Realistic Forecasting of Groundwater Level, Based on

the Eigenstructure of Aquifer Dynamics

V. J. Bidwell

Lincoln Environmental, Lincoln Ventures Ltd, Lincoln, New Zealand ([email protected])

Abstract: Short-term management of groundwater resources, especially during droughts, can be assisted by
forecasts of groundwater levels. Such forecasts need to account for the natural dynamic behaviour of the
aquifer, likely recharge scenarios, and recent but unknown abstractions. These requirements mean that
forecasts, at say monthly intervals, need to be updated with current observations on a real-time basis. One
established procedure for this kind of problem is to fit autoregressive, moving-average, exogenous-variable
(ARMAX) time-series models to the history of groundwater levels in response to estimates of land surface
recharge. The ARMAX difference equations are then converted into forecast equations that allow real-time
updating to include recent forecast errors as an additional source of information. Some disadvantages of this
pure time-series analysis approach are the apparent lack of physical concepts in the model formulation and
statistical aspects of model identification and calibration that are related to the inherent structure of ARMAX
equations. This paper addresses these issues by describing a method for formulating ARMAX forecast
equations from a linear system description based on the eigenvalues and eigenvectors (eigenstructure) of the
dynamic behaviour of an aquifer. For the piezometric response of a heterogeneous aquifer to a fixed spatial
distribution of land surface recharge, with time-varying magnitude, only a few eigenvalues are significant for
describing the dynamics. The resulting model has a simple robust parameter structure, and is easily
calibrated and implemented in spreadsheet form. The eigenstructure approach enables transfer of some
parameter information from locations with good data records to those with sparse data. This modelling
approach is demonstrated with monthly values of land surface recharge, estimated from a daily water balance
model, and groundwater level data from an observation well in a 2000 km2 alluvial aquifer in Canterbury,
New Zealand.

Keywords: Groundwater forecasting; Aquifer dynamics; Eigenvalue; ARMAX

are considered either as univariate stochastic

1 INTRODUCTION processes driven by random inputs, or as bivariate
processes that include other observed inputs. This
Groundwater storage in aquifers may be
essentially “black box” approach has been applied
considered, in the natural state, as the dynamic
to groundwater fluctuations by Tankersley and
balance between recharge, driven by climatic
Graham (1993), in both univariate form and
processes, and discharge to surface waters. Most
bivariate form with rainfall as the input variable.
of the dynamic behaviour is caused by variations
Ahn (2000) used Kalman filter updating of a
in recharge through the land surface rather than
time-series model for groundwater levels at
from rivers. This balance is modified by
multiple sites in a multi-layered aquifer system.
groundwater abstraction for human use. During
periods of climatic drought, land surface recharge Time-series analysis can also be considered as a
is zero, and groundwater levels decline at a rate statistical approach to identifying linearised
determined by natural storage-discharge dynamics descriptions of physical processes (e.g., Young,
and the effect of, usually, increased abstractions 1984). Bidwell et al. (1991) applied the method
for purposes such as irrigation. Management of of Young (1984) to a simplified conceptual model
the resource under these conditions can be of groundwater level variations in response to
assisted by short-term forecasts of groundwater recharge estimated as soil-water drainage from a
levels (Ahn, 2000). water balance model.
Forecasting of the uncertainties and dynamics of Most mathematical descriptions of dynamic
many economic and natural processes is based on groundwater behaviour are in the form of
the mathematics of time-series analysis (e.g., Box linearised partial differential equations. This
and Jenkins, 1970). The values of observed states spatially-distributed linear process can also be
represented as a linear system of interconnected referred to as an ARMAX equation, with general
discrete-space components, such as finite form:
difference or finite element schemes, or as a
system of conceptual linear components based on B( z −1 ) D( z −1 )
the eigenvalue solution to the distributed process
Yn = Xn + en (3)
A( z −1 ) C ( z −1 )
(Sahuquillo, 1983). Application of time-series
analysis to this multi-component system This model (3) is also referred to as a transfer-
description of groundwater dynamics offers the function-noise (TFN) description, when the
potential for completely linking aquifer emphasis is more on identifying a system in the
properties, boundary conditions, recharge presence of noise rather than simulating the entire
processes, and external stresses to the stochastic statistical structure.
forecasting equations. This paper describes an
approach that matches the stochastic difference
equation models of time-series analysis to the 2.2 Model identification and calibration
physically-based, linear system, groundwater Identification of (3) refers to the number of terms
model of Sahuquillo (1983). in the polynomials A, B, C, D in z-1, and
calibration refers to the values of their
2 TIME-SERIES MODELS coefficients, such as the ai and bi in (1). The
values of the observations Xn, Yn are assumed to
The dynamic behaviour of many natural systems be deviations from equilibrium values, such as
can be modelled in terms of stochastic linear averages or zero-input levels. However, in some
difference equations, for situations where situations these datum values may be unknown
observations about the system are available at and therefore become part of the parameter set.
regular time intervals. The general structure of a
difference equation that describes the dynamic Superficially, these difference equations would
relationship between an input time series Xn and appear to be amenable to least-squares regression,
an output series Yn is: but this is precluded by the nature of the noise and
interdependence among lagged variables, which
Yn = a1Yn −1 + a 2 Yn − 2 + Λ + a p Yn − p cause bias in parameter estimation. Two
(1) approaches to identification and calibration are
+ b1 X n + b2 X n −1 + Λ + bq X n − q +1
provided, respectively, by Box and Jenkins (1970)
Equation (1) relates the output series to past and Young (1984).
output values (autoregressive) and a “moving
average” of present and past input values. This 2.3 Forecast equation
autoregressive, moving-average (ARMA)
structure requires fewer model parameters than The stochastic linear dynamic model (3) can be
the equivalent purely AR or MA form. By means rewritten in the form:
of the z-operator (discrete-time equivalent of the
Laplace operator) in its time-shift form z-mYn A( z −1 )C ( z −1 )Yn = C ( z −1 ) B( z −1 ) X n
(4)
≡Yn-m, (1) can be expressed as the transform: + A( z −1 ) D ( z −1 )en

