Research On Intelligent Prediction Method Of Hydrological Time Series For Small Sample Data Set

Accurate hydrological time series prediction provides scientific basis and guidance for flood prevention and disaster reduction,reasonable allocation and scheduling of water resources.Monthly/annual precipitation and runoff prediction in hydrological prediction are all small sample prediction problems.At present,a large number of studies pay attention to mining the inherent logical relationship of large-scale sequence data,but ignore the small sample data with clear problem tendency.In order to make accurate prediction and intelligent decision,it is necessary to capture the rules with large sample data and match the scene with small sample data.However,the lack of data volume cannot provide the data space for deep learning model to learn,resulting in unsatisfactory prediction accuracy.Therefore,it is urgent to seek a time series prediction model suitable for small sample dataset.In order to improve the accuracy of hydrological time series prediction on small sample dataset,this paper attempts to improve and extend the deep learning model,and proposes a prediction model suitable for small sample dataset.The main contents and innovations are as follows:1)Several research have been conducted on the nonlinearity,instability,and severe volatility of hydrological time series,and it has been found that compared to time series data with large amounts of data,the volatility and nonlinearity of hydrological time series with small sample data sets are more significant.In this regard,two hydrological time series datasets with different data volumes at the same hydrological station(monthly rainfall data and daily rainfall data at Badong Station)were selected as the research basis,and Variational Mode Decomposition(VMD)was introduced for data preprocessing.There is currently no uniform standard for defining the K value of the VMD decomposition mode number.Through collation and induction,three applicable rules have been summarized.The hydrological time series is decomposed into multiple sub signals according to different central frequency information,and the temporal and spatial feature information extracted from multiple sub signals is preliminarily analyzed and verified using the optimal depth learning combination model.The experiment found that the deep learning model requires a large amount of data and is not suitable for the prediction and research of hydrological time series based on small sample data sets.Moreover,the deep structure of the deep learning combination model is greatly affected by data volatility,and there are problems such as hysteresis,inability to learn the change rules of data steep rises and falls under extreme weather conditions,and the prediction accuracy does not meet the reliability standard requirements.2)VMD provides a good training environment for processing nonlinear data,but still exists many problems in the combination model.VMD decomposition is greatly affected by the number of decomposition modes and the lagrange multiplier update step.At present,only the best mode number is found.In order to seek higher prediction accuracy,more stable model performance and faster operational efficiency,an improved VMD model(IVMD)is proposed。The update step in the VMD decomposition is optimized by combining the Residual Index Correction method(RIC).RIC is improved on the basis of the Mean Absolute Error,and uses the update step value as the iteration number to calculate the loss of the decomposition results and the original data after the superposition and reconstruction under the unsynchronized length to find the optimal update step.At the same time,it is noted that the Broad Learning System(BLS)has the advantages of solving the global optimization,directly calculating the weight,simple calculation and less parameter quantity,so it is considered to introduce the BLS model and the Long and Short-term Memory network model(LSTM).BLS belongs to the width network structure,which is different from the deep structure of the Convolution Neural Network model(CNN),reduces the impact of the deep structure on the fluctuation of hydrological time series.In order to fully study the model,the monthly rainfall data and daily rainfall data of Badong Station are selected to analyze the impact of data volume on the model,and the monthly runoff data of Shimen Street Hydrological Station is further introduced to verify the applicability of the new model to small sample data sets.The experimental results show that the new model can be applied to the prediction of small sample precipitation and runoff time series data sets.The overall performance is relatively stable and has high accuracy.It can be used as a prediction model of small sample hydrological time series.3)The above research puts forward a combined model suitable for small sample hydrological time series prediction.There are some problems in the analysis process.There is still room to improve the model performance and prediction accuracy:(1)The correlation between VMD decomposition mode and original data;(2)The initial weight of BLS is randomly generated,which has no relationship with the input data and is not interpretable.IVMD mode is further divided into reduced feature set and residual feature set,and an enhanced BLS(EBLS)model is proposed.Compared with IVMD-LSTM-BLS,new model removes the influence of VMD decomposition redundancy mode on model training.At the same time,EBLS updates the initial input weight and the final calculated weight value based on the idea of LSTM cell state,which enhances the correlation between the model and the input data.The prediction accuracy can reach up to 93%,and it has Class A reliability standard.Due to the removal of the influence of redundant modes and the advantages of BLS fast calculation,the efficiency has also been greatly improved.The model proposed in this paper has obvious advantages in prediction accuracy,model efficiency and computational efficiency.It is also applicable to different types of small sample hydrological time series data,which can make guidance for real hydrological situation and provide theoretical basis for flood prevention and water resources scheduling.