CG-SAWGAN-Based Sequence Sample Augmentation Method for RSS Well Path Prediction

Well path prediction is a key part of directional drilling. During the downhole operations of Rotary Steerable Systems (RSS), issues such as high data acquisition costs and limited effective samples severely restrict the accuracy of drilling well path prediction. To address this problem, a Clustering - guided Sequence - aware Wasserstein Generative Adversarial Network (CG - SAWGAN) is proposed in this paper by integrating clustering and time - series prediction algorithms.It is applied to the sequential augmentation of drilling samples, thereby breaking through the accuracy bottleneck of RSS drilling well path prediction under small - sample conditions via a data - driven approach.Firstly, a multi - dimensional time - series feature space is constructed by extracting downhole measurements during drilling (such as inclination, tool face angle, and weight on bit). Subsequently, RSS operational data are clustered based on criteria such as Dynamic Time Warping (DTW) and cosine similarity, thereby partitioning the feature subspaces of drilling parameter sequences. Finally, a Sequence - Analysis Generative Adversarial Network (SAWGAN) is developed. This network employs Gated Recurrent Units to capture the temporal dependencies of drilling parameters, incorporates self - attention mechanisms to identify the statistical distributions of key features such as well path inflection points, and enhances training stability through the use of Wasserstein distance constraints.Experiments have shown that the stability of all types of CG-SAWGAN models is significantly improved compared to the non-clustering SAWGAN model. After injecting pseudo-samples at ratios of 50%, 100%, and 150% into the LSTM prediction model, the R2 of wellbore inclination rate prediction is increased by 3% to 5% compared to that before augmentation. By integrating clustering-based prior knowledge with deep generative models, this method provides a reliable data augmentation solution for intelligent decision-making in RSS drilling under complex geological conditions.