基于DCT-CNN-GRU的短期电力负荷预测研究

刘伟; 蔡东升; 冯付勇; 韩昊; 黄琦

刘伟,蔡东升,冯付勇,韩昊,黄琦.基于DCT-CNN-GRU的短期电力负荷预测研究[J].电测与仪表,2026,63(2):138-147.
LIU Wei,CAI Dongsheng,FENG Fuyong,HAN Hao,HUANG Qi.Research on short-term electric load forecasting based on DCT-CNN-GRU[J].Electrical Measurement & Instrumentation,2026,63(2):138-147.

基于DCT-CNN-GRU的短期电力负荷预测研究

Research on short-term electric load forecasting based on DCT-CNN-GRU

DOI：10.19753/j.issn1001-1390.2026.02.015

中文关键词: 短期电力负荷预测 DCT变换卷积神经网络门控循环单元，时频结合

英文关键词:short-term electric load forecasting, DCT transform, convolutional neural network, gated recurrent unit, time-frequency combination

基金项目:国家自然科学基金（52007025）;四川省科技支撑计划（2022JDRC0025）

作者中文名	作者英文名	单位
刘伟	LIU Wei	成都理工大学计算机与网络安全学院(牛津布鲁克斯学院)
蔡东升	CAI Dongsheng	成都理工大学计算机与网络安全学院(牛津布鲁克斯学院)
冯付勇	FENG Fuyong	成都理工大学计算机与网络安全学院(牛津布鲁克斯学院)
韩昊	HAN Hao	成都理工大学计算机与网络安全学院(牛津布鲁克斯学院)
黄琦	HUANG Qi	成都理工大学计算机与网络安全学院(牛津布鲁克斯学院)

摘要点击次数: 425

中文摘要:

短期电力负荷预测具有非线性、周期性以及变化快等特点，因此需要一种强大的模型来有效地挖掘其中的信息。为了提高短期电力负荷的预测精度，挖掘其中的信息，文中提出了一种综合应用离散余弦变换(discrete cosine transform, DCT)、卷积神经网络(convolutional neural network, CNN)和门控循环单元(gated recurrent unit, GRU)的混合模型的预测方法。模型首先利用离散余弦变换，将时域信息转换成频域信息,这个步骤有助于捕捉数据的频域特性。然后，将包含时域和频域信息的数据输入到卷积神经网络和门控循环单元中进行训练和预测。在模型中，首先通过卷积神经网络，对具有时域和频域信息的数据进行特征提取，再将数据传递给门控循环单元，充分利用门控循环单元的循环特性，学习数据的周期性和时序特征，从而实现更准确地预测。文中以美国加利福尼亚州的负荷数据和国内某公司的负荷数据作为案例进行实验验证。实验结果表明，所提出的混合模型相对于门控循环单元GRU、长短期记忆（long short-term memory, LSTM）网络、时间卷积网络（temporal convolutional network， TCN)等传统方法，能够获得更高的预测准确性。

英文摘要:

The short-term electric load forecasting is featured with nonlinearity, periodicity and rapid changes. Therefore, it requires a powerful model to effectively extract information from it. In order to enhance the accuracy of short-term electric load forecasting and extract information from it, this paper proposes a predictive approach that integrates discrete cosine transform (DCT), convolutional neural network (CNN), and gated recurrent unit (GRU) in a hybrid model. The model first employs the DCT transformation to convert time-domain information into frequency-domain information, which aids in capturing the frequency domain characteristics of data. Subsequently, the data, which contains both time-domain and frequency-domain information, is input into convolutional neural networks and gated recurrent units for training and prediction. In the model, data containing both time-domain and frequency-domain information is first subjected to feature extraction using convolutional neural networks. The data is then passed to gated recurrent units to fully leverage the recurrent nature of GRU, allowing the model to learn the periodic and temporal characteristics of the data and achieve more accurate prediction. The load data from a domestic company and California, USA, are used in this work as examples for experimental verification. The experimental results demonstrate that the proposed hybrid model, in comparison to traditional methods such as GRU neural networks, long short-term memory (LSTM), temporal convolutional network (TCN), and others, achieves a higher level of predictive accuracy.

查看全文查看/发表评论下载PDF阅读器

关闭