Power Load Forecasting Based On The Graph Neural Networ K

Smart Grid has been developed rapidly,and lead to increasingly higher requirements on the accuracy of load forecasting of power system.It is required that the power grid should not only ensure the safe,reliable and uninterrupted power supply,but also ensure the economical and effective energy saving of the operation area.So that it can achieve the purpose of the rational use of resources,reducing the power grid pressure and,reducing the accident loss.However,changes in power loads are not only affected by their own changes,but also affected by various other factors.There are structural factors such as temperature,daily types,etc.,but also the influence of meteorological cloud maps and weather conditions.Traditional prediction methods cannot analyze the influencing factors of non-structured factors.Although the intelligent prediction method can be processed by non-structured factors,it cannot handle the sequence of dynamic time dependence,and it cannot analyze the degree of influence on load changes by other factors.Therefore,improving the accuracy of load forecasting of power system currently needs to be solved in the power market.In recent years,the neural network has shown a strong ability to obtain the space correlation of non-structured data with dynamic time-dependence,and can use time to pay attention to the mechanism to give different historical data to different factors.To solve the above problems,we use the public multivariate data set in New England in the United States and propose a model of load forecasting in power system based on graph neural networks.It captures the influence of the other factors of load forecasting to improve the accuracy of the prediction,the main results and conclusions are as follows:(1)There are problems with non-structured influencing factors in the load forecasting of power system and difficulty in obtaining dynamic time dependence.Based on the GNN,the influencing factors of diversified loads are integrated,and the accuracy of load forecasting of power system is improved.Through the relationship between the influencing factors and loads by the graph convolution network,the time dependence is captured by time convolution.For the time sequence,traditional methods cannot handle long-term sequences,we use Diffusion-Convolutional Neural Networks to handle them;for the relationship between the diverse variables of the power load,and the relationship between the traditional map of the neural network requires predetermined dependencies,we adopt a diagram learning structure to build a adjacent matrix of the dynamic relationship between the influencing of the power load.Finally,the reservoir calculation is added to reduce the calculation cost.The experimental results show that the use of the Graph Neural Network for power load prediction can significantly improve the accuracy.(2)The potential topology relationship between the influencing factor variables for the influencing factor variables in the neural network is low.The mechanism and experimental results show that the addition of multiple parameters has significantly improved the accuracy of power load prediction.(3)For the computing cost of the network is increased due to the long-term sequence of the power load prediction,we use the smooth and sparse structure to reduce the calculation of.For the decline in the accuracy of the long-term sequence prediction,the dynamic space relationship of multi-map extraction of the dynamic space of power load variables is used to generate multi-diagram power load variables.Better reflect the time-space relationship between different variable values during time changes.The experimental results show that adding a smooth and sparse structure and a multi-picture generation network can effectively reduce computing costs and improve the accuracy of forecasting.