CN-121980181-A - Direct current transmitting end system transient voltage dynamic prediction system and method based on LSTM network

CN121980181ACN 121980181 ACN121980181 ACN 121980181ACN-121980181-A

Abstract

The invention provides a transient voltage dynamic prediction system and a transient voltage dynamic prediction method of a direct current transmission end system based on an LSTM (least squares) network, which belong to the technical field of transient voltage prediction and comprise the steps of collecting electric quantity data of the direct current transmission end system in real time through a synchronous phasor measurement unit and executing data preprocessing; extracting time domain and frequency domain features, screening out a key feature subset which has obvious influence on transient voltage prediction, constructing an LSTM deep learning model, setting model super-parameters, training the model by utilizing a training set, adjusting parameters to minimize prediction errors, evaluating and optimizing the model by utilizing a testing set, preprocessing electric quantity data acquired in real time, inputting the preprocessed electric quantity data into the trained LSTM deep learning model, predicting transient voltage values at future time and outputting the predicted transient voltage values to an electric control system. The system and the method for dynamically predicting the transient voltage of the direct current transmitting end system based on the LSTM network solve the problems of low prediction precision and low response speed in the traditional prediction method.

Inventors

WANG XUEBIN
ZHANG JIE
WEN XISHAN
CHEN XIAOYUE
YIN XIYU
ZHANG LINYU
DING YUJIE
FU GUOBIN
SONG RUI
WANG SHENGJIE
WANG SHENGFU
YANG KAIXUAN
ZHAO DONGNING
ZHAO JINCHAO

Assignees

国网青海省电力公司电力科学研究院
国网青海省电力公司
武汉大学

Dates

Publication Date: 20260505
Application Date: 20260120

Claims (10)

1. The direct current transmitting end system transient voltage dynamic prediction method based on the LSTM network is characterized by comprising the following steps of: acquiring electrical quantity data of a direct current end system in real time, performing data processing on the acquired electrical quantity data, and performing normalization processing on the processed electrical quantity data; extracting time domain features and frequency domain features from the processed electrical quantity data, selecting key features, and selecting a feature subset which has influence on transient voltage prediction; Constructing an LSTM deep learning model comprising an input layer, an LSTM layer, a full connection layer and an output layer, and setting model super parameters; Dividing the processed electric quantity data into a training set and a testing set, training an LSTM deep learning model through the training set, adjusting parameters by means of a back propagation algorithm to minimize a prediction error, and evaluating the LSTM deep learning model by the testing set to optimize; the LSTM deep learning model predicts a transient voltage value at a future time according to the input data and outputs a prediction result to the electrical control system.
2. The LSTM network-based direct current transmission end system transient voltage dynamic prediction method of claim 1, wherein the data cleaning in S1 adopts a method based on statistical analysis and machine learning, data points exceeding a reasonable range are regarded as abnormal values by utilizing a sliding window statistical mean value and a standard deviation, interpolation correction is carried out according to adjacent data points, the normalization processing in S1 adopts a maximum and minimum normalization method, the data is mapped to a [0,1] interval, and the normalization expression is: ; Wherein, the Is the original data; is the data minimum; data maximum; is normalized data.
3. The method for dynamically predicting transient voltage of direct current transmission end system based on LSTM network as set forth in claim 1, wherein the frequency domain feature extraction extracts frequency domain features of voltage and current through Fourier transform, and the expression is: ; Wherein, the Is a time domain signal; Is a frequency domain signal.
4. The LSTM network-based direct current transmission end system transient voltage dynamic prediction method is characterized in that key feature selection comprises the steps of performing reduction and optimization on extracted features by utilizing improved principal component analysis PCA, introducing feature importance evaluation indexes on the basis of traditional PCA by the improved PCA method, performing weighting treatment on the features, and selecting a feature subset with the largest influence on transient voltage prediction.
5. The method for dynamically predicting transient voltage of direct current transmission end system based on LSTM network as recited in claim 1, wherein the adjustment parameters are self-adaptive moment estimation optimization algorithm, and the expression is: ; Wherein, the Estimating for a first moment; Estimating for a second moment; Is a gradient; And (3) with Are momentum parameters; is the learning rate; is a constant parameter; is a model parameter.
6. The method for dynamically predicting transient voltage of direct current transmission end system based on LSTM network as recited in claim 1, wherein the data prediction expression in S5 is: ; Wherein, the Is that A hidden state from time; 、 Are output layer parameters.
7. The method for dynamically predicting transient voltage of direct current transmission end system based on LSTM network as set forth in claim 1, wherein the prediction error between the predicted result and the actual monitored result is evaluated by mean square error and root mean square error, and the calculation formulas are respectively: ; ; Wherein, the Is the first Predicted values for the individual samples; Is the first True values of the individual samples; is the number of samples.
8. A prediction system of a direct current transmitting end system transient voltage dynamic prediction method based on an LSTM network is characterized by comprising the following steps: The data acquisition module adopts a high-precision synchronous phasor measurement unit, is arranged on a bus and a key node of the direct current converter station, and acquires voltage, current, active power and reactive power electric quantity data in real time; The data preprocessing module is connected with the data acquisition module and comprises a sliding window statistics sub-module, and the average value and the standard deviation are adopted for abnormal value detection; the feature extraction module is connected with the data preprocessing module and comprises a frequency domain feature selection sub-module; the LSTM model construction module comprises two layers of LSTM network structures, each layer of the LSTM network structure comprises 64 neurons, the full-connection layer comprises 32 neurons, the input layer adopts a ReLU activation function, and the output layer adopts a linear activation function; the model training and optimizing module is connected to the feature extraction module and the LSTM model construction module and comprises an adaptive moment estimation optimizer; The prediction result output module is connected to the LSTM model construction module, predicts the transient voltage value at the future moment, and outputs the prediction result to the electrical control system.
9. The LSTM network-based direct current transmission end system transient voltage dynamic prediction system according to claim 8, wherein the prediction system adopts a modular design, and each module performs data interaction through a standardized interface to support hot plug maintenance.
10. The LSTM network-based direct current transmission end system transient voltage dynamic prediction system of claim 8, wherein the prediction result output module includes a visualization sub-module, and the trend of comparison between the predicted voltage and the actual voltage is shown by a dynamic graph.