(b1 + b2 z −1 + Λ + bq z − q +1 ) Equation (4) includes present and past values of

Yn = −1 −2 −p
Xn en. These are also the forecast errors at one time
(1 − a1 z − a2 z −Λ − a p z )
(2) step ahead, given that the input Xn is known.
B( z −1 ) Inclusion of the past values of forecast error
= Xn provides a self-correction component that can
A( z −1 )
take account of the dynamic effect of unobserved
Traditional time-series analysis (Box and Jenkins, influences.
1970) considers a univariate stochastic series,
after removal of trends and seasonal effects, as 3 GROUNDWATER MODEL
the output Nn from an ARMA model (2) with
uncorrelated noise en as input. Time-series analysis is a “black box”, external
approach to modelling an assumed linear dynamic
system that links the observed input and output
2.1 ARMAX equations series. The following sections describe how the
When the stochastic difference equation includes structure of a linear systems model of piezometric
the exogenous (X) input Xn as well as the response to groundwater recharge can be
stochastic term Nn then the model structure is synthesised from an understanding of
groundwater flow and storage in aquifers. The
resulting model can then be expressed in forecast The eigenvalues are functions of transmissivity,
equation form (4), with some advantages in storativity, aquifer geometry, and boundary
identification and calibration relative to the time- conditions, but these need not be determined in
series approach. the present application. The corresponding
eigenvector coefficients gi(x,y) define the
contribution Ui(t) of each conceptual water
3.1 Eigenstructure approach storage to the dynamic piezometric effect U(x,y,t)
Sahuquillo (1983) shows how the dynamic relative to the steady piezometric effect D(x,y) at
behaviour of piezometric head at any location in location (x,y), so that:
an aquifer can be modelled in terms of the
U i (t ) = g i ( x, y )Wi (t )
eigenstructure of the solution to the groundwater
U ( x , y , t ) = ∑ U i (t ) (6)
flow equation. The eigenstructure, comprising
the eigenvalues and eigenvectors, is a means of i
representing the distributed linear aquifer system The transfer function of R(t) to Ui(t) by each
as an infinite series of first-order water storages water storage contribution, in terms of the
arranged in a parallel system structure. For the discrete-time samples Uni and Rn, can be
particular case of piezometric response to a fixed expressed in ARMA form (2) as:
spatial pattern P(x,y) of land surface recharge,
with time-varying magnitude R(t), only a few of βi
these water storage components are required for U ni = Rn (7)
(1 − α i z −1 )
acceptable model accuracy. These water storages
are conceptual and do not exist in any physically for which:
definable form.
α i = exp(−k i ∆t )
(8)
β i = g i ( x, y )(1 − α i )
3.2 Linear system model
Figure 1 shows the parallel system structure The relationship (8) also holds for the vadose
representing aquifer dynamics, and an additional zone storage, but kv(x,y) varies with location and
component in series for modelling the dynamic the gain coefficient is unity, so that βv =(1−αv).
effect of transport through the vadose zone and The piezometric response H(x,y,t), at a particular
perching of groundwater above aquitards. location, to the input series of land surface
recharge R(t) over the whole aquifer is usually
Land surface recharge R(t)
observed relative to an arbitrary datum.
Vadose zone
However, the dynamic response U(x,y,t) of the
& perched model in Figure 1 is superimposed on the
kv(x,y)
groundwater assumed steady piezometric effect of other
storage recharges (e.g., from rivers) and abstractions.
This steady effect D(x,y) is an additional
Aquifer parameter at each location. For convenience, the
k1 k2
k3 storage observed piezometric time series Hn(x,y) at a
particular location is transformed to:
g1(x,y) g2(x,y) g3(x,y)
U n = H n ( x, y ) − D ( x, y ) (9)
Piezometric level U(x,y,t) relative to D(x,y)

