Well-test model identification and, subsequently, model parameters determination are more complex in horizontal wells than in vertical wells. This is due to the increase in number of flow regimes occurring during a flow period and the fact that strong correlation exists between model parameters.
This study presents a new approach for automatic model identification and computer-aided well-test interpretation in horizontal wells. The new approach is based on using neural network to (1) identify the well-test interpretation model; (2) identify flow regimes; and (3) mark the position of identified flow regions on the derivative plot of well test data.
This work consists of first generating common model signatures using Ozkan and Ragavan analytical solutions for horizontal wells in various reservoir and inner boundary conditions assuming laterally boundless reservoirs. Next, these signatures are used to train neural networks for three identification stages; model identification, flow regime identification, and position of flow regime identification. Separate networks were trained, then tested and validated using synthetic as well as field data. Once the three identification stages are completed, specialized plots for data points falling into each flow regime are used to determine initial model parameters. Final model parameters are determined using nonlinear regression.
A comparative study was carried out using different network architectures. Modular approach with direct data utilization is found to be most suitable for implementation of our approach.