nonstationarities and at-site probabilistic forecasts of
TRANSCRIPT
ARTICLE
Nonstationarities and At-site Probabilistic Forecasts of SeasonalPrecipitation in the East River Basin, China
Peng Sun1,2,3 • Qiang Zhang2,3,4 • Xihui Gu5 • Peijun Shi2,3,4 • Vijay P. Singh6 •
Changqing Song2,3,4 • Xiuyu Zhang7
Published online: 8 March 2018
� The Author(s) 2018. This article is an open access publication
Abstract Seasonal precipitation changes under the influ-
ence of large-scale climate oscillations in the East River
basin were studied using daily precipitation data at 29 rain
stations during 1959–2010. Seasonal and global models
were developed and evaluated for probabilistic precipita-
tion forecasting. Generalized additive model for location,
scale, and shape was used for at-site precipitation fore-
casting. The results indicate that: (1) winter and spring
precipitation processes at most stations are nonstationary,
while summer and autumn precipitation processes at few of
the stations are stationary. In this sense, nonstationary
precipitation processes are dominant across the study
region; (2) the magnitude of precipitation is influenced
mainly by the Arctic Oscillation, the North Pacific Oscil-
lation, and the Pacific Decadal Oscillation (PDO). The El
Nino / Southern Oscillation (ENSO) also has a consider-
able effect on the variability of precipitation regimes across
the East River basin; (3) taking the seasonal precipitation
changes of the entire study period as a whole, the climate
oscillations influence precipitation magnitude, and this is
particularly clear for the PDO and the ENSO. The latter
also impacts the dispersion of precipitation changes; and
(4) the seasonal model is appropriate for modeling spring
precipitation, but the global model performs better for
summer, autumn, and winter precipitation.
Keywords China � ENSO regimes � GAMLSS
model � Nonstationarity � Probabilistic precipitation
forecasting
1 Introduction
An investigation of precipitation changes is one of the key
steps in understanding hydrological response to climate
change at the local, regional, and global scales, and is of
critical importance for the management of water resources
and agricultural development. This is especially important
in China (Xia et al. 1997; Zhang et al. 2012b)—the largest
agricultural country with the largest population in the
world. Precipitation forecasts play a crucial role in the
planning of water resource use, agricultural management
(Wei et al. 2005; Kisi and Cimen 2012; Ortiz-Garcıa et al.
2014), and flood forecasting (Francesco and Rebora 2015).
Multitudes of studies have addressed precipitation
forecasting using a variety of techniques. Ortiz-Garcıa
et al. (2014) predicted daily precipitation using the support
& Qiang Zhang
& Xiuyu Zhang
1 College of Territorial Resources and Tourism, Anhui Normal
University, Wuhu 241000, Anhui, China
2 Key Laboratory of Environmental Change and Natural
Disaster, Ministry of Education, Beijing Normal University,
Beijing 100875, China
3 State Key Laboratory of Earth Surface Processes and
Resource Ecology, Beijing Normal University,
Beijing 100875, China
4 Academy of Disaster Reduction and Emergency
Management, Beijing Normal University, Beijing 100875,
China
5 School of Environmental Studies, China University of
Geosciences, Wuhan 430074, Hubei, China
6 Department of Biological and Agricultural Engineering and
Zachry Department of Civil Engineering, Texas A&M
University, College Station, TX 77843, USA
7 South China Institute of Environmental Sciences,
Guangzhou 510655, Guangdong, China
123
Int J Disaster Risk Sci (2018) 9:100–115 www.ijdrs.com
https://doi.org/10.1007/s13753-018-0165-x www.springer.com/13753
vector machine (SVM) and pointed out its excellent per-
formance in comparison with other neural computation-
based approaches, such as multilayer perceptron, extreme
learning machine, and classical algorithms such as decision
trees and K-nearest neighbor classifier. Manzato (2007)
applied a neural computation approach to the short-term
forecasting of thunderstorm rainfall. Shukla et al. (2011)
indicated the importance of El Nino / Southern Oscillation
(ENSO) indices in precipitation forecasting. With the help
of a neural network, Shukla et al. (2011) forecasted pre-
cipitation in India due to the summer monsoon by con-
sidering ENSO indices. Nastos et al. (2013) applied an
artificial neural network to forecast precipitation in Athens,
Greece. Ingsrisawang et al. (2008) compared several
machine learning algorithms, such as decision trees (DT),
neural networks (ANN), and SVMs for short-term precip-
itation prediction in Thailand. Lu and Wang (2011) pre-
dicted monthly rainfall in China, using an SVM approach
with different kernel functions. These forecasts are deter-
ministic (point forecast representing our best guess) or
probabilistic (providing the forecast in terms of probability
of exceedance that reflects a range of possible values for
the variable of interest), and are statistical in nature or
based on numerical models.
In the above studies, deterministic forecast has been
widely used. However, statistical models have also been
widely used in precipitation forecasting (Villarini and
Serinaldi 2012; Wang et al. 2013). Advantages of statistical
models include the ability to identify changing patterns, the
estimation of parameters, uncertainty analysis, and diag-
nostic checking (Mishra and Desai 2005; Wang et al.
2013). Some statistical models, such as the autoregressive
moving average (ARMA) models, can predict the output
from the values of the output at previous time points. These
models are only suitable for a stationary system (Wang
et al. 2013), although it has been stated that seasonal
autoregressive integrated moving average (SARIMA)
models can model time series that exhibit nonstationary
behavior (Box et al. 2015).
The generalized additive model for location, scale, and
shape (GAMLSS) developed by Rigby and Stasinopoulos
(2005) has been used in this study. It can be devised to
cope with nonexponential distribution families and is able
to incorporate covariates not only in the position parameter
but also in the scale and shape parameters. Thus, GAMLSS
allows for a better control and representation of dispersion,
skewness, and kurtosis (Rigby and Stasinopoulos 2005;
Jones et al. 2013) and provides a suitable framework to
model the discrete–continuous distribution of daily pre-
cipitation accounting for seasonality and exogenous forc-
ing variables, such as alternative climate states (Francesco
and Kilsby 2014; Francesco and Rebora 2015). In this
study, the GMALSS model was used with exogenous
forcing variables such as the ENSO to simulate and fore-
cast precipitation in the East River basin, China.
