基于ARIMA模型的致密气田气井产量预测

Gas production prediction of tight gas fields based on ARIMA model

由于天然气井日产量数据存在随机的关井操作,是不可预测的人为因素,为克服人为因素造成的误差,以实际开井生产数据为基础,将累计产气量作为时间序列,基于ARIMA模型对鄂尔多斯盆地致密气田S区块44口生产井使用差分自回归移动平均模型建模,将以前的累产序列为训练集,预测300 d的累计产气量,交叉验证的结果显示:①ARIMA模型为线性模型,对于生产稳定的气井累计产气量拟合效果较好,方法简单,预测精度高,对于产量断崖式变化的单井预测效果较差,而对于以区块为单位的累计产量预测误差小,和实际生产数据拟合效果好,有应用价值;②以传统的自相关函数和偏相关函数拖尾和截尾的特征进行参数优化很难确定最优化参数,而且人为因素较大,在确定当前时间产气量受历史数据影响的最大阶数后,采用遍历的方法建立模型,以赤池信息准则和贝叶斯信息准则值最小作为模型选择的策略,预测的S区块300 d的累计产量和实际产量误差分别为0.40%(AIC最小)和2.11%(BIC最小),满足产气量预测的精度;③从S区块44口井的统计结果来看,在满足差分后数据序列稳定的前提下,单独一阶差分和二阶差分预测的数据序列偏离较大,为消除随机误差,文中取一阶差分和二阶差分的平均值作为预测的最终结果,预测效果明显提升。

There are random shut-in operations in the daily production data of natural gas Wells,which is an unpredictable human factor.In order to overcome the error caused by human factors,the cumulative gas production was taken as a time series based on the actual well opening production data,and the differential autoregressive moving average model(ARIMA)was used to model 44 production wells in Block S of the tight gas field in the Ordos Basin based on the ARIMA model.The previous cumulative production series was taken as a training set to predict the cumulative gas production of 300 days.The cross-validation results show that:①As a linear model,ARIMA model has a good fitting effect on cumulative gas production of gas wells with stable production,simple method and high prediction accuracy.It has a poor forecasting effect on single wells with cliff-type production changes,and a small error in forecasting output per block,which has a good fitting effect and application value.②It is difficult to determine the optimal parameters by using the trailing and truncated features of the traditional autocorrelation function and partial correlation function,and the human factors are large.After determining the maximum order of gas production affected by the historical data,the ergodic method is used to establish the model.Taking Akike information criterion(AIC)and Bayesian information criterion(BIC)as the minimum value as the model selection strategy,the predicted cumulative production and actual production errors of Block S in 300 days were 0.4%(AIC)and 2.11%(BIC),respectively,which met the prediction accuracy of gas production.③According to the statistical results of 44 wells in block S,on the premise that the data series is stable after the difference is satisfied,the data series predicted by the first and second order difference alone deviates greatly.In order to eliminate random errors,the average value of the first and second order difference is taken as the final prediction result,the prediction effect was improved.