Abstract
This paper presents a comprehensive framework for multi-source heterogeneous data fusion and intelligent prediction modeling in chemical engineering construction projects using improved Transformer architecture with enhanced attention mechanisms. The proposed methodology addresses critical challenges in integrating diverse data modalities including structured numerical measurements, semi-structured operational logs, and unstructured textual documentation through innovative multi-scale attention mechanisms and cross-modal feature alignment modules. Key technical contributions include an adaptive weight allocation algorithm for dynamic data source management and a multi-task learning framework enabling simultaneous progress estimation, quality assessment, and risk evaluation. Comprehensive experimental validation demonstrates prediction accuracies exceeding 91% across multiple tasks, representing improvements of up to 19.4% over conventional machine learning techniques and 6.1% over standard Transformer architectures. Real-world deployment in three major chemical engineering construction projects confirms practical viability with robust anomaly detection capabilities achieving 92% + detection rates and real-time processing performance under 200 ms. The integration of interpretability mechanisms through attention visualization and SHAP analysis provides transparent decision-making processes aligned with engineering domain expertise requirements.