基于NARX神经网络的千岛湖藻类短期预测模型构建

局部水域的藻类异常增殖现象逐渐成为千岛湖面临的水环境保护难题. 构建以数据驱动的水华预测模型，实现对重点水域叶绿素a (Chla)浓度短期动态变化的预测，是快速应对潜在水华风险的有效手段之一. 鉴于NARX神经网络在预测非平稳时间序列动态特征方面的优势，以千岛湖国控监测断面小金山2016—2019年Chla的高频时间序列作为研究对象，对Chla剖面数据进行沿深平均、缺失值插补后，分别以连续3 d和连续7 d的Chla浓度作为输入，构建了基于NARX神经网络的藻类预测模型，用于预测未来0.5~7 d Chla浓度的变化，探讨了相关参数设置、训练及评价方法，并针对不同的预见期分析了模型性能. 结果表明:① 模型预测性能稳定，预测值与实测值相关系数保持在0.8~0.9之间，均方误差在15~30之间. ②随着预见期的变化，模型性能不同. 其中，在未来0.5~4 d的预测中，使用连续3 d的 Chla浓度作为输入的预测效果较好；在未来4.5~7 d的预测中，使用连续7 d的Chla浓度作为输入的预测效果较好. 研究显示，该模型可以较为准确地预测未来0.5~7 d的Chla浓度，可为构建以数据驱动的千岛湖水华监测预警系统提供科学依据. 

Short-Term Forecasting Model for Algae Based on NARX Neural Network in Qiandaohu Reservoir

In recent years, Qiandaohu Reservoir has suffered from abnormal growth of algae. Short-term forecasting models of algal blooms driven by data can be built to predict the temporal variation of chlorophyll-a (Chla) concentration in some concern water. It is considered effective to deal with potential bloom risks. NARX (nonlinear autoregressive with external input) neural network has a competitive advantage in predicting the dynamic characteristics of non-stationary time series. In this paper, a high-frequency monitoring buoy dataset of Xiaojinshan Station, a state-controlled section in Qiandaohu Reservoir from 2016 to 2019 was handled with vertical integration and missing value interpolation. An algal bloom forecasting model based on NARX neural network was then built to predict the temporal variations of Chla over the next 0.5-7 days. To drive the model, Chla concentrations from the first 3 and 7 days were used as initial inputs, respectively. Parameters settings, trainings and evaluations were discussed, and model performance was analyzed according to different foreseeable periods. The results showed that: (1) The performance of the model was steady. The correlation coefficients between prediction and observation reached 0.8-0.9, and the corresponding mean square error values were 15-30. (2) In the 0.5-4 days of forecasting, the model using the first 3-day Chla concentration as the initial input had higher accuracy, while the model using the first 7-day chlorophyll concentration as initial input showed better precision in the forecast for days 4.5-7. It is therefore recommended to us models with different inputs for different forecast periods. The model contributed to provide a methodology for a bloom early warning system driven by data.