南海油气资源是我国重要的能源接替区，但储层埋深大多较深，高围压下岩层展现的强塑性和复杂的地质环境严重影响了钻井时效，精确预测钻速也十分困难。基于此，针对南海巨厚泥岩地层具有独特的黏弹性和强塑性特征，建立智能钻速预测模型。该模型以南海某区域10口井的实际数据为样本，首先进行预处理，寻找离群值、降噪和标准化后，排除了若干影响因素；其次对5种实测地层特征(含地震波速、孔隙压力、破裂压力、上覆压力和地层岩性)使用因子分析，得到5种地层特征在3个公共因子下的关系；随后基于K-Means++算法进行分析，利用轮廓系数为指标，得出了该区域的地层聚类主要划分为2种地层类型，分别为以泥岩和粉砂质泥岩为主的地层类型和以粉砂岩、细砂岩和中砂岩为主的地层类型；在此基础上，引入5种地层特征，训练KNN分类模型，实现了对地层类型的准确预测；最后针对不同的地层类型，使用随机森林就不同的地层类型分别建立钻速预测模型，并在建立时使用经过贝叶斯优化算法进行超参数优化，得到了最适合的超参数组合。测试结果表明，所提出的基于地层分类预测的钻速预测模型在测试集的数据环境下，R2达到0.991， ERMS达到0.018，EMA达到0.011，相比其他常规机器学习算法在该区域具有更高的预测精度。本研究可为寻找地层潜在分类对钻速预测精度的影响提供参考。

In terms of petroleum and gas resources, South China Sea is the important energy replacement area in China. However, most of the reservoirs are buried deep, and the strong plasticity of the formation under high confining pressure and the complex geological environment seriously affect the drilling efficiency. It is also very difficult to accurately predict the ROP. Hence, a set of intelligent ROP prediction model was established for the extremely thick mudstone formation with unique viscoelastic and strong plastic characteristics in South China Sea. The model took the actual data of 10 wells in an area of South China Sea as a sample. Firstly, the sample was preprocessed, and had the influencing factors excluded through outlier screening, noise reduction and standardization. Secondly, factor analysis was conducted for the five measured formation characteristics (including seismic velocity, pore pressure, fracture pressure, overbudrden pressure and formation lithology), obtaining the relationship between the five formation characteristics under three common factors. Then, based on the K-Means++ algorithm, it was concluded with the silhouette coefficient as the index that the formation clustering in this area was mainly divided into two types, the formation mainly composed of mudstone and silty mudstone, and the formation mainly composed of siltstone, fine sandstone and medium sandstone. On this basis, the KNN classification model was trained by introducing five formation characteristics to achieve the accurate prediction of formation types. Finally, Random Forest was used for different formation types to establish the ROP prediction models accordingly. Besides, Bayesian optimization algorithm was used to optimize the hyperparameters at the establishment of models, and in this way the most suitable combination of hyperparameters was found. The testing results indicate that the proposed ROP prediction model based on formation classification prediction had a R2 of 0.991, ERMS of 0.018, and EMA of 0.011 in the data environment of the testing set. Compared with other conventional machine learning algorithms, it had higher prediction accuracy in this area. Generally, this study could provide a reference for identifying the influence of formation potential classification on the prediction accuracy of ROP.