Electronic Theses and Dissertation

Pencemaran udara akibat partikulat halus (PM2.5) merupakan risiko lingkungan dan kesehatan masyarakat yang signifikan di kota besar seperti Jakarta. Dinamika PM2.5 dipengaruhi oleh kondisi meteorologi dengan interaksi nonlinier yang kompleks, sehingga sulit dijelaskan menggunakan pendekatan statistik linier konvensional. Penelitian ini bertujuan untuk menentukan kinerja algoritma machine learning terbaik dalam memprediksi konsentrasi PM2.5, menentukan variabel meteorologi utama yang berpengaruh, serta mengidentifikasi hubungan kausal nonlinier antara variabel meteorologi dan PM2.5 di Provinsi DKI Jakarta. Data yang digunakan berupa data harian periode 2021-2025 dengan total 1.826 observasi dan tujuh variabel, yaitu PM2.5, temperatur, titik embun, curah hujan, kelembapan relatif, tekanan udara, dan kecepatan angin. Nilai PM2.5 yang hilang diimputasi menggunakan metode musiman berbasis Day-of-Year. Pendekatan machine learning dilakukan menggunakan algoritma Random Forest, XGBoost, dan LightGBM, dengan evaluasi menggunakan RMSE, MAE, dan WAPE. Interpretasi algoritma terbaik dilakukan menggunakan SHAP, dan hubungan kausal nonlinier diidentifikasi menggunakan Convergent Cross Mapping (CCM). Hasil penelitian menunjukkan bahwa Random Forest memiliki kinerja terbaik (RMSE = 30,69; MAE = 25,05; WAPE = 27,62). Hasil SHAP mengidentifikasi kelembapan sebagai variabel paling berpengaruh, diikuti curah hujan dan kecepatan angin. Hasil CCM menunjukkan hubungan kausal dua arah antara PM2.5 dan sebagian besar variabel meteorologi, sementara tekanan udara hanya menunjukkan hubungan satu arah. Temuan ini menegaskan bahwa dinamika PM2.5 dikendalikan oleh interaksi atmosfer nonlinier yang kompleks dan pentingnya integrasi machine learning serta analisis kausal dalam pengelolaan kualitas udara perkotaan.

Air pollution caused by fine particulate matter (PM2.5) poses a significant environmental and public health risk in large cities such as Jakarta. The dynamics of PM2.5 are strongly influenced by meteorological conditions and involve complex nonlinear interactions, making them difficult to explain using conventional linear statistical approaches. This study aims to evaluate the performance of machine learning algorithms in predicting PM2.5 concentrations, determine the key meteorological variables influencing PM2.5, and identify nonlinear causal relationships between meteorological factors and PM2.5 in DKI Jakarta. The study utilizes daily data from 2021 to 2025, comprising 1,826 observations and seven variables: PM2.5, temperature, dew point, rainfall, relative humidity, air pressure, and wind speed. Missing PM2.5 values were imputed using a seasonal Day-of-Year approach. Machine learning approaches were implemented using Random Forest, XGBoost, and LightGBM, evaluated with RMSE, MAE, and WAPE. Model interpretability was examined using SHAP, and nonlinear causal relationships were identified using Convergent Cross Mapping (CCM). The results show that Random Forest achieved the best predictive performance (RMSE = 30.69; MAE = 25.05; WAPE = 27.62). SHAP results indicate that relative humidity is the most influential variable, followed by rainfall and wind speed. CCM results reveal bidirectional causal relationships between PM2.5 and most meteorological variables, while air pressure exhibits a unidirectional causal effect. These findings highlight that the dynamics of PM2.5 are governed by complex, nonlinear atmospheric interactions and underscore the importance of integrating machine learning and causal approaches in urban air quality management