Research On Marine Water Quality Prediction Model Based On LSTM Hybrid Network

The ocean is the largest ecosystem on earth and an important foundation for human survival and development,and is also an important part of national development strategies.In recent years,with the development of marine economy,marine water quality pollution has become increasingly serious,and in order to effectively respond to marine water pollution,it is necessary to accurately predict marine water quality.Marine water quality data has the characteristics of large volume,many species and high latitude,so the traditional time series prediction cannot meet the prediction requirements.In order to improve the accuracy of marine water quality prediction,this paper is based on machine learning and neural network prediction technology,and the research of combined marine water quality prediction model is carried out:Firstly,from the basic theory of seawater quality prediction,the necessity of machine learning in water quality prediction is analyzed,and the use of Long Short-Term Memory Network(LSTM)is proposed to predict marine water quality data;the source and collection of data are introduced,and the data set is constructed by processing missing values,outliers and normalization of the data,which provides data support for the construction of marine water quality prediction model.Secondly,for the marine water quality prediction problem,a combined prediction model based on XGBoost and LSTM is proposed,which improves the computational speed of the LSTM model and reduces the prediction error of the model.The combined model prediction index is determined by principal component analysis of marine water quality monitoring data,and the XGB and LSTM models are trained and combined with the XGB-LSTM model by the inverse of error method;the analyzed metrics are input to the combined model for prediction and compared with the individual XGB and LSTM,and the results show that the XGB-LSTM hybrid model has higher prediction accuracy and improved computing speed.Finally,for the problem of gradient explosion in LSTM networks in long-term water quality prediction,a combined SDPA-SLSTM model including a stacked long short-term memory network(SLSTM)and a Scaled Dot-Product Attention mechanism is proposed.SLSTM further enhances the stability of the model and improves the data processing capability by stacking multiple LSTMs to effectively satisfy the complex input data,while SDPA enhances the attention to key time series by assigning weights to different variables and effectively solves the long-range information loss problem.The input data obtained by the dimensionality reduction process are compared with the experiments,and the results show that the SDPA-SLSTM model effectively improves the prediction accuracy.