Short-Term Forecasts for Traffic Models

• Forschungsstelle: ASTRA SBT
• Projektnummer: ASTRA2006/019
• Projekttitel Englisch: Short-Term Forecasts for Traffic Models

Schlüsselwörter (Englisch)

Short Term Forecasts, Transport Model, Error Propagation

Kurzbeschreibung (Englisch)

How will the Federal Office of Roads be sure to use the best possible short-term forecasts in the new traffic management centre (VM-CH) by Jan 1, 2008?

This is the main question we intend to answer with this project. A short survey among SVI members should give us a picture of methods used in Switzerland. We will concentrate on 3+ methods, test their ability to predict traffic data and compare them. The positive findings of SVI 2004/015 “Error Propagation in Transport Models” will be used to monitor the quality of the forecasts through error propagation. The capacity of the different methods to “learn” through feedback will be analyzed as well.

All this would not be possible without a robust software. The program ERR_PROP from SVI 2004/015 will be extended to short term forecasts and renamed PRE_VISI. PRE_VISI will be used to analyze and test the individual methods and to compare them in parallel with exactly the same input data. This is the only way to guarantee the comparability.

Same as for ERR_PROP in SVI 2004/015, PRE_VISI is only a research tool and neither suited nor intended to be used commercially.

Projektbeschreibung (Englisch)

please see the detailed description in attachment 1

Methoden (Englisch)

The methodology can be resumed to 5 aspects:

1 Survey about short term forecasting in Switzerland

To get a view of the use of short term forecasts in Switzerland, we intend to start the study with a small survey among all members of the SVI (more details about the questions in attachment 1). We expect no more than 5-10 answers for whole Switzerland.

2 How is knowledge transferred from past to future?

The great advantage of short term forecasts is that values forecasted at t0 for the time t0 + t can be compared with real values at the latest at t0 + t !

How is information extracted from past data (daily variation curves)? How is this information stored? What does the system “know”?

How is the knowledge about the past applied to the future? What quality can be guaranteed at t0 for time t0 + t? How is the past forecast compared to the future when future has arrived? How is the quality measured at time t0 + t ?

Quality control will play a major role in this study. It will be used to measure how well a method performs in forecasting the near future, to optimize the parameter settings of each method and to compare the different methods among each other. Is it able to “learn” from past experiences, and how? Is it able to react to changes, and how?

3 Which methods will be analysed and compared?

Three methods will certainly belong to the analysis: Cluster analysis, regression analysis and Fourier series. If the survey or the literature search reveals that other methods are useful, they will be included too. However, the research will not be extended to methods a) existing only in theory, b) needing highly complex calculations like neuronal networks or c) with no reasonable chance to be of any utility for traffic applications.

Cluster analysis has been user for years by C. de Rham. It was also the subject of his PhD thesis. The method works well but is very probably implemented sub optimally. This research will help find the best options among metrics, distance functions, parameter settings, number of clusters, etc.

Regression analysis is also widely used for short term forecasting. Here too, it would be of great value to have at least some recommendations about parameters, acceptable R2 levels, number of time intervals etc. to assure a near optimal implementation.

Fourier series are new to traffic engineers and no applications are known to the research team. C. de Rham is electrical engineer and well off with Fourier techniques. The idea behind Fourier series is that every continuous curve can be expressed as a sum of sine waves of increasing frequency and varying amplitude. The whole information content of the curve is given by the “series” of frequencies and amplitudes of single sine waves which, once added, reproduce the original curve.

This has many advantages for the analysis of time series analysis. Low frequencies belong to real trends. The higher the frequencies the higher the probability than it is just noise. It is possible that forecasts could be estimated with the frequencies instead of the curve itself and the question to be answered is: Are forecasts in the frequency domain more accurate than in the time domain? A “yes” would be a break-through in favor of the method.

The comparison of the 3+ methods will be central to this study. Questions to be answered are: What is the optimal setting of parameters, etc. for each method? Does one method consistently produce better forecasts? If not, what can be recommended to users?

4 Error Propagation and Short-Term Forecasts

The research project SVI 2004/015 „Error Propagation in Transport Models“ produced interesting results for the assignment and calibration of online traffic situations.

Another finding is that error propagation can directly be used as tool for quality management. The idea is simple: if the quality (= errors) of the input is known and if the software is able to propagate this quality (= errors) throughout the whole calculation process, the output quality (= errors) will be known too. This is exactly what has been done for the actual situation.

There is no reason not to apply this idea to short term forecasts as well. We would like to show that by integrating error propagation throughout short term forecasting calculation, one can directly get the output data needed for quality control. This hypothesis has to be verified, but we actually do not see why it should not work.

The great advantage would be that all additional and external software to monitor the quality of results would become obsolete. Each step where a result can be obtained with less complexity and less software is a step in the right direction.

The software ERR_PROP was developed for SVI 2004/015. This will be renamed PRE_VISI and extended to apply short term forecasts with the 3+ methods mentioned above. PRE_VISI will be programmed to do short term forecasting from the same historical data with all methods within the same run in parallel. This is the only way to insure comparability.

5 Feedback and Learning

Another aspect which will be studied in parallel is to introduce feedback mechanisms to let the procedure “learn” from past experiences and to enhance the quality of results with time. What parameter can / should be monitored? How will they have to be modified to increase the forecasting precision over time? What precision can be reasonably obtained?

