Prediction of flow rates from different entries using PLT p-T measurements in a horizontal well by machine learning methods

Abstract

Oil, which is a source of energy and raw materials, has been maintaining its importance for almost a century, and although humanity has sought alternatives, it has not been able to find a source of energy and raw materials that can replace oil and natural gas to date. The origin of the oil production process is more than a century old. Oil, which was first produced with vertical wells, can now be produced from deviated and horizontal wells with the development of technology. Today, horizontal wells have become quite common. Production logging is basically a logging method used to determine how much fluids (water, oil, gas) flow through which intervals. It is possible to measure parameters such as flow profile, pressure, temperature, ID of the well thanks to traditional production logging tools. Production logging tools, originally developed for vertical wells, have lost their function with the spread of deviated and horizontal wells. Because the flow profile in deviated and horizontal wells is quite complex compared to vertical wells. Data obtained with traditional tools are often unhealthy or incomplete data. New devices have been developed to eliminate this disadvantage. Thanks to these devices, velocity and hold-up distribution across the cross-section can be obtained. However, incomplete-inaccurate data due to mechanical problems and high costs are their disadvantages. The spread of horizontal wells allows oil and natural gas production at high flow rates. It is quite obvious that this situation is very advantageous from an economic point of view, but it also brought with it a number of technical problems. Due to the high flow rate, the pressure drop due to friction increases along the well; thus, creating an unbalanced flow profile. Eventually, water and gas coning may occur. Fortunately, ICDs (inflow control device), a product of advanced sounding technology, help solve these problems. Machine learning is the ability to learn and improve a model using data input. It provides a better understanding of the data set by making inferences from the existing data set and allows predictions to be produced using input data. Another feature of machine learning is that it is fast. It can explain analyzes that can take days in a very short time using the right model. Machine learning provides a wide range of uses that can affect many areas such as image process, classification and linear regressions. A substantial amount of data is also obtained in the oil and gas sector. This indicates that the applications of ML in the oil and gas sector will increase. In high flow wells, the roughness value becomes important as the pressure drop due to friction is high. Also, the fact that the roughness is an unknown parameter, moreover, its variation with time makes it difficult to determine. In this study, a constant flow rate interval is determined. Then, the pressure difference due to inflow-outflow and acceleration is neglected. Since the well model is created by considering the pressure difference due to gravitation, only a pressure difference due to friction is detected in this constant flow rate interval. Then, by applying the reverse solution, the relative roughness is calculated. Defining the flow profile correctly is critical for for each member of subsurface team. Although modern production logging tools are helpful in understanding the flow profile, it is prone to mechanical problems and moreover, logging with optimum conveyance requires long operation times and singinificant logging budget. Therefore, some researchers have used solely pressure and temperature data to estimate flow and determine the contribution of perforations. In addition to empirical and analytical approaches, ML applications have started to emerge recently. Considering both the rapid development of ML and its successful applications, the applicability of ML techniques to the oil and gas industry is promising. In this study, the flow rate of six production zones was estimated by using PLT pressure and temperature data obtained from two different measurements (high-choke-low-choke). The aim of this study is to produce a simple, practical and cost effective solution to the limitations of traditional Array PLT caused by mechanical problems and not working correctly in low flow regions. In this context, a well model was created and the contribution of the perforations of this model was adjusted according to the high-choke PLT (high flow) data. Various skin factors are assigned to perforations to make this adjustment. According to this model, synthetic data was produced and flow rate estimates were made for both measurements using machine learning techniques. Machine learning decision tree regression, linear regression, ridge regression and random forest regression were used for flow estimation with high-choke pressure data. 70% of the synthetic data is used for training and 30% for testing. Decision tree regression and random forest regression based on test score are the two methods that give the best results. The high-choke pressure data and the flow estimation are quite well and almost perfectly matched. Flow estimation with high-choke temperature data also gives a good result. As for flow estimation with low-choke pressure and temperature data, the test scores of both methods are above %90, but the estimates are far from low-choke flow data. However, the learning model of both ML algorithms is promising for estimating flow rate.