A new deep federated learning model with high generalizability for electrical load forecasting against data integrity attacks

• A new SDAE-BiLSTM-FL model is proposed for STLF under data integrity attacks. • Deep FL clients use SDAE-BiLSTM instead of CNN for robust training and pre-processing. • The proposed model ensures privacy and resiliency against scaling cyber-attacks. • STLF generalization to unseen regions is achieved using global feature extraction. • Performance is validated using 12 state-of-the-art models on real-world data. Precise short-term load forecasting (STLF) is a key issue for power industry utilities and is essential to improving the security and economic performance of energy systems. In this study, a cyber-secure forecasting model with high generalizability, which is a combination of stacked denoising autoencoder (SDAE), bidirectional long short-term memory (BiLSTM), and federated learning (FL) algorithm, is introduced for STLF. The proposed model is called SDAE-BiLSTM-FL. By presenting a new structure of FL, it produces a robust STLF model that, in addition to being generalized to regions from which no training data is available, is resilient against data integrity attacks. The performance evaluation of the proposed model is carried out in two different scenarios (without and with data integrity attacks) using the real-world data of three study regions of Iran with different geographical and meteorological characteristics. In every scenario, the STLF performance of the proposed SDAE-BiLSTM-FL model is compared with the STLF performance of the state-of-the-art models used in recent studies. The results presented in each scenario show the superiority of the proposed SDAE-BiLSTM-FL model in terms of accuracy, generalizability, and cyber-resiliency compared to the state-of-the-art models.