Estimating emergency vehicle travel time

Sammanfattning

A large part of the total response time in case of an emergency is the travel time between the location where the emergency vehicle is stationed and the emergency location. Having a reliable estimate of travel time will give good conditions for the planning of emergency calls, which in turn can help to reduce total response time and reduce operating costs. This thesis investigates what factors influence emergency vehicle travel time using historical vehicle data and publicly available information such as road characteristics and weather. The vehicle position data available for this thesis span across all regions of Sweden. The positions were made into routes that were map matched to the existing road network from the national road database. An automatic route validation algorithm was implemented to remove poorly map matched routes. Routes were given a variety of properties describing them, for example, time of day, average speed limit, and precipitation. With the routes and their properties, a variety of speed estimation models were developed using methods such as linear regression and decision trees with different variable selections, which were then used for travel time estimation. The models were evaluated with mean absolute percentage error (MAPE) and root mean squared error (RMSE). A baseline model consisting only of the average speed limit was used for reference, the performance was 21.82% MAPE and 202.07 seconds RMSE with a decision tree. The thesis concluded that some of the most important variables for emergency vehicle travel time estimation were total distance, average speed limit, and response status. One of the best-performing models included a decision tree with variables that included total distance, average speed limit, hour of day, response status, visibility, and more, which had MAPE 16.21% and RMSE 181.81 seconds. A regression-based model with only four input variables in use with interaction terms and variable transformations also showed a relatively high prediction accuracy of 17.40% MAPE and 187.65 seconds RMSE.