Analysis And Application Of Runoff Time Series Based On Transfer Learning

Our country has relatively abundant hydropower resources,but the utilization rate of hydropower resources is relatively insufficient.Reliable forecasting of runoff time series can assist in decision-making on power supply schemes,optimize power generation resource allocation,and improve the utilization rate of water resources.Meanwhile,it has certain guiding significance for the Program of Flood control and drought resistance.In recent years,researchers have made brilliant achievements in the study of runoff forecasting,but the following problems still need to be solved:(1)The runoff sequence does not have obvious regularity,so sequence decomposition is required,but the traditional decomposition method is greatly affected by the endpoint value.(2)In reality,the amount of hydrological data is small,and the data of other hydrological monitoring stations with different distributions is difficult use.(3)In actual production,classic machine learning algorithms are often used for modeling,and these algorithms often cannot adequately represent the complex process of runoff changes.In response to the above problems,this thesis mainly did the following work:1.Propose a method of using approximate sequence extension value instead of unknown value of the endpoint to sequence decomposition for runoff sequence decomposition.The runoff sequence does not have obvious regularity,so it is necessary to decompose the runoff sequence to separate the periodic factors and random factors in the runoff sequence.The decomposition result of the traditional empirical mode decomposition algorithm is greatly affected by the unknown value of the endpoint.The approximate sequence of the original sequence is obtained by using Fourier transform,and the unknown value replacement value is obtained by extending it,so that the decomposition result is closer to the decomposition result of the original sequence.2.Propose a direct runoff forecasting model based on transfer learning.The model uses the transfer learning method combined with long short memory network or convolutional neural network to establish a model to directly predict that the runoff is high flow or low flow.The model process is divided into two stages: cyclic training and fine-tuning.In the cyclic training phase,the runoff data of other hydrological stations are sorted from small to large in order of similarity,and then the data of a hydrological station and the data of the target station are subjected to domain adaptation and all the processed data is sent in the model for training each time.In the fine-tuning stage,the data of the target site is input into the model for fine-tuning.This model makes full use of the data of other hydrological sites,so that the deep learning model can be modeled with less data.3.Propose an indirect runoff forecasting model based on transfer learning.Due to the high correlation between the adjacent months in the dry season,the magnitude of the change is relatively small;the flood months fluctuate sharply and the changes are more complex,so it is necessary to divide the flood season months and the dry season months.Firstly,the model divides the months within a year into flood season months(May to October)and dry season months(January to April,November to December),then calculates the runoff of all months respectively,finally calculates the type of runoff.The migration learning method is combined with the long-short-term memory network or the convolutional neural network for modeling in the flood season,and the linear regression model is used for the modeling in the dry season.The flood season model process is divided into two stages,the loop training stage and the fine-tuning stage,which are basically similar to the direct forecasting model.The model makes full use of the annual change discipline of runoff,making the forecast target more detailed.Experimentally,this thesis draws the following conclusions:(1)The method using approximate sequence extension value instead of unknown value to sequence decomposition has a significant improvement compared with the original decomposition algorithm.(2)The transfer learning method can solve the problem of deep learning model modeling under the premise of less runoff data.Compared with the traditional method,the effect is improved by more than 10%.(3)The indirect forecasting scheme forecast has a better effect on the forecast of the two types of abundance and dryness.This thesis mainly studies the prediction of runoff richness and dryness combined with transfer learning,and two transfer models are proposed to complete the prediction target.The models called transfer-LSTM and transfer-CNN.They use the Fourier-CEEMDAN decomposition algorithm,transfer learning algorithms,and deep learning models for modeling.Moreover,two methods,respectively direct method and indirect method,are used to predict runoff richness and dryness.The models solve the problem of difficulty in modeling deep learning models caused by the small amount of data in runoff sequence prediction and the end effect problem of traditional decomposition algorithms.At the same time,the granularity of prediction is refined.All of these are useful for improving runoff.The accuracy of the forecast of abundance and drought has obvious effects.Compared with the original method,the accuracy of the new decomposition method is increased by 5%,the accuracy of the direct prediction scheme of the migration model is increased by 11% compared with the traditional model,and the accuracy of the indirect prediction scheme of the migration model is increased by 6% compared with the traditional model.