Ph.D. Dissertation Defense: Advancing Time Series Forecasting - Innovative Deep Learning Approaches


The significance of multivariate time series (MTS) data is increasingly recognized in the realms of science and technology. Accurately modeling historical time series records and forecasting future events pose significant challenges in the field of Artificial Intelligence. Two key issues in this domain require further exploration. First, the probability distribution of time series data may change over time. Second, the challenge of effectively and efficiently modeling temporal dependencies remains. My research during my PhD training addresses the multivariate time series (MTS) forecasting problem using advanced deep neural network techniques. The presentation is structured 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 independently forecast normal and extreme time series instances. Additionally, we present a novel enhanced 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 consisting of multiple transformer encoder-decoder pairs, each level designed to capture temporal dependencies at a specific scale. MTS forecasting is based on multi-scale latent representations. We further examine the efficiency of the self-attention mechanism, the core component of the transformer, whose quadratic computational complexity constrains its use in MTS forecasting by limiting the size of the look-back window. To overcome this limitation, we introduce a sequence-adaptive sparse attention mechanism that reduces computational complexity to a linear scale.

Ph.D. Candidate: Yifan Zhang

Dial-In Information


Meeting ID: 891 1372 1021

Passcode: 441651

Friday, April 19 at 3:00 pm to 5:00 pm

William N. Pennington Engineering Building, WPEB 200