Multivariate Time-Series Deep Learning for Short-Term Forecasting of Lost Circulation in Drilling Operations

Abstract

Lost circulation is a frequent and costly problem in drilling operations, often leading to non-productive time, operational delays, and increased risk. Accurate prediction of lost circulation is challenging due to the complex, time-dependent interactions among drilling parameters, formation conditions, and operational states. This thesis investigates the use of multivariate time-series deep learning models for short-term lost circulation prediction based on real field drilling data. The dataset used in this study was collected during a DARPA sponsored field drilling project and includes controllable operational parameters, measured drilling responses, and environmental variables. A comprehensive data preprocessing work flow is developed to address sensor noise, missing data, inconsistent sampling rates, and non-drilling intervals, resulting in a structured and physically consistent dataset suitable for time-series analysis. The prediction task is formulated as a supervised multivariate time series forecasting problem. Multiple baseline models and advanced deep learning models are evaluated under consistent experimental settings. The results show that the Chronos-2 foundation model achieves the best overall performance, outperforming traditional statistical models and earlier deep learning approaches in terms of prediction accuracy and explanatory power. Notably, Chronos-2 demonstrates strong zero-shot capability and further performance improvements after fine-tuning, enabling reliable short-term prediction of lost circulation trends. These results indicate that modern time-series foundation models can effectively learn the temporal dynamics of drilling operations and provide accurate short-term predictions of lost circulation from field data.