Air Quality Index (AQI) prediction is vital for reducing the negative impacts of air pollution on both public health and the environment. While previous methods, particularly those using deep learning, have delivered impressive results, they often fall short in capturing temporal dependencies and managing datasets with numerous outliers. This study introduces a new hybrid model that combines Inception Time and Bidirectional Long Short-Term Memory (BiLSTM) networks to enhance AQI prediction performance. The InceptionTime model is adept at capturing multi-scale temporal patterns, whereas the BiLSTM component leverages sequential dependencies in the data. Furthermore, we use Huber Loss to optimize model training and increase robustness against outliers. Our model is evaluated on two public air quality datasets and outperforms existing approaches, demonstrating its effectiveness as a reliable tool for accurate AQI prediction.