Similarity-driven truncated aggregation framework for privacy-preserving short term load forecasting

Accurate short-term load forecasting (STLF) is essential for the efficient and reliable operation of power systems, enabling effective scheduling and integration of renewable energy sources. Federated learning (FL) offers a collaborative, privacy-preserving approach for distributed model training by avoiding data sharing among sources. However, existing FL methods for STLF often rely on clustering techniques for highly variable residential data, which struggle to effectively handle data diversity, privacy constraints, and anomalous model updates. This study addresses these concerns and presents a similarity-driven truncated aggregation (SDTA) algorithm designed for STLF at macro-level sub-stations. SDTA enhances model alignment by computing layer-wise cosine similarity among client updates and mitigates outliers through truncated mean aggregation, reducing overfitting and improving robustness. The algorithm integrates differential privacy (DP) mechanisms to protect model updates and applies cosine-similarity-based filtering to safeguard against adversarial attacks. Extensive simulations on real-world substation data validate that SDTA significantly outperforms standard FL algorithms such as federated averaging (FedAVG) and federated distance (FedDist). Under conditions without privacy constraints, SDTA achieves a mean absolute percentage error (MAPE) of 2.63%, surpassing FedAVG and FedDist with MAPE of 2.89% and 3.11%, respectively, with faster convergence. Under strict DP constraints, SDTA maintains high forecasting performance with a MAPE of 4.02%, outperforming FedDist and FedAVG by 9.7% and 20.4%, respectively. Furthermore, SDTA exhibits substantial resilience under adversarial conditions, achieving a MAPE reduction of 20.5% over FedAVG when 40% of edge nodes are compromised. Moreover, the study examines the robustness of SDTA against random client selection scenarios, illustrating its resilience and practical applicability in real-world settings, particularly when client selection rates are below 60%.