Differentially private deep learning for load forecasting on smart grid

Load forecasting is vital for a reliable and sustainable smart grid as it is used to predict the demand and make price adjustment accordingly. Electric consumption data which is gathered from IoT devices like smart meter or smart appliances is a key input to improve the accuracy of the forecasting task. However, this data can leak private information of the householders as the consumption data reflects the behavioral patterns of the individuals. Providing privacy for the data without compromising the utility of the forecast is a challenging problem and this is where the differential privacy comes in to play. In this work, we present a practical implementation of the privacy preserving load forecasting with differential privacy techniques using Tensorflow Privacy library. We show that privacy guarantee for the data can be achieved to varying degrees with a tolerable degradation in the forecast results. We provide privacy-utility tradeoff values in our experiments for different privacy levels.