Abstract

With the growing shift of conventional power systems toward decentralized smart grid electrical systems, many key challenges have come into action. Data privacy, scalability, and real-time decision making are major concerns. A vast amount of sensitive data is generated due to the widespread adoption of distributed energy resources (DERs), electric vehicles (EVs), and other intelligent devices. Although this data is useful for grid improvement, it also raises concerns about privacy, security, and communication, which, in turn, make centralized machine learning unsuitable. Because of the aforementioned issues, this thesis presents a dual framework contribution. Firstly, it offers a systematic and comprehensive review of the Federated Learning (FL) in smart grids. FL developments from 2010 to 2025 are categorized into five main areas using a novel thematic classification framework. These are 1) foundations, 2) load forecasting and demand response, 3) federated reinforcement learning for control, 4) privacy and security, and 5) emerging trends, including Explainable AI (XAI) and Quantum FL. The proposed work presents a comparative analysis of system models, application-specific goals, and technical methods. Various barriers including model heterogeneity, communication latency, and scalability limits are identified followed by the proposing a sustainable three-layer FL architecture for privacy preserving collective intelligence. Secondly, the thesis introduces a two-stage evaluation framework for short term load forecasting (STLF). Deep learning models perform very well, but still their operation under diverse load dynamics remains overlooked. This thesis presents a comparative analysis of frameworks, mathematical models and performance of different STLF-based deep learning models benchmarked against a reference model, namely the persistence baseline. Different models include long short-term memory (LSTM), convolutional neural networks (CNN), a hybrid CNN– LSTM, spatio-temporal convolutional networks (STCN), and an Attention-based CNN–GRU model. A multistep forecasting pipeline is designed that predicts the next 24 hours of demand from the previous one-week load values using the American Electric Power (AEP) hourly load dataset. Using a method inspired by phasor measurement unit (PMU) analysis, the models are evaluated. The focus is on how these models follow the rapid changes in the load in the three scenarios such as normal, ramp (slightly changing), and finally hard (highly variable). Their accuracy and behavior towards the errors are also observed. The models are further evaluated in terms of computational efficiency, including latency and prediction time, to assess their suitability for real-time smart grid applications. Under normal operating conditions, the LSTM model showed the highest accuracy with a mean absolute percentage error (MAPE) of 1.99%. On the other hand, CNN models perform well during ramp events. During the hard case, the hybrid CNN–LSTM remains reliable as load fluctuates heavily, maintaining a MAPE of 3.36%. The Attention CNN–GRU model also demonstrates strongiv performance across scenarios, particularly under dynamic conditions, achieving competitive error values while effectively capturing important temporal dependencies.Beyond standard metrics such as mean absolute error (MAE) and root mean square error (RMSE), the proposed framework gives deeper insight into forecasting behavior. Therefore, it becomes easier to implement these forecasting models in smart-grid environment systems.

Publication Date

5-2026

Document Type

Thesis

Student Type

Graduate

Degree Name

Electrical Engineering (MS)

Department, Program, or Center

Electrical Engineering

Advisor

Jinane Mounsef

Advisor/Committee Member

Haris M. Khalid

Advisor/Committee Member

Abdulla Ismail

Comments

This thesis has been embargoed. The full-text will be available on or around 2/6/2027.

Campus

RIT Dubai

Download

Share

COinS