High-Dimensional Probabilistic Time Series Forecasting Via Wavenet+Timegrad
Document Type
Conference Proceeding
Publication Date
1-1-2023
Abstract
Time series forecasting has been used extensively in different fields (e.g., audio signal/real-time video processing, stock price prediction, workload projection). Traditional time series models have difficulties in handling the complexity and non-linearity in high dimensional time-series data. In this work, we combine truncated WaveNet framework with TimeGrad process to provide probabilistic multi-step forecasting for high dimensional time series data. Dilated causal convolutional neural network (DC-CNN), skip connections, and residual networks are the major components of WaveNet framework. DC-CNN and residual networks are used to extract the trend and short-term movement information from time series data and solve the degradation problem, respectively. TimeGrad process functioned as generating probabilistic forecast that provide confidence information of the predictions. The experimental results on six large real-world datasets demonstrate that WaveNet+TimeGrad model performed better (lower CRPSsum value) than the benchmark models and previous WaveNet+t model. WaveNet+TimeGrad probabilistic model performed especially well (higher accuracy, less memory resource consumption) with 'Direct Method' when the dimension of time series data becomes larger.
Publication Source (Journal or Book title)
Proceedings - International Conference on Machine Learning and Cybernetics
First Page
187
Last Page
193
Recommended Citation
Sun, X., & Chen, J. (2023). High-Dimensional Probabilistic Time Series Forecasting Via Wavenet+Timegrad. Proceedings - International Conference on Machine Learning and Cybernetics, 187-193. https://doi.org/10.1109/ICMLC58545.2023.10327945
