Pemodelan Hybrid GRU-XGBOOST Dengan PSO-Hyperparameter Optimization dalam Peramalan Konsentrasi Polusi Udara PM2.5

Athaya, Pudja Billah (2026) Pemodelan Hybrid GRU-XGBOOST Dengan PSO-Hyperparameter Optimization dalam Peramalan Konsentrasi Polusi Udara PM2.5. Diploma thesis, Institut Teknologi Sepuluh Nopember.

Text 5006221079_Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Pencemaran polusi udara merupakan salah satu permasalahan lingkungan yang berdampak langsung terhadap kualitas hidup dan kesehatan masyarakat. Salah satu komponen polusi udara yang paling berbahaya adalah particulate matter (PM), khususnya PM2.5, yaitu partikel berdiameter sangat kecil yang mampu menembus saluran pernapasan hingga ke paru-paru dan aliran darah. Di Indonesia, permasalahan polusi udara merupakan hal serius, terutama di wilayah perkotaan dengan aktivitas industri dan transportasi yang tinggi. Salah satu upaya penting dalam penanganan polusi udara adalah melalui peramalan konsentrasi PM2.5 yang dapat memberikan antisipasi kualitas udara di masa depan dan mendukung pengambilan keputusan dalam perencanaan mitigasi risiko serta kebijakan lingkungan. Data PM2.5 memiliki bentuk yang kompleks, nonlinier, dan memiliki ketergantungan jangka panjang. Pada penelitian ini, dilakukan pemodelan dan peramalan terhadap konsentrasi PM2.5 menggunakan hybrid model GRU-XGBoost. Dinamika pemodelan akan dilakukan dengan deep learning berbasis Gated Recurrent Unit (GRU), yang dirancang untuk menangkap dinamika temporal dan dependensi jangka panjang pada data deret waktu. Kemudian, diterapkan strategi residual error correction menggunakan algoritma XGBoost untuk meningkatkan akurasi hasil peramalan. Hasil penelitian menunjukkan bahwa model hybrid GRU-XGBoost berhasil meningkatkan akurasi prediksi dengan nilai Mean Absolute Error (6,39011) dan Root Mean Squared Error (22,065%) yang lebih rendah dari single model GRU ataupun XGBoost. Model GRU cenderung memberikan prediksi yang lebih smooth, sedangkan model XGBoost menghasilkan prediksi yang lebih fluktuatif. Oleh karena itu, model hybrid GRU-XGBoost mampu menyeimbangkan penangkapan pola dependensi temporal dan non-linear yang lebih baik dengan hasil prediksi yang lebih adaptif dalam menangkap pola data fluktuatif PM2.5. Hasil peramalan menunjukkan bahwa nilai konsentrasi variabel PM2.5 di Desa Sukamahi, Kabupaten Bekasi menunjukkan kondisi yang moderat hingga tidak sehat pada periode Februari 2026. Penanganan dan mitigasi risiko perlu menjadi perhatian, mengingat nilai tersebut mungkin akan lebih ekstrem pada musim kemarau. Update berkala terkait pemodelan dan peramalan perlu dilakukan untuk mengantisipasi risiko di masa depan.
=============================================================================================================================
Air pollution is one of the major environmental problems that directly impacts public health and quality of life. One of the most dangerous components of air pollution is particulate matter (PM), particularly PM2.5, which consists of very fine particles capable of penetrating the respiratory tract into the lungs and bloodstream. In Indonesia, air pollution is a serious issue, especially in urban areas with high industrial and transportation activities. One important effort in addressing air pollution is through PM2.5 concentration forecasting. Accurate forecasting can provide information on future air quality conditions, thereby supporting decision-making in risk mitigation planning and environmental policies. PM2.5 data exhibits a complex time series form that is nonlinear and possesses long-term dependencies. Therefore, machine learning, deep learning, and hybrid models are frequently used in PM2.5 modeling and forecasting. In this research, modeling and forecasting of PM2.5 concentrations are conducted using a hybrid GRU-XGBoost model. The modeling dynamics will be performed with deep learning based on Gated Recurrent Unit (GRU), designed to capture temporal dynamics and long-term dependencies in time series data. Subsequently, a residual error correction strategy is applied using the XGBoost algorithm to enhance forecasting accuracy. The results showed that the hybrid GRU-XGBoost model successfully improved prediction accuracy, with lower Mean Absolute Error (6.39011) and Root Mean Squared Error (22.065%) compared to the single GRU or XGBoost models. The GRU model tended to produce smoother predictions, while the XGBoost model generated more fluctuating predictions. Therefore, the hybrid GRU-XGBoost model was able to balance the capture of temporal dependency patterns and nonlinear relationships more effectively, producing predictions that were more adaptive in capturing the fluctuating PM2.5 data patterns. The forecasting results indicate that the concentration values for the PM2.5 variable in Sukamahi Village, Bekasi Regency, are expected to be moderate to unhealthy during February 2026. Risk management and mitigation should therefore receive serious attention, considering that the values may become more extreme during the dry season. Regular updates to the modeling and forecasting process are necessary to anticipate future risks.

Item Type:	Thesis (Diploma)
Uncontrolled Keywords:	Deret Waktu, Gated Recurrent Unit (GRU), Koreksi Residual Error, Model Hybrid, Polusi Udara, XGBoost
Subjects:	G Geography. Anthropology. Recreation > GE Environmental Sciences > GE300 Environmental management Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > Q Science (General) > Q337.3 Swarm intelligence Q Science > QA Mathematics > QA276 Mathematical statistics. Time-series analysis. Failure time data analysis. Survival analysis (Biometry) Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Mathematics, Computation, and Data Science > Actuaria > 94203-(S1) Undergraduate Thesis
Depositing User:	Pudja Billah Athaya
Date Deposited:	17 Jul 2026 08:21
Last Modified:	17 Jul 2026 08:21
URI:	http://repository.its.ac.id/id/eprint/135369

Actions (login required)

View Item