A hybrid deep learning framework for urban air quality forecasting.

Deep learning models address air quality forecasting problems far more effectively and efficiently than the traditional machine learning models. Specifically, Long Short-Term Memory networks (LSTMs) constitute a significant breakthrough in understanding the complex sequential behavioral dependencies of the time series. Further, LSTM models justify well with the speed–accuracy tradeoff, among other deep learning models. However, there are several limitations of such deep learning models. Firstly, the addition of multiple hidden layers, on the one hand, improves the performance but, on the other hand, requires extensive hardware and computation capabilities. Secondly, most of the previous works that utilized LSTMs for air quality forecasting do not consider the issue of optimal hyperparameter calibration. While deciding the gradient, network learning parameters should be so fixed such that the model does not underfit or overfit. To address these issues, a stochastic optimization algorithm, mimicking the pattern of flocking birds, is utilized to find the most fitting solution in the parameter search space. Particle swarm optimization setup primarily models varying particles representing parameters to reach an optimum state. Furthermore, the Spatio-temporal instabilities of LSTM models are addressed in this work using preprocessing, segmentation and feature engineering to understand seasonal and trend characteristics along with the Spatio-temporal correlation of the time series. The proposed model is employed on the air quality dataset of 15 locations in India. A variety of experiments are performed to prove the superiority of the proposed method. Firstly, a comparison with traditional sequential models and deep learning models is done. Secondly, results are further evaluated over several existing benchmark dataset samples. Results suggest that the proposed method outperforms existing forecasting models when evaluated over a variety of performance metrics. • Complex spatio-temporal dependencies in the data renders model learning task difficult. • Long Short-Term Memory network are utilized to model complex sequential dependencies. • Optimal hyperparameters selection using swarm intelligence is done. [ABSTRACT FROM AUTHOR]