Irham, Ainal (2026) Power Plant Efficiency Prediction Using MrMiMaxG Feature Engineering and Consensus Voting Mechanism. Masters thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 6025231006-Master_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
6025231006-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (14MB) | Request a copy |
Abstract
Penilaian kinerja Pembangkit Listrik Tenaga Uap (PLTU) seperti Boiler Efficiency (BE) dan Net Plant Heat Rate (NPHR) melalui pengujian fisik membutuhkan biaya tinggi dan memiliki risiko terhadap stabilitas transmisi. Pendekatan machine learning menawarkan solusi pemantauan kontinu, namun dihadapkan pada tantangan data sensor operasional dan emisi yang berdimensi tinggi, bising, serta memiliki multikolinearitas ekstrem. Selain itu, metrik evaluasi konvensional kurang memadai untuk menangkap risiko kegagalan ekstrem pada sistem pembangkit yang kritis keselamatan. Penelitian ini mengusulkan kerangka kerja analitik prediktif menggunakan protokol seleksi fitur ensemble berbasis konsensus yang dinamakan MrMiMaxG (mengintegrasikan MI, mRMR, RFE, dan GA). Kerangka kerja ini mengintegrasikan data emisi gas buang sebagai fitur laten dan memperkenalkan metrik sadar risiko Weighted Penalty Performance Score (WPPS). Model prediksi XGBoost dan Stacking Regressor dievaluasi menggunakan data aktual dari PLTU supercritical dan divalidasi dengan dataset eksternal UCI. Hasil penelitian menunjukkan bahwa protokol MrMiMaxG mampu mereduksi dimensi fitur secara drastis hingga lebih dari 70% dan berhasil mengatasi multikolinearitas ekstrem dengan menekan nilai Variance Inflation Factor (VIF) di bawah 10. Penggunaan Stacking Regressor dengan fitur hasil seleksi MrMiMaxG mampu mempertahankan tingkat akurasi yang tinggi, mencatatkan nilai koefisien determinasi (R2) di atas 0,99 dan skor WPPS melebihi 90%. Selain itu, reduksi dimensi fitur mempercepat waktu komputasi pelatihan secara signifikan dari 160 detik menjadi di bawah 20 detik, sehingga sangat efisien untuk implementasi real-time. Metode konsensus yang diusulkan terbukti lebih stabil dan superior dibandingkan dengan metode Exploratory Factor Analysis (EFA) maupun seleksi fitur tunggal. ====================================================================================================================================
Performance evaluation of Coal-Fired Power Plants (CFPP), such as Boiler Efficiency (BE) and Net Plant Heat Rate (NPHR), through physical testing is costly and poses risks to transmission stability. Machine learning approaches offer continuous monitoring solutions but face significant challenges with high-dimensional, noisy, and extremely multicollinear operational and emission sensor data. Furthermore, conventional evaluation metrics are inadequate for capturing extreme failure risks in safety critical power generation systems. This research proposes a predictive analytics framework utilizing a consensus-based ensemble feature selection protocol named MrMiMaxG (integrating MI, mRMR, RFE, and GA). This framework incorporates exhaust gas emission data as latent features and introduces a risk-aware evaluation metric, the Weighted Penalty Performance Score (WPPS). XGBoost and Stacking Regressor predictive models were evaluated using actual data from a supercritical CFPP and validated with an external UCI dataset. The results indicate that the MrMiMaxG protocol successfully reduced feature dimensions drastically by more than 70% and resolved extreme multicollinearity by suppressing Variance Inflation Factor (VIF) values below 10. Employing Stacking Regressors with MrMiMaxG-selected features maintained a high level of accuracy, achieving a coefficient of determination (R2) above 0.99 and a WPPS score exceeding 90%. Additionally, feature reduction significantly accelerated computational training time from 160 seconds to under 20 seconds, proving highly efficient for real-time implementation. The proposed consensus method proved to be more stable and superior compared to Exploratory Factor Analysis (EFA) and other single feature selection methods.
Actions (login required)
 |
View Item |