Short-Term Load Forecasting Model of Distribution Transformer Based on CNN and LSTM
Simin Luo, Yi Rao, Jian Chen, Haijing Wang, Zhao Wang
Abstract
This paper proposed a novel short-term load forecasting method using a hybrid model of Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) for distribution transformers (DT). Traditional Short-term load forecasting are expected to improve the economic efficiency and reliability of the power grid. Thus, the traditional short-term load forecasting methods are mainly used for systemic or regional loads such as regions, cities, and counties. Scholars have proposed lots of mature and accurate prediction methods for traditional load forecasting. However, due to the small order of magnitude (kilowatts), real-time fluctuations and complexity, the load of DT is difficult to predict. Few researches focus on load forecasting for DT. The aim of load forecasting for DT is to find the DT that has a high risk of overloading in the future, thereby improving safe operation of distribution networks. Using a one-dimensional CNN as a pre-processing step can convert a long sequence into a short sequence of advanced features with benefit of high computational speed and low redundancy. Thus, this method is chosen to process long sequence. Besides, to evaluate the impact of three phase unbalance operation, this method separately deals with each phase load data instead of total load data of DT to obtain the different operation statues of each phase. The method can effectively avoid unexpected damage caused by overload of a certain phase. Moreover, the data collected from on-service DT are used to verify the effectiveness and accuracy of this model by comparison with the traditional prediction method. Most importantly, the method provided in this paper can be widely used for the short-term load forecasting for DT in practice.