基于卷积神经网络、小波神经网络与掩码多头注意力机制的全球辐照度预测模型

Accurate prediction of global irradiance is critical for optimizing energy management in photovoltaic (PV) systems, particularly in solar-powered electric vehicles (ESVs). However, traditional models struggle to capture the complex spatial and temporal dependencies in irradiance data, limiting prediction accuracy under varying weather conditions. Existing approaches, including statistical methods, conventional machine learning models, and standalone deep learning techniques like LSTM, fail to integrate local features and long-term dependencies simultaneously, creating a need for more robust solutions. This paper introduces a novel hybrid framework, CNN-WNN-MMHA, that combines Convolutional Neural Networks (CNN), Wavelet Neural Networks (WNN), and a Masked Multi-Head Attention (MMHA) mechanism. The CNN extracts spatial and local features, WNN performs frequency decomposition to capture multi-scale variations, and MMHA models temporal dependencies while encoding positional information.The model is trained and evaluated on a real-world climatic dataset from Tunisia, collected over eight years. Experimental results demonstrate that the proposed model significantly outperforms state-of-the-art methods such as LSTM, BiLSTM, and CNN-LSTM, achieving a 79% reduction in MAPE and superior generalization performance across diverse weather scenarios.This advancement enhances energy forecasting reliability, supporting smarter route planning and energy optimization for solar-powered vehicles, with potential extensions to other renewable energy systems.