Figure 1. Eigenstructure conceptual model. 3.3 ARMAX and forecast equations

The transfer-function-noise description of the
The eigenvalues ki are the same at all locations, system structure (Figure 1), with only the first
and can be considered as the discharge two eigenvalues, is:
coefficients of the conceptual water storages, for
which the response of the water content Wi(t) to βv
recharge R(t) can be described by the solution of a Un =
(1 − α v z −1 )
first-order differential equation over a time
interval ∆t:  β1 β2 
× −1
+ −1
 Rn (10)
Wi (t ) = Wi (t − ∆t ) exp(−k i ∆t )  (1 − α 1 z ) (1 − α 2 z ) 
(5)
+ R(t )[1 − exp(− k i ∆t )] + Nn
When (10) is multiplied out into the form of (2):
 (b1 + b2 z −1 ) calibrated at locations with good data may be
Un = Rn + N n (11) transferred to locations with less data. In practice,
(1 − a1 z −1 − a 2 z − 2 − a3 z −3 ) this is especially true of α1, corresponding to the
The relationships between the coefficients of (10) smallest eigenvalue k1.
and (11) are obtained by equating powers of z-1. Optimisation is conducted on subsets of the
The noise component Nn can be described parameters to facilitate stable convergence to final
adequately by a first-order structure, for many values, in the following order:
practical applications, and thus the complete D ( x, y )
ARMAX model becomes:
D( x, y ),α 1 , β1
(b1 + b2 z −1 ) D( x, y ),α1 , β1 ,α 2 , β 2
Un = Rn
(1 − a1 z −1 − a 2 z − 2 − a3 z −3 ) Λ Λ Λ Λ Λ
(12)
1 The best of the above, plus α v
+ en
(1 − f1 z −1 ) Then add the noise parameter f1
When (12) is multiplied out to the form of (4) and
transformed to the time domain, the resulting 4 APPLICATION
forecast equation is:
U n = c1u n −1 + c 2 u n − 2 + c 3 u n −3 + c 4 u n − 4 4.1 Aquifer description
+ d 1 R n + d 2 R n −1 + d 3 R n − 2 The Central Canterbury Plains, in the South
(13)
− a1e n −1 − a 2 e n − 2 − a 3 e n −3 Island of New Zealand, overlay an interconnected
+ en system of unconfined and semi-confined aquifers
in silt-sand-gravel alluvium up to about 500 m
for which ci, di are obtained from ai, bi, f1 in (12) thick. This 2000 km2 region is approximately
by equating powers of z-1. 50 km long from mountains to the sea and 40 km
wide between two large braided rivers that cross
The forecast equation (13) shows how, for this the plains from mountains to sea. These rivers are
example, the three previous values of forecast perched above the aquifers over much of their
error contribute to making a new forecast one length and are a significant source of recharge.
time step ahead.
The other significant source is land surface
recharge under a mixture of land uses including
3.4 Model identification and calibration irrigated and dryland pasture, crop farming, and
Equations (9) – (13) are implemented in an Excel forest.
spreadsheet, as difference equation formulas, Theoretical results from aquifer dynamics
together with the relationships (not shown in this indicate that variations in piezometric head are
paper) between the coefficients of (10), (11), (12) caused primarily by land surface recharge, and
and (13). Parameter estimation is conducted by that these are superimposed on a steady
means of the “solver” optimisation tool in Excel. piezometric surface caused by river recharge and
The objective function is minimisation of the the effect of average groundwater abstraction.
sums of squares of the forecast errors en. Therefore, land surface recharge is the only
The parameter search is conducted on D(x,y) and exogenous variable in the analysis.
the coefficients αv, βv, αi βi of (10), rather than
the coefficients of (13). The reasons for this 4.2 Estimation of land surface recharge
approach are, firstly, the former parameters are
structurally independent and, secondly, the setting Rainfall varies from about 1000 mm/y near the
of initial values and constraints is easier. All the mountains to about 600 mm/y near the sea, and
parameters are set to an initial value of zero, in there is considerable variation in soil type and
the absence of any prior knowledge. The set of land use throughout the region. However, the
parameter constraints is: eigenstructure approach requires only that land
surface recharge be of the form P(x,y)R(t). The
0 ≤ α v ,α i < 1 fixed spatial pattern P(x,y) forms part of the linear
(14)
βv , βi ≥ 0 system to be analysed, and only a single
magnitude series R(t) is required as input to the
Because the eigenvalues are, theoretically, the system.
same at all locations in the aquifer, values of αi
 -5 200
Observed
-6 Predicted 180
1-mth forecast
-7 160

