History Matching of Production Performance for Highly Faulted, Multi Layered, Clastic Oil Reservoirs using Artificial Intelligence and Data Analytics: A Novel Approach

Abstract

History Matching (HM) is one of the critical steps for dynamic reservoir modelling to establish a reliable predictive model. Numerous approaches have emerged over the decades to accomplish a robust history matched reservoir model ranging from the classical reservoir engineering approach to the widely accepted 3D numerical simulation approach and its variations. As geological complexity of the oil and gas field increases (multilayered reservoirs, heavily faulted) compounded with completion complexity (dual strings, commingle production), building a fully representative predictive reservoir model can be arduous to almost impossible task. Artificial Intelligence (AI) and machine learning has advanced almost all major industries, including the petroleum industry in general and reservoir engineering. The objective of this paper is to outline a novel approach in history matching using a data-driven approach through Artificial Intelligence via Artificial Neural Network (ANN) and Data-Driven Analytics. In this paper, a step by step methodology in building a reservoir model and history matching process using ANN will be described which includes data preparation and data QA/QC, spatiotemporal database formulation, reservoir model design, ANN architecture design, model training and history matching strategy. A case study of the implementation to Field "A" in Malaysian waters is presented where good to fair history matching quality was obtained for all production parameters. Field "A" is a 25kmx75km oil sandstone reservoirs of a highly geologically complex field (more than 200 major and minor faults, more than 30 reservoir layers) of more than 25 years of production. The challenges of history matching of this field does not only lie on its geologically complex structure and its corresponding subsurface uncertainties, but also on the production strategy of the wells that involved commingled dual strings production with several integrity issues that adds additional dimensions to the field's complexities. To date, Field "A" has no field wide history matched reservoir model using conventional numerical simulation method available due to the complexity of history matching. This long history matching woe is mitigated via the implementation of AI based reservoir model and Data Analytics. This novel approach is estimated to be more time and cost-efficient compared to the conventional method. The comparison of this AI based reservoir model and history matching methodology with the conventional numerical reservoir model approach will be discussed. Furthermore, the advantages, limitations and areas of improvements of this AI based history matching methodology will also be highlighted. The target audience of this paper would be to reservoir engineering practitioners and dynamic model simulators who is interested to learn the complementary or alternative approach in reservoir modelling apart from conventional numerical modelling in order to create time-efficient reservoir model and reducing the risks in their field development plans.

Applications of Machine Learning in Subsurface Reservoir Simulation—A Review—Part I

“…The results showed that the ANN was expensive enough regarding the number of simulations that were required; however, good-quality results were achieved, in terms of uncertainty reduction, accuracy and speed. Mohmad et al [119] focused on the risk analysis area by developing a simple ANN to match a highly complex faulted reservoir with dual-string wells, creating a more efficient model that minimizes the risk uncertainty related to production and management plans. Their results showcased that this approach is much more competent, as far as the calculation time is concerned, when compared to conventional simulators.…”

“…ANNs with stochastic optimization [106,107] ANNs with dimensionality reduction methods [108] Ensembles of ANNs [109] RBFNNs, Generalized Regression ANNs, FSSC and ANFIS stochastic optimization [110][111][112] Direct history matching Supervised ANNs [114][115][116][117][118][119][120]126] Bayesian ML models [121,122,124,125] MARS, DTs, single-pass GRNNs [127] Unsupervised Self-Organizing Map (SOM) [131] Supervised SVR with dimensionality reduction and optimization [135] RNN [142,143] CNN [148,158] Unsupervised GAN [149,159] Piecewise Reconstruction from a Dictionary (PRaD) with pluri-PCA [151] Convolutional AutoEncoders [152,157] Reinforcement learning Reinforcement learning models [103,162] Currently, dozens of professional products used to set up ML models are available to developers, might that be related to research or commercial products. This palette includes client tools developed by major players in the market such as Google and Microsoft (Google cloud AI platform, Azure machine learning) as well as free tools such as TensorFlow by Google and the Anaconda distribution for Python.…”