Well path lightweight prediction model construction method for rotary steerable system based on composite knowledge distillation

The rotary steerable system (RSS) is a critical directional drilling technology where accurate well path prediction is essential for operational precision and efficiency. Existing methods struggle to balance accuracy and model complexity: high-accuracy models are too large for downhole deployment, while lightweight models lack the accuracy required for closed-loop control. To address this, we propose a novel RSS well path prediction method combining time series forecasting (TSF) with knowledge distillation (KD). A high-precision teacher model is first constructed by integrating graph representation learning with TSF. A novel isomorphic knowledge distillation method, termed intermediate layers feature adaptive fusion supervised KD, is then employed to extract and transfer intermediate layer knowledge from the teacher. Finally, a composite distillation framework combining intermediate and output layer knowledge trains a lightweight student model, with the loss function dynamically optimized via a learnable weight allocation mechanism. The resulting student model achieves 98% compression while maintaining high accuracy, with only a 1.3% drop in R2 and a 16% increase in MAE compared to the teacher, realizing high-precision, lightweight RSS well path prediction.