Research Of Reservoir Production Index Forecast And Rapid Production Optimization Based On Data Driven

Reservoir index prediction and production optimization are important task in oilfield development and many methods have been developed by predecessors.A large amount of various types of data has been accumulated through long-term production process in oilfield,which provides valuable data resources for data mining research.In this paper,statistical method,time series analysis method,neural network algorithm and quantum-evolution intelligent optimization algorithm were used to carry out research on reservoir production index forecast and rapid production optimization.This article mainly has the following achievements and innovations:(1)The establishment of the learning sample library: k-means++ cluster analysis was applied to clean up the raw data.Then 20 features were constructed by performing data fusion for the dynamic and static indicators of reservoir.Finally,the principal component analysis method was used to transform it into linearly independent principal component dimensions.(2)Establishment and optimization of production index prediction model: two improved methods were proposed on the basis of deep LSTM neural network.On the one hand,the LSTM neural network was used as a base learner to perform cyclic learning on the residuals,on the other hand,it was combined with the ARIMA time series model.Then the layer number and hyperparameters of the model were optimized.In the hyperparameter optimization,the Bayesian algorithm was mainly used to optimize the 32 hyperparameters of the model.Finally,each oil well was modeled and the reservoir dynamic production prediction process was designed.(3)Rapid production optimization: firstly,the mathematical model of injection-production optimization was established,including clarity of the objective function and constraint conditions.Then the quantum evolution injection-production optimization method was formulated,which was combined with the reservoir dynamic prediction model.The iterative evolution of injectionproduction scheme was realized by the quantum gate rotation evolution,quantum crossover,quantum mutation in the quantum evolutionary algorithm,and the fitness of the injectionproduction scheme was evaluated by the prediction model.Finally,the optimal injectionproduction scheme was obtained.(4)Automatic prediction and optimization software programming: the modularization,automation and visualization of the above operations was mainly realized.Field example verification shows that:(1)The sample base processed by principal component analysis method contains more effective information,which can make the prediction result of the model more excellent;(2)The two-layer LSTM neural network model is more suitable for reservoir production index prediction.The proposed model improvement and optimization method can effectively improve the prediction accuracy.The result shows that the average relative error of the reservoir dynamic prediction model for predicting oil production in the next 12 months is only 3.91%;(3)The 4 comparatively excellent schemes obtained by the quantum evolution injection-production optimization method were verified by numerical simulation.The result shows that the two methods are not only in the same order of oil production calculated under different schemes,but also have an accuracy rate of 82% in predicting the increase or decrease trend of single well oil production,indicating that the method in this paper and the numerical simulation method have similar forecasting results.In summary,the dynamic prediction model and injection-production optimization method established in this article have high practical significance.