The East River basin is a major tributary of the Pearl
River basin in South China with a drainage area of
35,340 km2, which accounts for about 5.96% of the Pearl
River basin area (Fig. 1). Water resources in the East River
basin have been highly developed for a variety of uses,
such as municipal and rural water supply, hydropower
generation, navigation, irrigation, and suppression of sea-
water invasion (Chen et al. 2010). In recent years the East
River has been supplying water to meet about 80% of Hong
Kong’s annual hydrological demand (Chen et al. 2010;
Wong et al. 2010). Therefore, the vulnerability of the East
River water resource system is of great importance for the
sustainable social and economic development of the Pearl
River Delta, one of the economically most developed
regions in China, and also for the sustainable water supply
of Hong Kong (Chen et al. 2013). In this context, precip-
itation changes play a critical role. Investigations of pre-
cipitation structure indicate higher frequencies of heavy
precipitation (defined as events with C 50 mm/day pre-
cipitation) in the eastern parts of the Pearl River basin and
the East River basin in recent decades (Zhang et al. 2012a).
Therefore, forecasting of precipitation variations in the
East River basin is important for the planning and man-
agement of water resources. The influence of large-scale
climate oscillations, for example ENSO, on precipitation
changes in the East River basin has been established and
corroborated (Zhang et al. 2013; Huang et al. 2017). Thus,
large-scale climate oscillations are used as exogenous
factors in precipitation forecasts.
The objectives of this study are to: (1) screen the models
that can forecast the precipitation behavior at a seasonal
temporal scale; and (2) forecast precipitation changes while
taking account of nonstationarity and uncertainty, includ-
ing those stemming from climate oscillations. Results of
this study will provide a framework for precipitation
forecasting in humid regions and guidance for planning and
management of water resources and agricultural activities
at the regional scale.
2 Data
Monthly precipitation data, covering the period of
1959–2010 and derived from 29 meteorological stations in
the East River basin, were analyzed in this study. The
locations of these stations are shown in Fig. 1. The dataset
was compiled by the Climatic Data Center of the National
Meteorological Information Center at the China Meteoro-
logical Administration.1 The quality of the data was
1 http://www.nmic.cn/web/channel-8.htm.
123
Int J Disaster Risk Sci 101
controlled before its release (Zhang et al. 2013). Seasonal
average precipitation changes were also analyzed, based on
four seasons: winter (December to February of the subse-
quent year), spring (March to May), summer (June to
August), and autumn (September to November).
The climatic indices used in this study for seasonal
average precipitation forecasts were the Arctic Oscillation
(AO) index, the North Pacific Oscillation (NPO) index, the
Pacific Decadal Oscillation (PDO) index, and the Southern
Oscillation Index (SOI). Many researchers have found that
precipitation changes in the East River basin are signifi-
cantly influenced by these four climate oscillations (Feng
and Li 2011; Zhang et al. 2013, 2014; Ding et al. 2014).
Data on the climatic indices were obtained from an Internet
source of the NOAA Earth System Research Laboratory.2
3 Methodology
Due to the influence of human activities and climate
change itself, precipitation changes are considered non-
stationary (Li et al. 2016; Chang et al. 2017). Therefore, a
nonstationary model should be used with parameters
changing along with the explanatory variables. In this
study, the GAMLSS model was selected for at-site pre-
cipitation forecasting.
3.1 The Generalized Additive Model for Location,
Scale, and Shape (GAMLSS)
The GAMLSS model (Rigby and Stasinopoulos 2005)
assumes a parametric distribution for the response variable
Y, the seasonal precipitation amount in this study, and
models the distribution parameters as functions of an
explanatory variable, time ti and the climatic indices—AOi,
NPOi, PDOi, and SOIi. It is assumed that in GAMLSS,
random variables yi, for i = 1,…, n, are mutually inde-
pendent and are from the same distribution function,
FY(yi|hi) with a vector of p distribution parameters,
hi = (h1, h2,…, hp), accounting for location, scale, and
shape parameters. Generally, p B 4 in that a 4-parameter
distribution family can describe distribution characteristics
of different hydrometeorological variables. Assume that
gk(�) shows monotonic relations between hk, the explana-
tory variable Xk, and the random term, that is:
gkðhkÞ ¼ gk ¼ Xkbk þXJk
j¼1
Zjkcjk ð1Þ
Fig. 1 Location of the study
region—the East River basin—
and the 29 precipitation stations
used in the study
2 http://www.esrl.noaa.gov/psd/data/climateindices/list/.
123
102 Sun et al. Probabilistic Forecasts of Seasonal Precipitation in the East River Basin
where gk and hk are the vectors with the length of n,
bkT = {b1k, b2k, …, bJk} is the parameter vector with the
length of Jk, Xk is the explanatory variable matrix with the
length of n 9 Jk, Zjk is the fixed design matrix with the
length of n 9 qjk, and cjk is the random variable following
the normal distribution. In this study, a cubic spline func-
tion was used as the link function between parameters and
covariate variables. Five two-parameter functions were
selected and are displayed in Table 1: Gumbel (GU),
Weibull (WEI), Gamma (GA), Lognormal (LOGNO), and
Logistic (LO).
The Akaike information criterion (AIC) technique was
used to select the distribution function with the highest
goodness-of-fit. To ensure the quality of the fit, a worm
plot was applied to estimate the fitting performance of the
distribution functions. More details can be found in Rigby
and Stasinopoulos (2005).
