Short-Term Load Forecasting Based on CNN-BiLSTM with Bayesian Optimization and Attention Mechanism

• Published in: Parallel and Distributed Computing, Applications and Technologies Vol. 12606; pp. 116 - 128
• Main Authors: Miao, Kai, Hua, Qiang, Shi, Huifeng
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2021; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Attention Mechanism; Bayesian Optimization Algorithm; Bidirectional long short-term memory network; Convolutional Neural Networks; Short-term load forecasting
• ISBN: 9783030692438; 3030692434
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-030-69244-5_10

Short-term power load forecasting is quite vital in maintaining the balance between power production and power consumption of the power grid. Prediction accuracy not only affects the power grid construction, but also influences the economic development of the power grid. This paper proposes a short-term load forecasting based on Convolutional Neural Networks and Bidirectional Long Short-Term Memory (CNN-BiLSTM) with Bayesian Optimization (BO) and Attention Mechanism (AM). The BiLSTM is good at time series forecasting, and the Attention Mechanism can help the model to focus on the important part of the BiLSTM output. In order to make the forecasting performance of the model as good as possible, the Bayesian Optimization is used to tune the hyperparameters of the model. The input of the model is history load, time slot, and meteorological factors. In order to eliminate the seasonal influence, the data set is divided into four subsets with respect to four seasons. The performance of the proposed model is compared with other forecasting models by MAE, RMSE, MAPE, and R2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} score. The experiment results show that the proposed model fits the actual values best and has the best forecasting performance among the contrast models.