由于影响光伏发电的因素复杂多变,其功率表现出波动性与不确定性,这对电网安全稳定运行构成了重大影响。对中长期光伏发电功率进行精准的预测,既可以保障光伏发电安全并网,又能满足电网的中长期统筹与规划。本文提出了一种既能捕获时序周期性特征又能提取周期内波动性特征的功率预测模型DCN-GRU。该模型首先利用深度交叉网络(DCN)对光伏发电功率原始时序进行特征提取,将数值型特征与类别型特征进行高阶交叉组合,充分挖掘序列的周期内波动性特征；再通过门控循环单元(GRU)捕获中长期时序的周期性特征；最终将波动性特征与周期性特征拼接在一起,经过全连接层预测出中长期光伏发电功率结果。以我国华东及华南两个光伏电站为例的实验结果证明:DCN-GRU模型在1d、3d、5d、7d预测任务上相较于基准模型均具有较高的精度,其中平均绝对误差的整体均值为0.688、均方误差的整体均值为2.61、决定系数的整体均值为0.976。

Due to the complex and variable factors affecting photovoltaic (PV) power generation, its output power exhibits fluctuation and unpredictability, presenting considerable difficulties to the reliable operation of the power. Precise forecasting of medium and long-term PV power generation is crucial, it would not only ensures the safe grid-connection of PV power, but also meets the needs for medium and long-term planning of the power grid. This paper proposes a power prediction model called DCN-GRU, which could capture both the periodic features of time series and the volatility features within periods. The model first uses a Deep Cross Network (DCN) to extract features from the original time series of PV power generation, performing high-order cross of numerical and categorical features, in order to fully mining the volatility features within the sequence period. It then captures the periodic features of medium and long-term time series through Gated Recurrent Units (GRU). Finally, the volatility features are concatenated with the periodic features, passed through a fully connected layer to predict the medium and long-term PV power generation results. Experimental results using two PV power stations in East China and South China demonstrate that the DCN-GRU model has higher accuracy than benchmark models in 1d, 3d, 5d and 7d forecasting tasks, with overall mean values of Mean Absolute Error (MAE) being 0.688, Mean Square Error (MSE) being 2.61, and Coefficient of Determination (R2) being 0.976.