Abstract
This study explores the use of three machine learning models—Linear Regression, XGBoost, and Long Short-Term Memory (LSTM) to predict air pollutant concentrations based on meteorological factors in Almaty, Kazakhstan, from 2021 to 2024. The primary aim is to assess how well these models can capture the relationships between environmental variables such as temperature, humidity, and wind speed, and air quality metrics including PM2.5, PM10 and CO. Linear Regression was used as an initial approach to model linear dependencies, while XGBoost was employed to capture more complex, non-linear relationships, and LSTM was utilized to model temporal dynamics in the data. Performance was evaluated using Mean Square Error (MSE), Mean Absolute Error (MAE), and R-squared (R2) values, with XGBoost showing the best overall performance, particularly for PM2.5, with an R2 of 0.803. LSTM demonstrated strong predictive power for time-series data but had slightly lower R values compared to XGBoost. The study concludes that while XGBoost outperforms Linear Regression and LSTM in terms of explaining variance in air quality, all models have certain limitations, including data quality issues and a limited dataset. Further research with more comprehensive datasets could enhance the predictive capabilities and generalizability of these models for air quality forecasting.