The aim of the internship is to develop a method of traffic flow prediction based on Deep Learning technics. The dataset is provided by the California Department of Transportation (Caltrans) and its Performance Measurement System (PEMS).
IFSTTAR MLV : RER A / Noisy-Champs
GRETTIA lab wishes to investigate the potential of deep learning algorithms applied to short-term traffic flow prediction. Smart highway systems are commonly monitored using sensing technologies such as inductive loop buried in the pavement: they record the number of vehicles, and their speed, passing over the loop during a unit of time. In the case of the Caltrans Performance Measurement System (PEMS), data are collected from nearly 40000 individual detectors spanning the freeway system across all major metropolitan areas of California: over ten years of data are archived, with a variety of information (traffic detectors, speed estimation, incidents, lane closures, census traffic counts, toll tags…). Traffic flow (or volume) over a period of 15 minutes is expressed in Vehicles/15 minutes.
Vehicle Detector Stations (VDSs) are spatially distributed along a road, and the distance between stations can affect the spatial correlation between measurements of two VDS. Often time in publications, VDS measurements are aggregated for a given highway so that the traffic volume studied is actually the mean over the whole highway. Therefore, traffic flow forecasting of one highway is performed using past mean volume data of the highway network. If the prediction is based on k=4 previous times, and m=10 highways are considered, then forecasting of one highway volume is based on 40 measurements (40 inputs for 1 output). We believe it is possible to predict the traffic volume at the level of each VDS of a highway, this information being more valuable than the mean over one road. In this case, for m=10 highways with a total of s=100 stations, considering k=4 previous times, such a forecasting method require to design a network with 400 inputs and 100 outputs.
The contribution of this internship will be mainly experimental: implement neural networks learning to adapt recent methods of deep learning to the processing of data collected from the PEMS database.
The trainee's mission will be as follows:
• gather information and data from the PEMS website, and acquire an expertise on it. Parsing the files for number of vehicles/5 minutes of each VDS, finding their location from their identification number.
• implementing the traditional one output predictor with mean volumes of the highways as input
• design a multitask network predicting the traffic flow for several VDS stations
• In both cases, the model of Deep Stacked Auto-Encoder DSAE will be developed to compress the data in a non-linear way. In the latter, an extension with CNN (Convolutionnal Neural Network) or a convolutional layer will be investigated in order to model the correlations between neighbor stations.
The functionalities to be implemented will be based on the Theano library.
An important part of the subject regards the technical documentation, in order to ease further development of the project by other collaborators.
<!--[if gte mso 9]> <
(c) GdR 720 ISIS - CNRS - 2011-2018.