Research On Multi-Section Water Quality Prediction Method Of Watershed Based On Deep Multi-Task Learning

Efficient and real-time water quality prediction is the main decision-making basis for the precise prevention and control of water environmental pollution in river basins,which is of great significance for promoting the protection of ecological environment and high-quality development of river basins.The water environment of the basin is a grey system affected by many factors,which has complex multivariate correlation.Meanwhile,the water mobility of the basin makes the pollution of the upstream section have a certain influence on the water quality monitoring of the downstream section.At present,the prediction analysis of water quality mainly focuses on a single section of the basin,without considering the influence of correlation between different sections on water quality monitoring,lack of correlation analysis between the spatial and temporal characteristics of multiple sections of the basin.Aiming at the above problems,we takes the concentration of Chemical Oxygen Demand(COD)in four sections of Lanzhou Section of the Yellow River as the research object to carry out a study on the multi-section water quality prediction method of the basin.Specific studies include:(1)Deep learning-based watershed water quality prediction methods and data collection.From the perspective of data-driven watershed water quality prediction,a multi-sectional water quality prediction method based on deep learning is presented.The data collection,model selection,evaluation index and prediction process methods in watershed multi-sectional water quality prediction are analysed and studied,and a watershed water quality multi-task learning prediction method system considering the relevance of multi-basin sections is proposed.For the Lanzhou section of the Yellow River basin,four sections of water quality data were collected and pre-processed and correlation analysis was carried out,and the results of the study showed that the COD between the four sections had time series correlation,providing a methodological system and data basis for subsequent research work.(2)Research on water quality prediction model of river basin multi-section based on deep multi-task learning.To address the problem of insufficient consideration of the correlation of pollution characteristics between cross-sections in the existing water quality prediction of single cross-sections,a deep multi-task learning based multisectional water quality prediction model(MTL-CNN-LSTM,MCL)is proposed.Water quality prediction tasks of multiple sections are put into the model training at the same time to share water quality information of each section while retaining the heterogeneity of each.In addition,using the CNN-LSTM hybrid model to better exploit the timedependent features between water quality information from short-term to long-term.The model was applied to four sections in Lanzhou section of the Yellow River.The experimental results show that,compared with the COD prediction results of CNN,LSTM and CNN-LSTM at time t+1,MAE and RMSE of the model are the lowest.The MAE and RMSE values of the model decreased by 46.7% and 41.5% respectively in the overall prediction of four sections.(3)Research on water quality prediction model based on integrated empirical modal decomposition of watershed multi-section.The MCL model does not take into account the lag problem caused by the inconspicuous period of water quality data,and proposes a deep multi-task water quality prediction model(MTL-EEMD-LSTM-Attention,MELA)based on integrated empirical modal decomposition.The integrated empirical modal decomposition method is chosen to "decompose-reconstruct" different sections of data at the same time,smoothing the data to solve the model lag problem,and adding attention mechanism to the deep learning part of the MCL model to prevent the decay of water quality sequence data,which makes the model’s prediction effect better.The experimental results show that the MAE and RMSE values of this model are reduced by 46.1% and39.8% respectively compared to the MCL model,indicating that the inclusion of the EEMD method can improve the prediction performance of the model by reducing the coupling of different cycles of data in the original data series.And compared with existing deep learning prediction models,the MELA model is able to predict multiple crosssections of the basin with higher accuracy.