Research On Productivity Prediction Method Of Bedrock Buried Hill Reservoir Based On Machine Learning

In the development process of bedrock buried hill reservoirs,the analysis of factors affecting productivity and the dynamic prediction is an essential foundation for development plans,which have practical significance.The seepage law of bedrock buried hill reservoirs is complex,and conventional productivity prediction methods often assume too many conditions,and the applicable conditions are strict and cannot take into account the impact of production adjustments meanwhile.This article combines the field dynamically and static data of the bedrock buried hill reservoir and provides a new way of thinking for the analysis and dynamic prediction of productivity influencing factors based on the machine learning algorithm.This paper analyzes the geology and production situation of the target area and establishes a learning sample database.In order to improve the quality of learning samples,this thesis uses the isolated forest algorithm to remove outliers and utilizes the K-nearest neighbor algorithm to fill in missing values.For the processed data,this article analyzes and screens the influencing factors of productivity with the adaptive Lasso algorithm and concluded that the fracture density has the most significant impact on productivity.Combining the selected main control factors,this paper uses the K-means clustering algorithm to determine the model wells.The productivity diagnosis is performed through fuzzy pattern recognition to judge low-yield wells,middle,and high-yield wells.The test results show that the overall accuracy rate of the productivity diagnosis model is 80.93%,and the identification effect of low-yield wells is the best.This paper introduces production adjustment data such as nozzle size and daily opening and combined with production-related time steps and geological data to form a time series data set in the oil well productivity prediction.The first 80% of the data set is used as the training set to train the model,and the last 20% is used as the test set to test the model’s performance.The differential autoregressive moving average algorithm(ARIMA)determines the time step as 5,and constructs a production productivity prediction model that considers production adjustments with the recurrent neural network(RNN)and long short-term memory neural network(LSTM).This paper separately uses the Adam algorithm,SGD algorithm,and RMSProp algorithm to optimize the weights of the RNN and LSTM networks and optimizes the structural parameters by the leave-one-out method.Taking the average absolute error as the evaluation standard,the results show that the LSTM model has the best productivity prediction effect,which can accurately reflect the dynamic changes of daily oil production.Finally,this paper analyzes the impact of the production adjustment on the performance of the LSTM productivity prediction model and compares it with the deep learning model that does not consider the production adjustment process.The results show that the LSTM productivity prediction model considering production adjustment has higher accuracy and can more accurately characterize the change characteristics of actual output.At the same time,the LSTM productivity prediction model can effectively learn production adjustments such as opening and closing wells and changing the size of the nozzle,and predict the impact of production adjustments on production.