Spezielle Geräte und Installationen (Englisch)

None

Allgemeiner Stand der Forschung (Englisch)

Many practical applications of short term forecasting exist. However, there is no coherent overview on available methods with advices about their use, applicability, optimal setting of parameter limits, etc. What lacks is a global comparison of available methods with their respective advantages, disadvantages, forecasting precision, effect on error propagation and learning potential through feedback

Projektziele (Englisch)

Analyze, evaluate and compare 3+ short term forecasting methods like cluster analysis, regression, Fourier series and answer questions like:

1) How is knowledge built from past data?

2) How is this knowledge applied to the near future?

3) For Fourier series only: is forecast in the frequency domain possible / accurate?

4) How is quality monitored and controlled?

5) Will explicit error propagation simplify quality control? (by avoiding more SW to do it externally)

6) Will feedback theory help to “learn” from past and increase forecasting precision over time?

Forschungsplan (Englisch)

Step1 (2-3 months) Survey, collection of historical data, experiment set up

Survey among members of SVI about the use of methods

Literature search on internet

Decision about which methods will be analyzed in depth

Extension of software ERR_PROP è PRE_VISI to include short term forecasts

Collection of daily variation curves for different situations

Experiment setup for individual and comparative analysis of all methods

1^st meeting with the steering group

Step2 (5-6 months) Calculations, optimisation, comparison

Use of PRE_VISI for all calculations, production of results for analysis

Parameter optimization for each method

Conclusion for each individual method

2^nd meeting with the steering group

Step3 (2-3 months) Conclusions, Final Report

Comparison of all methods

Advices about the use of methods

3^rd and last meeting with the steering group

The research time will be strictly limited to a maximum of one year

Abstract (Deutsch)

.

Umsetzung und Anwendungen (Deutsch)

.

Umsetzung und Anwendungen (Englisch)

What method is most appropriate to a) build knowledge from past data, b) apply this knowledge to the future, c) apply quality monitoring by error propagation, d) learn from past through feedback mechanisms?

Give an overview about available methods for public and private administrations, software houses and other instances concerned with short-term forecasts.

- Help administrations to formulate specifications and evaluate bids.

- Help software houses to choose among methods.

Help users evaluate short-term forecast results.

Berichtsnummer (Deutsch)

1206

Literatur (Englisch)

[BAST] Bundesanstalt für Strassenwesen: Technische Lieferbedingungen für Strecken­stationen (TLS), Ausgabe 2002

[Bevington] Philip R. Bevington and D. Keith Robinson, Data Reduction and Error Analysis for the Physical Sciences, 3^rd ed., McGraw Hill 2003

[Boyce] Boyce, D.E., B. Ralevic-Dekic, H. Bar-Gara, Convergence of Traffic Assignments: How much is enough?, Journal of Transportation Engineering, 130(1) 49-55.

[Cooley, Lohnes] W.W. Cooley, P.R. Lohnes Multivariate Data Analysis, Wiley 1971

[Dietrich] , Karl Dietrich et al., Strassenprojektierung, IVT ETHZ 1993

[FGSV 365] Hinweise zur Schätzung von Verkehrsbeziehungen mit Hilfe von Querschnittszählungen, FGSV, AG Verkehrsführung und Verkehrssicherheit, Köln 1995

[Halpern] Joseph Y. Halpern, Reasoning about Uncertainty, The MIT Press 2003

[de Jong] G. de Jong, A. Daly, M. Pieters, S. Miller, R. Plasmeijer, F. Hofman, Uncertainty in Traffic Forecasts: Literature Review and New Results for the Netherlands. European Transport Conference, Strasbourg 2005

[Krempel] L. Krempel, Visualisierung komplexer Strukturen, Campus Verlag, Frankfurt 2005

[Ortuzar, Willumsen] J. de D. Ortuzar, L.G. Willumsen, Modelling Transport, Wiley 1990

[Rabinovich] Seymon G. Rabinovich, Measurement Errors and Uncertainties, Theory and Practice, 2^nd ed., Springer 2000, ISBN 0-387-98835-1

[de Rham 86] C. de Rham, Herleitung von Verkehrsbeziehungen des öffentlichen Verkehrs aus Verkehrszählungen in Zügen. Auftrag Nr. 5-A111 Stab für Gesamtverkehrsfragen 1986

[de Rham 87] C. de Rham, Umlegung, Überprüfung und Korrektur von Wunschlinien des Personenverkehrs anhand von Querschnittszählungen. Auftrag Nr. 5-A133 Stab für Gesamtverkehrsfragen 1987

[de Rham 03] C. de Rham, Incident Detection with Floating Car Data, paper #4069, 10^th World Congress on Intelligent Transport Systems 2003 Madrid

[de Rham 04] C. de Rham and Michelle Sisto, How to Quantify the Quality of Matrix Estimation with Traffic Counts?, paper #2799, Congress on Intelligent Transport Systems in Europe 2004 Budapest

[de Rham 05] C. de Rham and Michelle Sisto, From Sensors to Happy Users: a Long Way, paper #28779, 5^th European Congress on Intelligent Transport Systems 2005 Hannover

[Roughgarden] Tim Roughgarden, Selfish Routing and the Price of Anarchy, The MIT Press 2005

[Schnabel/Lohse] Schnabel W., Lohse D. Grundlagen der Strassenverkehrstechnik und der Verkehrsplanung, Band 2, Verlag für Bauwesen, Berlin 1997

[Taylor] John R. Taylor, An Introduction to Error Analysis, 2^nd ed., University Science Books 1997