基于相似日划分与常春藤优化的CNN-BiGRU-Attention光伏功率预测研究

Research on Photovoltaic Power Prediction Based on Similar-Day Clustering and Ivy Algorithm with CNN-BiGRU-Attention Hybrid Model

中文关键词:光伏发电功率  预测模型  神经网络  常春藤算法

keywords:photovoltaic power generation  prediction model  neural network  Ivy algorithm

中文摘要:为提高光伏发电功率预测的准确度与效率，提出一种基于相似日划分和常春藤算法（ivy algorithm，IA）优化的卷积神经网络（convolutional neural network，CNN）-双向门控循环单元（bidirectional gated recurrent unit，BiGRU）-注意力机制的光伏功率预测模型。首先，构建基于动态时间规整的K中心点相似日选取方法，采用动态时间规整（dynamic time warping，DTW）距离替代传统聚类方法的欧氏距离，实现对历史数据的非线性时间序列聚类，并将其划分为晴天、阴天和雨/雪天3种类型；其次，提出基于IA优化的CNN-BiGRU-注意力机制组合预测模型，利用IA自动优化超参数，并结合CNN的空间特征提取、BiGRU的时序建模能力以及注意力机制的关键信息聚焦功能，提升模型预测性能；最后，采用新疆某光伏电站的实测数据进行算例分析。实验结果表明，所提模型在长、短期预测中均表现出更高的准确性和时效性，同时验证了改进聚类方法的有效性。

Abstract:To enhance photovoltaic power forecasting performance，a novel prediction model is developed by integrating similarity day classification with an Ivy algorithm-optimized convolutional neural network （CNN）-bidirectional gated recurrent unit （BiGRU）-attention mechanism. First，a dynamic time warping （DTW）-based k-medoids method for selecting similar days is proposed. This approach replaces the Euclidean distance in traditional clustering with DTW distance to achieve nonlinear time series clustering of the historical data，which is classified into three categories：sunny，cloudy，and rainy/snowy. Second，a combined forecasting model based on an Ivy Algorithm-optimized Convolutional Neural Network-Bidirectional Gated Recurrent Unit-Attention Mechanism is introduced.The Ivy Algorithm automatically tunes the hyperparameters，while the CNN extracts spatial features，the BiGRU captures temporal dependencies，and the attention mechanism focuses on key information，thereby collectively improving prediction performance. Finally，a case study is conducted using real operational data from a PV plant in Xinjiang. Results demonstrate the proposed model delivers significantly higher accuracy and faster prediction speed for both long-term and short-term forecasting，while also confirming the effectiveness of the enhanced clustering method.