Perpustakaan Digital ITB

Reserve estimation is one of the crucial aspects needed to be well determined because it indirectly represents how big the field's economic value is. Reserve can be estimated with many techniques during exploration, development, and production stages to continuously updating the remaining reserve. Decline curve analysis (DCA) is the most straightforward method to determine the remaining reserve during production stages because it only utilizes production data. Unfortunately, the traditional DCA process is time-consuming and repetitive, which might take weeks to complete, especially in handling numerous wells, and highly subjective during the manual curve fitting process. Digital technologies, especially machine learning as part of artificial intelligence (AI), could overcome the drawbacks of the traditional DCA process. A novel automated DCA process is proposed, including data reading and preparation, feature engineering, event detection, curve fitting, forecasting, and performance determination. First, the event detection consists of two processes: zero rates and no longer zero rate production detection and detection based on statistical approach using local polynomial regression to find abrupt changes of production trend. Second, the curve fitting is done using quantile regression curve fitting, which is less sensitive to outliers and could accommodate the range of uncertainty of the subjectively traditional DCA process. Third, EUR result from forecasting is normalized and regressed linearly to determine the performance based on trend gradients which might indicate hidden EUR potential. The algorithm of the automated DCA process is applied to one of Indonesia's oil fields which have approximately 1000 actively producing wells of a total of around 2000 wells. Overall, all of the proposed workflows is shown a good result. It could detect zero rates and no longer zero production as breakpoints and detect sharp changes during the transition of downward and upward production trends for event detection. Also, the curve fitting result is consistent with the algorithm and could accommodate the range of uncertainty. In addition, the performance of field "A" and all its wells is well determined and indicates there are hidden potential EUR, especially wells classified as bad and worst performance wells. Moreover, the significant impact of automated DCA proposed is that all processes start from data reading until performance determination only took about 1 hour to complete, which is very time-efficient and could decrease the cost for the DCA process.