A Cybersecure Distribution-Free Learning Model for Interval Forecasting of Power Load Under Cyberattacks

Reliable interval forecasting of uncertain load with cyberattacks is critical for grid resilience and decision-making. Nevertheless, various constraints on the predictive distribution assumptions are introduced into existing probability prediction models, limiting the model's capability to approximate the actual distribution. Moreover, load forecasting involving cyberattacks may lead to inappropriate decision-making for power transmissions. To this end, a novel cybersecure distribution-free learning model is devised for load forecasting under cyberattacks. First, an isolation forest is implemented to counter load anomalies with cyberattacks automatically by unsupervised learning. Second, an attention-assisted gated recurrent unit embedded in the developed model is employed to extract long- and short-term temporal dependencies of load time series. Then, an optimal prediction interval (PI) construction strategy based on conformal inference is put forward to eliminate the quantile crossing rates, which approximates the actual cumulative distribution function of load data and attains effective marginal coverage. Finally, numerical experiments have demonstrated that the model surpasses existing models in both PI width and cybersecurity.