Antonio Andrade Marin, Salim Busaidy, Mohammed Ahsan Adib Murad, Issa Al Balushi · 19 authors
Abstract A failed Electrical Submersible Pump (ESP) well is generally identified when there is no flow to the surface. The process of reviving well production can take weeks leading to huge unwanted deferment. Through a Proof-Of-Concept (PoC), the objective is to prototype and evaluate the results of an early failure detection for ESP wells using Machine Learning (ML), without reserving focus on implementation. By demonstrating the feasibility of this approach and verifying that the concept has practical potential, the tool can be used to reduce deferment and identify failure prone component to either devise mitigation strategy for extending time-to-failure or work on an improved design before failure. The paper details all the work undertaken to develop a Predictive Analytics model based on ML algorithms using field sensor data, real time physics-based model calculated data and well failure history to predict ESP well failure and identify failed component in advance. The approach of database standardization, data pre-processing, machine-learning algorithm selection, supervised training and validation dataset creation shall be discussed. ESP domain knowledge used for Feature Engineering across multiple modeling iterations to consistently improve well and component level model metrics shall be detailed. After the evaluation by well owners at Petroleum Development Oman (PDO), refered as Operator's blind test, the prediction of the ML algorithm shows a good accuracy in its ability to capture historical failures ranging between days to months in advance. The Well Level Failure model captures failure prone wells with a precision of 90% and accuracy of 76%. The Component Level Failure model correctly identifies pump failure from other failures with a precision of 92% and accuracy of 88%. These numbers show the reliability of future predictions that could enable users to make high stake workover and operating envelope optimization decisions with confidence. Following benefits are estimated from both Well failure and Pump Component failure prediction models metrics respectively: 28.35% savings from total unscheduled ESP deferment1% increase in Overall Mean Time to Failure (MTTF) based on optimization of predicted pump component failure wells. In an organization where over thousand ESP wells are managed by limited production engineers, post ESP failure, the effort invested for hoist scheduling, raising new well proposal, rig mobilization, new ESP installation and commissioning utilizes huge time and leads to long undesired oil deferment. Implementation of engineered analytics to predict ESP failures and failed components in advance can support production engineers to plan early for workover operations, increase well run life and minimize oil deferment losses. Methodologically assessed by Senior Petroleum Engineers in selected clusters (using historical data and in the context of each failure and non-failure cases), the Predictive Analytics journey has started. It is ready to be operationalized at a small scale to build confidence as an advisory tool for Production Engineers in real-time to evaluate multiple wells’ failure probability on a daily basis and generate massive savings from well deferment. This agile journey focused on value generation is achieved with combined efforts between technology, domain knowledge and data.
Blockchain technology is a distributed ledger of data, vetted before acceptance, encrypted and shared among parties (depending on the configuration) authorised to view it. Due to the potential cost savings, revenue generation, efficiency gains and security; governments, organisations, companies and consortia across many industries are developing proofs of concept with two in oil and gas reportedly now in production. The value chain in the oil and gas industry is characterised by many remote working locations across different geographical regions; involving multiple stakeholders and regulators each with differing criteria to be met. Many opportunities for improvement across the oil and gas value chain have been identified and some are beginning to be addressed by proofs of concept and live blockchain platforms. However, these are being developed via discrete consortia and not as part of a cohesive industry strategy. Standards to promote understanding and adoption are being developed at the global, national and industry level, but this is an area currently lacking in the oil and gas industry. This paper highlights activity in other industries and suggests that the recently formed OOC Blockchain Consortia of Oil and Gas industry participants, if extended to statutory organisations, regulators, standards organisations and academia could be the forum required to accelerate understanding and adoption within the industry to release the currently untapped value.
New offshore oil recourses that are developed are more difficult to drill and increase the requirements to the technology in the offshore industry. A relatively new technology is Managed Pressure Drilling where a choke topside is used to control the bottom hole pressure. The bottom hole pressure measurement is unreliable, which motivates the need for an observer. Different methods for estimation will be presented and compared in this thesis. Proofs of convergence are outlined for the Stamnes observer, derived for the Grip observer and some stepping stones for further work are presented for the Optimal Polynomial Filter. Each observer is simulated with a simple step in the mud pump to verify the estimation laws. The results show that all observes estimated the bottom hole pressure correctly for this simple case. A more realistic case, a pipe connection, is also simulated for each observer. The case includes zero flow from the mud pump, which reveals that all observers miss the estimation convergence in this case, but that the estimate converges when there is flow from the mud pump. One of the states that affects the bottom hole pressure is the pressure loss due to friction in drill string and annulus. Earlier work modeled these losses as quadratic with respect to the flow through the bit, which are simplifications. To improve the estimate of the bottom hole pressure, new and better friction models are needed. Measurement data from Gullfaks C are analyzed to get new knowledge of friction loss in the drilling string and annulus. For the drill string the quadratic friction model is found to be good enough, catching the main behavior. On the other hand, the friction loss in the annulus is a more complicated function of flow. The annulus friction is approximated with sets of basis functions and the weighted sum of these functions gives an approximation to the friction curve. Each weight is estimated to get the friction loss estimate. The use of the weighted sum of four 1st-order b-spline functions give a good approximation to the real friction curve, and the weight for each basis function is estimated. This is tested in simulations both with a simple case and the pipe connection case. The simulations show that the annulus friction loss and the bit pressure are estimated correctly.