\chapter*{摘要}
\addcontentsline{toc}{chapter}{摘要}
\thispagestyle{abstract}

随着全球气候变暖，极端高温事件频发且强度持续增加，对公共卫生构成日益严峻的挑战。老年群体（65岁及以上）因体温调节功能减退、慢性病患病率高以及社会隔离等因素，是高温热浪最脆弱的群体之一。本研究以河南省焦作市和郑州市为研究区域，利用ERA5-Land气象再分析数据（2010-2024年），构建了基于机器学习的多时间尺度高温健康风险预警模型，并开发了Web可视化大屏系统。

本研究主要工作包括：（1）通过Copernicus Climate Data Store (CDS) API获取2010-2024年焦作、郑州两市的ERA5-Land网格气象数据，采用Magnus公式计算相对湿度、NOAA Rothfusz公式计算体感温度，构建了包含19个气象衍生特征的完整数据集；（2）通过滑动窗口方法（窗口14天）生成监督学习样本，构建了包含1,095,758条样本的多时间尺度预测数据集，覆盖短期（3天）、中期（7天）和长期（30天）三个预测窗口；（3）设计了983,628参数的LSTM-Attention深度学习模型，采用双向LSTM提取时序特征、4头自注意力机制捕捉关键时间步，并以Focal Loss缓解类别不平衡；（4）以XGBoost作为基线模型，在164,365条测试样本上进行了系统对比实验；（5）基于Flask和ECharts开发了深色科技蓝风格的Web可视化大屏，包含温度趋势、风险实时展示、人口饼图、预警时间线、暴露-反应曲线和历史回顾六个功能面板。

实验结果表明，XGBoost模型在三个时间尺度上均取得优异性能：短期（3天）F1-Macro达0.9325、中期（7天）达0.9195、长期（30天）达0.8576。LSTM-Attention模型（F1=0.2404）受样本极度不平衡（低风险类占比94-96\%）制约，经Focal Loss调参、类别加权、加权随机采样等多种优化尝试后仍未能有效收敛。该对比揭示了梯度提升树模型在表格型时序预测任务中相对深度序列模型的优势。本研究构建的可视化大屏系统为面向银发群体的高温健康防护提供了直观的决策支持工具。

\textbf{关键词：}高温热浪；银发群体；多时间尺度预警；XGBoost；LSTM-Attention；体感温度；可视化

\newpage

\chapter*{Abstract}
\addcontentsline{toc}{chapter}{Abstract}
\thispagestyle{abstract-en}

Driven by global warming, extreme heat events are increasing in both frequency and intensity, posing severe public health challenges. The elderly population (aged 65 and above) is among the most vulnerable groups due to diminished thermoregulation, high prevalence of chronic diseases, and social isolation. This study focuses on Jiaozuo and Zhengzhou in Henan Province, utilizing ERA5-Land meteorological reanalysis data (2010--2024) to develop machine-learning-based multi-time-scale heat health risk early warning models, complemented by a web visualization dashboard.

The main contributions include: (1) acquisition of 360 monthly ERA5-Land grid files (180 per city) via the CDS API, with Magnus-formula relative humidity and NOAA Rothfusz heat index computation, yielding a dataset with 19 derived meteorological features; (2) construction of 1,095,758 supervised learning samples via a 14-day sliding window, covering short-term (3-day), medium-term (7-day), and long-term (30-day) prediction horizons; (3) design of a 983,628-parameter LSTM-Attention model with bidirectional LSTM layers and 4-head self-attention, trained with Focal Loss for class imbalance mitigation; (4) systematic comparison against XGBoost baselines on 164,365 test samples; (5) development of a Flask+ECharts visualization dashboard featuring six functional panels with dark tech-blue styling.

Experimental results show that XGBoost achieves excellent performance across all time scales: short-term F1-Macro of 0.9325, medium-term 0.9195, and long-term 0.8576. The LSTM-Attention model (F1=0.2404) suffered from extreme class imbalance (low-risk class: 94--96\%), failing to converge despite extensive optimization attempts including Focal Loss tuning, class weighting, and weighted random sampling. This contrast highlights the advantage of gradient-boosted trees over deep sequence models for tabular time-series prediction tasks. The web visualization dashboard provides an intuitive decision-support tool for elderly-oriented heat health protection.

\textbf{Keywords:} Heatwave; Elderly Population; Multi-time-scale Early Warning; XGBoost; LSTM-Attention; Heat Index; Visualization
\newpage