Abstract
This paper presents a workflow that evaluates whether the production rate decline or abnormal pressure behavior is attributed to formation damage. The workflow uses artificial neural networks (ANN) as the engine behind the evaluation process. All the training and validation data comes from a reservoir simulator. The required parameters that are used for building the neural networks are the completion factor (RF) and skin. These parameters are then used to match the production and pressure profiles of the subject well.
Formation damage refers to the reduction of the effective permeability near the wellbore. There are several causes of formation damage including but not limited to the completion and workover fluid, fines movement, and scale deposition. In this paper, the effect of plugged perforation sets and skin are examined. If the reservoir simulation model has a suitable history match, the model response should not considerably deviate from the real well response. If that deviation occurs, the developed ANN is introduced to check if that deviation is a result of formation damage. The reservoir simulator is used to generate several data sets with variable completion factors (RF) and skin parameters to be used for training and testing the ANN. The RF and skin can vary through time to reflect the actual changes in production response. The ANN then outputs the RF and skin factor based on the observed pressure and production profiles. The recommended output of the ANN is eventually validated with the reservoir simulator.
This workflow has been tested on a synthetic, heterogeneous reservoir model. Results show good prediction capabilities for the developed ANN in terms of producing formation damage parameters that correspond the simulated data. The workflow also suggests which parameter (the RF factor or skin) should be updated in the reservoir simulation model to reflect the observed production response.
The results indicate that the data for the subject wells in the reservoir simulation model can be updated in a fast and efficient way using ANN. The ability to vary the number of layers, number of neurons, learning rate, and training algorithm makes the ANN a suitable tool for tackling problems where the relationship between parameters is difficult to comprehend.
