Research

Greedy Copula Segmentation of Multivariate Non-Stationary Time Series for Climate Change Adaptation 

This study targets judicious near-future modeling of non-stationary climate processes using past observations optimally.  The methodology proposed seeks to divide up observed data into non-overlapping segments, each of which are separately treated as stationary but with underlying probability and dependence structure, while the long time series collectively yield multiple such segments that are mutually independent; the segments are not known a priori but need to be statistically established for this analysis.  A Greedy Copula Segmentation (GCS) algorithm is developed that employs best-fit multivariate probability distributions and copula functions after the data-driven time series segmentation is undertaken.  Predictions based on the GCS approach are closer to the actual future than those made by a traditional model using all the available data.

Optimizing Operations and Maintenance for Offshore Multi-Purpose Platforms in Variable Weather Conditions

A concept of offshore floating multi-purpose platforms (MPPs) has emerged that offer benefits from shared use of infrastructure assets for multiple services including resource extraction activities such as renewable energy generation, aquaculture, leisure, and transport functions. This study is the formulation of a Markov decision process (MDP) to provide an optimized policy that guides the scheduling of operation and maintenance (O&M) activities of MPPs.  Satisfactory performance of multiple functions with MPPs requires dealing with a wide-ranging set of O&M activities that can take different amounts of time and require different levels of calmness in weather conditions.  The right decision on the start or delay/postponement of an issued O&M activity is key to resolving the O&M problem quickly, without interruption and accident, while carrying out work tasks in changeable weather.  Historically, such decisions have been made by lead operators, based largely on experience.  The formulated MDP involves a stochastic weather window analysis that operators can employ to decide upon the scheduling of work activities.  By following the provided policy, the overall loss of revenue and costs of O&M are inherently minimized.  The robustness of the method has been validated by demonstrating that the optimized policy produces lower accumulated costs than naïve policies for a wide range of general metocean conditions.

Sustainable Reuse of Decommissioned Jacket Platforms for Offshore Wind Energy Accounting for Accumulated Fatigue Damage

This study aims to maximize the benefits of sustainable reuse of oil and gas platform for wind energy generation by establishing an optimized plan that accounts for the remaining life of the repurposed platform, the overall construction and platform retrofit costs, and an expectation of a period of clean energy generation after the wind turbine installation.  The proposed framework incorporates metocean data, spectral fatigue damage assessment, reliability-based life cycle estimation, and economic revenue evaluation.  For the choice of wind turbine, we consider various options with different associated power ratings and dimensions.  An optimization problem is defined that considers large turbines with higher output ratings but possibly shorter expected service lives as well as contrasting alternatives.  A realistic case study and sustainable reuse scenario for a site in Porto, Portugal, are employed to illustrate the advantages of the model developed.

SOS WindEnergy 

Learning Generative Embeddings using an Optimal Subsampling Policy for Tensor Sketching

Data tensors of orders 3 and greater are routinely being generated. These data collections are increasingly huge and growing. For instance, a North American Regional Reanalysis (NARR) has been collecting 70 climate variables every 3 hours from 1979 to the present, and it is currently at a total size of 29.4 Terabytes. Directly accessing such large data tensor collections for information has become increasingly prohibitive. We learn approximate full-rank and compact tensor sketches with decompositive representations providing compact space, time and spectral embeddings of the original tensor. All subsequent information querying with high accuracy is performed on the generative sketches. We produce optimal rank-r Tucker decompositions of arbitrary order data tensors by building tensor sketches from a sample-efficient sub-sampling of tensor slices. Our sample efficient policy is learned via an adaptable stochastic Thompson sampling using Dirichlet distributions with conjugate priors.