Use of Neural Networks in Transport Models



Use of Neural Networks in Transport Models

Authors

Anan Allos, Atkins, Sam Brooks, Atkins, Dominic Dachs, Atkins

Description

This project undertakes the calibration of trip generation and distribution models to provide the best parity between synthetic and observed inter-zonal trips

Abstract

Use of Neural Networks in Transport Models
Anan Allos, Sam Brooks and Dom Dachs, Atkins

The traditional method of developing a typical four stage transport models is to have a trip generation model, with mode choice and trip distribution in the demand model, and each has to be calibrated separately. Trip generation is a function of land use and socio-economic variables, and the distribution model is typically some sort of a gravity model - a function of trip generation and inter-zonal deterrence.

An alternative is proposed which has been partially tested with machine learning using neural network techniques. It effectively simultaneously undertakes the calibration of the trip generation and distribution models to provide the best parity between synthetic and observed inter-zonal trips.

This novel technique was going to be applied to car person trips from the RSI data collected for a national multi-mode transport model being developed by Atkins for Turkey. Conventional trip generation models were built using regression for production and attraction, as a function of zonal variables. A gravity model was calibrated using a Tanner function which gave a good fit given the usual criteria. But the Turkey data was too frugal to train the model, so the South East Regional Model was used for network skims alongside Journey to Work Census data for travel patterns.

A neural network model was built and the inputs were:
• Numerous zonal attributes such as population, cars owned, employment by economy sector, number of students, average income;
• The partial trip matrix obtained from the combined RSI surveys; and
• An inter-zonal deterrence representing generalised time.
The neural network model was then ‘trained’ so that the synthetic inter-zonal trips simulate as much as possible those in the partial observed trip matrix. Several alternate models were developed with varying numbers of hidden layers, nodes per layer, and different loss functions. The use of training, validation and test sets enabled the model to be verified on data it had never seen before and avoid overfitting, a common problem with neural networks.

Initial results are encouraging in that there is excellent compatibility between the synthetic and observed trip length distribution of the data. Some stress testing has been undertaken to estimate the impact of positive and negative changes in the zonal variables and inter-zonal deterrence on the number of trips, and many of the preliminary results are encouraging and rational.

Further training of the neural network model is required possibly adding public transport skim costs, using the model in back-casting and to another area. .
One shortcoming of the technique is that the underlying model formulation is not known and the model sensitivity and response has to be carefully sense checked before it can be used as a forecasting tool. However, if the technique is successful, it can be a powerful and innovative way forward which can be expanded to cover for example mode choice as well as generation and distribution.

Publisher

Association for European Transport