School of Electrical Information Engineering,Northeast Petroleum University
单井日产量趋势预测研究在油田生产中具有重要意义。由于油井生产工况复杂，难以准确预测日产量，本文建立了基于多变量时序数据的产量模型。基于卷积门控循环单元（Convolutional Neural Network – Gate Recurrent Unit, CNN-GRU）提取深层特征进行时序预测，基于梯度提升框架的决策树模型（Light Gradient Boosting Machine, LightGBM）从回归预测角度进行预测，两者结果相互融合，进一步提高产量预测精度。同时，本文提出了一种可以实现多变量时序预测或回归预测模型在未知输入特征情况下准确预测产量的方法—超前参数递归预测策略。采用该方法对影响产量的重要特征进行超前预测，并将预测到的重要特征应用于预测产量的仿真测试中。仿真结果表明：本文所建立模型与超前参数递归策略配合最好，在测试集上的预测准确度最高。相比单变量时序预测和回归预测模型，可显著提高预测精度。
A study on prediction for daily production per well is of great significance in oilfield production. Under complex production conditions of oil wells, it is difficult to accurately predict daily production. A production model based on multi-variable time series data is built in this paper. Based on the convolutional neural networks - gate recurrent unit (CNN-GRU), deep features are extracted for timing prediction, and a decision tree model (LightGBM) based on the gradient boosting framework provides prediction results from the perspective of regression prediction. The results of the two are integrated with each other to further improve the accuracy of production prediction. A method that supports multi-variable time series prediction or regression prediction to accurately predict production under unknown input characteristics – strategy for recursive prediction of advance parameters is proposed. This method is used to predict important features that affect production in advance, and the predicted important features are used in simulation tests on production prediction. The simulation results show that the model established in this paper works best with the advance parameter recursive strategy and has the maximum prediction accuracy on the test set. Compared with univariate time series prediction and regression prediction models, the prediction accuracy can be significantly improved.
杨莉,周子希,王婷婷,等. 基于CNN-GRU-LightGBM模型的单井产量 预测方法[J]. 科学技术与工程, , ():复制