A Study On The Prediction Of Chlorophyll Concentration In Water Based On Temporal Convolutional Network

Human industrial and agricultural activities have caused certain disturbances to rivers,coasts,etc.,and more and more waters are becoming eutrophic,which greatly increases the risk of algal blooms.Chlorophyll concentration is an important indicator of eutrophication and plankton density,and an accurate prediction of chlorophyll concentration can be of great significance for water environment management and the prevention of algal blooms.Chlorophyll concentration is influenced by a combination of factors and has complex properties such as time-series and chaos.Traditional prediction methods can only learn the linear features of them.A single algorithm prediction model is able to learn the nonlinear features,but the feature extraction mode is single.The effect of using the features of a single mode to predict is not ideal.In order to more fully extract the hidden patterns of water quality data,this paper investigates the single-step and multi-step prediction models of chlorophyll concentration based on deep learning principles and techniques,and the main research contents include:Firstly,pre-process the multiple water quality parameters collected by the marine buoy monitoring system.Find and replace the abnormal values and missing values contained in the original water quality parameters to improve the quality of data.Using min-max normalization to dimensionless processing of multivariate water quality parameters to solve the problem of comparability of different parameters.Use Pearson coefficient method for feature selection,select the water quality parameters with high correlation with chlorophyll concentration,eliminate the interference of irrelevant parameters.Data pre-processing provides reliable data for data modeling.Secondly,for single-step prediction of chlorophyll concentration,a DAE-TCN(Deep auto encoder Temporal convolutional network)prediction model is used.The Auto-Encoders are embedded in the Temporal Convolutional Network to fuse information on multiple input water quality parameters to enhance the spatial feature extraction capability of the model,while the dimensionality reduction effect of the encoder relieves the computational burden of the temporal convolutional hidden layer.One-dimensional convolution with the incremental expansion coefficient of the convolution layer can extract temporal features from multiple scales and increase the richness of temporal features.Finally,the spatio-temporal features are mapped to the target space through the fully connected layer to obtain the prediction results.Comparative experiments were conducted with the water quality dataset of the coastal waters of the North Sea.The results show that DAE-TCN has the highest prediction accuracy,the improved speed,and has better stability.Finally,for multi-step prediction of chlorophyll concentration with more complex mapping relationship,Graph-TCN(Graph-Temporal convolutional network)prediction model is used.It mainly includes graph convolution and temporal convolution structures.Graph convolution adaptively learns the influence relationship between water quality parameters on the graph structure and updates the state of water quality parameters.Temporal convolution mainly captures the rich temporal features.In addition,in order to simplify the complex mapping relationship,two model components,sequence decomposition and period enhancement,are designed.The sequence decomposition decomposes the original input data into trend components and period components,and the period enhancement module reduces the interference of noise on the period components and helps the feature extraction structure of the model to extract data features more carefully.Comparative experiments were conducted with the water quality dataset of the coastal waters of the North Sea.The results show that Graph-TCN is able to give longer output while maintaining a certain prediction accuracy condition.The beneficial effects of each model component were demonstrated by ablation experiments.