WATER RESOURCES BULLETIN
VOL. 24, NO. 5 AMERICAN WATER RESOURCES ASSOCIATION
SHORT-TERM FORECASTING OF SNOWMELT RUNOFF USING ARMAX MODELS1
J. P. Haltiner and J. D. Saks2
ABSTRACT: Time series models of the ARMAX class were investi-
gated for use in forecasting daily riverfiow resulting from combined
snowmelt/rainfall. The Snowmelt Runoff Model (Martinec-Rango
Model) is shown to have a form similar to the ARMAX modeL The
advantage of the ARMAX approach is that analytical model identifica-
tion and parameter estimation techniques are available. In addition,
previous forecast errors can be included to improve forecasts and
confidence limits can be estimated for the forecasts. Diagnostic checks
are available to determine if the model is performing properly.
Finally, Kalinan filtering can be used to allow the model parameters
to vary continuously to reflect changing basin runoff conditions. The
above advantages result in improved flow forecasts with fewer model
parameters.
(KEY TERMS: streaniflow forecasting; snowmelt runoff; time series
modeling; Kalman filtering; real-time forecasting.)
INTRODUCTION
Short-term streamfiow forecasting (several hours to several
days in advance) is important in flood warning, reservoir
operation, recreation, and water quality applications. Rainfall-
runoff forecasting at this time scale has been quite inten-
sively studied, while combined snowmelt/rainfall runoff model-
ing has received less attention.
The use of deterministic physical process models for
short-term forecasting has a number of drawbacks including
large computer storage requirements, difficulty in including
real-time measured data in the forecast, and overparameteri-
zation (more model parameters than can be justified given the
available data). A simpler conceptual model, the "Snowmelt
Runoff Model" (SRM) developed by Martinec and Rango
(Martinec, 1960; Rango and Martinec, 1979) has been shown
to give good results on a number of basins of varying size
(Rango, 1983; Martinec and Rango, 1986). Reliable modeling
using the SRM requires accurate parameter estimation, proper
determination of the model form (number of previous runoff
and snowmelt/precipitation terms to include), and measure-
ment of snow-covered areas (SCA). Analytical methods to
address the first two requirements have not been previously
presented. In addition, measurement of SCA are expensive
and difficult to obtain on a regular basis. Finally, use of the
'Paper No. 87063 of the Water Resources Bulletin. Discussions are open until June 1, 1989.
2Respectively, Principal, Philip Williams and Associates, Pier 35, The Embarcadero, San Francisco, California 94133; and Professor, Colorado State
University, Dept. of Civil Engineering, Ft. Coffins, Colorado 80521.
1083
OCTOBER 1988
model in a forecasting mode could be improved if real-time
information could be used to update the model parameters
on a regular basis.
In this paper, a class of stochastic, time-series models re-
ferred to as ARMAX (Auto Regressive-Moving Average with
Exogenous Inputs) or transfer function models (Box and
Jenkins, 1970) are used. It is shown that the Snowmelt
Runoff Model can be viewed as a particular case within this
general class of linear stochastic models. This offers a num-
ber of advantages: (1) analytical techniques are available for
adapting the model form to basins of varying characteristics
and to varying time frames (daily, hourly, etc.); (2) a num-
ber of efficient parameter estimation techniques are available;
(3) confidence limits describing the accuracy of the forecast
can be included with the forecast; (4) diagnostic checks are
available to determine if the model is performing properly;
(5) previous forecast errors can be included in the model to
improve future forecasts; and (6) the model can be cast in
"systems format" and the Kalman filter can be used to up-
date the parameters or status of the model in real time.
MODEL FORMULATION
The general univariate ARMAX model for river flow
(adapted from Box and Jenkins, 1970) can be expressed as:
p q r
= + (1)
i=1 j=0 LI 'tj + k=0 0k 6t—k
where:
= streamfiow;
Q
I = "moisture input" (combined snowmelt/rainfall);
e = white noise process;
p,q,r = number of previous river flow, input and noise
terms included in the model;
WATER RESOURCES BULLETIN
 Haltiner and Salas

