Trustworthy Distributed Load Forecasting in Resource-Limited Smart Grids and Buildings via Random Layer Aggregation

Federated learning (FL) is a privacy-preserving method for short-term load forecasting in energy networks. However, current defense mechanisms against adversarial attacks often depend on supplementary machine learning frameworks, such as anomaly detection models or Byzantine-robust aggregators. These frameworks add significant computational overhead, straining edge devices such as smart meters and IoT systems with limited processing power. To solve this issue, we propose a new defense-free framework called federated random layer aggregation (FedRLA). By aggregating only one randomly chosen neural network layer per communication round, FedRLA limits adversarial influence to isolated layers. This reduces attack surfaces by 66% compared to full-model aggregation (FedAvg). Using 8-bit quantization, FedRLA cuts data transmission by 92.97% without accuracy loss (MAE: 0.08 kWh vs. FedAvg’s 0.076 kWh). Under four model poisoning attacks, it reduces forecasting errors by 19%–35% compared to FedAvg. FedRLA also uses 24% less CPU and 13% less memory than frameworks such as FedProx, while training 58% faster. It combines communication efficiency (0.195 MB/round), adversarial robustness (MAE ≤ 0.11 kWh under ϵ = 0.2 DP), and low resource consumption, offering a scalable solution for secure FL in resource-constrained energy networks.