A well production prediction method based on blending heterogeneous ensemble learning optimized by OOA

Abstract

Oil well production prediction is an important research content in oilfield development, and constructing a scientific and good prediction model is a key issue. In this paper, a blending ensemble learning oil well production prediction model combining Random Forest, LGBM, and TCN is established and optimized by the Osprey optimization algorithm. Firstly, ANOVA and Pearson's correlation coefficient is used to filter the relevant features affecting the well yield prediction, the autocorrelation coefficient is used to determine the lag order of the well yield prediction, and RF, LGBM, and TCN networks with good performance are selected as the base learners. Linear regression is used as the meta-learning to form the blending ensemble learning model, and the Osprey optimization algorithm is used to optimize the whole blending ensemble learning model. The Osprey optimization algorithm is used to optimize the whole blending model, and the optimal hyperparameters are derived and brought into the ensemble learning prediction model. After the experiment, it is shown that the ensemble oil well production prediction model can reduce the MAE by about 12.75% compared with the single prediction model, and can effectively predict the future production of oil wells.