POTENTIAL HYDROCARBON ZONES PREDICTION USING ARTIFICIAL NEURAL NETWORK APPLICATION - CASE STUDY : FIELD X BACHELOR THESIS David Kristianto 12215079 Submitted as partial fulfillment of the requirements for the degree of BACHELOR OF ENGINEERING in Petroleum Engineering study program PETROLEUM ENGINEERING STUDY PROGRAM FACULTY OF MINING AND PETROLEUM ENGINEERING INSTITUT TEKNOLOGI BANDUNG 2019 1 POTENTIAL HYDROCARBON ZONES PREDICTION USING ARTIFICIAL NEURAL NETWORK APPLICATION - CASE STUDY : FIELD X David Kristianto* and Amega Yasutra** Copyright 2019, Institut Teknologi Bandung Abstract One important aspect in petroleum engineering is to define the zone interest, which indicate that there is potential value of hydrocarbon reserve. Up until now, well log analysis is the standard method to determine those potential hydrocarbon zones by estimating water saturation parameter using empirical correlations developed by experts, such as Archie equation for clean sandstone, Simandoux and Waxman-Smits for shaly sandstone. However, pay zones interpretation from well log analysis is heavily depended upon human interpretation and field evaluation. Result of created water saturation log are limited by assumptions of used correlations, therefore it might be possible that human errors are happening when interpreting pay zones. The approach of this study is to eliminate those assumptions created when analysing potential pay zones, using machine learning application to create prediction based on well log input data without further interpretation. An Artificial Neural Network (ANN) model is created using well log data for input and production rate data for output, the neural network model are then tested to give prediction in multiple wells. Based on production rate data, fractional flow data can then be determined and being used as output data. Created ANN model is composed of 2 layer and 5 optimum hidden neurons with feedforward back-propagation learning algorithm. ANN model gives correlation factor result of 0.8990 for training, 0.9233 for validation and 0.9910 for testing. The model is then used for predicting pay zones in which two kinds of prediction test are conducted. The first prediction is applied on 30 samples of production rate data test as comparison, producing a correlation factor result of 0.8432. The second prediction is applied on a well without data test as comparison. Overall, created ANN model is able to give good prediction of pay zones matched with data test result with limitations that the model is dependent on the quality and numbers of data being used. Keywords: Pay Zone, Artificial Neural Network, Feedforward Back-Propagation, Machine Learning, Fractional Flow Sari Salah satu aspek penting dalam teknik perminyakan adalah dalam hal pendefinisian zona interest, yang menunjukkan bahwa terdapat potensi cadangan hidrokarbon di dalamnya. Sampai saat ini, metode evaluasi zona yang berpotensi mengandung hidrokarbon ini adalah melalui analisa log sumur dengan menentukan nilai saturasi air menggunakan korelasi yang dikembangkan para ahli, seperti persamaan Archie untuk clean sandstone, Simandoux dan Waxman-Smits untuk shaly sandstone. Namun, penentuan pay zone melalui well log analysis sangat bergantung pada interpretasi manusia dan evaluasi lapangan. Hasil penentuan saturasi air berdasarkan analisa log terbatas oleh asumsi - asumsi yang ditetapkan berdasarkan persamaan yang digunakan, sehingga sangat mungkin terjadi kesalahan pada saat menginterpretasi pay zone. Di dalam studi ini akan dilakukan pendekatan untuk mengeliminasi asumsi-asumsi yang dilibatkan pada saat menganalisa potensi cadangan pada pay zone, dengan menggunakan aplikasi dari machine learning untuk menghasilkan prediksi berdasarkan data dari analisa log sumur tanpa memerlukan interpretasi manusia. Sebuah model Sistem Saraf Jaringan dihasilkan dengan input data dari analisa log sumur dan output data dari data laju alir produksi, kemudian model sistem saraf jaringan diuji untuk memberikan prediksi pada beberapa sumur. Berdasarkan data laju alir produksi, nilai dari aliran fraksional dapat ditentukan dan dijadikan sebagai output data. Model Sistem Saraf Jaringan Buatan yang dibuat terdiri dari 2 layer dan 5 jumlah neuron optimum dengan menggunakan algoritma pembelajaran feedforward back-propagation. Model Sistem Saraf Jaringan Buatan memberikan hasil correlation factor sebesar 0.8990 untuk training, 