제안된 저전압 수락 자동화 플랫폼은 사물인터넷(IoT), 시각적 워크플로우 오케스트레이션, 엣지 및 클라우드 컴퓨팅, 피처 엔지니어링, 머신러닝을 통합하여 저전압(LV) 시스템 수락 테스트를 자동화합니다. 이 플랫폼은 지능형 결함 검출, 실시간 의사결정, 자동 수락, 높은 정확도, 낮은 운영 지연 시간 및 확장 가능한 배포를 가능하게 합니다.
이 논문에서
요약
초록
저전압(LV) 네트워크의 급격한 확장과 분산 에너지 자원과의 통합으로 인해 지능형 및 자동화된 관리 솔루션이 필요하게 되었습니다. 클라우드-엣지 협업 사물인터넷(IoT) 플랫폼은 실시간 모니터링, 제어 및 데이터 수집을 지원합니다. 그러나 기존 플랫폼들은 일반적으로 워크플로우 자동화, 시각적 프로세스 오케스트레이션, 사용자 가이드 기반 의사결정 지원 및 종합적인 검증 기능이 부족합니다. 결과적으로, 이러한 플랫폼들은 자동화된 LV 네트워크 인수 테스트를 위한 프로세스 중심의 솔루션을 제공하지 못하고 있습니다. 본 연구에서는 시각화된 프로세스 캔버스 기반의 저전압 인수 자동화 플랫폼의 설계 및 평가를 제시합니다. 제안된 플랫폼은 인수 워크플로우를 시각적으로 생성, 관리 및 실행하는 프로세스 중심 아키텍처를 채택합니다. 시각적 프로세스 캔버스는 실시간 워크플로우 실행, 검증 및 의사결정을 가능하게 함으로써 기존의 정적 모니터링을 동적 워크플로우 자동화로 전환합니다. 이 프레임워크는 전기 부하 동작의 실제적인 모델링을 지원하기 위해 Open LV Network & Smart Meter 데이터셋을 통합하였습니다. 워크플로우에는 IoT 기반 데이터 수집, 데이터 전처리, 특성 공학, 시각적 프로세스 캔버스를 이용한 워크플로우 오케스트레이션, 머신러닝 기반 검증 및 실시간 대시보드 시각화가 포함됩니다. 제안된 프레임워크는 98.4%의 결함 검출률, 0.968의 수신자 조작 특성 곡선 아래 면적(ROC-AUC) 및 31 ms의 운전 의사결정 지연 시간을 달성하였으며, 이는 기존의 클라우드 중심 및 IoT 기반 베이스라인 접근 방식보다 약 30%–40% 더 우수한 성능을 보였습니다.
서론
산업 4.0 (I4.0)1 여러 신흥 기술들에 의해 가능해졌으며, 그중 산업용 사물 인터넷(IIoT)은 가장 중요한 기술 중 하나입니다. IIoT는 산업 환경에서 사물 인터넷(IoT) 기술을 적용하는 것을 의미하며, 여기에서는 서로 연결된 장치와 기계들이 주로 기계 간 상호작용을 통해 통신합니다. 산업 시스템의 고장은 상당한 운영 손실을 초래할 수 있으므로, IIoT 응용 프로그램은 기존 IoT 시스템보다 훨씬 더 많은 양의 데이터를 생성하는 동시에 매우 높은 신뢰성의 통신을 필요로 합니다.2또한, 일부 연구에서는 IIoT를 산업 4.0과 밀접하게 연관된 개념으로 간주합니다.3IoT 기술은 원활한 데이터 교환이 가능한 더욱 유연한 시스템을 구현함으로써 전통적인 계층 구조의 아키텍처를 변화시키고 있습니다.4최근 다수의 연구에서 기존의 산업 사물 인터넷(IIoT) 아키텍처를 검토하고 새로운 프레임워크 설계를 제안하였습니다.5,6,7,8,9이러한 발전에도 불구하고, 두 가지 주요한 한계점이 남아 있습니다. 첫째, 이전 아키텍처의 단점을 극복하기 위해 새로운 프레임워크가 빈번하게 제안됨에 따라, 아키텍처의 다양성과 구현 복잡성을 증가시키는 대안 모델의 수가 계속해서 늘어나고 있습니다. 둘째, 제안된 많은 아키텍처와 관련된 높은 추상화 수준으로 인해 산업 현장에서의 실질적인 도입이 제한적입니다.10.
문헌에는 여러 IIoT 아키텍처가 보고되어 있습니다. OPC(Open Platform Communications) 표준 기반의 IIoT 아키텍처가 OPC 통신 관리를 위한 Python 미들웨어와 함께 제시되었습니다11. 7계층 IIoT 아키텍처가 제안되었으나, 구현 세부 사항과 검증 실험은 제공되지 않았습니다12. 5G와 IIoT 기술을 통합한 프레임워크가 대표적인 유스케이스와 함께 설명되었지만, 실험적 검증은 제한적이었습니다13. 인더스트리 4.0을 위한 감성 디자인 프레임워크가 소개되었으며, 실험적 확인 없이 그 잠재적 영향이 논의되었습니다14. 다층 IIoT 플랫폼이 통신 측면의 고려 사항에 중점을 두고 실제 테스트베드에서 평가되었습니다15. 이와 유사하게, 실내 공기질 모니터링을 위한 모듈형 IoT 프레임워크가 제시되었으며, 어린이가 있는 84가구에서 수집한 데이터를 사용하여 그 성능을 평가하였습니다16.
사물인터넷(IoT)은 전력 부문에서 응용 가능성이 매우 큰, 빠르게 진화하는 기술이 되었습니다17,18. 전력 시스템 노드에 IoT 기술을 통합하면 스마트 그리드 내 전력 흐름에 대한 지속적인 실시간 모니터링을 기반으로 자율적인 의사결정이 가능해져, 더욱 지능적이고 효율적인 발전, 송전, 배전, 운용 및 활용이 가능합니다. 결과적으로, 이러한 지능형 노드들은 전력망의 전반적인 운용 성능을 향상시킵니다19. 스마트 그리드 구현에 있어 통신 기술은 핵심적인 구성 요소인데, 이는 센서 측정값과 제어 신호라는 두 가지 주요 범주의 정보가 안정적으로 전송되어야 하기 때문입니다. 스마트 그리드 응용 분야에서는 신뢰할 수 있는 운용을 보장하기 위해 낮은 통신 지연 시간뿐만 아니라 적절한 서비스 품질(QoS)과 충분한 데이터 보안이 요구됩니다20.
스마트 그리드 애플리케이션에 적합한 통신 기술을 선택하려면 통신 간격, 데이터 전송 속도 및 통신 비용을 고려해야 합니다21. IoT 통신 네트워크는 일반적으로 소프트웨어, 보안, 서비스, 시스템 및 연결 계층으로 구성됩니다. 연결 계층 내에서 통신 프로토콜은 일반적으로 세 가지 범주로 분류됩니다. 장거리 프로토콜에는 Long Range (LoRa) 및 Narrowband-IoT (NB-IoT)가 포함되며, 중거리 프로토콜에는 Wi-Fi, 4G/LTE 및 5G가 포함되고, 단거리 프로토콜에는 Bluetooth 및 ZigBee가 포함됩니다. 전력 시장의 규제가 점차 완화됨에 따라, 기상 조건, 장비 고장, 규제 변경 및 기타 운영 요인이 전력 서비스 비용에 미치는 영향을 평가하기 위한 부하 예측이 더욱 중요해졌습니다. 결과적으로, 정확한 주거용 부하 예측은 스마트 그리드, 마이크로그리드 및 스마트 빌딩의 운영과 감독에 필수적이 되었습니다22.
저압(LV) 네트워크의 신뢰할 수 있는 운용은 높은 전력 품질을 유지하는 것에 달려 있습니다23. 전력 품질은 불평형 운전 조건 하에서 전압 조정, 전류 균형 및 기술적 교란 완화의 효과성에 의해 결정됩니다24. 이러한 요인들 중 상 평형(phase balancing)은 각 상에 걸쳐 전압과 전류의 균일한 분포를 결정하기 때문에 특히 중요합니다25. 불량한 상 평형은 도체 가열, 전압 편차 및 중성선 전류를 증가시켜 시스템 효율성과 운용 신뢰성을 저하시킬 수 있습니다26. 따라서 시스템 성능, 운용 안정성 및 장기적 신뢰성을 향상시키기 위해 설계 과정에서 상 평형을 고려해야 합니다. 그럼에도 불구하고, 이러한 측면은 기존의 미니그리드 설계 방법론에 거의 반영되지 않고 있습니다.
