【摘要】 针对光伏功率预测中多变量交互作用、多尺度时间动态和非线性依赖关系的复杂挑战，本研究提出了一种先进的混合神经网络架构.该模型采用一维卷积层提取原始时序数据的局部特征，集成层级注意力机制实现自适应特征选择与噪声抑制，并利用循环单元建立长期时序关联.相较于传统CNN-LSTM模型，本框架具有三大创新点：(1)增强时空特征融合能力；(2)关键时间节点的优化信息加权机制；(3)提升系统扰动适应性能.大量实验验证表明，该模型在不同天气条件和运行状态下均表现出优异的鲁棒性和泛化能力.本研究为复杂工业时序预测任务提供了切实可行的解决方案.更多还原

【Abstract】 To address the intricate challenges in photovoltaic power forecasting involving multivariate interactions, multi-scale temporal dynamics, and nonlinear dependencies, this study proposed an advanced hybrid neural network architecture.The proposed model employs one-dimensional convolutional layers to extract local features from raw time-series data, incorporates a hierachical attention mechanism for adaptive feature selection and noise suppression, and utilizes recurrent units to establish long-term temporal dependencies.Three key innovations distinguish this framework from conventional CNN-LSTM models:(1)enhanced spatiotemporal feature integration;(2) optimized information weighting mechanism for critical time steps;(3)improved adaptability to system disturbances.Comprehensive experimental results confirm the model’s robust performance and superior generalization capacity across diverse weather conditions and operational states.This research offers a practically viable solution for complex industrial time-series prediction tasks.更多还原