Deep Learning Models for Visibility Forecasting

Abstract

This dissertation addresses the task of visibility forecasting via deep learning models using data from weather stations. Visibility is one of the most critical weather impacts on transportation systems. Low visibility conditions can seriously impact safety and traffic operations, leading to adverse scenarios, causing accidents, and jeopardizing transportation systems. Accurate visibility forecasting plays a key role in decision-making and management of transportation systems. However, due to the complexity and variability of weather variables, visibility forecasting remains a highly challenging task and a matter of significant interest for transportation agencies nationwide. This dissertation explores the use of deep learning models for the task of single-step visibility forecasting (i.e., estimation of visibility distance for the next hour) using time series data from ground weather stations. The aforementioned task has not been fully addressed in the literature, thus, this work represents a baseline for further research. The author explores five neural network architectures: multilayer perceptron (MLP), traditional convolutional neural network (TCNN), fully convolutional neural network (FCNN), multi-input convolutional neural network (MICNN), and long short-term memory (LSTM) network. Models were evaluated using two datasets from Florida (one of the top states across the US dealing with visibility problems). Three cases of lag observations were considered: three, six, and nine hours input data.