2016 INFORMS Annual Meeting Program

TA05

INFORMS Nashville – 2016

3 - Reliable Renewable Generation And Transmission Expansion Planning: Co-optimizing System’s Resources For Meeting Renewable Targets David Pozo, Pontificia Universidad Católica de Chile, Santiago, Chile, davidpozocamara@gmail.com, Alexandre Moreira, Alexandre Street, Enzo E Sauma We propose a two-stage renewable generation and transmission expansion planning model that jointly finds the best subset of new transmission assets and renewable sites to be developed. The main goal of this co-optimization planning model is to address renewable targets while accounting for the least-cost reserve scheduling to ensure reserve deliverability under generation and transmission outages and renewable variability. A case study with realistic data from the Chilean system is presented and solutions obtained with different level of security are tested against a set of 10,000 simulated scenarios of renewable injections and system component outages. TA05 101E-MCC Short-term Operation, Maintenance, and Long-term Planning for Power Systems Sponsored: Energy, Natural Res & the Environment, Energy I Electricity Sponsored Session Chair: Murat Yildirim, Georgia Institute of Technology, 755 Ferst Dr, Atlanta, GA, 30312, United States, murat.v.yildirim@gmail.com 1 - Topics On Optimal Power Flow Richard O’Neill, FERC, Richard.O’Neill@ferc.gov 2 - Load-dependent Sensor-driven Maintenance And Operations In Power Systems Murat Yildirim, Georgia Institute of Technology, Atlanta, GA, 30332, United States, murat@gatech.edu, Andy Sun, Nagi Gebraeel The operational loads on the generating units have a significant impact on how fast they degrade. For instance, the frequency of start-up and shut-down cycles can change the lifetime of combined-cycle power plants by an order of magnitude. In this talk, we use in-situ sensor based signals to provide i) an accurate load-dependent degradation model for generating units, and ii) a flexible framework whereby the scheduler gains some control on how fast the generating units are degrading. The proposed framework achieves significant improvements in cost and reliability. 3 - Impact Of Short-term Variability And Uncertainty On Long-term Planning Problems Henrik Bylling, University of Copenhagen, Universitetsparken 5, Copenhagen, DK-2100, Denmark, bylling@math.ku.dk, Salvador Pineda, Trine Krogh Boomsma Considering a detailed representation of short-term system operations turns long- term planning problems, such as generation expansion, computationally intractable. Simplified models reduce the computational burden by focusing on a particular aspect of the short-term operation. We compare existing simplified models in terms of i) their ability to capture the impact of both short-term variability and short-term uncertainty on long-term planning decisions and ii) their computational complexity. We also propose a new procedure that outperforms existing ones in these two aspects.

2 - Online Feature Importance Ranking Based On Sensitivity Analysis Alaleh Razmjoo, University of Central Florida, Orlando, FL, 32765, United States, alaleh.razmjoo@Knights.ucf.edu, Petros Xanthopoulos In this paper, we present a fast and efficient incremental online feature ranking and feature selection. We employ the concept of global sensitivity and rank features based on their impact on the outcome of classification model. In the feature selection part, we use a two stage filtering method to first eliminate highly correlated and redundant features and then eliminating irrelevant features in the second stage. It can be implemented along with any online classification method. The proposed method is primarily developed for online tasks, however, significant experimental results in comparison with popular feature selection methods suggest that it can be also used in batch learning tasks. 3 - A Novel Weighting Policy For Unsupervised Ensemble Learning Based On Mean-variance Portfolio Optimization Method. Ramazan Unlu, UCF, ramazanunlu@gmail.com Unsupervised ensemble learning is an optimal combination strategy of individual clustering methods to create a model that fits to data better. Determining proper weights for clustering methods is a crucial step to build a well-combined partition. Recently, an approach was proposed based on concept of internal validity measures that has profound advantages over traditional ensemble learning. Despite its robust properties this approach consider only index values itself, but not variation of them. In this paper, we propose a better weighting policy for this problem that is based on mean-variance portfolio optimization method and compare against other popular approaches. 4 - Nonlinear Dimensionality Reduction For Analysis Ofelectroencephalography Records Anton Kocheturov, University of Florida, Gainesville, FL, United States, antrubler@gmail.com We suggest using nonlinear dimensionality reduction technique called the Local Linear Embedding for analysis of EEG records. This approach enabled us to distinguish between different states of the brain in a more efficient way comparing to the existing machine learning techniques since it is faster and doesn’t require training of the algorithm. We also detected evidence for local linearity of the brain in the resting state and introduced a new model of the brain based on it.

TA07 102B-MCC Retail Analytics Sponsored: Data Mining Sponsored Session

Chair: Matthew Lanham, Virginia Tech, Pamplin 1007, Blacksburg, VA, 24061, United States, lanham@vt.edu 1 - Analytics on The Edge Of Retail Aaron Burciaga, Accenture, adburciaga@gmail.com

The fervor of big data and business analytics have led to a bumper crop of education, training, tools, and methods. It’s has become increasingly difficult to detect the signal from the noise of those same people, processes, and tools that purportedly exist to distinguish signals from noise. This presentation will review several case studies of how commercial and national government programs are developing (or stumbling) in their analytics programs. Emergent technologies and methods, including the application of Machine Learning and Artificial Intelligence on edge devices will be presented, showing how the last mile and last dollar can be closed in both new and traditional challenges. 2 - An Investigation Of Cluster Analysis Of Retail Stores To Improve Predictive Modeling Of Sales Linda Schumacher, Merchandise Scientist, Raleigh, NC, 27604, United States, schumachers@bellsouth.net Data mining clustering algorithms are used to identify similar groups of retail stores for segmenting data to improve predictive modeling results. Clustering methods including centroid-based, hierarchical, two-step and probabilistic clustering are considered. The performance of these clustering methods is evaluated and compared with calculated metrics. Using data from a national retailer, the impact of segmenting the data to improve overall predictive performance is reported. 3 - Investigating Sparse Demand Models To Support The Assortment Planning Decision Matthew Lanham, Clinical Assistant Professor, Purdue University, West Lafayette, IN, United States, malanham@gmail.com We present research examining the performance of substitution-based multi- classification models currently being researched and employed in practice by major retailers, versus more naïve binary classification models to understand purchase propensity. We discuss how these models would yield different assortments for sparse demand products.

TA06 102A-MCC Optimization Models in Data Mining Sponsored: Data Mining Sponsored Session

Chair: Petros Xanthopoulos, University of Central Florida, 4000 Central Florida Blvd., P.O. BOX 162993, Orlando, FL, United States, petrosx@ucf.edu 1 - Relaxing Support Vector Machines

Orestis P. Panagopoulos, California State University, Stanislaus, Turlock, CA, United States, orepana@gmail.com, Talayeh Razzaghi, Petros Xanthopoulos, Onur Seref In this paper, we extend Relaxed Support Vector Machines (RSVM) to perform regression as well as one-class classification tasks. Our models, Relaxed Support Vector Regression (RSVR) and One-Class Relaxed Support Vector Machines (ORSVM) are formulated using both linear and quadratic loss functions and are solved with sequential minimal optimization. Their performance is measured on several publicly available datasets and are compared to other state-of-the-art regression and classification methods.

234