基于随机森林和长短期记忆网络模型的高压气井环空带压预测方法

A prediction method of annular pressure in high-pressure gas wells based on the RF and LSTM network models

高压气井在生产过程中持续的环空带压容易引起套管柱变形或挤毁,是高压气井完整性失效的主要原因之一。为解决传统方法环空带压预测精度不高的问题,以鄂尔多斯盆地苏里格气田某高压气井为例,首先利用主成分分析法和相关系数法找到影响环空带压的主要因素,然后使用高压气井井筒温压场理论值和孤立森林模型对主成分进行物理解释和数据清洗,再对清洗后的数据使用随机森林(RF)和长短期记忆网络(LSTM)模型建立了环空带压定量预测模型,并对两类模型进行权重组合,最终建立了精确度高于任意单一模型的RF—LSTM组合环空带压预测新模型。研究结果表明:(1)环空带压的主要影响因子有温度分量、压力分量、产量分量、腐蚀程度、生产状态,而温度分量与环空带压间存在最高关联性;(2)通过错误格式、离群点及基于井筒温压场的数据清洗,可以得到数据清洗后的环空带压影响因素训练集;(3)通过平均绝对误差法(MAE)能够建立误差分数小于任意单一模型,而拟合优度介于两者之间的组合模型,因此可以将具有高拟合优度和低误差分数的两类模型结合,从而组合出同时满足两种分数的组合模型。结论认为:(1)运用大数据挖掘技术及算法进行环空带压定量预测,方法新颖,预测精度高,结果可行;(2)该方法为现场环空带压预测和风险管控提供了决策工具参考,为实现环空带压风险实时预测、预警和管控提供了理论支撑。

The continuous annular pressure of a high-pressure gas well in the production may cause casing string deformation or collapse,which is one of the main reasons for wellbore integrity failure.However,traditional methods for predicting the annular pressure yield unexpected accuracy.This paper presents a new prediction method of annular pressure based on random forest(RF)and long short-term memory(LSTM)network model.By taking a high-pressure gas well in the Sulige gas field of the Ordos Basin as an example,the method is verified.Firstly,the principal component analysis(PCA)and correlation coefficient method are used to identify the main factors affecting the annulus pressure.Then,the theoretical value of wellbore temperature and pressure field of high-pressure gas wells and an isolated forest model are used to physically interpret the principal components and clean the data.Based on the cleaned data,a quantitative prediction model of annulus pressure is established by using the RF and LSTM models.Finally,a combined RF-LSTM annulus pressure prediction model with a higher accuracy than any single model is established by combining the weights of the two models.The following results are obtained.First,the main influencing factors of annulus pressure are temperature component,pressure component,yield component,corrosion degree and production state.The temperature component has the highest correlation with annulus pressure.Second,through the error format,outliers and data cleaning based on wellbore temperature and pressure field,the training set of influencing factors of annulus pressure after data cleaning can be obtained.Third,the mean absolute error(MAE)method can be used to establish a combined model with error scores less than any single model and goodness of fit between the two.Therefore,the two types of models with high goodness of fit and low error scores can be combined to form a combined model that meets the two scores at the same time.In conclusion,the proposed method for quantitatively predicting annular pressure based on big data technology and algorithm is innovative,accurate and feasible.It provides a referential decision-making tool for field annular pressure prediction and risk control,and also a theoretical support for the realization of real-time prediction,early warning and control of annulus pressure risk.