3.2 At-site Probabilistic Forecast
At-site (rain station) probabilistic forecasting was done
using the GAMLSS model, taking time, t, and the climatic
indices AO, NPO, PDO, and SOI as covariates. Two
forecasting modes were developed: (1) Taking the seasonal
precipitation series of each individual season—spring,
summer, autumn, and winter—as a sole series, this kind of
forecasting was denoted as the seasonal model; and (2)
Filtering out each seasonal precipitation component such as
spring, summer, autumn, and winter with a seasonal factor,
for example, 1, 2, 3, and 4 standing for spring, summer,
autumn, and winter, respectively, from each year and forms
the annual series of spring, summer, autumn, and winter
precipitation series. The forecasts were done on the reor-
ganized annual variations of seasonal precipitation com-
ponents. This kind of forecasting was denoted as the global
model. The seasonal and global models were then com-
pared to determine which model produces better forecasts.
Two precipitation forecasting practices were used: the
track precipitation forecast and the cross-validation pre-
cipitation forecast. In track precipitation forecasting,
parameter estimation is done using the seasonal precipita-
tion data covering the period of 1959–2000 and the esti-
mated parameter(s) is(are) used to forecast the precipitation
in 2001. The forecasted seasonal precipitation in 2001 is
added to the seasonal precipitation series of the period of
1959–2000 and parameters are estimated again, based on
the newly produced seasonal precipitation series. Then, the
seasonal precipitation is forecasted for 2002, using the
newly estimated parameters based on the newly produced
seasonal precipitation series of the period of 1959–2001.
This procedure is repeated until the forecasted seasonal
precipitation for 2010 is attained. In cross-validation pre-
cipitation forecasting, parameter estimation is done using
the seasonal precipitation data covering the calibration
Table 1 The five distribution functions considered in this study
Distribution
function
Probability density function Distribution moment Link functions
h1 h2
GumbelfY y h1; h2jð Þ ¼ 1
h2exp � y � h1
h2
� �� exp � y � h1ð Þ
h2
� �� �
�1\y\1;�1\h1\1; h2 [ 0
E Y½ � ¼ h1 þ ch2 ffi h1 þ 0:57722h2
Var½Y � ¼ p2h22=6 ffi 1:64493h22
Identity Log
WeibullfY y h1; h2jð Þ ¼ h2yh2�1
hh21exp � y
h1
� �h2( )
y[ 0; h1 [ 0; h2 [ 0
E Y½ � ¼ h1C1
h1þ 1
� �
Var½Y � ¼ h21 C2
h2þ 1
� �� C
1
h2þ 1
� �� �2( )
Log Log
Gamma
fY y h1; h2jð Þ ¼ 1
h22h1� 1=h22
y1
h22
�1
exp �y= h22h1� �
Cð1=h22Þ
y[ 0; h1 [ 0; h2 [ 0
E½Y � ¼ h1
Var½Y � ¼ h21h22
Log Log
LognormalfY y h1; h2jð Þ ¼ 1ffiffiffiffiffiffiffiffiffiffi
2ph22
q 1
yexp � logðyÞ � h1j j2
2h22
( )
y[ 0; h1 [ 0; h2 [ 0
E½Y � ¼ x1=2eh1
Var½Y � ¼ xðx� 1Þe2h1 ; where x ¼ exp h22�
Identity Identity
LogisticfY y l;rjð Þ ¼
exp � y�lr
�
r 1þ exp � y�lr
� �2
�1\y\1;�1\l\1;r[ 0
E½Y � ¼ r
Var½Y � ¼ r2p2
3
Identity Log
123
Int J Disaster Risk Sci 103
period, such as 1959–2000, and then the estimated
parameter(s) is(are) used to forecast the seasonal precipi-
tation for the verification period, such as 2001–2010.
In track precipitation forecasting, the naıve model
(simple linear regression) is introduced and a comparison
of the forecast performances of the seasonal and global
models and the naıve model (Mason and Baddour 2008;
Gabriele and Francesco 2012) is done. In cross-validation
precipitation forecasting, time, t, is taken as the covariate
variable for seasonal precipitation forecasting, called the
time model. Comparison is done for evaluating the fore-
casting performances of seasonal and global models as well
as the time model.
3.3 Evaluation of the At-site Probabilistic Forecast
Eight evaluation criteria were applied to evaluate the
forecasting accuracy of seasonal precipitation. Since each
method has its own advantages and limitations, it is
assumed that these methods can combine to evaluate the
forecasting accuracy of the models considered, thus
avoiding the uncertainty and unreliability due to limitations
of the individual method. These methods (Table 2) are:
absolute mean error (AME), mean absolute error (MAE),
median absolute error (MdAE), root mean squared error
(RMSE), mean absolute percentage error (MAPE), median
absolute percentage error (MdAPE), and root mean squared
percentage error (RMSPE). It should be noted that AME,
MAE, MdAE, and RMSE are absolute error indices and
MAPE, MdAPE, and RMSPE are comparative error indi-
ces. The geometric reliability index (GRI) was also used to
evaluate the performance of models from a geometric
perspective.
4 Results
Based on the aforementioned analysis procedure, we
obtained the following interesting and important results
and findings.
4.1 Nonstationarity of Seasonal Average
Precipitation Changes and Teleconnections
with ENSO, AO, PDO, and NPO Events
The AIC method was used to screen the models that sat-
isfactorily simulate seasonal and integrated seasonal pre-
cipitation and the related distribution functions (Fig. 2). It
can be seen from Fig. 2a that the distribution functions
with the highest goodness-of-fit for spring precipitation
series (seasonal model) are Gamma and Logistic models
for 19 out of the 29 stations. The location parameter, l, wasinfluenced mainly by PDO and AO (Fig. 3a) and had a
linear relation with AO but a nonlinear relation with PDO
(Fig. 3a). The scale parameter, d, had a close relation with
AO, PDO, and SOI (Fig. 4a). A significant influence of AO
on scale parameter, d, was observed for 16 out of the 29
stations (Fig. 4a). AO had a linear relation with the scale
parameter and a nonlinear relation was detected between d,and PDO and SOI. In general, spring precipitation in the
East River basin was influenced mainly by AO and PDO.
