Travel time prediction using LSTM

• Main Author: Narayanan Aakash Kumar
• Other Authors: Justin Dauwels
• Format: Theses and Dissertations
• Language: English
• Published: 2019
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: http://hdl.handle.net/10356/78673
• Institution: Nanyang Technological University

The aim of the research is to predict the time it will take for a vehicle to travel across a stretch of road. This prediction will be useful for autonomous cars to estimate the travel times in multiple roads and to choose the fastest route possible thereby reducing congestion in traffic as well as improving efficiency. Also, this prediction will be useful to improve the service quality of commercial delivery by delivering goods with given period. The main requirement for this research is the historical data of travel times collected on a stretch of road. So, the travel time taken by vehicles in 29 different road segments were recorded for a period of 3 years with a frequency of 3 minutes and this data was used to train the deep learning model. Multiple deep learning techniques were evaluated to solve this time series prediction problem and LSTM (Long Short-Term Memory) stood out as a genuinely suitable model. Due to its long-term memory, the LSTM model can recognize trends. Multiple researches on LSTM have shown that it reads the trends in the data and each neuron looks out for a pattern in the data. When trained over a piece of C code, the LSTM can even write a new code without any syntax error although without much meaning. Thus, an LSTM model was trained over the historical data. For the prediction, multiple models were trained for different scenarios like different prediction horizons and different number of input features. The performances were compared against each other to find out the best prediction horizon and number of input features to get the most accurate prediction. Some other popular time series prediction models are also trained on the same historical data and the MAPE values of all the algorithms are compared at the end to identify the best performing algorithm. 8 The data was divided into two segments. 90% was used for training and 10% was used for testing. The outliers in the data were chosen as the values which were away from the mean by more than 3 standard deviations. Since outliers can’t be directly removed in a time series data, it is replaced by linearly interpolating the nearest non-outlier data. After cleaning the data, separate models are used to predict travel time in different road segments and the MAPE (Mean Absolute Percentage Error) is called for all the segments under different scenarios. It is also possible for a single LSTM model to predict travel time on multiple road segments, but the accuracy reduces by a small fraction. Based on experimental prediction results on the real-life traffic data, the LSTM model has proved that it can be used in real life scenarios for predicting travel time.