Abstract: Conventional methods for reservoir lithology identification suffer low precision and efficiency since reservoir lithologies have various types and complex compositions and alternate frequently.This study proposed a reservoir lithology identification method based on multi-scale time-frequency-space feature combination.Based on the original logging characteristics,this method introduced the multi-scale frequency-domain components from the complementary ensemble empirical mode decomposition (CEEMD) to improve the longitudinal resolution of log curves.Moreover,a multi-scale convolutional neural network-bidirectional gated recurrent unit-attention mechanism (CNN-BiGRU-AT) model was constructed to extract the spatio-temporal features of log data containing multi-scale frequency-domain components.In this way,the joint learning of time-frequency-space features of log data was realized.Finally,the model output was optimized using the attention mechanism to reduce the propagation of error information.To verify the reliability of this method,an experimental analysis was conducted using the data from five wells that have relatively complete data.As revealed by the analysis results,the identification accuracy of training and verification sets containing multi-scale frequency-domain components was increased by 9.50% and 8.66%,respectively in the comparative experiments of different data combinations.The method proposed in this study yielded sample identification accuracy of 94.11%.Compared with support vector machine (SVM),backpropagation (BP) neural network,convolutional neural network (CNN),bidirectional gated recurrent unit (BiGRU),and CNN-BiGRU fusion models, the identification accuracy of this method increased by 16.21%,14.54%,11.69%,5.05%,and 3.38%,respectively.