Medium And Long-term Runoff Prediction Based On Deep Learning

Streamflow in the water resources system is an important component of the hydrological cycle in the river basin,and reliable streamflow prediction is of great significance to water resources management and water resources dispatch.However,the hydrological process is a complex process involving hydrology and meteorology.It is affected by both deterministic factors and random factors.The hydrological process exhibits non-linear and non-stationary characteristics.In this context,high-precision runoff prediction is full of challenges.How to use limited data to improve prediction accuracy and prediction stability.Establishing a prediction model with good performance and robustness has become a hot topic in related professional research.This paper takes the monthly streamflow of Shangjingyou Station and Fenhe Reservoir Station from 1958 to 2016 as the research objects,and explores the prediction effects of the single Extreme Learning Machine(ELM)model,the single Least Squares Support Vector Machine(LSSVM)model and the single Gated Recurrent Unit(GRU)model.In addition,the improved Complete Ensemble Empirical Mode Decomposition(ICEEMDAN)and the Variational Mode Decomposition(VMD)and the Improved Grey Wolf optimizer(IGWO)are used to establish combined forecasting models.The two prediction structure models are by established by sequential monthly runoff sequence and single-month monthly runoff sequence.This paper studies the applicability of these models and provides new methods for streamflow prediction.The main contents and results of the research are as follows:(1)The monthly streamflow of the four stations on the upper reaches of the Fenhe River is tested and analyzed for non-stationarity,using the ADF(Augmented Dickey-Fuller)unit root test and the PP(Phillips&Perron)unit root test method.The result is that the statistics of monthly streamflow are all more than 5%critical value,which quantitatively describes the non-stationarity of the monthly streamflow time series.(2)In the single model,the deep learning model can better simulate the internal relationship of the streamflow time series.Compared with the shallow neural network methods(ELM and LSSVM),the GRU model test period prediction has better results,where the indexes all show the superiority of the deep learning model.(3)In the "decomposition-prediction-collection" model,VMD or CEEMDAN are combined with LSSVM model.The VMD method is more accurate than the combined model of the CEEMDAN decomposition method,and the variable mode and improved empirical mode decomposition are more capable of obtaining the hydrological information contained in the monthly streamflow series.The VMD-LSSVM model achieves the highest accuracy in all the comparison models during the calibration period and the test period.It can also draw a spectrogram for each component decomposed by VMD to find out the main frequency component and other components corresponding to each component.In the and comparison of the two frequency components'weights and prediction error,the main frequency component have high accuracy and high weight,which can reflect the advantages of the"decomposition-prediction-collection" model.(4)For the double-processing strategy prediction model,the noise reduction method can remove the noise in the streamflow series.Then decomposed the original series and combined the deep learning model for monthly streamflow prediction.The deep learning model based on the ICEEWT method(ICEEWT-IGWO-GRU)has an average increase of 43%in the NSE index compared with the single GRU model.The qr of ICEEWT-IGWO-GRU model obtained at the Fenhe Reservoir Station is 97.22%and at the Shangjingyou Station is 86.11%,which respectively correspond to the "very good" and "good" prediction accuracy in the hydrological forecasting.(5)Compared with the sequential streamflow prediction model,the monthly streamflow prediction model has a different sequence structure.In terms of the prediction accuracy,the monthly IGWO-GRU model is obtained the largest r,NSE,and smallest RMSE values,among the sequential IGOW-GRU model,the single ELM and the single LSSVM model.In addition,the qr values predicted of the monthly IGWO-GRU model by the two hydrological stations(Fenhe reservoir station and Shangjingyou station)are within "very good" and "good",respectively.