Road surface condition forecasting from historical data and weather forecasts

Rok Kršmanc (2013) Road surface condition forecasting from historical data and weather forecasts. PhD thesis.

Preview

PDF Download (2287Kb)

Abstract

The dissertation presents a new method for construction of hierarchical regression models by combining the predetermined groups of data based on the knowledge of compatibility and/or similarity of models. Regression models are widely used in many areas. The simplest, linear regression models are often chosen because of their robustness. In this work we consider a case in which the data can be split into multiple subsets on which the predicted outcome is linear. While such data cannot be modelled with a single linear model, inducing a separate linear model for each subset would yield inaccurate models due to small samples. This problem can be nicely framed in the theory of bias-variance decomposition of error. A single model would have an unacceptably high bias, while separate models would have high variances. Based on this theoretical justification, we developed a method for hierarchical construction of linear models, which starts with separate models for each subgroup and then iteratively merges them until the effect of decreased variance due to larger data samples available for each model begins to be overcome by the increased bias due to treating a non-linear relation as linear. We tested the method on controlled synthetic data, which proved the correctness of our approach. The method was then tested on the data on road meteorology: we were able to successfully predict the road surface temperature for several hours ahead. This result is interesting for the field of road meteorology as it shows that it is possible to construct models with good forecasting accuracy with statistical methods alone. The advantage of such modelling compared to physical models based on energy balance equation is that they do not require any knowledge about the road construction properties. Their weakness is that they require past data from road weather station at a particular location. To overcome this problem, we investigated a problem in which the data for several locations is available and the task is to find a predictive model for a new location based on the known physical, but easily obtainable similarities between this and other locations, such as sky visibility and similar. We first checked that the attributes that we chose to describe locations are indeed correlated with coefficients from regression models. Based on positive findings of this study, we defined a modelling technique that can construct a linear model for the location based on linear models for other locations. We again first empirically tested the method on synthetic data constructed in such a way that it fulfills the assumptions of the method, and then on the actual data from the road weather stations. As expected, the accuracy of such models is below those constructed from the actual data, yet still quite in the acceptable range for its potential practical use. Results of the dissertation may find practical use. Forecasts of road weather conditions are a valuable resource for drivers, as well as for road maintenance services, in particular in winter. More accurate forecasts can provide safer roads while cutting down the maintenance costs and minimizing the environmental damage from over-salting.

Actions (login required)