Most process mining study focuses on analysis of past data. This differs from predictive process monitoring, which, as a part of operational support, has as one of its focuses on the prediction of a running case [1]. Although there are several measures of interest that can be provided, in the present study, we focused on the remaining time of a running case. Results produced by Deep Neural Network (DNN) [2], despite its acknowledged power for various problems, typically are no better than those of other supervised algorithms with problems involving categorical variables in tabular data [3]. Because the dataset extracted from event logs that contain categorical variables can be constructed and categorized in tabular form, it is unwise to use only ordinary DNN. In this study, we showed that we can increase the accuracy of DNN on tabular data that contains categorical variables by using a technique known as Entity Embedding. To show the robustness of the method, we conducted experiments with two types of dataset, synthesis data and real-world data, and also compared its performance with other supervised learning algorithms for regression problems. The experimental results showed that it is true that the proposed method can increase the accuracy of DNN predictions on remaining time prediction problem involving categorical variables and beats all baseline methods used as comparison.