See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/253300826

Kp forecast models

Article in Journal of Geophysical Research Atmospheres · April 2005

DOI: 10.1029/2004JA010500

CITATIONS READS

127 354

11 authors, including:

S. Wing Jay R Johnson

Johns Hopkins University Andrews University
210 PUBLICATIONS 4,973 CITATIONS 179 PUBLICATIONS 2,976 CITATIONS

SEE PROFILE SEE PROFILE

D. G. Sibeck
NASA
589 PUBLICATIONS 16,703 CITATIONS

SEE PROFILE

All content following this page was uploaded by S. Wing on 21 October 2016.

The user has requested enhancement of the downloaded file.

 JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 110, A04203, doi:10.1029/2004JA010500, 2005

Kp forecast models
S. Wing,1 J. R. Johnson,2 J. Jen,1 C.-I. Meng,1 D. G. Sibeck,3 K. Bechtold,1 J. Freeman,5
K. Costello,6 M. Balikhin,4 and K. Takahashi1
Received 25 March 2004; revised 21 December 2004; accepted 27 January 2005; published 9 April 2005.

[1] Magnetically active times, e.g., Kp > 5, are notoriously difficult to predict, precisely
the times when such predictions are crucial to the space weather users. Taking advantage
of the routinely available solar wind measurements at Langrangian point (L1) and nowcast
Kps, Kp forecast models based on neural networks were developed with the focus on
improving the forecast for active times. To satisfy different needs and operational
constraints, three models were developed: (1) a model that inputs nowcast Kp and solar
wind parameters and predicts Kp 1 hour ahead; (2) a model with the same input as
model 1 and predicts Kp 4 hour ahead; and (3) a model that inputs only solar wind
parameters and predicts Kp 1 hour ahead (the exact prediction lead time depends on the
solar wind speed and the location of the solar wind monitor). Extensive evaluations of
these models and other major operational Kp forecast models show that while the new
models can predict Kps more accurately for all activities, the most dramatic improvements
occur for moderate and active times. Information dynamics analysis of Kp suggests
that geospace is more dominated by internal dynamics near solar minimum than near solar
maximum, when it is more directly driven by external inputs, namely solar wind and
interplanetary magnetic field (IMF).
Citation: Wing, S., J. R. Johnson, J. Jen, C.-I. Meng, D. G. Sibeck, K. Bechtold, J. Freeman, K. Costello, M. Balikhin, and
K. Takahashi (2005), Kp forecast models, J. Geophys. Res., 110, A04203, doi:10.1029/2004JA010500.

