Research On Accurate Prediction Of Distribution Network Voltage Based On Deep Neural Architecture

With the proposal of the Party Central Committee’s strategic goal of "achieving carbon peaks by 2030 and achieving carbon neutrality by 2060",the transformation needs of the power industry have become particularly urgent.Under the current circumstances,the power industry’s carbon emissions account for about 35% of the country’s total carbon emissions,and the large-scale integration of renewable energy(photovoltaics,wind power,etc.)has become a general trend.However,the grid connection of high-density distributed photovoltaics has also brought serious challenges to the security and stability of the power grid.The direct control of operation and over-limit conditions has very important guiding significance for the power system dispatching department in the optimization and management of power distribution and future planning.In line with the digital upgrade trend of the power industry,combined with digital technologies such as big data and artificial intelligence neural networks,this article starts from a data-driven perspective and analyzes the regularity of photovoltaic power generation to predict changes in the distribution network voltage.The research object of this paper is a distribution network with high-density distributed photovoltaics.Relying on the actual historical voltage-related data,the related research work of the thesis is carried out based on the deep neural architecture theory.This article first introduces the principle of the high-density distributed photovoltaic power generation system,analyzes its impact on the voltage of the distribution network based on the grid-connected structure of the distributed photovoltaic,and studies how many photovoltaic power generations are connected to the distribution network.Voltage fluctuations,explore the causes of voltage fluctuations or rises.It is concluded that the changes in the voltage of the distribution network are closely related to the grid connection of distributed photovoltaics.The influence of distributed photovoltaics on the distribution network is analyzed,and the necessity of predicting the voltage of the distribution network is pointed out from the data level.Secondly,in order to solve the problems of voltage-related data from different data sources,such as data complexity and duplication,the characteristics of multi-source data are studied,and the Back Propagation Neural Network(BPNN)is used to realize the voltage-related data.Data fusion in multiple dimensions and levels;the quality of the data directly affects the accuracy of the subsequent voltage prediction research work.In order to improve the quality of the fused voltage-related data,from the perspective of data cleaning technology,the confidence interval is used to achieve outliers Identify and fill in missing values based on cubic spline interpolation,improve the overall sample data set,and visualize the results after data processing;at the same time,considering the difficulty of reducing the prediction model and further improving the quality of the input data,the ant colony is adopted the Ant Colony Optimization(ACO)realizes the feature selection of voltage influencing factors,and finds the optimal feature subset after multiple trainings,thereby reducing the number of features,and providing a highly relevant feature subset for subsequent prediction models.Finally,considering that the Neural Basis Expansion Analysis for Interpretable Time Series Forecasting(N-BEATS)algorithm based on deep neural network in the field of deep learning has superiority in dealing with the problem of time series forecasting This article is the first to use this method to study the voltage prediction.Take the voltage-related feature subsets that have undergone feature engineering as input,train the N-BEATS preliminary voltage prediction model,and obtain the corresponding preliminary voltage prediction results and actual error values;in order to further improve the voltage prediction accuracy,the neural network prediction results There must be an error characteristic.A voltage prediction model based on the Temporal Convolutional Network(TCN)error correction is proposed.The actual error output by the N-BEATS model and the training set are used as the basis,and the residual error of the TCN model is used.The module fully extracts historical data features,predicts voltage errors based on time and seasonal features and outputs the error prediction values for each season,and superimposes the error prediction results with the preliminary voltage prediction results to obtain the final voltage prediction results.Finally,an example is used to analyze the rationality of the optimal feature subset,and compare the prediction results with other traditional prediction models under different seasons and time scales.Compared with these prediction models,the voltage preliminary prediction model based on N-BEATS has accurate performance.Calculated by Mean Absolute Error Percentage(MAPE)and Root Mean Square Error(RMSE)as error evaluation standards.It is increased by about 25%.Compared with the prediction result of the N-BEATS model,the average prediction accuracy of the voltage prediction value output by the TCN error correction model can be increased by about 22%.The actual prediction results have proved the effectiveness and high accuracy of the proposed model in the prediction of distribution network voltage.