A Simple Data-Driven Approach to Production Estimation and Optimization

Abstract

In this paper we describe an approach to real-time decision support that is completely data-driven. The approach can be viewed as a sequence of transformations acting on available production data. The transformations form a data pipeline from sensors to operational advice, and can be summarized as follows. Historical and real-time production data, experimental data such as well tests, and any other operational metadata, enters the data pipeline on one end. The data is then synchronized, cleansed and compressed, while retaining important uncertainty measures such as standard errors. The resulting information is used to build models using regression analysis or machine learning. These models are updated in real time as new data enters the pipeline. In this fashion, the models are adjusted frequently to the operational situation, making up for their limited prediction capabilities compared to advanced first-principle models. This approach may reduce the need for advanced fluid modeling and model calibration efforts in real-time applications. Arguably, the approach may provide automation under uncertainty, and hence sustainability, to a higher degree than traditional approaches based on process simulation. A great advantage of the data-driven approach is that the uncertainty of measurements and models is tracked. By considering this information, production estimation and optimization may give operational advice with uncertainty measures. Furthermore, by tracking uncertainty, it is possible to advise experiments (e.g. step-tests) that may yield valuable operational information by entering unexplored operational regions. In our experience, information about the uncertainty of alternative production advice may ultimately alter the final decision. Furthermore, production advice without accompanying uncertainty measures may be misleading.

Bayesian Neural Networks for Virtual Flow Metering: An Empirical Study

“…We preprocess the experiment data using a data squashing technology [39] to avoid short-scale instabilities. After preprocessing, each example represents an experiment performed during steady-state operation, which is suitable for modeling steady-state production rates.…”

On gray-box modeling for virtual flow metering

“…The datasets for each well are preprocessed in two steps. First, the processing technology in Grimstad et al (2016), is utilized to generate a compressed dataset of steady-state operating points suitable for steady-state modeling. Secondly, a set of filters are applied to remove data samples that likely originate from erroneous sensor data, such as negative pressures or choke openings.…”

Multi-task learning for virtual flow metering

“…We are interested in the steady state behaviour of the flow rates. All observation are therefore averages taken over 3-9 hour intervals of stable production [23]. Observations are shifted and scaled to lie approximately in the unit interval before model training and evaluation.…”