1. Introduction 1949. It was extended backward to 1932 in 1951. Thus the

name Kp has a German origin and is an acronym for
[2] As technology advances, space weather prediction has
‘‘planetarische Kennziffer,’’ which simply means planetary
become increasingly important to our nation’s defense,
index. One Kp value is produced for each 3 hour interval of
commerce, and research activities. For example, space
universal time (UT) at 0 – 3, 3 – 6, 6 – 9, . . .21 – 24. The
weather can affect communications, navigation systems,
previous month or older Kp values and other Kp related
satellite health, power grids, and space travel. Kp is one
information are available at GeoForschungsZentrum (GFZ),
of the most common indices used to indicate the severity of
Postdam, Germany Web site (http://www.gfz-potsdam.de/
the global magnetic disturbances in near-Earth space. Kp is
pb2/pb23/GeoMag/niemegk/kp_index/) and several of its
an index based on the average of weighted K indices at
mirror sites, including at the USA’s National Oceanographic
13 ground magnetic field observatories. It is based on the
and Atmospheric Administration (NOAA) (ftp://ftp.ngdc.
range of the magnetic field variation within 3 hour intervals
noaa.gov/STP/GEOMAGNETIC_DATA/INDICES/
that is caused by phenomena other than the diurnal variation
KP_AP/). A review of the Kp derivation method was given
and the long-term components of the storm time variations.
by Rostoker [1972].
The values of the Kp range from 0 (very quiet) to 9 (very
[3] The relatively long, uninterrupted Kp record since
disturbed) in 28 discrete steps, resulting in values of 0, 0+,
1932 makes this index useful for studying solar wind-
1
, 1, 1+, 2
, 2, 2+, . . .9. The Kp index was first
magnetosphere interactions and space weather. There have
introduced by Bartels [1949] and has been published by
been many studies that show the correlations between Kp
the Institut für Geophysik der Universität Göttingen since
and various parameters of the solar wind and interplanetary
magnetic field (IMF) [e.g., Papitashvili et al., 2000;
1
Johns Hopkins University Applied Physics Laboratory, Laurel, Crooker and Gringauz, 1993; Garrett et al., 1974]. Kp
Maryland, USA.
2
has also been shown to correlate to many geospace
Princeton Plasma Physics Laboratory, Princeton University, Princeton, phenomena. For example, the location of the substorm
New Jersey, USA.
3
NASA Goddard Space Flight Center, Greenbelt, Maryland, USA.
injection has been shown to have a Kp dependence [Mauk
4
Department of Automatic Control and System Engineering, University and McIlwain, 1974]. The stretching of the field line, the
of Sheffield, Sheffield, UK. earthward boundary of the plasma sheet, and the ion
5
Physics and Astronomy Department, Rice University, Houston, Texas, isotropy boundary have been shown to correlate well with
USA.
6
NASA Lyndon B. Johnson Space Flight Center, Houston, Texas, USA.
Kp [Wing and Newell, 2003; Newell et al., 1998; Sergeev et
al., 1993]. Kp also plays significant roles in space weather,
Copyright 2005 by the American Geophysical Union. e.g., many satellite operators use Kp to estimate satellite
0148-0227/05/2004JA010500$09.00 drag. Many magnetospheric and ionospheric models require

