Prediksi Konsentrasi PM2.5 Berbasis Light Gradient Boosting Machine Dengan Feature Engineering Temporal Di Kawasan Industri Benowo, Surabaya

Bangun, Meidiantha Nataniel (2026) Prediksi Konsentrasi PM2.5 Berbasis Light Gradient Boosting Machine Dengan Feature Engineering Temporal Di Kawasan Industri Benowo, Surabaya. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5003221117-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Data sensor kualitas udara beresolusi 15 menit dapat memuat rangkaian missing value yang panjang. Kondisi ini memutus riwayat yang diperlukan untuk membentuk fitur lag dan dapat mencampurkan error rekonstruksi dengan error prediksi. Penelitian ini bertujuan memprediksi konsentrasi PM₂.₅ 15 menit ke depan di Benowo, Surabaya, sekaligus mengevaluasi imputasi celah panjang serta kontribusi temperatur dan kelembapan. Data AirVisual Pro periode 2022–2025 diselaraskan ke grid 15 menit dan diperiksa terhadap anomali sensor. Data tersebut memuat satu celah sepanjang 131 hari. Celah pendek diisi menggunakan interpolasi waktu, sedangkan median diurnal, day donor terarah, dan transition smoothed dibandingkan untuk celah panjang melalui artificial masking pada data training. Deret hasil imputasi digunakan untuk membentuk Model 1 dengan riwayat PM₂.₅ dan fitur waktu serta Model 2 yang menambahkan riwayat meteorologi. Konfigurasi dipilih melalui validasi temporal lima fold dengan expanding window, kemudian dinilai pada target testing terukur. Median diurnal menghasilkan error rekonstruksi terendah, tetapi Model 1 dengan imputasi transition smoothed memperoleh RMSE validasi terendah. Pada testing, model menghasilkan RMSE 10,836 µg/m³ dan R² 0,873 serta menurunkan RMSE 24,74% dibandingkan persistence. Penambahan fitur meteorologi tidak memperbaiki hasil validasi, sedangkan PM25_lag_1 menjadi fitur paling berpengaruh. Model masih mengalami underprediction pada konsentrasi tinggi dan kinerjanya menurun pada prediksi rekursif 24 jam. Hasil penelitian terbatas pada satu sensor dan prediksi satu langkah ketika PM₂.₅ terbaru tersedia. Nilai pada celah panjang merupakan hasil rekonstruksi, bukan pengamatan aktual.
===================================================================================================================================
Air-quality sensor records at 15-minute resolution may contain long sequences of missing observations. These gaps interrupt the history required for lag features and can confound reconstruction error with prediction error. This study aimed to predict PM₂.₅ concentrations 15 minutes ahead in Benowo, Surabaya, while evaluating long-gap imputation and the contribution of temperature and humidity. AirVisual Pro data from 2022 to 2025 were aligned to a 15-minute grid and checked for sensor anomalies. The data contained one continuous 131-day gap. Short gaps were filled using time interpolation, whereas median diurnal imputation, directed day donor, and transition smoothed were compared for longer gaps through artificial masking on the training data. The reconstructed series were used to build Model 1 from historical PM₂.₅ and time features and Model 2 by adding meteorological histories. Configurations were selected through five-fold expanding-window temporal validation and assessed on observed testing targets. Median diurnal imputation produced the lowest reconstruction error, but Model 1 using transition smoothed achieved the lowest validation RMSE. On the testing data, the model obtained an RMSE of 10.836 µg/m³ and an R² of 0.873, reducing RMSE by 24.74% relative to persistence. Meteorological features did not improve validation performance, while PM25_lag_1 was the most influential feature. The model still underpredicted high concentrations, and its performance deteriorated under 24-hour recursive prediction. The findings are limited to one sensor and one-step prediction when the latest PM₂.₅ observation is available. Values within the long gap are reconstructions rather than actual observations.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	LightGBM, Missing Value Imputation, PM₂.₅, Prediksi 15 Menit, Temporal Feature Engineering, 15-Minute Prediction, LightGBM, Missing Value Imputation, PM₂.₅, Temporal Feature Engineering
Subjects:	Q Science > QA Mathematics > QA276 Mathematical statistics. Time-series analysis. Failure time data analysis. Survival analysis (Biometry)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49201-(S1) Undergraduate Thesis
Depositing User:	Meidiantha Nataniel Bangun
Date Deposited:	04 Aug 2026 01:06
Last Modified:	04 Aug 2026 01:06
URI:	http://repository.its.ac.id/id/eprint/140749

Actions (login required)

View Item