Research On Runoff Prediction In Data-scarce Areas Based On Few-shot Learning

Runoff simulation and prediction is a basic problem in the field of hydrology and water resources.Accurate and reliable runoff simulation and prediction is beneficial to deeply understand the evolution of the hydrological system in the basin,and provide decision support for scientific water resources planning and management.However,traditional hydrological models rely on a large number of inputs and calibration of parameters,and their application in data-scarce areas is limited.The process-based datadriven hydrological model models and fits the input and output through the black-box model,which reduces the difficulty of hydrological simulation and prediction in datascarce areas and improves the feasibility.However,due to the spatiotemporal heterogeneity and uncertainty of river systems,as well as the scarcity of data in datascarce areas,the large amount of supervised data and prior knowledge required for process-based data-driven models cannot be met.Humans are capable of rapid learning of new knowledge and cognitive reasoning of new concepts based on a small amount of data.Inspired by it,few-shot learning and metalearning techniques have developed rapidly,and a new paradigm of machine learning has emerged that utilizes a small amount of supervised data for rapid learning.Data-driven hydrological models based on few-shot learning and meta-learning are beneficial to improve the accuracy of runoff and flood forecasting in data-scarce areas,increase their generalization and robustness,and reduce the model’s dependence on data volume.Metric learning is an implementation of the few-shot learning task.Few-shot learning is performed by learning a metric space.In this thesis,a few-shot（LSTMPrototype）fusion model based on metric learning is proposed and used for runoff prediction in data-scarce areas.By integrating the metric learning algorithm and the Long Short-Term Memory（LSTM）network,the dependence of the data-driven model on the amount of data is reduced.The LSTM is used to extract the time-dependent features of the input and output,and the metric learning algorithm aims to guide the model to learn a more accurate metric space.Taking the Lancang River Basin as an example,the experimental results of monthly runoff prediction in this basin show that:（1）The LSTMPrototype fusion model proposed in this thesis is superior to the five data-driven models（LSTM,SVR,ARMA,random forest,ANN）,the best NSE of the fusion model is 0.86.（2）Comparing the fusion model with LSTM,the results show that in the case of data scarcity,the metric learning strategy effectively improves the performance of LSTM and reduces the dependence of the model on the amount of data.Aiming at the problem of insufficient prediction performance of LSTM-Prototype fusion model for peak runoff,this thesis combines the idea of meta-learning and the network parameter optimization algorithm of Model-Agnostic Meta-Learning（MAML）on the basis of the original LSTM-Prototype fusion model.A meta-learning-based fewshot（MAML-LSTM-Prototype）fusion model is proposed for flood event prediction in data-scarce areas.Taking the Yellow River source area as the experimental basin,the experimental results show that:（1）The MAML-LSTM-Prototype fusion model proposed in this thesis effectively improves the prediction performance of the original fusion model in flood events.The indicators are more stable,the best prediction accuracy is 100%,and the model generalization is improved.（2）After 1000 repeated stability experiments,comparing the MAML network parameter optimization and random initialization of the fusion model,the former shows very strong stability,the results of 1000 experiments are exactly the same,and the stability of random initialization parameters is insufficient,which shows that the fusion model has better robustness.