DualEncDecoder: Federated Short-Term Load Forecasting under Heterogeneous Data

Sasmita Harini S, Naman Srivastava, Priyanka Nihalchandani, Varun Ojha, Pandarasamy Arjunan

Keywords: Federated Learning, Short term Load Forecasting, Client Sampling, Time Series

Abstract: Accurate short-term load forecasting (STLF) is a critical application for modern energy systems, enabling efficient scheduling, demand response, and grid reliability. Traditional centralized approaches raise privacy concerns and often degrade under heterogeneous building-level data. To address these challenges, this study explores federated learning (FL) with a novel \textit{DualEncDecoder} architecture, a Difficulty-Aware Sampling (DAS) strategy, and an autoencoder-based clustering method for client grouping. Experimental results show that DualEncDecoder consistently outperforms baseline models, achieving the lowest SMAPE, while the clustering improves performance for simpler models such as LSTM and GRU. DualEncDecoder, also surpass centralized models under highly varying client data. Overall, the integration of advanced architectures, adaptive sampling, and representation-driven clustering establishes a robust framework for federated energy forecasting in privacy-sensitive, heterogeneous environments.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 23