Research On Power Quality Time Series Data Trend Prediction Algorithm Based On Deep Learning

The demand for intelligent development of the power grid is increasing with the expanded power grid scale and more and more advanced technologies in China.Among a great deal of data collected by power grid,the time series data of power quality is an important indicator to supervise the operation status and equipment operation and maintenance conditions of power grid,which affects the safe operation and stable maintenance of power grid.Having been put into operation successively in each province,the power quality detection system has collected a great deal of observation data at monitoring stations all over the country.In related research fields such as computer vision and natural language processing,the ability of traditional methods to analyze and extract features from mass data for prediction has been surpassed by new algorithms generated in the era of artificial intelligence.Therefore,the core starting point of this study is to combine the powerful deep learning algorithm and the prediction of time series data of power quality.In view of the above-mentioned technical background and focuses,a series of studies on the core issue of trend prediction of power quality time series data were carried out on the basis of deep learning technology in this paper.The main work and contributions are as follows:First of all,data missing of power quality features as well as the difficult data cleaning of the training data lead to data imbalance problem.In this paper,the generative adversarial network with excellent performance in the generation of image and sequence simulation data is migrated to deal with incremental generation of missing features and generate new simulation data to add to the original data set through training on different data of power quality features,thus solving the problem of small data volume in deep neural network training.In addition,its performance advantage has been verified through a contrast experiment with the traditional SMOTE algorithm.Its mean average percentage error₈8)8)8)8)8)8)8))in the generation of several types of power quality data is about 5.29%,which is better than that of oversampling technique-based SMOTE algorithm,that is,about 11.38%.Secondly,in terms of trend prediction algorithm for the power quality time series data,the recurrent neural network to process series data in deep learning technology was utilized to first analyze the suitability of its two variant networks of GRU and LSTM to this problem.After selecting the LSTM network for this study,the LSTM unit with optimized peephole was selected to construct the deep neural network after referring to the peephole connection of the LSTM network.Afterwards,in terms of the construction of network structure,a contrast experiment was carried out between the Encoder-Decoder structure which is good at dealing with Sep2Sep problem in the field of natural language processing,and two classical variant networks of LSTM,stacking LSTM network and bi-directional LSTM network.The research results show that the Encoder-Decoder LSTM network performs better than other two LSTM network structures in two types of evaluation indicators,among which the normalized mean square error₉9)8)8)9)9)8)8))decreased by more than 3?6,and the mean average percentage error₈8)8)8)8)8)8)8))decreases by more than 1.2%,indicating that it has better accuracy in the trend prediction of power quality time series data.In conclusion,this paper designed and implemented a trend prediction display system of power quality time series data,which can make it more convenient and quicker to configure the hyper-parameters of training and to supervise the training situation of the model in real time.As for the test results,it can reflect the performance of model prediction in a visualized way.