For the Screening Complexity in the Statistical Downscaling Model (SDSM) Nurul Nadrah Aqilah Tukimat1, Sobri Harun2 Dept
Total Page:16
File Type:pdf, Size:1020Kb
ISSN: 2319-5967 ISO 9001:2008 Certified International Journal of Engineering Science and Innovative Technology (IJESIT) Volume 2, Issue 6, November 2013 Multi-Correlation Matrix (M-CM) for the Screening Complexity in the Statistical Downscaling Model (SDSM) Nurul Nadrah Aqilah Tukimat1, Sobri Harun2 Dept. of Hydraulics and Hydrology, Universiti Teknologi Malaysia, Skudai, Malaysia Abstract— The statistical downscaling model (SDSM) has been applied for the projection of future climate pattern in Kedah, Malaysia. But it is quite difficult to make a correct decision on the potential correlation between multi-site predict ands and multi-predictors during the screening process based on the SDSM tool, because of its limited ability. In this regard, the M-CM analysis has been used to determine the correlation between 26 predictors and 20 predictand (rainfall station) in a single running. The concept of M-CM is sufficient to show the capability and reliability of the predictors based on the correlation value that can be explained in the dependent variable using the independent variable. The potential of predictor selection based on this method has been tested using MAE, MSE, and StD results. Results revealed the simulated value produced by these predictors set was closer to the observed value except at stn.IBT, KT, SL, SIK, SG and Kg.LS. It was consistent to the discrepancies (MAE and MSE) and StD results that showed bigger error compared to others rainfall stations. However, the error is still can be acceptable because produced less than 10% of discrepancies. Therefore, the future climate trend at this region was generated using constant predictors provided by HadCM3 under A2 scenarios. Index Terms— climate change; statistical downscaling; multiple correlation matrix; calibration process. I. INTRODUCTION Downscaling has become the imperative model to bridge the spatial and temporal resolution of General Circulation Models (GCMs) in the direction of the local-scale surface weather. The representation of the atmospheric parameters, ocean circulation, cryosphere, and land surface in the larger resolution (typically 50,000 km2) restricting the GCMs model ability to resolve important sub-grid scale features such as clouds and topography [2].The climate results after downscaled may produce different outcomes rather than using coarse-scale as the input. By using downscaling method, it assists the climate simulation downscales in the finer resolution focus on the specific sub-grid to interprets the local climate and not as general trend. Moreover, the gridded observed line could prevent the simulation process being damped by the ocean grid-box effect [3]. Two prominent techniques to derive local scale surface weather from regional scale atmosphere there are dynamical downscaling (DD) and statistical downscaling (SD). These techniques have their own pros and cons while evaluating the potential of climate change impacts arising from future increases in greenhouse gases concentration [10]. In the researched by [16] revealed each model performance were influenced by the seasonal effects that make DD and SD executed well in different season. Therefore, the combination between SD and DD are estimated to produce the best skill scores in some of the analyzed cases [5]. Meanwhile, [24] were classified these comparison techniques as a contentious due to the skill metrics and transparency of experimental design and producing inconsistent results at different cases. In this study, statistical downscaling model (SDSM) was applied in effort to generate the current/future local climate trend. The SDSM provides station-scale climate information from grid resolution GCM-scale with using multiple regression techniques between atmospheric circulation pattern (predictors) and local-scale parameters (predictands). The advantages of using SDSM tools in the projection of current/future climate are computationally undemanding, low cost and simple assess, which make this model the most popular model among the researchers. The potential of SD, also studied by [18], [11], [6], [13], [19] proved the capability and the reliability of the SD simulation results. Yet, the accuracy of the climate simulation in SDSM is depends on the predictors’ selections that have better association with the particular local surface climate. The predictors are typically derived from sea level pressure, geopotential height, wind fields, temperature variables, specific and relative humidity that represented in the GCMs parameters. The selection of predictors refers to the some criteria and behaviors. They should be reliably simulated by GCM; readily available from archives of GCM output, and strongly correlated with the surface 331 ISSN: 2319-5967 ISO 9001:2008 Certified International Journal of Engineering Science and Innovative Technology (IJESIT) Volume 2, Issue 6, November 2013 variable of interest [20].The different predictand may optimally choose the different set of predictors’ domain. Though, few climatologist appointed the downscaling local precipitation correlated well to the surface divergence and specific humidity as precipitation predictors meanwhile the local temperature has good relation with the geopotential height or air temperature as temperature predictors [21]. As mentioned by [7] in his article: “the changing climate should simulate with the predictors that carry the climate change signal”. Previous study applied various methods in effort to achieve the closest calibration among predictors and predictands such as Classical Least Square (CLS), Multiple Linear Regression (MLR), Non-linear Programming (NLP), singular value decomposition analysis, and Canonical Correlation Analysis (CCA) [1], [2], [25], [9], [4]. The CLS method is a traditional techniques practicing in the calibration process using regression technique. It evaluates the entire involved component with simultaneously or known as total calibration [8]. The CLS application uses simple equation and faster in calculation than other calibration method makes it more popular among the researchers. The selection of the wavelength is out of the consideration in the calibration as long as the number of wavelengths beyond the number of constituents. However, the conceptual of CLS that considers all the constituents in the calculation (total calibration) is impractical in the real world. The equation is also non-functional when the predictors are non-normal distribution [12]. Thus, MLR was introduced to overcome the limitation. It is an upgrading version of CLS. The method applies when involve multiple X (independent) and multiple Y (dependent) variables. It uses statistical techniques to find a mathematical relationship between selected factors (independent) simulated with the local condition (dependent) in the form of a linear. The measurement of MLR was influenced by intercepts and slopes/coefficient of predictors-predictand response around the mean values. The screening process in the SDSM model is applying MLR concept to express the predictor-predictand relationship as general. But, the fundamental problem in the SDSM analysis is that only probable to analyze multiple predictors with a single-predictand in a single simulation run. It becomes more complicated to choose the predictors that correlated well with the whole climate stations (predictand) involved in a study region. Like in this case study that have 20 of rainfall stations surrounding study area that have potential to forecast different climate trend at different locations. It will take longer time enough to screen, analyze, and select the predictors set represents for the whole rainfall stations. Besides, the rainfall pattern in Malaysia is non-uniform and very sensitive to the GCMs parameters that could influence the rainfall intensity. Due to this concern, the MLR idea is continue apply in this study but focus on multiple-predictand to multiple-predictors relationships in one time of running presented in multi correlation matrix (M-CM). The function is to measure the empirical relationship between 26 of predictors with 20 of predictands so that the correlation of each bond can show completely. II. METHODOLOGY A. Statistical Downscaling Model (SDSM) Statistical downscaling (SD) is analogous to the Model Output Statistic (MOS) and perfect prog approach used for short range numerical weather prediction [23]. The model was developed by Robert L. Wilby and Christian W. Dawson from United Kingdom and it uses weather generator method to produce multiple realization of synthetic daily weather sequence. This software calculates the statistical relationship based on multiple regression techniques between large scale (predictor) and local climate (predictand). These relationships developed using an observed weather data through the relationship of GCM-derived predictor. It produce a maximum and minimum temperature, precipitation and humidity of site specific daily scenarios for selected region and range of statistical parameter such as variance and frequencies of extremes. SDSM model allows different types of data to be transformed into standard predictor variables before being downscaled and calibrated to produce nonlinear regression models. Data series can also be shifted forward or backward by any number of time steps to produce lagged predictor variables. Regressions models can also be built on a monthly or annual basis. To generate the most ideal downscaled model, SDSM can reduce the standard error of estimate and increase the number of explained variance