INTEGRASI PEMODELAN TERMODINAMIKA DENGAN MACHINE LEARNING UNTUK MENINGKATKAN EFISIENSI TERMAL PLTP DARAJAT
OPAN ROPANDI, Dr.-Ing. Ir. Sihana; Dr.Eng. Ir. Dwi Joko Suroso, S.T., M. Eng., IPP.
2026 | Tesis | MAGISTER TEKNIK FISIKA
The thermal efficiency of the Darajat Geothermal Power Plant (GPP) decreased from 21% at design condition to 18?fore the overhaul (OH). This decline is primarily caused by increased condenser pressure due to the decreasing effectiveness of the cooling system as the plant's operating age has exceeded thirty years. The decrease in thermal efficiency has worsened the heat rate to 19.940 kJ/kWh, far below the commissioning value of 17.713 kJ/kWh, making it difficult to maintain the generator power target of 55 MW. The thermodynamic modeling conducted by Sukra et al. (2023) describes the plant's operating conditions but remains limited as it is not integrated with real-time operational parameter data and lacks a predictive mechanism to generate optimal operational parameter recommendations.
Thermodynamic modeling was performed using Cycle-Tempo and MS Excel to simulate four operational scenarios: design, before OH, after OH, and after maintenance outage, with average validation errors of 0,38% for MS Excel and 1,1% for Cycle-Tempo. A machine learning model was developed using the AdaBoost algorithm with seven input features (MWgen, T17, T10, P10, m4, CV10, and P11) to generate real-time recommendations for the opening position of the CV24 control valve to maintain the target MWgen of 55 MW. Integration was accomplished through the PI Vision platform as the operator interface.
The AdaBoost model achieved an R² = 88,12%, RMSE = 0,27 MW, and MAPE = 0,38%. The recommendation system successfully predicted the optimal CV24 position under actual conditions at 72,05% to achieve the 55 MW generator power target. The main advantage of this system is the reduction in detection and response time to generator power decrease from 2–3 days to less than one hour.
The integration of thermodynamic modeling and machine learning successfully developed a real-time CV24 position recommendation system integrated with PI Vision, enabling operators to maintain the 55 MW generator power target with faster and more accurate response compared to existing procedures.
Kata Kunci : Efisiensi termal, AdaBoost, Condenser, Cycle-Tempo, machine learning.