Land surface recharge (mm/mth)

Land surface recharge
-8 140
Groundwater depth (m)

-9 Drought seasons 120

-10 100

-11 80

-12 60

-13 40

-14 20

-15 0
1995 1996 1997 1998 1999 2000
Figure 2. Comparison between the observed, predicted, and forecast piezometric levels at
monthly intervals for Observation Well M36/0255, in response to land surface recharge.

A monthly time-series Rn of land surface recharge piezometric datum D(x,y), and noise parameter f1
was estimated from a water balance model are shown in Table 1.
calculated on a daily basis. Rainfall and climatic
Table 1. Eigenstructure-noise parameters for
data for Penman estimates of evaporation, at one
Observation Well M36/0255.
location, were used for calculating water
consumption and drainage R(t) from dryland kv(x,y) (mth-1) 0.970
pasture. The mean value of R(t) was 196 mm/y.
-1
k1 (mth ) 0.046
4.3 Observation well k2 (mth-1) 0.970
-1
The eigenstructure modelling approach has been g1(x,y) (m mm mth) 0.279
applied to several observation well records in the g2(x,y) (m mm mth) -1
0.037
region (Bidwell and Morgan, 2002), of which one
(M36/0255) was selected to illustrate this D(x,y) (m) -15.02
forecasting application. The piezometric f1(x,y) 0.711
observations at this well are a good indicator of
low flow in the Halswell River, about 13 km
distant, which is supplied by the aquifer. The transfer function part of the model (2)
Therefore, the ability to forecast at this well accounted for 87% of the variance of the
enables early detection of likely environmental piezometric time-series Un, and addition of the
effects in the river caused by excessively low noise component Nn raised this to 94% for a
flow during droughts, because the likely cause is one-step-ahead forecast. The remaining 6% was
lowering of piezometric levels by additional forecast error en, which had a standard deviation
groundwater abstraction in the region. A 17-year of 0.341 m. The randomness of the en series was
record of monthly observations (1983-2000) was demonstrated by means of the portmanteau lack
used, with a few missing data that are easily of fit test (Box and Jenkins, 1970; 8.2.2) applied
handled by the method. to the first 12 autocorrelation coefficients.
Figure 2 shows a comparison between observed
5 RESULTS piezometric head, prediction from the transfer
The significant parameter values of the function model, and one-month-ahead forecast
eigenmodel system (Figure 1), steady local from the complete transfer-function-noise model.
The five-year record includes two drought
seasons when groundwater abstractions caused
significant departure of piezometric head from flow equations. Decisions about the number of
that predicted on the basis of climatic data. terms to be included in (12) are made in (10), for
which the parameters are relatively independent
under structural change. This approach contrasts
6 DISCUSSION AND CONCLUSIONS with conventional time-series analysis in which
Only the first two eigenvalues were required to the parameters of (12) all change with each
account for the dynamics of the aquifer system at structural addition.
monthly time intervals. The dynamic behaviour
The physical realism of the eigenstructure means
is dominated by the first eigenvalue k1, with its
that prior knowledge of the first eigenvalue (at
associated gain (0.279) equal to 88% of the total
least) can be transferred between observation
g1(x,y)+g2(x,y). The value of k1 = 0.046 mth-1
