The accelerating global energy transition and rapid expansion of renewable energy sources，presents both opportunities and challenges. This transformation has introduced new concerns related to the “safety and stability” of power grids，particularly as large-scale integration of renewable energy sources such as wind and solar power results in issues including frequency overruns and voltage instability. This study explores the impact of renewable energy output and weather conditions on equipment failures and establishes a comprehensive scenario for power grids under renewable energy integration. A novel multihead graph-attention neural network model is proposed that integrates graph neural networks with multihead attention mechanisms. By incorporating parallel training methods，the proposed model is utilized in renewable energy power systems with the aim of improving risk assessment efficiency while maintaining accuracy in grid risk assessments. The model is trained and tested using data obtained from a provincial power grid within an electrical power simulation system. Results，derived from integrating real-world data from a provincial power grid in China with that of the electrical power simulation system，demonstrate that the attention-based graph neural network method approach substantially improves the robustness and efficiency of risk assessments compared to other artificial intelligence methods. This approach shows considerable promise in renewable energy power systems for enhancing risk assessment.