Improving the Forecast Accuracy of PM2.5 Using SETAR-Tree Method: Case Study in Jakarta, Indonesia

Air pollution in Jakarta, one of the most polluted cities globally, has reached critical levels, with PM2.5 concentrations exceeding the WHO guidelines and posing significant health risks. Accurate forecasting of PM2.5 is crucial for effective air quality management and public health interventions. PM2.5 exhibits significant nonlinear fluctuations; thus, this study employed two machine learning approaches: self-exciting threshold autoregressive tree (SETAR-Tree) and long short-term memory (LSTM). The SETAR-Tree model integrates regime-switching capabilities with decision tree principles to capture nonlinear patterns, while LSTM models long-term dependencies in time-series data. The results showed that: (1) SETAR-Tree outperformed LSTM, achieving lower RMSE (0.1691 in-sample, 0.2159 out-sample) and MAPE (2.83% in-sample, 2.98% out-sample) compared to LSTM’s RMSE (0.2038 in-sample, 0.2399 out-sample) and MAPE (3.48% in-sample, 4.05% out-sample); (2) SETAR-Tree demonstrated better responsiveness to sudden regime changes, capturing complex pollution patterns influenced by meteorological and anthropogenic factors; (3) PM2.5 in Jakarta often exceeds the WHO limits, highlighting this study’s importance in supporting strategic planning and providing an early warning system to reduce outdoor activity during extreme pollution.