Doctor of Philosophy (PhD)


Division of Computer Science and Engineering

Document Type



A time series refers to a collection of data points that are arranged in chronological order and collected over a sequence of time. Time series prediction involves predicting future values based on historical records. The applications of time series prediction span across diverse domains, including stock market trends, traffic flow patterns, monitoring of COVID-19 cases, and tracking global temperature changes. The advent of the IoT era results in the generation of high-dimensional time series data, which are acquired simultaneously from various sources through multiple sensors or channels. Traditional statistical models have limitations in handling the complexity, non-linearity, adaptability, and high dimensionality of real-world time series data.

To overcome these challenges, advanced approaches such as machine learning algorithms and deep learning models were developed to provide more accurate and flexible solutions for time series prediction. Deep learning models have demonstrated superior performance compared to traditional methods. This study primarily focuses on leveraging deep learning models for the analysis of high-dimensional time series data.

We leverage the recent development and develop new models to further mitigate the above issues. The truncated WaveNet + Probabilitic hybrid models, including WaveNet+t, WaveNet+TimeGrad, and GraphWaveNet+TimeGrad, serve as powerful tools for tackling complex problems. These models combine the strengths of WaveNet and harness the power of probability density estimation through Student’s distribution and TimeGrad to effectively mitigate overfitting issues through interval estimation. Additionally, two prediction strategies (Iterative vs. Predict_All) were implemented. These hybrid models outperform nine competitive benchmark models across six real-world datasets based on the metric.

Furthermore, transfer learning concept was explored in this dissertation. ResNet-18 was introduced as a feature pre-processing component for the TransferLearning+WaveNet+TimeGrad model. Although it did not meet our previous expectations, it provides valuable insights for further improvement. Future research will focus on investigating an alternative parameter transfer approach to enhance the model's performance.



Committee Chair

Chen, Jianhua

Available for download on Thursday, December 26, 2030