Document Type : Original Article
Authors
1 Department of Computer Sciences, Shahrekord University, Shahrekord, Iran
2 Department of Geology, Faculty of Sciences, Urmia University, Urmia, Iran
Abstract
This study presents a comprehensive benchmarking analysis of state-of-the-art recurrent neural network (RNN) architectures for sequential prediction tasks. A wide range of recurrent models, including Simple RNN, Independently Recurrent Neural Network (IndRNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional LSTM (BiLSTM), Residual BiLSTM (ResBiLSTM) are systematically investigated under a unified experimental framework. The proposed benchmarking framework evaluates all models on a real-world air pollution prediction task using multivariate environmental and meteorological data. Performance is assessed using multiple standard metrics, including root mean square error (RMSE), mean absolute error (MAE), and the coefficient of determination (R²), enabling a comprehensive evaluation of predictive accuracy and generalization capability. Experimental results demonstrate that IndRNN achieves the best overall performance, providing superior accuracy while maintaining a relatively simple and computationally efficient architecture compared to other deep recurrent variants. The actual-versus-predicted analysis further supports this conclusion and IndRNN exhibits the closest alignment between predicted and observed values, achieving the highest coefficient of determination (R² = 0.6375) and the lowest prediction error (RMSE = 17.0203) among the evaluated recurrent architectures. These findings highlight the effectiveness of IndRNN as a strong candidate for practical sequential forecasting applications.
Highlights
- Comprehensive benchmarking of six recurrent neural network architectures for sequential prediction.
- Unified evaluation using real-world multivariate air pollution and meteorological data.
- Comparative assessment based on RMSE, MAE, and $R^2$ predictive performance metrics.
- IndRNN demonstrates superior predictive accuracy across the evaluated recurrent architectures.
- IndRNN achieves $R^2=0.6375$ and RMSE$=17.0203$, supporting its practical forecasting potential.
Keywords
[10] Karmoude, M., Munhungewarwa, B., Chiraira, I., et al. (2025). Machine learning for air quality prediction and data analysis: Review on recent advancements, challenges, and outlooks. Science of the Total Environment, 1002, 180593. https://doi.org/10.1016/j.scitotenv.2025.180593.