Deteksi Anomali Pada Konsumsi Energi Smart Building Menggunakan Long Short-Term Memory Autoencoder Dengan Rule-Based Labeling

Putri, Santi Oktaviana (2026) Deteksi Anomali Pada Konsumsi Energi Smart Building Menggunakan Long Short-Term Memory Autoencoder Dengan Rule-Based Labeling. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221138-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy

Abstract

Efisiensi energi merupakan aspek penting dalam pengelolaan smart building, sehingga diperlukan deteksi anomali penggunaan energi. Salah satu tantangannya adalah variasi pola konsumsi energi antar musim dan zona bangunan serta keterbatasan label anomali pada data operasional. Penelitian ini bertujuan menganalisis pola konsumsi energi pada smart building lantai 2 dataset CU-BEMS, mengevaluasi kemampuan
model Long Short-Term Memory Autoencoder (LSTM-AE) dalam mendeteksi anomali penggunaan energi, serta menganalisis perbedaan hasil deteksi pada setiap zona beserta interpretasinya terhadap pengelolaan energi. Anomali didefinisikan sebagai kondisi AC tetap beroperasi pada ruang yang tidak terindikasi dihuni selama lebih dari 30 menit
berdasarkan konsumsi daya lighting dan plug load menggunakan pendekatan rule-based labeling. Data diproses menggunakan beberapa konfigurasi window (W), kemudian model dievaluasi berdasarkan reconstruction error. Hasil penelitian menunjukkan, model LSTM Autoencoder memperoleh nilai recall hingga 0,8333 pada zona 4 dan 0,4375 pada zona 2. Setelah dilakukan random undersampling, nilai F1-score meningkat yang menunjukkan bahwa penyeimbangan kelas membantu meningkatkan kemampuan model dalam membedakan data normal dan anomali berdasarkan ground truth hasil rule-based labeling. Dibandingkan CNN Autoencoder sebagai model pembanding, LSTM Autoencoder menunjukkan kemampuan deteksi anomali yang lebih baik pada zona dengan proporsi anomali rendah, sedangkan CNN Autoencoder memperoleh nilai AUPRC
yang lebih tinggi pada zona dengan proporsi anomali lebih besar. Secara keseluruhan, LSTM Autoencoder memberikan performa deteksi yang lebih baik dibandingkan model baseline. Hasil penelitian menunjukkan bahwa LSTM Autoencoder berpotensi mendukung peningkatan efisiensi energi pada smart building melalui deteksi dini anomali
penggunaan AC yang mempertimbangkan karakteristik musim dan zona bangunan.
======================================================================================================================================
Energy efficiency is an important aspect in smart building management, so that energy usage anomaly detection is needed. One of the challenges is the variation in energy consumption patterns between seasons and building zones as well as the limitations of anomaly labels in operational data. This study aims to analyze energy consumption patterns on the 2nd floor of a smart building using the CU-BEMS dataset, to determine the ability of the Long Short-Term Memory Autoencoder (LSTM-AE) model to detect energy usage anomalies, and to analyze the differences in detection results in each zone along with their understanding of energy management. Anomalies are defined as conditions where the AC remains operating in a room that does not indicate pregnancy for more than 30 minutes based on the power consumption of lighting and plug load using a rule-based labeling approach. Data is processed using several window configurations (W), then the model is evaluated based on reconstruction error. The results showed that the LSTM Autoencoder model obtained a recall value of up to 0.8333 in zone 4 and 0.4375 in zone 2. After random undersampling, the F1-score value increased, indicating that class balancing helped improve the model’s ability to distinguish normal
and anomalous data based on the ground truth results of rule-based labeling. Compared to CNN Autoencoder as a comparison model, the LSTM Autoencoder showed better anomaly
detection capabilities in zones with a low anomaly proportion, while the CNN Autoencoder obtained a higher AUPRC value in zones with a larger anomaly proportion. Overall, the LSTM Autoencoder provided better detection performance than the baseline model. The results showed that the LSTM Autoencoder has the potential to support increased energy efficiency in smart buildings through early detection of AC usage anomalies that consider the characteristics of the season and building zone.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deteksi Anomali, Konsumsi Energi, Smart Building, LSTM Autoencoder, Rule-based Labeling, Anomaly Detection, Energy Consumption, Smart Building, LSTM Autoencoder, Rule-based Labeling
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Santi Oktaviana Putri
Date Deposited:	03 Aug 2026 09:01
Last Modified:	03 Aug 2026 09:01
URI:	http://repository.its.ac.id/id/eprint/142762

Actions (login required)

View Item