混合图神经网络和门控循环网络的短期光伏功率预测

SHORT-TERM PHOTOVOLTAIC POWER PREDICTION METHOD BASED ON MIXED GRAPH NEURAL NETWORK AND GATED RECURRENT UNIT NETWORK

为了能从大量历史光伏发电数据中提取出有效的时序特征以及在非欧几里得域中的关联,建立了基于混合图神经网络以及门控循环网络的短期光伏功率预测模型。该模型首先通过最邻近分类算法生成气象及出力数据的最邻近图,再将其结合图神经网络作为编码器对气象及出力数据进行编码形成时间序列,最后通过门控循环网络以及全连接层解码输出光伏功率预测结果。通过仿真分析验证,该模型具有更优的特征挖掘能力和分析性能,能更好地突出某时间节点的气象及出力数据特征,适应天气突变带来特征变化,从而提升光伏预测整体模型的表达能力。

In order to extract effective temporal features and connections between non Euclidean domains from a large amount of historical photovoltaic power generation data,a short-term photovoltaic power prediction model based on mixed graph neural network and gated recurrent network is established.The model first generates the K-nearest neighbor graph of meteorological and output data through the K-nearest neighbor classification algorithm,and then uses the graph neural network as an encoder to encode the meteorological and output data to form a time series,and finally outputs the photovoltaic power prediction results through the gated recurrent network and the full connection layer decoding.Through simulation and analysis,the model has better feature mining ability and analysis performance,can better highlight the meteorological and output data characteristics of a certain time node,adapt to the feature changes caused by sudden changes in weather,and thus improve the expression ability of the overall model of photovoltaic prediction.