Abstract: Rate of penetration (ROP) is a key indicator for evaluating drilling efficiency. Its variation is jointly influenced by multiple drilling parameters and exhibits strong coupling and pronounced nonlinear characteristics. The ROP at the current time step is determined not only by the instantaneous drilling parameters and formation conditions, but also closely related to the historical evolution of ROP, drilling parameters, and drilling states over previous time steps, showing evident temporal dependence and sequence memory effects. To address the challenges posed by complex operating conditions, high noise levels in real drilling data, and the difficulty of traditional models in effectively capturing long-term temporal dependencies, this study investigates real drilling operations from a drilling site in Xiangyang and proposes a dual-layer CNN- LSTM-based ROP modeling method. First, to cope with the coexistence of multiple drilling conditions in raw data, a normal drilling condition identification and automatic well-section segmentation approach is developed. Combined with abnormal data cleaning, segment-wise filtering, and scale transformation, a systematic data preprocessing scheme is established. Second, a time-lagged mutual information analysis is employed to quantitatively analyze the nonlinear and temporal correlations between ROP and multiple drilling parameters, based on which the model input variables and the length of the time window are determined. On this basis, a ROP modeling framework is constructed by integrating multi-scale convolutional feature extraction with temporal sequence modeling through parallel dual-layer CNNs cascaded with an LSTM network. Experimental results demonstrate that the proposed model achieves superior accuracy and stability on real drilling data compared with traditional machine learning models and deep learning models with single network structures. The results indicate that the proposed approach can effectively capture the temporal evolution characteristics of ROP during drilling operations, providing a feasible data-driven solution for ROP modeling under complex drilling conditions.