of the period to obtain more accurate prediction results.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor is capable of implementing a multiple periodic load prediction method based on a combined event relationship network as claimed in any of the preceding claims 1 to 8 when the program is executed.
10. A computer readable storage medium having stored thereon computer instructions which when executed by a processor implement a multiple periodic load prediction method based on a combined event relationship network as claimed in any of claims 1 to 8.

Description

Multiple periodic load prediction method based on combined event relation network Technical Field The invention relates to the field of intelligent power grid data prediction, in particular to a multiple periodic load prediction method based on a combined event relation network. Background In a national network cloud data system, the accuracy of load prediction is directly related to the efficiency of cloud resource scheduling and the stability of power business. As the power grid business migrates to the cloud primary architecture, the micro-service cluster scale continues to expand, the load change presents complex characteristics of multiple periodicity and sudden event superposition, and the traditional prediction method based on a single time sequence is difficult to meet the demands. The existing load prediction technology mainly has the following limitations that on one hand, the traditional time sequence model is difficult to effectively extract and fuse multiple periodic characteristics, so that the fitting of periodic changes is insufficient, and on the other hand, the modeling capability of sudden independent events and the combined effect thereof is lacking, so that the aperiodic load fluctuation cannot be accurately predicted. Therefore, there is a need to design a multiple periodic load prediction method based on a combined event relationship network to solve the above-mentioned technical problems. Disclosure of Invention Aiming at the defects of the technology, the invention aims to provide a prediction method for multiple periodic time series data, which is used for solving the problems that the traditional model is insufficient in multiple periodic feature extraction, the independent event combination effect is difficult to quantify and the prediction adaptability is poor in a mutation scene. In order to achieve the purpose, the technical scheme adopted by the invention is as follows, the multiple periodic load prediction method based on the combined event relation network comprises the following steps: Firstly, extracting load information from a national network cloud history operation log, and extracting a multi-dimensional periodic characteristic weight on the basis of a Prophet model; Secondly, extracting independent events, constructing a correlation graph structure network, mapping the pooled vectors into a combined effect coefficient through MLP (maximum likelihood) by analyzing the context relation of the graph structure through GNN, and predicting the load change when the combined events occur; and thirdly, predicting mutation through residual analysis based on the deviation index of the prediction result, and optimizing the model. Further, extracting time sequence data in the historical running load log of the national network cloud, and passing the first 80% of the time sequence data based on a Prophet modelThe change pointsEqually spaced apart into linear segments; For each point in time Creating a length ofVector of (3)Determining the time of each element representation according to the formulaWhether or not to tag the time period after each change point, after the corresponding change point: Wherein the method comprises the steps of As candidate change points,Is thatThe j-th component of (a); Compared with the prior art, the method has the advantages that the existing time sequence prediction model generally assumes that the data trend is single and stable, and multiple non-stable trend mutation of the power grid cloud platform load caused by various planned events is difficult to effectively process. The whole trend is decomposed into a plurality of linear segments by introducing a structural change point, and the fitting capacity of the model to complex and segmented trends is improved; further, the trend term is calculated according to the formula segmentation: Wherein the method comprises the steps of Is the initial slope, the trend term is a combination of a plurality of linear segments, the change points are the connection points of the segments, the slope of the trend is allowed to have an adjustment quantity at each change point,Is at the point of changeA vector of slope adjustment values at which,Is the intercept point of the beam,Is a vector of offset adjustments to ensure that the trend term function is continuous at the point of change. In order to make the function at the change pointWhere continuous, i.e.The values approximated from the left and right sides are equal and must be matchedConstraint is carried out; At the point of change Where the slope suddenly changesTo compensate for this jump and to make the curve continuous, the intercept must be adjustedI.e.Will beSubstituting the original formula can obtain a form which is more convenient to calculate: According to the formula, the slope change quantity Applying a laplace a priori distribution avoids overfitting: Wherein the method comprises the steps of Is a scale parameter, ifThe larger the a priori distrib