Abstract: Accurate identification and evaluation of water flooded layers are key technical bottlenecks for improving oil recovery in the middle and later stages of oilfield development. The original resistivity of the formation, as an important parameter reflecting the initial electrical characteristics of the reservoir, is of great significance for identifying water flooding and analyzing remaining oil. Considering the complex reservoir conditions of the target block, traditional inversion methods suffer from low precision and weak anti-interference capability, this paper proposes a deep learning combination model (CNN-BiLSTM-ATT) that integrates convolutional neural network (CNN), bidirectional long short-term memory network (BiLSTM), and attention mechanism (Attention) for high-precision inversion of original resistivity. The model fully combines the local feature extraction ability of CNN, the temporal modeling ability of BiLSTM, and the key feature focusing ability of Attention mechanism, improving the nonlinear modeling ability and overall robustness. The X oilfield in the Pearl River Mouth Basin is selected as the target block, and the model training is carried out in combination with conventional logging curves. The experimental results show that this model has significant advantages over other comparative models. On this basis, the resistivity decay rate (M) and the comprehensive parameter (GPR) are further introduced to establish a water-flooded zone identification chart, achieving fine classification of different water flooding levels with a recognition accuracy rate of 92.2%. The research results provide new and efficient technical ideas and means for the evaluation and fine development of water flooded layers in high water cut oil fields.