Deliverability is recognized as the primary productivity indicator of salt cavern gas storage. Accurate prediction of deliverability based on input parameters such as total gas storage capacity and base gas capacity is essential for ensuring operational safety and achieving precise injection-withdrawal regulation. To address existing limitations of the conventional random forest (RF) model in deliverability prediction, including long optimization time, incomplete feature engineering, and insufficient generalization capability, a multi-sampling-rate RF (MSR-RF) prediction model was proposed. The model was designed under a “triple-optimization” technical framework that integrated a dual-engine hyperparameter optimization strategy, a physics-informed and data-driven feature engineering scheme, and a three-tier adaptive sampling ensemble mechanism. These enhancements collectively improved the predictive accuracy, interpretability, and robustness of the model. Validation was performed using 1 334 samples. The results show that the proposed model achieves a coefficient of determination of 0.999 7 on the test set, demonstrating high prediction accuracy. Compared with the single-sampling-rate RF model, reductions of 48.4% in root mean square error and 52.1% in mean absolute error are observed, and stable performance is maintained across various noise levels. The research results provide new insights into the application of machine learning for intelligent storage management and predictive control in salt cavern gas storage systems.
PENG Jia-qi, YANG Jing-ze, FU Bin-bin, YAO Hong.
Deliverability Prediction of Salt Cavern Gas Storage Based on Optimized Random Forest Model[J].
Science Technology and Engineering, 2026, 26(24): 10394-10403 DOI:10.12404/j.issn.1671-1815.2507121
当前,人工智能和机器学习技术已在能源预测领域展现出广泛的适用性,如风电功率预测[6]、充电站负荷预测[7]、地质灾害预测[8]及地下岩层岩性预测[9]等。基于自学习机制与复杂的非线性问题处理能力,人工智能技术正发展成为具有工程潜力的新一代预测手段。然而,人工智能在地下储气库采输能力(deliverability)预测中的应用较为有限。目前研究多基于美国能源信息署(U.S. Energy Information Administration,EIA)提供的天然气储气库月度数据,选取易获取的关键储量参数,包括总库容(total gas storage capacity)、垫气量(base gas capacity)和工作气量(working gas capacity),从而构建储气库采输能力预测模型[10-13]。
为了评估机器学习模型的性能,选用决定系数(coefficient of determination,R2)、均方根误差(root mean square error,RMSE)和平均绝对误差(mean absolute error,MAE)作为预测评价指标[19]。R2用于衡量模型的拟合优度,反映模型预测结果的准确性,其值趋近于1时表明模型拟合效果越好[20]。RMSE和MAE用于衡量预测值与实际值之间的偏差,其值趋近于0时表明模型性能越好[21]。R2、RMSE和MAE的计算公式分别为
经超参数优化及特征工程优化后,进一步采用三级自适应采样集成机制构建了多采样率集成RF(multi-sampling-rate random forest,MSR-RF)预测模型,其三层采样架构分别由采样率为95%、85%和75%的子模型搭建而成,具体参数配置及不同噪声环境中各子模型的权重分配如表7所示。
为了进一步分析MSR-RF模型的鲁棒性,通过严格控制变量实验同步设置了单一采样率RF(single-sampling-rate random forest,SSR-RF)预测模型作为对照组,即在控制其他参数、特征一致的情况下,分别经过超参数优化和特征工程优化后所形成的单一采样率为97%的SSR-RF模型。
YangChunhe, WangTongtao. Opportunities, challenges, and development suggestions for deep underground energy storage in China[J]. Chinese Science Bulletin, 2023, 68 (36): 4887-4894.
YangChunhe, WangGuibin, ShiXilin, et al. Demands and cha-llenges of large-scale salt cavern hydrogen storage in China[J]. Rock and Soil Mechanics, 2024, 45(1): 1-19.
[6]
WeiW, HouJ, LiuX. Evaluating and predicting deliverability of natural gas storage sites using stacking machine learning models[J]. Geoenergy Science and Engineering, 2025, 249: 21377.
