RISK COMMUNICATIONS: AROUND THE WORLD Neural Network Models for Assessing Road Suitability for Dangerous Goods Transport
J. Taboada,1 J.M. Mat´as,2 A. Saavedra,2 C. Ordo˜ ez,1 and R. Mart´nez-Alegr´a3 ı ´n ı ı Department ofEnvironmental Engineering, University of Vigo, Vigo, Spain; 2 Department of Statistics & Operational Research, University of Vigo, Vigo, Spain; 3 Civil Protection Service, Regional Government of Castilla & Leon, Spain ´ ABSTRACT This article describes a methodology for assessing the degree of remedial action required to make short stretches of a roadway suitable for dangerous goods transport (DGT).The methodology is based on the evaluation of a set of variables that have a bearing on DGT risk. The large number of variables involved made it necessary to apply a supervised approach based on expert criteria. The result was a knowledge base that can be used both to estimate DGT risk for new stretches of roadway and to determine sources of risk without having to rely on an expert. A number ofmultivariate statistical analysis techniques were tested for the construction of the model, namely linear discriminant analysis with a prior reduction in dimensionality, multilayer perceptrons, and support vector machines. The results obtained from a test sample show that the support vector machines represented expert knowledge most reliably. A graphic representation of the risk index for a studiedstretch of roadway results in a map of the level of DGT risk for that roadway. Key Words: transportation, dangerous goods, multivariate statistics, neural networks, SVM.
INTRODUCTION The continuing increase in the volume of hazardous materials (hazmats) transported by road implies a growing risk for populations and the environment. One of the first steps toward identifying and quantifyingthis risk is to develop and perfect risk analysis methodologies that can subsequently be used to minimize risk, whether by reducing the probability of the occurrence of accidents or by reducing the level of exposure of the more vulnerable environmental and populational elements.
Received 19 September 2004; revised manuscript accepted 20 January 2005. Address correspondence to Celestino Ord´ nezGal´ n, Universidad de Vigo, Escuela T´ cnica o˜ a e Superior de Ingenieros de Minas de Vigo, Lagoas-Marcosende 9, 36200 Vigo, Spain. E-mail: firstname.lastname@example.org 174
Assessing Road Suitability for Dangerous Goods Transport
Mart´nez-Alegr´a et al. (2003) recently proposed a conceptual model for analyzing ı ı the risks associated with DGT that identifies the highest-risk roads in a network.This method was based on statistical accident rates, total traffic density, specific traffic density for vehicles transporting hazmats, and vulnerability ratings for environmental and populational elements in relation to different types of hazmats. The formula for this model takes the following form:
R = P ∗G (1)
where R is risk, P is the probability of an accident, and G is the gravity of thepossible damage. To estimate G, the following expression is used: G = Pe (Hhg + Vp + Va ) where Vp represents a series of populational vulnerability factors, Va represents a series of environmental vulnerability factors, Hhg is the potential damage inherent in a product (which depends on properties such as flammability, reactivity, toxicity, corrosion, and oxidation), and where Pe is the hazardimplied in an accident, which depends on considerations such as whether or not the accident involves leakage of the transported substance, and whether or not an explosion occurs. Publicly available historical data on general traffic density, specific traffic density for vehicles transporting hazmats, kinds of accident, and the proportion of traffic represented by each type of hazmat were used to...