Meanwhile, PDO mainly impacted the location parameter,
l, implying that PDO altered the changing magnitude of
spring precipitation in the East River basin. However, AO
and ENSO mainly influenced the scale parameter, d, that isthe statistical dispersion properties of spring precipitation
changes (Figs. 3a and 4a). Besides, Figs. 3a and 4a show
that l and d for 4 and 3 out of the 29 stations were constant,
Table 2 Eight estimation methods for evaluation of modeling accuracy
Estimation INDICES Equations LB UB No errors
AME 1n
Py � yð Þ
0 Inf 0
MAE 1n
Py � yj j 0 Inf 0
MdAE Median ( y � yj j) 0 Inf 0
RMSEffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1n
Py � yð Þ2
q0 Inf 0
MAPE 1n
P y�yy
� 100 0 Inf 0
MdAPE Median y�yy
� 100
� �0 Inf 0
RMSPEffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1n
P y�yy
� �2r
� 1000 Inf 0
GRI1þQ1�Q
, where Q ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1n
P y�yyþy
� �2r
1 Inf 1
Low boundary (LB) and Upper boundary (UB) stand for the maximum and minimum values in theory, respectively; y stands for the measured
values, and y stands for the modeled values
123
104 Sun et al. Probabilistic Forecasts of Seasonal Precipitation in the East River Basin
indicating that spring precipitation changes at these stations
were stationary. In this sense, spring precipitation changes
at most of the stations were nonstationary.
Figure 2b indicates that Gamma distribution is the best
model for summer precipitation changes (seasonal model)
for 18 out of the 29 stations. In comparison, different
Fig. 2 Spatial distribution of
the best models in the East
River basin study area in
southeastern China based on
seasonal and global models with
the highest goodness-of-fit
123
Int J Disaster Risk Sci 105
changing properties of summer precipitation were detected
when compared to spring precipitation changes. The
influence of AO on scale parameter, d, was linear for 18 outof 22 stations (Fig. 4b). However, the location parameter,
l, was largely free from the influence of climate oscilla-
tions (Fig. 3b), implying that the mean value of summer
precipitation had stationary changes to a certain degree.
Figure 2c indicates that the Gamma distribution func-
tion was also the best model for autumn precipitation
changes (seasonal model) for 21 out of the 29 stations. A
closer look at Figs. 2c, b, 3c, and 4c indicates similar
changes in the properties of autumn precipitation changes
when compared to summer precipitation. However, com-
plicated changing characteristics were detected for winter
precipitation changes.
Figure 2d shows that the Weibull distribution function
was the best choice for modeling winter precipitation
changes (seasonal model) for 10 out of the 29 stations in
the East River basin. When compared to climatic indices
that are related to seasonal precipitation changes in summer
and autumn, more climate oscillations had considerable
impacts on winter precipitation changes, especially
Fig. 3 The climatic indices that
are closely related to the
location parameter in the
seasonal models in the East
River basin study area in
southeastern China. The
numbers in the parentheses
denote the number of stations
with nonlinear relations
between seasonal precipitation
and climatic indices; ct means
that the location parameter (l) isconstant
123
106 Sun et al. Probabilistic Forecasts of Seasonal Precipitation in the East River Basin
location and scale parameters. Specifically, AO, NPO,
PDO, and ENSO influenced the location parameter, l, ofwinter precipitation changes with different degrees and
magnitudes in both space and time (Fig. 3d). Statistically,
the impact of ENSO on the location parameter, l, wasnonlinear for 8 out of the 29 stations but the influence of
AO and NPO on l was respectively mostly linear and
linear. The impact of AO and PDO on the scale parameter
was linear for 13 and 17 out of the 29 stations, respectively.
ENSO had little impact on the scale parameter (Fig. 4d).
Based on the spatiotemporal properties of l and d, winter
precipitation changes were nonstationary. Therefore, a
nonstationary model is necessary to describe the properties
of seasonal precipitation changes in the East River basin.
In general, stationarity properties of seasonal precipita-
tion changes were different from one season to another and
climatic indices that were evidently related to seasonal
precipitation changes also differ. Specifically, summer and
autumn precipitation changes shared similar changing
properties, being influenced mainly by AO. Summer and
autumn precipitation changes were roughly stationary;
spring and winter precipitation changes were
Fig. 4 The climatic indices that
are closely related to the scale
parameter in the seasonal
models in the East River basin
study area in southeastern
China. The numbers in the
parentheses denote the number
of stations with nonlinear
relations between seasonal
precipitation and climatic
indices; ct means that the scale
parameter (d) is constant
123
Int J Disaster Risk Sci 107
nonstationary. Spring precipitation changes were influ-
enced mainly by PDO. Winter precipitation changes were
influenced by more climate oscillations. The impacts of
ENSO and PDO on winter precipitation changes were
pronounced. ENSO began to influence precipitation chan-
ges during autumn, and the influence of ENSO became
significant during the winter season.
In addition to the seasonal analysis, the spring, summer,
autumn, and winter season precipitation series were inte-
grated into one series and seasonal precipitation changes
were analyzed, based on the identification of seasonal
components, that is, the global model method (Fig. 5). It
can be seen from Fig. 2e that the Weibull distribution
function satisfactorily described the integrated seasonal
precipitation changes for 20 out of the 29 stations. PDO
and ENSO influenced the location parameter, implying that
these two climate oscillations impacted the changes in the
precipitation amount. The scale parameter, however, was
impacted mainly by ENSO.
4.2 Seasonal Precipitation Modeling for Spring,
Summer, Autumn, and Winter
For concise presentation, Xunwu, Dongyuan, Lingxia, and
Shenzhen stations were selected to illustrate the results of
the seasonal and global models because these four stations
are located along the main stream of the East River. Fig-
ure 6 shows the results of seasonal and global models for
spring precipitation changes. Due to the length limit of the
article, the results of the remaining 25 stations that were
also considered in this study are not presented here. The
seasonal model better captured the spring precipitation
changes than the global model, mirrored by the narrower
confidence interval circled by the 25 and 75% percentile
curves. The seasonal model captured higher and lower
amounts in the precipitation variations. This is particularly
important in the analysis of precipitation extremes. This
analysis was also corroborated by the worm plots in Fig. 7.
