Machine Learning-Based Fuel Prediction Method Comparison on Bulk Carrier Ship

Adilah, Akbar (2026) Machine Learning-Based Fuel Prediction Method Comparison on Bulk Carrier Ship. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 04211941000009-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Strict decarbonization mandates like the Carbon Intensity Indicator (CII) require precise fuel management, yet traditional hydrodynamic models often fail when applied to the noisy data found in manual "Noon Reports". This research addresses that reliability gap by developing a hybrid prediction framework for "Bulk Carrier Ship X." Instead of training on raw logs, the study applied the STEAM 2 physics-based method to filter out physically impossible data points, ensuring the machine learning algorithms learned only from valid operational states. Three ensemble models Random Forest, XGBoost, and CatBoost were benchmarked using 10-Fold Cross-Validation to ensure stability rather than just raw accuracy. Results highlighted a critical trade-off: while XGBoost achieved the highest raw accuracy (R^2 0.9027), it proved unstable during cross-validation. CatBoost emerged as the superior operational solution, delivering exceptional stability (CV R^2 0.9323) and a daily error margin of just 0.58 tons. Validated by its ability to prioritize efficiency metrics over simple speed, the winning model was deployed as a verified operational calculator for the vessel's crew.
===================================================================================================================================
Regulasi dekarbonisasi ketat seperti CII menuntut manajemen bahan bakar yang presisi, namun model tradisional sering gagal menghadapi data "Noon Reports" yang penuh noise. Penelitian ini mengatasi ketidakpastian tersebut melalui kerangka kerja prediksi hibrida untuk "Kapal Bulk Carrier X". Alih-alih menggunakan data mentah, studi ini menerapkan metode berbasis fisika STEAM 2 untuk memfilter data yang mustahil secara teknis, memastikan algoritma hanya mempelajari kondisi operasional yang valid. Tiga model ensemble Random Forest, XGBoost, dan CatBoost diuji menggunakan 10-Fold Cross-Validation untuk memastikan stabilitas model. Hasil penelitian menunjukkan sebuah trade-off krusial: meskipun XGBoost mencatat akurasi mentah tertinggi (R^2 0,9027), model ini terbukti tidak stabil saat validasi silang. CatBoost terbukti sebagai solusi terbaik dengan stabilitas tinggi (CV R^2 0,9323) dan margin kesalahan harian hanya 0,58 ton. Tervalidasi oleh kemampuannya memprioritaskan metrik efisiensi di atas sekadar kecepatan, model ini diimplementasikan sebagai kalkulator operasional bagi kru kapal.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Fuel Prediction, Machine Learning, CatBoost, STEAM 2, Bulk Carrier.Prediksi Bahan Bakar, Machine Learning, CatBoost, STEAM 2, Bulk Carrier.
Subjects:	V Naval Science > VM Naval architecture. Shipbuilding. Marine engineering > VM276.A1 Fuel (Including supplies, costs, etc.)
Divisions:	Faculty of Marine Technology (MARTECH) > Marine Engineering > 36202-(S1) Undergraduate Thesis
Depositing User:	Akbar Adilah
Date Deposited:	05 Feb 2026 04:29
Last Modified:	05 Feb 2026 04:29
URI:	http://repository.its.ac.id/id/eprint/132168

Actions (login required)

View Item