Construction and Application of a Reservoir Analogy Intelligent Agent Based on the Integration of Large and Small Models

With the continuous increase in the complexity of newly developed oil fields globally, traditional reservoir analog methods, constrained by static empirical formulas and discrete data analysis models, are faced with industry bottlenecks such as low matching accuracy of multi-dimensional features and the disruption of interdisciplinary decision-making chains. In this study, an innovative reservoir analog intelligent agent based on the integration of large and small models is proposed. Its innovativeness is reflected in the following aspects: Firstly, a task chain-driven calculation paradigm for reservoir membership degree is established. It deeply integrates the semantic analysis capability of large language models and the computational accuracy of professional models, achieving precise service scheduling through dynamic tool invocation. Secondly, a spatiotemporal adaptive weight allocation algorithm is developed. Twelve-dimensional features—such as structural complexity index and reservoir heterogeneity coefficient—are embedded into the membership function, and a time decay factor is introduced to capture differences across development stages. Thirdly, a task decomposition-recombination mechanism controlled by the intelligent agent is designed. Natural language instructions from users are transformed into a standardized four-stage process: “semantic analysis → task decomposition → tool invocation → solution generation,” forming a closed-loop link from demand to computation to decision-making. Field-tested on real reservoir datasets, the intelligent agent achieved 92.4% accuracy in instruction parsing and 87.6% Top-3 precision in analogy recommendation. In addition, solution generation time was significantly reduced to the minute level. These results validate the effectiveness of large–small model collaboration in bridging the “data–knowledge–decision-making” gap, providing a quantifiable and interpretable framework for intelligent oilfield development, and promoting the industry’s transition from experience-driven to model-driven workflows.