워크플로 오케스트레이션 시스템의 최근 발전은 산업용 IoT 환경 내에서 이벤트 기반 자동화와 시각적 프로세스 관리의 이점을 입증해 왔습니다. 그럼에도 불구하고, 배전망 검증 및 저압 수락 시험에 대한 이러한 시스템의 적용은 여전히 제한적입니다. 대부분의 기존 저압 자동화 시스템은 주로 모니터링 및 감시 제어에 집중하고 있는 반면, 워크플로 기반의 수락 관리나 자동화된 준수 여부 검증을 조사한 연구는 상대적으로 적습니다. 마찬가지로, 머신러닝 기술이 스마트 미터 이상 탐지 및 전력 품질 평가 분야에서 상당한 관심을 끌었지만, 이러한 접근 방식은 일반적으로 통합 워크플로 오케스트레이션 프레임워크의 구성 요소라기보다 독립적인 분석 모델로 개발되었습니다.
최근 몇 년 동안 스마트 빌딩 관리 및 홈 오토메이션을 위한 다양한 스마트 에너지 미터 솔루션이 제안되었습니다27. 대학 수준의 교육 개념을 스마트 미터링 기술에 통합하는 것을 입증하기 위해 참고문헌27에서 스마트 미터 프로토타입이 제시되었습니다. 개발된 스마트 미터는 에너지 관리 및 제어 분야에서 상당한 잠재력을 보여주었으며, 양방향 IoT 통신을 통해 에너지 소비를 줄임으로써 유틸리티 기업과 소비자 모두에게 이점을 제공했습니다. 이후 지능형 에너지 모니터링을 지원하기 위해 마이크로컨트롤러 기반의 스마트 플러그 시스템이 도입되었습니다28.
제안된 스마트 플러그 시스템은 Android 기반의 그래픽 사용자 인터페이스를 통해 전력 소비량을 실시간으로 표시합니다. 또한, 이 장치는 연결된 전기 제품의 존재 여부를 감지할 수 있으며, 이러한 기능을 비침습적 부하 측정(invasive load measurement)28이라고 합니다. 구현을 위해 저비용의 상용 스마트 플러그가 사용되었습니다. 향후 개선 사항으로는 실효치(RMS) 전류, 유효 전력, 무효 전력 및 전압과 전류 사이의 위상각 모니터링이 포함됩니다. 이전 연구에서 설명한 바와 같이, 부분적으로 그리드에 연결된 태양광 마이크로그리드는 ZigBee 및 LoRa 통신 프로토콜을 사용하여 스마트 미터 장치와 정보를 교환함으로써 분산 에너지 시스템을 위한 IoT 기반 통신의 가능성을 입증하였습니다29,30.
기존 배전 시스템 계획의 한계를 해결하기 위해 공학 실무와 과학 연구는 여러 상호 보완적인 방향으로 발전해 왔습니다. 한 연구 방향은 부지 선정 및 예비 자원 평가를 위한 지리 정보 시스템(GIS)을 이용한 고수준 계획에 집중해 왔습니다31,32. 또 다른 방향은 도체 선정, 최적 발전 및 에너지 저장 용량 산정, 그리고 교류(AC) 및 직류(DC) 배전 시스템의 비교 평가를 포함하여 개별 구성 요소와 하위 시스템을 개선하는 데 집중했습니다33,34,35,36. 세 번째 연구 방향은 최적화 알고리즘을 배전망 계획, 피더 경로 설정 및 네트워크 토폴로지 설계에 적용한 것입니다. 최근의 발전 사례로는 장치 배치와 네트워크 구성을 통합한 연구진의 uGrid 플랫폼과 같은 프레임워크가 있으나, 검증 및 최적화를 위한 표준 조류 계산 분석이 부족한 실정입니다. 다른 접근 방식들은 토폴로지 생성을 위해 수동 매개변수 튜닝에 의존하거나 저전력 DC 마이크로그리드와 같은 특수 응용 분야에 집중하고 있습니다37,38,39. 이러한 접근 방식들이 상당한 진전을 나타내고 있음에도 불구하고, 저압 배전 시스템의 엔드투엔드 설계를 위한 통합 방법론을 구축하기 위해서는 추가적인 연구가 필요합니다. 또한 단상, 3상 및 하이브리드 저압 시스템에 대한 종합적인 기술 경제성 평가는 여전히 제한적인 상태입니다.
인공지능(AI) 및 머신러닝의 최근 발전은 지능형 전력 시스템의 역량을 더욱 확장시켰습니다. 한 종합적인 리뷰에서는 연합 학습(Federated Learning), 생성형 AI(Generative AI), 대규모 언어 모델(Large Language Models), 지능형 사물인터넷(AIoT) 및 디지털 트윈(Digital Twins)과 같은 신흥 기술을 포함하여 스마트 그리드에서의 AI 및 머신러닝 적용 사례를 요약하여 제시하였습니다40. 해당 리뷰는 부하 예측, 예측 유지보수, 이상 징후 탐지, 수요 측면 관리 및 전기차 통합 분야의 응용 사례를 강조하는 한편, 상호 운용성, 개인정보 보호, 확장성, 강건성 및 윤리적 고려 사항과 관련된 지속적인 과제들을 역설하였습니다. 또 다른 최근 리뷰에서는 분산형 에너지 제어를 위한 머신러닝, 강화 학습 및 다중 에이전트 시스템에 초점을 맞추어, IoT-엣지-클라우드 인프라 내 분산 에너지 관리를 위한 AI 방법론을 평가하였습니다41.
종합적으로, 이전의 연구들은 저압(LV) 배전 시스템을 위한 IoT 기반 모니터링, 스마트 미터링, 클라우드-엣지 컴퓨팅, 인공지능 및 지능형 에너지 관리를 발전시켜 왔습니다. 하지만 이러한 기술들은 대부분 독립적으로 개발되었습니다. 워크플로우 오케스트레이션, 자동화된 수락 관리, 프로세스 기반 검증 및 통합 의사결정 지원은 여전히 제한적인 상황이며, 이는 통합된 워크플로우 중심 자동화 프레임워크의 필요성을 강조합니다.
따라서 본 연구에서는 시각화된 프로세스 캔버스를 기반으로 하는 저압 수전 자동화 플랫폼을 제시합니다. 제안된 프레임워크는 실제 스마트 미터 데이터를 사용하여 자동화되고 신뢰할 수 있는 LV 네트워크 수전 시험을 지원하기 위해 IoT 기반 데이터 수집, 데이터 전처리, 특성 공학, 워크플로 오케스트레이션, 머신러닝 기반 검증, 클라우드-엣지 컴퓨팅 및 실시간 시각화를 통합 아키텍처 내에 통합합니다. 그림 1은 제안된 플랫폼의 워크플로 지향 아키텍처에 대한 개요를 제공합니다. 이 워크플로는 데이터 수집, 전처리, 특성 공학, 워크플로 오케스트레이션, 지능형 의사결정 및 클라우드-엣지 통합으로 구성되어, 프로세스 지향 프레임워크를 통한 LV 배전망의 자동 검증을 가능하게 합니다. 이 플랫폼은 기존 방식과 비교하여 운영 의사결정 지연 시간과 워크플로 실행 시간을 단축하는 동시에 높은 검증 정확도를 입증했습니다.
그림 1. 제안된 저압 수락 자동화 플랫폼의 워크플로우 중심 아키텍처. 제안된 프로세스 기반 저압(LV) 수락 자동화 플랫폼의 상위 수준 아키텍처. 이 프레임워크는 LV 배전망 검증을 위한 IoT 기반 데이터 수집, 데이터 전처리, 특성 공학, 워크플로우 오케스트레이션, 지능형 의사결정, 클라우드-엣지 협업 및 자동 수락의 순차적 통합을 보여줍니다. IoT = 사물 인터넷; LV = 저압; ML = 머신 러닝. 이 그림의 더 큰 버전을 보려면 여기를 클릭하십시오.
프로토콜
This study did not involve human participants, animals, or identifiable personal data. The analyses were performed using a publicly available dataset (UK Power Networks Open LV Network & Smart Meter dataset) together with synthetically generated data for validation; therefore, ethics committee approval and informed consent were not required. The research tools used in the protocol are listed in the Table of Materials.
1. Data acquisition
The Low-Voltage (LV) Acceptance Automation Platform was developed using electrical measurements acquired from smart meters, IoT sensors, and the Open LV Network & Smart Meter Dataset. Electrical parameters, including voltage, current, active power, reactive power, and energy consumption, were collected through Wi-Fi-, PLC-, and RS-485-based communication protocols. Real-time measurements were used for live system monitoring, whereas historical data supported model development and validation.
