Heterogeneous Multi-Task Learning for Real-Time Collaborative Prediction of Multiple Interrelated Drilling Parameters

Accurate real-time prediction of drilling parameters, such as Rate of Penetration (ROP), Bottom Hole Pressure (BHP), and Formation Pressure (FP), is vital for optimizing drilling efficiency and ensuring operational safety. However, these parameters are interrelated, and traditional machine learning models often overlook their vital interdependencies, limiting predictive performance. We propose a novel Heterogeneous Multi-Task Learning (HMTL) framework to capture these complex non-linear interdependencies and enable synergistic co-prediction of these crucial drilling parameters. The HMTL framework introduces a novel architecture that intelligently integrates heterogeneous individual predictive models for ROP, BHP, and FP, each grounded in mechanism-data fusion. This architecture comprises: (1) task-specific expert networks, leveraging the strengths of diverse base models tailored to each parameter's unique characteristics; (2) a shared expert layer, distilling common underlying features from comprehensive drilling data; and (3) adaptive gating mechanisms, dynamically balancing task-specific insights with shared knowledge. Moreover, four loss balancing strategies were evaluated with hyperparameter tuning using Bayesian Optimization. Validation on field drilling datasets from the Tarim Oilfield demonstrates the HMTL model's superior performance. The Adaptive Weighting (AW) loss balancing strategy yielded the best results, achieving a Mean Absolute Percentage Error (MAPE) of 3.74% for ROP, 0.36% for BHP, and 1.45% for FP, with corresponding R2 values of 0.985, 0.976, and 0.929, respectively. Notably, the HMTL approach significantly outperformed the standalone Single-Task Learning (STL) models, reducing the prediction error for ROP by approximately 50%, and for BHP and FP by approximately 30% and 28%, respectively. This enhanced accuracy was achieved while also reducing the overall model training time by approximately 28%, underscoring the efficiency and effectiveness gains of joint parameter prediction. This paper introduces a novel HMTL architecture that leverages interdependencies among multiple drilling parameters for enhanced collaborative prediction. It uniquely addresses the challenge of integrating diverse, specialized single-parameter models into a cohesive, efficient predictive system. This offers a transformative framework for holistic, accurate, and computationally efficient intelligent drilling operations, scalable for future parameter additions.