A Novel Approach to Sand Volume Prediction Using Machine Learning Algorithms

The objective of the paper is to discuss the application of different Machine Learning (ML) algorithms to predict sand volume during oil production from a weak sandstone reservoir in Kazakhstan. The field data consists of the data set from 10 wells comprising such parameters as fluid flow rate, water cut value, depth of the reservoir, and thickness of the producing zone. Six different algorithms were applied and root-mean-square error (RMSE) was used to compare different algorithms. The algorithms were trained with the data from 8 wells and tested on the data from the other two wells. Variable selection methods were used to identify the most important input parameters. The results show that the KNN algorithm has the best performance. The analysis suggests that the ML algorithm can be successfully used for the prediction of transient and non-transient sand production behavior. The algorithm is especially useful for transient sand production, where sand burst is followed by abrupt decline and finally stops. The results show that the algorithm can fairly predict the peak sand volumes which is useful for sand management measures. The variable selection studies suggest that water cut value and fluid flow rate are the most important parameters both for the sand volume amount and accuracy of the algorithm. The novelty of the paper is an attempt to predict sand volume using ML algorithms while existing studies focused only on sanding onset prediction.