A study on connectivity path search in fractured-vuggy reservoirs based on multi-agent system

Abstract

Due to the complex and heterogeneous spatial structure of fractured-vuggy reservoirs caused by multiple tectonic movements, understanding inter-well connectivity paths is challenging. Explicit connectivity paths are crucial for designing injection and production schemes and enhancing oil recovery. Traditional methods often fail to adequately characterize geological structures, making it difficult to represent preferential pathways. This study proposes a novel algorithm that integrates seismic multi-attribute data and reinforcement learning to automatically search for 3D inter-well connectivity paths. A multi-agent deep reinforcement learning model based on actor-critic is employed, with each agent representing a flow direction in the multi-phase carbonate rock system. Game theory is used to identify connectivity paths that align with geological structures, while fluid flow laws are incorporated into the reward function to improve search accuracy. A multi-head self-attention mechanism is introduced to capture global state information and the correlation between fluid flows in different directions. Variational Bayesian estimation is utilized to improve search efficiency by leveraging prior geological data. The algorithm is applied to a typical oilfield in China, where it successfully identifies connectivity paths. The results are validated by comparing the identified paths with tracer concentration production curves, showing improved accuracy in representing the spatial distribution characteristics of the reservoir.