Abstract: With the development of the Internet of Things,numerous sensors can collect a large number of time series with rich data correlation,providing powerful data support for various data mining applications.However,some objective or subjective reasons(such as equipment failure,sparse sensing) often lead to the loss of collected data to varying degrees.Although many approaches have been proposed to solve this problem,data correlation is either not fully considered or computationally expensive.In addition,existing methods only focus on the completion of missing values,and fail to take into account downstream applications.Aiming at the above shortcomings,this paper designs a two-channel echo state network to achieve both the completion task and the prediction task.Although the two channels share the input layer,they have their own reservoir and output layer.The biggest difference between them is that the output layer of the left/right channels respectively represents the target value or prefilled va-lue corresponding to the moment before/after the input layer.Finally,by fusing the estimates of the two channels,the data correlation from before and after the missing moments is fully utilized to further improve performance.Experimental results of diffe-rent missing rates with two missing mechanisms(random missing and piecewise missing) show that the proposed model is superior to the current methods in both completion accuracy and prediction accuracy.

Key words: Data correlation, Time series, Exogenous variables, Two-channel echo state network, Missing value completion, One-step