The proposed framework used the Open LV Network & Smart Meter Dataset provided by UK Power Networks42. The dataset comprises household smart meter consumption, feeder-level load profiles, and feeder voltage measurements. Time-series subsets containing load consumption and voltage data were selected for model training and evaluation. In addition, a simulated smart meter dataset was generated to represent operational conditions and fault scenarios required for workflow validation and machine learning model evaluation. The simulated dataset included timestamped measurements of voltage, current, active power, reactive power, energy consumption, total harmonic distortion (THD), and voltage imbalance. Fault samples were generated by simulating undervoltage, overvoltage, excessive harmonic distortion, and voltage imbalance based on established low-voltage operational thresholds. The dataset characteristics are summarized in Table 1.
| Attribute | Description |
| Dataset Name | Open LV Network & Smart Meter Dataset |
| Simulated Smart Meter Dataset | Geneated 1000 smaples for Fault simulation and workflow validation |
| Data Type | Time-series (historical smart meter data) |
| Data Granularity | Household-level and feeder-level measurements |
| Key Features | Voltage, current, active power, reactive power, energy consumption, timestamps |
| Data Frequency | High-resolution (e.g., half-hourly / minute-level depending on subset) |
| Coverage | Residential low-voltage distribution networks |
| Usage in Proposed Work | Model training, validation, load behavior analysis, and acceptance workflow evaluation |
Table 1: Dataset description. Summary of the Open LV Network & Smart Meter dataset used for model development and performance evaluation. The table presents the dataset components, measured electrical variables, sampling characteristics, and dataset partitioning for training, validation, and testing.
Because the Open LV Network & Smart Meter Dataset does not contain predefined fault labels, a rule-based labeling strategy was applied to both the real and simulated datasets. Samples satisfying normal operating limits were assigned a label of 0 (Normal), whereas samples exceeding operational thresholds (e.g., voltage deviations greater than ±10% of the nominal voltage, THD > 5%, or voltage unbalance > 2%) were assigned a label of 1 (Fault). These labels served as the ground truth during model training, validation, and testing. The proposed rule engine also allows threshold values to be configured in accordance with regional standards or utility-specific acceptance criteria.
The dataset was partitioned into training (70%), validation (15%), and testing (15%) subsets. Missing values were addressed using linear interpolation, mean imputation, and forward-fill techniques as described in the preprocessing stage. To ensure statistical robustness and reproducibility, all experiments were repeated five times using different random seeds.
2. Data preprocessing
Raw data from the Open LV Network & Smart Meter Dataset and IoT sensor measurements were preprocessed to improve data quality prior to feature extraction and model development. Missing values, noise, and inconsistencies caused by communication interruptions and sensor malfunctions were identified and corrected using interpolation, mean imputation, and forward-fill techniques. The detailed mathematical derivations are provided in Supplementary File 1 (Section S1).
The preprocessed time-series data were segmented into representative operational intervals (S₁–S₅ and Sk). Statistical and electrical features, including average load, peak load, minimum load, standard deviation, voltage deviation, and power consumption, were extracted from each segment. The extracted features are summarized in Table 2 and served as inputs for the subsequent feature engineering stage.
| Segment ID | Mean Load (kW) | Peak Load (kW) | Min Load (kW) | Std Deviation (kW) | Voltage Deviation (p.u.) | Power Consumption (kWh) | Time Interval |
| S₁ | 2.45 | 3.8 | 1.2 | 0.65 | 0.032 | 58.4 | 00:00–01:00 |
| S₂ | 2.9 | 4.1 | 1.5 | 0.72 | 0.028 | 64.2 | 01:00–02:00 |
| S₃ | 3.25 | 4.85 | 1.8 | 0.81 | 0.035 | 71.6 | 02:00–03:00 |
| S₄ | 3.8 | 5.2 | 2.1 | 0.9 | 0.041 | 78.9 | 03:00–04:00 |
| S₅ | 4.1 | 5.75 | 2.4 | 0.95 | 0.045 | 85.3 | 04:00–05:00 |
| Sk | 3.6 | 5.1 | 2 | 0.88 | 0.038 | 76.5 | tk |
Table 2: Representative feature representation extracted from time-series segments. Representative statistical, electrical, load behavior, voltage stability, and power quality features extracted from segmented LV time-series data during feature engineering. The values illustrate the engineered feature representation and are not experimental performance results. LF = Load Factor; PAR = Peak-to-Average Ratio; VSI = Voltage Stability Index; PF = Power Factor; THD = Total Harmonic Distortion.
3. Feature engineering
Feature engineering was performed to improve the representation of the preprocessed data for workflow orchestration and intelligent decision-making. Principal Component Analysis (PCA)^43 was applied to reduce data dimensionality by removing redundant and highly correlated variables while retaining the most informative feature components. The mathematical derivations are provided in Supplementary File 1 (Section S2).
Statistical descriptors, load behavior indicators, voltage stability indices, and power quality metrics were derived from the segmented time-series data. These engineered features included Load Factor (LF), Peak-to-Average Ratio (PAR), Voltage Stability Index (VSI), Voltage Unbalance Index, Power Factor (PF), and Total Harmonic Distortion (THD). The resulting feature set was used as the input to the workflow orchestration and intelligent decision-making modules.
The feature engineering workflow is summarized in Supplementary File 1 (Supplementary Algorithm 1). The procedure consisted of segmenting the preprocessed data, extracting statistical and electrical features, calculating load behavior, voltage stability, and power quality indicators, and applying PCA to generate the final feature vector used for model training and validation.
4. Visualized process canvas design
The Visualized Process Canvas was developed as a graph-based workflow orchestration framework for automating low-voltage (LV) acceptance procedures. The workflow was represented as a directed graph, G = (N, E), where N = {n1, n2, …, nm}denotes the workflow nodes corresponding to individual processing stages (e.g., data input, preprocessing, validation, and decision-making), and E = {eij}represents the directed edges defining the execution sequence between nodes. This graph-based representation enabled the creation of configurable workflows and their automated execution without manual coding.
Each workflow node performed a specific processing task by receiving input data, executing predefined operations, and forwarding the processed output to subsequent nodes. Acceptance criteria, including voltage and load thresholds, were implemented as configurable validation rules that could be updated without modifying the workflow structure. The workflow supported both sequential and parallel execution to improve computational efficiency. Machine learning models were integrated into designated workflow nodes to perform anomaly detection and automated acceptance validation (Figure 2).
Figure 2: Visualized process canvas architecture for LV acceptance automation. Overview of the visual workflow orchestration framework used for LV acceptance automation. The process canvas illustrates the graphical organization of workflow nodes for data processing, feature engineering, rule-based validation, intelligent decision-making, and output generation to support workflow design, execution, and monitoring. Please click here to view a larger version of this figure.
The mathematical formulation describing information flow through the workflow graph is provided in Supplementary File 1 (Section S3). Supplementary File 1 (Supplementary Algorithm 2) describes the workflow orchestration procedure, including workflow traversal, node execution, rule validation, decision generation, and output aggregation.
The Visualized Process Canvas was implemented using a microservice-oriented software architecture to support scalability and modular deployment. The graphical workflow interface was developed using React.js to provide drag-and-drop workflow configuration. Backend orchestration services were implemented using Flask/FastAPI, while Eclipse Mosquitto served as the MQTT message broker for communication between workflow modules. EdgeX Foundry provided edge device management and IoT integration. Machine learning services were implemented with Scikit-learn, TensorFlow, and PyTorch, and containerized with Docker and Kubernetes.
5. Workflow orchestration engine
The workflow orchestration engine coordinated execution of the configured acceptance workflow. A detailed description is provided in Supplementary File 1 (Section S4), and the execution sequence is summarized in Supplementary Figure 1.
The workflow execution process consisted of validating the workflow definition, initializing the execution environment, scheduling workflow nodes according to their dependencies, executing nodes sequentially or in parallel, applying rule-based validation, aggregating intermediate outputs, and generating the final acceptance or rejection decision based on rule-based and machine learning evaluation. Execution results were published through dashboards, reports, application programming interfaces (APIs), and data storage services. The overall workflow orchestration process is illustrated in Figure 3.
Figure 3: Comparison of classification performance metrics across different methods. Comparison of the classification performance of IoT-LV, RBM, Support Vector Machine (SVM), Random Forest (RF), Long Short-Term Memory (LSTM), and the proposed framework using the Open LV Network & Smart Meter dataset. Performance is evaluated using Accuracy, Precision, Recall, and F1-score. Results are presented as mean ± standard deviation (SD) from five independent experimental runs using a 70%/15%/15% training/validation/testing split. Error bars represent one standard deviation. Please click here to view a larger version of this figure.
6. Intelligent decision module
The Intelligent Decision Module was developed to automate low-voltage (LV) acceptance decisions using supervised machine learning models. Classification and regression algorithms were applied to the engineered feature vectors to support operational decision-making. Classification models identified normal and abnormal operating conditions, whereas regression models estimated future energy consumption and operational trends.
The input feature vector, X ∈ ℝd, comprised the statistical, voltage stability, load behavior, and power quality features extracted during the feature engineering stage. Standalone Support Vector Machine (SVM)44, Random Forest (RF)45, and Long Short-Term Memory (LSTM)46 models were implemented as baseline classifiers for comparative evaluation. The mathematical formulation of the learning models is provided in Supplementary File 1 (Section S5).