= model parameters to be estimated; and point was at the Del Norte, Colorado, gaging station. Drain-
age area above this point is 3419 km2 (1320 mi2), with
t time step. elevations ranging from 2432 m (7980 ft) to 4215 m (13,830
ft). Riverflow behavior is dominated by snowpack accumula-
Next, consider the Snowmelt Runoff Model equation in a tion and melt, resulting in a low-flow (snowpack accumula-
commonly used form (Shafer, et a!., 1982): tion) period between September and April and a high-flow
(snowpack melting) period, which begins in early April and
persists through the summer. This basin was selected because
(T_Tb) S + P } A' (1—Kt)
=
Qt_iKt + C extensive modeling with the SRM has already been conducted
on the basin (Shafer, et a!., 1982), allowing use of identical
(2) data for model comparison and evaluation.
where: Daily precipitation, streamfiow, average daily temperature,
and snow-covered area data for the period 1973 to 1980 were
= streamfiow, used in the forecasting evaluation. Only the peak flow
period (April to September) in each year was examined.
C= runoff coefficient, Precipitation, temperature, and snow-covered area data were
used to calculate the moisture input, I, for each of three
= degree day factor,
at elevation zones using the SRM approach shown in Equation
= air temperature, (4). These zonal moisture inputs were then combined to
Tt represent a single basin moisture input, I, for use in Equa-
Tb = base temperature, tion (1). Although the SRM approach to snowmelt was used
in this study to facilitate model comparison, any available
t = time step (typically, one day), snowmelt/precipitation model could be used to calculate the
S= snow-covered area, input term in Equation (1).
= precipitation,
MODEL ESTIMATION
A= watershed area,
= recession coefficient, and
Model Format
The first step in estimation is to determine the model
= elevation zone.
order (the values of p, q, and r) in Equation (1). This can
It can be seen that the forecast of flow at the next time step be done empirically, analytically, or using a combination of
is a function of the flow at the current time step and the ef- both. Empirical identification is based on examining the
fective snowmelt/rainfall during the next time step. If we physical processes involved. For example, Shafer, et aL
denote this combined snowmelt/rainfall input from all eleva- (1982) examined hourly flow records for the Rio Grande at
tion zones as I, let C1.Ø equal the product of C and (1—Kt) Del Norte and estimated that about 35 percent of snowmelt/
and call the recession coefficient çb1, Equation (2) becomes: rainfall for a particular day appears as runoff on that same

=
i Qt—i + ''o 1t (3)
day and 65 percent on the following day. Using this informa-
tion, they determined that the SRM model should include
two input terms. The analytical approach uses the cross-
correlation structure between the input and output (flow)
where:
series to identify model order. In this approach, the input
= 3 ii i ii
i=l (at(Tt_Tb) St + Pt) A (4)
series is "prewhitened" (Box and Jenkins, 1970) by fitting
a standard time series model to the data, resulting in a white
noise process. This same model is then applied to the output
series, and the cross-correlation at various time lags is calcu-
Equation (3) can be seen to represent a special case of lated. This cross-correlation is scaled using the standard
Equation (1), without the inclusion of a noise term. The deviation of the transformed input and output processes to
advantages of using the ARMAX formulation will be presented
produce an impulse response function:
in the following sections.

= k=0, 1, 2.
/L1I Q? (k)
APPLICATION
The ARMAX model described by Equation (1) was used
to forecast daily streamfiows in the headwaters of the Rio
Grande River in southern Colorado. The riverflow forecast

1084 WATER RESOURCES BULLETIN

 Short-Term Forecasting of Snowmelt Runoff Using ARMAX Models

where: considered suboptimal. Nonetheless, the NLS approach was

found to provide consistent, stable parameter estimates.
v = impulse response function, Several schemes to account for the variation of basin
= standard hydrologic conditions with the ARMAX model were investi-
and deviation of the transformed in-
gated. In the seasonal version, parameters were estimated
put and flow series, on a monthly and biweekly basis. In a second approach, the
= cross-correlation between the transformed level of flow in the river was used as an index of antecedent
input and flow, and moisture conditions. Six thresholds of flow were chosen
= time lag. and a set of parameters was calculated for each level. The
k
Ø parameter is inversely proportional to flow while the 1o
and o parameters (accounting for the current and previous
The impulse response function for daily input/flow data
for the Rio Grande is shown in Figure 1. When compared day's input) are generally proportional to flow (Figure 2).
Thus, increasing flows indicate greater antecedent soil moisture
with known impulse response curves (Box and Jenkins, 1970), and correspondingly, an increased amount of snowmelt/rain-
it indicates that the model should include one past flow term
fall appearing as streamflow. One interpretation of the sea-
and two input terms, the same as that identified in Shafer, sonal behavior of (..,ij and w1 (Figure 3) is that early in snow-
et a!. (1982). The analytical approach is especially useful in melt season, the proximity of the snowpack to the river re-
complex cases, such as hourly modeling or large basin model- sults in a majority of a current day's snowmelt appearing as
ing, where there may be several time periods of pure delay runoff on that same day. As the season progresses, the snow-
and a large number of previous flow and input terms to be pack recedes and most of the snowmelt produced on a given
included.
day does not appear as runoff until the following day (repre-
sented by the fact that w > c).
0.30 Two additional model parameterizations were evaluated.
In the first, the snowmelt and rainfall terms are included in
0.25 the model separately. The model for the Rio Grande River
then becomes:
0.20
C
0.15

