Peramalan Suhu Berbasis CNN-BiLSTM dengan Attention Mechanism pada Data Cuaca AWOS Tarakan

Rachmawati, Putri Meyliya and Handika, Allen Keyo (2025) Peramalan Suhu Berbasis CNN-BiLSTM dengan Attention Mechanism pada Data Cuaca AWOS Tarakan. Project Report. [s.n.], [s.l.]. (Unpublished)

Text 5025221062_5025221298-Project_Report.pdf - Accepted Version Restricted to Repository staff only Download (1MB) | Request a copy

Abstract

Prakiraan cuaca, khususnya prediksi suhu udara, krusial untuk pertanian, mitigasi bencana, dan perencanaan energi. Namun, kompleksitas data time-series meteorologi menimbulkan tantangan akurasi. Penelitian ini mengusulkan model hybrid CNN-BiLSTM-Attention untuk meningkatkan prakiraan suhu jangka pendek hingga panjang. Dataset dari AWOS Stasiun Cuaca Tarakan (11.104 baris, Januari 2023–April 2024) diproses dengan median imputation untuk missing values, deteksi outlier via IQR dan rolling median, serta normalisasi. Model dievaluasi pada horizon 96, 192, 336, dan 720 jam, dibandingkan baseline (LSTM, BiLSTM, MLP, Transformer, Informer) menggunakan MAE, MSE, dan CORR. Hasil: CNN-BiLSTM-Attention unggul, dengan reduksi MSE rata-rata 2,68% dan peningkatan CORR 0,25% terhadap BiLSTM, meskipun penurunan ringan pada 192 jam. Model efektif menangkap pola lokal-global.
======================================================================================================================================
Weather forecasts, especially air temperature predictions, are crucial for agriculture, disaster mitigation, and energy planning. However, the complexity of meteorological time-series data poses accuracy challenges. This study proposes a hybrid CNN-BiLSTM-Attention model to improve short- to long-term temperature forecasts. The dataset from the Tarakan Weather Station AWOS (11,104 rows, January 2023–April 2024) was processed with median imputation for missing values, outlier detection via IQR and rolling median, and normalization. The model was evaluated over 96, 192, 336, and 720-hour horizons, compared to baselines (LSTM, BiLSTM, MLP, Transformer, Informer) using MAE, MSE, and CORR. Results: CNN-BiLSTM-Attention outperformed, with an average MSE reduction of 2.68% and a CORR improvement of 0.25% over BiLSTM, despite a slight decline at 192 hours. The model effectively captured local-global patterns.

Item Type:	Monograph (Project Report)
Uncontrolled Keywords:	Prakiraan Suhu, Deep Learning, CNN-BiLSTM, Mekanisme Attention, dan Time-Series. Temperature Forecast, Deep Learning, CNN-BiLSTM, Attention Mechanism, and Time Series.
Subjects:	T Technology > T Technology (General) > T174 Technological forecasting
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Putri Meyliya Rachmawati
Date Deposited:	05 Nov 2025 00:49
Last Modified:	05 Nov 2025 00:49
URI:	http://repository.its.ac.id/id/eprint/128732

Actions (login required)

View Item