Spatial-Temporal Dependency of Traffic Flow and Its Implications for Short-Term Traffic Forecasting

This dissertation, "Spatial-temporal Dependency of Traffic Flow and Its Implications for Short-term Traffic Forecasting" by Yang, Yue, 樂陽, was obtained from The University of Hong Kong (Pokfulam, Hong Kong) and is being sold pursuant to Creative Commons: Attribution 3.0 Hong Kong License. The content of this dissertation has not been altered in any way. We have altered the formatting in order to facilitate the ease of printing and reading of the dissertation. All rights not granted by the above license are retained by the author. Abstract: Abstract of thesis entitled Spatial‐temporal Dependency of Traffic Flow and Its Implications for Short‐term Traffic Forecasting Submitted by Yue Yang for the degree of Doctor of Philosophy at The University of Hong Kong in September 2005 Short-term traffic forecasting is of great significance to modern transport management. It can help management centres and individual travellers to make better travel decisions and offers a more rational and effective way to alleviate traffic congestion and redistribute traffic more evenly over a road network. Several forecasting models have been proposed in recent decades, but no single method has been able to consistently outperform its rivals under all conditions. Some of these models have produced even worse results than the traditional historical average method, making them practically valueless. Attempts at econometric forecasting also support the above observations. This thesis, which emphasizes the importance of a systematic approach in understanding traffic phenomena, first examines the spatial-temporal dependency of traffic flow (i.e. how traffic flows are related in both dimensions) by Cross Correlation Function (CCF) analysis which quantifies this dependency using correlation coefficient and time lag. Based on the implications of the spatial-temporal relationship of traffic flow, it then proposes a short-term traffic forecasting framework, using an adaptive forecasting model selection strategy which blends effective real-time and historical data in different proportions to suit different forecasting horizons instead of attempting to create a single model to deal with all forecasting settings. It also develops an improved Kalman filter - spatial-temporal Kalman filter (STKF), whose parameters and coefficients are determined by the effective real-time data and their weights. The forecasting strategy is examined using three typical examples, in which the forecasting horizon is respectively within, equal to, and beyond the network maximum up-trace time. The STKF is used for the first two situations, where higher forecasting accuracy is desirable. In the third situation, the historical average method is used, and the thesis demonstrates that this compares favourably with ARIMA (Autoregressive Integrated Moving Average) and NN (neural network) in terms of forecasting accuracy and overall costs. The thesis does not seek to propose an innovative forecasting model. Rather, it focuses on the traffic flow generation and evolution process among road links, and emphasizes that the propagation of traffic flow must be properly understood if traffic forecasting is to be effective. It makes the following novel contributions: (1) it explicitly quantifies the spatial-temporal relationships of traffic flows observed at different locations; (2) it suggests that the spatial-temporal dependency of traffic flows should be a major factor in traffic forecasting method; (3) it examines the effectivity of real-time data and forecastability of traffic conditions in forecasting; (4) it proposes an adaptive forecasting strategy based on these notions, in which the forecasting time horizon and the network storage for the provision of effective real- time data determine the choice of forecasting method; (5) it develops a spatial- temporal Kalman filter which outperforms ARIMA and