多尺度注意力加权的CNN-BiGRU光伏功率预测方法

Photovoltaic Power Prediction Method Based on Multi-Scale Attention Weighted CNN-BiGRU

摘要: 光伏功率预测是保障光伏电站高效运维与电网优化调度的核心技术，其精度直接影响新能源消纳水平与并网安全性。针对传统CNN-BiGRU模型在光伏功率预测中存在的单尺度注意力表征能力不足、特征融合维度受限及网络结构布局不合理等问题，文章提出一种基于多尺度注意力加权的CNN-BiGRU光伏功率预测模型。该模型以“局部特征提取–多尺度特征融合–时序依赖建模–精准预测输出”为主干结构，首先通过卷积神经网络提取输入时序数据的局部空间特征；其次设计并行多尺度注意力模块，采用3、5、7三种卷积核分别捕捉短期波动、中期趋势与长期周期特性，并在192维高维特征空间中进行自适应加权，显著提升模型对复杂气象条件下功率变化的拟合能力；随后将优化后的特征输入BiGRU网络以建模双向时序依赖关系，最后通过全连接层输出单步预测结果。实验结果表明，所提模型在保证轻量化的同时有效降低了预测误差，相较于传统CNN-BiGRU及现有改进模型，在多项评价指标上均取得更优表现，验证了多尺度注意力机制与结构优化策略的有效性。

Abstract: Photovoltaic power prediction is a core technology for ensuring efficient operation and maintenance of photovoltaic power stations and optimized dispatch of power grids, and its accuracy directly affects the level of renewable energy accommodation and grid-connected security. To address the limitations of traditional CNN-BiGRU models in PV power prediction, such as insufficient single-scale attention representation capability, restricted feature fusion dimensions, and unreasonable network structural layout, this paper proposes a CNN-BiGRU photovoltaic power prediction model based on multi-scale attention weighting. The proposed model follows a main structure of “local feature extraction, multi-scale feature fusion, temporal dependency modeling, and accurate prediction output”. Firstly, a Convolutional Neural Network (CNN) is utilized to extract local spatial features from the input time-series data. Secondly, a parallel multi-scale attention module is designed, employing convolution kernels of sizes 3, 5, and 7 to capture short-term fluctuations, medium-term trends, and long-term periodic characteristics, respectively. Adaptive weighting is performed in a 192-dimensional high-dimensional feature space, significantly enhancing the model’s fitting capability for power variations under complex meteorological conditions. Subsequently, the optimized features are fed into a Bidirectional Gated Recurrent Unit (BiGRU) network to model bidirectional temporal dependencies. Finally, the single-step prediction results are output through a fully connected layer. Experimental results demonstrate that the proposed model effectively reduces prediction errors while maintaining a lightweight architecture. Compared with traditional CNN-BiGRU and existing improved models, the proposed method achieves superior performance across multiple evaluation metrics, verifying the effectiveness of the multi-scale attention mechanism and structural optimization strategies.