In chemical plants, soft sensors are widely used to estimate process variables that are difficult to measure online. The predictive accuracy of soft sensors decreases over time because of changes in the state of chemical plants, and soft sensor models based on time difference (TD) have been constructed to reduce the effects of deterioration with time, such as drift. However, although a TD interval of training data affects the predictive accuracy of TD models and the predictive accuracy depends on a TD interval of new data, the way to determine an optimal TD interval of training data for that of new data remains to be clarified. We, therefore, propose an automatic selection method of an appropriate TD interval of training data based on a new cross-validation method. When training data are divided into data for the model construction and validation data in cross-validation, we change not only TD intervals of model construction data, but also those of validation data, and TD models and their predictive accuracy were saved for each TD interval. In prediction, TD of new data are entered into the TD model with the highest predictive accuracy for the validation data TD interval that is the same as the new data TD interval. We analyzed dynamic simulation data and real industrial data and confirmed that the predictive accuracy of TD models increase using the proposed method.