Deep learning schemes for predicting deliverability from different underground gas storage reservoir types; Gas Science and Engineering; Vol. 145
| Parent link: | Gas Science and Engineering.— .— Amsterdam: Elsevier Science Publishing Company Inc. Vol. 145.— 2026.— Article number 205805, 20 p. |
|---|---|
| Other Authors: | , , , , , |
| Summary: | Title screen Underground gas storage (UGS) is vital for maintaining energy security and grid stability. Accurate forecasting of maximum daily deliverability (MDD) is crucial for optimizing UGS operations. However, traditional methods for forecasting MDD are often time-consuming or rely on difficult-to-obtain reservoir data, creating a need for more efficient and accessible predictive tools. This study develops robust deep learning (DL) models to predict MDD using a large dataset of over 3000 samples from UGS facilities across the USA. The models utilize only four publicly available features: field type (FTP), working gas capacity (WGC), base gas volume (BG), and total field storage capacity (TFC). Standalone long short-term memory (LSTM) and convolutional neural network (CNN) models are implemented, then hybridized with grey wolf optimization (GWO) and growth optimizer (GO) to fine-tune their hyperparameters and internal weights. A comprehensive evaluation, including a detailed error distribution analysis, demonstrates that hybrid models significantly outperform standalone ones. The LSTM-GO model achieves the highest prediction performance (test root mean square error: 25.98 million cubic feet (MMcf)/day) and superior generalization with the lowest overfitting index. Interpretability analysis (Shapley additive explanation/partial dependent plot) identifies FTP as the most influential feature on MDD, with salt caverns yielding the highest deliverability. This work provides a novel, accurate, and efficient DL framework for screening-level UGS deliverability forecasting, offering a valuable tool for energy planners to enhance the reliability and integration of natural gas storage within the evolving energy landscape Текстовый файл AM_Agreement |
| Language: | English |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://doi.org/10.1016/j.jgsce.2025.205805 |
| Format: | Electronic Book Chapter |
| KOHA link: | https://koha.lib.tpu.ru/cgi-bin/koha/opac-detail.pl?biblionumber=687214 |
| Summary: | Title screen Underground gas storage (UGS) is vital for maintaining energy security and grid stability. Accurate forecasting of maximum daily deliverability (MDD) is crucial for optimizing UGS operations. However, traditional methods for forecasting MDD are often time-consuming or rely on difficult-to-obtain reservoir data, creating a need for more efficient and accessible predictive tools. This study develops robust deep learning (DL) models to predict MDD using a large dataset of over 3000 samples from UGS facilities across the USA. The models utilize only four publicly available features: field type (FTP), working gas capacity (WGC), base gas volume (BG), and total field storage capacity (TFC). Standalone long short-term memory (LSTM) and convolutional neural network (CNN) models are implemented, then hybridized with grey wolf optimization (GWO) and growth optimizer (GO) to fine-tune their hyperparameters and internal weights. A comprehensive evaluation, including a detailed error distribution analysis, demonstrates that hybrid models significantly outperform standalone ones. The LSTM-GO model achieves the highest prediction performance (test root mean square error: 25.98 million cubic feet (MMcf)/day) and superior generalization with the lowest overfitting index. Interpretability analysis (Shapley additive explanation/partial dependent plot) identifies FTP as the most influential feature on MDD, with salt caverns yielding the highest deliverability. This work provides a novel, accurate, and efficient DL framework for screening-level UGS deliverability forecasting, offering a valuable tool for energy planners to enhance the reliability and integration of natural gas storage within the evolving energy landscape Текстовый файл AM_Agreement |
|---|---|
| DOI: | 10.1016/j.jgsce.2025.205805 |