digilib@itb.ac.id +62 812 2508 8800

ABSTRAK Rocky Yan Classica
PUBLIC Suharsiyah

This study discusses the development of a Pressure Transient Analysis (PTA) method using the Convolutional Neural Network (CNN) algorithm to identify well test models automatically. The well test model is a combination of wellbore, reservoir, and boundary models. Although the PTA process currently utilizes commercial well testing software. However, this method proves to be less efficient as it requires long iterations in the process. In this study, the CNN program will be trained using a database of pressure and pressure derivative response images obtained from the test design process using commercial well testing software. The database consists of 8 classes with a total of 1000 images. For the purposes of training and testing the CNN program, 70% of the data will be used as training data, while the remaining 30% will be used as test data. The number of layers used in the CNN network is 12 layers, consisting of 5 convolution layers, 5 pooling layers, and 2 fully connected layers. The type of activation function used in the convolution and first fully connected layers is Exponential Linear Units (ELU). Meanwhile, in the second fully connected layer, the type of activation function used is Soft-Max. In addition, the developed CNN program will also use Dropout as a regularization method, Cross-Entropy as a loss function, and Adam as an optimization method. In the well test models prediction process, the program requires input data in the form of images of pressure and pressure derivative response derived from commercial well testing software. After that, the program will automatically provide predictions in the form of appropriate wellbore, reservoir, and boundary types for each image inputted into the program. The study results show that the CNN program has good performance in predicting the well test models. Evaluation using F1-score shows a high average level of prediction accuracy, which is 87%. Meanwhile, through training and validation plots, it shows that the CNN program is safe from overfitting or underfitting conditions. In addition, testing using case study data from Field-X shows that the CNN program can identify the well test models quickly and accurately. Thus, through the CNN program that has been developed, it can make the process of identifying well test models through PTA more efficient.