Water Quality Prediction Methods Based On Wavelet Decomposition And Deep Neural Network

Water quality is an important basis for social activities and human health.China has always attached great importance to water quality assurance.Water quality prediction is based on the historical data of water quality parameter,using various methods to extract its trend and predict the future water quality,such as some water indexes or water quality grade.For the security of water environment,water quality prediction can provide important data support and scientific basis for relevant government departments in water management and risk prevention.At present,in the field of water quality prediction,the multi-step prediction of water quality parameter and water quality evaluation grade still needs to be developed.This study combines wavelet decomposition and artificial neural network,so that it can strengthens the data characteristics and excavates the data connotation.This study establishes a prediction model of water quality parameters and a prediction model of water quality grade with high accuracy and strong applicability,which effectively improves the prediction effect.This study focuses on the problem of water quality prediction,it combines wavelet decomposition and deep neural network to improve the prediction effect of water quality parameters and water quality grade.The main contents of this study can be divided into three aspects as mentioned below.First,a multi-step prediction model of water quality parameter based on historical time series is proposed.Aiming at continuous time series,this model uses wavelet decomposition to decompose the original sequence data into signals with two dimensions in time and frequency,it can separate high and low frequency information,and reduce the data noise.This model uses the encoder-decoder architecture of recurrent neural network to realize multi-step prediction,it also integrates the attention mechanism into the encoder-decoder architecture to adaptively focus on different time steps in the input series.Using all of the above techniques,this model realizes the endto-end multi-step prediction task,and improves the accuracy and stability of prediction results.Second,an end-to-end prediction model from historical multi-dimensional water quality parameter data to future water quality grade is proposed.This model proposes a dual-stage attention mechanism of time-feature dimension.By combining the attention mechanism with the encoder-decoder architecture,temporal attention is realized,which ensures the grasp of the internal fluctuation in a single parameter time series;by using multi-head self-attention mechanism,featural attention is realized,which ensures the feature learning of multi representation subspace in several water quality parameters,so as to improve the accuracy and minority category sensitivity of the model.Third,empirical study on water quality data analysis and water quality prediction for the Yangtze River Basin is done.Based on the weekly monitoring data of 23 water quality monitoring stations in the Yangtze River Basin from 2004 to 2018,the above two models are used for empirical prediction of water quality parameters and water quality grades.The experimental result show that these two models have better reliability and effectiveness.At the same time,performance differences and variation characteristics in various scenarios of these models are discussed in depth.