Quality estimation of reservoir fluid compositions from optical sensor measurements (OPTs) is crucial for real-time downhole fluid analysis. Conventional workflow, which performs multiple single-analyte predictions with neural networks (NNs), often suffers the issue of mass balance in estimated outputs. In this study, single-analyte fluid characterization (SAFC) is compared with a novel multi-analyte fluid characterization (MAFC) method. The investigation aims to evaluate whether the new method can maximize the balance of the density concentrations of fluid compositions and provide self-consistent solutions to support early decision making on reservoir exploitation and production optimization.
As an inversion problem, fluid characterization in this paper is accomplished in two stages using NNs. The first stage converts actual sensor measurements into synthetic optical responses. After the optical data transformation is completed, then fluid characterization at the second stage is applied using synthetic optical data inputs. The combination of optical data transformation and fluid characterization networks forms a concatenated optical computing network (COCN) for downhole fluid analysis. The performance evaluation is conducted on the calibration database first, followed by comprehensive testing assessment using simulation data and laboratory measurement data.
Fluid characterization networks developed on a large database pertain to low uncertainty for both multi-analyte and single-analyte calibrations. SAFC networks allow for selecting analyte-specific network inputs and manipulating the network performance under its own connecting parameter space. However, the inherent data relationship and constraints among the fluid compositions and properties are not utilized to calibrate individual models. This limitation often leads to inconsistent compositional predictions on the new data and hence requires post-processing of output data from multiple single-analyte networks. MAFC networks, in comparison, have constraints of fluid compositions built into training outputs. By minimizing overall training error on all compositional outputs, MAFC networks structurally ensure the self-consistent predictions on analyte density concentrations. Nevertheless, sharing network parameters in the hidden layers necessitates the use of more complex MAFC network structures to improve overall prediction accuracy. The quality of optical data transformation, which affects all fluid compositional outputs with less flexibility in input selection, can present challenges during MAFC network application.
Conventional optical data transformation networks calibrated on a small number of reference fluids are capable of supporting sensor-based fluid characterization but cannot justify the robust mapping for sensor-independent MAFC networks because of the limitation associated with sparse-data-based transformation model development. New transformation networks, which facilitate fluid analysis with generic optical inputs, are built through two-directional progressive calibration to improve data mapping. The resulting reverse and forward transformation networks can be optimized using measurement and simulation data and validated in real-time signal processing. Applying reverse and forward transformation networks alternatively, for example, can help determine the degree of matching between simulated and measured optical sensor data. Using the quality of matching as a performance index can also enhance uncertainty assessment of predictions from fluid characterization networks. It has been demonstrated that the novel transformation networks developed during this study provide more robust optical sensor data mapping and improve predictions for both MAFC and SAFC networks.