Abstract: Accurate production prediction remains difficult for hydraulically fractured horizontal wells in mud shale oil reservoirs because numerous geological and engineering factors jointly affect well performance.To solve this problem,a big-data preprocessing method for shale oil production prediction was proposed by combining retrieval-augmented generation(RAG) technology with the DeepSeek large language model,increasing data collection efficiency by six times.A RIME-XGBoost model suitable for production prediction in mud shale reservoirs was constructed to achieve intelligent prediction of cumulative oil production from hydraulically fractured horizontal wells,and SHAP-based interpretability analysis was used to identify the key factors controlling production.The results show that the main controlling factors of mud shale oil production are vitrinite reflectance (R_o),compensated neutron,proppant intensity,sand-fluid ratio,logging-derived effective porosity,maximum horizontal principal stress,and total hydrocarbon from gas logging.The RIME-XGBoost model yields a coefficient of determination(R²) of 0.936 and a prediction accuracy of 91.04%,demonstrating both high accuracy and strong interpretability.This study effectively addresses problems such as low efficiency in data preprocessing and poor model interpretability in traditional production prediction,and provides a data-driven solution for accurate productivity prediction and fracturing parameter optimization in mud shale oil development,which is of great significance for the efficient development of mud shale oil.

Key words: mud shale oil, machine learning, production prediction, intelligent fracturing, geology-engineering integration