A Few-shot Deep Learning Reservoir Prediction Method Based on Kriging Data Enhancement

The depositional environments of conglomerate reservoirs exhibit features of diverse provenance inputs and restricted transport distances, accompanied by abrupt lateral and vertical variations in sand body geometry. To address the issue of substantial inter-block reservoir parameter variations, this study develops a machine learning-driven predictive framework incorporating fault-block constraints. A Kriging-based data augmentation strategy is formulated to mitigate the limited sample size challenge caused by sparse well distribution patterns, thereby enhancing the representativeness of training datasets. Utilizing well-derived sandstone thickness measurements from target intervals as desired outputs, a multi-layer perceptron architecture is optimized through augmented training data, enabling subsequent spatial prediction of sandstone distribution within well-covered domains. Field datasets from Oilfield A demonstrate that the calibrated algorithm achieves over 95% thickness prediction accuracy at control wells, with horizontal well applications attaining 80%+ sandstone penetration success rates in post-drilling evaluations. These operational validations confirm the methodology's exceptional predictive fidelity and robust extrapolation capabilities.