Abstract: Conventional linear regression and empirical formulas for predicting initial productivity have limited application scope and large prediction error, and are difficult to characterize the nonlinear variation of initial production under the influence of multiple factors. Therefore,machine learning algorithm was introduced to improve the accuracy of the prediction. Taking an ultra-low permeability oilfield as an example, ten factors affecting the initial productivity were selected from three aspects of geology, development,and engineering,and the linear correlation among these factors was analyzed by Pearson correlation. Random Forests was applied to identify the main controlling factors of initial productivity. Support vector machine(SVM)optimized by grey wolf optimizer(GWO)was used to establish the prediction model of initial productivity of oil well for the first time. The results show that the main controlling factors of initial productivity in ultra-low permeability oilfields are sand volume of the fracturing,perforation thickness, initial water saturation, effective oil layer thickness and sand intensity of fracturing. The GWO-SVM model has faster operation speed and higher accuracy compared with multivaluable linear regression and SVM optimized by grid searching. The research results can provide reference for the initial productivity evaluation of oil wells.

Key words: machine learning, Random Forest, grey wolf optimizer, support vector machine, productivity forecast