Reservoir gas-bearing identification is a crucial link in natural gas exploration and development. However, under the conditions of sparse well distribution, challenges such as insufficient prior information and low identification accuracy often arise. In response, we have developed a method for identifying the gas-bearing of tight clastic reservoirs based on the fusion of well logging and seismic data. Initially, feature engineering processing is applied to P-wave slowness, S-wave slowness, and density to calculate petrophysical parameters and fluid indicators, using the random forest algorithm to refine reservoir-sensitive parameters. Subsequently, a deep-learning network named modern temporal convolutional network (ModernTCN) is constructed, which uses reservoir-sensitive parameters as inputs for model training and testing. This network features a decoupled design to segregate temporal and feature information, effectively capturing the longitudinal gas-bearing characteristics of the reservoir. Next, we compare the ModernTCN with a fully convolutional network, a multiscale fully convolutional network, a self-attention bidirectional long short-term memory network, a self-attention bidirectional gated recurrent unit, and a 1D deep residual network to determine the reliability of this method. Furthermore, a probabilistic weighted voting algorithm based on the multi to single strategy captures the spatial correlation among adjacent seismic traces to predict lateral distribution characteristics of the reservoir. Ultimately, the methodology is applied using seismic and well data from the Huangyan structural belt of Xihu Sag, China, to identify gas-bearing zones in tight clastic rock reservoirs. The results are validated through the reflection characteristics of the test well logs and along-horizon slicing. This approach determines substantial well-seismic concordance, confirming its potential to support the exploration and development of tight clastic gas reservoirs.
Read full abstract