The Research Of Reservoir Prediction Method Based On Random Regression Forests

Facing more and more complicated oil and gas reservoir,reservoir prediction needs one kind of nonlinear algorithm to mine underlying relation between seismic data and reservoir feature parameter for achieving the information of physical,lithological properties,etc.At present,substantial intelligent algorithms have been applied and studied in seismic reservoir prediction,having some advantages and disadvantages.Based on statistical theory,random forest algorithm performs convenient operation,high efficiency,strong ability to tolerate noise,not needing to worry about overfitting problem in many fields and resulting in favor of many scholars,but it has been seldom applied and researched in seismic reservoir prediction.If it proved to depict the geological features of complicated reservoir effectively,random forest will give convenience and help to oilfield technology research and production service.This paper focused on the applicability issue of random regression forest to seismic reservoir prediction.Combining actual data and simulated data,specific analysis and experiments did from feasibility,noise tolerance,sensitivity and limitation.Aiming at the problem found in the limitation study,this paper had done some corresponding improvement research.For the seismic attributes redundancy problem discovered in the sensitivity research,we studied seismic attributes selection based on variable importance of random forest.The studied results indicate that random forest algorithm has feasibility in seismic reservoir prediction and can be able to depict the geological features of complicated reservoir effectively.The seismic reservoir prediction technology based on random regression forest embodies the advantages of convenient operation,high efficiency and not needing to worry about overfitting problem.The error analysis results of test samples not reflect good or bad of prediction effect completely.The input variable data of sample has great influence in accuracy and prediction of random regression forest model.