Pengaruh Metode Optimasi Terhadap RNN, LSTM, Dan Gru Pada Prediksi Data Time Series Dari Persamaan Diferensial

Alamsyah, Melvin (2026) Pengaruh Metode Optimasi Terhadap RNN, LSTM, Dan Gru Pada Prediksi Data Time Series Dari Persamaan Diferensial. Other thesis, Instittut Teknologi Sepuluh Nopember.

Text 5002221020-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Penelitian ini mengkaji pengaruh delapan metode optimasi terhadap kinerja tiga arsitektur jaringan saraf berulang, yaitu Recurrent Neural Network(RNN), Long Short Term Memory (LSTM), dan Gated Recurrent Unit(GRU), dalam memprediksi data deret waktu yang dihasilkan dari sistem persamaan diferensial. Data time series dibangun melalui simulasi numerik dari tiga model dinamika nonlinier: model SIR dan sistem Lotka–Volterra menggunakan metode Euler, sedangkan sistem Lorenz menggunakan metode Runge–Kutta orde empat (RK4) karena karakteristik chaos-nya yang memerlukan akurasi integrasi tinggi. Fokus utama penelitian ini diarahkan pada analisis perbedaan karakteristik konvergensi, stabilitas pelatihan, dan akurasi prediksi yang dihasilkan oleh metode optimasi orde pertama (Gradient Descent, Steepest Descent, dan Adam) serta metode optimasi orde kedua (Newton, Levenberg–Marquardt, BFGS, L-BFGS, dan QR Gauss-Newton). Melalui pendekatan komparatif, penelitian ini diharapkan memaparkan keterkaitan antara karakteristik dinamika sistem diferensial, arsitektur jaringan saraf berulang, dan algoritma optimasi, sehingga memberikan landasan sistematis dalam pemilihan metode optimasi yang sesuai untuk prediksi data time series berbasis persamaan diferensial.
======================================================================================================================================
This study investigates the impact of eight optimization methods on the performance of three recurrent neural network architectures, namely Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), in predicting time series data generated from differential equation systems. The time series data are constructed through numerical simulation of three nonlinear dynamical models: the SIR model and the Lotka–Volterra system using the Euler method, while the Lorenz system uses the fourth-order Runge–Kutta method (RK4) due to its chaotic nature requiring high integration accuracy. The main focus is a comparative analysis of convergence behavior, training stability, and prediction accuracy resulting from first-order optimization methods (Gradient Descent, SteepestDescent, and Adam) and second-order optimization methods (Newton,Levenberg–Marquardt, BFGS, L-BFGS, and QR-Gauss-Newton). Through this comparative framework,the study aims to reveal the interrelationship between differential equation dynamics, recurrent neural network architectures, and optimization algorithms,thereby providing a more systematic foundation for selecting appropriate optimization methods in differential equation-based time series prediction.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Persamaan diferensial, data time series, jaringan saraf berulang, RNN, LSTM, GRU, metode optimasi, QR-Gauss-Newton
Subjects:	Q Science > QA Mathematics Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Melvin Alamsyah
Date Deposited:	03 Aug 2026 03:35
Last Modified:	03 Aug 2026 03:35
URI:	http://repository.its.ac.id/id/eprint/142742

Actions (login required)

View Item