Professional Documents
Culture Documents
ABSTRACT: In this study, the ability of the seasonal autoregressive integrated moving average (SARIMA) and autoregressive
integrated moving average (ARIMA) models was investigated for long-term runoff forecasting in the United States. In the first
stage, the amount of runoff is forecasted for 2011 in each US state using the data from 1901 to 2010 (mean of all stations in
each state). The results show that the accuracy of the SARIMA model is better than that of the ARIMA model. The relative
error of the SARIMA model for all states is <5%. In the second stage, the runoff is forecasted for 2001 to 2011 by using the
average annual runoff data from 1901 to 2000. The SARIMA model with periodic term equal to 20, R2 = 0.91, and mean bias
error (MBE) = 1.29 mm is the best model in this stage. According to the obtained results, a trend is observed between annual
runoff data in the United States every 20 years or almost a quarter century.
KEY WORDS
1.
Schar et al. (2004) forecasted a seasonal runoff using precipitation from meteorological data assimilation systems. The high
correlations suggest that a reliable seasonal runoff forecasting system can be constructed from the statistical relationship between the model-assimilated precipitation and subsequent
runoff. Artificial neural network is one of the other approaches
for runoff forecasting. In addition, seasonal runoff was simulated
using Hydrologic Engineering Center (HEC) models (Anderson
et al., 2002). Patry and Marino (1984c) presented a two-stage
urban runoff forecast model. This approach was found to be
particularly useful for lead times of up to 30 min. Wang et al.
(2011) developed a modified rational equation for arid-region
runoff estimation successfully. Habib et al. (2007) investigated
the effect of local systematic and random errors of the commonly
used tipping-bucket (TB) rain gauges on the accuracy of runoff
predictions. The computed runoff differences caused by the TB
random errors were dependent on the magnitude of the runoff
discharge, and on the temporal resolution of the rainfall input. In
recent decades, radar information helps researchers in accurate
forecasting of runoff data (Rahimi et al., 2014). Patry and Marino
(1984a, 1984b) investigated the sensitivity and application of different equation models to real-time urban runoff forecasting.
2012a), and rough theory, autoregressive integrated moving average (ARIMA) and seasonal autoregressive integrated moving
average (SARIMA) models have been applied based on stochastic theory; considering the serial correlation among observations, and by providing for systematic searching in each stage
(identification, estimation and diagnostic check) an appropriate
model was always applied to ensure that the hydrological forecast
exhibited a strong random component (Zhang et al., 2011).
Valipour et al. (2013) compared autoregressive moving average (ARMA), ARIMA and the autoregressive artificial neural
network models in forecasting the monthly inflow of the Dez
reservoir. The inflow to the reservoir shows that the ARIMA
models results match the forecast compared with the ARMA
model. Valipour (2012b), using time series models, determined
the required observation data for rainfall forecasting according
to the climate conditions. By comparing the R2 of the models,
it was determined that time series models are more appropriate
for rainfall forecasting in semi-arid climates. In addition, considering the importance of an accurate estimation of evapotranspiration (Psilovikos and Elhag 2013, Valipour 2014a; Valipour
2015; Valipour and Eslamian, 2014), ARIMA can be a powerful model for forecasting evapotranspiration in hydrometeorology and irrigation water requirement (Mahdizadeh Khasraghi
et al., 2014; Valipour, 2014b, 2014c; Valipour et al., 2014).
Valipour (2013a, 2013b) forecasted snow water equivalent using
the ARIMA model with a relation error of <10%. Valipour
(2012a, 2012c) determined the critical areas in Iran for agricultural water management according to the annual rainfall by using
ARIMA model. Using the collected data, a rainfall forecast for
the next 1 year was made by the ARIMA model.
3.
* Correspondence: M. Valipour, Young Researchers and Elite Club, Kermanshah Branch, Islamic Azad University, Kermanshah, Iran. E-mail:
vali-pour@hotmail.com
M. Valipour
4.
(RO RF )2 (RO RA )2
(5)
R2 = 1
In this study, the SARIMA and ARIMA models were used for
runoff forecasting as follows:
Seasonal autoregressive integrated moving average
= SARIMA (p, d, q) (P, D, Q)s
(1)
(4)
MBE =
(RO RF )n
(6)
5.
Figure 1. The map designed based on gathered annual hydrological data for temperature, relative humidity, rainfall and runoff. Note that ranges
in the figure legend include values greater than or equal to the lower bound and less than the upper bound indicated. See http://en.wikipedia.org/
wiki/List_of_U.S._state_abbreviations for definition US state abbreviations.
M. Valipour
Table 1. The best runoff forecasting models for each US state in 2011.
