| dc.description.abstract | Production data analysis for low permeability shale reservoirs is crucial in characterizing flow regimes as well as reservoir properties, and the forecasting of production is essential for portfolio and reservoir management. However, traditional methods have failed due to incorrect physics or complicated convolution from the well control history. In this research, we investigate several novel machine learning approaches and their variations to better facilitate production surveillance as well as accelerate the analysis of variable and noisy historical production data. Current production data analyses are on the basis of either simple constant drawdown-pressure rate solutions or constant rate pressure-drop solutions in which we cannot handle field data variation. Therefore, deconvolution methods have been developed to transform the variable pressure variable rate profiles into those two simple solutions. Pressure-rate or rate-pressure deconvolution is an ill-posed, complex time-series problems. Much research has indicated that Echo-State Networks (ESN) and Long Short-Term Memory (LSTM) are useful for dynamic time-series problems, however, the connections between physics and machine-learning-based solutions remain unpublished. Thus, one of the motivations of this research is to establish the connection between transient physics and machine learning algorithms specifically using ESN and LSTM paradigms. Traditional decline curve analysis (DCA) models have played an important role in the oil and gas industry for nearly a century because of their simplicity. However, forecasting models with too much simplicity sometimes rely on unnecessary assumptions and fail to honor important realities. Unlike those DCA methods which rely solely on rate histories, this research first introduces novel analytic methods to accomplish forecasting using pressure-rate-time information while still maintaining the reduced complexities. By switching the application from rate transient analysis to pressure transient analysis, these analytic models are demonstrated to be useful in identifying the reservoir model with only pressure-rate information.
In general, the purpose of this study is to present alternative methodologies for deconvolution to facilitate the production data analysis. We forecast future production in addition to applying the ‘inverse’ process for reservoir characterization. We also utilize the merits of physics-based training features to further enhance the production surveillance and finally, we illustrate the proposed workflows for production analysis, production forecasting and production surveillance using variants under the ESN and LSTM paradigms with synthetic cases and field examples. | en |