Description

Direct current transmitting end system transient voltage dynamic prediction system and method based on LSTM network Technical Field The invention relates to the technical field of transient voltage prediction, in particular to a direct current transmitting end system transient voltage dynamic prediction system and method based on an LSTM network. Background In an electric power system, a direct current transmitting end system is used as a key link of electric energy transmission, and stable operation of the direct current transmitting end system is important for guaranteeing safety and efficiency of the whole power grid. However, since direct current end systems involve complex electro-physical processes and are susceptible to external disturbances and internal faults, system transient voltage fluctuations are frequent and difficult to predict. The traditional transient voltage prediction methods mainly depend on an empirical formula and a simplified model, and the methods are unsmooth in processing the working conditions of a complicated and changeable power system and have the limitations of low prediction precision, low response speed and the like. The traditional method often fails to fully pay attention to the cleaning and preprocessing work of the original electric quantity data, so that noise and abnormal values mixed in the data are directly used for predictive analysis without effective processing, and the accuracy and reliability of a predicted result are seriously affected. In the feature engineering stage, the traditional method is mostly limited to the extraction of time domain features, but ignores frequency domain features and other factors which can have important influence on transient voltage. The one-sided feature extraction method causes incomplete selected feature subsets, and dynamic characteristics of the power system cannot be comprehensively reflected, so that improvement of prediction accuracy is limited. Most of traditional prediction models are static models, and it is difficult to effectively capture the dynamic change characteristics of a power system. When facing complex working conditions such as sudden faults or large-scale load changes, the static models often cannot timely adjust the prediction strategy, so that the prediction effect is obviously reduced, and the actual engineering requirements cannot be met. In the model parameter optimization stage, the traditional method mostly adopts low-efficiency algorithms such as a trial-and-error method or a simple gradient descent method, and the algorithms are not only low in convergence speed, but also easy to fall into a local optimal solution, so that the model cannot reach a global optimal state, and further stability and accuracy of prediction performance are affected. Disclosure of Invention The invention aims to provide a transient voltage dynamic prediction system and a transient voltage dynamic prediction method of a direct current transmitting end system based on an LSTM network, which solve the problems of low prediction precision and low response speed in the traditional prediction method. In order to achieve the above purpose, the present invention provides a method for dynamically predicting transient voltage of a direct current transmitting terminal system based on an LSTM network, comprising: acquiring electrical quantity data of a direct current end system in real time, performing data processing on the acquired electrical quantity data, and performing normalization processing on the processed electrical quantity data; extracting time domain features and frequency domain features from the processed electrical quantity data, selecting key features, and selecting a feature subset which has influence on transient voltage prediction; Constructing an LSTM deep learning model comprising an input layer, an LSTM layer, a full connection layer and an output layer, and setting model super parameters; Dividing the processed electric quantity data into a training set and a testing set, training an LSTM deep learning model through the training set, adjusting parameters by means of a back propagation algorithm to minimize a prediction error, and evaluating the LSTM deep learning model by the testing set to optimize; the LSTM deep learning model predicts a transient voltage value at a future time according to the input data and outputs a prediction result to the electrical control system. Preferably, the data cleaning in S1 adopts a method based on statistical analysis and machine learning, and uses a sliding window statistical mean value and standard deviation to treat data points exceeding a reasonable range as abnormal values and carry out interpolation correction according to adjacent data points, the normalization processing in S1 adopts a maximum and minimum normalization method to map the data to a [0,1] interval, and the normalization expression is as follows: ; Wherein, the Is the original data; is the data minimum;