AIR QUALITY INDEX FORECASTING USING HYBRID NEURAL NETWORK MODEL WITH LSTM ON AQI SEQUENCES

Abstract

This paper presents an approach to forecasting air pollution levels measured as Air Quality Index (AQI) metric using hybrid Long Short-Term Memory (LSTM) models. The pollution levels have been found to vary in a particular pattern that depends on both the overall climate or season as well as the hour of the day. The hybrid model captures these 2 patterns and makes the prediction of AQI of some future hour. It employs 2 separate LSTM models that are trained on time-series data of AQI gathered at different time lags i.e. hourly and daily. The final output is given as a weighted sum of the 2 outputs produced by LSTM model. Upon comparing the performance of the standalone hour-wise forecasting LSTM model and the hybrid model it was found the latter gives the minimum error metric given an appropriate weight is chosen.