Research On Migration Of Watershed Runoff Model Based On Geographic Similarity

The forecast of accurate watershed runoff plays a significant role in water resources management,flood warning,and other tasks.The existing hydrological modeling methods can be classified into two categories: physics-based models and data-driven models.Data-driven models have advantages with fewer restricted conditions and shorter forecast periods comparing with physics-based models,which are mainly used for daily runoff prediction to satisfy the requirements of high precision and high reliability.However,data over a period of time and with several key types of monitoring are necessary when the watersheds runoff prediction model is built regardless of whether it uses a physics-based model or a data-driven model.At present,some basins have good monitoring equipment and can provide important data,so various modeling methods can be used.However,some watersheds are difficult to build the runoff predicting model since they lack the key monitor sensors to result in necessary information data not being obtained.Therefore,it is an urgent problem how to efficiently forecast runoff in the ungauged watershed.To solve the problem,this paper uses the CAMELS watershed dataset as the study area,the recurrent neural network is employed to construct the watershed runoff prediction model.The autoencoder models are used to construct the extraction model of the watershed spatial-temporal comprehensive characteristics,and the geographic similarity of watersheds can be measured.Moreover,the relationship between the migration performance of runoff prediction model and the geographic similarity of the watershed is analyzed.The proposed approach can be used as an alternative method for runoff prediction in ungauged basins worldwide.The main research contents and results in this paper can be summarized as follows:(1)Constructing watershed runoff prediction model based on a machine learning methodIn order to achieve high-precision prediction of the daily runoff sequence of the watershed and provide support for the subsequent model migration performance research,the characteristics of typical machine learning models for runoff prediction are compared and analyzed.The recurrent neural network model is used to construct runoff prediction models based on the CAMELS watershed dataset.The migration performance can be obtained by transferring the models to other basins.The results show that the recurrent neural network algorithm can achieve high-precision prediction of daily runoff.Its migration performance is affected by the applicability of the model and the differences between river basins.(2)Extracting spatiotemporal comprehensive characters of watersheds and measuring geographic similarityIn order to obtain the watershed features used to measure the geographic similarity of watersheds,a convolutional autoencoder was designed based on the principle of autoencoders to extract watershed spatial features using watershed geography and land use data,and the spatial pyramid pooling algorithm was used to obtain the hierarchical spatial features of the watershed.At the same time,a time series auto-encoder is designed to realize the extraction of watershed time characteristics using watershed hydrological time series.On this basis,a sparse autoencoder is used to design a watershed spatiotemporal comprehensive feature extraction model that integrates watershed time and space characteristics,and the watershed spatiotemporal comprehensive feature is obtained.Since the comprehensive spatial and temporal characteristics of the watershed are not directly related to the migration performance of the recurrent neural network watershed runoff prediction model,a genetic algorithmbased model for the selection and weighting of comprehensive temporal and spatial characteristics of the watershed is designed.Based on this,the temporal and spatial characteristics of the watershed that can support the migration of the recurrent neural network runoff prediction model are obtained,and the geographic similarity of the watershed is calculated by using various geographic features of the watershed.(3)Analysis of migration performance of runoff prediction model based on geographic similarityIn order to use the geographical similarity of the watershed to assist the migration of watershed runoff prediction model,a recurrent neural network runoff prediction model is designed,which includes six aspects: coverage rate of recommended results,availability rate,average percentage error of Nash efficiency of recommended results,the available number of recommended results,the joint distribution of migration percentage error and various types of similarity,and joint distribution of comprehensive spatiotemporal similarity and other types of similarity.Based on this,the relationship between basin geographic similarity and model migration ability is analyzed,and the influence of basin geographic similarity calculated by different characteristics on runoff prediction model migration ability is compared and analyzed.The results show that when 25 recommendation results are provided,the coverage rate of recommendation results is 16.8%,the availability rate is 34.3%,and the average percentage error of Nash efficiency is 30.9%.For the top 5% of the similarity,the availability of spatiotemporal similarity is more than 73%.The comparative analysis shows that the comprehensive spatiotemporal similarity is better than other types of similarity,which can better assist the target migration watershed selection of recurrent neural network.