Brian M. Hartman

Storm outage modeling for an electric distribution network in Northeastern USA

Abstract The interaction of severe weather, overhead lines and surrounding trees is the leading cause of outages to electric distribution networks in forested areas. In this paper, we show how utility-specific infrastructure and land cover data, aggregated around overhead lines, can improve outage predictions for Eversource Energy (formerly Connecticut Light and Power), the largest electric utility in Connecticut. Eighty-nine storms from different seasons (cold weather, warm weather, transition months) in the period 2005–2014, representing varying types (thunderstorms, blizzards, nor’easters, hurricanes) and outage severity, were simulated using the Weather Research and Forecasting (WRF) atmospheric model. WRF simulations were joined with utility outage data to calibrate four types of models: a decision tree (DT), random forest (RF), boosted gradient tree (BT) and an ensemble (ENS) decision tree regression that combined predictions from DT, RF and BT. The study shows that the ENS model forced with weather, infrastructure and land cover data was superior to the other models we evaluated, especially in terms of predicting the spatial distribution of outages. This framework could be used for predicting outages to other types of critical infrastructure networks with benefits for emergency-preparedness functions in terms of equipment staging and resource allocation.

Nonparametric Tree‐Based Predictive Modeling of Storm Outages on an Electric Distribution Network

This article compares two nonparametric tree-based models, quantile regression forests (QRF) and Bayesian additive regression trees (BART), for predicting storm outages on an electric distribution network in Connecticut, USA. We evaluated point estimates and prediction intervals of outage predictions for both models using high-resolution weather, infrastructure, and land use data for 89 storm events (including hurricanes, blizzards, and thunderstorms). We found that spatially BART predicted more accurate point estimates than QRF. However, QRF produced better prediction intervals for high spatial resolutions (2-km grid cells and towns), while BART predictions aggregated to coarser resolutions (divisions and service territory) more effectively. We also found that the predictive accuracy was dependent on the season (e.g., tree-leaf condition, storm characteristics), and that the predictions were most accurate for winter storms. Given the merits of each individual model, we suggest that BART and QRF be implemented together to show the complete picture of a storms potential impact on the electric distribution network, which would allow for a utility to make better decisions about allocating prestorm resources.

Model Selection and Averaging in Financial Risk Management

Simulated asset returns are used in many areas of actuarial science. For example, life insurers use them to price annuities, life insurance, and investment guarantees. The quality of those simulations has come under increased scrutiny during the current financial crisis. When simulating the asset price process, properly choosing which model or models to use, and accounting for the uncertainty in that choice, is essential. We investigate how best to choose a model from a flexible set of models. In our regime-switching models the individual regimes are not constrained to be from the same distributional family. Even with larger sample sizes, the standard model-selection methods (AIC, BIC, and DIC) incorrectly identify the models far too often. Rather than trying to identify the best model and limiting the simulation to a single distribution, we show that the simulations can be made more realistic by explicitly modeling the uncertainty in the model-selection process. Specifically, we consider a parallel model-selection method that provides the posterior probabilities of each model being the best, enabling model averaging and providing deeper insights into the relationships between the models. The value of the method is demonstrated through a simulation study, and the method is then applied to total return data from the S&P 500.

Credibility in Loss Reserving

This article proposes using credibility theory in the context of stochastic claims reserving. We consider the situation where an insurer has access to the claims experience of its peer competitors and has the potential to improve prediction of outstanding liabilities by incorporating information from other insurers. Based on the framework of Bayesian linear models, we show that the development factor in the classical chain-ladder setting has a credibility expression: a weighted average of the prior mean and the best estimate from the data. In the empirical analysis, we examine loss triangles for the line of commercial auto insurance from a portfolio of insurers in the United States. We employ hierarchical model for the specification of prior and show that prediction could be improved through borrowing strength among insurers based on a hold-out sample validation.

Estimated Time of Restoration (ETR) Guidance for Electric Distribution Networks

Abstract Electric distribution utilities have an obligation to inform the public and government regulators about when they expect to complete service restoration after a major storm. In this study, we explore methods for calculating the estimated time of restoration (ETR) from weather impacts, defined as the time it will take for 99.5% of customers to be restored. Actual data from Storm Irene (2011), the October Nor’easter (2011) and Hurricane Sandy (2012) within the Eversource Energy-Connecticut service territory were used to calibrate and test the methods; data used included predicted outages, the peak number of customers affected, a ratio of how many outages a restoration crew can repair per day, and the count of crews working per day. Data known before a storm strikes (such as predicted outages and available crews) can be used to calculate ETR and support pre-storm allocation of crews and resources, while data available immediately after the storm passes (such as customers affected) can be used as motivation for securing or releasing crews to complete the restoration in a timely manner. Used together, the methods presented in this paper will help utilities provide a reasonable, data-driven ETR without relying solely on qualitative past experiences or instinct.

