Investigating the Effect of Wavelet Decomposition on the Performance of the Optimized Support Vector Regression in Precipitation Simulation

Considering the climatic changes and the increase of extreme values in recent years, in this study, the effect of time series decomposition based on wavelet transform in improving the performance of the optimized support vector regression model in the simulation of annual precipitation in Dashband and Tapik stations has been discussed and investigated in the Lake Urmia Basin in the period of 1971-2020. In this study, the Ant colony algorithm was used to optimize the parameters of the support vector regression model. Daubechies 4 wavelet with three decomposition levels 1, 2 and 3 was used to decomposition the time series of precipitation in the studied stations. The SVR model takes in annual precipitation data as input, while the decomposition-based models take in decomposed precipitation values. The results of investigation the error rate and efficiency of the 4 investigated models include optimized SVR, W1-SVR (optimized SVR based on level 1 decomposition), W2-SVR (optimized SVR based on level 2 decomposition) and W3-SVR (optimized SVR based on level 3 analysis) showed that the error rate of all 4 mentioned models is acceptable and the observed values are in the 95% confidence interval. The error rate of 5.20 and 6.68 mm in the simulation of precipitation in Dashband and Tapik stations using the optimized SVR model by time series decomposition based on wavelet theory in level 1 decomposition in the mentioned stations, 31 and 35 percent improvement has been found. The level 2 decomposition of the time series of precipitation obtained the lowest error among the different levels of decomposition, which was 3.42 and 3.26 mm in Dashband and Tapik stations, respectively. Considering the increase in simulation complexity with the involvement of wavelet theory, the error rate improvement and model performance are acceptable. The hybrid W-SVR model in this study provides reliable results for precipitation simulation. Analyzing the annual precipitation series makes it possible to develop the dimensions of the optimized SVR model.