Correlation-guided multivariate time-series framework for tight gas well production forecasting

Accurate short-term production forecasting is critical for tight sandstone gas well performance evaluation and production management, but reservoir heterogeneity, nonlinear dynamics and coupled parameter variations pose major challenges. Existing studies focus heavily on forecasting architecture improvement, while the adaptation of pretrained time-series models to production variable characteristics remains underexplored. This study develops a correlation-guided multivariate forecasting framework. Based on 330-day mature production data from a fractured horizontal well in the Sulige Gas Field, four feature sets are constructed using four complementary correlation metrics. Six conventional and pretrained time-series models are evaluated under a direct 7-day forecasting task with 30-day historical inputs, combined with paired statistical tests and SHAP analysis. The strong-correlation feature set achieves the best overall performance. The pretrained Chronos-2 model yields the highest accuracy, with a MAPE of 4.60% and R2 of 0.9547. Weakly correlated auxiliary variables provide limited incremental predictive gain, and feature contribution patterns remain consistent across architectures. The framework highlights systematic matching between production dynamics and forecasting models, demonstrating the promise of pretrained temporal representations. Further multi-well validation is required to assess its robustness and transferability.