PIAD-SRNN: Physics-Informed Adaptive Decomposition in State-Space RNN

TL;DR

PIAD-SRNN is a physics-informed, adaptive decomposition RNN that effectively separates components of time series data, leading to superior forecasting accuracy and efficiency over state-of-the-art models, especially in indoor air quality prediction.

Contribution

It introduces a novel physics-informed adaptive decomposition framework within a state-space RNN for improved time series forecasting.

Findings

Outperforms SOTA models in MSE and MAE across various horizons

Balances accuracy and computational efficiency

Provides four curated indoor air quality datasets

Abstract

Time series forecasting often demands a trade-off between accuracy and efficiency. While recent Transformer models have improved forecasting capabilities, they come with high computational costs. Linear-based models have shown better accuracy than Transformers but still fall short of ideal performance. We propose PIAD-SRNN, a physics-informed adaptive decomposition state-space RNN, that separates seasonal and trend components and embeds domain equations in a recurrent framework. We evaluate PIAD-SRNN's performance on indoor air quality datasets, focusing on CO2 concentration prediction across various forecasting horizons, and results demonstrate that it consistently outperforms SoTA models in both long-term and short-term time series forecasting, including transformer-based architectures, in terms of both MSE and MAE. Besides proposing PIAD-SRNN which balances accuracy with efficiency, this paper also provides four curated datasets. Code and data: https://github.com/ahmad-shirazi/DSSRNN