Cluster-oriented multi-feature time series data preprocessing method
Xiangjun Cheng, Hongmei Zhang, Xilang Tang, Ruiqi Zhang · 2024
Aiming at the problem of multi-feature time series data preprocessing with cluster characteristics, the difference between cluster-oriented multi-feature time series and general time series under the background of big data is expounded. Considering the differences and time-varying of individuals in the cluster, as well as the non-uniformity of time granularity between the dependent variable and the independent variable, the descriptive statistical method is introduced to characterize the complex multi-features of the cluster under the time series, and the standardized processing including flow of data cleaning, integration and transformation, and progressive statistics is established. The experimental results on a real engine data set show that the LSTM model using the cluster-oriented method for data preprocessing method is superior to the simple method in prediction accuracy. On three datasets, the cluster-oriented method makes the RMSE and MAE of the model decrease by 10.71% and 18.50%, 8.06% and 9.82%, 6.73% and 11.06% respectively, which verifies the feasibility and superiority of the proposed method. It has reference value for processing time series data with cluster characteristics and standardizing the processing flow.