Machine Learnign Model for the Detection of Electricity Energy Fraud Using and Edge-Fog Computing Architecture
Download as PPTX, PDF0 likes54 views
This document proposes a machine learning model using an edge-fog computing architecture to detect electric energy fraud from smart meter data. It tests various machine learning algorithms on smart meter data from Mexico and finds that a multi-layer perceptron regression model achieves the lowest error rates. The results indicate the model can adequately detect anomalies in energy consumption and production. While human intervention is still needed, the proposed model is designed to operate on embedded devices with limited computing capabilities. The use of fog computing through data centers also improves forecasting performance.
Machine Learnign Model for the Detection of Electricity Energy Fraud Using and Edge-Fog Computing Architecture
- 1. Machine Learning Model for the Detection of Electric Energy Fraud Using and Edge-Fog Computing Architecture Juan C. Olivares Rojas, Enrique Reyes Archundia, Noel E. Rodríguez Maya, José A. Gutiérrez Gnecchi, Ismael Molina Moreno, Jaime Cerda Jacobo
- 2. The importance of electricity
- 3. Industry 4.0
- 4. ¿Mejora o empeora la seguridad?DX: Digital Transformation Robotics Digital Twins Integration Systems IoT CibersecurityCloud Computing 3D Printing AR/VR Big Data Technologies 4.0
- 5. Smart City / Smart Grid
- 7. PRODEREI 2018-2032 Reduce Non-Thecnical losses at 13.38% 1,407 Millions MXN of Thief and Energy Fraud Non-Technical Losses
- 8. Tampering
- 9. Solution
- 10. Literature Review • It’s a well-know problem • The are different works using different AI/ML/ Analytics Techniques • Most of the works are focused in Big Data and Cloud Computing
- 12. Smart Metering Systems Architecture Edge Fog Cloud
- 13. Implementing a SMS Architecture
- 14. SM1 DC1 DC2 SM2 SM3 SM4 SM5 SM6 SM7 Edge Computing Fog Computing Implementing a SMS Architecture
- 15. SM measurements Consumption/ Production Patterns DC A, SAs, DER Fraud Events Report Machine Learning Machine Learning Consumption/ Production Data Consumption/ Production Patterns Consumption/ Production Data11 2 3 3 Check model accurancy 4 5 6 SMS Edge-Fog Architecture
- 16. Structure of SM Database Timestamp Consumption (kWh) Production (kWh) 01/01/2018 00:15 0.02652 0 01/01/2018 00:30 0 0. 00048 01/01/2018 00:45 0.04563 0.000458 … … … 01/31/2019 23:45 0.23181 0.000475
- 17. Comparative of ML Models Decision Tree Regression (DTR), Linear Regression (LR), Sequential Neural Network (SNN) and MultiLayer Perceptron Regression (MLPR). Scikit-learn and Keras libraries of Python were used.
- 18. Method MAPE SM MAPE DC MAPE DC-SM DTR 16.47% 16.84% 16.51% LR 18.27% 17.97% 17.66% SNN 20.76% 21.12% 20.93% MLPR 13.72% 12.36% 12.29% Test with the complete database
- 19. Method MAPE SM MAPE DC MAPE DC-SM DTR 17.39% 16.23% 16.21% LR 22.17% 21.49% 21.17% SNN 25.33% 26.12% 26.43% MLPR 23.47% 22.13% 20.57% Test with database per year
- 20. Method MAPE SM MAPE DC MAPE DC-SM DTR 16.27% 16.64% 16.19% LR 17.88% 17.71% 17.53% SNN 19.23% 19.12% 19.23% MLPR 12.66% 12.16% 11.99% Test with database per month
- 21. Method MAPE SM MAPE DC MAPE DC-SM DTR 16.99% 17.14% 17.07% LR 18.15% 17.97% 17.93% SNN 20.53% 20.97% 20.99% MLPR 13.28% 12.98% 12.74% Test with database per week
- 22. Method MAPE SM MAPE DC MAPE DC- SM DTR 16.69% 16.73% 16.45% LR 18.05% 17.83% 17.80% SNN 19.53% 20.74% 20.63% MLPR 13.57% 12.99% 12.91% Test with database per day
- 23. This work shows that it is possible to detect energy fraud using data and smart meter processing. The results indicate that the detection of anomalies of consumption/production of electrical energy is adequate. Still, like any forecast model, human intervention plays an essential role in decision making. It is important to emphasize that the proposed model adjusts its parameters and amount of training/fit values so that they are executed properly in embedded devices with limited computing capabilities. Finally, the use of Fog Computing though DC in general terms, improves the results of forecasting. Conclusions
- 24. This work is partially supported by the TecnológicoNacional de México under the grant 8000.20-P and 9002.20-P. Questions? Thanks you so much! juan.or@morelia.tecnm.mx, enrique.ra@morelia.tecnm.mx, noel.rm@zitacuaro.tecnm.mx, jose.gg3@morelia.tecnm.mx, ismael.mm@Morelia.tecnm.mx, jaime.cerda@umich.mx