基于GF-LSTM和GAN网络的小样本集人工水体溶解氧浓度预测

采用推流方式改善人工水体溶解氧分布不均衡以防止富营养化时,需要对其分布进行预测来提高推流效率,为此构建了基于生成式对抗网络(GAN,Generative Adversarial Networks)和长短期记忆神经网络(LSTM,Long-Short Term Memory Network)的溶解氧浓度预测模型。以广西大学镜湖35 m2的一片水体区域为研究对象,采用不同电压直流水泵推流,用无人船搭载在线检测仪在一段时间内定时定点采集水体溶解氧浓度数据作为原始数据样本,并采用GAN扩充数据样本。利用遗传算法和改进的一阶滤波算法进行溶解氧的噪声数据处理,结合LSTM网络构建溶解氧浓度预测模型GF-LSTM(Genetic And Filtering Algorithm-Long Short Term Memory Network)。结果表明:相比常用的BP网络,GF-LSTM网络预测的平均误差降低了62%,均方误差降低了75%;相比传统的LSTM网络,GF-LSTM网络预测的平均误差降低了22%,均方误差降低了50%。

Forecast for dissolved oxygen concen-tration in artificial water body based on GF LSTM and GAN network

This paper intends to propose a prediction strategy of the dissolved oxygen(DO) concentration in the man-built water bodies based on the long-short-term memory(LSTM) neural network in combination with the generative adversarial network(GAN),genetic algorithm and filtering algorithm to deal with the sampling data. The above mentioned strategy can be expected to improve the efficiency of the man-made flowing method designed to balance the DO distribution to prevent from the water eutrophication. And,we have first of all done the artificial flowing experiments in the direct current pumps of the different rated voltages on the proposed water body( a 35 m2 area of Jinghu Lake in Guangxi University),and,then,an unmanned vessel has been taken to collect the DO data on the designed spots transferred to our computer program as soon as possible,for the vessel is equipped with an online water quality monitoring probe.Such a kind of experiments can be continuously done during 4 days togainsets of original data. And,next,the data sets can be extended by training the original data sets with the generative adversarial networks(GAN) to get the adequate new reliable data of similar nature. Besides,there can usually be found noise in the original data,which may be related to some other hydrological parameters of the dissolved oxygen contents. And,just because of such a need,we have modified the first-order filtering algorithm in addition to the error compensation items,applying the genetic algorithm to optimize the filtering parameters and error compensation parameters. Combined with the above data processing methods,we have also used a novelty LSTM network( GF-LSTM) to predict the DO concentration rate in the water area by comparing it with the other 2 traditional network methods,i. e. BP and LSTM. The experimental results we have done above indicate that: as compared with the commonly adopted BP network methods,the average error of GF-LSTM network prediction can be reduced by 62% with the mean square error being cut off by 75%. On the other hand,the average error of GFLSTM network prediction can be cut short by 22%,in comparison with the traditional LSTM one by reducing the mean square error by 50%. In such a case,GF-LSTM network can be expected to display higher accuracy and better generalizing capability.And,thus,it can be concluded that,based on the predicted DO data,a lot more of data analysis methods,such as the pseudocolor mapping of the data distribution,can be introduced to transfermore filtering algorithm artificially and efficiently.