Research On Cyanobacteria Density Prediction Technology In Taihu Lake

The cyanobacterial blooms in Taihu Lake seriously affect the lives of residents along the lake,and it is very important to predict the cyanobacteria density in advance.Although many cyanobacteria density prediction models,based on traditional methods,have been established.The prediction accuracy of these models still needs to be Improve.This study is mainly different from the traditional cyanobacteria density monitoring methods.It adopts deep learning and transfer learning methods to model the cyanobacteria density through environmental variables in the Taihu Lake waters to achieve accurate prediction of cyanobacteria density.The main research work is carried out from the following three aspects:1.Firstly,research on the prediction method of cyanobacteria density in Taihu Lake,based on deep learning is done.This research method is aimed at the prediction of cyanobacteria in Taihu Lake at a single monitoring point.A deep learning model 1D CNN-SLSTM-ATTE model is proposed to predict the density of cyanobacteria in Taihu Lake.In the process of establishing the model,after correlation analysis and wavelet transform of the data,the traditional long-short term memory(LSTM)model is used.The LSTM model is improved,adding the 1D Convolutional Neural Networks(1D CNN)and Attention Mechanism(AM).The activation function is also improved.Compared with traditional machine learning and deep learning algorithms,the prediction results of the 1D CNN-SLSTM-ATTE model are more accurate.The root mean square error(RMSE)decreased by 272.21cell/ml on average,the mean absolute percentage error(MAPE)decreased by 26.23%on average,the mean absolute error(MAE)decreased by 142.28cell/ml on average,the Nash-Sutcliffe efficiency coefficient(NSE)increased by 0.217 on average.2.Secondly,the research on the prediction method of cyanobacteria density in Taihu Lake based on transfer learning is done.This study is aimed at the prediction of cyanobacteria in Taihu Lake at multiple monitoring sites.Deep learning and transfer learning methods is applied to predict the trend of cyanobacterial density for multiple targets by establishing a large available dataset.Based on this data set,the pre-trained cyanobacteria density,including deep learning algorithms,such as Residual Network(Res Net),Gated Recurrent Unit(GRU)and Multi-Head Attention mechanism,was firstly established.The knowledge,gained and retained by the pretrained model,is then transformed into a cyanobacteria density prediction model for the target system with a much smaller amount of available data.In order to reflect the advantages of transfer learning,the accuracy of a cyanobacterial density prediction model with the same structure was established without transfer learning by using the data of the target system itself.Compared with the traditional transfer learning,based on the deep learning model,after using the transfer learning model studied in this paper,the RMSE,MAPE,and MAE,predicted by cyanobacteria,decreased by 55.67cell/ml,12.67%,and30.71cell/ml,respectively,and the NSE increased by 0.126.3.Thirdly,the research on cyanobacteria bloom prediction method based on convolutional long short-term memory neural network is done.In this study,aiming at the prediction and monitoring of cyanobacteria density in many lake bays of Taihu Lake,based on the convolutional long short-term memory(Conv LSTM)and CNN algorithm,a deep learning model CNN+Conv LSTM+CNN for cyanobacteria density prediction is proposed.The model is compared with the deep learning model Conv LSTM,which only considering spatiotemporal features,the deep learning model CNN,which only considers spatial features,and the deep learning models LSTM,Bi LSTM,and 1D CNN+LSTM,whcih only consider temporal features.The test results show that the CNN+Conv LSTM+CNN model finally reduces RMSE,MAPE,and MAE by 98.56 cell/ml,7.13%,and 74.59 cell/ml,respectively,and improves NSE by 0.33.