An integrated AI-based software architecture for load forecasting, controlled anomaly detection, and predictive energy management
Artykuł w czasopiśmie
MNiSW
20
Lista 2024
| Status: | |
| Autorzy: | Alibekkyzy Karlygash, Massabayeva Gulnur, Majka Michał, Bazarova Madina, Kaliyev Daniyar, Keribayeva Talshyn, Belginova Saule, Bugubayeva Alina |
| Dyscypliny: | |
| Aby zobaczyć szczegóły należy się zalogować. | |
| Rok wydania: | 2026 |
| Wersja dokumentu: | Drukowana | Elektroniczna |
| Język: | angielski |
| Wolumen/Tom: | 8 |
| Strony: | 1 - 23 |
| Impact Factor: | 3,4 |
| Scopus® Cytowania: | 0 |
| Bazy: | Scopus |
| Efekt badań statutowych | NIE |
| Materiał konferencyjny: | NIE |
| Publikacja OA: | TAK |
| Licencja: | |
| Sposób udostępnienia: | Witryna wydawcy |
| Wersja tekstu: | Ostateczna wersja opublikowana |
| Czas opublikowania: | W momencie opublikowania |
| Data opublikowania w OA: | 15 października 2026 |
| Abstrakty: | angielski |
| The increasing deployment of smart meters and Internet of Things technologies has created a demand for intelligent computational systems capable of jointly forecasting electricity consumption, detecting abnormal behavior, and optimizing load management. This study proposes an integrated artificial intelligence-based framework for automated energy metering and consumption management in smart-grid environments. The framework combines a Long Short-Term Memory network for short-term load forecasting, an autoencoder-based model for anomaly detection, and Model Predictive Control for demand-side optimization. The system architecture includes data acquisition, communication, preprocessing, predictive analytics, anomaly monitoring, and decision-support layers. The primary empirical evaluation used 4,380 measured hourly observations (approximately 6 months) from a single administrative building. A separately flagged synthetic temporal extension was retained only as an auxiliary dataset and was excluded from all primary performance metrics. Across five independent model initializations, the final LSTM achieved a MAPE of 3.423 ± 0.555%, an RMSE of 0.336 ± 0.048 kW, and an R2 of 0.973 ± 0.007 on the 633-sequence independent measured test subset. The synthetic extension was maintained as a separately identified auxiliary component and was excluded from all primary training, validation, and independent test metrics. Using a validation-selected reconstruction-error threshold, the autoencoder detected 12 of 30 controlled anomalies in the independent test subset, yielding 100.00% precision, 40.00% recall, and an F1-score of 57.14%. Under the same controlled protocol, One-Class SVM achieved higher recall (53.33%) and F1-score (69.57%), indicating that the autoencoder should be interpreted as an architectural component rather than as the best-performing anomaly detector for the present dataset. A historical proof-of-concept MPC simulation, retained for architectural illustration, reported a 15% reduction in simulated peak demand and a 12% reduction in estimated electricity cost; these values are not interpreted as measured outcomes from the administrative-building experiment. Controlled anomaly injection was used to establish reproducible validation and test labels for the anomaly-detection component. The results demonstrate that the joint integration of forecasting, anomaly detection, and predictive control provides a computationally coherent framework for supporting energy-management efficiency, operational reliability, and data-driven decision-making in intelligent energy systems. |
