Water Quality Prediction Models Of Yellow River Basin Based On Deep Learning

Water is the basis for human and other organisms’ survival,growth,and development.The national strategy for ecological protection and high-quality development of the Yellow River basin and other watershed management policies have also put forward new requirements for precise pollution prevention and control in the basin.Establishing a valid and reliable water quality prediction model can help water environment managers grasp the trend changes in water quality in a timely manner and provide a basis for decisionmaking for the control of water pollution in the basin.But the current monitoring station collection of water quality data has many vacant and abnormal values,and other noise data will affect the data distribution of the water quality series.And in the water quality data,there are sudden changes in the prediction points that can impact the value,which needs to focus on.This thesis takes the Lanzhou section of the Yellow River basin as the study area,uses water quality data collected from several monitoring stations of Gansu Provincial Environmental Science and Design Institute,and uses deep learning methods for single-step and multi-step prediction models of water quality in the Yellow River basin.The main work includes:(1)Research the noise reduction method of watershed water quality series based on SG filter.Affected by the measurement environment and measurement instruments,there will be some vacant data and abnormal data in the data set,resulting in the trend characteristics in the data cannot be accurately characterized,affecting the prediction accuracy of the model.In this thesis,we first use the linear interpolation method,isolated forest algorithm,and data normalization method to pre-process the data set;then construct the SG-LSTM water quality prediction model composed of SG filter and LSTM network by comparing with the water quality prediction methods been built by different filters to test the sequence noise reduction effect of varying noise reduction methods.The results show that: the constructed SG-LSTM water quality prediction model has the highest accuracy,and MAE is 0.26.The model can capture the subtle and rapid changes in the time series data and reduce the noise while retaining adequate information,improving the trend characteristics of water quality data.(2)A single-step water quality prediction model for watersheds based on attention mechanism and Bi-LSTM.Water quality data has time series characteristics,and existing research uses time series deep learning models such as LSTM to carry out water quality prediction research.Still,it does not consider the role of reverse time series on the model.This thesis combines the attention mechanism and Bi-LSTM network to construct an ATBILSTM model to model the watershed single-step water quality prediction model on the water quality data after SG filter processing.The model mainly contains the Bi-LSTM layer and temporal attention layer.After the bi-directional feature extraction of water quality time series data,the attention mechanism is introduced to highlight the data series that critically impact the prediction results.The results show that: compared with the benchmark model,the RMSE and MAE of the model are reduced to 0.101 and 0.059,respectively,with the best prediction performance in the water quality single-step prediction task of four cross-sections.The model can effectively focus on the essential features that impact the prediction node and make a more accurate prediction of the water quality data in the next 1 hour.(3)A multi-step water quality prediction model for watersheds based on DILATE loss function and Transformer network.The multi-step water quality prediction is more practical.Still,the Bi-LSTM network used in the AT-BILSTM model is usually computed considering the extended position along the sequence.The adequate information will be gradually compressed with the increasing number of input features,generating the fundamental constraint problem of sequential computation.And the loss function based on the vertical Euclidean distance does not consider the existence of time distance,which will produce a certain time lag.In this paper,we use the Transformer network to improve the sequential computation problem existing in Bi-LSTM,retain the attention mechanism feature,and introduce DILATE loss function to calculate the error loss from both shape and time aspects to improve the time lag problem.The results show that the constructed SG-Transformer model has the best prediction performance with RMSE and MAE improved to 0.153 and 0.123 in the multi-step prediction task at T+3 moments.The model can give extra attention to different moments of water quality data,better grasp the serial correlation while improving the generalization ability,and be the more accurate multistep water quality prediction.