Application of Neural Networks for Improved Gravel-Pack Design

Abstract

Use of effective sand-control practices has sustained oil and gas production from wells that would otherwise have shut-in. However, a large number of gravel-packs fail early after installation necessitating expensive well intervention. A basic prerequisite to effective control among others is a good gravel-pack design and execution1. This includes obtaining a representative sample of the formation sand, analysing the grain-size distribution and selecting an optimum gravel-size. Gravel-size selection is carried out in relation to formation grain-size to control formation sand movement and using the optimum screen slot to retain the gravel. However, core samples are generally not taken in all wells in a field. In the absence of specific well information, it has become accepted practice to use offset-well data for designs creating a potential for ineffective gravel-packs. This paper discusses the application of neural networks to obtain real-time, well specific, grain-size distributions and how this input could be used to improve gravel-pack design and achieve optimum sand control. Neural networks have been applied with success to predict grain-size distributions from well logs. The availability of a continuous grain-size profile across an entire reservoir has facilitated comparison between the size of gravel on existing gravel-packs where offset-well information was used and gravel-size obtained based on neural network estimates. The results indicate significant differences. They further demonstrate that the use of estimates of grain-size distribution across the entire reservoir rather than offset well grain-size can lead to improved gravel pack performance and thus significant savings on life-cycle costs. Introduction For decades gravel packing has proven to be a succesfull technique for sand control. However, the method has been associated with a reduction in well productivity. The key to successful gravel packing involves selecting gravel of the proper size and quantity and placing the gravel without contamination at the proper location2. It is therefore evident that gravel pack impairement can result from a combination of these factors. One of the most compromised factors of all is the selection of proper gravel size. Early literature on use of formation grain-size for gravel pack design is based on the work of Coberly and Wagner (1938)3. Saucier (1974)4 published results of tests with physical models; he stated that ‘To apply the recommended design criteria, the important information required - knowledge of formation grain-size - is often the least available’. The reality of the unavailability of formation grain-size data is due to the fact that coring is expensive and therefore seldom carried out on development wells. Gravel pack designs have had to be based on grain-size data obtained from offset wells rather than the well in question and so negating the effect of reservoir heterogeneity. In some cases designs are based on very scanty sieve data from the subject wells - in some cases as few as five formation samples over large reservoir sections are used. Poor data from sources such as bailers, sidewall samples or production samples have also been used. These practices create a potential for the gravel pack system so designed to be ineffective or fail in early life. In an earlier paper (SPE 56626)5 the authors showed that grain-size prediction is feasible from wireline logs using neural networks. Using this technique, a backpropagation neural network can be trained with available grain-size distribution and well logs in a field and used to characterize grain-size distribution in subsequent wells in that field. Thisin situ approach means the just-in-time availability of a continuous log of grain-size over an entire reservoir section.

Results of Thermal Horizontal Completions with Sand Control in Lake Maracaibo, Venezuela: Case Histories of Horizontal Gravel Packs in Bachaquero-01 Reservoir

“…Recommended clearance between washpipe OD and the screen ID is 0.8. 12,13 The following completion procedures were used:…”

Recent Advances in Granular-Based Sand Control Design Operations and Their Performance Evaluation

“…Faga et al (1,2) has, for instance, presented a method for determining grain size distribution using Neural Networks. The methodology, based on log data, utilizes a "learning" process by which the grain size prediction is continuously being calibrated against actual laboratory results.…”