Javad ramezani moghadam; Mostafa Yaghoubzadeh; Ahmad Jafarzadeh
Abstract
Introduction & Background: Assessment of climate change impacts on hydrology is relied on the information of climate changes in adequate scale. Due to outputs of GCMs (General Circulation Models) that are the most confident tools for simulating climate change impacts but are available in coarse resolution. ...
Read More
Introduction & Background: Assessment of climate change impacts on hydrology is relied on the information of climate changes in adequate scale. Due to outputs of GCMs (General Circulation Models) that are the most confident tools for simulating climate change impacts but are available in coarse resolution. Downscaling process which is classified to several methods such as transfer function, weather generator and weather typing is performed for improving of GCMs projection and using them in local scale. Meanwhile feature selection is the main essential step in downscaling with transfer function. Because the main goal of downscaling is the improvement of GCMs projections, several researches examined vary approaches for feature selection. This study aims to assess performance of downscaling daily precipitation under four different selection methods such as PCA, CA, SRA and ParCA using comprehensive comparison tests.
Materials and Methods: Measured daily rainfall for Ardebil (with cold semi-arid climate) and Birjand (arid climates) were collected for the period from 1977 to 2004. The CanESM2 (Canadian Earth System Model) outputs were used as GCM for simulating of climate change impacts on precipitation pattern. So of CanESM2 outputs (large scale predictors) and measured daily precipitation (local scale predictants) were considered as input and target for downscaling respectively. The Artificial Neural Network (ANN) which widely has been used in climate change researches was selected as downscaling method. Despite of the most of literature have used only efficiency criteria for distinguishing from different approaches in downscaling, this study reveals performance of feature selection methods based on either them or statistical tests. The comparison tests between measured and downscaled rainfall such as assessment criteria, statistics characteristics comparison, contingency table event for wet and dry series diagnostics and Violin plot were used as tools for skill assessment of feature selection approaches.
Results and Discussion: Results showed that although different methods of predictor selection had includes various subsets, predictors such as relative humidity at surface and zonal velocity component at 500-hPa pressure levels in Birjand and mean temperature at 2m, mean sea level pressure and rotation of the air in Ardebil are the most descriptive features which have more relationship with measured daily precipitation. The efficiency criteria of comparing measured and downscaled precipitation indicated that CA method is superior to other in Birjand station and SRA’s results were better than those of other in Ardebil station. Value of RMSE, R and NSE was achieved 1.2 mm/day, 0.55 and 0.25 in Birjand and 1.75 mm/day, 0.14 and 0.013 in Ardebil respectively. The examination of measured and downscaled statistical characteristics reveals that CA has the better influence on downscaling than those of others in Birjand station. In this comparative test most of downscaled statistical components such as mean, median and skewness under CA have more similarity to measured values. But in Ardebil, with cold and arid climate, performance of SRA to downscale was the same as performance of CA to it. Also both SRA and CA were better than ParCA. The skill assessment of different methods to fit measured and downscaled variability by violin plot showed that generally ParCA outperformed other method in Birjand station. The comparison of violin plots, in Ardebil, revealed that no one of predictor selection methods has acceptable accuracy for fitting measured variability. Outcomes of contingency table event showed although all feature selection methods have not remarkable capability for distinguishing from the measured wet and dry series in Ardebil station, performance of ParCA and SRA were acceptable in Birjand station. The values of CSI for ParCA and SRA were calculated 0.25 and 0.22 in Birjand and it shows that more of 20 percent of ParCA and SRA’s diagnostics was correct.
Conclusions: By assessing of results, it can be inferred that generally downscaling of daily rainfall in Birjand station is outperforming Ardebil. In other expression daily downscaling of precipitation in arid climate has better results than cold and arid climate. Also different tests have various results about feature selection methods. In Ardebil, SRA in efficiency criteria test and both SRA and CA in statistics characteristics have better performance than others. But in this region no methods have remarkable performance in violin and dry and wet tests.
Abazar Solgi; Amir Pourhaghi; Heidar Zarei; Hadi Ansari
Abstract
Introduction: Chemical pollution of surface water is one of the serious issues that threaten the quality of water. This would be more important when the surface waters used for human drinking supply. One of the key parameters used to measure water pollution is BOD. Because many variables affect the water ...
Read More
Introduction: Chemical pollution of surface water is one of the serious issues that threaten the quality of water. This would be more important when the surface waters used for human drinking supply. One of the key parameters used to measure water pollution is BOD. Because many variables affect the water quality parameters and a complex nonlinear relationship between them is established conventional methods can not solve the problem of quality management of water resources. For years, the Artificial Intelligence methods were used for prediction of nonlinear time series and a good performance of them has been reported. Recently, the wavelet transform that is a signal processing method, has shown good performance in hydrological modeling and is widely used. Extensive research has been globally provided in use of Artificial Neural Network and Adaptive Neural Fuzzy Inference System models to forecast the BOD. But support vector machine has not yet been extensively studied. For this purpose, in this study the ability of support vector machine to predict the monthly BOD parameter based on the available data, temperature, river flow, DO and BOD was evaluated.
Materials and Methods: SVM was introduced in 1992 by Vapnik that was a Russian mathematician. This method has been built based on the statistical learning theory. In recent years the use of SVM, is highly taken into consideration. SVM was used in applications such as handwriting recognition, face recognition and has good results. Linear SVM is simplest type of SVM, consists of a hyperplane that dataset of positive and negative is separated with maximum distance. The suitable separator has maximum distance from every one of two dataset. So about this machine that its output groups label (here -1 to +1), the aim is to obtain the maximum distance between categories. This is interpreted to have a maximum margin. Wavelet transform is one of methods in the mathematical science that its main idea was given from Fourier transform that was introduced in the nineteenth-century. Overall, concept of wavelet transform for current theory was presented by Morlet and a team under the supervision of Alex Grossman at the Research Center for Theoretical Physics Marcel in France. After the parameters decomposition using wavelet analysis and using principal component analysis (PCA), the main components were determined. These components are then used as input to the support vector machine model to obtain a hybrid model of Wavelet-SVM (WSVM). For this study, a series of monthly of BOD in Karun River in Molasani station and auxiliary variables dissolved oxygen (DO), temperature and monthly river flow in a 13 years period (2002-2014) were used.
Results and Discussion: To run the SVM model, seven different combinations were evaluated. Combination 6 which was contained of 4 parameters including BOD, dissolved oxygen (DO), temperature and monthly river flow with a time lag have best performance. The best structure had RMSE equal to 0.0338 and the coefficient of determination equal to 0.84. For achieving the results of the WSVM, the wavelet transform and input parameters were decomposed to sub-signal, then this sub-signals were studied with Principal component analysis (PCA) method and important components were entered as inputs to SVM model to obtain the hybrid model WSVM. After numerous run this program in certain modes and compare them with each other, the results was obtained. One of the key points about the choice of the mother wavelet is the time series. So, the patterns of the mother wavelet functions that can better adapt to diagram curved of time series can do the mappings operation and therefore will have better results. In this study, according to different wavelet tests and according to the above note, four types of mother wavelet functions Haar, Db2, Db7 and Sym3 were selected.
Conclusions: Compare the results of the monthly modeling indicate that the use of wavelet transforms can increase the performance about 5%. Different structures and sensitivity analysis showed that the most important parameter which used in this study was parameter BOD, and then flow, DO and temperature were important. This means that the most effective BOD and temperature with minimum impact. Also between different kernels types, RBF kernel showed the best performance. So, combined wavelet with support vector machine is a new idea to predict BOD value in the Karun River.
