Abstract
The main purpose of the work is to create a system for effective monitoring of well operation by analysing the data coming from sensors. This algorithm is focused on automation of intraday monitoring of unstable wells, as well as search for deviations in operation of wells and downhole equipment.
This paper applies an algorithm to reconstruct flow rate dynamics in an unstable wells based on deep learning models, specifically using a neural network model that learns from the large amount of information coming from telemetry sensors installed on wells and downhole equipment. In addition, static well data, such as well construction, PVT properties, and various data by results of well surveys, are also used. This approach allows all information from all wells in a particular field or group of fields to be used, and also allows the model to find complex relationships between parameters that sufficiently characterise well performance.
For wells with unstable working conditions, the approach presented in the work has a significant advantage, as it allows to form a sample with participation of a large number of wells on different regimes within the same method of exploitation, taking into account intentional or unintentional changes in operating regimes on these wells. The implementation of this algorithm achieved a convergence for unstable wells of up to 85% compared to the standard algorithms, which did not exceed 75% convergence. Despite the statistical model, this algorithm absorbed and summarised the vast expert experience of specialists, based on which the feature space was built, additional calculated parameters were added and parameters that did not affect the dynamics of well operation were removed from the sample. This algorithm has been successfully tested on gas lift unstable wells and has been translated to other artificial lifting methods, including complex periodic modes on ESP wells and wells operating modes under annular flow conditions. This algorithm helps field engineers solve challenges related to intraday analysis of well operations, such as regime control, wellbore understanding of deviations from planned operation modes, anomalies in operation and fast solution of production allocation problem.
The approach of building a general statistical model, compared to standard models that focus on knocking down each individual well, is a leap forward in the development of intraday stock analysis. With a timely trained model, the analysis is done extremely quickly on the entire field at once, and more importantly, quite efficiently. Where simple standard modeling approaches do not allow reproducing nonstationary processes with sufficient accuracy or take large amount of computing resources, the neural network is able to automatically filter data from all well telemetry, generalize features and identify complex patterns in the well operation behavior.