For classification, the output variable y ∈ {0,1} represented normal (0) and abnormal (1) operating conditions. When abnormal operating conditions or overload events were detected, the workflow generated a fault notification and flagged the acceptance process as invalid; otherwise, the process was classified as accepted.
The workflow orchestration engine integrated machine-learning predictions with rule-based validation, using a priority-based conflict-resolution strategy. Safety-critical operational rules and regulatory acceptance criteria were treated as hard constraints and therefore took precedence over machine-learning predictions when conflicts occurred. Final acceptance decisions were generated using the combined outputs of the rule-based validation and intelligent decision modules.
7. Cloud–edge integration and visualization dashboard
A cloud-edge computing architecture was implemented to support scalable data processing, storage, and visualization for the LV acceptance automation platform. Smart meters, SCADA systems, and IoT devices continuously collect electrical measurements, which are processed at the edge layer via filtering, normalization, feature extraction, and real-time inference to reduce communication overhead and enable low-latency decision-making. The mathematical derivations are provided in Supplementary File 1 (Section S6).
Processed data were transmitted to the cloud layer for long-term storage, advanced analytics, feature engineering, model training, and model management. Interactive dashboards provided real-time system monitoring, alerts, reporting, and operational decision support. The cloud-edge architecture is illustrated in Supplementary Figure 2, and the implementation details are summarized in Table 3.
| Layer | Component | Hardware / Software | Specification / Version | Purpose / Description |
| Edge Layer | Edge Device | Raspberry Pi 4 / NVIDIA Jetson Nano | Quad-core CPU, 4–8 GB RAM | Real-time data acquisition and local processing |
| IoT Gateway | Industrial IoT Gateway | MQTT / Modbus / OPC-UA support | Communication between sensors and edge/cloud | |
| Sensors / Smart Meters | Digital Smart Meters | Voltage, Current, Power Measurement | Data collection from LV network | |
| Edge Software | Python, EdgeX Foundry, MQTT Broker | Python 3.x | Data preprocessing, filtering, local inference | |
| Communication Protocols | Wi-Fi / PLC / RS-485 | Standard protocols | Data transmission to cloud layer | |
| Cloud Layer | Cloud Platform | AWS / Microsoft Azure / Google Cloud | Scalable cloud services | Data storage, analytics, and model training |
| Storage | AWS S3 / Azure Blob / InfluxDB | Time-series database | Storing historical and real-time data | |
| Compute & ML | Python, TensorFlow, PyTorch, Scikit-learn | ML libraries (latest versions) | Model training, evaluation, deployment | |
| Containerization | Docker, Kubernetes | Latest stable versions | Workflow orchestration and scalability | |
| Data Processing | Apache Spark / Pandas | Distributed processing | Large-scale data analytics | |
| Visualization Layer | Dashboard Tools | Grafana / Apache Superset / Plotly Dash | Web-based tools | Real-time monitoring and visualization |
| Web Framework | React.js / Angular / Flask / FastAPI | Modern web stack | UI development and API integration | |
| Notification System | Email / SMS Gateway / Push Notifications | Twilio / Firebase | Alert generation and communication | |
| API Services | REST APIs | JSON-based communication | Integration between modules |
Table 3: System implementation details. Summary of the hardware configuration, software environment, communication protocols, and cloud–edge computing components used for implementing the proposed LV acceptance automation platform.
결과
The performance of the proposed Low-Voltage (LV) Acceptance Automation Platform was evaluated using the Open LV Network & Smart Meter Dataset together with a synthetic smart meter dataset generated to represent additional operating conditions and fault scenarios. The synthetic dataset enabled controlled evaluation of workflow orchestration, intelligent decision-making, and fault detection under abnormal operating conditions. The Open LV Network & Smart Meter Dataset comprised household smart meter measurements, feeder-level load profiles, and voltage measurements. Time-series segments corresponding to load consumption and voltage variations were extracted for model development and evaluation.
The dataset was partitioned into training (70%), validation (15%), and testing (15%) subsets. The proposed LV Acceptance Automation Platform achieved high classification accuracy (approximately 96%–98%), effective fault detection capability, and reduced operational decision latency. To ensure statistical robustness and reproducibility, all experiments were repeated five independent times using different random seeds. The resulting performance distributions were subsequently analyzed using paired t-tests, one-way analysis of variance (ANOVA), Tukey's honestly significant difference (HSD) post-hoc analysis, and 95% confidence intervals. The simulation environment used for implementing the proposed framework is summarized in Table 4.
| Parameter | Value | Description |
| Dataset | Open LV Network & Smart Meter Dataset + Synthetic Smart Meter Dataset | Real-world and simulated LV network data |
| Total Samples | 1000 synthetic samples + selected Open LV subsets | Data used for validation and testing |
| Data Split | 70% Training / 15% Validation / 15% Testing | Model development and evaluation |
| Number of Independent Runs | 5 | Statistical reliability analysis |
| Random Seeds | 42, 52, 62, 72, 82 | Reproducible experimental setup |
| Total Statistical Observations | 30 observations (6 models × 5 runs) | Used for t-test, ANOVA, and Tukey HSD |
| Sampling Rate | 30-min / 1-min intervals | Time-series resolution |
| Features | Load, Voltage, Current, Active Power, Reactive Power, THD, Voltage Unbalance, Derived Features | Input features for ML models |
| ML Models | Random Forest, SVM, XGBoost, LSTM | Intelligent decision module |
| Hardware | NVIDIA RTX 3090 GPU, Intel Core i9 CPU, 64 GB RAM | Training and simulation environment |
| Edge Device | Raspberry Pi 4 / Jetson Nano | Edge inference and real-time processing |
| Software | Python 3.10, TensorFlow 2.12+, Scikit-learn 1.3+, PyTorch 2.0+ | Model implementation |
| Communication Protocols | MQTT / HTTP / RS-485 | Data transmission |
| Threshold Values | Voltage ±10%, THD > 5%, Voltage Unbalance > 2%, Feeder Load > 90% capacity | Fault labeling and validation rules |
| Statistical Tests | Paired t-test, One-way ANOVA, Tukey HSD, 95% Confidence Interval | Performance significance evaluation |
Table 4: Simulation environment. Configuration of the experimental environment, including dataset partitioning, machine learning models, hardware specifications, software environment, and validation protocol. Experiments were performed using a 70% training, 15% validation, and 15% testing split and repeated five times using different random seeds.
Classification performance
The proposed framework was evaluated using multiple performance metrics, including accuracy, precision, recall, F1-score, latency, and fault detection rate. Accuracy was used to assess overall classification performance, whereas precision, recall, and F1-score evaluated the effectiveness of fault detection. Latency measured the operational decision response time of the workflow, and the fault detection rate quantified the ability of the framework to identify abnormal operating conditions within the LV distribution network. Scalability was further assessed by progressively increasing the number of input data samples.
The proposed LV Acceptance Automation Platform was compared with five benchmark approaches, namely IoT-LV, Rule-Based Monitoring (RBM), Support Vector Machine (SVM), Random Forest (RF), and Long Short-Term Memory (LSTM). The IoT-LV framework primarily supports data acquisition and basic monitoring using predefined threshold criteria without intelligent decision-making capabilities. Similarly, the RBM approach relies on fixed operational thresholds and therefore lacks adaptive learning. Although SVM and RF effectively classify operational data through supervised learning, they do not incorporate workflow orchestration or integrated acceptance management. LSTM provides improved temporal learning for time-series analysis but does not support automated workflow execution. In contrast, the proposed framework combines workflow orchestration, hybrid rule-based and machine learning decision-making, feature engineering, and cloud-edge computing within a unified automation architecture.
To ensure a fair comparison, all baseline methods were evaluated using the same training, validation, and testing partitions (70%, 15%, and 15%, respectively). Hyperparameter optimization for the machine learning baselines was performed using the validation dataset, and all models employed identical preprocessing and feature-scaling procedures. Unlike the standalone SVM44, RF45, and LSTM46 classifiers, which were trained directly on the original feature space, the proposed framework incorporated advanced feature engineering, workflow validation, and acceptance-rule orchestration prior to classification.
The classification performance of all methods is summarized in Table 5. Among the conventional approaches, IoT-LV achieved the lowest classification accuracy (82.4%) because of its reliance on conventional monitoring without intelligent decision-making. RBM slightly improved performance, reaching an accuracy of 85.7%; however, its fixed validation thresholds limited adaptability. Machine learning approaches demonstrated progressively better performance, with SVM and RF achieving accuracies of 91.6% and 93.4%, respectively, owing to their ability to learn discriminative patterns from the data. LSTM further improved the classification accuracy to 95.2% by effectively capturing temporal dependencies within the smart meter time-series data.
