Traffic speed prediction and privacy model with vehicle data

In this thesis study, three novel and new models have been developed for the prediction of traffic parameters and providing trajectory privacy, tested in heavy and medium-heavy traffic time zones on two main roads in Ankara, and successful results have been obtained. In the first proposed traffic speed prediction model, the vehicle data, which were passed through the filtering, data cleaning and map-matching processes for the traffic speed prediction, were density-based clustered using OPTICS algorithm and thus characteristic features of each of roads were generated, predictions were made with LSTM and GRU neural network models considering the recurrent patterns of road characteristic information, these models were tested on Eskisehir and Istanbul roads, and RMSE error values in km/h were obtained with the highest performance of 7.814 in heavy traffic and 4.372 in medium heavy traffic, respectively. The privacy-preserved second prediction model was developed based on the first proposed model with applying differential privacy (DP) method to the model an intermediate stage, DP method was adopted for the first time to a speed prediction model, was tested on Eskisehir and Istanbul roads with two different datasets compared with the original cluster speed features, the RMSE error values with the highest performance were measured as 6.127 km/h in heavy traffic and 8.789 km/h in medium heavy traffic on the Istanbul road. The third one introduces a new privacy-preserving model which is window-based integrated with DP applied to ensure privacy when publishing the trajectory information. Like the previous two models, this model was also tested on Eskisehir and Istanbul roads, the RMSE error values of x-y coordinates were obtained as 0.00031973 and 0.00013009 degrees, 0.00011643 and 0.00008601 degrees, respectively, for the sample vehicles producing the highest GPS signal at the value of epsilon and window size 2. The proposed models in this thesis were developed within the scope of TÜBİTAK project given number 3191873 and, improvements in the studies still continue in order to apply for real-time implementation. As a result, it is concluded that the proposed models might offer different and new solutions in the prediction of real-time traffic parameters for future studies and can be used in real applications, will guide privacy-oriented studies, and can be applied in different data publishing research fields.