Open AccessResearch on Power Load Forecasting and Visualization Method Based on Deep Neural Network
Author(s) -
Yifei Wang,
Xian Li,
Shuang Mo
Publication year - 2020
Publication title -
iop conference series. materials science and engineering
Language(s) - English
Resource type - Journals
eISSN - 1757-899X
pISSN - 1757-8981
DOI - 10.1088/1757-899x/768/7/072091
Subject(s) - computer science , artificial neural network , scheduling (production processes) , electric power system , visualization , deep learning , artificial intelligence , industrial engineering , power (physics) , big data , grid , machine learning , operations research , data mining , real time computing , engineering , operations management , physics , geometry , mathematics , quantum mechanics
Short-term load forecasting (STLF) has attracted much attention in the last decade, which uses data from historical power grid operation monitor to forecast future power demand for a few days. Previous works on STLF mainly focus on features, which are widely used to capturing external factors in power system. There are also approaches based on physics, which use linear combination of active and reactive power consumption to model dynamics of power demand. However, in many real-world applications they don’t work well. In the self-service power grid, STLF is a very important part in power scheduling and planning. But STLF is a challenging task for its complex power demand and the effects of weather. In this paper, we study the problem of combination of complex and seasonality of power demand, as well as effect of external weather. Therefore, we come up with an approach, called DNNCast, it’s a model based on deep neural network, which can achieve a good accuracy and fast training speed. In the structure it uses a deep neural network to model various power loads and external weather conditions. Finally, with visualization tools we can wisely make decision about power scheduling and planning. Compared with the current STLF methods, our model has a better prediction effect, and faster in the experiments of the extended experiment on 133 real power datasets.