Advancing Time Series Forecasting: Innovative Machine Learning Approaches

Loading...
Thumbnail Image

Authors

Zhang, Yifan

Issue Date

2024

Type

Dissertation

Language

en_US

Keywords

Deep Learning , Recurrent Neural Networks , Time Series , Transformer

Research Projects

Organizational Units

Journal Issue

Alternative Title

Abstract

The importance of multivariate time series (MTS) data is increasingly acknowledged in scientific and technological domains. Effectively modeling historical time series records and accurately forecasting future events remain critical challenges within the field of artificial intelligence. Two key issues in this domain require further investigation. First, the probability distribution of time series data may vary over time. Second, effectively and efficiently modeling temporal dependencies remains a significant challenge. This dissertation tackles the problem of MTS forecasting using advanced deep neural network techniques and is organized into three parts. The first part provides an overview of the historical development of time series forecasting and highlights recent advancements. We also discuss potential future directions for addressing this challenge and outline our motivations. In the second part, we explore the use of recurrent neural networks (RNN) to model the evolving distributions of time series data under extreme conditions. We introduce a framework that integrates RNN with other machine learning models to forecast normal and extreme time series instances independently. Additionally, we present a novel extreme event adaptive gated recurrent unit (eGRU) architecture that models both normal and extreme events using the same RNN cell. The effectiveness of the first approach is demonstrated using real-world datasets obtained from our collaborators, while the second approach is validated using publicly available benchmarks. In the third part, we investigate the application of transformer neural networks for modeling temporal patterns in time series data. We propose a pyramid architecture comprising multiple transformer encoder-decoder pairs to capture temporal dependencies at various scales. Additionally, we investigate the efficiency of the self-attention mechanism, the core component of the transformer, whose quadratic computational complexity limits its use in MTS forecasting by constraining the size of the look-back window. To address this limitation, we introduce a sequence-adaptive sparse attention mechanism that reduces computational complexity to a linear scale.

Description

Citation

Publisher

License

CC BY-NC

Journal

Volume

Issue

PubMed ID

DOI

ISSN

EISSN