Sistem Deteksi Drift Pada Data Pengukuran Power Meter Menggunakan Long Short-Term Memory Autoencoder

Rohman, Alvin Haque (2026) Sistem Deteksi Drift Pada Data Pengukuran Power Meter Menggunakan Long Short-Term Memory Autoencoder. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 2040221092-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Pemantauan konsumsi energi pada sistem distribusi daya menghasilkan data pengukuran dalam jumlah besar yang perlu dipastikan tetap merepresentasikan kondisi operasional sebenarnya. Namun, perubahan karakteristik data pengukuran sering kali sulit dikenali karena proses pemantauan masih bergantung pada pengamatan manual dan belum didukung oleh baseline pola normal yang representatif. Kondisi tersebut menyebabkan proses identifikasi drift menjadi kurang konsisten serta berpotensi menghambat pengambilan keputusan dalam pemantauan energi. Untuk mengatasi permasalahan tersebut, proyek akhir ini mengembangkan sistem deteksi drift pada data pengukuran power meter menggunakan Long Short-Term Memory (LSTM) Autoencoder. Karena setiap power meter memiliki karakteristik pola konsumsi energi yang berbeda, dilakukan tahap pengelompokan data berdasarkan kemiripan pola konsumsi energi sebelum proses pemodelan sehingga model dapat mempelajari pola data yang lebih seragam. Deteksi drift dilakukan berdasarkan nilai Reconstruction Error, sedangkan penentuan threshold menggunakan persentil ke-95 terhadap Reconstruction Error yang telah difilter dengan metode Interquartile Range (IQR). Kinerja sistem dievaluasi menggunakan metrik Accuracy, Precision, Recall, dan F1-Score. Hasil pengujian menunjukkan bahwa sistem berhasil membentuk baseline pola normal data pengukuran dengan rata rata Reconstruction Error scaled sebesar 0,108071 hingga 0,120047 pada PM2 dan 0,104296 hingga 0,114024 pada PM5, serta berhasil mendeteksi drift pada data pengukuran power meter. Pada data pengujian, pendekatan individual memberikan performa terbaik pada PM2 dengan Accuracy sebesar 95,65%, Precision sebesar 62,50%, Recall sebesar 100%, dan F1-Score sebesar 76,92%, sedangkan pendekatan 4 Cluster memberikan keseimbangan performa terbaik pada PM5. Selain itu, pengujian tambahan menggunakan data Injection menunjukkan bahwa sistem mampu mendeteksi drift pada berbagai pola konsumsi energi yang di ukur oleh power meter dengan Accuracy berkisar antara 86,56% hingga 96,66%, Recall antara 50% hingga 100%, serta F1-Score antara 40% hingga 80%. ====================================================================================================================================
Energy consumption monitoring in power distribution systems generates large volumes of measurement data that must accurately represent actual operating conditions. However, changes in measurement data characteristics are often difficult to identify because the monitoring process still relies on manual observation and lacks a representative normal baseline. This condition makes drift identification less consistent and may hinder decision-making in energy monitoring. To address this problem, this final project develops a drift detection system for power meter measurement data using a Long Short-Term Memory (LSTM) Autoencoder. Since each power meter exhibits different energy consumption characteristics, the measurement data are first grouped based on the similarity of their consumption patterns before model training, enabling the model to learn more homogeneous data patterns. Drift detection is performed based on the Reconstruction Error, while the detection threshold is determined using the 95th percentile of the Reconstruction Error after filtering with the Interquartile Range (IQR) method. The system performance is evaluated using Accuracy, Precision, Recall, and F1-Score. The experimental results show that the proposed system successfully constructs a normal baseline for measurement data, with an average scaled Reconstruction Error ranging from 0.108071 to 0.120047 for PM2 and from 0.104296 to 0.114024 for PM5, and successfully detects drift in power meter measurement data. For the testing data, the individual approach achieved the best performance for PM2, with an Accuracy of 95.65%, Precision of 62.50%, Recall of 100%, and an F1-Score of 76.92%, while the 4-Cluster approach provided the best balance of detection performance for PM5. Furthermore, additional validation using ±25% injected data demonstrated that the proposed system was capable of detecting drift across various power meter energy consumption characteristics, achieving an Accuracy ranging from 86.56% to 96.66%, a Recall ranging from 50% to 100%, and an F1-Score ranging from 40% to 80%.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Drift Data Pengukuran, LSTM Autoencoder, Power meter, Reconstruction Error. Measurement Drift, LSTM Autoencoder, Power meter, Reconstruction Error.
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines.
Divisions:	Faculty of Vocational > 36304-Automation Electronic Engineering
Depositing User:	Alvin Haque Rohman
Date Deposited:	10 Aug 2026 02:19
Last Modified:	10 Aug 2026 02:19
URI:	http://repository.its.ac.id/id/eprint/144011

Actions (login required)

View Item