针对传统循环神经网络（RNN）长时间使用会存在梯度爆炸以及在处理长时间序列时容易忽略重要时序信息的不足，本文提出一种结合注意力机制（Attention）的双重选择循环神经网络（Double selection Recurrent Neural Network，DsRNN），面向短期光伏发电功率预测的模型。首先，引入气象影响因子数据并根据相关性大小进行修正处理，改变原有单一输入源建立新的数据集；然后，融合注意力机制，提取光伏发电功率的时序特征，挖掘数据之间的深层联系；最终，实现对分布式光伏发电进行较有效、精准的短期功率预测。仿真结果表明：气象数据的输入以及DsRNN光伏发电功率预测模型的使用能完成较高精度的预测任务，误差更小。

Aiming at the shortcomings that the traditional recurrent neural network (RNN) will have gradient explosion for longterm use and easily ignore important time series information when processing longterm sequences, this paper proposed a double selection recurrent neural network (DsRNN) combined with attention mechanism, which was oriented to the shortterm photovoltaic power prediction model. First, the meteorological impact factor data was introduced and corrected according to the correlation size, and a new data set was established by changing the original single input source; then, the attention mechanism was integrated to extract the time series features of photovoltaic power, and mine the deep relationship between the data; finally, a more effective and accurate shortterm power prediction for distributed photovoltaic power generation was realized. The simulation results show that the input of meteorological data and the use of the DsRNN photovoltaic power prediction model can complete the prediction task with higher precision and the error is smaller.

TK221

国家自然科学基金青年项目（51807079）；中国博士后面上基金（2020M673339）