Optimization of Rate of Penetration in a Convoluted Drilling Framework using Ant Colony Optimization

Abstract

Abstract This paper presents a combination of artificial neural networks (ANN) and ant colony optimization (ACO) used to determine optimal rate of penetration (ROP). The Bayesian regularization neural network (BR-NN) was trained and compared to the modified Warren model for ROP. The efficiency and effectiveness of ACO was evaluated for the optimization of ROP. There is no acceptable universal mathematical model that accurately describes ROP because of complex downhole conditions. Analytical models or real-time data analytics are usually used separately to estimate and predict ROP. Combining and using both methods simultaneously enables efficient ROP estimates and predictions. ANN can calculate ROP as output; inputs include depth, weight on bit (WOB or W), rotational speed (N), mud flow rate (Q), and gamma ray. ANN can potentially be used with real-time raw data without an explicit mathematical model for ROP. ACO is an algorithm that mimics ants determining the shortest path between their nest and food source. The ACO algorithm for the travelling salesperson problem (TSP) is modified to optimize ROPs. BR-NN with three hidden layers is selected. The trained BR-NN compares to benchmark data with a correlation coefficient > 0.99. ANN is considered suitable to calculate ROP (or ROP0 without the bit wear term) from drilling parameters, including W, N, Q, depth, and gamma ray. Similar to a typical ACO algorithm for TSP, the pheromone update includes three terms: decay, recent ant activity, and elite path of shortest distance. In the probability function, the ROP reciprocal (i.e., drilling time) replaces the distance. For the test example, ANN-based ACO reproduces the global optimal ROPs, which are obtained by brute force for the test example. ROPs can be overestimated because bit wear is assumed constant for each depth. By separating the bit wear term from ROP, ACO provides an optimal overall drilling time as compared to actual drilling time. The paper presents field examples to predict and estimate optimal drilling parameters. The ROP estimate using the new model was performed using actual drilling parameters. The calculated ROP profile closely matched the actual data with an error rate of less than 5%. The paper proposes specific procedures to estimate drilling parameters and predict ROP and bit life. ANN was used to calculate ROP from input data with high accuracy. This is the first time an ACO algorithm is examined for ROP optimization, and ACO reproduces the optimal ROPs for the test example. This study combines ACO and ANN and shows their great potentials in optimizing real-time ROP based on real-time measurements.