The proposed LV Acceptance Automation Platform achieved the best overall classification performance, with an accuracy of 97.3%, precision of 96.8%, recall of 96.1%, and an F1-score of 96.4%. These improvements are attributable to the integration of advanced preprocessing, feature engineering, workflow orchestration, hybrid rule-based and machine learning decision-making, and cloud-edge collaborative processing rather than the application of the machine learning algorithms alone. Collectively, these components enhanced the separability between normal and abnormal operating conditions while improving the efficiency of automated acceptance validation.
The comparative classification performance of all evaluated methods is illustrated in Figure 3.
| Method | Accuracy (%) | Precision (%) | Recall (%) | F1-Score (%) |
| IoT-LV | 82.4 | 80.2 | 78.5 | 79.3 |
| RBM | 85.7 | 83.9 | 82.1 | 83 |
| SVM | 91.6 | 90.8 | 89.7 | 90.2 |
| RF | 93.4 | 92.6 | 91.8 | 92.2 |
| LSTM | 95.2 | 94.5 | 93.8 | 94.1 |
| Proposed Method | 97.3 | 96.8 | 96.1 | 96.4 |
Table 5: Comparison of classification metrics. Comparison of the classification performance of IoT-LV, RBM, SVM, RF, LSTM, and the proposed framework using Accuracy, Precision, Recall, and F1-score. Results are reported as mean ± standard deviation (SD) obtained from five independent experimental runs.
Statistical significance analysis
To further evaluate the reliability of the proposed LV Acceptance Automation Platform, statistical significance testing was performed to compare its performance with the benchmark methods, including IoT-LV, RBM, SVM, RF, and LSTM. Multiple experimental runs were conducted using different train-validation-test partitions of the Open LV Network & Smart Meter Dataset. The resulting classification accuracies were used to calculate the mean and standard deviation for each method, followed by paired t-tests to determine whether the observed performance differences were statistically significant. The mathematical formulation of the paired t-test is provided in Supplementary File 1 (Section S7).
A significance level of 0.05 was adopted for all statistical analyses. To ensure reproducibility, each experiment was repeated five independent times using different random seeds and data partitions. The performance metrics obtained from each run were subsequently used for statistical evaluation.
The paired t-test results are summarized in Table 6. The proposed framework significantly outperformed all benchmark methods (p < 0.05), demonstrating statistically significant improvements in accuracy, precision, recall, and F1-score. Significant performance gains were observed not only over conventional approaches such as IoT-LV and RBM but also over the more advanced machine learning models, including SVM, RF, and LSTM. These findings indicate that the observed improvements were not attributable to random variation but resulted from the integrated workflow-oriented automation framework.
| Comparison | p-value | Significance |
| Proposed vs IoT-LV | 0.0008 | Significant |
| Proposed vs RBM | 0.0015 | Significant |
| Proposed vs SVM | 0.012 | Significant |
| Proposed vs RF | 0.021 | Significant |
| Proposed vs LSTM | 0.034 | Significant |
Table 6: Pairwise statistical significance analysis (p-values). Pairwise statistical comparison between the proposed framework and the benchmark methods. The table reports p-values obtained from statistical significance testing, where p < 0.05 indicates a statistically significant difference.
Receiver operating characteristic (ROC) analysis
Receiver operating characteristic (ROC) analysis was performed to further evaluate the classification capability of the proposed LV Acceptance Automation Platform in comparison with IoT-LV, RBM, SVM, RF, and LSTM. The area under the ROC curve (AUC) was used to quantify the discrimination capability of each classifier by evaluating the trade-off between the true positive rate (TPR) and false positive rate (FPR).
The IoT-LV framework achieved the lowest AUC (approximately 0.690), reflecting its reliance on conventional monitoring without intelligent decision-making. RBM produced a moderate AUC of approximately 0.738 because of its dependence on fixed threshold-based validation rules. The machine learning classifiers demonstrated progressively improved performance, with SVM and RF achieving AUC values of approximately 0.833 and 0.861, respectively, owing to their ability to learn discriminative patterns from operational data. LSTM further improved classification performance, achieving an AUC of approximately 0.885 by effectively modeling temporal dependencies in smart meter time-series data.
The proposed LV Acceptance Automation Platform achieved the highest AUC (approximately 0.910), demonstrating superior discrimination between normal and abnormal operating conditions. This improvement is attributed to the combined effects of advanced feature engineering, workflow orchestration, hybrid rule-based and machine learning decision-making, and cloud-edge collaborative processing. The comparative ROC curves for all evaluated methods are presented in Figure 4.
Figure 4: Receiver operating characteristic (ROC) curve comparison for LV acceptance Automation. Comparison of the Receiver Operating Characteristic (ROC) curves for IoT-LV, RBM, SVM, RF, LSTM, and the proposed framework. The Area Under the ROC Curve (ROC-AUC) evaluates classification performance across different decision thresholds. TPR denotes True Positive Rate, and FPR denotes False Positive Rate. Please click here to view a larger version of this figure.
ANOVA-based statistical significance analysis
To further validate the experimental results, a one-way analysis of variance (ANOVA) was performed to compare the performance of the proposed framework with IoT-LV, RBM, SVM, RF, and LSTM. Whereas the paired t-test evaluates differences between two methods, ANOVA determines whether statistically significant differences exist among multiple groups simultaneously. The ANOVA analysis was conducted using repeated experiments with different train-validation-test partitions of the Open LV Network & Smart Meter Dataset. The mathematical derivations are presented in Supplementary File 1 (Section S8).
The ANOVA results are presented in Table 7. A large F-statistic (approximately 18.7) with p < 0.001, obtained from 30 observations across five independent experimental runs, demonstrated statistically significant differences among the evaluated methods.
| Source of Variation | SS | df | MS | F-value | p-value |
| Between Groups | 412.6 | 5 | 82.52 | 18.7 | < 0.001 |
| Within Groups | 132.4 | 30 | 4.41 | — | — |
| Total | 545 | 35 | — | — | — |
Table 7: One-Way analysis of variance (ANOVA) results. Results of the one-way ANOVA performed to evaluate statistical differences among the compared methods. The table reports the F-statistic, degrees of freedom (df), and corresponding p-values.
To identify the specific methods responsible for these differences, Tukey's honestly significant difference (HSD) post-hoc analysis was subsequently performed. The results are summarized in Table 8 and showed that the proposed LV Acceptance Automation Platform significantly outperformed IoT-LV, RBM, SVM, and RF while also achieving statistically significant improvements over LSTM. These findings further confirm the effectiveness of integrating workflow orchestration, feature engineering, hybrid rule-based and machine learning decision-making, and cloud-edge computing within a unified acceptance automation framework.
| Comparison | p-value | Significance |
| Proposed vs IoT-LV | < 0.001 | Significant |
| Proposed vs RBM | < 0.001 | Significant |
| Proposed vs SVM | 0.009 | Significant |
| Proposed vs RF | 0.018 | Significant |
| Proposed vs LSTM | 0.041 | Significant |
Table 8: Tukey honestly significant difference (HSD) post-hoc analysis. Pairwise comparisons among the evaluated methods using the Tukey HSD post-hoc test following ANOVA. The table reports mean differences, confidence intervals (CI), adjusted p-values, and statistical significance.
Confidence interval analysis
The reliability and consistency of the proposed framework were further evaluated by calculating 95% confidence intervals (CI) for the classification accuracy obtained from repeated experiments. The computational procedure is described in Supplementary File 1 (Section S8).
The confidence interval analysis is summarized in Table 9. Conventional approaches, including IoT-LV and RBM, exhibited relatively wide confidence intervals, indicating greater variability in classification performance. In contrast, the machine learning-based methods (SVM, RF, and LSTM) demonstrated narrower confidence intervals, reflecting improved stability across repeated experiments.
| Method | Mean Accuracy (%) | Std. Dev (%) | 95% Confidence Interval (%) |
| IoT-LV | 82.4 | 1.8 | [81.0, 83.8] |
| RBM | 85.7 | 1.6 | [84.5, 86.9] |
| SVM | 91.6 | 1.3 | [90.6, 92.6] |
| RF | 93.4 | 1.2 | [92.5, 94.3] |
| LSTM | 95.2 | 1 | [94.5, 95.9] |
| Proposed Method | 97.3 | 0.8 | [96.7, 97.9] |
Table 9: Confidence intervals for classification accuracy. Mean classification accuracy, standard deviation (SD), and 95% confidence interval (CI) for each evaluated method based on five independent experimental runs. The table summarizes the statistical variability and reliability of the classification performance.
The proposed LV Acceptance Automation Platform achieved the highest mean classification accuracy (97.3%) together with the narrowest 95% confidence interval (96.7%–97.9%), indicating excellent consistency and robustness. Furthermore, the absence of overlap between the confidence interval of the proposed framework and those of the benchmark methods provides additional evidence of its statistically superior performance.