0.10
=
i —i + + + ''o't + 'lptl
C

0.05 + k0 0kt—k (6)

0.00 .1. ._._.— ..—.--.—.- I .1.

0 2 3 4 5 6 7 where:
Log (doys)

Figure 1. Impulse-Response Function for the Mt = . (T—Tb)(S) ---

Snowmelt/Precipitation Input (impulse) and i-i A
Streamfiow (response) on the Rio Grande River.
AT = total area, and

The number of noise terms to be included in the model i

'A
3 Al
will be determined by analyzing the residual series, Ct, after P= P -fl-—
preliminary estimation of the and c model parameters.

Parameter Estimation In Equation (6), Mt represents an areal average value of

The model parameters (0, and 0) were estimated using
a nonlinear least squares (NLS) algorithm. In the NLS ap-
proach, Equation (1) is solved for e, which represents the
difference between the observed and forecast flows. Param-
eters are selected using a Newton-Raphson iterative tech-
nique, which minimizes Under the conditions of sta-
tionarity and normality of the data, nonlinear least squares
is an approximate maximum likelthood estimator. Since
these conditions are not strictly met, the estimator must be
the product of degree days and snow-covered area and re-
presents an areal average value of precipitation. The degree
day factor is now included in the tj and parameters. In
addition to eliminating the need for estimating the degree
day factor, the model form of Equation (6) allows separate
consideration of the effects of snowmelt and rainfall on river
flow. Perhaps the primary difficulty in using the SRM ap-
proach is obtaining values of the snow-covered area (see Equa-
tion 2). Because aerial photography is expensive for use on a
regular basis, satellite imagery has received primary emphasis.
Problems associated with satellite data include cloud cover,

1085 WATER RESOURCES BULLETIN

 Haltiner and Salas

1.0 .10

0.8 .08

C
a
U .06
a 0.6
U C
a

53
1
53

0.4 .04

0.2 .02

0.0.

0. 1000. 2000. 3000. 4000.

______
5000. 6000. 7000.
00
8000.

Flow (cfs)

Figure 2. ARMAX Model Parameter Estimates Based on Flow Thresholds.

41
wo 0.I
-.-.-wl

0.8 0.08

Q
0.6 0.06

F-.

0.4 0.04

0.2 0.02

0.0 0.0

April May June July Aug. Sept.

Figure 3. Biweekly Estimates of the ARMAX Model Parameters.

1086 WATER RESOURCES BULLETIN

 Short-Term Forecasting of Snowmelt Runoff Using ARMAX Models

estimation between time of satellite passes, and difficulty
in obtaining imagery quickly for use in real-time forecasting.
Alternatives to sateffite imagery include snow-cover depletion
curves based on accumulated degree days, or possibly rela-
ting areal snow cover to daily snow water equivalent (SWE)
as measured at SNOTEL sites. An approach investigated
here was to eliminate the snow-covered area term from the
model. The melt term in Equation (6) then becomes:
Mt =
3 i i
-a--—
(7)
i-i (It—Tb)
AT
The parameters çj and c in Equation (6) must also now
account for the areal extent of snow available for melting, the
"effectiveness" of degree days in causing melt and the amount
of melt to appear as runoff. The parameters w0 and ''
could not be seasonally estimated because of the dramatic
differences year to year in the SCA. It was determined that
season, the parameters are approximately constant, as degree
day effectiveness continues to increase, but snow-covered
area is decreasing. This latter factor becomes dominant as
the 'season progresses, and the parameters decrease. It can
also be seen that in the early season, crj (current day's melt)
is dominant. As the season progresses, (previous day's
melt) becomes more important. The application of Kalman
filtering has some inherent problems. The user must deter-
mine the noise covariance matrices in the filter, which deter-
mine how rapidly the parameters change. Analytical methods
of estimating these matrices resulted in overestimation and
subsequent filter instability. It appears that the problem re-
sults from inaccuracies in the measurement and areal averaging
of precipitation and degree days. The fairly large errors in-
herent in these data may cause the parameter values to
fluctuate too rapidly.
Two additional advantages of the ARMAX formulation are
the availability of diagnostic checks to determine model
adequacy and the calculation of confidence limits for a fore-
•

