Abstract
Tight oil fields are affected by factors such as geology, technology, and development, so it is difficult to directly obtain an accurate recovery rate. The accurate prediction of the recovery rate is very important for measuring reservoir development effects and dynamic analysis. Traditional tight oil recovery predictions are obtained by conventional formula calculations and curve fitting, which are less applicable and very different from actual conditions. Machine learning can make accurate predictions based on a large amount of data, so it is used to predict the recovery rate of tight oil reservoirs. The recovery rate of 200 wells in M tight oil reservoirs ranges widely between 8.8 and 27.6%, with more than 14 factors affecting the recovery rate, and the overall declining rule is not clear. Therefore, this research combines the production data of horizontal wells with random forest, support vector regression (SVR), and other methods, establishing recovery prediction models to gain more accurate recovery predictions. First, the Pearson correlation coefficient and the random forest (RF) machine learning method are used to measure and calculate the degree of nonlinear influence of factors on oil well recovery. Second, SVR and optimization of support vector regression by particle swarm (PSO-SVR) recovery prediction models are developed and tested, with 75% of the data being used to train SVR and PSO-SVR recovery prediction models and 25% to verify the model. Third, the accuracy of the results of these two SVR oil recovery prediction models is compared, suggesting that when the data are scarce, the optimized model is more accurate than the unoptimized one by 10.85%. Thus, this model can assure a relatively more accurate prediction of oil recovery. Machine learning recovery prediction, being more accurate and applicable, enables the data of factors such as construction and production systems to be optimized in the future, enhancing the oil recovery rate.