基于多参数时间序列及粒子群优化算法的油藏产量动态建模预测方法

Reservoir production performance prediction model based on multi-parameter time series and particle swarm optimization algorithm

王娟 梅启亮 邹永玲 蔡亮 苏建华 田榆杰 黄瑞

WANG Juan MEI Qiliang ZOU Yongling CAI Liang SU Jianhua TIAN Yujie HUANG Rui

中国石油长庆油田分公司数字和智能化事业部 中国石油天然气集团有限公司勘探开发人工智能技术研发中心 中国石油长庆油田分公司勘探开发研究院 清华四川能源互联网研究院 北京思达威云石油工程技术研究院有限公司

Digital and Intelligent Division of PetroChina Changqing Oilfield Company, Xi’an 710018, Shaanxi, China Artificial Intelligence Technology R&D Center for Exploration and Development, CNPC, Beijing 100007, China Research Institute of Exploration and Development of PetroChina Changqing Oilfield Company, Xi’an 710018, Shaanxi, China Sichuan Energy Internet Research Institute, Tsinghua University, Chengdu 610200, Sichuan, China Beijing Startwellcloud Petroleum Engineering Research Institute Co., Ltd., Beijing 100007, China

在油田开发过程中，油藏产量预测方法的研究对开发方案的动态调整具有重要意义。针对利用机器学习算法进行油藏产量预测过程中，因缺乏考虑时间序列模型的参数调整优化技术，以及新数据叠加进行预测模型动态更新技术，导致产量预测的准确率不高且时效性不强，难以满足实际生产应用需求等问题，研究了基于长短期记忆神经网络模型的多参数时间序列预测方法及粒子群参数优化算法，构建了随时间动态更新的油藏产量预测模型，从而进一步提升油藏产量预测的准确率与实用性，并在长庆油田多个油藏的生产过程中进行了应用。应用结果表明，模型预测结果的准确率较高，且模型具有实时训练和自动更新的特点，在实际生产中展现出了较高的应用价值。

The research on reservoir performance prediction method is of great significance for the dynamic adjustment of reservoir development plan. The production prediction based on machine learning suffers from low accuracy and low time effectiveness, due to the absence of the parameter adjustment and optimization technology featuring the time series model and the dynamic update technology of prediction model by introducing new data, making it insufficient for practical applications. Through investigating the multi-parameter time series prediction method based on the long short-term memory (LSTM) neural network model and the particle swarm optimization (PSO) algorithm, the reservoir production performance prediction model that is dynamically updated with time was constructed. Application of the presented model to several reservoirs of the Changqing oilfield shows that the model exhibits a high accuracy and enables real-time training and automatic updating. The presented model is of high significant for practical application.

10.13639/j.odpt.2023.02.010