Enhanced real-time drilling trajectory prediction via SAC-optimized LSTM Networks

Abstract Accurate prediction of wellbore trajectory is critical for drilling safety and efficiency. However, strong parameter coupling and rapid trajectory variation in complex geological conditions challenge conventional models. Long Short-Term Memory (LSTM) networks are highly sensitive to hyperparameter settings, and fixed configurations often lead to overfitting and a delayed response to dynamic drilling environments. To address this issue, this paper proposes a SAC-LSTM framework integrating the Soft Actor-Critic (SAC) algorithm with LSTM. Unlike traditional offline optimization strategies, SAC enables continuous hyperparameter adjustment within a continuous action space, maintains exploration via entropy regularization, and dynamically adapts parameters using real-time data feedback. Combined with incremental training, the model continuously learns from new measurements. Field data from three wells demonstrate R 2 improvements of up to 0.162 (inclination) and 0.129 (azimuth) over conventional LSTM, confirming the superior accuracy and adaptability of the proposed method in real-time trajectory prediction under evolving drilling conditions.