Comparison of Decline Curve Analysis DCA with Recursive Neural Networks RNN for Production Forecast of Multiple Wells

摘要

Production forecast can significantly influence field development planning and economic evaluation. Traditional methods including numerical simulations and decline curve analysis models (DCA) requires extensive domain knowledge or lack of flexibility in modeling complex physics. However, data-driven techniques using recursive neural networks (RNN) have proven very efficient and accurate in time-series forecasting related applications. This study implemented and compared RNN with DCA in production forecast of single and multiple wells. A typical RNN based long short-term memory (LSTM) models were first developed with various input and output sequences. Then, well-known DCA models such as Duong, Stretched Exponential Decline (SEPD), Power Law Exponential Decline (PLE) were implemented as reference solutions. Moreover, data cleaning process involves preparation of history production rates and well constraints for existing wells. For multiple wells, similar input parameters were aggregated together for adjacent wells before declining forecast using the former model. Finally, hold-out training and validation were performed, followed by comparison of model accuracy and efficiency. Various LSTM based sequence-to-sequence models such as one-to-one, many-to-one, and many-to-many were successfully implemented for production forecast. Feature engineering was performed to generate additional features to facilitate training process. It was observed better agreement for the blind-forecasting validation dataset (i.e., last 20% of the given history) between LSTM model prediction and history production than DCA based models. LSTM models captured the overall trend whereas DCA only produced smooth curves. In addition, LSTM based models yielded good matches for all three-phase rates whereas DCA was usually limited to a certain phase. Moreover, for multiple wells, a group of neighboring wells with variable history lengths were used for training the model to forecast the production rates, where the modeling process is similar as character translation in natural language processing. Finally, it was demonstrated that the developed RNN based sequence-to-sequence models will be readily extended to model other time-series related problems such as condition-based maintenance and failure prediction. This study proposed a novel approach to model time-series related problems (e.g., production forecast) using the RNN based sequence-to-sequence models. The developed data-driven approach makes the process of history matching and forecasting efficiency and accurate for assets with or without decent operation history information. In addition, the algorithms and case studies herein were developed with open-source libraries, which could be readily incorporated into either in-house or commercial packages.