Time Series Analysis Based Air Pollution Level Prediction Using Long Short Term Memory Recurrent Neural Network with Attention Mechanism

Abstract

Air pollution has grown significantly as a threat to the environment during the past few decades. It has had a profound effect on the ecosystem, public health, and safety. For government agencies to design plans for pollution prevention, predicting the extent of air pollution becomes essential for maintaining the environment and environmental protection. One of the efficient approaches to solving the problem of predicting the level of air pollutants is time series analysis. The drawbacks of conventional machine learning algorithms in time series analysis are solved by employing deep learning algorithms in a variety of ways. Deep Learning models learn features and dynamics exclusively and directly from the data, in contrast to Machine Learning models, such as autoregressive models (AR) or exponential smoothing, where feature engineering is conducted manually and frequently certain parameters are adjusted while taking domain expertise into account. To forecast the level of pollution in the following hour using the current hour's weather and pollution levels, a Recurrent Neural Network (RNN) model based on Long Short Term Memory (LSTM) units has been trained, tested, and optimized. The predictive model has been developed with both uni-variate, which takes into account only one feature, and multi-variate, which takes into account many features. To improve the performance of the model on long input sequences attention mechanism has been implemented.