| In recent years,to effectively solve water environment safety issues and promote excellent improvement of water environment quality,ecological environment monitoring stations in all river basins across the country have carried out intensive water quality monitoring to ensure that environmental protection departments dynamically obtain water quality conditions and conduct water quality early warning decision-making analysis.Therefore,how to accurately monitor water quality and provide early warning of water quality data has become one of the current research hotspots.However,since most water quality sensors are usually affected by factors such as routine maintenance,database entry errors,sensor measurement errors,et al,there are usually two types of water quality data errors: noise data and missing data.It is noted that the existing traditional methods are only well applied into simple and small amount of error data cleaning.Taking time series characteristics as the starting point,this paper pay more attention to the study of water quality data cleaning framework from two aspects: adaptive noise removal based on the complex and changeable distribution characteristics in Qiantang River water quality data,and large-scale water quality missing data imputation based on transfer learning.Then,this study carries out the early warning of water quality abnormalities in the Qiantang River Basin as a decision analysis case.The main work and innovations are as follows:Firstly,various types of noise occur in the collected data because the current sensors are susceptible to environmental factors such as measurement errors.Besides,traditional noise removal algorithms have poor anti-interference and robustness when facing different data.Therefore,this paper proposes an adaptive noise removal method based on empirical wavelet transform and multiscale fuzzy entropy,where the empirical wavelet transform can adaptively decompose spectral data,and multiscale fuzzy entropy can construct an adaptive threshold function.The comparison results show that this adaptive method has the better performance of noise removal.Secondly,in view of the problems that missing data usually exists in the collected water quality data,and the traditional imputation methods can only solve the small-scale and random missing data imputation issues.Therefore,this paper takes large-scale continuous missing data as the research object,and proposes a new imputation missing data method based on transfer learning theory and long short-term memory model.The experiments results show that the accuracy of missing data imputation of the proposed method has been greatly improved.Finally,data cleaning aims to improve the accuracy of water quality anomaly warning results.This paper also studies a dynamic water quality anomaly early warning method based on autoregressive integrated moving average and isolation forest.The online dynamic prediction is based on the autoregressive integrated moving average algorithm,and then the isolation forest algorithm is adopted to detect abnormal early warning data points beyond the threshold.The results show that this method has better performance of early warning of water quality abnormalities.In addition,the algorithm has been implemented in engineering as the support of the project’s water quality early warning module. |