the parameters could be allowed to vary on a daily basis
using a Kalman filter to update the parameters based on
measured streamfiow. This has been used in rainfall-runoff
models for a number of years (Todini and Bouillot, 1975)
and initial attempts in snowmelt runoff (Burn and McBean,
1985) appear promising. Daily values of co and oi for the
years 1975 and 1976 are shown in Figure 4. It can be seen
that the parameters increase rapidly in the early snowmelt
season, as the degree day effectiveness in causing melt and
soil moisture are both increasing. At the peak of the snowmelt
cast. The presence of significant autocorrelation in the fore-
cast error sequence is an indication of model inadequacy.
This represents the presence of additional "information," in
the data which the model is not using. Tests of the SRM
model and ARMAX models without moving average terms
showed the presence of residual autocorrelation. The inclu-
sion of moving average terms increased the forecasting ac-
curacy of the ARMAX model and eliminated the autocorrela-
tion. Confidence limits for the forecast can be obtained using

150 —
WI

100

0>
,_ 50
w
4-
a,
(a) E
0
00
0

i i i .i 'J 1 .1. .1 1 i. • .1.

Apr. May Jun. Jul. Aucj. Sep. Apr. May Jun. Jul. Aug. Sep.
1975 1976

Figure 4. Kalman Filter Estimates of Snowmelt Parameters (Co and i) in the ARMAX
Model Described by Equation (6) and Equation (7) for the Years 1975-76.

1087 WATER RESOURCES BULLETIN

 Haitmer and Salas
the variance of the forecast errors. Theory and results are
available in Haltiner (1985).
RESULTS
The previously described models were used to make one-
and three-day ahead flow forecasts for the period 1973-1980.
In addition, a simple ARMA(1,l) model which only uses pre-
vious values of streamfiow (no snowmelt/rainfall inputs) was
also included; it represents the forecasting accuracy which
can be achieved relatively simpiy without attempting to model
the rainfall or snowmelt. Models were compared using a mean
squared error criterion (MSE). The model forecasting results
are presented in Table 1. This table also includes the num-
ber of parameters which are estimated during the forecasting
period and an indication of the presence of significant auto-
correlation in the forecast errors. Model 1 represents the
Snowmelt Runoff Model described by Shafer, et a!. (1982),
adapted to include measured data. Average parameters were
calculated for each 15-day period. Model 2 is the simple time
series model with six sets of parameters based on flow thres-
holds. Models 3 and 4 correspond to the ARMAX models
described in Equation (1), using a combined snowmelt/rainfall
input. Model 5 is the ARMAX model described in Equation
(6), with separate snowmelt and rainfall inputs.
A comparison of the MSE shows that the ARMA(1,1),
which does not include rain or snowmelt, gives the poorest
forecasts. The ARMAX Models 3, 4, and 5 give better fore-
casts with fewer parameters than the SRM model. Since the
form of the SRM and ARMAX models is similar, the smaller
MSE of the latter result from the analytical parameter estima-
tion techniques and the inclusion of moving average terms.
The presence of residual autocorrelation in the SRM model
indicates that the model forecasts could be improved by in-
cluding the previous model errors in the forecasts.
TABLE 1. Comparison of One-Day-Ahead Forecasting Results for the Period 1973-1980.
Model MSE*
1. SRM 41.3
(seasonal param.)
2. ARMA(1,1) 56.3
(threshold param.)
3. ARMAX-1 38.8
(combined input/threshold param.)
4. ARMAX-2 35.1
(combined input/seasonal param.)
5. ARMAX-3 33.5
(sep. snow/ppt input, seasonal param.)
*MSE = mean square error.
1088
Figure 5 presents an example of the one-day-ahead fore-
cast and measured streamfiows for the 1973 snowmelt season
using Model 4. (It should be noted that in using the ARMAX
and SRM models in this study, the temperature and precipita-
tion input data on the forecast day are assumed known. In
operational use, these data would be forecast, using either
the National Weather Service forecast or autoregressive type
forecasts.)
The results in Table 1 were obtained using all eight years
of data for parameter estimation and forecasting. In a true
forecasting situation, the parameters must be estimated sequen-
tially, using data available up to the time of the forecast.
To accomplish this, the seasonal and threshold ARMAX
models were reformatted so that the first year's data was
used to estimate model parameters and the model then used
to forecast flows during the second year. Following this, the
model parameters were re-estimated using the first two years
of data and the model used to forecast the third year of
flow, etc. The Kalman filter parameter estimation scheme
operates recursively, re-estimating the parameters each day
based on past parameter values and the most recent measure-
ments. The SRM model could not be included in this com-
parison as there is no version currently available which allows
analytical sequential updating of the model parameters.
The results of this experiment for the one-, two-, and
three-day ahead forecasts are shown in Table 2. The seasonal
model provided the most accurate forecasts, followed by the
Kalman filter scheme. The threshold model resulted in the
highest MSE values.
CONCLUSIONS
The ARMAX formulation of the SRM model appears to
offer a number of advantages over current procedures.
Analytical techniques are available to identify model format
No.of
Parameters Res. Corr. Test
84 Failed
12 Passed
60 Passed
88 Passed
78 Passed
WATER RESOURCES BULLETIN
 Short-Term Forecasting of Snowmelt Runoff Using ARMAX Models

