Enhanced Deep Networks for Short-Term and Medium-Term Load Forecasting

• Main Authors: Lingyi Han, Yuexing Peng, Yonghui Li, Binbin Yong, Qingguo Zhou, Lei Shu
• Format: Article
• Language: English
• Published: IEEE 2019-01-01
• Series: IEEE Access
• Subjects: MTLF; long-term historical load data; spatial correlation; CNN; LSTM
• Online Access: https://ieeexplore.ieee.org/document/8584438/

Electric load forecasting (ELF) is vitally beneficial for electrical power planning and economical running in smart grid. However, the medium-term load forecasting has been rarely studied. In addition, existing ELF models mainly consider the impact of limited external factors, which are usually difficult to forecast accurately. In this paper, the characteristics of the electric loads are analyzed and used as a guideline for the design of the proposed methods. To fully exploit the quasi-periodicity with different time ranges, i.e., year, quarter, month, and week, two deep learning methods, time-dependency convolutional neural network (TD-CNN), and cycle-based long short-term memory (C-LSTM) network are proposed to improve the forecasting performance of short-term load forecasting and MTLF with a little payload of computational complexity. Both of them only utilize the historical electric load and can mine the underlying load patterns by extracting the long-term global integrated features and short-term local similar features. By representing the loads as pixels and rearranging them into a 2-D photograph, TD-CNN transforms the temporal correlation of load series into the spatial correlation and keeps the long-term memory. Specifically, the convolutional kernel with special size targeted to load data is designed to extract the local pattern with similar characteristic, while the pooling layer is removed in order to keep the finer features. Moreover, in order to extract the temporal correlation between the long-term sequences with lower complexity, the proposed C-LSTM method generates a new short series from the original long load series without information loss. The LSTM is then applied to model the dynamical relationship of the load series with shorter time steps. The experimental results show that the proposed methods outperform the existing method with greatly reduced computation complexity, whose training time is about two-to-five times shorter than the existing method.