Time Series Data Mining Based On Machine Learning And Its Application In Power Regulation

Time series data runs through all aspects of production and management in all fields of society,contains a lot of useful information,and plays a vast and unique role in an intelligent society.At the same time,there are many problems in time series data.Abnormal data will be generated due to instrument errors,human factors,and other reasons.A large amount of anomalous data not only affects data mining,but may also cause economic losses in production and life;There are many factors affect time-series data,and a single factor cannot fully describe the time series characteristics,which affects the prediction effect of time series data;A single time series prediction model is easily restricted by its algorithm,and it isn’t easy to improve the prediction accuracy further.Therefore,it is necessary to detect anomalies in time series data and analyze abnormal values;analyze the relationship between multivariate data,construct a multivariate forecast model to make predictions and reduce forecast errors;build a combined forecast model to make up for the shortcomings of a single forecast model.To solve the problem of incomplete detection by a single unsupervised learning anomaly detection algorithm,this thesis proposes a secondary detection method.The DBSCAN anomaly detection model used for clustering to detect outliers out of the cluster.The i Forest-MIE model is proposed,and the maximum interval evaluation system is used to evaluate the detection effect of i Forest and identify the outliers in each cluster.Correct the erroneous data obtained through anomaly detection,comprehensively consider multi-factor variables,construct an Attention-CNNLSTM model,extract feature information from multiple variables through the attention mechanism for prediction;create a TCN network prediction model based on the attention mechanism,and use attention the force mechanism extracts time series features in multivariate data.Time-training and prediction are performed on time-series data combined with multiple variables to improve prediction accuracy further.To solve the problem that a single prediction model often cannot fully describe the time series characteristics,the TCN prediction model based on the attention mechanism and the XGBoost model are weighted and combined through the reciprocal error method to reduce the data with significant errors in a single model and improve the overall prediction accuracy.The electric power data of the 2016 Electrical Mathematical Modeling Competition is used to verify the secondary detection method proposed in this article,which shows that the secondary detection method proposed in this article can accurately and comprehensively detect outliers in the time series data.The multi-factor time series prediction models proposed in this paper,such as LSTM model,Attention-CNN-LSTM model,TCN,Attention-TCN model,and Attention-TCN and XGBoost combined model,are compared and analyzed.Experiments show that the Attention-CNN-LSTM model and Attention-TCN prediction model constructed in this article can effectively extract meteorological factor information and improve the prediction accuracy.The prediction accuracy of the Attention-TCN and XGBoost combined prediction model is the best among many models,able to achieve higher prediction accuracy.