Research on Load Forecasting of Smart Grid Based on Machine Learning

With the continuous popularization of smart grid and the establishment of advanced measurement system, a large amount of electricity consumption data has been generated in smart grid. These data can effectively help power companies to effectively forecast the load in the grid. Accurate and safe load forecasting is essential for power production, transmission and maintenance. At present, machine learning models have replaced other classical models to become the most popular prediction models. However, the deep prediction model requires users to provide a large amount of private power consumption data, which has potential privacy disclosure risks. Edge nodes can use machine learning to federate the global model through aggregation. As a new type of distributed machine learning technology, machine learning only sends model parameters and does not share raw data. However, existing machine learning-based prediction methods still face challenges of data heterogeneity, privacy breaches, and communication bottlenecks. To solve these problems, this paper further improves the load forecasting method based on machine learning, and proposes the load forecasting method based on personalized machine learning and the efficient load forecasting method based on single machine learning to solve the corresponding problems. This paper presents a user-level load forecasting method based on personalized machine learning (PFL). The personalized model trained by this method is better than the global model in predicting local data. In addition, we introduce a novel differential privacy algorithm into the method to provide additional privacy protection. Based on the generative adversarial network principle, the algorithm achieves a balance between privacy assurance and prediction accuracy in the whole game process. We carried out simulation experiments on real data sets, and the experimental results show that this method can meet the requirements of accuracy and privacy in actual load prediction scenarios. Aiming at the communication bottleneck of machine learning, we designed an efficient load prediction method based on one-shotFL. Aiming at the scenario of non-IID client data in load forecasting, this method uses the transmission of distillation data to complete the single communication between client and server, and finally trains the global model on distillation data to complete the task of user load forecasting. While greatly reducing the communication cost, the machine learning framework uses distilled data instead of gradient for transmission, which avoids the possibility of reverse-pushing private data sets by gradient leakage, and improves the privacy protection performance of the method.