رسائل ماجيستير
English
2022
Prediction of Well Production Rates For Multiphase Flow Using Machine Learning Techniques
Qinzhuo Liao , Shirish Patil , مياو جين
كلية هندسة البترول وعلوم الأرض-جامعة الملك فهد للبترول والمعادن · السعودية
Accurate estimation of well production rate of multiphase flow is critical for enhanced oil recovery and economic evaluation in the oil and gas industry. Numerical simulations, analytical methods, and empirical models (e.g., traditional decline curve analysis) are typically applied to predict well production data. However, numerical simulation is time consuming, while analytical methods and empirical models oversimplified the reservoir condition and may not reflect the dynamic production change.Machine learning (ML) techniques are comprehensively analyzed and applied to the petroleum industry. Several ML models have been adopted to predict well production rates, such as Artificial Neural Networks (ANN), Decision Tree, Support Vector Machine (SVM), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), etc.Manual operations, especially the frequent opening and closing operations, often lead to the unexpected variation of reservoir parameters and thus become one of the primary obstacles to actual well performance forecast. However, shut-in periods are rarely considered in the well production rate forecast in previous research. Therefore, after comparing different ML algorithms, we applied the multivariate LSTM to predict the actual well production performance, including shut-in periods. Specifically, we propose to improve the LSTM via two approaches: (1) Optimizing the hyperparameters of LSTM itself, such as neuron number, the number of epochs, batch size, etc. (2) Combining LSTM with ANN to construct the hybrid model.Through the hyperparameter tunning, the optimal hyperparameters were selected for each output. The oil LSTM model has one hidden layer with 41 neurons, the activation function is the sigmoid function, and the weight and bias optimizer is adamax. The gas LSTM model has one hidden layer with 45 neurons, the activation function is softmax, and the optimizer is adam. The water LSTM model has one hidden layer with 7 neurons, the activation function is softplus, and the optimizer is rmsprop. The optimal models' average relative mean absolute error (RMAE) is 15.92%, 16.73%, and 13.84% for oil, gas, and water production.A hybrid model was successfully developed to improve the prediction performance of the oil and gas production. The hybrid model integrated the LSTM and ANN models to consider the influence of historical performance and daily operation conditions for production rates. After the hyperparameters optimization, the hybrid models' average RMAE reached 12.08% and 12.14% for oil and gas production, respectively.