LSTM-based prediction of PM2.5 and PM10 air pollution: A case study of İstanbul, Turkiye

Abstract

This thesis focuses on the important task of forecasting PM10 and PM2.5 levels in Istanbul. It utilizes LSTM-based models to generate precise predictions for the next hour. The models demonstrate a noteworthy level of precision, with PM10 achieving an accuracy of around 0.87 and PM2.5 achieving an accuracy of 0.89. This highlights their potential importance in providing timely alerts to the public, especially those with respiratory conditions. Moreover, these models have the additional benefit of estimating PM levels when sensors are not operational, guaranteeing continuous monitoring and response strategies. The research commences by providing a context for the influence of air pollution on public health, with a particular focus on the detrimental consequences of particulate matter (PM). Istanbul, chosen as the study area, features 39 strategically positioned air monitoring stations throughout the city. Data acquisition involves the combination of hourly weather data sourced from NASA and air pollution data obtained from the Turkish Ministry of Environment. The Ümraniye station was chosen for modeling due to its extensive data availability, which includes monitoring of PM10 and PM2.5, as well as other pollutants, from January 2018 to August 2023. The pre-processing steps involve detecting and eliminating outliers using the Interquartile Range method, and then filling in missing values using the K-nearest neighbors (KNN) algorithm. After normalizing the data and carefully dividing it into a 60-20-20 split for training, validation, and testing, three different LSTM architectures (Vanilla, Stacked, and Bidirectional) were tested using different sizes of sliding input windows. The results demonstrate that the Stacked LSTM model outperforms other models in predicting PM10, with a Root Mean Square Error (RMSE) of 6.96 and a coefficient of determination (R2) of 0.871. The Vanilla LSTM model is the best performer for PM2.5, with an RMSE of 2.74 and an R2 of 0.894. Tests conducted on other monitoring stations show promising results in terms of generalization, with R2 scores ranging from 0.77 to 0.924 and RMSE values spanning from 4.3 to 13.4. The Üsküdar station is particularly noteworthy for its exceptional performance in terms of having the lowest RMSE and the highest R2 scores, as observed in both Vanilla and Stacked models. This outstanding performance could possibly be attributed to its close proximity to the Ümraniye district. The study openly acknowledges certain limitations while acknowledging the success of the models. Incorporating estimated weather data, alongside actual ground-measured air pollution data, can introduce potential biases that may result in increased root mean square error (RMSE) and decreased coefficient of determination (R2) scores. The study also highlights the need for caution when interpreting hourly predictions and proposes future research directions, such as the creation of multi-step models to improve forecasting for longer periods of time. Moreover, it emphasizes the significance of determining feature importance in order to distinguish the precise contributions of individual factors to the prediction models. To sum up, the LSTM-based models for PM10 and PM2.5 in Istanbul show encouraging accuracy and provide important information for environmental and public health monitoring. Although the study has certain limits that exist, it establishes the basis for future research by indicating the need for multi-step models and analysis of feature importance. This research makes a valuable contribution to the overall discussion on predicting air quality in urban areas, and it paves the way for future progress in this field.