Elevating Air Quality Forecasting: Integrating Hybrid Clustering Techniques with Long Short-Term Memory Networks

Air pollution is a pressing issue in many urban areas, including Jakarta, Indonesia, where it poses significant health and environmental challenges. This research addresses the air pollution problem in Jakarta by proposing a hybrid approach that combines K-Means and K-Medoids clustering algorithms with a Long Short-Term Memory (LSTM) predictive model to forecast Air Pollution Standard Index (ISPU) datasets. Developed using data mining methodologies and the CRISP-DM framework, the hybrid model is implemented in stages, with the hybrid clustering method effectively reducing the root mean square error (RMSE) score, thereby improving model accuracy. To determine optimal clusters, the study conducted several iterations of clustering, measuring cluster distance using both Euclidean distance and Dynamic Time Warping (DTW). Systematic RMSE-based comparisons of predictive results were performed to identify the most accurate models, highlighting the significant influence of the optimizer choice on performance. Notably, the LSTM univariate results for the cluster 2 ISPU dataset from 2021 to 2023 consistently demonstrated the lowest RMSE scores, ranging from 2.91984 to 9.53943, indicating its effectiveness as the preferred model configuration. This hybrid approach presents a robust solution for improving air quality forecasting accuracy, offering valuable insights for mitigating air pollution in urban environments like Jakarta.