0.9233 untuk validasi dan 0.9910 untuk tes. Kemudian model digunakan untuk melakukan prediksi zona potensial, yang meliputi dua jenis tes prediksi. Tes prediksi pertama diaplikasikan terhadap 30 data sampel data tes produksi sebagai pembanding, memberikan hasil berupa correlation factor sebesar 0.8432. Prediksi kedua dilakukan pada satu sumur tanpa 2 data tes sebagai pembanding. Secara keseluruhan, model Sistem Saraf Jaringan Buatan mampu memberikan hasil prediksi zona potensial yang sesuai dengan data tes dengan keterbatasan bahwa model sangat bergantung terhadap kualitas data dan jumlah data yang digunakan. Kata Kunci : Pay Zone, Sistem Saraf Jaringan Buatan, Feedforward Back-Propagation, Machine Learning, Aliran Fraksional *) Student of Petroleum Engineering Study Program, Institut Teknologi Bandung, 2015 batch **) Thesis Adviser in Petroleum Engineering Study Program, Institut Teknologi Bandung 1. Introduction The fluctuating condition of oil and gas prices contributes to uncertain and unpredictable outcomes of almost all of oil and gas industries. This state leads to creative and beyond-traditional approach of solutions from many innovators all around the world. One considerable problem faced in oil and gas industry is the enormous amount of data that needs processing. Currently, Artificial Intelligence (AI) is growing correspondingly with technological advancements and one method that acts wells on solving big data problem is known as the Artificial Neural Network (ANN). There are many ongoing researches regarding the prediction of water saturation, in which the output are mostly acquired through analytical equation and setting certain constants to default values, by using ANN. This method can be done for the reason that water saturation is closely related to well-log data such as resistivity log, and neutron density log. One question comes to mind knowing that water saturation is a dependent function of fractional flow, “Is it possible to predict the values of fractional flow simply by using well-log data with ANN ?” Even further, “If it is possible to do so, then it is definitely possible to define zone of interest simply by using fractional flow predictions”. Basically, fractional flow is chosen as output data instead of water saturation because of its capability to predict the threshold of movement of fluids. It is more convenient to set the cutoff value when fractional flow data are available. Traditionally, analytical method is used to calculate the fractional flow, this approach however might contain limitation because some variable are still assumed. Fractional flow’s assumptions include displacement in horizontal plane, neglected capillary pressure and gravitational forces. These assumptions are taken in for fractional flow calculation based on well log’s water saturation interpretation as the former is dependent on the latter. Aside from the analytical method, there is an alternative to predict the movement performance of immiscible fluids, by using computational simulation. Although it is more accurate than the analytical method, it is cost inefficient and more time consuming. In this study, an ANN model is built to give prediction of fractional flow values based on well-log data and production rate (RFT) data without the need to give further interpretation, but instead by creating a complex non-linear equation that is heavily affected by the range and complexity of input data. This study limits well log interpretation for ANN prediction, using its raw log data and interpreted well log data are given at the start of this study. Some wells are not taken into consideration for prediction because they do not contain potential hydrocarbon zones and other wells that are being evaluated are treated as having a two-phase oil water system in their reservoir. 2. Basic Theory Artificial Neural Network Originally, artificial neural network (ANN) is created to mimic biologic neurons behavior in human brain using its own artificial neurons. Each neuron is linked by a weight and a bias, thus creating the ANN structure. ANN works by creating a complex non - linear equation based on the relations of input variables (Ali, 1994). This complex equation comes from the learning process of ANN, similar to that of human brain, by arranging ANN’s neuron structure which is dependent on input and output parameters that come in through a training process (Shahkarami. 2014).