Overview

• Enabled accurate prediction of traffic speed under various factors (e.g., floating car data, weather condition, air quality, weekend & holiday) using HDL-net, a hybrid deep learning model

• Built, trained, and validated a multi-layer perceptron (MLP) neural network as a plugin to the HDL-net for predicting the impact of various factors

• Transformed the traffic speed data to 2D tensors; built convolutional neural network (CNN) and recurrent neural network (RNN) models to extract the spatial-temporal relationship between traffic speeds and specific road segments

• Introduced the Convolutional Block Attention Module to improve the effectiveness and interpretability of the models

• Developed a software app to facilitate intelligent traffic monitor based on the model prediction; employed Unified Modeling Language (UML) to guide software design, and GitHub to promote team collaboration and version control

• Enabled multi-scale traffic speed prediction based on a suite of neural network models in TensorFlow:

  • Collected transportation data from multiple sources, followed by data cleaning to analyze the road characteristics and to extract their topological information

  • Performed 1) short-term (5 minutes to 1 hour) speed prediction using LSTM and GCN models, 2) near-term (1 day) speed prediction via RNN and KNN models, and 3) long-term speed prediction using random forest, neural network, and BiLSTM models

  • Systematically compared the pros and cons of the aforementioned models, focusing on the time efficiency and prediction accuracy across various types of roads (expressway, main lane, side lane, and alley)

• Designed and prototyped a software app in PyQt5 to enable fast query and visualization of predictive results for specified times and road segments