基于MA-FOSELM-OTF的海洋养殖环境数据在线预测

Online prediction of marine aquaculture environmental data based on MA-FOSELM-OTF

为降低过时海洋养殖环境数据对后续学习的影响和解决单一模型随机初始化输入权重的问题,提出一种基于MA-FOSELM-OTF的海洋养殖环境在线预测模型.采用模型平均化(model averaging,MA)算法对全在线顺序极限学习机(fully online sequential extreme learning machine,FOSELM)进行集成,以降低输入权重初始化引起的随机性;在FOSELM中引入过时遗忘机制(obsolete to forget,OTF),对过时的数据进行遗忘加权,降低其对顺序学习的影响;利用FOSELM递推计算所得输出结果集成所有输出结果,取其平均值作为MA-FOSELM-OTF在线预测模型的最终输出.结果表明,MA-FOSELM-OTF在海洋养殖环境数据在线预测任务中的预测性能优于其他对比模型,可为海洋养殖预警平台提供参考.

In order to reduce the influence of outdated marine aquaculture environmental data on subsequent learning and solve the problem of randomly initializing the input weight of a single model,an online marine aquaculture environment prediction model based on MA-FOSELM-OTF is proposed.This model uses the model averaging(MA)algorithm to integrate the fully online sequential extreme learning machine(FOSELM).To reduce the randomness caused by input weight initialization;This paper introduces obsolete to forget(OTF)mechanism in FOSELM to weight the obsolete data,which can reduce the influence of obsolete data on sequential learning.The output results obtained by FOSELM recursion calculation are integrated with all the output results,and the average value is taken as the final output of MA-FOSELM-OTF online prediction model.The results show that the prediction performance of MA-FOSELM-OTF is better than other comparison models in the online prediction task of marine aquaculture environmental data,which can provide a reference for the marine aquaculture early warning platform.