Research On Time Series Prediction Algorithm And Its Servicelization Based On Association Analysis

With the rapid development of the industrial Internet of things in recent years,a large number of sensors are deployed on the industrial equipment of thermal power plants to monitor the operation of the equipment in real time.The data recorded by the sensors are multi-dimensional time series data.Device anomalies can be found by detecting anomalies in the data,but the common anomaly detection methods don’t fully consider the correlation in the multi-dimensional time series data,and the accuracy of the detection result still has a large space to improve.Anomaly detection can also be performed by predicting key variables in the time series data and comparing the difference between the predicted value and the true value.When the predicted value and the true value differ too much at a certain moment,it can be considered that the device is abnormal at that moment,but this detection method is very dependent on the performance of the prediction algorithm,so the key to detecting anomalies through prediction is how to improve the effectiveness of the prediction algorithm.The main goal of this paper is to improve the algorithm for the time series data recorded by sensors of industrial equipment in thermal power plants to improve the prediction performance of key variables.The data of thermal power plant includes multi-dimensional variables,which are related to each other.At the same time,the data are also related in time sequence.This prediction problem belongs to the time series prediction problem based on multi variables.After investigation,it is found that if we want to improve the prediction effect,we can usually start from two aspects:feature selection algorithm and time series prediction algorithm.Feature selection can remove some redundant feature variables from the original data,and improve the prediction effect.The selection of model structure of time series prediction is also an effective means to improve the prediction effect.Based on the above problems,the specific work researched in this paper includes:A feature selection algorithm based on the combination of Max-Relevance and Min-Redundancy(MRMR)and Backward Search is proposed for the feature selection of data in thermal power plant;A time series prediction model based on the combination of Long Short-Term Memory Network(LSTM)and Gradient Boosting Decision Tree(GBDT)is proposed,this algorithm also draws on the idea of Embedding;In order to facilitate users to use the algorithm in real industrial scenarios,this paper also encapsulates the above algorithm into a servicelization system.In addition,a large number of experiments are carried out to verify the effectiveness of the algorithm,R-squared(R2)was used in the experiment as an evaluation criteria of algorithm performance.The experimental results show that the feature selection algorithm proposed in this paper can reduce the running time consumption of the algorithm in addition to ensuring the performance of the existing algorithm.The R2 score of the time series prediction algorithm model proposed in this paper is 0.9315 on the test set,which is significantly improved compared with the GBDT(R2 is 0.8427)and single-layer LSTM(R2 is 0.8863),and it’s also better than stacked LSTM(R2 is 0.9089).At the same time,the training time and prediction time of the time series prediction algorithm in this paper are only slightly higher than single-layer LSTM,and reduced by 1 times compared to stacked LSTM.