Pengembangan Sistem Peringatan Dini Berbasis AI Untuk Penggunaan Daya Distribusi Listrik Dengan Metode Long Short-Term Memory (LSTM)

Supomo, Wibaga Traya Scotikotama (2025) Pengembangan Sistem Peringatan Dini Berbasis AI Untuk Penggunaan Daya Distribusi Listrik Dengan Metode Long Short-Term Memory (LSTM). Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5022211207-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Ketidakstabilan konsumsi daya listrik pada bangunan bertingkat dapat menyebabkan gangguan operasional dan inefisiensi energi. Penelitian ini mengembangkan sistem peringatan dini berbasis kecerdasan buatan untuk mendeteksi anomali konsumsi daya secara real-time. Sistem menggunakan model Long Short-Term Memory (LSTM) dengan pendekatan recursive one-step-ahead forecasting untuk dua skala waktu: jangka pendek (30 menit) dengan resolusi data 1 menit dan lookback 120 menit, serta jangka panjang (1 minggu) dengan resolusi 1 jam dan lookback 168 jam. Data dikumpulkan dari sistem distribusi listrik menggunakan protokol Modbus, disimpan dalam InfluxDB, dan divisualisasikan melalui Grafana yang diakses publik melalui tunneling HTTPS Cloudflare. Proses deteksi anomali dilakukan melalui dua tahapan: pertama, membandingkan data aktual terhadap batas kendali statistik (UCL/LCL) untuk mendeteksi warning; kedua, mengevaluasi error prediksi menggunakan Mean Absolute Error (MAE). Sebuah kondisi diklasifikasikan sebagai anomali apabila ditemukan ≥15 warning berturut-turut dalam satu window, atau nilai MAE melebihi ambang batas tertentu. Pengujian dilakukan terhadap 10.080 titik data per fitur selama periode satu minggu. Sistem mendeteksi total 4.222 anomali pada daya aktif total gedung, 3.395 anomali pada lantai 1, 1.450 anomali pada lantai 2, 4.354 anomali pada lantai 3, dan 3.265 anomali pada lantai 4. Deteksi sebagian besar berasal dari evaluasi terhadap peta kendali, sedangkan sisanya berasal dari akumulasi error prediksi. Model jangka pendek menunjukkan performa terbaik dengan MAE sebesar 0.4601 dan MSE sebesar 0.8341, sementara model jangka panjang menghasilkan MAE sebesar 0.6561 dan MSE sebesar 1.8354. Sistem ini terbukti efektif dalam memberikan peringatan dini terhadap kondisi overload dan underload secara presisi, serta mendukung pengambilan keputusan dalam manajemen energi bangunan berbasis data
=======================================================================================================================================
Instability in electricity consumption in multi-story buildings can lead to operational disruptions and energy inefficiencies. This study develops an artificial intelligence-based early warning system to detect real-time anomalies in power consumption. The system employs a Long Short-Term Memory (LSTM) model using a recursive one-step-ahead forecasting approach on two time scales: short-term (30 minutes) with 1-minute resolution and a 120-minute lookback, and long-term (1 week) with 1-hour resolution and a 168-hour lookback. Data is collected from the electrical distribution system via the Modbus protocol, stored in InfluxDB, and visualized through Grafana, which is publicly accessible via a secure HTTPS tunnel using Cloudflare. Anomaly detection is performed in two stages: first, by comparing actual data against statistical control limits (UCL/LCL) to identify warnings; second, by evaluating prediction errors using Mean Absolute Error (MAE). A condition is classified as anomalous if there are ≥15 consecutive warnings within a sliding window or if the MAE exceeds a defined threshold. Testing was conducted on 10,080 data points per feature over a one-week period. The system detected 4,222 anomalies in total building power consumption, 3,395 on floor 1, 1,450 on floor 2, 4,354 on floor 3, and 3,265 on floor 4. Most anomalies were detected through statistical control chart evaluation, with the remainder identified through accumulated prediction errors. The short-term model demonstrated superior performance with an MAE of 0.4601 and an MSE of 0.8341, while the long-term model yielded an MAE of 0.6561 and an MSE of 1.8354. These results demonstrate the system's effectiveness in providing accurate early warnings of overload and underload conditions, supporting data-driven decision-making in building energy management.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Sistem Peringatan Dini, Deteksi Anomali, Konsumsi Daya Listrik, Long Short-Term Memory (LSTM), Peta Kendali Statistik, Mean Absolute Error (MAE), Distribusi Energi, IoT Monitoring, Cloudflare Tunnel, InfluxDB, Grafana, Early Warning System, Anomaly Detection, Electrical Power Consumption, Long Short-Term Memory (LSTM), Statistical Process Control, Mean Absolute Error (MAE), Energy Distribution, IoT Monitoring, Cloudflare Tunnel, InfluxDB, Grafana.
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK201 Electric Power Transmission T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK3030 Electric power distribution systems T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK351 Electric measurements. T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7878 Electronic instruments
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20201-(S1) Undergraduate Thesis
Depositing User:	Wibaga Traya Scotikotama Supomo
Date Deposited:	24 Jul 2025 02:59
Last Modified:	24 Jul 2025 02:59
URI:	http://repository.its.ac.id/id/eprint/120797

Actions (login required)

View Item