sites, and sensible initial values and constraints
corresponds to a mean storage time (1/ k1) of 22
can be set for parameter optimisation. These
months. This value is similar to the first
enable effective use of the optimisation (solver)
eigenvalue for other observation wells in this
function available in Microsoft Excel, rather than
aquifer system (Bidwell and Morgan, 2002). The
dedicated time-series-analysis software based on
second eigenvalue k2 corresponds to a storage
a statistical approach.
time of about one month, and accounts for minor
transient behaviour.
7 REFERENCES
The value (0.970) of the vadose zone coefficient
kv(x,y) also corresponds to a mean storage time of Ahn, H., Ground water drought management by a
about one month for the system component that feedforward control method, Journal of the
accounts for hydraulic transport lag at this American Water Resources Association,
location, as well as any effects of the one-month 36(3), 501-510, 2000.
sampling interval. There is no apparent effect of
Bidwell, V.J., P.F. Callander and C.R. Moore, An
any perched groundwater at this well, and
application of time-series analysis to
groundwater level variations occur above the
groundwater investigation and
steady-state effect at 15 m below ground level
management in Central Canterbury, New
(given by D(x,y)).
Zealand, Journal of Hydrology (New
The first-order noise model (with f1 = 0.711) was Zealand), 30(1), 16-36, 1991.
quite satisfactory for simulating the dependence
Bidwell, V.J. and M.J. Morgan, The eigenvalue
structure, as shown by the test result for lack of
approach to groundwater modelling for
dependence in the residuals en.
resource evaluation at regional scale,
Figure 2 shows that the transfer function (first Proceedings 4th International Conference
term of (12)) can be a good predictor of on Calibration and Reliability in
piezometric level during seasons of “normal” Groundwater Modelling, Prague, 92-95,
climatic stress, but not during the drought seasons 2002.
of 1997/98 and 1998/99. During these periods the
Box, G.E.P., and G.M. Jenkins, Time Series
levels were significantly lower, probably due to
Analysis Forecasting and Control,
increased groundwater abstraction for irrigation.
Holden-Day, 553 pp., San Francisco, 1970.
However, the one-month-ahead forecasts from
(13) effectively tracked the observed departure Sahuquillo, A., An eigenvalue numerical
and would have provided indication of increased technique for solving unsteady linear
flow depletion in the Halswell River, on a groundwater models continuously in time,
real-time basis. Water Resources Research, 19(1), 87-93,
1983.
The forecast series (Figure 2) has been calculated
as if the future input of land surface recharge is Tankersley, C.D., W.D. Graham and K. Hatfield,
already known. For practical application during Comparison of univariate and transfer
drought periods, the future input may be function models of groundwater
considered to be zero for worst-case management fluctuations, Water Resources Research,
purposes, or likely non-zero scenario values may 29(10), 3517-3533, 1993.
used for a broader strategic view.
Young, P., Recursive Estimation and Time-Series
The eigenstructure approach provides a sound Analysis, Springer-Verlag, 300 pp., Berlin,
basis for identifying the structure of the ARMAX 1984.
equation (12), because the coefficients in (12) are
derived directly from the parallel system (10)
based on theoretical solutions of the groundwater