كليدواژه :
تراوايي , رخسارۀ الكتريكي , رگرسيون , شبكۀ عصبي مصنوعي , ماشين بردار پشتيبان , نمودارهاي پتروفيزيكي
چكيده فارسي :
تراوايي از مؤلفه‌هاي اساسي در ارزيابي مخازن هيدروكربني است كه عمدتاً از طريق اندازه‌گيري‌هاي آزمايشگاهي از مغزه يا داده‌هاي چاه‌آزمايي به دست مي‌آيد. با اين حال، به دليل هزينۀ زياد و فراواني كم اين نوع از داده‌ها، پيش‌بيني تراوايي با استفاده از داده‌هاي چاه‌نگاري از جايگاه ويژه‌اي برخوردار است. در اين مطالعه، براي تخمين تراوايي، ابتدا داده‌هاي چاه‌نگارها با توجه به مطالعات زمين‌شناسي صورت گرفته بر روي ميدان مورد مطالعه به چهار گروه رخساره‌هاي الكتريكي دسته‌بندي مي‌شوند: پكستون - وكستون – مادستون، پكستون – وكستون، گرينستون – پكستون و گرينستون – پكستون – وكستون. در اين مطالعه، از شبكه‌هاي عصبي مصنوعي و ماشين بردار پشتيبان براي تخمين تراوايي در يكي از مخازن ناهمگون كربناته با استفاده از داده‌هاي چهار چاه در ميدان مذكور استفاده شده است. جهت تخمين تراوايي، ابتدا داده‌هاي نگاره‌هاي چاه با استفاده از روش‌هاي «تجزيه و تحليل مؤلفه‌هاي اصلي» و «تجزيه و تحليل خوشۀ مبتني بر مدل» به رخساره‌هاي الكتريكي تقسيم‌بندي شده‌اند. سپس هر رخسارۀ الكتريكي به‌عنوان ورودي شبكۀ عصبي مصنوعي و ماشين بردار پشتيبان جهت تخمين تراوايي در نظر گرفته شده‌اند. شبكۀ عصبي مصنوعي با استفاده از «توابع پس انتشار لونبرگ»، «گراديان نزولي با تكانه وزني» و «تابع يادگيري بياس» با ده لايۀ مخفي آموزش داده شده است. از ماشين بردار پشتيبان با رگرسيون‌هاي اپسيلون و نو با توابع كرنلي مختلف استفاده شده است. در اين مطالعه، تابع كرنل شعاعي ماشين بردار پشتيبان داراي خطاي كمتري در مقايسه با شبكۀ عصبي است. خطاي حاصل از ماشين بردار پشتيبان براي رخساره‌هاي الكتريكي گروه اول تا چهارم به ترتيب برابر است با: 0.0065، 0.0242، 3.6587 و 0.0195.
چكيده لاتين :
Permeability is one of the main parameters in the oil reservoir evaluation that is usually estimated by using well test data and laboratory measurements from the reservoir core samples. However, these methods are very expensive and time consuming, and usually a few number of wells have such information to obtain permeability and other reservoir parameters. Therefore, the prediction and assessment of the reservoir rock permeability using other non-expensive and indirect methods can effectively reduce the exploration and production costs and give us useful information about the permeability of the hydrocarbon reservoirs. Nevertheless, we have to consider that this kind of information may suffer in resolution and the results may have some unacceptable errors in estimation of the permeability. Thus, using proper prediction methods and comparing the obtained results with the permeability from the well test data and laboratory measurements leads to better and reasonable predictions of the permeability in oil and gas reservoirs. Moreover, the type of the reservoir rocks can also severely affect the estimated permeability. Usually the permeability estimation in the sand stone reservoirs is much easier than in carbonate reservoirs, especially in the heterogeneous carbonate reservoirs. This is mostly because of the porosity type and the conditions of depositional environments.
In this regard, using well log data also has important role in the permeability prediction. This is mostly because the well logging tools run in many wells and well log data are more available. Including more data in the prediction process will result in better constrained permeability estimation. Common methods of permeability prediction use empirical equations based on not always sufficient core data. These equations are usually used for a special type of reservoir and may not applicable to various types of reservoirs.
In this study, Artificial Neural Networks (ANN) and Support Vector Machine (SVM) methods are used to estimate permeability parameter in one the Iranian heterogeneous carbonate oil reservoir using well log data from the 4 wells, located in the given oilfield. These wells have 7 common logs that are incorporated in the permeability prediction process. The well log data firstly are classified into 4 electrofacies based on geological studies carried out on the field. The classified electrofacies are as follow: packstone-wackestone, mudstone-packstone, wackstone-grainstone-packstone, grainstone-packstone-wackstone. The classification is done by using Principle Component Analysis (PCA) and Model Based Cluster Analysis (MCA) methods. Then, each group of elecrtofacies is used as input data for Artificial Neural Networks and Support Vector Machine methods to predict permeability.
Artificial Neural Network (ANN) is trained by using Levenberg-Marquardt back propagation algorithm and Gradient Descent method with Momentum Weight and Bias Learning Function with 10 hidden layers. The Support Vector Machine (SVM) method is implemented using Nu and Epsilon algorithms and different types of kernel functions, such as linear, radial based functions, polynomial and sigmoid functions. Usually, the radial based kernel function gives the best regression with minimum error values. Our results show that, for all of the electrofacies, Support Vector Machine (SVM) method has less error than Artificial Neural Network (ANN) in the regression process. The Support Vector Machine (SVM) errors for the above mentioned Electrofacies are as following: 0.0065, 0.0242, 3.6587 and 0.0195 respectively.