State
Best model
Alabama
SARIMA11
(5,1,1)(0,1,1)
SARIMA55
(4,0,5)(0,0,1)
SARIMA22
(5,0,1)(0,1,1)
SARIMA5
(5,1,3)(0,1,1)
SARIMA22
(5,3,1)(0,1,1)
SARIMA55
(5,0,1)(0,1,1)
SARIMA11
(4,1,2)(0,1,3)
SARIMA5
(2,0,3)(3,0,1)
SARIMA2
(1,1,4)(3,0,1)
SARIMA22
(4,0,4)(1,1,0)
SARIMA11
(5,1,3)(0,1,2)
SARIMA11
(1,1,2)(2,0,1)
SARIMA22
(5,2,0)(4,0,1)
SARIMA11
(0,1,1)(2,1,1)
SARIMA22
(3,1,0)(0,1,2)
SARIMA11
(3,1,1)(3,0,1)
SARIMA22
(2,0,0)(1,2,0)
SARIMA11
(0,0,2)(2,0,1)
SARIMA22
(2,3,1)(1,1,1)
SARIMA22
(4,1,2)(2,1,0)
SARIMA2
(0,0,4)(3,0,1)
SARIMA22
(0,3,2)(0,0,1)
SARIMA22
(1,0,2)(2,0,1)
SARIMA22
(0,2,2)(2,1,1)
SARIMA22
(2,0,2)(3,1,2)
SARIMA5
(0,1,1)(3,1,1)
Alaska
Arizona
Arkansas
California
Colorado
Connecticut
Delaware
District of Columbia
Florida
Georgia
Hawaii
Idaho
Illinois
Indiana
Iowa
Kansas
Kentucky
Louisiana
Maine
Maryland
Massachusetts
Michigan
Minnesota
Mississippi
Missouri
Observed
(mm)
Forecasted
(mm)
RE
(%)
368.52
370.39
0.51
Montana
599.50
586.79
2.12
Nebraska
8.55
8.65
1.17
Nevada
377.02
372.10
1.31
New Hampshire
338.60
349.11
3.10
New Jersey
92.95
90.83
2.28
New Mexico
1034.30
1033.60
0.07
New York
453.26
453.12
0.03
North Carolina
422.22
424.79
0.61
North Dakota
145.60
151.48
4.04
Ohio
186.84
180.26
3.52
Oklahoma
1550.06
1545.11
0.32
Oregon
359.74
360.17
0.12
Pennsylvania
412.08
413.07
0.24
Puerto Rico
495.40
504.48
1.83
Rhode Island
305.45
306.63
0.39
South Carolina
36.85
35.03
4.94
South Dakota
631.48
620.61
1.72
Tennessee
111.18
116.34
4.64
Texas
957.13
925.61
3.29
Utah
455.51
455.12
0.09
Vermont
850.86
812.16
4.55
Virginia
333.96
333.75
0.06
Washington
287.77
300.14
4.30
West Virginia
299.20
284.35
4.96
Wisconsin
331.70
326.17
1.67
Wyoming
State
Best model
SARIMA22
(2,4,1)(1,1,1)
SARIMA5
(0,0,4)(3,0,1)
SARIMA22
(0,3,1)(1,1,1)
SARIMA22
(0,1,1)(2,1,1)
SARIMA22
(0,2,2)(0,1,1)
SARIMA11
(0,0,4)(1,1,2)
SARIMA22
(3,1,2)(2,1,2)
SARIMA11
(0,1,1)(5,1,0)
SARIMA22
(4,4,0)(1,0,2)
SARIMA22
(2,2,1)(0,1,1)
SARIMA55
(1,0,0)(1,0,0)
SARIMA22
(2,2,0)(1,0,0)
SARIMA22
(2,2,2)(2,1,1)
SARIMA11
(4,0,2)(0,0,1)
SARIMA5
(0,0,3)(2,0,3)
SARIMA22
(3,0,2)(2,0,1)
SARIMA22
(0,3,1)(2,0,0)
SARIMA11
(1,1,3)(3,0,1)
SARIMA55
(5,0,0)(1,0,0)
SARIMA22
(2,3,2)(2,1,0)
SARIMA22
(2,2,2)(2,1,1)
SARIMA2
(0,0,4)(4,0,1)
SARIMA11
(0,0,5)(4,1,0)
SARIMA22
(0,0,4)(3,0,1)
SARIMA11
(2,1,0)(3,0,1)
SARIMA22
(2,4,1)(1,1,1)
Observed
(mm)
Forecasted
(mm)
RE
(%)
214.82
216.52
0.79
56.60
56.65
0.09
68.90
71.50
3.78
863.97
865.15
0.14
880.27
843.26
4.20
11.06
11.28
2.01
952.39
914.77
3.95
271.89
276.04
1.53
110.14
105.85
3.89
593.05
613.69
3.48
66.23
67.36
1.71
599.54
576.13
3.90
826.04
864.23
4.62
998.96
1045.94
4.70
642.83
643.03
0.03
171.35
177.90
3.82
91.66
94.61
3.22
581.12
583.22
0.36
12.30
11.79
4.11
105.21
109.68
4.25
1036.29
1024.72
1.12
334.93
345.44
3.14
1103.22
1134.69
2.85
596.16
595.10
0.18
356.72
350.94
1.62
141.23
140.90
0.24
processes (Figure 1) and runoff rate between 200 and 600 mm,
all runoff events were forecasted by choosing s = 2. However,
SARIMA22 also obtained acceptable values for these states
(Figure 2). For many states more than one model could be used;
however, only in five states relative error values were <10% in
all models (Tennessee, Virginia, West Virginia, Delaware and
Maryland). In Nebraska, Arkansas, Rhode Island and Delaware,
although SARIMA55 obtained a better forecasting, SARIMA22
was also able to forecast runoff with RE < 10%. However, in
Meteorol. Appl. (2015)
Figure 2. Performance of runoff forecasting models applied in this study based on forecasting accuracy.
