Short-term congestion prediction: comparing time series with neural networks

Neural network applications in transportation have been the subject of research for the past two decades. Often they produce nice results, but also very often these results are not compared with results of other methods. In this study we try to make an effort to compare performances of an ARMA time series analysis method with those of a multilayer feedforward (MLF) neural network method. It turns out that the MLF neural network method gives a better performance in both the congestion prediction and the processor time efficiency. A few remarks must be made, however. It is very likely that both methods will give a worse performance in the case of an incident. The time series analysis method then probably will outperform the neural network method because its parameters are set on a single point in space whereas the weights of the neural network are trained on recognising spatial patterns. The neural network simply will not recognise these events, because they were never learned. If one thus uses neural network methods as described above, they can be operate under recurrent congestion conditions and it would be recommended to use incident detection methods at the same time. Another remark concerning this research is that this is just the start of a bigger picture. Other method´s performances should be estimated with the same data set, e.g. fuzzy logic, recurrent neural networks, and/or genetic algorithms to name a few. Then these methods should also be used on different data sets before one can generalise conclusions. In short: further research is needed