A1769
Title: Probabilistic forecast for time series with transformer-based models
Authors: Thu Nguyen - University of Maryland Baltimore County (United States) [presenting]
Abstract: Time series forecasting plays a pivotal role in contemporary data-driven sectors, spanning domains such as finance, energy management, and manufacturing, among others. The aim is centred on the long sequence forecasting (LSFT) task within the realm of time series data, introducing a new approach that leverages Transformer-based neural architectures. While Transformers are celebrated for their adaptability in sequence modelling, they grapple with substantial challenges when applied to LSFT, including quadratic time complexity, memory utilization, and inherent limitations in the encoder-decoder architecture. Although recent research has made strides in mitigating these issues, intricate challenges endure. A novel Transformer-based LSFT model architecture is proposed. The model incorporates new attention mechanisms designed to capture the intricacies of time series characteristics. Additionally, a hybrid modelling approach is adopted that combines statistical components with deep learning methodologies.