minimum amount of runoff in the entire statistical period. However, other models had a RE >30% for this state. From the other
peak values forecasted by the SARIMA models it can be seen
that runoffs in Connecticut, Minnesota, Montana, New York and
Vermont were the maximum in the last 111 years. In addition, the
accuracy of the models is not acceptable in New Jersey, North
Dakota and Ohio during the last century.
5.2.
The time series related to the annual runoff of each state show
that due to climatic changes the runoff could not be forecasted for
more than 1 year with a good accuracy for each state. However,
in some situations that need a comprehensive study, for example,
for flood control and use of runoff caused by rainfall, it is required
to forecast with a time horizon longer than 1 year.
While the average annual runoff of all the states was considered
as the runoff of the United States, because of neutralization
of drastic climatic changes and reduction in the peak points,
Meteorol. Appl. (2015)
M. Valipour
Figure 3. Performance of ARIMA and SARIMA models for runoff forecasting in each US state.
Table 2. The best structures of each model for runoff forecasting
(20012011) in all US states.
Model
ARIMA
SARIMA2
SARIMA4
SARIMA5
SARIMA10
SARIMA20
SARIMA25
SARIMA50
Structure
R2
MBE (mm)
(5,1,3)
(4,1,4)(1,0,1)
(3,1,4)(1,0,1)
(3,1,3)(2,1,1)
(3,1,3)(2,0,2)
(0,1,4)(1,0,1)
(1,1,4)(1,0,1)
(1,1,3)(1,0,1)
0.86
0.86
0.86
0.87
0.87
0.91
0.90
0.88
4.06
1.72
2.90
8.09
1.86
1.29
4.39
2.49
6.
Recommendation
Although the results show that the obtained accuracy is currently acceptable (the RE of the SARIMA models for all the
states is <5% and R2 = 0.91 for the runoff estimation in the
United States), a hybrid forecasting that combines nonlinear
and machine-learning approaches can improve or enhance
Figure 4. Performance of the SARIMA20 model in long-term runoff forecasting for the United States. (a) Observed and forecasted runoff versus
time and (b) forecasted runoff versus observed runoff.
2015 Royal Meteorological Society
Conclusions
References
Anderson M, Chen Z, Kavvas M, Feldman A. 2002. Coupling
HEC-HMS with atmospheric models for prediction of watershed
runoff. J. Hydrol. Eng. 7(4): 312318.
Banihabib ME, Valipour M, Behbahani SMR. 2012. Comparison of
autoregressive static and artificial dynamic neural network for the
forecasting of monthly inflow of Dez reservoir. J. Environ. Sci. Technol. 13: 114. http://jest.srbiau.ac.ir/?_action=articleInfo&article=84
(accessed 15 February 2012).
Dong SH, Zhou HC, Xu HJ. 2004. A forecast model of hydrologic single
element medium and long-period based on rough set theory. Water
Resour. Manage. 18(5): 483495.
Habib E, Aduvala AV, Meselhe EA. 2007. Effect of radar-rainfall
errors on rainfall-runoff modeling. World Environmental and Water
Resources Congress, 1519 May 2007, Tampa, FL, DOI: 10.1061/
40927(243)285.
Mahdizadeh Khasraghi M, Gholami Sefidkouhi MA, Valipour M.
2014. Simulation of open- and closed-end border irrigation
systems using SIRMOD. Arch. Agron. Soil Sci. DOI: 10.1080/
03650340.2014.981163 (in press).
Patry G, Marino M. 1984a. Two-stage urban runoff forecast model.
J. Water Resour. Plann. Manage. 110(4): 479496.