基于超参数优化的极限学习机区域水资源短缺风险评价

Risk Assessment of Regional Water Resource Shortage Using Extreme Learning Machine Based on Hyperparameter Optimization

为科学评价区域水资源短缺风险水平,改进极限学习机(ELM)评价性能,提出晶体结构算法(CryStAl)、鹈鹕优化算法(POA)与ELM组合的水资源短缺风险评价模型,并通过云南省水资源短缺风险评价实例进行验证。首先,简要介绍CryStAl、POA原理,通过4个标准函数对CryStAl、POA进行仿真测试;其次,建立水资源短缺风险评价指标体系和等级标准,采用线性内插和随机选取的方法生成样本,并构建ELM超参数优化适应度函数;最后,采用CryStAl、POA对适应度函数进行寻优,利用寻优获得的最佳ELM超参数建立CryStAl-ELM、POA-ELM模型对实例各年度水资源短缺风险进行评价,结果与模糊综合评价法、CryStAl-SVM、POA-SVM、ELM、SVM模型的评价结果作对比。结果表明:CryStAl、POA具有较好的寻优精度及全局搜索能力;CryStAl-ELM、POA-ELM模型对检验样本评价的平均绝对百分比误差(MAPE)分别为0.077%、0.083%,评价精度较CryStAl-SVM、POA-SVM模型提高57.7%以上,较SELM、SVM模型提高83.5%以上;CryStAl、POA能有效优化ELM超参数,提高ELM的评价性能。CryStAl-ELM、POA-ELM模型评价结果表明,实例2006年~2008年水资源短缺风险为“较高风险”,2009年~2012年为“中风险”,2013年~2019年为“较低风险”,2020年~2025年为“低风险”;近15年来云南省水资源短缺风险水平呈下降趋势,且下降趋势显著。

In order to evaluate the risk level of regional water shortage scientifically and improve the performance of Extreme Learning Machine(ELM),two water shortage risk assessment models combining Crystal Structure Algorithm(CryStAl) and Pelican Optimization Algorithm(POA) with ELM is proposed,respectively.The model is validated through an example of water shortage risk assessment in Yunnan Province.Firstly,the principles of CryStAl and POA are briefly introduced,and the simulation tests on them are conducted through four standard functions.Secondly,the risk assessment index system and level standards for water resource shortage are established,the samples are generated using linear interpolation and random selection methods,and an ELM hyperparameter optimization fitness function is constructed.Finally,the CryStAl and POA are used to optimize the fitness function,and the optimal ELM hyperparameters obtained from the optimization are used to establish CryStAl-ELM and POA-ELM models to evaluate the annual water shortage risks.The results are compared with the evaluation results of fuzzy comprehensive evaluation method,CryStAl-SVM,POA-SVM,ELM and SVM models.The results show that:(a) the CryStAl and POA have good optimization accuracy and global search ability;(b) the average absolute percentage error(MAPE) of the CryStAl-ELM and POA-ELM in evaluating test samples are 0.077% and 0.083%,respectively,the evaluation accuracy is more than 57.7% higher than that of CryStAl-SVM and POA-SVM,and more than 83.5% higher than that of SELM and SVM;and(c) the CryStAl and POA can effectively optimize ELM hyperparameters and improve ELM prediction performance.The evaluation results of CryStAl-ELM and POA-ELM models indicate that:(a) the water resource shortage risk in Yunnan Province is as higher-level between 2006 to 2008,medium-level between 2009 to 2012,lower-level between 2013 to 2019 and low-level between 2020 to 2025;and(b) in the past 15 years,the risk level of water resource shortage in Yunnan Province has shown a downward trend,and the downward trend is significant.