Date of Award
9-6-2024
Publication Type
Thesis
Degree Name
M.A.Sc.
Department
Mechanical, Automotive, and Materials Engineering
Keywords
Decomposition;GAN;Multi-Agent Systems;Time Series Forecasting;Transformers
Supervisor
Afshin Rahimi
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Abstract
Time-series forecasting is crucial in various industrial systems' efficient operation and decision-making processes. This thesis presents a novel multi-agent decomposition model with a Generative Adversarial Network, called MAD-GAN, to predict the long- and short-term multivariate time series data. Firstly, the various time series forecasting models used in manufacturing systems, including statistical models such as Moving Average, machine learning models such as Support Vector Machines, Gradient Boosting, etc., and the modern deep learning and generative models have been reviewed and compared. A case study was performed using a manufacturing systems dataset to understand the implications of Prophet and regression-based machine learning models on short- and long-term forecasting. Finally, a multi-agent architecture is introduced that uses agent abstractions to work on direct multi-step forecasting. This architecture consists of three Agents. Decomp-Agent incorporates raw time series data decomposition into more predictable seasonality and cyclic trend components. Model-Agent provides the best model configuration from various linear models, and Adversarial-Agent picks the best model from Model-Agent, enhancing its prediction accuracy and robustness with adversarial training. The model is validated through extensive empirical evaluations with eleven benchmark multivariate time series datasets, including the manufacturing, weather, and financial datasets, and compared against state-of-the-art models. The results suggest that MAD-GAN outperforms existing transformer-based models for long-term forecasting and shows promising short-term forecasting results. This research underscores the potential of multi-agent architectures in time series forecasting and the impact of adversarial training on the existing deterministic models, opening new avenues for further exploration in this field. Furthermore, the robustness of the proposed model can be extremely useful for industrial manufacturing systems that involve datasets of varied distributions.
Recommended Citation
Fatima, Syeda Sitara Wishal, "MAD-GAN: Multi-Agent Decomposition with GANs for Time Series Forecasting in Manufacturing Systems" (2024). Electronic Theses and Dissertations. 9400.
https://scholar.uwindsor.ca/etd/9400