The results of the seasonal and global models for sum-
mer precipitation are shown in Fig. 8. The fluctuations of
summer precipitation were smaller than the spring precip-
itation changes. Generally, the global model had better
modeling efficacy than the seasonal model (Fig. 8). Sta-
tionary changes are shown by the seasonal model at the
Fig. 5 The climatic indices that are closely related to the location (l)and scale (d) parameters in the global model. The numbers in the
brackets denote the number of stations with nonlinear relations
between integrated seasonal precipitation and climatic indices; ct
means that the parameter is constant
123
108 Sun et al. Probabilistic Forecasts of Seasonal Precipitation in the East River Basin
Dongyuan and Shenzhen stations (Fig. 8b-1, d-1). How-
ever, the global model, as shown by the 50% percentile
curves, mirrored the average changes of summer precipi-
tation amounts. The confidence interval of the global
model defined by the 25 and 75% percentiles was narrower
than for the seasonal model. The 5 and 95% percentiles of
the global model better captured the changing features of
summer precipitation than did the seasonal model, partic-
ularly the peak and trough values of summer precipitation
variations (for example, Fig. 8d-2). The worm plots also
verified these results.
Figure 9 illustrates the results of the seasonal and global
models for autumn precipitation. Stationary changes of
autumn precipitation were revealed by the seasonal model
with no evident temporal changes of percentiles. The glo-
bal model satisfactorily depicted the changes in the mean
autumn precipitation, as reflected by the 50% percentile
curves (Fig. 9). The confidence intervals, defined by the 25
and 75% percentiles and also by the 5 and 95% percentiles,
well represented the changes in autumn precipitation,
especially precipitation magnitudes. The global model
better depicted the peak and trough values of the autumn
precipitation than did the seasonal model, as mirrored by
the 5 and 95% percentiles. The seasonal model showed
stationary changes in autumn precipitation, similar to
summer precipitation for some of the stations, and as a
result, the seasonal model did capture details of the
changing properties of autumn precipitation.
The results of the seasonal and global models for winter
precipitation changes are shown in Fig. 10. Winter pre-
cipitation had larger fluctuations than had spring, summer,
and autumn precipitation. A closer look at Fig. 10 indicates
the global model had higher modeling efficacy than the
seasonal model. The temporal variations of winter precip-
itation were simulated well as shown by the 50% per-
centile. The global model had narrower confidence
intervals than the seasonal model as defined by the 25 and
75% percentiles for simulation of the winter precipitation
changes. The global model also mirrored the nonstationary
changes in winter precipitation.
Figures 6, 7, 8, 9 and 10 show that both the seasonal and
global models can model temporal changes in seasonal
precipitation at 4 out of the 29 stations. Similar results can
be obtained for other stations considered in this study. A
similar efficacy in the simulation of the spring precipitation
changes was observed for the seasonal and global models.
However, the results of these two models are subject to
large discrepancies in the modeling of seasonal precipita-
tion changes in summer, autumn, and winter. Stationary
changes of summer, autumn, and winter precipitation were
detected by the seasonal model, though the precipitation
changes were nonstationary in the global model output. In
Fig. 6 Modeled spring
precipitation by the seasonal
model (left) and the global
model (right) at four selected
stations (Xunwu, Dongyuan,
Lingxia, and Shenzhen) in the
East River basin in southeastern
China. The black dashed line
denotes the 50% percentile, the
pink region denotes the interval
defined by 25 and 75%
percentiles, and the greenish-
blue region denotes the interval
defined by the 5 and 95%
percentiles. The grey points
denote the observed
precipitation data
123
Int J Disaster Risk Sci 109
summary, the seasonal model was the best choice for the
modeling of spring precipitation, and the global model was
the best choice for the modeling of seasonal precipitation
changes in summer, autumn, and winter. Thus, the global
model should be used for the modeling, simulation, and
forecasting of seasonal precipitation changes in the East
River basin, China. These results do not indicate, however,
which model is the best choice for forecasting regional
precipitation changes.
4.3 Integrated Precipitation Forecasting
by Seasonal and Global Models
The seasonal precipitation series in spring, summer,
autumn, and winter, respectively, were integrated into an
entire seasonal precipitation series, and then the seasonal
and global models were developed, based on the integrated
seasonal precipitation series, to model and forecast pre-
cipitation changes using the track precipitation forecasting
and cross-validation precipitation forecasting techniques.
In this study, the forecasting accuracy was evaluated by
eight methods (Table 2). In the track precipitation fore-
casting practice, the naıve model was used to evaluate the
forecasting performance of the seasonal and global models;
in the cross-validation precipitation forecasting practice,
the time model was used to compare and evaluate the
forecasting performance of the seasonal and global models.
To evaluate the overall performances of these models in
a region, Fig. 11 displays the performances of the seasonal,
global, and naıve models in the track precipitation fore-
casting practice for the forecasting of precipitation changes
during the period of 2006–2010. The seasonal model better
forecasted precipitation changes in spring, autumn, and
winter than did the global model, but did worse than the
global model in forecasting precipitation changes in sum-
mer. Based on AME and MAE, the seasonal and global
models had the smallest error in their modeling perfor-
mance. However, MAPE indicated differently, that is, the
seasonal and global models better forecasted precipitation
changes in spring and summer than in autumn and winter.
However, in general, the seasonal and global models
forecasted better than the naıve model (Fig. 11). Figure 12
shows the performance of the seasonal, global, and naıve
models in the track precipitation forecasting practice for
the forecasting of precipitation changes during the period
of 2001–2010. These results indicate that the longer the
forecasting time interval, the lower the forecasting
accuracy.
