A TCN–LSTM–Attention Framework for Rate-of-Penetration Prediction from Multivariate Drilling Sequences

Rate of penetration (ROP) depends on interacting drilling conditions and formation properties, making prediction from sequential measurements challenging. This study develops a framework combining a temporal convolutional network (TCN), long short-term memory (LSTM), and attention for ROP prediction using logging data from the offshore BY area in China. Convolutional encoding and recurrent integration represent historical drilling sequences, while attention adaptively aggregates their hidden states. In the primary four-model comparison on a held-out depth interval, the framework achieves R2 = 0.9093 and MAE = 0.8981, against 0.7474, 0.7737, and 0.7915 for LSTM, TCN, and TCN–LSTM. Configuration selection is separated from testing. The controlled comparisons needed to attribute this difference to individual components, to uniform aggregation, or to model capacity have not been completed, and the reported difference is, therefore, described as an observed comparison rather than as an established component effect. Attention weights characterize internal aggregation and are not interpreted as direct evidence of lithological change or pressure event recognition. Generalization to independent wells and practical deployment require further validation.