LIQUIDAI Framework for Smart Grid Optimization: A Dynamic Framework for Distributed Energy Optimization
Mahmud, Md Anisul Islam (2025)
2025
Tietojenkäsittelyopin maisteriohjelma - Master's Programme in Computer Science
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2025-12-17
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2025121711817
https://urn.fi/URN:NBN:fi:tuni-2025121711817
Tiivistelmä
Modern energy systems are facing increasing difficulties concerning scalability, data protection, and adaptability. Traditional centralized methods frequently face difficulties in managing the dynamic characteristics of smart grids, resulting in problems such as delay, model drift, and restricted local decision making capacity. These problems highlight the necessity for decentralized and self-adaptive frameworks capable of functioning efficiently within distributed energy environments.
This thesis introduces a LiquidAI-based framework, where LiquidAI refers to an emerging architecture that enables AI models to operate across cloud and edge environments while continuously adapting to changing conditions. The proposed framework aimed at improving adaptability, decentralization, and continuous learning in the optimization of smart grids. The framework contains necessary components such as edge intelligence, which involves localized computation on lightweight devices; federated learning, enabling collaborative model training without sharing of raw data; and dynamic retraining, which updates models in response to performance declines. To support experimentation, synthetic datasets that match Fingrid standards were generated to imitate authentic grid conditions. Ensemble machine learning models, specifically Random Forest and Gradient Boosting, were trained and implemented on lightweight edge devices utilizing the ONNX format for enabling low-latency inference. A federated learning framework was used for collaborative model enhancement without the exchange of raw data. In addition, a feedback mechanism tracked performance indicators; mean squared error (MSE), root mean squared error (RMSE), and the coefficient of determination (R²) and initiated retraining upon detection of performance decline.
Experimental evaluations showed significant enhancement in both load and solar forecasting accuracy after retraining. These results demonstrate the framework's capacity to adapt to changing operating conditions over time. The smart grid was chosen as a representative use case to demonstrate the framework's capability. However, the exact same concepts can be extended to other distributed, data-driven domains where scalability, security, and self-adaptation are critical.
This thesis introduces a LiquidAI-based framework, where LiquidAI refers to an emerging architecture that enables AI models to operate across cloud and edge environments while continuously adapting to changing conditions. The proposed framework aimed at improving adaptability, decentralization, and continuous learning in the optimization of smart grids. The framework contains necessary components such as edge intelligence, which involves localized computation on lightweight devices; federated learning, enabling collaborative model training without sharing of raw data; and dynamic retraining, which updates models in response to performance declines. To support experimentation, synthetic datasets that match Fingrid standards were generated to imitate authentic grid conditions. Ensemble machine learning models, specifically Random Forest and Gradient Boosting, were trained and implemented on lightweight edge devices utilizing the ONNX format for enabling low-latency inference. A federated learning framework was used for collaborative model enhancement without the exchange of raw data. In addition, a feedback mechanism tracked performance indicators; mean squared error (MSE), root mean squared error (RMSE), and the coefficient of determination (R²) and initiated retraining upon detection of performance decline.
Experimental evaluations showed significant enhancement in both load and solar forecasting accuracy after retraining. These results demonstrate the framework's capacity to adapt to changing operating conditions over time. The smart grid was chosen as a representative use case to demonstrate the framework's capability. However, the exact same concepts can be extended to other distributed, data-driven domains where scalability, security, and self-adaptation are critical.