Exploring a deep learning based spatio-temporal forecasting model: testing for scalability and other features

Abstract

In recent years, there has been a rapid increase in demand for forecasting services in the fields of solar power gathering and city pollution control, among many other spatio-temporal variable use cases. It is important that we anticipate this demand by developing new models and architectures capable of accurately predicting these variables. But our efforts must not be limited towards achieving greater accuracy alone. Other non-functional characteristics such as flexibility, robustness and scalability are also required. In this thesis, a novel architecture based on deep learning technologies will be explored. The main aim being to test the scaling capabilities of such a model, as well as implementing features that would allow a successful deployment of the model in a real-world like scenario.