结合改进的DCGAN和Attention-LSTM的光伏功率预测

针对新建光伏发电站在光伏功率预测过程中由于缺少训练数据导致预测精度较低以及光伏发电功率的不稳定等问题，本文提出一种结合改进的深度卷积生成对抗网络（DCGAN)、 注意力机制（Attention）和LSTM网络组合的光伏功率预测方法。首先，将DCGAN中生成器的特征提取网络由二维卷积改为一维卷积,更好的学习一维时序数据，并用改进的DCGAN对光伏数据进行扩充，其次，采用Attention模块和LSTM模型相结合，先通过Attention模块提高重要特征的权重占比，生成新的输入特征，再通过LSTM模型进行功率预测。采用澳大利亚沙漠知识太阳能中心(DKASC)Alice Springs电站的数据进行验证，实验结果表明，结合深层卷积生成对抗网络与Attention-LSTM的混合预测方法能有效提升预测的精度。

Photovoltaic power prediction combining improved DCGAN and Attention-LSTM

To solve the problems of low prediction accuracy and instability of photovoltaic power generation caused by lacking training data in the process of photovoltaic power prediction of new photovoltaic power stations, this paper proposes a day-ahead photovoltaic power prediction method combining the improved Deep Convolutional Generative Adversarial Network (DCGAN), Attention Mechanism and LSTM network. Firstly, the feature extraction network of the generator in DCGAN is changed from two-dimensional convolution to one-dimensional convolution to better learn one-dimensional time series data. And the photovoltaic data is expanded with the improved DCGAN. Secondly, combine the attention module and the LSTM model. The weight ratio of important features is increased through the attention module to generate new input features, and then using the the LSTM model to make power prediction. By adopting the data of Alice Springs power station of Desert Knowledge Australia Solar Centre(DKSAC) in verification, the experimental results show that the prediction accuracy can be effectively improved by the hybrid prediction method combining Deep Convolutional Generative Adversarial Network and Attention-LSTM.