This is an exclusive member-only content. Login here to view the video.
Using Machine Learning for Building Multivariate IPR Models from High Frequency Streaming DataAug 15, 2021
This presentation in English is from the 2020 SPE Latin American and Caribbean Petroleum Engineering Conference. Our archived 2020-21 event content is now available free to SPE members!
- Paper Number: SPE-199132-MS
- Published: 20 July 2020
- Author: Mario Antonio del Pino Fiorillo
Dealing with the volume of data at the field has become increasingly challenging, as conventional computational tools are not suited for dealing with such high-frequency data. On this work, we explore an innovative way to address this issue when trying to build productivity models for an asset.
The objective of this work was to develop an algorithm capable of building IPR-like models with high-frequency data generated from producing wells. The code handles data retrieval, filtering, model fitting and results presentation. It accounts for depletion-driven changes and can accurately predict short- and medium-term evolutions in productivity. This technique was tested on a gas condensate field constrained to a daily nominal rate of around 200MMscfd of dry gas. Machine learning techniques were used to create multiple models. Metrics were computed to quantify model accuracy and select the best fit model. The resulting code successfully built IPR curves with great prediction capabilities.
This approach gives a tool to automate the creation of mathematical models for data-rich assets. There are numerous potential applications: real-time production strategy updates, integrated interpretation of asset development, better design and scheduling of surface facilities, among many others.