Figure 13 shows the performance of the seasonal, glo-
bal, and time models in the cross-validation precipitation
Fig. 7 The worm plot for the
simulation of spring
precipitation by the seasonal
model (left) and the global
model (right) at four selected
stations (Xunwu (a-1, a-2),Dongyuan (b-1, b-2), Lingxia(c-1, c-2), and Shenzhen (d-1,d-2)) in the East River basin in
southeastern China. The two
black dashed lines denote the
95% confidence intervals
123
110 Sun et al. Probabilistic Forecasts of Seasonal Precipitation in the East River Basin
Fig. 8 Modeled summer
precipitation by the seasonal
model (left) and the global
model (right) at four selected
stations (Xunwu, Dongyuan,
Lingxia, and Shenzhen) in the
East River basin in southeastern
China. The black dashed line
denotes the 50% percentile, the
pink region stands for the
interval defined by the 25 and
75% percentiles, and the
greenish-blue region displays
the interval defined by the 5 and
95% percentiles. The grey
points denote the observed
precipitation data
Fig. 9 Modeled autumn
precipitation by the seasonal
model (left) and the global
model (right) at four selected
stations (Xunwu, Dongyuan,
Lingxia, and Shenzhen) in the
East River basin in southeastern
China. The black dashed line
denotes the 50% percentile, the
pink region shows the interval
defined by the 25 and 75%
percentiles, and the greenish-
blue region indicates the
interval defined by the 5 and
95% percentiles. The grey
points denote the observed
precipitation data
123
Int J Disaster Risk Sci 111
forecasting practice for the forecasting of precipitation
changes during the period of 2006–2010. In general, the
seasonal model better forecasted precipitation in spring,
autumn, and winter than the global model; but the global
model did better in forecasting precipitation in summer.
Based on AME, MAE, MdAE, and RMSE, the forecasting
accuracy of the global and seasonal models for winter
precipitation was the lowest. MAPE seemed to show the
highest efficacy of the global and seasonal models for the
forecasting of precipitation in spring and summer. Higher
forecasting efficacy was detected for the seasonal model
than the time model, but no distinct differences in the
forecasting performance were observed between the global
and time models. Other evaluation indices, except MAPE
and RMSPE, indicated the larger forecasting accuracy of
the seasonal model than the time model. Figure 14 shows
the forecasting performance of the seasonal, global, and
time models in the cross-validation precipitation forecast-
ing practice for the forecasting of precipitation changes
during the period of 2001–2010. Similar conclusions were
Fig. 10 Modeled winter
precipitation by the seasonal
model (left) and the global
model (right) at four selected
stations (Xunwu, Dongyuan,
Lingxia, and Shenzhen) in the
East River basin in southeastern
China. The black dashed line
denotes 50% percentile, the
pink region represents the
interval defined by the 25 and
75% percentiles, and the cyanic
region occupies the interval
defined by the 5 and 95%
percentiles. The grey points
denote the observed
precipitation data
SGNSGNSGNSGN0
100
200
300
AME
SGNSGNSGNSGN0
200
400
MAE
SGNSGNSGNSGN0
200
400
MdA
E
SGNSGNSGNSGN0
200
400
600
RMSE
SGNSGNSGNSGN0
50
100
150
200
MAPE
SGNSGNSGNSGN0
50
100
150
MdA
PE
SGNSGNSGNSGN0
50
100
150
200
250
RMSP
E
SGNSGNSGNSGN1
2
3
GRI
S:Seasonal model G:Global model N:Naive model
(a) (b) (c) (d)
(e) (f) (g) (h)
Fig. 11 The eight modeling
accuracy indices for the track
predictions of seasonal
precipitation (from left to right:
spring, summer, autumn, and
winter) during 2006–2010,
using the seasonal, global, and
naıve models in the track
precipitation forecasting
practice for all 29 stations in the
East River basin of southeastern
China
123
112 Sun et al. Probabilistic Forecasts of Seasonal Precipitation in the East River Basin
obtained from Fig. 14. However, when compared to
Figs. 13, and 14 indicates that the longer the forecasting
time, the smaller the advantage of the time model over the
seasonal and global models in precipitation forecasting
practice—for example, MAPE and RMSPE—indicate that
the seasonal model was better than the time model in terms
of forecasting accuracy.
SGNSGNSGNSGN0
100
200
AME
SGNSGNSGNSGN0
200
400
MAE
SGNSGNSGNSGN0
200
400
600
MdA
E
SGNSGNSGNSGN0
200
400
600
RMSE
SGNSGNSGNSGN0
100
200
MAPE
SGNSGNSGNSGN0
100
200
MdA
PE
SGNSGNSGNSGN0
100
200
300
RMSP
E
SGNSGNSGNSGN
1.52
2.53
GRI
S:Seasonal model G:Global model N:Naive model
(a)
(e) (f) (g) (h)
(b) (c) (d)Fig. 12 The eight modeling
accuracy indices for the track
predictions of seasonal
precipitation (from left to right:
spring, summer, autumn and
winter) during 2001–2010,
using the seasonal, global, and
naıve models in the track
precipitation forecasting
practice using for all 29 stations
in the East River basin of
southeastern China
TSGTSGTSGTSG0
100
200
300
AME
TSGTSGTSGTSG0
100
200
300
400
MAE
TSGTSGTSGTSG0
200
400
MdA
E
TSGTSGTSGTSG0
200
400
600
RMSE
TSGTSGTSGTSG0
50
100
150
MAPE
TSGTSGTSGTSG0
50
100
MdA
PE
TSGTSGTSGTSG0
50
100
150
200RMSP
E
T SGTSGTSGTSG
1.5
2
2.5
GRI
T:Time model S:Seasonal model G:Global model
(a) (b) (c) (d)
(e) (f) (g) (h)
Fig. 13 The eight modeling
accuracy indices for the
seasonal precipitation (from left
to right: spring, summer,
autumn, and winter) during
2006–2010, using the time,
seasonal, and global models in
the cross-validation
precipitation forecasting
practice for all 29 stations in the
East River basin of southeastern
China
TSGTSGTSGTSG0
50100150200250
AME
TSGTSGTSGTSG
100
200
300
400
MAE
TSGTSGTSGTSG0
100
200
300
400
MdA
E
TSGTSGTSGTSG
100200300400500
RMSE
TSGTSGTSGTSG0
100
200
300
400
MAPE
TSGTSGTSGTSG0
50
100
150
200
250
MdA
PE
TSGTSGTSGTSG0
200
400
RMSP
E
T SGTSGTSGTSG
1.52
2.53
3.54
GRI
T:Time model S:Seasonal model G:Global model
(a) (b) (c) (d)
(e) (f) (g) (h)
Fig. 14 The eight modeling
accuracy indices for the
seasonal precipitation (from left
