NRL4AQF: Noise-Resistant Learning for Long-Sequence Air Quality Forecasting using Cross-Time Embedding

Recently, deep learning (DL) has greatly advanced data analysis, including time-series forecasting. Recurrent neural networks (RNNs) are commonly used for DL-based forecasting, but they face limitations in capturing complex, long-range dependencies, especially with cross-dimensional input data. To address these challenges, transformer-based architectures have been introduced, offering improved performance for long-range predictions. However, they still struggle with modeling long input sequences and managing noise in temporal feature learning. To overcome these issues, we propose the NRL4AQF (Noise-Resistant Learning for Air Quality Forecasting) model, a novel denoising transformer-based approach for time-series forecasting. By combining cross-dimensional transformer-based embeddings with a radial basis function neural network (RBFNN), NRL4AQF effectively reduces noise and captures more accurate temporal patterns. Applied to long-sequence air quality forecasting, the model achieves superior results compared to traditional and state-of-the-art DL techniques, as demonstrated by experiments on a real-world dataset.