Chapter 4
The early forecasting of seasonal climate focused mainly on mean temperatures or total rainfall (Ahago 1992, Briggs and Wilks 1996). Another application of seasonal prediction was the forecast of the frequency of cyclones, such as the north Atlantic hurricanes (Gray et al. 1994, Lehmiller et al. 1997, Mielke and Berry 2000) and the tropical cyclones (taifuns) in the northwest Pacific (Chan et al. 1998), the southwest Indian Ocean (Jury 1998) and the southwest Pacific (Basher and Zheng 1995). The prediction of the monsoon season was later carried out by means of a seasonal forecasting scheme (Thapliyal 1981, Navone and Ceccatto 1994, Goswami and Srividya 1996, Sahai et al. 2003, Rajeevan et al. 2007, Pattanaik and Kumar 2010). Nonetheless, the prediction of inter-annual climate variations has rarely been endeavored (White 2000).
Meanwhile, most of the few longer-lead predictions done up to date are based merely on the extrapolation of trends and cycles (Dyer and Tyson 1977, Currie 1993).
Another group of prediction tool encompasses dynamical forecasting models which use comprehensive coupled GCMs (Rosati et al. 1997, Stockdale et al. 1998b, Segschneider et al.
2000), including atmospheric- and oceanic-state variables to predict the climate and the ocean state - abbreviated here as SST which, though it stands for sea surface temperatures, comprises also other ocean-atmospheric indices, such as the pressure oscillation index (ENSO), mentioned earlier (Stockdale et al. 1998a)-. However, because of their intrinsic non-linearity, these models are very complex and require huge computational resources.
The applicability of GCM- output for predicting seasonal climate was alternatively studied (Palmer and Anderson 1994, Kumar and Hoerling 1995, Stern and Miyakoda 1995, Kumar et al.
1996). Although some studies of obtaining GCM- predictors in predicting short-term climate demonstrated that large-scale atmospheric flow pattern are not capable to simulate the local-scale variability in the short-term (Goddard et al. 2001), statistical forecasts which include GCM-outputs appear to be an alternative way to improve the skill of short-term prediction (Krishnamurti et al. 2006, Chakraborty and Krishnamurti 2009, Krishnamurti et al. 2009, Kar et al. 2012). To enhance the prediction performance of a single model further, the combination of results from different models was suggested (Thompson 1957, Clemen and Murphy 1986, Fraedrich and Leslie 1987, Kharin and Zwiers 2002, Palmer et al. 2004, Palmer et al. 2005, Semenov and Stratonovitch 2010). Especially, for seasonal climate prediction, these combination schemes were applied successfully (Fraedrich and Leslie 1987, Livezey and Barnston 1988, Casey 1995, Doblas-Reyes et al. 2000, Palmer et al. 2000, Pavan and Doblas-Reyes 2000, Palmer et al.
2004, Yun et al. 2005, Prasad et al. 2010). Therefore, following this approach also in the present study, the predictors from multi-domain GCMs will subsequently be used for seasonal climate prediction.
El Niño conditions over the tropical Pacific, in conjunction with the coupled changes in ENSO, are known to affect climate conditions by teleconnection over hemispherical distances (Ropelewski and Halpert, 1987; 1989,;1996) and, more so, over the closer circum-pacific region (Allan et al. 1996, Orlove et al. 2000, Guyot et al. 2009). Predictands persisting over longer-time in the atmospheric and land boundary layers can have a higher teleconnective predictability (Huang et al. 1996, Rosati et al. 1997, Zeng 1999, Goddard et al. 2001). For example, with the forerunner ability of the ENSO- atmospheric variability, the prediction of the monsoon-rainfall’s inter-decadal variability is possible (e.g. Parthasarathy et al. 1991, Hastenrath et al. 1995, Hastenrath 1995, Kumar et al. 1999, Sahai et al. 2000). However, even though local climate conditions are influenced by the large-scale atmospheric circulation, using initial atmospheric predictors alone is still insufficient for predicting seasonal variations (Branković et al. 1990, Barnett 1995, Branković and Palmer 1997).