to right: spring, summer,
autumn, and winter) during
2001–2010, using the time,
seasonal, and global models in
the cross-validation
precipitation forecasting
practice using all 29 stations in
the East River basin of
southeastern China
123
Int J Disaster Risk Sci 113
5 Conclusions
Precipitation forecasting plays a critical role in the man-
agement of water resources and agricultural activities, and
that is particularly true for the East River basin in China
where water resources have been highly developed and
exploited. Hong Kong, for example, meets approximately
four-fifths of its current annual water demand by with-
drawals from the East River basin. In this study, modeling
of precipitation changes is done using GAMLSS-based
nonstationary techniques with consideration of the influ-
ence of large-scale climate oscillations on precipitation
changes in the East River basin. Important conclusions of
the study are as follows:
(1) Spring precipitation follows the gamma and logistic
distributions and exhibits nonstationary changes. The
changing magnitude of spring precipitation is influ-
enced mainly by AO and PDO; and the changing
dispersion is impacted by AO. The summer and
autumn precipitation changes follow the gamma
distribution. Summer and autumn precipitation at
most of the stations exhibits stationary changes. The
changing magnitude of summer and autumn precip-
itation is effected by NPO and ENSO, respectively. In
contrast, the precipitation variability in summer and
autumn is influenced mainly by AO. Winter precip-
itation follows the Weibull distribution and shows
nonstationary changes. The causes behind winter
precipitation changes are complicated when com-
pared to those in spring, summer, and autumn. More
or less, NPO, ENSO, AO, and PDO have evident
influences on the changing magnitude of winter
precipitation. But the influence of ENSO on the
changing magnitude of winter precipitation is signif-
icant when compared to NPO, AO, and PDO.
Integrating seasonal precipitation series into one
whole series may tell a different story. The whole
seasonal precipitation series, that is, the precipitation
series after integration of the precipitation series of
spring, summer, autumn, and winter, follows the
Weibull distribution. The location parameter, that is,
the changing magnitude, is influenced mainly by PDO
and ENSO. However, the precipitation variability is
influenced mainly by ENSO alone.
(2) Both seasonal and global models depict temporal
changes in seasonal precipitation in the East River
basin. The seasonal model is the best choice for
modeling spring precipitation, and the global model is
the best choice for modeling seasonal precipitation
changes in summer, autumn, and winter.
(3) In both the track precipitation forecasting and cross-
validation precipitation forecasting practices, the
seasonal model better forecasts precipitation changes
in spring, autumn, and winter; however, the global
model better forecasts summer precipitation changes.
In this sense, the global model performs better than
the seasonal model in the simulation of seasonal
precipitation changes, however, that is not the case in
precipitation forecasting. It is not a good idea to use
only one model, either the global model or the
seasonal model, in precipitation forecasting. In the
track precipitation forecasting and cross-validation
precipitation forecasting practices, the seasonal model
should be the choice in the forecasting of seasonal
precipitation changes. Results of this study help
clarify the selection of models in modeling, simula-
tion, and forecasting of seasonal precipitation varia-
tions, particularly considering the influence of ENSO
regimes. Because spatial dependence cannot be
neglected, a regional method that considers the
influence of spatial dependence may be of help in
improving the performance of the seasonal precipita-
tion forecast, and this will be addressed with consid-
erable consideration in our subsequent research.
Acknowledgements This work is financially supported by the Fund
for Creative Research Groups of the National Natural Science
Foundation of China (Grant No. 41621061), the National Science
Foundation for Distinguished Young Scholars of China (Grant No.
51425903), and the National Science Foundation of China (Grant
Nos. 41601023; 41401052). Our cordial gratitude should be extended
to the editor, Dr. Ying Li, and anonymous reviewers for their pro-
fessional and pertinent comments and suggestions, which were
greatly helpful for further quality improvement of this manuscript.
Open Access This article is distributed under the terms of the
Creative Commons Attribution 4.0 International License (http://crea
tivecommons.org/licenses/by/4.0/), which permits unrestricted use,
distribution, and reproduction in any medium, provided you give
appropriate credit to the original author(s) and the source, provide a
link to the Creative Commons license, and indicate if changes were
made.
References
Box, G.E.P., G.M. Jenkins, G.C. Reinsel, and G.M. Ljung. 2015. Time
series analysis: Forecasting and control. Hoboken, NJ: John
Wiley & Sons.
Chang, N.B., Y.J. Yang, I. Sanaz, and L. Mullona. 2017. Multi-scale
quantitative precipitation forecasting using nonlinear and non-
stationary teleconnection signals and artificial neural network
models. Journal of Hydrology 548: 305–321.
Chen, Y.D., T. Yang, C.Y. Xu, Q. Zhang, X. Chen, and Z.C. Hao.
2010. Hydrologic alteration along the Middle and Upper East
River (Dongjiang) basin, South China: A visually enhanced
mining on the results of RVA method. Stochastic Environmental
Research and Risk Assessment 24(1): 9–18.
Chen, Y.D., Q. Zhang, M.Z. Xiao, and V.P. Singh. 2013. Evaluation
of risk of hydrological droughts by the trivariate Plackett copula
123
114 Sun et al. Probabilistic Forecasts of Seasonal Precipitation in the East River Basin
in the East River basin (China). Natural Hazards 68(2):
529–547.