[7]
AliA, GuoL. Adaptive neuro-fuzzy approach for prediction of dewpoint pressure for gas condensate reservoirs[J]. Petroleum Science and Technology, 2020, 38(9): 673-681.
ZhaoWanming, ZhengZuolong. Adaptive frequency domain multi-layer perceptron based new method for short-term wind power prediction[J]. Science Technology and Engineering, 2025, 25(30): 12945-12953.
LiBo, WangNing, LüYelin, et al. Real-time load prediction model of electric vehicle charging station considering environmental factors[J]. Journal of Tongji University(Natural Science), 2024, 52(6): 962-969.
FanChengqiang, XiaYuanyou, ZhangHongwei, et al. Tunnel rockburst prediction based on LLE-DBSCAN-SMOTE data processing[J]. China Safety Science Journal, 2024, 34(12): 140-148.
ZengRui, ShaoYanlin, HuangYu, et al. Intelligent identification of logging lithology based on deep learning[J]. Science Technology and Engineering, 2025, 25(30): 12804-12812.
[16]
AliA. Data-driven based machine learning models for predicting the deliverability of underground natural gas storage in salt caverns[J]. Energy, 2021, 229: 120648.
[17]
AliA, AliyudaK, ElmitwallyN, et al. Towards more accurate and explainable supervised learning-based prediction of deliverability for underground natural gas storage[J]. Applied Energy, 2022, 327: 120098.
[18]
Vo ThanhH, Safaei-FaroujiM, WeiN, et al. Knowledge-based rigorous machine learning techniques to predict the deliverability of underground natural gas storage sites for contributing to sustainable development goals[J]. Energy Reports, 2022, 8: 7643-7656.
[19]
Vo ThanhH, ZamanyadA, Safaei-FaroujiM, et al. Application of hybrid artificial intelligent models to predict deliverability of underground natural gas storage sites[J]. Renew Energy, 2022, 200: 169-184.
DengYixuan, HuangYuping, HuangZhouchun. Charging and discharging capacity forecasting of electric vehicles based on random forest algorithm[J]. Automation of Electric Power Systems, 2021, 45(21): 181-188.
WangLaigang, ZhengGuoqing, GuoYan, et al. Prediction of winter wheat yield based on fusing multi-source spatio-temporal data[J]. Transactions of the Chinese Society for Agricultural Machinery, 2022, 53(1): 198-204, 458.
MaoJing, ChenZhongxu, LiuJiayuan, et al. Productivity prediction based on random forest optimized by differential evolution[J]. Henan Science and Technology, 2024, 51(21): 32-37.
ZhangWengang, TangLibin, ChenFuyong, et al. Prediction for TBM penetration rate using four hyperparameter optimization methods and random forest model[J]. Journal of Basic Science and Engineering, 2021, 29(5): 1186-1200.
ZhouGuxin, HuGuikai. Parameter optimization of random forest based on grid algorithm and particle swarm optimization[J]. Journal of Anhui University(Natural Science Edition), 2025, 49(3): 27-34.
ChenMingqiang, ZhengWenhao, SunYanjun, et al. Fuel consumption prediction of general aviation aircraft based on regression tree and random forest[J]. Science Technology and Engineering, 2025, 25(7): 3026-3034.
[32]
ChiccoD, WarrensM J, JurmanG. The coefficient of determination R-squared is more informative than SMAPE, MAE, MAPE, MSE and RMSE in regression analysis evaluation[J]. PeerJ Computer Science, 2021, 7: e623.
[33]
HodsonT O. Root-mean-square error (RMSE) or mean absolute error (MAE): when to use them or not[J]. Geoscientific Model Development, 2022, 15(14): 5481-5487.
ChenMingsheng, HongXiaoyan, PengQingxia, et al. Agricultural carbon footprint evaluation and its influencing factors based on grid search-random forest algorithm: taking Fujian Province as an example[J]. Acta Ecologica Sinica, 2025, 45(14): 6871-6885.
ZhaoGuoliang, LiuQiang, ChenZeping, et al. Driving fatigue detection based on interpretable sparrow search algorithm optimized random forest algorithm[J]. Science Technology and Engineering, 2024, 24(30): 13161-13169.