Research On Deep Learning Prediction Method For Non-point Source Pollution In Watershed Based On Differential Manifold

In recent years,deep learning models had been widely used in the field of time series prediction,and more and more scholars had applied deep learning models to environment field.At present,non-point source pollution output load had become an increasingly serious environmental problem in water environment.Therefore,to guarantee ensure the water resources security effectively,water quality prediction has become an important research content.Due to the problems of redundant parameters and complex structure,the traditional physical models have some limitations in practical application.With the rapid development of science and technology,deep learning models has become the mainstream choice of time series prediction,which can overcome the shortcomings of traditional prediction methods effectively.However,the deep learning models also has some practical problems.When extreme values appear in the data,the prediction of the models are usually affected,and even have large errors.In the prediction of water quality change caused by non-point source pollution,the influence of spatial factors such as slope,vegetation and land use should be considered.At the same time,high-dimensional nonlinear features were presented in these spatial factors,and it was necessary to extract internal key information from the massive and complicated high-dimensional nonlinear data.In order to solve the above problems,the specific work of this thesis is as follows:(1)Firstly,to effectively solve the problem that the non-point source pollution in the basin was affected by spatial factors such as terrain slope,this thesis adopted the spatial feature extraction model of VGG model.More detailed spatial image information could be obtained by the VGG model,and spatial features such as terrain slope could be extracted effectively.In order to better reflect the influence of watershed spatial characteristics on water quality change prediction caused by non-point source pollutions,the spatial feature factors were transformed into spatial high-dimensional time series features by VGG.(2)Then,after the spatial high-dimensional temporal features are extracted by the VGG model,the nonlinear dimensionality reduction method manifold learning(ML)was used for dimensionality reduction.The manifold learning dimensionality reduction method was used to transform the spatial high-dimensional time series features into more compact spatial multi-dimensional time series data,remove irrelevant or subtle influence features,and only retain features that had important impact on the results.Then,the spatial multi-dimensional time series data was combined with hydrometeorological data and pollutant data to form a continuous time series data set,which was used as the data input of the subsequent time series deep learning model GRU,and the multi-source data information was used to improve the accuracy and prediction performance of the model.(3)Finally,a time-series deep learning model for predicting non-point source pollution in watersheds was constructed,and a VGG-ML-GRU combined model was constructed by combining VGG,ML and GRU models.The experimental results showed that the proposed model performed well in extreme value prediction.Compared with the GRU model without considering spatial factors and the VGG-GRU model without considering the dimensionality reduction of spatial high-dimensional temporal features,the proposed model showed effectiveness and superiority.At the same time,the differential autoregressive moving average model(ARIMA),support vector regression model(SVR)and long short-term memory neural network(LSTM)model were constructed for comparative experiments.Taking dissolved oxygen(DO)as an example,the MAE,RMSE and SMAPE of the proposed model were 0.191,0.257 and 0.022,respectively.Compared with other time series comparison models,MAE increased by an average of 28.14%,RMSE increased by an average of 25.10%,and SMAPE increased by an average of 33.56%,indicated that the VGGML-GRU combination model had good prediction performance.(4)In summary,this thesis constructed a combined water quality prediction model based on manifold learning and deep learning models by exploring the spatial feature information of the basin and historical water quality time series data.Experimental results showed that the proposed model could obtain good extreme value prediction results.Compared with other prediction models,the impact of watershed spatial feature information on water quality changes caused by non-point source pollution could be reflected by the proposed model,and the effectiveness of spatial high-dimensional time series feature dimensionality reduction processing could be reflected.This provided a new direction and approach for the control and prediction of non-point source pollution in the future,and also provided an important reference for the protection of water resources security and the healthy development of the ecological environment.