Data Driven Production Forecasting Using Machine Learning

TLDR

Forecasting production in unconventional prospects has attracted significant attention, yet challenges remain in rapidly deploying many wells, managing completion‑driven productivity, and understanding uncertain fluid flow physics. This study applies machine learning algorithms to forecast production for existing and new wells in unconventional assets using geological maps, production history, pressure data, and operational constraints. Artificial Neural Networks are trained on large datasets of well history, geological and operational data, then used to predict performance of existing wells from their own history and new wells from nearby wells with similar geology. The ANN‑based approach reduces data conditioning and model building effort, enabling users to focus on what‑if scenarios and performance assessment.

Abstract

Abstract Forecasting of production in unconventional prospects has gained a lot of attention in the recent years. The key challenges in unconventional reservoirs have been the requirement to put online a) a large number of wells in a short period of time, b) well productivity significantly driven by completion characteristics and that c) the physics of fluid flow in these prospects still remain uncertain. In this paper, machine learning algorithms are used to forecast production for existing and new wells in unconventional assets using inputs like geological maps, production history, pressure data and operational constraints. One of the most popular Machine Learning methods – Artificial Neural Network (ANN) is employed for this purpose. ANN can learn from large volume of data points without assuming a predetermined model and can adapt to newer data as and when it becomes available. The workflow involves using these data sets to train and optimize the ANN model which, subsequently, is used to predict the well production performance of both existing wells using their own history and new wells by using the history of nearby wells which were drilled in analogous geological locations. The proposed technique requires users to do less data conditioning and model building and focus more on analyzing what-if scenarios and determining the well performance.