The integration of AI and ML in energy forecasting is pivotal for modern energy management. Federated Learning (FL) stands out by enhancing data privacy and collaboration among distributed energy resources, enabling distributed model training while reducing reliance on centralized servers and data transfers. Despite its merits, FL faces substantial security challenges, particularly from adversarial attacks that can compromise the integrity and reliability of the models. This paper aims to address these security concerns by examining the efficiency of Centralized Federated Learning (CFL) and Decentralized Federated Learning (DFL) for distributed load forecasting. Through comparative analysis utilizing publicly available household datasets for short-term load forecasting, our study reveals that DFL demonstrates superior resilience against adversarial attacks compared to CFL. Notably, our findings indicate that the impact of adversarial model poisoning attacks is confined to the targeted client in DFL, while CFL exhibits broader susceptibility across all clients. When attacked, CFL's averaged client Mean Absolute Error (MAE) increased from 0.076 to 0.22 kWh, whereas DFL maintained a lower MAE of 0.116 kWh. Additionally, we present Decentralized Random Layer Aggregation (DRLA) to augment DFL's robustness, offering further insights into enhancing FL methodologies within energy contexts.

The integration of artificial intelligence (AI) into energy networks significantly advanced short-term forecasting, particularly in smart meter applications. However, as distributed energy resources proliferated and energy systems grew in complexity, traditional centralized approaches to data analysis became insufficient in addressing privacy-preserving challenges. Federated learning (FL) emerged as a promising solution, leveraging distributed data sources while safeguarding user privacy. Nonetheless, FL encountered inherent vulnerabilities to adversarial attacks during model training, undermining its reliability and effectiveness. Existing techniques to eliminate these attacks often required additional frameworks for detection, imposing an added burden on devices. To address this issue, we proposed a novel method called federated random layer aggregation (FedRLA). It aggregated only one randomly chosen neural network layer on the server in a privacy-aware manner, leaving the remaining layers unchanged. FedRLA exhibited superior resilience against adversarial attacks by confining attackers to a single neural network layer. Our simulations, focusing on household energy consumption, demonstrated that FedRLA achieved 3.56 times less data transmission compared to FedAvg during global model training. This enhanced efficiency translated to improved energy usage and resource conservation. Furthermore, FedRLA performed better in the presence of differential privacy under attack and no attack conditions.

Federated Learning (FL) in load forecasting improves predictive accuracy by leveraging data from distributed load networks while preserving data privacy. However, the heterogeneous nature of smart grid load forecasting poses significant challenges to conventional FL methods, which are also unsuitable for resource-constrained devices. To address these issues, we propose a novel Lightweight Single Layer Aggregation (LSLA) framework tailored for smart grid networks. The LSLA framework mitigates the issue of data heterogeneity in load forecasting by emphasizing local learning and partially using updates from local devices for model aggregation. Additionally, the framework is optimized for resource-constrained devices by introducing a stopping criterion during model training and weight quantization. Our results show that after quantization, there is an acceptable degradation of 0.01% in Mean Absolute Error (MAE). Compared to traditional FL, up to a 1000fold reduction in communication overhead is achieved with LSLA. Moreover, with an 8-bit fixed-point data representation of neural network weights, a 75% reduction in storage/memory requirement is achieved, and computational cost is reduced due to the replacement of complex floating-point computational units with simpler fixed-point counterparts. By addressing data heterogeneity and minimizing data storage, computation, and communication overheads, our novel framework is well-suited for resource-constrained devices in smart grid networks.