CN-121997300-A - Power distribution network load pseudo-measurement generation method combining meteorological and scene labels

CN121997300ACN 121997300 ACN121997300 ACN 121997300ACN-121997300-A

Abstract

The invention discloses a power distribution network load pseudo-measurement generation method combining weather and scene labels, which comprises the steps of collecting historical data of a plurality of groups of installed table nodes in a power distribution network, constructing weather feature vectors corresponding to historical moments according to the weather data, mapping time data of the historical moments into time period scene label vectors, mapping a user behavior mode into the user behavior mode label vectors, constructing multidimensional feature vectors, taking history load data actually measured by the corresponding nodes at the corresponding historical moments as labels of the corresponding multidimensional feature vectors, so as to obtain a training set, training a multitask deep learning model by utilizing samples in the training set, and obtaining an active power predicted value and a reactive power predicted value of an unloaded table node at the predicted moment by utilizing the multitask deep learning model. The method can effectively improve the accuracy and the robustness of the load pseudo measurement generation aiming at the non-tabulated nodes.

Inventors

CAI JUNYU
HAN JIAJIA
SUN XIN
CHEN XIN
WANG ZIXIANG
YAO YING
YAN YONG
WANG LIUWANG
LIU RONGHAO

Assignees

国网浙江省电力有限公司电力科学研究院

Dates

Publication Date: 20260508
Application Date: 20260409

Claims (10)

1. A power distribution network load pseudo-measurement generation method combining weather and scene labels is characterized by comprising the following steps: Collecting historical data of a plurality of groups of installed meter nodes in the power distribution network, wherein the historical data comprises meteorological data, time data, user behavior patterns, holiday marks and historical load data of all the installed meter nodes at a plurality of historical moments; constructing weather feature vectors corresponding to the historical moments according to the weather data, and mapping the time data of the historical moments into time period scene tag vectors; Constructing a multidimensional feature vector according to the meteorological feature vector, the time period scene tag vector, the user behavior mode tag vector and the holiday mark at the historical moment, and taking the historical load data actually measured by the corresponding node at the corresponding historical moment as the tag of the corresponding multidimensional feature vector so as to obtain a training set; the input of the multi-task deep learning model is a multi-dimensional feature vector of the node at the corresponding moment, and the multi-task deep learning model is output as an active power predicted value and a reactive power predicted value corresponding to the node; And for the node without the meter, constructing a multidimensional feature vector according to weather data, time data, a user behavior mode and holiday marks corresponding to the prediction time, and inputting the multidimensional feature vector into a trained multitask deep learning model, so as to obtain an active power predicted value and a reactive power predicted value of the node at the prediction time.
2. The method for generating pseudo-measurement of load of power distribution network by combining weather and scene tags according to claim 1, wherein the types of the installed table nodes comprise resident nodes, business/office nodes and industrial production nodes.
3. The method for generating the pseudo-measurement of the load of the power distribution network by combining the meteorological and scene tags according to claim 1, wherein for the node n, the node n is at the following position The expression of the weather feature vector at the moment is: ; in the formula, 、、 And Respectively representing a temperature normalized value, a relative humidity normalized value, a solar irradiance normalized value and a wind speed normalized value of the position of the node n at the moment t, Is the meteorological characteristic vector of the node n at the time t.
4. The method for generating the pseudo-measurement of the load of the power distribution network by combining the meteorological and scene tags according to claim 1, wherein a long-term memory network or a time convolution network is adopted by a sharing layer of the multi-task deep learning model.
5. The method for generating the pseudo-measurement of the load of the power distribution network by combining the meteorological and scene tags according to claim 1, wherein the expression of a loss function L adopted for training the multi-task deep learning model is as follows: ; in the formula, And The weights of the first task and the second task are respectively, MSE is mean square error, and P and Q are respectively a node active power measurement value and a node reactive power measurement value; And And respectively representing the node active power predicted value and the node reactive power predicted value.
6. The method of generating a pseudo-metric of load of a power distribution network combining weather and scene tags as recited in claim 1, wherein mapping time data of historical moments into time-period scene tag vectors comprises: extracting the hour number of the current moment from the time stamp; Defining scene meanings corresponding to different time intervals, wherein time period scenes corresponding to time periods 00:00-06:00 are late night/deep valley, time period scenes corresponding to time periods 06:00-10:00 are early peak, time period scenes corresponding to time periods 10:00-16:00 are waist load, time period scenes corresponding to time periods 16:00-21:00 are late peak, and time period scenes corresponding to time periods 21:00-24:00 are night transition; And acquiring a corresponding time period scene according to the hour number of the current moment, and mapping the time period scene into an integer index.
7. The method for generating the pseudo-measurement of the load of the power distribution network by combining the meteorological and scene tags according to claim 1, wherein the holiday mark is obtained by obtaining date information corresponding to the historical moment, judging whether the holiday is the holiday, setting the holiday mark to be 1 if the holiday is the holiday, and setting the holiday mark to be 0 if the holiday is the holiday.
8. The method for generating the pseudo-measurement of the load of the power distribution network by combining the meteorological and scene tags according to any one of claims 1 to 7, wherein the multidimensional feature vector expression of the node n at the time t is as follows: ; Wherein: The weather feature vector of the node n at the time t is obtained; A time period scene label vector of the node n at the time t; a user behavior pattern tag vector for node n; Is a holiday sign.
9. The method for generating power distribution network load artifact measurement combining meteorological and scene tags of claim 8, further comprising noise disturbance based data enhancement of the training set prior to training the multi-task deep learning model with samples of the training set, comprising: superimposing random noise subjected to normal distribution on the meteorological feature vector as a newly added meteorological feature vector ; Acquiring a new multidimensional feature vector of a training sample by using the new meteorological feature vector, and taking a set of the new multidimensional feature vector and an original multidimensional feature vector of the training sample as an enhanced training sample, wherein the expression of the new multidimensional feature vector of the node n at the time t is as follows: 。
10. The method for generating power distribution network load pseudo-measurement by combining meteorological and scene tags according to claim 9, wherein the method for generating power distribution network load pseudo-measurement by combining meteorological and scene tags is characterized in that the training set is subjected to data enhancement based on noise disturbance and further comprises the step of setting one of the multidimensional feature vectors to zero according to a preset probability so as to obtain a new multidimensional feature vector of a training sample.

