In unconventional reservoirs, several horizontal wells are drilled within each well pad, and could have different production profiles. Automatic history matching (AHM) is the process of inversely calibrating fracture geometry based on field production. The main challenge of AHM is to reduce the uncertainty of model parameters for a single well. This challenge is amplified for multi-well history matching, and limited research has been dedicated to AHM of multiple wells. This work presents a systematic workflow to be the first capable to characterize fracture configurations and other properties of a well-pad with only 6 hours of computational cost for 2 wells' calibration.
We initialize a mean vector and a variance-covariance matrix with zeros on the off-diagonals of model parameters for each well and sample 50 sets of uncertain parameters from the initialized Multivariate Gaussian distributions (MGD). Each sample of model parameters is fed into the embedded discrete fracture model (EDFM) along with a reservoir simulator to obtain the modeling result. A multi-objective loss function is used to compute the global error where 10 best modeling results with minimum global errors are selected, and these 10 sets of model parameters are used to update the mean vector and the variance-covariance matrix of the MGD. In the second iteration, the model parameters are sampled from the updated MGD, and the process repeats until the modeling results converge to history value based on a user-defined error threshold. We applied this novel workflow to two shale oil horizontal wells in Uinta Basin. The results show exceptional match between the best simulation model and the field production observation. The final variance-covariance matrix shows significantly reduced uncertainty in all model parameters compared to the initial variance-covariance matrix. The variance-covariance matrix also captures the inter-correlations between model parameters where the inter-correlations act as a sampling constraint which eliminates non-physical samples and significantly improves sampling efficiency. The proposed workflow's performance is robust against poor initial parametrizations. This is because the mean vector is updated through each iteration and always shifts it towards the optimal combination of model parameters, even when the initial iteration samples are sub-optimal. The proposed workflow has improved the matching between modeling results and the field observations at only a quarter of computational cost compared to other AHM workflows. Results show that our workflow obtains better global optimums of model parameters with high precision and allows us to provide superior characterizations of fracture properties/geometries of multi-wells in a well pad setting, providing valuable suggestions for well spacing optimizations.
