Abstract: Power industry requires accurate short-term load forecasting to provide precise load requirements for power system control and scheduling. In order to improve the accuracy of short-term power load forecasting, a method based on FFT optimized ResNet model is proposed. The model first defines power load forecasting as a time series problem, then introduces one-dimensional ResNet for power load regression prediction, and uses FFT to optimize ResNet. The FFT transform of a layer of convolution results gives the model the ability to extract periodic features in the data. Experiments show that the prediction accuracy of FFT-ResNet is better than several benchmark models in 6-hour power load forecasting, which indicates that this method has a good application prospect in power load forecasting.

Key words: Fast fourier transform, residual network, short-term electric load forecasting, convolutional neural network, time series

摘要： 短期负荷预测在为电力系统制定经济、可靠和安全的运行策略中起着关键作用。为了提高预测精度，提出了一种基于多模型融合神经网络的短期负荷预测方法。首先将三种不同的神经网络单独训练；再将单独模型的输出作为输入，训练一个顶层全连接神经网络；最后并行排列三种单独模型，再将3个模型的输出层组合作为顶层全连接神经网络的输入层，使4个模型融合为一个模型，并进行精调训练。短期负荷预测的实验结果表明，该方法的精度优于单个全连接神经网络、长短期记忆网络或残差网络。说明该方法具有良好的实用价值。

关键词: 模型融合, 循环神经网络, 短期负荷预测, 卷积神经网络, 时间序列