Joint Inversion of DTS and DAS Data Through Deep Learning and Intelligent Multi-Objective Optimization Algorithm: A Case Study of Horizontal Injection Well in Middle East Carbonate Reservoir

Abstract

Abstract Distributed optical fiber sensing for real-time downhole monitoring is an essential technology in the efficient development of Middle Eastern carbonate reservoirs, in which distributed temperature sensing (DTS) and distributed acoustic sensing (DAS) are two frequently utilized monitoring techniques. Efficiently and accurately inversing DTS and DAS data is important in identifying key water injection channels, capitalizing on residual oil reserves, and accurately forecasting production metrics. Meanwhile, there are two aspects of challenges in inversing DTS and DAS data, the first one is the inversion algorithms developed so far lack robustness and efficiency when facing an extensive set of parameters and computationally expensive forward models. The other one is that existing inversion techniques for distributed fiber optic monitoring data rely solely on either DTS or DAS data, with no research conducted on the combined inversion of DTS and DAS data. With those in mind, a joint inversion method coupling deep learning (DL) and multi-objective optimization (MOO) algorithm called DL-MOO is proposed for simultaneous inversion DTS and DAS so as to obtain the comprehensive inversing results with reservoir parameters including reservoir permeability, water saturation, and grid well indices. The proposed DL-MOO method integrates DL and MOO to address the joint inverse problem of DTS and DAS data with an extensive set of parameters and the computationally expensive forward model. In detail, the Long Short-Term Memory auto-encoder (LSTMAE) technique effectively condenses interpretation parameter sets into compact latent vector representations to achieve the goal of reducing the dimensionality of the parameter space. Subsequently, the inversion process is conducted within the neural network's latent variable space rather than the conventional parameter space of the forward model, leading to notable enhancements in efficiency and robustness. After that, the hybrid multi-objective particle swarm optimization algorithm (HMPSO) is adopted to search and update latent variables into the forward model to obtain the Pareto front (PF) for maximum R2 of temperature profile with DTS data and the R2of frequency band extracted with DAS data. Furthermore, a case study is conducted on a horizontal injection well in the Middle East carbonate reservoir to demonstrate the superior performance of the DL-MOO method. The results indicate that the PF of the DL-MOO method matched well with the PF of the commercial software-based MOO method, which validates its effectiveness and reliability. Additionally, a series of comparison analyses among the DL-MOO method against, the DL-MOPSO (Multi-objective Particle Swarm Optimization) method and the DL-NSGA-II (non-dominated sorting genetic algorithm-II) are executed to demonstrate the remarkable enhancements in the quality of inversion results achieved by the DL-MOO method. Under the same iteration steps, the convergence and diversity of the PF the DL-MOPSO and the DL- NSGA-II method are dominated by the PF of DL-MOO method. To the best of our knowledge, this is the first time that the joint inversion of DTS and DAS data for interpreting reservoir parameters. Through the integrated inversion of DTS and DAS data, the DL-MOO method realizes the purpose of robustness and efficient interpretation of parameter sets along the wellbore direction, encompassing reservoir permeability, water saturation, and grid well indices. Moreover, the precise interpretation results attained through the DL-MOO method could substantially enhance the effectiveness and accuracy of evaluating and monitoring horizontal well performance, which holds significant importance for optimizing the development of water-flooding carbonate reservoirs with horizontal wells.