Accurate production forecasting is crucial to the decision making process and evaluation of investment scenario both in conventional and unconventional reservoirs. Deep Learning (DL), which is a subset of Machine Learning (ML), show promising results in improving forecasting accuracy. Despite these promising results, DL algorithms are difficult to train due to their inherent weaknesses known as vanishing and exploding gradient problems. The main goals of this comprehensive review paper are: (1) encourage improvement in DL applications utilized in production forecasting by evaluating and outlining the strengths and weaknesses of both traditional methods and DL applications, and (2) provide viable solutions to overcome the weaknesses in DL applications in production forecasting tasks.
This comprehensive literature review study investigates DL resources both within and outside the spectrum of petroleum engineering to identify and remedy the weaknesses of DL applications in production forecasting. This paper is divided into four main sections: (1) an overview of the general process of data preparation for training DL algorithms, (2) an overview of relevant DL algorithms applied to production forecasting, (3) strengths and weaknesses of both traditional methods and DL algorithms in production forecasting, and (4) viable solutions to overcome the weaknesses of DL algorithms to improve forecasting accuracy. Additionally, reviewing and analyzing all of this information allows for comparisons of traditional methods and DL applications in production forecasting.
There are two major conclusions in this paper: (1) the DL algorithms are difficult to train due to their inherent weaknesses and, (2) if these weaknesses are remedied effectively, DL algorithms are hard to replace with other traditional methods due to their potential for improving forecasting accuracy. Additional observations include: (1) the deeper the network gets, the easier it gets for the vanishing and exploding gradient problems to arise and, (2) arbitrary implementations of the solutions to remedy the weaknesses of DL algorithms can create additional problems that are more difficult to solve. These additional observations outline the main reasons behind the difficulty of training DL algorithms.
This paper provides the much needed viable solutions to remedy the weaknesses of DL algorithms to improve production forecasting accuracy of both conventional and unconventional reservoirs. Most ML related review papers are not focused on a single subject, such as production forecasting. The value of this paper is more significant to researchers and industry professionals that are specifically interested in production forecasting. Additionally, this study serves as a guide to help improve the DL applications in production forecasting.