莱州湾南岸海水入侵特征及基于WA-NARX混合模型的动态预测

随着我国海岸带经济的快速发展与人类活动的加剧，地下水超采现象十分严峻，由此引发的海水入侵已成为滨海地区普遍存在的地质问题。本研究以莱州湾南岸海水入侵区为研究对象，根据地下水连续监测数据分析了地下水位和电导率的动态变化特征。在此基础上，基于降雨、蒸发、潮汐及农业排灌用电量等影响地下水动态变化的因素，建立了小波分析（WA）与具有外部输入的非线性自回归神经网络（NARX）的混合模型，对地下水水位和电导率进行动态预测，并采用均方根误差和拟合度对预测结果进行评价。研究结果表明，莱州湾南岸地下水年内动态变化特征为降雨入渗–开采型，地下水位和潮汐之间在0.5 d频率上呈现较高相关性，潮汐对地下水电导率的影响要弱于对地下水位的影响；WA-NARX混合模型在训练和测试阶段的均方根误差均小于0.03且拟合度大于0.98，可有效预测研究区海水入侵的变化程度。同时，为验证模型适用性，对比了不同影响因素作为模型输入参数对预测结果的影响。结果表明，降雨和潮汐参数是影响海岸带地下水位和电导率的主要变量，蒸发以及农业排灌用电量反映的部分抽水信息对地下水位和电导率也有影响，其影响程度与观测频率相关。本文研究结果可为海岸带海水入侵的实时监测、预测、预警提供理论与技术支撑。

Seawater intrusion characterization and dynamics prediction based on WA-NARX hybrid model in the south of Laizhou Bay

With the rapid economic development and increasing anthropogenic activities, the groundwater in the coastal area has been excessively exploited. The resulting seawater intrusion has become a widely distributed environmental geological problem. Taken the coastal area of Laizhou Bay as a research area, the dynamics of groundwater level (GWL) and electrical conductivity (EC) were analyzed with the continuous monitoring data. Based on the rainfall, evaporation, tide and agricultural irrigation and drainage electricity consumption that affect the groundwater variation, the hybrid model of wavelet analysis (WA) and NARX neural network was introduced to predict the dynamics of GWL and EC. The root mean square error (RMSE) and goodness of fit (R2) were used to measure the prediction accuracy. The results showed that the annual variation of GWL was characterized by a type of rainfall infiltration-exploitation. A significant correlation at the frequency of 0.5 d was observed between groundwater level and tide, and the influence of tide on EC was weaker than that on GWL. For the dynamics prediction with WA-NARX method, the RMSE was less than 0.03 and R2 was greater than 0.98 in both the training and testing stages. The results indicated the hybrid model had a good performance and could effectively predict the dynamics of GWL and EC. The effects of different influencing factors as model input parameters on the prediction results were further compared. The results showed that rainfall and tide parameters were the main variables affecting the GWL and EC variations in the coastal zone. The pumping information reflected by the evaporation and agricultural drainage and irrigation power consumption also affected the groundwater dynamics. The degree of influence was related to the observation frequency. The research results can provide theoretical and technical support for real-time monitoring, prediction and early warning of seawater intrusion in coastal zone.