A Simulated Annealing Accelerated Bayesian Optimization (SAABO) for Real-Time Deep Neural Network Model with Range Constraints

Abstract

optimization issues in many fields involve finding optimal solutions over surfaces that are highly non-linear, discontinuous and lumpy. They also necessitate integration of domain-specific reduced form physics-based engineering models with data-driven optimization techniques. Real-time drilling modeling and optimization is used as a prototype for demonstrating the new algorithm but can be applied to several real-time systems. This paper provides a new real-time model with deep neural network (DNN) automated closed loop control drilling optimization using a new simulated annealing accelerated Bayesian optimization (SAABO) algorithm. This approach offers an improved and efficient methodology to arrive at the maximum for optimization surfaces that are highly non-linear and lumpy (i.e., might contain multiple local minima). This method was applied to predict rate of penetration (ROP) with automatic optimum parameter tuning of weight on bit (WOB) and rotations per minute (RPM) for drilling optimization along with range constraints. These range constraints are obtained from the engineering models produced from domain insight. Thus, the optimization integrates reduced form physics-based engineering models into its framework. As such, the new real-time ROP DNN model used for optimization is a hybrid engineering and data-driven model. The hybrid DNN model uses real-time data and the generated data from the engineering model. The generated data are needed to fill the void space in the surface not covered by the real-time measured data. The stochastic optimization algorithm SAABO is fast, as the exploration points needed for Bayesian optimization (BO) are obtained from the simulated annealing (SA)-based algorithm. SAABO uses the distribution of exploration points from SA to speed up the convergence and improve the accuracy of the BO algorithm to optimize the ROP with constraints. The methodology comprising real-time ROP model prediction and stochastic optimization is fast and efficient for the computation of the necessary optimum controllable parameters for drilling. The new hybrid modeling approach and the SAABO algorithm developed in this paper can be applied to any real-time modeling system with optimization and closed loop control.