Technology Forecasting: A Secure Solar Power Generation Forecasting Framework for Recurrent Neural Networks
Murat Kuzlu, Salih Sarp, Ferhat Özgür Çatak, Ümit Cali, Yanxiao Zhao · 2024
Solar photovoltaics (PV) power generation forecasting has become more crucial with the high use of solar PV resources and high impact on grid stability. The use of solar PV generation in smart grids causes reliability concerns due to its high dependence on weather conditions. This study demonstrates a secure solar power generation forecasting framework embedded in adversarial attack and training methods for recurrent neural networks (RNNs), i.e. simple RNN, long- and short-term memory (LSTM), bidirectional (BiLSTM), gated recurrent unit (GRU), and LSTM/BiLSTM, with attention using publicly available time-series solar power generation dataset. The main objective of the framework is to provide an environment that allows researchers to evaluate the RNN-based solar PV generation models with adversarial attacks and training methods. Results are evaluated based on the model performance, i.e. the accuracy of the model results against adversarial attacks with and without adversarial training. The findings indicate that RNN-based models are robust against adversarial attacks with an attack power of 0.10 (epsilon) without adversarial training, but they become even more robust with an attack power of 0.15 after adversarial training. This suggests that the adversarial training process significantly improves the defense of RNN-based models in solar power generation against adversarial attacks.