Description

Power distribution network load pseudo-measurement generation method combining meteorological and scene labels Technical Field The invention relates to the technical field of power distribution network data processing, in particular to a power distribution network load pseudo-measurement generation method combining weather and scene labels. Background The distribution network is used as the tail end of a power system, and the accurate perception of the running state of the distribution network is important to guaranteeing the power supply reliability. Because the distribution network has numerous nodes and wide distribution, the distribution network is limited by construction cost and communication conditions, and the coverage of real-time measurement equipment on all the nodes is difficult to realize. Therefore, the generation of high-precision load pseudo-measurement data for the non-tabulated nodes by using the existing measurement data is a key link for realizing comprehensive observability and accurate state estimation of the power distribution network. The load of the power distribution network is jointly influenced by meteorological conditions, time rules, user behavior patterns, holidays and other factors, and the power distribution network has strong nonlinear, fluctuation and randomness characteristics. In the prior art, a deep learning algorithm based on data driving is generally adopted to solve the problem, for example, a long-short-term memory network (LSTM) or a time convolution network is utilized to generate pseudo-measurement data by mining time sequence dependency relations in historical load data and establishing a load prediction model. However, the above prior art still has a certain limitation in practical application. Firstly, most of traditional deep learning models adopt a single-task learning mode, only prediction of active power is usually focused, and an inherent physical relation and a strong coupling relation existing between active power and reactive power are often ignored, so that the generated pseudo measurement data of the active power has lower precision. Secondly, the existing method is rough in feature construction, original time data or meteorological data are generally directly input into the model, concrete scene semantics (such as peaks, deep valleys and the like in the morning and evening) behind a time period and differential influence of combination of meteorological factors and specific scene labels on load change cannot be fully mined, and therefore generalization capability and prediction accuracy of the model under complex and changeable scenes are limited. Disclosure of Invention In order to solve the technical problem of low accuracy of a generation result of a power distribution network pseudo measurement data generation method in the prior art, the invention provides a power distribution network load pseudo measurement generation method combined with meteorological and scene labels, so as to effectively improve the accuracy and robustness of load pseudo measurement generation for non-tabulated nodes. The invention adopts the following technical scheme that the power distribution network load pseudo-measurement generation method combining meteorological and scene labels comprises the following steps: Collecting historical data of a plurality of groups of installed meter nodes in the power distribution network, wherein the historical data comprises meteorological data, time data, user behavior patterns, holiday marks and historical load data of all the installed meter nodes at a plurality of historical moments; constructing weather feature vectors corresponding to the historical moments according to the weather data, and mapping the time data of the historical moments into time period scene tag vectors; Constructing a multidimensional feature vector according to the meteorological feature vector, the time period scene tag vector, the user behavior mode tag vector and the holiday mark at the historical moment, and taking the historical load data actually measured by the corresponding node at the corresponding historical moment as the tag of the corresponding multidimensional feature vector so as to obtain a training set; the input of the multi-task deep learning model is a multi-dimensional feature vector of the node at the corresponding moment, and the multi-task deep learning model is output as an active power predicted value and a reactive power predicted value corresponding to the node; And for the node without the meter, constructing a multidimensional feature vector according to weather data, time data, a user behavior mode and holiday marks corresponding to the prediction time, and inputting the multidimensional feature vector into a trained multitask deep learning model, so as to obtain an active power predicted value and a reactive power predicted value of the node at the prediction time. Preferably, the types of the tabulated nodes include residen