A Hybrid, Neuro-Genetic Approach to Hydraulic Fracture Treatment Design and Optimization

Abstract

Summary This paper summarizes the efforts conducted toward the development of a new and novel methodology for optimal design of hydraulic fracture treatments in a gas storage field. What makes this methodology unique is its capability to provide engineers with a near optimum design of a frac job despite very little (almost none) reservoir data availability. Lack of engineering data for hydraulic fracture design and evaluation had made use of 2D or 3D hydraulic fracture simulators impractical. As a result, prior designs of hydraulic frac jobs had been reduced to guess works and in some cases dependent on engineers with many years of experience on this particular field, who had developed an intuition about this formation and its possible response to different treatments. This was the main cause of several frac job failures every year. On the other hand, in case of relocation of engineers with experience on this particular field the risk of even more frac job failures was imminent. The unique design optimization method presented here is a logical continuation of the study that was covered in two previous papers. To thoroughly understand this methodology, reading of these two references are highly recommended. This method will accept available data on each well, which includes basic well information and production history, and provides engineer with a detail optimum hydraulic fracture design unique to that well, along with the expected post-frac deliverability. Please note that there are no simulated data throughout this study and all data used for development and verification of all methods are done using actual field data. Background In two previous papers a systematic approach using a three-layer back propagation neural network was introduced. The approach assisted engineers in predicting post-stimulation well performance and to select candidate wells for stimulation treatment. In those papers it was mentioned that this approach can also be used to optimize the stimulation design parameters. The optimization of frac design is the subject of this paper. Unlike conventional simulators that are based on mathematical modeling of the fracturing process, the process introduced in these papers used no specific mathematical model. As a result, access to explicit reservoir data such as stress profile, porosity, permeability and thickness is not essential. This was the major advantage over conventional hydraulic fracturing simulators, which can translate to considerable savings since it eliminates the need for expensive data collection. The application of this methodology to a gas storage field was presented. It was demonstrated that the developed neural network can predict the post-fracture well deliverability with better than 95% accuracy. These results were achieved in the absence of reservoir data (permeability, porosity, thickness and stress profiles) that makes conventional fracture simulation impossible. A complete version of these papers can be downloaded directly from our World Wide Web site at http://www.pe.wvu.edu. Genetic Algorithm Many problems in life are solved through some kind of searching process. In a world of almost unlimited combinations, we need to find the best time to schedule meetings, the best mix of chemicals, the best way to frac a well, the best stocks to pick, or the best way to stack boxes. The most common way we solve simple problems is the "trial and error" method. The problem is that search spaces are frequently too large for us to examine every possible combination. The model being investigated in this study has seventeen parameters which has been encoded into a 74 bit long chromosome. A chromosome is the binary form of all parameters concatenated to form one member of the genetic population. All the possible combinations of genes within this chromosome makes this problem to have 1021 distinct possible solutions. If we could examine 106 solutions per second, it would still take us 1015 seconds (about 300 million years) to exhaustively search the model space. In the past, people would solve problems like this by making intelligent guesses about the values of the parameters. and with whatever trial and error as they could afford, time-wise. P. 287