Studi Identifikasi Faktor Hidrologi Dominan dan Proyeksi Sedimentasi Waduk Menggunakan Pendekatan Machine Learning (Studi Kasus: PLTA Mrica)
Bill Rufus Simbolon, Dr. Eng. Ir. Dwi Joko Suroso, S.T., M.Eng., IPP; Dr. Eng. Ir. Mohammad Kholid Ridwan, S.T., M.Sc., IPU
2026 | Tesis | MAGISTER TEKNIK FISIKA
Waduk PLTA Mrica (180 MW) telah kehilangan 73,0% kapasitas tampungan efektifnya akibat sedimentasi dengan laju 3,769 juta m³/tahun hingga 2025. Penelitian ini mengkombinasikan pemodelan Machine Learning (ML) dengan analisis feature importance, analisis hubungan inflow–sedimentasi melalui Sediment Rating Curve (SRC), dan forecasting probabilistik untuk menganalisis sedimentasi dan memproyeksikan volume sedimen masuk periode 2026–2030.
Data operasional harian PLTA Mrica dan NASA Power periode 2015–2025 (3.998 observasi) digunakan untuk melatih tiga model ML — Linier Regression (LR), Random Forest (RF), dan Support Vector Regression (SVR) — dengan skema pelatihan 2015–2020 validasi 2021-2022 dan pengujian 2023–2025. RF (R² = 0,9566) unggul dibanding SVR (R² = 0,9491) dan LR (R² = 0,8746), dengan inflow terbukti sebagai variabel dominan (82,4% rata-rata tiga metode feature importance). Hubungan matematis antara inflow dan sedimen selanjutnya dianalisis menggunakan Sediment Rating Curve (SRC), menghasilkan persamaan Qs = 1,6580 × Qw¹·?¹?? (r = 1,000, p < 0>Prophet harian dengan Annual Scale Factor berbasis variasi hidrologi historis, menghasilkan MAE = 10,58 m³/s, MAPE = 17,68%, dan Coverage P10–P90 = 97,2%. Proyeksi P50 kumulatif mencapai 14,567 juta m³ (2026–2030), dengan active storage turun dari 12,49 menjadi 8,99 juta m³ pada 2030.
Mrica Hydroelectric Power Plant (HEPP) reservoir (180 MW) has lost 73.0% of its effective storage capacity due to sedimentation at a rate of 3.769 million m³/year through 2025. This study combines Machine Learning (ML) modeling with feature importance analysis, inflow–sedimentation relationship analysis through the Sediment Rating Curve (SRC), and probabilistic forecasting to analyze sedimentation and project incoming sediment volume for 2026–2030.
Daily operational data from Mrica HEPP and NASA Power for 2015–2025 (3,998 observations) were used to train three ML models – Linear Regression (LR), Random Forest (RF), and Support Vector Regression (SVR) – with training 2015–2020, validation 2021–2022, and testing 2023–2025. RF (R² = 0.9566) outperformed SVR (R² = 0.9491) and LR (R² = 0.8746), with inflow identified as the dominant variable (82.4% average across three feature importance methods). The mathematical relationship between inflow and sediment was further analyzed using the Sediment Rating Curve (SRC), yielding Qs = 1.6580 × Qw¹·?¹?? (r = 1.000, p < 0 xss=removed xss=removed xss=removed>
Kata Kunci : sedimentasi waduk, machine learning, feature importance, forecasting Prophet, PLTA Mrica