Machine Learning Methods In Hydrological Time Series Forecasting

Accurate water level prediction is of great importance for the safety of water infrastructure such as dams and embankments.However,accurate prediction of water levels is a challenging problem due to the nonlinear,non-stationary nature of water level information.With the current development of machine learning methods,neural networks are able to adapt to the nonlinear characteristics of data and thus have been widely used in the field of hydrological information prediction.The prediction accuracy of the traditional neural network single model can no longer meet today’s prediction accuracy requirements,and the combined model based on machine learning methods shows certain advantages.This thesis proposes two new combined neural network-based models to predict water level information in order to improve the accuracy and reliability of water level prediction in response to the above problems.The research work in this paper focuses on the following two aspects:(1)A combined model(VMD-GA-ELMAN-VMD-ARIMA)based on Variational mode decomposition(VMD),ELMAN neural network,Genetic Algorithm(GA)and Autoregressive Integrated Moving Average(ARIMA)model is proposed.The water level was first preprocessed data with VMD,the parameters of ELMAN were optimized with GA,and then each subsequence was predicted,while the error sequence was decomposed with VMD and predicted with ARIMA model,and finally the previously predicted water level was corrected.In the thesis,data from three different sets of water level stations in the Heihe River in Zhangye City was used to build 10 comparative models to compare the performance of the model.The results show that the combined model of VMD decomposition algorithm combined with GA-ELMAN and ARIMA models has the performance to improve the prediction results.In addition,the VMD double processing can largely improve the prediction accuracy.(2)A combined model(ICEEMDAN-VMD-WOA-ELM)is proposed based on improved adaptive noise complete ensemble empirical mode decomposition(ICEEMDAN),variational mode decomposition(VMD),extreme learning machine(ELM)and whale optimization algorithm(WOA).First,the hierarchical clustering method is used to extract the historical data with high similarity to the forecast day,and the extracted data is analyzed by box plot.Then ICEEMDAN is used to preprocess the original water level data of different Heihe stations,and the obtained high-frequency subsequence IMF1 is decomposed secondary by VMD.Afterwards,the ELM parameters are optimized during the prediction of each subsequence using the global optimization capability of WOA,and the predicted values are accumulated to generate the ultimate water level prediction results.The results show that:(a)The model uses hierarchical cluster method to find the periodic fluctuation of water level from historical data,and uses box plot to detect and correct it,which can greatly reduce the influence of outliers on the sequence while maintaining its mainstream trend.(b)The prediction error of secondary decomposition of ICEEMDAN-VMD is smaller than that of single decomposition of ICEEMDAN.(c)The addition of WOA can also improve the prediction accuracy of water level.The two new models validate the stability and accuracy of the combined models based on machine learning methods,which are significant for future water level research.