Forecasting performance
In addition to classification performance, the forecasting capability of the proposed LV Acceptance Automation Platform was evaluated using the root mean square error (RMSE) and mean absolute error (MAE). These metrics quantify the accuracy of energy consumption prediction by measuring the deviation between the predicted and observed values.
The forecasting performance of the proposed framework and the benchmark methods is summarized in Table 10. The corresponding confusion matrix heatmaps for the evaluated classification methods are presented in Figure 5, illustrating the distributions of true positive, true negative, false positive, and false negative predictions. The proposed framework achieved the lowest prediction errors, with an RMSE of 2.1 and an MAE of 1.6, demonstrating superior forecasting performance compared with IoT-LV, RBM, SVM, RF, and LSTM. Prediction errors progressively decreased from the conventional IoT-LV and RBM approaches to the machine learning-based SVM, RF, and LSTM models, highlighting the advantages of intelligent learning techniques. The proposed framework further reduced prediction errors by integrating advanced feature engineering, workflow orchestration, hybrid rule-based and machine learning decision-making, and cloud-edge collaborative processing, thereby improving the accuracy and reliability of energy consumption forecasting. Operational decision latency for the evaluated methods is compared in Figure 6. The proposed framework achieved the shortest decision latency, demonstrating the effectiveness of workflow orchestration and cloud-edge collaborative processing.
| Method | RMSE | MAE |
| IoT-LV | 8.9 | 7.2 |
| RBM | 7.6 | 6.1 |
| SVM | 5.1 | 4.3 |
| RF | 4.2 | 3.5 |
| LSTM | 3.4 | 2.8 |
| Proposed Method | 2.1 | 1.6 |
Table 10: Comparison of forecasting error metrics (RMSE and MAE). Comparison of the forecasting performance of IoT-LV, RBM, SVM, RF, LSTM, and the proposed framework using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). Lower RMSE and MAE values indicate smaller prediction errors.
Figure 5: Comparison of confusion matrix heatmaps. Confusion matrices for IoT-LV, RBM, SVM, RF, LSTM, and the proposed framework showing classification outcomes in terms of True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). The heatmaps illustrate the distribution of correctly and incorrectly classified samples for each method. Please click here to view a larger version of this figure.
Figure 6: Comparison of operational decision latency across different methods. Comparison of the operational decision latency of IoT-LV, RBM, SVM, RF, LSTM, and the proposed framework during LV acceptance automation. Latency is measured in milliseconds (ms) and represents the operational decision time required for workflow execution, rule validation, and acceptance decision generation. Please click here to view a larger version of this figure.
Ablation study
The contribution of each major component of the proposed LV Acceptance Automation Platform was evaluated through an ablation study. The framework was progressively constructed by enabling feature engineering, intelligent machine-learning decision-making, workflow orchestration, and cloud-edge integration. The baseline configuration, in which none of these modules was enabled, achieved an accuracy of 82.6%, an F1-score of 80.4%, an RMSE of 8.7, and an operational latency of 96 ms, demonstrating the limitations of conventional processing without intelligent automation.
The ablation results are summarized in Table 11. Progressive incorporation of feature engineering substantially improved classification performance by enhancing the representation of operational characteristics extracted from the smart meter data. Adding the intelligent decision module further increased classification accuracy and reduced prediction errors through data-driven decision-making. Integration of workflow orchestration improved automation efficiency by coordinating validation, execution, and decision processes, while cloud-edge collaboration reduced operational decision latency through distributed processing. Collectively, these modules produced consistent improvements in accuracy and F1-score while simultaneously reducing RMSE and execution latency.
The complete LV Acceptance Automation Platform achieved the best overall performance, with a classification accuracy of 97.3%, an F1-score of 96.4%, an RMSE of 2.1, and an operational latency of 31 ms. These findings demonstrate that the observed performance improvements result from the integration of workflow orchestration, advanced feature engineering, hybrid rule-based and machine-learning decision-making, and cloud-edge collaboration, rather than from any individual component alone. The fault detection performance of the evaluated methods is presented in Figure 7, where the proposed framework achieved the highest fault detection rate. Figure 8 illustrates the scalability analysis, demonstrating that the proposed framework maintained the lowest processing time as the dataset size increased, confirming its suitability for large-scale LV acceptance automation.
| Configuration | Workflow Orchestration | Feature Engineering | Cloud–Edge Processing | Intelligent ML Decision Module | Accuracy (%) | F1-Score (%) | RMSE | Latency (ms) |
| Baseline System | ✗ | ✗ | ✗ | ✗ | 82.6 | 80.4 | 8.7 | 96 |
| + Feature Engineering | ✗ | ✓ | ✗ | ✗ | 88.9 | 87.5 | 6.2 | 84 |
| + ML Decision Module | ✗ | ✓ | ✗ | ✓ | 92.8 | 91.6 | 4.5 | 63 |
| + Workflow Orchestration | ✓ | ✓ | ✗ | ✓ | 95.1 | 94.2 | 3.3 | 48 |
| + Cloud–Edge Integration | ✓ | ✓ | ✓ | ✓ | 97.3 | 96.4 | 2.1 | 31 |
Table 11: Ablation study of the proposed framework. Experimental evaluation of the contribution of feature engineering, intelligent decision-making, workflow orchestration, and cloud-edge integration to the overall performance of the proposed framework. Performance is reported for each incremental configuration using Accuracy, F1-score, RMSE, and Operational Decision Latency.
Figure 7. Comparison of fault detection rate across different methods. Comparison of the fault detection performance of IoT-LV, RBM, SVM, RF, LSTM, and the proposed framework using the Open LV Network & Smart Meter dataset. Fault Detection Rate (%) represents the proportion of correctly identified fault conditions. Please click here to view a larger version of this figure.
Figure 8. Scalability analysis of different methods. Comparison of the scalability of IoT-LV, RBM, SVM, RF, LSTM, and the proposed framework under increasing dataset sizes. Scalability is evaluated using processing time (ms) as computational workload increases. Please click here to view a larger version of this figure.
DATA AVAILABILITY:
The Open LV Network & Smart Meter dataset used in this study is publicly available through the UK Power Networks Open Data Portal42 at: https://ukpowernetworks.opendatasoft.com/explore/assets/ukpn-smart-meter-consumption-lv-feeder/. All study-generated resources, including the simulated smart meter dataset, example workflow configurations, deployment artifacts, research prototype implementation, source code, and evaluation scripts, are publicly available through the Zenodo repository at: https://zenodo.org/records/21187749.
Supplementary Figure 1: Flowchart of the workflow orchestration engine. The diagram illustrates the sequence of operations performed by the workflow orchestration engine, including data acquisition, preprocessing, workflow execution, machine learning-based validation, decision generation, and result reporting for automated low-voltage acceptance testing.Please click here to download this file.
Supplementary Figure 2: Cloud–edge architecture for low-voltage acceptance automation. The architecture illustrates the interaction between edge devices and cloud services. Electrical measurements collected from smart meters, SCADA systems, and IoT devices are processed at the edge for preprocessing and real-time inference, then transmitted to the cloud for storage, analytics, model management, and visualization via an interactive monitoring dashboard.Please click here to download this file.
토론
실험적 평가 결과, 제안된 워크플로우 지향적 저압(LV) 수락 자동화 플랫폼이 분류, 예측, 결함 검출, 운영 결정 지연 시간 및 확장성 측면에서 벤치마크 방식보다 일관되게 우수한 성능을 보였다. 이러한 성능 향상은 단일 머신러닝 모델의 사용보다는 고급 특성 공학, 워크플로우 오케스트레이션, 클라우드-엣지 협업, 그리고 규칙 기반과 머신러닝이 결합된 하이브리드 의사결정의 통합에서 비롯되었다. 이러한 결과는 워크플로우 지향적 자동화와 지능형 IoT 아키텍처가 산업 및 스마트 그리드 응용 분야에서 운영 효율성, 상호 운용성 및 의사결정 지원을 개선한다는 이전 연구 결과들과 일치한다2,3,9,10,18.
통계 분석을 통해 실험 결과의 신뢰성을 더욱 강화하였습니다. 쌍별 유의성 검정, 일원 분산 분석(ANOVA), 신뢰 구간 분석 및 ROC(Receiver Operating Characteristic) 분석을 통해, 제안된 프레임워크가 다양한 결정 임계값에서 우수한 분류 능력을 유지하면서 벤치마크 방법 대비 통계적으로 유의미한 개선을 달성했음을 확인하였습니다47,48,49. 이러한 통계적 평가는 관찰된 성능 향상이 무작위 변동에 의한 것이 아니라는 추가적인 근거를 제공합니다.
