Application of LSTM Neural Networks for Air Quality Index Class Forecasting

The problem of forecasting the Air Quality Index (AQI) is crucial for preventing the adverse effects of air pollution on people and improving the environmental situation. This issue is particularly relevant for Bishkek, a city significantly affected by air pollution. This work addresses the task of forecasting AQI as a classification problem and develops a classifier based on an LSTM neural network for its solution. The classification was based on meteorological parameters corresponding to AQI values measured every 3 h, which serve as the input variables for the neural network. Additionally, the daily amount of coal burned at the thermal power plant in Bishkek, a major pollution source, was used as an extra input variable. The developed classifier can predict AQI as one of four possible classes: AQI ≤ 50 (“Good”), 50 < AQI ≤ 100 (“Moderate”), 100 < AQI ≤ 150 (“Unhealthy for sensitive groups”), and AQI > 150 (“Unhealthy”, “Very Unhealthy”, “Hazardous”). Computational experiments were conducted, varying the depth of AQI forecasting (up to 2 days ahead) and the length of input variable sequences (up to 4 days of data history) due to the inertia of air pollution processes. Incorporating the pollution factor as an additional input increased the accuracy of AQI class prediction to 83%. Further expansion of the observational history will enable the creation of a classifier for predicting AQI classes for AQI > 200. #CSOC1120.