Therefore, another crucial and convenient method for predicting short-time climate variations is statistical forecasting which derives the future climate state from previously observed climate time series (Goddard et al. 2001). These statistical forecasting models predict the future climate by mathematical procedures, which produce relationships between oceanic predictors, i.e. air and sea surface temperature, and regional climate predictands, i.e. temperature, precipitation and solar radiation. Among this group of statistical prediction tools, the regression model is the main representative. In this method ocean-SST boundary, forcing predictors are used for seasonal climate prediction, i.e. the determination of some climate predictands (Palmer and Anderson 1994, Goddard et al. 2001).
According to the ‘two-tiered’ climate prediction approach (Hunt 1997, Bengtsson et al. 1993), the atmospheric climate is dominated by prescribed boundary conditions, such as SSTs as well as the characteristics of the land surface (Goddard et al. 2001). The SSTs are a predictor source that are widely used for forecasting seasonal climate in many areas across the world (Bah 1987, Rao and Goswami 1988, Folland et al. 1991, Barnston et al. 1994, Shabbar and Barnston 1996, Uvo et al. 1998, Colman and Davey 1999, Landman and Mason 1999, Thiaw et al. 1999), as well as in the Asia Pacific (Yu et al. 1997, Aldrian and Dwi Susanto 2003, Sahai et al. 2003, Xu et al. 2007).
For Thailand, a study of the usefulness of SSTs in climate prediction was carried out by Singhrattna et al. (2005a), using the regression method with ENSO-predictors to predict summer rainfall at one- to three-month lead times. Also, a k-nearest neighbor model associated with atmospheric variables, such as surface air temperature and sea level pressure derived from 2.5x2.5 global grid of NCEP/NOAA reanalysis was later used by the same author (Singhrattna et al. 2012) to forecast 7-9 month-ahead rainfall.
Since the SSTs are playing an important role for seasonal climate prediction worldwide and for Thailand, in particular, in this study, teleconnections of the ocean climate with Thai regional weather will also be examined as potential predictors, for use in short-term climate prediction in the study region. For the development of a short-term forecasting tool, various statistical time series methods, i.e. linear regression and autoregressive models will be employed, wherefore both teleconnection- and multi-GCM predictors are investigated for their applicability in short-term climate prediction in the area.
4.1.2 Applications of autoregressive (AR) models in climate prediction
Another famous technique used in recent years to predict local climate, is the autoregressive (AR) model, which is a data-driven technique for time-series modeling. This technique was applied in numerous climatic time-series analyses, namely, precipitation (e.g. El-Fandy et al. 1994, Chu et al. 1995, Mentz et al. 2000, Rebora et al. 2006, Kwon et al. 2007, Barbulescu and Pelican 2009, Sigrist et al. 2011) and temperature (e.g. Gu and Jiang 2005, Kwon et al. 2007, Malvestuto et al. 2011). Extensions of the basic AR-model are the autoregressive moving average (ARMA) and autoregressive integrated moving average (ARIMA) models which, later in this section, are also applied for the analysis of seasonal and inter-annual time-series (Goddard et al. 2001).
The ARMA-model was developed to handle a weak stationary stochastic process (Whittle 1983).
It has been widely applied in hydrologic time series analysis (e.g. Chang et al. 1984, Wu et al.
2009). The ARIMA-model is a further extension of ARMA, to deal with non-stationary trends in the time-series, wherefore, by appropriate differencing (corresponding to reverse integrating, therefore the name) of the original series, the latter is made stationary and so that it can subsequently be analyzed by regular ARMA-modeling (Box and Jenkins 1976, Box et al. 2008).
