Engineering-Constrained Explainable Machine Learning for Development Policy Optimization in the Y284 Tight Oil Block, Ordos Basin

Tight-oil performance is controlled by a coupled reservoir-engineering system: reservoir quality, fracturing design, well pattern, and operating policy jointly determine whether geological potential becomes sustained production. This study develops an engineering-constrained, explainable machine-learning workflow for the Y284 block of the Longdong Chang 6 reservoir and identifies the controls on five-year cumulative oil. Four constraints are imposed: pre-decision and post-hoc variables are separated to avoid leakage; decline fitting, water-breakthrough classification, and injection-effect diagnosis serve as physical constraints; models are selected by accuracy and engineering plausibility; and SHAP controls are turned into field-readable policies. Of 939 well records, 338 with a complete geological and engineering record form the model-ready sample; five-fold cross-validation stratified by well type gives an out-of-fold R2 of 0.55 for the retained random forest. Engineering communication intensity, set by pump rate, stage count, and per-stage fluid and proppant volume, dominates production (69.9% of the mean absolute SHAP value), followed by geological storage capacity (21.6%) and injection and production control (8.6%); stage count and pump rate rank first. The framework is a block-scale screening and ranking tool, not a high-accuracy predictor.