Prediksi Konsentrasi PM2.5 Resolusi 15 Menit di Kabupaten Brebes Menggunakan Transformer dan GEOS-CF NASA

Abstract

Polusi udara partikulat halus (PM2,5) merupakan ancaman serius bagi kesehatan masyarakat di Kabupaten Brebes, Jawa Tengah. Faktor penyumbang utamanya adalah emisi kendaraan di jalur Pantura, aktivitas industri perikanan, serta konsentrasi tinggi selama musim kemarau (Juni–November). Tidak adanya model peramalan sub-jam yang akurat menghambat pengembangan sistem peringatan dini yang efektif. Penelitian ini mengembangkan dan mengevaluasi model deep learning berbasis Transformer untuk memprediksi konsentrasi PM2,5 dengan resolusi waktu 15 menit. Data yang digunakan berasal dari NASA GEOS-CF (band PM25_RH35_GCC) yang diakses melalui Google Earth Engine menggunakan API Python. Dataset mencakup periode 1 Januari hingga 22 November 2025, menghasilkan 7.813 observasi per jam, yang kemudian diinterpolasi linear menjadi 31.249 titik data dengan resolusi 15 menit. Arsitektur Transformer terdiri dari 3 lapis enkoder, 4 kepala perhatian multi-head, dimensi embedding 128, dimensi feed-forward 256, panjang sekuen 60 timestep, dan augmentasi fitur menggunakan rerata bergulir (*rolling mean*, jendela = 3) dan beda pertama (*first difference*). Pelatihan dilakukan dengan TensorFlow-Keras, pengoptimal Adam, penjadwal peluruhan kosinus (*cosine decay scheduler*), dan fungsi kerugian Huber. Pembagian data dilakukan secara kronologis: 70% pelatihan, 30% validasi. Evaluasi pada set uji independen (16 Agustus–21 November 2025, 9.357 observasi atau 97 hari 11 jam 15 menit) menghasilkan MAE 0,7691 µg/m³, RMSE 1,2052 µg/m³, R² 0,9945, dan *Explained Variance Score* 0,9948. Model ini mampu menggambarkan variasi diurnal dan anomali musiman secara akurat, jauh melampaui model LSTM dan GTWR konvensional. Penelitian ini memberikan kontribusi signifikan di bidang Teknologi Informasi melalui kerangka kerja pengolahan *big data* satelit untuk aplikasi lingkungan.Fine particulate matter (PM2.5) air pollution poses a serious public health threat in Brebes Regency, Central Java. The main contributing factors are vehicle emissions on the Pantura route, fishing industry activities, and high concentrations during the dry season (June–November). The lack of an accurate sub-hourly forecast model hinders the development of an effective early warning system. This study develops and evaluates a Transformer-based deep learning model to predict PM2.5 concentrations with a 15-minute time resolution. The data used came from NASA GEOS-CF (PM25_RH35_GCC band) accessed through Google Earth Engine using the Python API. The dataset covered the period from 1 January to 22 November 2025, resulting in 7,813 observations per hour, which were then linearly interpolated into 31,249 data points with a resolution of 15 minutes. The Transformer architecture consists of 3 encoder layers, 4 multi-head attention heads, 128 embedding dimensions, 256 feed-forward dimensions, 60 timestep sequence length, and feature augmentation using rolling mean (window = 3) and first difference. Training was performed with TensorFlow-Keras, Adam optimiser, cosine decay scheduler, and Huber loss. Data division was chronological: 70% training, 30% validation. Evaluation on an independent test set (16 August–21 November 2025, 9,357 observations or 97 days 11 hours 15 minutes) resulted in MAE 0.7691 μg/m³, RMSE 1.2052 μg/m³, R² 0.9945, and Explained Variance Score 0.9948. The model is capable of accurately depicting diurnal variations and seasonal anomalies, far surpassing conventional LSTM and GTWR models. This research makes a significant contribution to the field of Information Technology through its satellite big data processing framework for environmental applications,