基于混沌优化的支持向量机地下水位动态预测

地下水位动态受到自然因素和人为因素的影响,随机性明显,因此在地下水物理过程分析的基础上构建地下水位动态预测的随机性模型对地下水资源评价具有重要意义。本文将小样本机器学习理论——统计学习理论中的支持向量机理论引入地下水位动态预测。最小二乘支持向量机是支持向量机的一种,考虑到地下水位动态序列的长度和峰值突变性的特点,本文提出一种改进的支持向量机-峰值识别最小二乘支持向量机;并针对支持向量机算法存在的参数优化、训练和测试速度等问题,结合混沌优化方法,建立了基于混沌优化的峰值识别最小二乘支持向量机地下水位动态预测模型;最后本文以内蒙古河套灌区义长灌域1990年~2004年3个灌期(夏灌(4月~6月)、秋灌(7月~9月)和秋浇(10月~11月)降水量、平均气温、蒸发量、引水量、地下水开采量、地下水排泄量和地下水位埋深共15年45个样本资料为数据源,将该模型和原最小二乘支持向量机模型分别用于义长灌域地下水位动态预测。结果表明,该模型的拟合值、检验值和预测值与实际值复合的很好,拟合的平均相对误差绝对值为2.0868%,检验的平均相对误差绝对值为3.4777%,预测的平均相对误差绝对值为6.8589%,且训练和测试速度快,而原最小二乘支持向量机模型预测的平均相对误差绝对值为20.6767%。因此,该模型用于地下水位动态预测是可行和有效的。

Dynamic Prediction of Groundwater Level based on Chaos Optimization and Support Vector Machine

Groundwater level has random characters because of influences of natural and anthropogenic factors.So study on random dynamic prediction model of groundwater level on the basis of groundwater physical process analysis is important to groundwater appraisal.The theory of supporting vector machine in small-sample machine learning theory was introduced into dynamic prediction of groundwater level.Considering groundwater level dynamic series length and peak mutation characters,the least squares support vector machine arithmetic based on peak value identification was proposed.Aiming at parameter optimization,training and speed test of supporting vector machine arithmetic,a least square supporting vector machine groundwater level dynamic forecasting model based on chaos optimization peak value identification was built up.At last,based on precipitation,average temperature,transpiration rate,amount of water diversion,ground water mining quantity,ground water excretion quantity and groundwater level burying depth data of summer irrigation periods(April to June),autumn irrigation periods(July to September,October to November)from 1990 to 2004 in Yichang irrigation sub-district of Hetao irrigation district in Inner Mongolia,dynamic prediction model of groundwater level was built up.The results show that the fitted values,the tested values and the predicted values of this model have little difference from their real values.The absolute value of the fitting mean relative error is 2.08 percent;The absolute value of the testing mean relative error is 3.48 percent;The absolute value of the predicting mean relative error is 6.86 percent.At the same time,the model has high training and testing speed.But the absolute value of the predicting mean relative error of the least squares support vector machine model is 20.68 percent.So the model proposed in this paper can provide a new tool for groundwater level dynamic forecasting.