基于LSTM的缝洞型油藏水淹异常模式识别及预警

大缝大洞的存在和频繁的工作制度导致缝洞型油藏含水率变化特征多样，暴性水淹预警难度大。针对传统预警方法存在的时滞性问题，采用K线理论刻画含水率生产指标变化趋势，总结出充沛型、突破型、反转型等水淹前异常模式；由于循环神经网络能够刻画生产数据间的长程相关性，采用基于循环神经网络的长短期记忆网络（LSTM）自动识别水淹异常模式特征实现暴性水淹预警。仿真实验表明基于LSTM的水淹异常模式识别模型通过变换数据尺度，较好地捕获暴性水淹前数据的整体变化趋势，识别精度明显优于支持向量机、朴素贝叶斯等模型。K线理论刻画的各类异常模式有效解决了传统预测的时滞难题，提前1~3周实现水淹预警，可以为缝洞型油藏水淹预警研究提供新的研究思路。 

Abnormal pattern recognition and early warning of water flooding in fractured-vuggy reservoir based on LSTM

The existence of large fractures and caves and frequent working system adjustments have resulted in the diverse characteristics of water cut in the fractured-vuggy reservoir, which makes it difficult to early warn the water flooding.Aiming at the problem of time delay of traditional early warning methods, this paper uses K-line theory to describe the change trend of water cut production indicators, and summarizes the pre-flooding abnormal patterns such as abundant type, breakthrough type and reversal type.Since the recurrent neural network can memorize the long-term correlation between production data, LSTM is used to automatically identify the features of abnormal pattern to realize early warning of water flooding.The experimental results show that by transforming the data scale, the proposed abnormal pattern recognition model based on LSTM can successfully extract the overall trend of data before water flooding.The recognition accuracy of the proposed model is significantly higher than that of support vector machine and na?ve Bayes and other models.Various kinds of abnormal patterns described by K-line can effectively solve the traditional problem of prediction delay.The proposed model realizes early warning of water flooding one to three weeks in advance, and provides new ideas for early warning of water flooding.