A novel discussion on two long-term forecast mechanisms for hydro-meteorological signals using hybrid wavelet-NN model

According to the different selection principles of model inputs in the testing period of a time series forecast, two kinds of long-term forecast mechanisms, the “seeming” and “true” long-term (SLT and TLT) forecasts, for different hydro-meteorological time series signals predicting and their forecast performance variations with corresponding driving mechanisms are proposed and discussed for the first time with this study. Daily precipitation and evaporation data of one station and river stage data of two stations are used as case studies, and six kinds of popular hybrid and pure models are used to compare both kinds of forecast performances. Results show that because of the forecast mechanism variations conventional SLT forecast models have abnomally overall high and similar performances. For meteorological signals, especially for precipitation signal, the signal features with larger numbers of zero value data and weak short-term periodicities, revealed by the Continuous Wavelet Transform (CWT) method, lead to the overall poor performances of different TLT forecast models, but make the Discrete Wavelet Transform (DWT) method significantly effective on SLT forecasts. With respect to hydrological river stage signals, the signal features with significant short-term periodicities and without interference of zero value data can finely reveal the significant advantage of DWT-NF hybrid model, combining DWT and Neuro-Fuzzy, on TLT forecasts, but weaken the advantages of DWT method and neural network models on SLT forecasts. Since the TLT forecast has higher practical value but lower performance than the conventional SLT forecast, the DWT-NF hybrid model has been demonstrated as a better predictor than other hybrid and pure models for effectively improving the hydro-meteorological signal TLT forecast performance.