基于人工智能技术的抽水蓄能电站预选址研究

Research on pre-site selection of pumped storage power stations based on artificial intelligence

摘要

【目的】为解决目前抽水蓄能电站预选址大多依靠人工比选，存在耗时费力、自动化程度低等问题，提出一种将多模态大模型应用于抽水蓄能电站预选址的方法。【方法】基于抽水蓄能电站选址规范，构建站址评价体系，结合模糊综合评价法计算各站址综合得分，作为电站标签。通过设计特定提示引导GPT模型生成与遥感图像相关的指令微调数据。在此基础上，采用提示工程(prompt)与LoRA微调技术对多模态大语言模型LLaVA进行训练，并将模型应用于安徽绩溪抽水蓄能电站预选址工程，对模型实施系统评估。【结果】结果显示：模型对安徽绩溪抽水蓄能电站工程水文、地形、经济等方面的关键指标评分准确，得出其综合得分为84.4分，符合理想站址标准；在1091个样本的测试集上进行验证时，模型能识别出74.1%的理想站址样本和82.4%的非理想站址样本；微调后的LLaVA模型的AUC值为0.822,相比Qwen-VL-Chat、InternLM-XComposer-VL、VisualGLM和InstructBLIP模型分别高出了0.106、0.152、0.205和0.207。【结论】结果表明：利用上述方法进行微调的LLaVA模型在站址分类方面的精准率、召回率和误检率相较于多模态通用模型有显著提升，并在实际的选址中展现出良好的站址评分能力，具有较高的推广应用价值。通过对LLaVA模型进行领域微调和应用，充分展示了多模态大模型在提高选址工作效率和自动化程度中的独特作用，可为抽水蓄能行业的智能化转型提供有力支撑。

Abstract

[Objective] Current pre-site selection of pumped storage power stations heavily relies on manual comparison and selection, which suffers from time-consuming processes and low-automation levels. To address these issues, a method integrating multimodal large models into the pre-site selection of pumped storage power stations is proposed. [Methods] Based on site selection criteria for pumped storage power stations, an evaluation system for potential sites was established. The fuzzy comprehensive evaluation method was employed to calculate an overall score for each site, which served as the station label. Then, specific prompts were designed to guide the GPT model in generating prompt fine-tuning data associated with remote sensing images. Based on this, prompt engineering and Low-Rank Adaptation(LoRA) fine-tuning techniques were used to train the multimodal large language model LLaVA. Subsequently, the trained model was applied to the pre-site selection of the Jixi Pumped Storage Power Station in Anhui Province, followed by a systematic evaluation of the model performance. [Results] The result showed that the model accurately scored for key indicators such as hydrology, topography, and economic factors for the Jixi Pumped Storage Power Station, yielding a comprehensive score of 84.4 that met the criteria for an ideal site. When validated on a test set of 1 091 samples, the model successfully identified 74.1% of ideal site samples and 82.4% of non-ideal site samples. The fine-tuned LLaVA model achieved an Area Under the Curve(AUC) value of 0.822, outperforming Qwen-VL-Chat, InternLM-XComposer-VL, VisualGLM, and InstructBLIP models by 0.106, 0.152, 0.205, and 0.207, respectively. [Conclusion] The findings indicate that the LLaVA model fine-tuned by the proposed method achieves significant improvements in accuracy, recall, and false detection rates for site classification compared to general-purpose multimodal models. Additionally, it demonstrates excellent site evaluation in practical applications, showing high potential for broader application. The domain-specific fine-tuning and application of the LLaVA model effectively highlight the unique advantages of multimodal large models in improving the efficiency and automation level of site selection, providing robust support for the intelligent transformation of the pumped storage industry.

关键词

抽水蓄能电站 / LLaVA模型 / 预选址 / 指令微调 / 影响因素

Key words

pumped storage power station / LLaVA Model / pre-site selection / prompt fine-tuning / influencing factors

基金资助

国家重点研发计划项目(2019YFCO409000)

教育部人文社会科学研究规划基金项目(23YJAZH225)

江苏省高等学校大学生创新创业训练项目(202410294250Y)