Intelligent Production Monitoring with Continuous Deep Learning Models

Abstract

Summary Monitoring of production rates is essential for reservoir management, history matching, and production optimization. Traditionally, such information is provided by multiphase flowmeters or test separators. The growth of the availability of data, combined with the rapid development of computational resources, enabled the inception of digital techniques, which estimate oil, gas, and water rates indirectly. This paper discusses the application of continuous deep learning models, capable of reproducing multiphase flow dynamics for production monitoring purposes. This technique combines the time evolution properties of a dynamical system and the ability of neural networks to quantitively describe poorly understood multiphase phenomena and can be considered as a hybrid solution between data-driven and mechanistic approaches. The continuous latent ordinary differential equation (Latent ODE) approach is compared with other known machine learning methods, such as linear regression, ensemble-based model, and recurrent neural network (RNN). In addition, the framework of virtual flowmetering (VFM) is analyzed from the point of multiphase measurement technology. In this work, the application of Latent ODEs for the problem of multiphase flow rate estimation is introduced. The considered example refers to a scenario in which the topside oil, gas, and water flow rates are estimated using the data from several downhole pressure sensors. The predictive capabilities of different types of machine learning and deep learning instruments are explored using simulated production data from a multiphase flow simulator. The results demonstrate the satisfactory performance of the continuous deep learning models in comparison with other machine learning methods in terms of accuracy, where the normalized root-mean-squared error (RMSE) and mean absolute error (MAE) of prediction below 5% were achieved. While Latent ODE demonstrates the significant time required to train the model, it outperforms other methods for irregularly sampled time series, which makes it especially attractive to forecast values of multiphase rates.