Production forecasting of wells in unconventional reservoirs is an important problem for the industry. Type well construction is a commonly used procedure to forecast aggregated wells. However, "averaging" observed production profiles, a commonly used type well construction method, is strongly affected by the presence of wells with noisy production history. Forecasting individual wells on the other hand is time consuming, and inaccurate for wells with limited production history. In this work, we propose model-based type wells for forecasting production for individual wells, using scaling factors.
We cluster wells based on their observed flow regimes. We then fit a model to a representative well in each cluster. This model could be analytical or numerical, depending on the data available. We then use Bayesian hierarchical modeling to scale the model to each well. We also obtain the distribution of the mean scaling factor for the wells in the cluster. This accounts for the uncertainties in the model fit, the individual well scaling, and the spread of wells within the cluster. Finally, we obtain the P10, P50 and P90 production profiles and forecasts for the wells from the respective scaling factors.
We show that our technique has lower uncertainty and better prediction for noisy wells and wells with limited production data, when compared to individually fitted models. This is because our technique uses partial pooling to "borrow" information from wells with longer production histories. We demonstrate that our technique is robust to the inclusion of noisy wells in the cluster. We demonstrate the effect of the noise and the number of wells in the cluster, on the uncertainty of the forecast. We validate this technique with simulated data, with and without added noise.
An important goal in production forecasting is to avoid systematic overestimation of reserves. This goal can be achieved when the uncertainties in the estimating process are accounted for in a principled manner. In this work, we account for model, well and cluster uncertainties in the process of type well construction. At late times, when the number of producing wells to be averaged is lower, the widely used method of averaging type wells fails. Our technique is not affected by this problem, since we consider the entire production history of a well when calculating the shift factors. Using our method, we calculate production forecasts that are well calibrated to any required cumulative probability.