Production Forecast Modeling And Application Based On Machine Learning

With the Development of big data technology,data-drive analysis has been widely applying to oil and gas exploration and development area.Data-drive analysis provides new solution for many problems which traditional approach can't deal with,one of the problems is Production forecast.Because the main controlling factors of Production in various development zones are very different,traditional mechanism-drive models are lack of generalization,that makes forecast effect unsatisfied when geological and engineering conditions are complex.Instead,data-drive approach more focuses on mining Production data characteristics rather than physical mechanism thus may quick obtains forecasting models with general applicability and reduce the over reliance on model and data condition.Among data-drive approaches,machine learning is an effective forecast method which can be used to solve regression and classification problems.Machine learning is popular in many areas and now becoming a common approach in Production forecast.At present many researches of Production forecast are based on machine learning,but some have common defects.The first is the limitation of generalization.For example,some researches concentrate on choosing optimal learner and optimize model parameters.While because of the difference of development zone,the model optimized by one dataset may not fit for other dataset thus doesn't take the strong generalization advantage of data-drive approach.The second is the application scenarios are not enough.One important application scenario of Production forecast is using static well parameters and existing production curve to forecast future production curve,which should be describe as a delay time series encoding-decoding problem,but relate research are lack or wrong in modeling.The third is the lack of application study.Production forecast models are not only valuable for economic evaluation and risk assessment,but also apply to analyze main controlling factors of Production and optimize production design.How to improve the Production and reduce costs is always overlooked in researches.To solve these defects,we complete following improvements studies:1.Study of Production forecast frame design and data pre-processing.We analyzed the similarities and differences between conventional big data and petroleum big data,then compared advantages and disadvantages between data-drive model and traditional mechanism-drive model in Production forecast task.According to the analysis results,we summarized the general process of Production forecast modeling and represented various Production forecast problems in machine learning.Besides,we take the feature engineering of horizon records in practical datasets as example to explain how to encoding petroleum data using petroleum knowledge,thus we can use more unstructured features while modeling.The Production forecast frame and the preprocessing method are general,which apply to other problems and datasets.2.Study of Production forecast machine learning modeling.We built static or dynamic model for various Production forecast problems and analyzed the applicable conditions of machine learning models in datasets under different data conditions.Specially,to predict future production curve by static well parameters and existing production curve,we designed BPNN-RNN and BPNN-RNN-RNN time series models and applied them,then integrated non-time series model to improve forecast performance.3.Study of Production forecast application.We make univariate and multivariate parameters analysis to find the main control factors of Production,then compared the difference of conclusions and applicable conditions between machine learning approach and traditional correlation analysis approach.At last we proposed the approaches to optimize the engineering parameters of new wells and the production measures of old wells that can help reduce production costs and improve Production.