基于长短期记忆网络和注意力机制的油井产油量预测

Oil Production Prediction of Oil Wells Based on Long Short-term Memory Networks and Attention

准确预测油井产油量在油田生产中具有非常重要的意义。针对传统的线性预测方法中存在的适应性差问题,以及在时序处理时难以很好地拟合历史数据问题,提出使用长短期记忆网络和注意力机制来提取油田生产数据中存在的时序关系和增强油井产油量预测模型的可移植性,并分析了时间滞后、学习率衰减和神经元随机失活3个参数对油井产油量预测模型的影响,发现当这3个参数分别为36、0.3和0.8时,油井产油量预测模型的表现最佳。在利用随机森林方法补全动液面的缺失数据后,使用获得的3个最优参数建立油井产油量预测模型,并将该模型应用于中国南方某油田3口油井的产油量预测中。具体的预测结果是:H3-32井后期的实际产油总量为1470.5 t,预测值为1442.33 t,相对误差为1.92%;H3-34井后期的实际产油总量为1564.5 t,预测值为1545.98 t,相对误差为1.20%;H3-35井后期的实际产油总量为742.2 t,预测值为772.12 t,相对误差为4.05%。由此可见,基于长短期记忆网络和注意力机制的油井产油量预测模型具有较高的准确率。研究结果可应用于中国油田生产开发方案的制订,对中国油田科技水平的进步具有非常重要的理论与现实意义。

It is very important to predict oil production of oil wells accurately in oilfield.In view of the poor adaptability of traditional linear prediction methods and the difficulty of fitting well in processing time series data,the long short-term memory neural networks(LSTM)and attention mechanism were proposed to extract the relationship of time series data in oilfield production and enhance the portability of prediction model for oil well production.Moreover,the effects of the three parameters of time lag,learning rate attenuation and random neuron deactivation on the prediction model for oil well production were analyzed,and it is found that when the three parameters were 36,0.3 and 0.8,respectively,the prediction model for oil well production performed best.After the random forest algorithm(RF)was used to complete the missing data of dynamic fluid level,the prediction model of oil production for a single well has been established with the three optimal parameters,and the model has been applied to the oil production prediction of three oil wells in an oilfield of southern China.The detailed forecast results are as follows.The actual total oil production of H3-32 well in the later period is 1470.5 t,the predicted value is 1442.33 t,and the relative error is 1.92%.In the later stage of development of H3-34 well,the actual total oil production is 1564.5 t,the predicted value is 1545.98 t,and the relative error is 1.20%.In the later stage of development of H3-35 well,the actual total oil production is 742.2 t,the predicted value is 772.12 t,and the relative error is 4.05%.It can be seen that the prediction model for oil well production based on LSTM and the attention mechanism has a high accuracy rate.The research results can be applied to the formulation of production and development plans for oilfields of China since it has very important theoretical and practical significance on the progress of oilfield science and technology in China.