PAI-Informer: A Production-Anchored Incremental Informer for Multistep Shale Gas Production Forecasting

Abstract

Abstract Reliable multistep forecasting of shale gas production is hindered by nonstationary decline, operational disturbances, and error accumulation. This study develops a Production-Anchored Incremental Informer (PAI-Informer) for 1–7-day production forecasting. Instead of directly predicting absolute production, the model estimates log-production increments anchored to the latest observation and reconstructs the production trajectory through cumulative summation and inverse log transformation. A joint level-trend (JLT) loss constrains both production magnitude and adjacent-step variation. The model was evaluated using daily production and pressure data from three shale gas wells and compared with Persistence, CNN-LSTM, CNN-GRU-AM, Informer, Autoformer, and PatchTST across five random seeds. Across the three wells, with evaluation metrics calculated independently for each well and then averaged, PAI-Informer achieved mean values of R2 = 0.9071, RMSE = 0.6384 × 104 m3/d, MAE = 0.3838 × 104 m3/d, and sMAPE = 31.29%. Compared with Informer, it reduced RMSE, MAE, and sMAPE by 11.46%, 23.33%, and 30.03%, respectively, and ranked among the top two models in RMSE for every well and forecast horizon. Ablation experiments identified production-anchored incremental decoding as the principal source of improvement, while the joint loss enhanced prediction stability. These modifications introduced no additional trainable parameters and increased inference time by only 5.6%. SHAP analysis attributed 71.45% of the global importance to historical gas production. Leave-one-well-out testing revealed sensitivity to unseen high-rate regimes, highlighting the need for broader multiwell data and domain adaptation. PAI-Informer provides a lightweight and interpretable framework that reduces multistep forecasting drift through production-anchored incremental prediction and improves trajectory consistency through JLT optimization without increasing model complexity.