Transformer-based forecasting for high-frequency natural gas production data

Abstract

Accurate prediction of natural gas well production data is crucial for effective resource management and innovation, particularly amid the global transition to sustainable energy. Traditional models struggle with high-frequency, high-dimensional datasets generated by digital transformation in the oil and gas industry. This study explores the application of Transformer-based models — Transformer, Informer, Autoformer, and Patch Time Series Transformer (PatchTST) — for forecasting high-frequency natural gas production data. These models utilize self-attention mechanisms to capture long-term dependencies and efficiently process large-scale datasets. Autoformer achieves predictive success through its Seasonal Decomposition Attention mechanism, which effectively extracts trend-seasonality patterns. However, our experiments show that Autoformer exhibits sensitivity to dataset changes, as performance declines when using old parameters compared to retrained models, highlighting its reliance on dataset-specific retraining. Experimental results demonstrate that increasing sampling frequency significantly enhances prediction accuracy, reducing MAPE from 0.556 to 0.239. Additionally, these models consistently track actual production trends across extended forecast horizons. Notably, PatchTST maintains stable performance using either pretrained or retrained parameters, showcasing superior adaptability and generalization. This makes it particularly suitable for real-world applications where frequent retraining may not be feasible. Overall, the findings validate the applicability of Transformer-based models, particularly PatchTST, in dynamic and precise natural gas production forecasting. This study provides valuable insights for advancing adaptive, data-driven resource management strategies.