A Differential Privacy-enhanced Federated Learning Method for Short-Term Household Load Forecasting in Smart Grid

Household Short-term load forecasting is critical to the operation and maintenance of smart grids. As more and more fine-grained electricity consumption data from customers are collected by smart meters deployed in various households, the LSTM recurrent neural network-based load forecasting methods show promising results. However, these methods require the centralized collection and storage of consumers’ electricity consumption data to train forecasting models, which threatens the privacy of customers. Therefore, we proposed a differential privacy-enhanced Federated Learning(DPEFL) method, which collaboratively learns LSTM load forecasting model from data distributed across multiple consumer households. DPEFL provides two levels of privacy protection: first, it keeps the raw consumer data local by federated learning, avoiding the privacy threat posed by centralized data collection and storage; second, it adds noise at the model training process to achieve differential privacy, avoiding the privacy threat posed by attacks on the model. We performed simulations on the data from the Pecan street dataset collected from real households in Texas, USA, and the forecasting results show that DPEFL can create models with high forecasting performance while achieving different levels of privacy guarantees.