FEDERATED LEARNING WITH STOCHASTIC GRADIENT DESCENT FOR SMART METER ENERGY FORECASTING

• Published in: Електроніка та інформаційні технологіі Vol. 30; no. 30; pp. 15 - 32
• Main Author: Khushalani, Bharat
• Format: Journal Article
• Language: English
• Published: Ivan Franko National University of Lviv 01.06.2025
• Subjects: decentralized optimization; energy forecasting; federated learning; privacy-preserving machine learning; smart meters; stochastic gradient descent
• ISSN: 2224-087X; 2224-0888; 2224-0888
• DOI: 10.30970/eli.30.2

Background. Smart meters are widely used to monitor household energy consumption and help improve energy efficiency. However, collecting this data in a centralized location raises privacy concerns, as detailed consumption records can reveal sensitive household behavior. Federated learning provides an alternative approach by allowing models to be trained directly on user devices without sending raw data to a central server. Materials and Methods. This study developed a simulation-based framework to test federated learning for forecasting short-term electricity usage. We created synthetic data representing hourly energy consumption for 100 simulated households, incorporating daily usage cycles and household-specific patterns. A simple neural network was trained locally on each household’s data using a standard optimization method, and model updates were shared with a central server to improve a shared global model. Results and Discussion. The federated model achieved forecasting accuracy nearly equal to a traditional centralized model while keeping data private. Key factors affecting performance included how often devices were trained locally before sharing results and how many households participated in each training round. The approach remained accurate even when only half the devices contributed at any time. Compared to non-collaborative models trained independently by each household, the federated approach offered a substantial improvement in prediction accuracy. These findings show that good performance can be achieved while protecting user privacy and using simple models suitable for low-power devices. Conclusions. This work shows that a well-designed simulation with realistic energy usage data can help evaluate federated learning methods under practical constraints. Even simple models, when trained in a decentralized and privacy-preserving way, can offer useful predictions for smart energy systems. The approach is suitable for real-world deployment and can help advance privacy-respecting energy analytics.