and estimate model parameters, and diagnostic checks are LITERATURE CITED
available to ensure that the model is operating properly. The
Box, G. E. P. and G. M. Jenkins, 1970. Time Series Analysis, Fore-
model can be updated in real time using previous forecast casting and Control. Holden-Day Inc., San Francisco, California.
errors; in addition, the Kalman filter can be used to update Burn, D. H. and E. A. McBean, 1985. River Flow Forecasting Model
the model parameters in real time. for Sturgeon River. ASCE J. of Hyd. Eng., VoL III, No. 2.
Haltiner, J. P., 1985. Stochastic Modeling of Seasonal and Daily
Streaniflow. Ph.D. Dissertation, Colorado State University, Fort
TABLE 2. Comparison of Forecasting Results (1974-1980) Collins, Colorado.
Using Sequential Parameter Estimation. Martinec, J., 1960. The Degree Day Factor for Snowmelt-Runoff
Forecasting. Proceedings: General Assembly of Helsinki, 1960.
Mean Square Error Commission on Surface Runoff. PubI. lASH No. 51, pp. 468-477.
Model
Martinec, J. and A. Rango, 1986. Parameter Values for Snowmelt
One-Day Two-Day Three-Day Runoff Modeling. J. Hydrol. 84:197-219.
Rango, A., 1983. Application of a Simple Snowmelt-Runoff Model to
1. Threshold Model 54.7 134.8 208.0 Large River Basins. Proceedings of the 51st Western Snow Con-
2. Seasonal Model 49.6 112.9 184.6 ference, Vancouver, Washington, pp. 89-99.
3. Kalman Filter Model 50.3 123.7 202.5 Rango, A. and J. Martinec, 1979. Application of a Snowmelt-Runoff
Model Using Landsat Data. Nordic Hydrology 10:225-238.
Shafer, B. A., E. B. Jones, and D. M. Frick, 1982. Snowmelt Runoff
Modeling in Simulation and Forecasting Modes with the Martinec-
Rango ModeL NASA Report CR 170452, Goddard Space Flight
ACKNOWLEDGMENTS Center, Greenbelt, Maryland.
Todini, E. and D. Bouillot, 1975. A Rainfall-Runoff Kalman Filter
This work was supported by the Colorado Agricultural Experiment
Station Projects COLO 0114 and COLO 0357. Portions of this paper
ModeL In: System Simulation in Water Resources, G. C. Van-
steenkiste (Editor). North Holland, Amsterdam.
were presented at the 53rd Annual Western Snow Conference, held
April 16-18, 1985, at Boulder, Colorado.

U.. — &asuqD FLOU

I DAY FORECAST

U.S

, 4...

2$SS

April May June July August September

Figure 5. Comparison of Measured Flows and One-Day-Ahead

Forecasts for 1973 Using Model 4 of Table 1.

1089 WATER RESOURCES BULLETIN