Penerapan Model Hybrid CNN-Residual BiLSTM dengan Optimasi Optuna Untuk Prediksi Time-Series: Studi Kasus Peramalan Degradasi Pressure Compressor Discharge Pada Turbin Gas

Pratama, Andika (2025) Penerapan Model Hybrid CNN-Residual BiLSTM dengan Optimasi Optuna Untuk Prediksi Time-Series: Studi Kasus Peramalan Degradasi Pressure Compressor Discharge Pada Turbin Gas. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6032231200-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Prediksi akurat degradasi Compressor Discharge Pressure (PCD) krusial untuk keandalan operasional dan optimalisasi pemeliharaan turbin gas di Central Processing Plant. Penelitian ini mengintroduksi kerangka prediktif canggih berbasis arsitektur Hybrid Convolutional Neural Network-Residual Bidirectional Long Short-Term Memory (Hybrid CNN-Residual BiLSTM) yang dioptimasi sistematis menggunakan Optuna (“Hybrid Optuna”). Model ini dirancang untuk menangkap dependensi temporal jangka panjang dan pola non-linier kompleks dari 24 fitur data operasional turbin gas multivariat. Evaluasi komprehensif menunjukkan Residual Bi-LSTM standar sebagai baseline terbaik (MAE = 0,0708 bar, R² = 0,9228). Namun, “Hybrid Optuna” secara signifikan melampauinya, mencapai MAE 0,0298, RMSE 0,0611, dan R² 0,9601 di mana pengurangan MAE 58%. Uji Wilcoxon Signed-Rank dan Cliff’s Delta mengkonfirmasi keunggulan statistik (p < 0,05) dan dampak praktis besar. Implikasi praktis terwujud melalui integrasi peramalan PCD multi-horizon (1–24 jam) ke dalam logika keputusan dinamis untuk pemeliharaan berbasis kondisi dan analisis skenario, dengan estimasi penghematan USD 5.000–10.000 per unit per bulan. Prototipe dashboard interaktif Streamlit memvisualisasikan tren dan rekomendasi. Rekomendasi meliputi pemantauan real-time, transfer learning, dan integrasi XAI. Metodologi ini menjadi acuan pengembangan Sistem Pendukung Keputusan Adaptif untuk keberlanjutan infrastruktur energi.
======================================================================================================================================
Accurate PCD degradation prediction is vital for gas turbine reliability and optimized maintenance in critical facilities. This research presents an advanced predictive framework with a Hybrid CNN-Residual BiLSTM architecture, optimized via Optuna (“Hybrid Optuna”). The model captures long-range temporal dependencies and non-Linier patterns in 24-feature multivariate gas turbine data. A standard Residual Bi-LSTM was the best baseline (MAE=0.0708, R²=0.9228). “Hybrid Optuna,” however, significantly outperformed this, achieving MAE 0.0298, RMSE 0.0611, and R² 0.9601—a 58% MAE reduction. Statistical validation (Wilcoxon test, Cliff’s Delta) confirmed its significant (p<0.05) and large practical impact. Practical implications involve integrating multi-horizon PCD forecasts (1–24 hours) into dynamic decision logic for Condition-Based Maintenance, estimating USD 5,000–10,000 monthly per-unit savings. An interactive Streamlit dashboard visualizes trends and recommendations. Future work includes real-time monitoring and transfer learning. This methodology benchmarks data-driven gas turbine maintenance, supporting Adaptive DSS for sustainable energy infrastructure.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Compressor Discharge Pressure (PCD), Residual Bi-LSTM, Time-Series Analytics, Predictive Maintenance, Hyperparameter Optimization, Decision Support System, Gas Turbine,Sistem Pendukung Keputusan, Turbin Gas
Subjects:	H Social Sciences > HD Industries. Land use. Labor > HD9490.A2 Essences and essential oils industry.
Divisions:	Interdisciplinary School of Management and Technology (SIMT) > 61101-Master of Technology Management (MMT)
Depositing User:	Andika Pratama
Date Deposited:	29 Jul 2025 10:30
Last Modified:	29 Jul 2025 10:32
URI:	http://repository.its.ac.id/id/eprint/122983

Actions (login required)

View Item