Security Concerns of Adversarial Attack for LSTM/BiLSTM Based Solar Power Forecasting

Solar photovoltaics (PV) power generation forecasting has become more crucial with the high use of solar PV resources and high impact on grid stability and reliability. The integration of solar PV can cause extreme stability and reliability issues due to its high dependence on weather conditions. This study investigates the impact of adversarial attack and training methods on Long Short-Term Memory (LSTM) and Bidirectional (BiLSTM) based solar power generation forecasts. Results are evaluated for models' responses against adversarial attacks with and without adversarial training in terms of the selected performance metrics. The results show that LSTM and BiLSTM-based models are dramatically vulnerable to adversarial attacks without adversarial training. It is also indicated that the adversarial training method effectively defends the LSTM and BiLSTM-based solar power generation models against adversarial attacks.