1. Controlled time series forecasting for oil reservoir management.
- Author
-
de Souza Jr., Alexandre, Tueros, Juan A. R., Machado, Mateus G., Santos, Rafael F. V. C., Willmersdorf, Ramiro B., Afonso, Silvana M. B., Oliveira, Leonardo C., and Horowitz, Bernardo
- Subjects
- *
RECURRENT neural networks , *CONVOLUTIONAL neural networks , *STANDARD deviations , *PETROLEUM reservoirs , *TIME series analysis - Abstract
Management of oil reservoirs requires many high computational cost simulations for decision-making. This work aims to develop fast machine learning-based proxies based on production history data to replace the reservoir simulator. The proxy must accurately forecast reservoir responses by analyzing time series of fluid phase rates or bottom hole pressures, defined as a function of informed controls of the injector and producer wells. It is crucial to note that this is not a regular problem of forecasting the future of time series based on past trends but the prediction of reservoir response time series as a function of informed new controls. The proposed strategy is to use neural networks to learn reservoir dynamics based only on samples of production history data. Multi-head architectures based on recurrent neural networks (RNNs), convolutional neural networks (CNNs) configurations, and hybridization are considered. Parallel architectures involving separate RNNs are tested, using long short-term memory (LSTM) and CNN concatenated sub-networks. Eight different architectures are tested on two reservoir models of different sizes and heterogeneity complexities. Several training samples are generated to assess their impact on accuracy and precision. For the smaller example, the CNN architecture produced the most precise results with fewer trainable parameters. The proposed Parallel CNN-LSTM architecture is generally the most successful for the larger and more complex reservoir. In any case, the proposed techniques have been demonstrated to be very promising, with root mean square error in the order of 2.25%. [ABSTRACT FROM AUTHOR]
- Published
- 2024
- Full Text
- View/download PDF