Innovative Deep Autoencoder and Machine Learning Algorithms Applied in Production Metering for Sucker-Rod Pumping Wells

Abstract

With the depletion of reservoirs, it is inevitable that once very prolific conventional and unconventional wells become stripper wells characterized as producing no more than 35 barrels of liquid equivalent per day over a 12-month period. Conventional metering strategy is not economic for its huge investment on facilities, equipment, sensors and ongoing maintenance compared to current low production. This paper presents an innovative method utilizing state-of-art artificial intelligent algorithms to predict the production rate from real-time IIot testing and producing data with very low cost and reliable accuracy.

Abundant real-time field and well data acquired from IIot and digital field facilities establish a fundamental foundation for developing a machine learning application. This paper presents a method to predict the real-time production rate from real-time IIot data. In our method, we start with constructing our datasets from different data sources by combining the dynamometer cards, pumping stroke and rate, pump, rod, wellbore and reservoir parameters as inputs and the corresponded production rate as targets. The machine-learning model contains two neural networks: first, a deep autoencoder to extract the feature representations from all the dynamometer cards; then another neural network combining all related features to predict the real-time production.

The deep autoencoder derived features from dynamometer cards are used as parts of inputs to further real-time production prediction model, which eliminate the disadvantage of conventional hand-crafted features. Hand-crafted features can lose important information whereas autoencoder is designed to minimize information loss by learning high level features that can be used to reconstruct the cards. The production prediction model with pump and producing data combining more informative abstract features generate a good accordance with the history data. After tested testing and validating data in several fields in one operator's fields in China, the model demonstrate very high accuracy and with R² more than 0.92, MAE less than 0.5 of wells producing less 5m³/day, RMSE less than 1.6 of wells producing 5-10m³/day and RMSE less than 2.2 of wells producing more than 10m³/day. The model has also been tested on hundreds of newly producing wells and with error in 10%, comparing with high resolution real-time metering equipment.

The method described in this paper can be fully utilized to metering the real-time production with ultra-low cost in wells as long as acquiring real-time dynamometer card. With the fast development of artificial intelligence technology and expanding training datasets, artificial intelligence is a good choice to lower the investment and maintenance cost for conventional and unconventional fields in the low oil price trend.