Discriminative Scanner Enhanced Selective Partial Aggregation-Based Federated Learning for Resilient and Efficient Wind Power Forecasting

Federated learning (FL)-based wind power forecasting has become a promising solution in recent years for collaborative deep learning (DL) models while preserving clients’ data privacy, resulting in improved forecasting accuracy. However, these forecasting schemes can be energy-consuming and susceptible to cyber-attacks, that require further research and investigation. This work proposes a computationally efficient and resilient FL scheme that reduces energy usage requirements while protecting the model parameters against single- and multi-client targeted denial of service (DoS) and poisoning attacks. A selective partial aggregation (SPA) scheme is proposed to optimize energy usage by reducing the number of clients’ models for aggregation in FL. Additionally, a discriminative scanner-based pre-processor (DiSP) is proposed to detect and remove corrupted or missing models for robust forecasting. Seven DL models, including four benchmark models, are investigated for two aggregation schemes to find the optimum forecasting framework. Experimental analysis reveals that cyber-attacks cause significant forecasting accuracy deviations in terms of mean absolute error (MAE) and root mean squared error (RMSE), especially during coordinated cyber-attacks. The proposed scheme can ensure up to 60% computational energy efficiency in model training and transmission to the server with negligible changes in forecasting accuracy, while statistically outperforming two benchmark defensive measures.