Research And System Integration Of Medium-and-long Range Runoff Forecast Based On CEEMDAN-LSTM Model

Everything is born from water,and water is the source of everything.In recent years,more and more frequent human activities and extreme climates have profoundly changed the atmospheric hydrological cycle,which has also made the randomness,ambiguity and chaos of runoff sequences more obvious.Improving the accuracy of mid-and long-term hydrological forecasting is the key foundation for the realization of reasonable water resources allocation.Based on this,this paper studies the mid-and long-term runoff forecasting methods on the basis of runoff analysis.In this paper,the control station Shazing Ridge Station in Xujiang River,the main branch of the upper Fuhe River Basin,was used as the research object.Using the historical data of 1980-2015 annual runoff and monthly runoff from Shaziling Station in Xujiang River,this paper studied and verified the CEEMDAN-LSTM model in mid-and long-term runoff forecasting,and through the comparison of models under different schemes,summarizes certain model forecasting accuracy rules.The research content and results of this article are as follows:(1)Taking Shaziling Station as an example,the deterministic components in the historical annual runoff series are analyzed,and it is concluded that there is no obvious trend in the annual runoff series.The mutation years are 1991 and 2002,and the main cycle of existence is about 3.5 years.And the conclusion that there are multiple sub-periods with different frequencies reveals the complexity and non-stationarity of the runoff sequence.At the same time,the estimation of the runoff period provides a reference for the selection of the neural network model time window.Characteristic analysis of the monthly runoff series shows that among the historical runoffs,the distribution of runoff in 2010 is the most uneven,and the distribution difference between different years is large;the difference in runoff during the same period in history is obvious,and the unevenness is large;2002(except 2010)The conclusion that the monthly runoff difference is relatively reduced and the overall is relatively stable.(2)The goal is to improve the accuracy of mid-and long-term runoff forecasting,and to provide scientific decision-making basis for the allocation of water resources in the basin.Aiming at the problem of non-stationary and non-linear prediction of runoff series,combined with the characteristics of CEEMDAN algorithm with high reconstruction accuracy and effective overcoming modal aliasing,a runoff prediction model combining CEEMDAN signal processing method with long and short-term memory neural network LSTM is proposed.From the perspective of the two key predictive factors,data and model,the CEEMDAN-LSTM model for mid-and long-term runoff forecasting is considered.By using the original runoff sequence and the steady runoff sequence considering the mutation as the model data basis,the CEEMDAN algorithm and the EMD algorithm decompose the prediction model,the signal decomposition after the superimposed prediction model and the single prediction model are compared in three dimensions,and the model prediction is analyzed.The effect is good,and the prediction accuracy of the CEEMDAN-LSTM model using stationary year and monthly runoff series considering mutation is higher than that of the original year and monthly runoff CEEMDAN-LSTM model;the prediction accuracy of the CEEMDAN-LSTM model to overcome the problems of aliasing modal and reconstruction errors It is higher than the EMD-LSTM model,and its advantages are prominent when the sequence is short;the prediction accuracy of the superimposed prediction model after signal decomposition is significantly improved compared to the single prediction model.(3)Based on the above research results of runoff forecasting models,a set of demonstration systems for mid-and long-term runoff forecasting in the basin is designed and implemented.The background model data processing fully considers the actual prediction situation,changes the original data unified normalization method,and adopts the maximum and minimum normalization method based on the maximum and minimum values of the training set data;the system increases the parameter rate The grading function can set a set of parameters more in line with the characteristics of runoff according to the real-time update of the data,which increases the life cycle of the system.Aiming at the difficulty of secondary development and complicated installation of traditional system design,the system adopts B/S architecture,the front-end interface uses lightweight Vue.js framework,and the back-end uses Python-based Flask framework to separate front-end and back-end development,enhancing the system's Scalability and portability.