Reservoir characterization has been generally performed by using simple average, geostatistical and stochastic methods with local static data such as core and logging data. However since the mid-1990s, more precise reservoir characterization has been ...
Reservoir characterization has been generally performed by using simple average, geostatistical and stochastic methods with local static data such as core and logging data. However since the mid-1990s, more precise reservoir characterization has been achieved by the inverse calculation with dynamic data such as production data from producers.
This study presents an inverse model that integrates a reservoir simulator for forward modeling with a genetic algorithm as an optimization method. To reduce calculation time and resolve a problem of a memory lack, individual- and memory-shared parallel processing methods were applied to make the genetic algorithm parallel in the developed model. To execute the parallel inverse model, a cluster was constructed using personal computers.
The validity and the feasibility of the developed inverse model were analyzed by carrying out the characterization with simulated production data on a heterogeneous reservoir system. For the first phase, the porosity and permeability distributions were calculated using the Kriging method with just static data, such as core data like porosity and permeability. Then, the results of the reservoir simulation, which used these distributions as its input data, were compared with those of inverse calculation. The difference between the calculated and the observed pressure of producer OP-4 was 3.02% with the Kriging method, and 0.57% for the inverse calculation of the memory-shared method. Thus, the inverse calculation yields better results of reservoir characterization than the Kriging method.
On the other hand, the inverse calculation has been performed by the individual-shared method to use not only static data, but also pressure form the producers and water-oil ratio. Results of the inverse calculation shows that, for a parallel computing system used in this study, the individual- shared method reduce the calculation time of the conventional serial method by a 3.8 times.
Next the inverse calculation results of the individual- and memory-shared methods were compared. As for calculation time, the memory-shared method reached convergence after 249 generations, much slower than the 93 generations of the individual-shared method. This is due to the creation of new individuals caused by exchanging the individuals between the nodes at each generation, and the need for additional calculations for them in memory-shared method. When the pressure matching results calculated by the two parallel methods were compared, the individual-shared method was faster in calculation time as mentioned above but with the reservoir characterization, the memory-shared method showed a relative error of 0.57% at producer OP-4, superior to the 1.54% of the individual-shared method.
Lastly, the memory-shared method equally divides the individuals at each node for calculation, while the individual-shared method carries out calculations for all individuals at a master node. It may introduce that a problem of a memory lack as the number of individuals increase. To confirm the assumption, we increased the number of individuals from 200 to 400 in the inverse calculation. As a result, the required memory for the individual-shared method was 1024 MB, while the computing system used had 512 MB, which made it impossible. However, the memory-shared method required only 256 MB per node, safely running under the same conditions.