近海多元时序智能预测模型评估——以赤潮与旅游区场景为例

Evaluation of intelligent multivariate time series forecasting models incoastal waters: a case study of red tide and tourist area scenarios

中文关键词:  近海水质预测  赤潮预警  海水浴场适宜性  时序依赖特征提取  深度学习  模型可解释性

英文关键词:coastal water quality forecasting  red tide early warning  bathing water suitability  temporal dependency feature extraction  deep learning  model interpretability

中文摘要:

      基于机器学习的定点浮标监测时序预测已成为实现近海水质高时效、低成本预报的重要研究方向。然而，近海水体环境变化具有显著的非线性与多变量耦合特征，如何在不同应用场景下平衡预测精度与使用成本，选择更适配业务需求的智能预测模型，是海洋管理面临的实际问题。为此，本研究构建了一套近海时序智能预测模型评估框架，旨在从预测精度、多步长稳健性、计算成本、时效性与可解释能力等维度，系统评估前沿智能预测模型的业务适用性。面向浙江近岸海域的赤潮预警和滨海预报两大典型场景，研究设计了统一的数据预处理流程、超参数优化搜索与评估指标体系，对 10种代表性的算 法进行了多要素预测与对比分析。结果表明，在赤潮预警场景中，PatchTST在Chl-a和溶解氧饱和度预测中表现稳健，能够有效捕捉藻华的突发性波动；在海水浴场适宜性预报场景中，XGBoost 在水温、pH、Chl-a和浊度等指标上综合表现优异，且具有计算效率高、部署成本低的优势，更适合业务化运行需求。进一步通过可解释分析技术，研究发现两大场景中多个水质要素均在约 12小时前呈现较高的时间权重，可能与浙江近岸海域典型的半日潮周期相关。因此，本研究提供了一套面向不同业务场景的近海预测模型评估方法，对提升近岸生态风险预警能力与支持海洋环境精细化管理具有实际参考价值。

英文摘要:

      Machine learning-based time series forecasting using fixed-point buoy monitoring has become an important research direction for achieving high-temporal-resolution and cost-effective coastal water quality predictions. However, coastal water environments exhibit significant nonlinearity and multivariate coupling characteristics. A practical challenge for marine management is how to balance predictive accuracy and computational cost under different application scenarios, and how to select intelligent forecasting models that better meet operational requirements. To address this, the present study develops a comprehensive framework for evaluating intelligent time series forecasting models for coastal waters, systematically assessing their applicability in terms of prediction accuracy, multi-step robustness, computational cost, timeliness, and interpretability. Focusing on two representative scenarios in the coastal waters of Zhejiang Province—red tide early warning and bathing water suitability forecasting—we designed a unified data preprocessing workflow, hyperparameter optimization and search strategy, and evaluation metric system, and conducted multi-factor forecasting and comparative analysis for ten representative algorithms. Results indicate that, in the red tide early warning scenario, PatchTST demonstrates robust performance in predicting chlorophyll-a (Chl-a) and dissolved oxygen saturation, effectively capturing sudden bloom fluctuations. In the bathing water suitability forecasting scenario, XGBoost shows superior overall performance in water temperature, pH, Chla, and turbidity predictions, while also offering high computational efficiency and low deployment cost, making it more suitable for operational applications. Further interpretability analysis reveals that multiple water quality variables in both scenarios exhibit high temporal importance approximately 12 hours in advance, which may be related to the semi-diurnal tidal cycle typical of Zhejiang's coastal waters. Therefore, this study provides a scenario-oriented evaluation methodology for coastal forecasting models, offering practical guidance for enhancing nearshore ecological risk warning capabilities and supporting refined marine environmental management.