Privacy-Preserved Framework for Short-Term Probabilistic Net Energy Forecasting

This article develops a differential privacy (DP) model for short-term probabilistic energy forecasting at the low-aggregate level. The method first takes probabilistic forecasting elements from a predictor and injects noise into the mean of the forecast based on the Laplace mechanism to guarantee privacy on the mean. The standard deviation is then carefully perturbed (enlarged) to ensure the forecast to be released contains a desirable confidence interval (CI), here 95%. By doing so, customers' privacy is protected, while the user, such as grid operators or retail providers, receives the forecast containing 95% CI of the original forecast. The predictor used to capture the model and data uncertainties is based on a Bayesian neural network, which is also benchmarked against a Gaussian Process. The simulations are carried out for different levels of privacy,
$\varepsilon$
, and the obtained trend provides decision-makers with a clear idea of determining the appropriate
$\varepsilon$
. The study found that the forecasting error is smoothed out on the used dataset for
$\varepsilon \geq 0.6$
. The proposed model is an output perturbation approach. Accordingly, the obtained results are compared with an input perturbation approach, showing that the proposed DP model gives the best accuracy/privacy tradeoff for users. Furthermore, this study includes rooftop photovoltaic (PV) generation behind the meter, which is one of the recent transitioning challenges in energy forecasting. Detailed analysis reveals that forecasting is more challenging for periods before sunset when PV generation drop coincides with the common changing point of households' activities.