Research And Application Of Petroleum Peak Model Based On Machine Learning

Oil resources don't only affect a country's economic development, and it is the indispensable strategic resource of a country. Peak oil prediction and the prediction of time, can provide valuable strategic decision for oil early warning information. To carry out the theory of machine learning methods in the research of peak prediction model, can better enhance the rationality of description ability and improve prediction.This paper introduces the related model of peak oil and machine learning methods. At first, based on the data deleted model anomaly detection and its processing method has been researched and practiced. And then, Hubbert model suitable for China's oil production forecast is established based on the thought of gradual regression forward. At last, on the basis of a large number of studies of peak oil prediction at home and abroad, combining with the different machine learning algorithms, a variety of modeling method for prediction of peak oil were studied. Particularly the multimodal forecast models of describing multi cycle production trend characteristics were put forward. This article main research content is as follows:(1) The data deleted model detection method is designed to do anomaly detection of China's oil production data.(2) Hubbert model suitable for China's oil production forecast is established by using the idea stepwise regression. In both cases of that outliers is not processed and corrected by the use of linear extrapolation, different URR(the Ultimate Recoverable Reserves)value is obtained based on the idea stepwise regression. And several more representative URR values are selected out for comparison of model building and forecasting. Finally the Hubbert peak prediction model when the corrected URR is 13.272 billion tons is more suitable for Chinese oil production forecast.(3) Segment Prediction Model based on unimodal model, Multimodal prediction model based on piecewise linear approximation, Multimodal Prediction Model based on polynomial curve fitting and Multimodal Prediction Model based on dynamic programming are proposed. During the model process, the automatic detection of the number of peaks and the split point method based on trough feature recognition, polynomial curve fitting and dynamic planning strategy are also proposed. These four multimodal prediction models are studied by using of a zone of a domestic oilfield production data. The results showed that in complex degree, peak block recognition methods, trends, etc., different models have different advantages and disadvantages:Simple polynomial curve fitting to deviate from the actual future prediction data, such data does not apply to oil production and the small amount of volatile data in the data set; Multimodal Hubbert prediction model based on piecewise linear approximation is the most simple, but its peak identification method is not so exact as the peak identification method based on trough characteristics; Segment Prediction Model based on unimodal prediction model is more suitable for the multimodal production data which overall growth trend is smooth; Multimodal Prediction Model based on dynamic programming is more suitable for the multimodal production data which overall growth trend is upward.