Innayah, Andina Safitri (2026) Analisis Efektivitas Continual Learning Pada Workload Forecasting Menggunakan iTransformer Dan FRNet Untuk Mendukung Cloud Resource Provisioning. Other thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 5025221204-Undergraduate_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
5025221204-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (4MB) | Request a copy |
Abstract
Perubahan pola workload pada lingkungan cloud computing menyebabkan distribusi data bersifat dinamis sehingga dapat menurunkan akurasi model workload forecasting yang dilatih secara statis. Pendekatan Continual Learning menawarkan kemampuan adaptasi terhadap perubahan distribusi data tanpa melakukan pelatihan ulang secara menyeluruh. Penelitian ini bertujuan mengevaluasi efektivitas Continual Learning berbasis Experience Replay pada workload forecasting menggunakan arsitektur iTransformer dan FRNet untuk mendukung resource provisioning di lingkungan cloud. Penelitian menggunakan dataset Alibaba Cluster Trace v2020 yang diolah menjadi data time-series dengan interval 30 menit. Evaluasi dilakukan pada dua feature set, yaitu demand only dan demand capacity, dengan tiga strategi pembelajaran, yaitu Experience Replay, Periodic Retraining, dan Static Baseline. Kinerja model dievaluasi menggunakan metrik Mean Squared Error (MSE), Mean Absolute Error (MAE), Average Accuracy, dan Backward Transfer (BWT). Selanjutnya, untuk kualitas simulasi resource provisioning, dianalisis menggunakan Under-Provisioning Rate, Over-Provisioning Rate, SLA Violation Rate, dan ΩSLA-VR. Hasil penelitian menunjukkan bahwa strategi Experience Replay menghasilkan performa yang lebih baik dibandingkan Static Baseline pada seluruh kombinasi arsitektur dan feature set. Experience Replay menghasilkan performa learning terbaik berdasarkan metrik Backward Transfer. Kemudian, feature set demand capacity memberikan hasil yang lebih baik dibandingkan demand only. Pada simulasi resource provisioning, penggunaan safety margin secara konsisten menurunkan under-provisioning rate, over-provisioning rate, SLA violation rate, dan ΩSLA-VR. Berdasarkan hasil tersebut, Continual Learning berbasis Experience Replay efektif meningkatkan kemampuan belajar model, sedangkan penggunaan feature set demand capacity memberikan kualitas prediksi dan simulasi resource provisioning yang lebih baik pada lingkungan cloud yang dinamis.
================================================================================================================================
The dynamic nature of workload patterns in cloud computing environments leads to changes in data distribution, which can degrade the performance of statically trained workload forecasting models. Continual Learning offers a learning paradigm that enables models to accommodate evolving data distributions without requiring complete retraining. This study aims to evaluate the effectiveness of Experience Replay-based Continual Learning for workload forecasting using the iTransformer and FRNet architectures to support cloud resource provisioning. The experiments were conducted using the Alibaba Cluster Trace v2020 dataset, which was processed into a time-series dataset with a 30-minute interval. The evaluation considered two feature sets, namely demand only and demand capacity, under three learning strategies: Experience Replay, Periodic Retraining, and Static Baseline. Forecasting performance was evaluated using Mean Squared Error (MSE), Mean Absolute Error (MAE), Average Accuracy, and Backward Transfer (BWT). The quality of the resource provisioning simulation was assessed using Under-Provisioning Rate, Over-Provisioning Rate, SLA Violation Rate, and ΩSLA-VR. The results show that Experience Replay consistently outperformed the Static Baseline across all combinations of architectures and feature sets. Experience Replay achieved the best learning performance in terms of Backward Transfer. The demand capacity feature set consistently outperformed the demand only feature set. In the resource provisioning simulation, increasing the safety margin consistently reduced the Under-Provisioning Rate, Over-Provisioning Rate, SLA Violation Rate, and ΩSLA-VR. These findings indicate that Experience Replay-based Continual Learning effectively improves the adaptability of workload forecasting models. Furthermore, incorporating capacity-related features leads to better forecasting performance and higher-quality resource provisioning simulation in dynamic cloud environments.
| Item Type: | Thesis (Other) |
|---|---|
| Uncontrolled Keywords: | Cloud Resource Provisioning, Continual Learning, Experience Replay, FRNet, iTransformer, Workload Forecasting |
| Subjects: | T Technology > T Technology (General) > T174 Technological forecasting |
| Divisions: | Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis |
| Depositing User: | Andina Safitri Innayah |
| Date Deposited: | 28 Jul 2026 04:25 |
| Last Modified: | 28 Jul 2026 04:28 |
| URI: | http://repository.its.ac.id/id/eprint/138524 |
Available Versions of this Item
- Analisis Efektivitas Continual Learning Pada Workload Forecasting Menggunakan iTransformer Dan FRNet Untuk Mendukung Cloud Resource Provisioning. (deposited 28 Jul 2026 04:25) [Currently Displayed]
Actions (login required)
 |
View Item |