The Section on Statistics in Defense and National Security sponsored the student poster awards at the Conference on Data Analysis (CODA), which was held March 2–4 in Santa Fe, New Mexico. CODA highlights data-driven problems of interest to the U.S. Department of Energy. This year, there were attendees from 10 national laboratories, 20 universities, and a variety of companies.
Poster Award Winners
First prize ($400) was awarded to Thomas Catanach from Caltech for “Power System Dynamic Estimation.”
Abstract: Because power systems are becoming increasingly complex and subject to disturbances, developing methods for state estimation and system identification is essential for increasing the reliability of the power grid. Currently, this problem is solved on slower, steady-state, time scales; however, faster estimation is now possible with the deployment of phasor measurement units (PMUs). Many methods have been studied for dynamic state estimation, including both local and global filtering methods. One of the main challenges for global methods is how the filter integrates the differential algebraic equations (DAE) that describe the power system. This work applies implicit methods used to solve DAEs to improve the performance and robustness of an extended Kalman filter, making it an attractive state estimation choice. Further, we introduce techniques to reduce the effect of temporary disturbances on the state estimated to help track the state through these faults where the network model is no longer accurate by creating a layered estimation architecture. This architecture integrates state estimation, change point detection, and classification of disturbances.
Second prize ($100) was awarded to Nicholas Michaud from Iowa State for “A Bayesian Hierarchical Model for Estimating Influenza Severity.”
Abstract: Timely monitoring and prediction of the trajectory of seasonal influenza epidemics allows hospitals and medical centers to prepare for and provide better service to patients with influenza. The U.S. Centers for Disease Control and Prevention’s ILINet system collects data on influenza-like illnesses from more than 3,300 health care providers and uses this data to produce accurate indicators of current influenza epidemic severity. However, ILINet indicators are typically reported at a lag of 1–2 weeks. Another source of severity data, Google Flu Trends (GFT), is calculated by aggregating Google searches for certain influenza-related terms. GFT data is provided in near-real time, but is a less direct measurement of severity than ILINet indicators and is likely to suffer from bias. We create a hierarchical model to estimate epidemic severity for the 2014–2015 epidemic season, which incorporates current and historical data from both ILINet and GFT, allowing our model to benefit from both the timeliness of GFT data and the accuracy of ILINet data. To forecast for the 2014–2015 influenza epidemic season, we provide our model with both ILINet and GFT data from previous seasons, starting with the 2004–2005 epidemic season and going through the 2013–2104 epidemic season. The hierarchical structure of our model allows ILINet and GFT data from previous seasons to inform epidemic severity prediction in the current season. ILINet data is modeled as being an unbiased but noisy estimate of the true, unknown influenza severity. GFT severity measurements, on the other hand, are influenced by external factors such as media coverage. These factors could consistently bias GFT severity estimates to over- or under-estimate the true epidemic severity, depending on the intensity of media influenza coverage in a season. To account for this potential bias in GFT data, we include a temporally correlated error term that allows over- or under-predictions made by GFT data in one week to carry over into the next. Estimation is performed using the Bayesian statistical software JAGS. We examine the increase in forecast accuracy that GFT data provides by comparing the forecasting ability of our model using both GFT and ILINet data to that of a model given only ILINet data. The two models are evaluated for their ability to predict epidemic severity multiple weeks into the future, and we find that combining up-to-date GFT data with accurate ILINet data improves epidemic severity forecasting ability significantly.
Honorable mentions went to Michael Grosskopf from Simon Fraser University and Ben Newton from The University of North Carolina at Chapel Hill.
Leave a Reply