Model Selection and Averaging of Health Costs in Episode Treatment Groups

Episode Treatment Groups (ETGs) classify related services into medically relevant and distinct units describing an episode of care. Proper model selection for those ETG based costs is essential to adequately price and manage health insurance risks. The optimal loss model (or model probabilities) can vary depending on the disease. We compare four potential models (lognormal, gamma, log-skew-t, and Lomax) using four different metrics (AIC and BIC weights, Random Forest feature classification, and Bayesian model averaging) on 320 episode treatment groups. Using the data from a major health insurer, which consists of more than 33 million observations from 9 million claimants, we compare the various methods on both speed and precision, and also examine the wide range of selected models for the different ETGs. Several case studies are provided for illustration. It is found that Random Forest feature selection is computationally efficient and sufficiently accurate, hence being preferred in this large data set. When feasible (on smaller data sets), Bayesian model averaging is preferred because of the posterior model probabilities.

A Case Study on Power Outage Impacts from Future Hurricane Sandy Scenarios

AbstractHurricane Sandy (2012, referred to as Current Sandy) was among the most devastating storms to impact Connecticut’s overhead electric distribution network, resulting in over 15 000 outage locations that affected more than 500 000 customers. In this paper, the severity of tree-caused outages in Connecticut is estimated under future-climate Hurricane Sandy simulations, each exhibiting strengthened winds and heavier rain accumulation over the study area from large-scale thermodynamic changes in the atmosphere and track changes in the year ~2100 (referred to as Future Sandy). Three machine-learning models used five weather simulations and the ensemble mean of Current and Future Sandy, along with land-use and overhead utility infrastructure data, to predict the severity and spatial distribution of outages across the Eversource Energy service territory in Connecticut. To assess the influence of increased precipitation from Future Sandy, two approaches were compared: an outage model fit with a full set of...

Using a Bayesian Regression Approach on Dual-Model Windstorm Simulations to Improve Wind Speed Prediction

AbstractWeather prediction accuracy is very important given the devastating effects of extreme-weather events in recent years. Numerical weather prediction systems are used to build strategies to prevent catastrophic losses of human lives and the environment and have evolved with the use of multimodel or single-model ensembles and data-assimilation techniques in an attempt to improve the forecast skill. These techniques require increased computational power (thousands of CPUs) because of the number of model simulations and ingestion of observational data from a wide variety of sources. In this study, the combination of predictions from two state-of-the-science atmospheric models [WRF and RAMS/Integrated Community Limited Area Modeling System (ICLAMS)] using Bayesian and simple linear regression techniques is examined, and wind speed prediction for the northeastern United States is improved using regression techniques. Retrospective simulations of 17 storms that affected the northeastern United States duri...

Predictive Modeling in Long-Term Care Insurance

The accurate prediction of long-term care insurance (LTCI) mortality, lapse, and claim rates is essential when making informed pricing and risk management decisions. Unfortunately, academic literature on the subject is sparse and industry practice is limited by software and time constraints. In this article, we review current LTCI industry modeling methodology, which is typically Poisson regression with covariate banding/modification and stepwise variable selection. We test the claim that covariate banding improves predictive accuracy, examine the potential downfalls of stepwise selection, and contend that the assumptions required for Poisson regression are not appropriate for LTCI data. We propose several alternative models specifically tailored toward count responses with an excess of zeros and overdispersion. Using data from a large LTCI provider, we evaluate the predictive capacity of random forests and generalized linear and additive models with zero-inflated Poisson, negative binomial, and Tweedie errors. These alternatives are compared to previously developed Poisson regression models. Our study confirms that variable modification is unnecessary at best and automatic stepwise model selection is dangerous. After demonstrating severe overprediction of LTCI mortality and lapse rates under the Poisson assumption, we show that a Tweedie GLM enables much more accurate predictions. Our Tweedie regression models improve average predictive accuracy (measured by several prediction error statistics) over Poisson regression models by as much as four times for mortality rates and 17 times for lapse rates.

Investigating international new product diffusion speed: A semiparametric approach

Supported in part by National Science foundation CMG research Grants DMS-07-24704, DMS-09-14951 and by Award Number KUS-CI-016-04 made by King Abdullah University of Science and Technology (KAUST).Supported in part through the Deans Faculty Research Fellowship award from the School of Management, State University of New York at Buffalo.