확장성 및 어블레이션 연구(ablation studies)를 통해 계산 부하가 증가하고 시스템 구성이 달라지는 상황에서도 제안된 아키텍처의 효과가 더욱 입증되었습니다. 확장성 분석 결과, 데이터셋 크기가 1,000–5,000개 샘플로 증가함에 따라 본 프레임워크가 벤치마크 방법들보다 더 낮은 처리 시간을 유지하는 것으로 나타났습니다. 마찬가지로, 어블레이션 연구를 통해 각 주요 구성 요소가 전체 시스템 성능에 기여함을 확인했습니다. 특성 공학(Feature engineering)은 스마트 미터 데이터에서 추출된 운전 특성의 표현력을 향상시켰으며, 지능형 결정 모듈은 고장 예측 및 자동 검증 기능을 강화했고, 워크플로우 오케스트레이션은 작업 스케줄링 및 실행을 최적화했으며, 클라우드-엣지 협업은 통신 오버헤드와 운전 결정 지연 시간을 단축시켰습니다. 이러한 관찰 결과는 확장 가능하고 지연 시간이 낮은 산업용 IoT 시스템을 구현하는 데 있어 클라우드-엣지 아키텍처와 프로세스 오케스트레이션의 장점을 강조한 기존 보고들9,15,19,20,21과 일치합니다.
제안된 프레임워크는 IoT 기반의 LV 환경 내에서 작동하므로, 실제 배포를 위해서는 사이버 보안과 데이터 프라이버시가 중요한 고려 사항입니다. 스마트 미터, 에지 디바이스, 게이트웨이 및 클라우드 서비스 간의 보안 통신은 Transport Layer Security (TLS)가 적용된 MQTT 및 HTTPS 프로토콜을 사용하여 구현할 수 있습니다. 역할 기반 액세스 제어 (RBAC)는 워크플로 관리와 시스템 리소스를 위한 인증 및 권한 부여를 제공하며, Advanced Encryption Standard (AES)-256 암호화는 전송 및 저장 과정에서 운용 데이터를 보호합니다. 감사 로깅과 보안 클라우드-에지 동기화는 수락 워크플로 전반에 걸쳐 데이터 무결성, 기밀성 및 추적성을 추가적으로 지원합니다. 이러한 보안 메커니즘은 보안 산업용 사물 인터넷 (IIoT) 및 스마트 그리드 통신 인프라에 대한 현재의 권장 사항과 일치합니다6,7,19,20,21.
제안된 프레임워크는 클라우드-에지 협업, 워크플로우 오케스트레이션, 특성 공학 및 머신러닝 기반 의사결정을 단일 자동화 플랫폼 내에 통합하는 통합 워크플로우 지향 아키텍처를 도입함으로써 지능형 전력 배전 시스템의 발전에 기여합니다. 주로 데이터 획득과 중앙 집중식 처리에 중점을 두는 기존의 IoT 기반 모니터링 방식1,10,11,12와 달리, 제안된 프레임워크는 자동화된 검증, 지능형 고장 감지 및 프로세스 기반 의사결정을 지원합니다. 확장성 및 어블레이션 연구는 향후 스마트 에너지 관리 및 Industry 4.0 응용 분야를 위한 모듈형 지능형 아키텍처의 이점을 더욱 입증합니다3,9,14,15.
실무적인 관점에서, 제안된 LV 수락 자동화 플랫폼은 최소한의 인적 개입으로 실시간 모니터링, 자동 검증, 지능형 결함 검출 및 수락 의사결정을 가능하게 합니다. 달성된 31 ms의 운영 의사결정 지연 시간과 98.4%의 결함 검출률은 실시간 산업 응용 분야를 지원하는 해당 프레임워크의 역량을 입증합니다. 또한, 클라우드-엣지 협업 아키텍처는 통신 오버헤드를 줄이는 동시에 스마트 미터, 분산 에너지 자원 및 IoT 기반 장치에서 수집된 데이터를 처리하는 데 필요한 확장성을 제공합니다. 이러한 기능들은 스마트 그리드 자동화 및 지능형 에너지 관리 시스템의 현재 추세와 일치합니다18,19,20,21,27,29,30.
유망한 결과에도 불구하고, 몇 가지 한계점을 인정해야 합니다. 실험적 평가는 Open LV Network를 사용하여 수행되었습니다. &공개적으로 이용 가능한 저압(LV) 데이터셋은 포괄적인 결함 주석을 제공하지 않기 때문에, 스마트 미터 데이터셋과 합성 결함 데이터를 함께 사용하였습니다. 결과적으로, 시뮬레이션된 결함 시나리오와 규칙 기반 레이블이 실제 운영 조건의 다양성을 완전히 대변하지 못할 수 있습니다. 또한, 단일 데이터셋을 사용하여 검증을 수행하였으며, 실제 배포 시에는 통신 제약, 센서 오류 및 이기종 네트워크 조건에서 작동하는 엣지 디바이스의 계산 능력 제한으로 인해 영향을 받을 수 있습니다. 제안된 아키텍처 내에서 사이버 보안 메커니즘이 고려되었으나, 이에 대한 구현 및 평가는 본 연구의 범위를 벗어납니다. 따라서 향후 연구는 독립적인 실제 데이터셋, 다양한 저압 네트워크 토폴로지 및 현장 배포 시나리오를 사용하여 프레임워크를 검증하는 데 집중할 것입니다. 추가 연구를 통해 적응형 지능형 에너지 관리, 설명 가능한 인공지능, 디지털 트윈 지원 모니터링 및 강화된 클라우드-엣지 협업을 조사하여 지능형 저압 수전 자동화 시스템의 강건성, 확장성 및 해석 가능성을 더욱 향상시킬 계획입니다.40,41.
공개 사항
저자들은 경쟁 관계에 있는 재정적 또는 비재정적 이해관계가 없음을 선언합니다.
감사의 글
The authors thank the UK Power Networks Open Data Portal for providing the publicly available Open LV Network & Smart Meter dataset used in this study. This research received no external funding.
재료
| 이름 | 회사 | 카탈로그 번호 | 댓글 |
|---|---|---|---|
| Apache Spark | Apache Software Foundation | Version 3.4+ | 대규모 데이터 분석 |
| Apache Superset | Apache Software Foundation | Version 3.x | 분석 대시보드 |
| Docker Engine | Docker Inc. | Version 24.x | 애플리케이션 배포 |
| Eclipse Mosquitto MQTT Broker | Eclipse Foundation | Version 2.x | 데이터 교환 |
| EdgeX Foundry (Jakarta Release or later) | LF Edge (Linux Foundation) | Version Jakarta / Geneva | 에지 장치 관리 |
| Flask / FastAPI | Pallets Projects / Sebastián Ramírez | Flask 2.3+ / FastAPI 0.100+ | 백엔드 API 통합 |
| Grafana | Grafana Labs | Version 10.x | 시각화 및 모니터링 |
| Industrial IoT Gateway | Advantech / Siemens | ECU-1251 / SIMATIC IOT2040 | 센서-클라우드 통신 |
| IoT Sensors (Voltage/Current Sensors) | Texas Instruments / Analog Devices | INA219 / ACS712 | 실시간 네트워크 모니터링 |
| Jetson Nano Developer Kit | NVIDIA Corporation | 945-13450-0000-100 | 에지 AI 추론 |
| Kubernetes | Cloud Native Computing Foundation (CNCF) | Version 1.27+ | 워크플로우 확장성 |
| LSTM Model (Keras/TensorFlow) | Google (TensorFlow/Keras) | Included in TF 2.12+ | 부하 예측 및 전망 |
| PLC (Power Line Communication) Module | STMicroelectronics | ST7580 | 스마트 그리드 통신 |
| Plotly Dash | Plotly Inc. | Version 2.x | 대화형 시각화 |
| Python 3.10 | Python Software Foundation | Version 3.10.x | 데이터 처리 및 분석 |
| PyTorch | Meta AI | Version 2.0+ | 딥러닝 구현 |
| Random Forest Algorithm | Scikit-learn | Included in v1.3+ | 결함 분류 |
| Raspberry Pi 4 Model B (4GB RAM) | Raspberry Pi Foundation | RPI4-MODBP-4GB | 에지 프로세싱 |
| React.js / Angular | Meta / Google | React 18.x / Angular 16+ | 사용자 인터페이스 개발 |
| RS-485 Transceiver Module | Maxim Integrated (Analog Devices) | MAX485 | 산업용 장치 연결성 |
| Scikit-learn | Scikit-learn Developers | Version 1.3+ | 머신러닝 모델 |
| Support Vector Machine (SVM) | Scikit-learn | Included in v1.3+ | 분류 및 검증 |
| TensorFlow | Google LLC | Version 2.12+ | 모델 학습 |
| Twilio / Firebase Cloud Messaging | Twilio Inc. / Google | 해당 없음 | 경고 및 알림 |
| Wi-Fi Module | Espressif Systems | ESP8266 / ESP32 | 데이터 전송 |
| XGBoost | DMLC (Distributed ML Community) | Version 1.7+ | 지능형 의사결정 지원 |
참고문헌
- Tong L, Zhou J, Li J. IoT-based low-voltage power distribution system management and control platform. Front Energy Res. 2022;10:902715. https://doi.org/10.3389/fenrg.2022.902715
- Babayigit B, Abubaker M. Industrial Internet of Things: A review of improvements over traditional SCADA systems for industrial automation. IEEE Syst J. 2023;18:120–133.
