基于天气分类的改进CNN-LSTM区域光伏功率单值和概率预测方法

Improved CNN-LSTM Method for Single Value and Interval Prediction of Regional Photovaltaic Power Based on Weather Classification

在大规模分布式光伏电站不断并入配电网的背景下,光伏功率的准确预测能指导配电网的安全运行和调控。目前针对单一光伏场站功率的预测较多,区域光伏功率单值和概率预测起步较晚。为了提高区域光伏功率的预测精度,该文提出一种根据天气进行分类的方法,基于非线性分位数回归改进的卷积长短期神经网络(convolutional neural networks-long short term memory,CNN-LSTM)神经网络,对区域光伏场站的功率进行预测,不同场站使用独立的卷积层和池化层,提取各自场站和场站间的特征,再输入到长短期记忆神经网络(long short term memory,LSTM)提取时间序列上的特征,以便于快速得到天气因素与区域功率之间的关系,再结合分位数回归方法对功率进行区间预测。为了证明所提方法的有效性和可靠性,使用澳大利亚5个相近光伏场站的发电数据进行验证,结果表明所提出的方法与现有方法相比,不仅提高了功率单值预测精度和区间预测准确性,还缩短了训练模型的时间。区域光伏功率精确的单值预测和区间预测为配电网的安全运行管理提供有效保障。

The accurate prediction of photovoltaic power can guide the safe operation and regulation of the distribution network under the background of the continuous integration of large-scale distributed photovoltaic power stations into the distribution network.At present,there are many predictions for the power of single photovoltaic station,while the single point and probabilistic prediction of regional photovoltaic power start late.In order to improve the accuracy of regional photovoltaic power prediction,a classification method based on weather is proposed in this paper.CNN-LSTM(convolutional neural networks-long short term memory)neural network improved by quantile regression is used to predict the power of regional photovoltaic stations.Multiple stations use independent convolution layer and pooling layer to extract the features of each station and between the stations.Then input the LSTM(Long short term memory)to extract the features on the time series,so that the relationship between weather factors and regional power can be found more quickly and accurately,and then combined with quantile regression method for power interval prediction.In order to prove the validity and reliability of the proposed method,the power generation data of five similar photovoltaic stations in Australia were used for verification.The results show that compared with the existing methods,the proposed method not only improves the accuracy of single value power prediction and interval power prediction,but also reduces the training time of the model.Accurate single point prediction and interval prediction of regional PV provide guarantee for safe operation and management of distribution network.