The ARIMA-model has been widely applied for the analysis and prediction of climatic and hydrologic time-series in many parts of the world, especially, of precipitation (e.g. Adamowski et al. 1987, Weesakul and Lowanichcha 2005, Liu and Shao 2006, Somvanshi et al. 2006, Yurekli and Kurunc 2006, Hurile et al. 2008, Barbulescu and Pelican 2009, Chattopadhyay and Chattopadhyay 2010, Kim et al. 2011), temperature (e.g. Kim et al. 2011) and streamflow (e.g.
Carlson et al. 1970, Salas et al. 1985, Haltiner 1988, Yu and Tseng 1996, Kothyari and Singh 1999, Huang et al. 2004, Castellanomendez et al. 2004, Modarres 2007, Fernández et al. 2009). In Thailand, Weesakul and Lowanichcha (2005) employed ARMA and ARIMA for annual rainfall forecasting at 21 stations across the entire country for yearly agricultural water allocation.
Tantanee (2006) also used an AR-model to predict annual rainfall over the northeastern part of Thailand for 1-year annual rainfall distribution. A review of various applications of autoregressive techniques in forecasting climate is provided in Table 4.1.
Table 4.1. Uses of autoregressive models for the prediction of temperature and precipitation.
predictand model* data used reference
temperature AR(3) monthly temperature in mainland China (Tantanee 2006) AR(1) daily minimum and maximum
temperatures in Lazio, Italy (Malvestuto et al. 2011) precipitation AR(1) annual rainfall in North Africa (Peel et al. 2004)
ARIMA(1,1,1) annual rainfall in the KY- basin (Weesakul and Lowanichcha 2005)
ARIMA(2,1,2) ARIMA(2,1,1) ARIMA(1,1,2)
annual rainfall in Thailand (Weesakul and Lowanichcha 2005)
AR(4) annual rainfall in north Thailand (Tantanee 2006) ARIMA(3,1,2) annual rainfall in China (Liu and Shao 2006) ARIMA(0,1,1) monsoon rainfall over India (Chattopadhyay and Chattopadhyay 2010) ARIMA (2,0,3) monthly rainfall in Mongolia (Kim et al. 2011)
*for notation see section 4.2
Another interesting extension of the classical autoregressive model is acquired by adding an external regressor variable (Hyndman and Khandakar 2008). Such ARIMA-models with external regressors, called from now on here "ARIMAex" (Nau 2012) have been applied in several studies of climate time-series analysis (e.g. Luz et al. 2008, Chun et al. 2009, Fernández et al. 2009, Chun et al. 2013). As external regressors, the teleconnections of the ocean state indices have been used in some of these ARIMAex- model applications (e.g. Luz et al. 2008, Chun et al. 2013). Although atmospheric predictors from global climate models (GCMs) form another predictor-set which is often used in long-term climate prediction, they have also been applied in shorter-time-scale climate predictions (Semenov and Stratonovitch 2010) for the simulation of seasonal and inter-annual variability (Palmer et al. 2005, Doblas-Reyes et al. 2006), but also for inter-season forecasting (Palmer et al. 2004, Palmer et al. 2008, Acharya et al. 2011).
The above studies indicate that predictor variables obtained from teleconnections as well as from GCMs can serve as potential external regressors in an ARIMAex- model. This relatively novel forecasting approach will be applied later in this chapter for the short-term prediction of rainfall and temperature in the study area.
4.1.3 Application of teleconnections and GCMs for seasonal climate prediction
Since both the atmospheric climate variables and ocean indices (SSTs) are physical parameters that are somehow transported through the global atmosphere-ocean system, it is of no surprise
that time shifts for a climate event between different regions can exist (Palmer and Anderson 1994). As a matter of fact, SSTs are useful for the prediction of a variety of climate variables and for the establishment of empirical, statistical models for regional climate and large-scale climate variability (Goddard et al. 2001). Although a lack of understanding of the proper predictor–
predictand relationships might probably indicate the missing of some explicit information on the relevant physical processes, such a statistical model may still be helpful for seasonal climate prediction.
