DSSRNN: Decomposition-Enhanced State-Space Recurrent Neural Network for Time-Series Analysis
Published in ICLM (submitted), 2025
Recommended citation: Mohammadshirazi, A., Nosratifiroozsalari, A., & Ramnath, R. (2024). DSSRNN: Decomposition-Enhanced State-Space Recurrent Neural Network for Time-Series Analysis. arXiv preprint arXiv:2412.00994. https://arxiv.org/pdf/2412.00994
Summary
The Decomposition-Enhanced State-Space Recurrent Neural Network (DSSRNN) is an advanced framework for time series forecasting, combining decomposition techniques with state-space and recurrent neural networks to handle seasonal and trend components effectively. Designed for both long-term and short-term forecasting, DSSRNN excels in predicting indoor air quality metrics like CO₂ concentrations, achieving superior accuracy and computational efficiency compared to state-of-the-art transformer-based and linear models. Its unified architecture supports both prediction and imputation tasks, ensuring robust handling of missing data while maintaining fast inference times and low resource requirements. This innovative model establishes a new standard for physics-informed machine learning, with applications in environmental forecasting and broader time series analysis.
Authors
Ahmad Mohammadshirazi, Ali Nosratifiroozsalari, Rajiv Ramnath