Construction method and application of a data-sharing agent for large-scale hydropower projects

Large-scale hydropower projects generate substantial amounts of heterogeneous data dispersed across design, construction, and supervision units. The interoperability among stakeholders is suboptimal due to the heterogeneity of data structures and professional contexts. Consequently, information sharing remains inefficient. Existing studies have typically focused on specific data types or lifecycle stages, lacking a unifying framework to facilitate comprehensive, full-cycle data sharing. To address this issue, this study proposes the development of a data-sharing agent tailored to the needs of hydropower engineering. The proposed agent is designed to accommodate structured, semi-structured, and unstructured data, and it integrates external tools such as time-series databases, knowledge graphs, and text vector databases. This integration enables accurate, on-demand data retrieval. By enhancing the tool-learning capabilities of large language models, the agent bridges data silos, enhances cross-domain collaboration, and lays a solid technical foundation for intelligent construction in complex hydropower projects.

The research commences with a systematic analysis of data-sharing requirements across the full lifecycle of hydropower projects, encompassing time-series monitoring data, technical documentation, and parametric design files. Based on this analysis, a comprehensive agent framework is designed to support multi-modal data interoperability. To ensure its practicality, a supporting tool system is constructed that integrates intelligent modules for database retrieval, knowledge graph querying, and rule-based inference. Furthermore, an action-planning dataset comprising over 4000 samples is developed to train the agent in decision-making and tool invocation. Two versions of the DeepSeek-R1-Distill-Qwen model (1.5B and 7.0B parameters) are fine-tuned using this dataset to enhance structured parameter extraction, multi-step reasoning, and action planning capabilities. To assess performance, a benchmark testing dataset comprising hundreds of real-world business queries derived from hydropower project workflows is established and manually annotated to ensure fairness and reproducibility.

Experimental results demonstrated that the fine-tuned models substantially improved planning and reasoning performance. A comparative analysis revealed that the 1.5B and 7.0B models achieved 270% and 104% improvements in planning accuracy, respectively, compared with their pre-fine-tuning counterparts. On the business query test set, the overall output accuracies were 65.83% and 90.83%, respectively, thereby confirming a significant enhancement in model reliability and practical utility through fine-tuning. Notably, the 7.0B model consistently outperformed the smaller version, highlighting the larger model's capacity to handle complex, multi-step reasoning tasks. A practical deployment of the agent-based data-sharing platform was conducted for a real hydropower project in a representative watershed. Under static and structured data-sharing conditions, the agent maintained an average response time of less than 20s. Conversely, dynamic monitoring scenarios involving high-frequency data streams exhibited average latencies exceeding 30s, with peaks exceeding 60s under intensive analytical loads.

This study proposes a comprehensive framework for constructing a data-sharing agent that effectively addresses critical challenges in current hydropower data-sharing practices, particularly in high-altitude, data-scarce environments. By aligning agent design with engineering-specific requirements and integrating a highly refined large language model with a domain-oriented tool ecosystem, the proposed method significantly enhances the efficiency, intelligence, and semantic interoperability of data sharing. The agent reduces cross-disciplinary access barriers, improves system responsiveness, and supports knowledge-driven decision-making. The results from field applications confirm its considerable potential for practical implementation in intelligent construction platforms. Furthermore, the findings of this study provide a scalable, generalizable technical foundation for the future development of data-driven management and intelligent decision-support systems in complex hydropower projects.