Recently, machine and deep learning algorithms have been proposed as alternatives to statistical methods for production time series forecasting of unconventional reservoirs. Although most efforts provide timeseries forecasts using machine learning (ML) algorithms for unconventional reservoirs with acceptable performance, no comparison has been made with classical methods in terms of computational requirements and accuracy. In addition, most studies have focused on a single time series or short-term forecasting period, which raises concerns about the generalization of results.
Here, a detailed comparative analysis of the performance of ML algorithms and statistical methods is presented to predict the oil production profiles of four hydraulically fractured horizontal wells. To construct the production time-series database, four numerical flow simulations of a tight oil Stimulated Reservoir Volume (SRV) were conducted. The attributes included reservoir characteristics (porosity, permeability, relative permeability endpoints, saturation) and operating parameters (fracture half-length and fracture locations).
A comparative study was then conducted of six modern ML networks, including Multilayer Perceptrons (MLP), Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Convolutional Neural Network (CNN), Long-term Recurrent Convolutional Network (LRCN), and Gated Recurrent Unit (GRU) with two statistical methods (Exponential Smoothing, and Seasonal Autoregressive Integrated Moving Average). For performance evaluation, the mean squared error, and the execution time for hyperparameter tuning were examined.
After tuning the hyperparameters of each algorithm, the results showed that the statistical methods outperformed complex and sophisticated ML algorithms in terms of accuracy, while MLPs showed the lowest mean squared error among the considered ML algorithms. We also observed that the computational requirements, including hyperparameter tuning execution time, for the ML models were greater than those of statistical methods. The findings of this study also suggest the statistical methods to be considered as the baseline for assessing the performance of the predictive models for production profiles in unconventional reservoirs.