| Reservoir characterization is important for reservoir management and performance optimization. For waterflood optimization, traditionally several techniques have been suggested, most of which are either too time-consuming or the data needed are often unavailable. There is a new research trend to overcome these limitations by applying advanced statistical techniques on only injection and production data, which are often readily available for any waterflood operations.;In this work, we follow this trend to formulate the reservoir characterization and forecasting problems in waterflood projects using a system identification framework: the injection rates are seen as inputs; the production rates are seen as the outputs; and the reservoir is considered as a dynamic system. By addressing the properties of general linear dynamic systems, we discuss the limitations of previous models and build three new predictive models to characterize some reservoir behavior, such as producer-to-producer interactions, which was not considered in previous literature. Then we discuss a general parameter estimation approach under the prediction-error framework.;For model evaluation, we propose two techniques: one is based on evaluating their prediction ability on a fresh data set, while the other is based on comparing the interpretations they provide about certain reservoir characteristics with the ground truth. All proposed models are verified by these two approaches. To perform a comparative analysis, we provide a practical metric to compare the prediction performance of different proposed models under various scenarios. From the results, we make several observations and suggestions for reservoir engineers to use the models.;To clarify the relationship between different models, we develop a general linear modeling framework and demonstrate that all proposed models can be considered as special cases within this framework. Moreover, the transfer function of the general linear model can be interpreted to provide insight on reservoir characteristics. Also, the relationship between different models can easily be built from this work.;We propose a multivariate autoregressive model to characterize situations in which a producer is shutting-in or a new producer is being brought online. As a totally new application, we introduce a novel "constrained producer" approach which that only requires minimal changes in production rates (e.g., limiting them to some level below their normal production capacity) to predict the performance after a producer is shut-in. This allows us to handle various "what if" scenarios in waterflood management.;Finally, to achieve a better model estimation, the patterns of injection rates play an important role. We addressed the problem of designing a set of injection rates to achieve a better estimation of target parameters in the reservoir. Two different approaches, deterministic and stochastic approaches, are discussed. For the deterministic approach, we propose a new procedure using a set of inverse-repeat signals to design a set of signals with zero cross-correlation property. For stochastic approach, we applied a common approach in system identification and evaluate all design procedures on some predictive model. |