- Nakagawa EY, et al. Industry 4.0 reference architectures: State of the art and future trends. Comput Ind Eng. 2021;156:107241. https://doi.org/10.1016/j.cie.2021.107241
- Ladegourdie M, Kua J. Performance analysis of OPC UA for industrial interoperability towards Industry 4.0. Internet Things. 2022;3:507–525.
- Bader SR, Maleshkova M, Lohmann S. Structuring reference architectures for the industrial Internet of Things. Future Internet. 2019;11:151. https://doi.org/10.3390/fi11070151
- Fun TS, Samsudin A. Recent technologies, security countermeasure and ongoing challenges of industrial Internet of Things (IIoT): A survey. Sensors (Basel). 2021;21:6647. https://doi.org/10.3390/s21196647
- Mirani AA, et al. Key challenges and emerging technologies in industrial IoT architectures: A review. Sensors (Basel). 2022;22:5836. https://doi.org/10.3390/s22155836
- Fraile F, et al. Reference models for digital manufacturing platforms. Appl Sci. 2019;9:4433. https://doi.org/10.3390/app9204433
- Serôdio C, et al. Software and architecture orchestration for process control in Industry 4.0 enabled by cyber-physical systems technologies. Appl Sci. 2024;14:2160. https://doi.org/10.3390/app14052160
- Ungurean I, Gaitan NC. A software architecture for the industrial Internet of Things—A conceptual model. Sensors (Basel). 2020;20:5603. https://doi.org/10.3390/s20195603
- Hsiao CH, Lee WP. OPIIoT: Design and implementation of an open communication protocol platform for industrial Internet of Things. Internet Things. 2021;16:100441. https://doi.org/10.1016/j.iot.2021.100441
- Domínguez-Bolaño T, et al. An overview of IoT architectures, technologies, and existing open-source projects. Internet Things. 2022;20:100626. https://doi.org/10.1016/j.iot.2022.100626
- Sasiain J, et al. Towards flexible integration of 5G and IIoT technologies in Industry 4.0: A practical use case. Appl Sci. 2020;10:7670. https://doi.org/10.3390/app10217670
- Cañas H, et al. A conceptual framework for smart production planning and control in Industry 4.0. Comput Ind Eng. 2022;173:108659. https://doi.org/10.1016/j.cie.2022.108659
- Cabrini FH, et al. Enabling the industrial Internet of Things to cloud continuum in a real city environment. Sensors (Basel). 2021;21:7707. https://doi.org/10.3390/s21227707
- Gabriel MF, et al. Implementation of an IoT architecture for promoting healthy air quality in 84 homes of families with children. Build Environ. 2024;266:112040. https://doi.org/10.1016/j.buildenv.2024.112040
- Silva BN, Khan M, Han K. Internet of Things: A comprehensive review of enabling technologies, architecture, and challenges. IETE Tech Rev. 2018;35:205–220.
- Dhaou ISB, et al. Internet of Things technologies for smart grid. In: Research Anthology on Smart Grid and Microgrid Development. IGI Global; Hershey; 2022. p. 805–832.
- Kondoro A, et al. A low latency secure communication architecture for microgrid control. Energies. 2021;14:6262. https://doi.org/10.3390/en14196262
- Tightiz L, Yang H. A comprehensive review on IoT protocols' features in smart grid communication. Energies. 2020;13:2762. https://doi.org/10.3390/en13112762
- Kondoro A, et al. Real-time performance analysis of secure IoT protocols for microgrid communication. Future Gener Comput Syst. 2021;116:1–12.
- Rodrigues R, et al. A review of solid-state circuit breakers. IEEE Trans Power Electron. 2020;36:364–377.
- IEEE. IEEE Std 1159-2019: IEEE recommended practice for monitoring electric power quality. IEEE; Piscataway (NJ); 2019. https://doi.org/10.1109/IEEESTD.2019.8800097
- Otcenasova A, et al. The influence of power quality indices on active power losses in a local distribution grid. Energies. 2019;12:1389. https://doi.org/10.3390/en12071389
- Ma K, et al. Review of distribution network phase unbalance: Scale, causes, consequences, solutions, and future research direction. CSEE J Power Energy Syst. 2020;6:479–488.
- Peppanen J, et al. Modeling power electronic equipment to improve distribution feeder current and voltage balance [Presentation]. Presented at: IEEE PES Transmission and Distribution Conference and Exposition (T&D); Denver, CO; 2018. Available from: https://ieeexplore.ieee.org/ https://doi.org/10.1109/TDC.2018.8440261
- Sayed S, et al. Design and realization of an open-source and modular smart meter. Energy Sci Eng. 2019;7:1405–1422.
- Satapathy LM, et al. Arduino based home automation using Internet of Things (IoT). Int J Pure Appl Math. 2018;118:769–778.
- Shaban M, et al. Design of a partially grid-connected photovoltaic microgrid using IoT technology. Appl Sci. 2021;11:11651. https://doi.org/10.3390/app112411651
- Shaban M, Dhaou IB. Design of an IoT-enabled microgrid architecture for a partial grid-connected mode [Presentation]. Presented at: 18th International Multi-Conference on Systems, Signals and Devices (SSD); Virtual; 2021. Available from: https://ieeexplore.ieee.org/ https://doi.org/10.1109/SSD52085.2021.9429443
- Taye BZ, et al. Rural electrification planning using geographic information system (GIS). Cogent Eng. 2020;7:1836730. https://doi.org/10.1080/23311916.2020.1836730
- Huld T, et al. Geospatial analysis of photovoltaic mini-grid system performance. Energies. 2017;10:218. https://doi.org/10.3390/en10020218
- Domínguez J, et al. IntiGIS-Local: A geospatial approach to assessing rural electrification alternatives for sustainable socio-economic development in isolated communities—A case study of Guasasa, Cuba. Energies. 2024;17:3835. https://doi.org/10.3390/en17153835
- Tafula JE, et al. Multicriteria decision-making approach for optimum site selection for off-grid solar photovoltaic microgrids in Mozambique. Energies. 2023;16:2894. https://doi.org/10.3390/en16062894
- Isihak S, et al. Evolution of GIS-based rural electrification planning models and an application of OnSSET in Nigeria. Renew Sustain Energy Transit. 2022;2:100019. https://doi.org/10.1016/j.rset.2022.100019
- Korkovelos A, et al. The role of open access data in geospatial electrification planning and the achievement of SDG7: An OnSSET-based case study for Malawi. Energies. 2019;12:1395. https://doi.org/10.3390/en12071395
- Cicilio P, et al. uGrid: Reliable minigrid design and planning toolset for rural electrification. IEEE Access. 2019;7:163988–163999.
- Nasir M, et al. Optimal planning and design of low-voltage low-power solar DC microgrids. IEEE Trans Power Syst. 2018;33:2919–2928.
- Vai V, et al. Optimal low-voltage distribution topology with integration of PV and storage for rural electrification in developing countries: A case study of Cambodia. J Mod Power Syst Clean Energy. 2020;8:531–539.
- Banad Y, Sharif S, Rezaei Z. Artificial intelligence and machine learning for smart grids: from foundational paradigms to emerging technologies with digital twin and large language model-driven intelligence. Energy Convers Manag X. 2025;28:101329. https://doi.org/10.1016/j.ecmx.2025.101329
- Álvarez-López C, González-Briones A, Li T. Explainable AI and multi-agent systems for energy management in IoT-edge environments: a state-of-the-art review. Electronics. 2026;15(2):385. https://doi.org/10.3390/electronics15020385
- UK Power Networks. Open data portal [Online]. Available from: https://www.ukpowernetworks.co.uk/our-company/open-data-portal (Accessed June 2026).
- Jolliffe IT. Principal Component Analysis. 2nd ed. Springer; New York; 2002. https://doi.org/10.1007/b98835
- Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20:273–297.
- Breiman L. Random forests. Mach Learn. 2001;45:5–32.
- Hochreiter S, Schmidhuber J. Long short-term memory. Neural Comput. 1997;9:1735–1780.
- Fawcett T. An introduction to ROC analysis. Pattern Recognit Lett. 2006;27:861–874.
- Montgomery DC, Runger GC. Applied Statistics and Probability for Engineers. 8th ed. John Wiley & Sons; Hoboken (NJ); 2020.
- Montgomery DC. Design and Analysis of Experiments. 10th ed. John Wiley & Sons; Hoboken (NJ); 2020.