Variable and time-lag selection using empirical data

The paper proposes a method to select the best variables and respective time-lags for industrial applications when the objective is the estimation of a target variable using the information content of empirical data. No further information is assumed about the process. The problem of jointly selecting the best variables and the respective time-lags is treated as a variable selection problem. This assumption implies an increase of input dimensionality and multicollinearity into input space. Then, a multidimensional mutual information estimator based on the l-nearest neighbor algorithm is used in a forward search procedure to select the best variables and and respective time-lags. To verify the performance of selected variables and delays, the method was successfully applied in two data sets. A least squares support vector machine was used as the main model for the soft sensor in both cases.