Various statistical climate forecast studies have been carried out in the last decades that used several ocean SST variables to predict the seasonal climate in adjacent land areas (Webster et al.
1998, Goddard et al. 2001). In the early studies of this kind only a single ENSO- index was used to predict seasonal climate variations (Hutchinson 1992) and typical features for the latter have then been associated with average responses to ENSO-events (Ropelewski and Halpert 1996).
However, later on a combination of various ocean state indices, such as SSTs, was introduced to increase the skill of short-term climate prediction (DelSole and Shukla 2002, Sadhuram and Ramana Murthy 2008), and doing so, the teleconnective nature of such ENSO/SST- regional climate relationships has then increasingly been recognized (Goddard et al. 2001).
The usefulness of a teleconnection application in climate forecasting is to be validated by various statistical methods (Barnston and van den Dool 1993, Zhang and Casey 2000). One is the cross-correlation analysis of ocean indices and local climate variables, as presented in Chapter 2. The teleconnections established there indicate that these have the ability of long-lead forecasting of regional climate variations in Thailand, including the EST study region. In fact, from the cross-correlation maps examined in Section 2.6, the various ocean indices were successively screened as potential climate predictors for the study region. One of the most important relationships found there is the high correlation of 0.4 for a lag of -3 months between El-Nino 1.2 SST and the local precipitation (see Figure 2.30). Thus, the time-series of this Pacific ocean index which has some persistence can be used to predict the local climate for one or more seasons ahead (Graham et al. 2000, Goddard et al. 2001).
However, it should also be noted that the teleconnective influence of the ocean SST on the regional climate, i.e. the performance of a seasonal climate forecast, is reduced for non-ENSO years (Barnston et al. 1999, Landman and Mason 1999). On the other hand, in addition to the use of simple SSTs, some improvements in the prediction skills were claimed when atmospheric predictors, e.g. zonal wind and pressure, are included in the prediction models (Hastenrath et al.
1995, Makarau and Jury 1997, Francis and Renwick 1998, Jury et al. 1999, Philippon and Fontaine 1999). The addition of such atmospheric predictors also appears to improve the prediction power of the forecasting model at seasonal scales (Goddard et al. 2001). In fact, by including these atmospheric predictors in statistical models, the early prediction of the monsoon rainfall has been possible (Shukla and Paolino 1983, Shukla and Mooley 1987), even in the presence of weak ENSO-events (Yang et al. 1998). These atmospheric predictor-signals were also applied to determine the variability of large-scale meteorological/climate pattern in Asia (e.g.
Hastenrath 1987, Vernekar et al. 1995, Webster 1995). In any case, such teleconnective relationships emphatically indicate that there is some predictability of the climate, when using the atmospheric state as a boundary forcing process.
An alternative source of potential predictor variables is derived from general circulation models (GCMs) which simulate the atmospheric climate state by the numerical solution of the equations governing the various physical processes in the ocean-atmosphere system under different greenhouse gas emission scenarios (for details see Chapter 3). The GCMs provide information
on various climate variables, such as temperature or precipitation, either directly or in terms of some other atmospheric predictors.
Climate predictors from GCMs have been widely used to predict seasonal climate variations (e.g.
Stern and Miyakoda 1995, Kumar et al. 1996, Graham et al. 2000, Acharya et al. 2011) where, moreover, the use of multi-model GCM-ensemble output has been recommended to increase the prediction skill (Graham et al. 2000). Nevertheless, GCMs have some limitations, as far as the short-term prediction of the small-scale climate variability from large-scale atmospheric flow pattern is concerned (Palmer and Anderson 1994, Goddard et al. 2001, Alexander et al. 2002). It is for these situations where the use of teleconnections in seasonal prediction can partly fill this deficiency gap of the GCM-predictor data and, so, improve the climate prediction at both the regional and seasonal scale.