Ding, Y., Y. Liu, S. Liang, X. Ma, Y. Zhang, D. Si, P. Liang, Y. Song,
and J. Zhang. 2014. Interdecadal variability of the East Asian
winter monsoon and its possible links to global climate change.
Journal of Meteorological Research 28(5): 693–713.
Feng, J., and J. Li. 2011. Influence of El Nino Modoki on spring
rainfall over south China. Journal of Geophysical Research
116(D13): 1–10.
Francesco, S., and C.G. Kilsby. 2014. Simulating daily rainfall fields
over large areas for collective risk estimation. Journal of
Hydrology 512: 285–302.
Francesco, S., and N. Rebora. 2015. Impact of precipitation forecast
uncertainties and initial soil moisture conditions on a proba-
bilistic flood forecasting chain. Journal of Hydrology 519:
1052–1067.
Gabriele, V., and S. Francesco. 2012. Development of statistical
models for at-site probabilistic seasonal rainfall forecast. Inter-
national Journal of Climatology 32(4): 2197–2212.
Huang, C., Q. Zhang, V.P. Singh, X.H. Gu, and P.J. Shi. 2017. Spatio-
temporal variation of dryness/wetness across the Pearl River
basin, China, and relation to climate indices. International
Journal of Climatology 37(1): 318–332.
Ingsrisawang, L., S. Ingsriswang, S. Somchit, P. Aungsuratana, and
W. Khantiyanan. 2008. Machine learning techniques for short-
term rain forecasting system in the northeastern part of Thailand.
World Academy of Science, Engineering and Technology 41:
248–253.
Jones, M.R., H.J. Fowler, C.G. Kilsby, and S. Blenkinsop. 2013. An
assessment of changes in seasonal and annual extreme rainfall in
the UK between 1961 and 2009. International Journal of
Climatology 33(5): 1178–1194.
Kisi, O., and M. Cimen. 2012. Precipitation forecasting by using
wavelet-support vector machine conjunction model. Engineering
Applications of Artificial Intelligence 25(4): 783–792.
Li, J.Z., Y.X. Wang, S.F. Li, and R. Hu. 2016. A nonstationary
standardized precipitation index incorporating climate indices as
covariates. Journal of Geophysical Research Atmospheres
120(23): 12082–12095.
Lu, K., and L. Wang. 2011. A novel nonlinear combination model
based on support vector machine for rainfall prediction.
Proceedings of the IEEE 4th International Joint Conference on
Computational Sciences and Optimization: 1343–1347.
Manzato, A. 2007. Sounding-derived indices for neural network based
short-term thunderstorm and rainfall forecasts. Atmospheric
Research 83(2–4): 349–365.
Mason, S.J., and O. Baddour. 2008. Statistical modelling. In Seasonal
Climate: Forecasting and Managing Risk, ed. A. Troccoli, M.
Harrison, D.L.T. Andersen, S.J. Mason, 189–192. Dordrecht,
The Netherlands: Springer.
Mishra, A.K., and V.R. Desai. 2005. Spatial and temporal drought
analysis in the Kansabati River Basin, India. International
Journal of River Basin Management 3(1): 31–41.
Nastos, P.T., K.P. Moustris, I.K. Larissi, and A.G. Paliatsos. 2013.
Rain intensity forecast using artificial neural networks in Athens,
Greece. Atmospheric Research 119: 153–160.
Ortiz-Garcıa, E.G., S. Salcedo-Sanz, and C. Casanova-Mateo. 2014.
Accurate precipitation prediction with support vector classifiers:
A study including novel predictive variables and observational
data. Atmospheric Research 139: 128–136.
Rigby, R.A., and D.M. Stasinopoulos. 2005. Generalized additive
models for location, scale and shape. Journal of the Royal
Statistical Society 54(3): 507–554.
Shukla, R.P., K.C. Tripathi, A.C. Pandey, and M.I. Das. 2011.
Prediction of Indian summer monsoon rainfall using Nino
indices: a neural network approach. Atmospheric Research 102:
99–109.
Villarini, G., and F. Serinaldi. 2012. Development of statistical
models for at-site probabilistic seasonal rainfall forecast. Inter-
national Journal of Climatology 32(14): 2197–2212.
Wang, S.W., J. Feng, and G. Liu. 2013. Application of seasonal time
series model in the precipitation forecast. Mathematical and
Computer Modelling 58(3–4): 677–683.
Wei, H., J.L. Li, and T.G. Liang 2005. Study on the estimation of
precipitation resources for rainwater harvesting agriculture in
semi-arid land of China. Agricultural Water Management 71(1):
33–45.
Wong, J.S., Q. Zhang, and Y.D. Chen. 2010. Statistical modeling of
daily urban water consumption in Hong Kong: Trend, changing
patterns, and forecast. Water Resources Research 46(3): Article
W03506.
Xia, J., K.M. O’Connor, R.K. Kachroo, and G.C. Liang. 1997. A non-
linear perturbation model considering catchment wetness and its
application in river flow forecasting. Journal of Hydrology 200:
164–178.
Zhang, Q., J.F. Li, V.P. Singh, C.Y. Xu, and J. Deng. 2013. Influence
of ENSO on precipitation in the East River basin, South China.
Journal of Geophysical Research 118(5): 2207–2219.
Zhang, Q., V.P. Singh, J.T. Peng, and Y.D. Chen. 2012a. Spatial-
temporal changes of precipitation structure across the Pearl
River basin, China. Journal of Hydrology 440–441: 113–122.
Zhang, Q., P. Sun, V.P. Singh, and X. Chen. 2012b. Spatial-temporal
precipitation changes (1956–2000) and their implications for
agriculture in China. Global and Planetary Change 82–83:
86–95.
Zhang, Q., M.Z. Xiao, V.P. Singh, and Y.D. Chen. 2014. Max-
stable based evaluation of impacts of climate indices on extreme
precipitation processes across the Poyang Lake basin, China.
Global and Planetary Change 122: 271–281.
123
Int J Disaster Risk Sci 115