Implementasi Long Short-Term Memory untuk Prediksi Tekanan Vakum sebagai Dasar Predictive Maintenance serta Monitoring Efektivitas pada Kondenser

., Norhalimah (2026) Implementasi Long Short-Term Memory untuk Prediksi Tekanan Vakum sebagai Dasar Predictive Maintenance serta Monitoring Efektivitas pada Kondenser. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5009221019_Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Kemampuan kondenser pada PLTP untuk mempertahankan kondisi vakum menjadi salah satu faktor yang memengaruhi proses ekspansi uap pada turbin dan performa pembangkit. Penurunan kinerja kondenser dapat menyebabkan peningkatan tekanan vakum yang berdampak pada kenaikan back pressure turbin dan penurunan daya keluaran pembangkit. Pendekatan berbasis Long Short-Term Memory (LSTM) digunakan untuk menghasilkan prediksi tekanan vakum kondenser tiga jam ke depan yang selanjutnya dimanfaatkan sebagai dasar predictive maintenance serta diintegrasikan dengan sistem monitoring efektivitas kondenser dan early warning system (EWS). Data operasional yang digunakan berasal dari PLTP selama periode Januari 2023 hingga Desember 2024 dengan interval pencatatan setiap satu jam. Variabel input yang digunakan meliputi laju aliran uap (m ̇_Steam), laju aliran air pendingin (m ̇_Water), temperatur air pendingin masuk (T_ci), temperatur air pendingin keluar (T_co), dan temperatur exhaust turbin (T_hi). Tahapan penelitian meliputi pre-processing data, normalisasi menggunakan metode Min-Max Scaling, pelatihan model LSTM, evaluasi performa menggunakan RMSE, MAE, NRMSE dan NMAE, serta implementasi sistem monitoring menggunakan MATLAB App Designer. Hasil penelitian menunjukkan bahwa konfigurasi terbaik diperoleh pada model LSTM dengan 64 neuron, learning rate 0,0001, dan dropout sebesar 0 yang menghasilkan nilai RMSE sebesar 3,648 mbar dan MAE sebesar 2,783 mbar, NRMSE sebesar 0,052, dan NMAE sebesar 0,040, yang menunjukkan bahwa model memiliki tingkat kesalahan prediksi yang sangat rendah berdasarkan kriteria evaluasi model prediksi. Sistem monitoring yang dikembangkan mampu menampilkan kondisi operasional kondenser, hasil prediksi tekanan vakum, tren efektivitas, serta klasifikasi kondisi operasi ke dalam kategori normal, warning, dan critical. Hasil tersebut menunjukkan bahwa sistem yang dikembangkan mampu mendukung identifikasi dini penurunan kinerja kondenser dan pengambilan keputusan pemeliharaan yang lebih proaktif. Selain itu, penelitian ini mendukung pencapaian SDG 9 (Industry, Innovation and Infrastructure) melalui penerapan teknologi berbasis machine learning dan condition monitoring pada sistem pembangkit.
================================================================================================================================
The ability of condensers in geothermal power plants to maintain vacuum conditions is one of the factors influencing steam expansion in turbines and overall plant performance. A decline in condenser performance may lead to an increase in vacuum pressure, resulting in higher turbine back pressure and reduced power output. A Long Short-Term Memory (LSTM) based approach was employed to generate condenser vacuum pressure predictions three hours ahead, which were subsequently utilized as a basis for predictive maintenance and integrated with a condenser effectiveness monitoring system and an early warning system (EWS). The operational data used in this study were collected from a geothermal power plant during the period from January 2023 to December 2024 with an hourly sampling interval. The input variables consisted of steam flow rate (m ̇_Steam), cooling water flow rate (m ̇_Water), cooling water inlet temperature (T_ci), cooling water outlet temperature (T_co), and turbine exhaust temperature (T_hi). The research procedure included data preprocessing, normalization using the Min-Max Scaling method, LSTM model training, performance evaluation using RMSE, MAE, NRMSE and NMAE, and implementation of the monitoring system using MATLAB App Designer. The results showed that the best model configuration was achieved using 64 neurons, a learning rate of 0.0001, and a dropout value of 0, resulting in an RMSE of 3,648 mbar, an MAE of 2,783 mbar, an NRMSE of 0,052, and an NMAE of 0,040, indicating that the model achieved a very low prediction error based on standard model evaluation criteria. The developed monitoring system was capable of displaying condenser operational conditions, vacuum pressure predictions, effectiveness trends, and operational status classifications into normal, warning, and critical categories. These findings indicate that the developed system can support early identification of condenser performance degradation and facilitate more proactive maintenance decision-making. Furthermore, this study contributes to the achievement of SDG 9 (Industry, Innovation and Infrastructure) through the implementation of machine learning and condition monitoring technologies in power generation systems.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Early Warning System, Efektivitas Kondenser, Long Short-Term Memory, Predictive Maintenance, Tekanan Vakum, Condenser Effectiveness, Vacuum Pressure
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Industrial Technology > Physics Engineering > 30201-(S1) Undergraduate Thesis
Depositing User:	Norhalimah Norhalimah
Date Deposited:	01 Aug 2026 06:50
Last Modified:	01 Aug 2026 06:52
URI:	http://repository.its.ac.id/id/eprint/141122

Actions (login required)

View Item