Reliable Short-Term Load Forecasting using Robust Federated Split Learning Framework

Federated learning (FL) was developed to address the limitations of centralized machine learning, particularly issues related to data privacy and security threats such as model poisoning. However, FL remains vulnerable to certain types of attacks specifically designed to exploit its architecture, such as model flipping attack. Moreover, FL can incur high communication costs, especially in environments with constrained connectivity. To address these challenges, we adopt Split Learning (SL), a technique designed to reduce communication overhead and improve model training efficiency in settings with limited client resources. Building upon this, we propose Federated Split Learning (FSL), a hybrid framework that combines the advantages of both FL and SL to enhance robustness in distributed load forecasting. Experimental results show that in non-adversarial settings, FSL demonstrates faster early-stage convergence compared to FL, reducing the Mean Absolute Error (MAE) from 10 MW to 5 MW and the Mean Absolute Percentage Error (MAPE) from 100% to 40% within the first 25 communication rounds. In a targeted model flipping attack that affects only one client, FL experiences severe performance degradation, MAE rising to 12 MW and MAPE remaining at 100% for all clients. In contrast, FSL maintains stable performance for unaffected clients, keeping MAE below 5 MW and MAPE under 40% after 10 training rounds. FSL reduces the attack surface to only the affected client.