A04203 1 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

Kp as an input parameter. For example, the Tsyganenko developed such a model based on an artificial neural
[1989] magnetic field model, atmospheric density models network (NN) algorithm. This model is now routinely
[e.g., Hedin, 1987], ring current – radiation belt models [e.g., operational, and its predictions can be obtained at the
Fok et al., 2001], conductivity models [e.g., Hardy et al., NOAA Web site (http://www.sec.noaa.gov/rpc/costello/
1987], and the Magnetospheric Specification Forecast Mod- index.html). More recently, Boberg et al. [2000] developed
els (MSFM) (J. W. Freeman, The magnetospheric specifi- their own NN Kp model that also inputs solar wind.
cation and forecast model, unpublished manuscript, 1995, [8] Moderate and high magnetic activities are notoriously
available at http://hydra.rice.edu/freeman/ding/www/ difficult to predict [Joselyn, 1995], precisely when predic-
msfm95/index.html) all require Kp as an input parameter. tions become more crucial for space weather users. The
Also, the recently developed Oval Variation, Assessment, previous Kp models also exhibit this typical behavior
Tracking, Intensity, and Online Nowcasting (OVATION) [e.g., Detman and Joselyn, 1999]. These models, which
model can use Kp as an input parameter to determine the were developed at the time when nowcast Kp models were
equatorial boundary of the auroral oval [Newell et al., not yet operational, are driven solely by solar wind/IMF.
2002]. For space weather, what the models such as MSFM However, it can be reasonably expected that if the model
and OVATION do, in effect, is to take Kp, which provides also inputs parameters that give the present and/or the
qualitative alerts and use it to produce more quantitative history of the state of the magnetosphere in addition to
alerts, such as magnetospheric/ionospheric particle fluxes the solar wind/IMF driver, the predictions would improve.
and electromagnetic fields, auroral oval location and fluxes, For example, Johnson and Wing [2004, 2005] show that
etc. As a result, precautionary measures could be taken to past Kps have strong linear and nonlinear correlations with
avoid or reduce catastrophic damage to power grids and future Kps.
satellites. [9] This paper presents new Kp models that have signif-
[4] As a global geomagnetic activity index, Kp has some icantly higher forecast accuracies than the previous Kp
flaws, although other indices have their own difficulties. operational models. In order to satisfy different needs and
The midlatitude locations (48
– 63
magnetic latitude) of operational constraints, three different models were devel-
the 13 midlatitude stations used to compute Kp render it oped: (1) a model that inputs nowcast Kp and solar wind
rather difficult for deducing the source(s) of its variations, parameters and predicts Kp 1 hour ahead; (2) a model with
leading to some ambiguities in its interpretations. Therefore the same input as model 1 and predicts Kp 4 hours ahead;
as space physics advances, Kp will probably be supplanted and (3) a model that inputs only solar wind parameters and
by a newer index or parameter that can indicate less predicts Kp 1 hour ahead. The 1 hour and 4 hour prediction
ambiguously the state of the magnetosphere. However, for lead times are just rough estimates for solar wind monitor at
the reasons mentioned in the previous paragraph and L1. Of course, the actual prediction lead times may vary,
continuity of the space weather operations, for now and in depending on the solar wind V and the location of the solar
the near future, Kp is likely to continue to play some roles in wind monitor.
space physics and space weather.
[5] As mentioned, the official Kp is published with a few
weeks’ delay, which is usually more than adequate for space
2. Data Set
physics research that analyzes old observations. Unfortu- [10] One shortcoming of many early models is that they
nately, this long delay makes it less useful for the space were trained on a limited data set. For example, the Costello
weather operations. For that reason, Gehred et al. [1995] NN was trained on 7 years of data from 1970, 1976, 1978,
developed a nowcast Kp algorithm, which takes real-time 1980, 1981, 1982, and 1989, but the limited availability of
data from several magnetometer stations, not necessarily the the solar wind data from IMP-8 effectively reduces the
same ones used for official Kp, and applies a similar method usable Kp data by more than 50% [e.g., see Wing et al.,
employed by official Kp to derive Kp estimates. The 1995]. The amount of historical solar wind data has multi-
resulting Kp estimates do not always exactly match the plied in the time since the Costello model was developed.
official Kps, but the advantage of this algorithm is that it Because there are dynamical variations over the course of
can produce estimated Kps in near real-time. This nowcast the solar cycle, it is necessary to build a database that can
Kp has been routinely produced by the United States Air capture such variations to the extent possible. Therefore we
Force (USAF) 55th Space Weather Squadron and made built a solar wind and Kp database that spans more than two
publicly available by NOAA through its Web site. solar cycles, 1975– 2001.
[6] More recently, Takahashi et al. [2001] developed a [11] We used the solar wind and IMF data from IMP-8
more sophisticated Kp nowcast algorithm that calculates Kp (1975– 1999), Wind (1994– 2000), and ACE (2000 – 2001).
fairly accurately. The Kp estimates from this model will The data are publicly available from NASA CDAWeb (http://
soon be made publicly available at the Johns Hopkins cdaweb.gsfc.nasa.gov/). The IMP-8 plasma data are publicly
University Applied Physics Laboratory (JHU/APL) Web available at the MIT IMP-8 Web site (ftp://space. mit.edu/pub/
site (http://sd-www.jhuapl.edu/UPOS/spaceweather.html). plasma/imp/www/imp.html).
[7] For the past few years, the Advanced Composition [12] The historical official Kp index is publicly available
Explorer (ACE) spacecraft, located upstream at Lagrangian at GFZ Postdam Web site (ftp://ftp.gfz-postdam.de/pub/
point (L1), has been reliably providing solar wind measure- home/obs/kp-ap/). Although the historical Kp database is
ments up to 30– 60 min in advance of their arrival at virtually continuous and uninterrupted from 1975 to 2001,
the near-Earth space environment. Therefore Kp forecast the solar wind and IMF data have gaps, effectively reducing
models based on solar wind input could use ACE observa- the number of Kp data points that can be used for model
tions to make short-term forecasts. In fact, Costello [1997] development.

2 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

propagation networks. Therefore our final models use

recurrent network architecture. In this study, all the NN
models use one hidden layer. The optimal number of hidden
nodes was empirically determined by systematically varying
the number of hidden nodes between 4 and 20, or higher if
needed. It turns out that most models reach the optimal
performances when the number of hidden nodes ranges
from 8 to 12. The NN only has one output node, which is
Kp. In addition, we developed a simple linear empirical
function that maps the NN output to observed Kp. The
function has the form of observed Kp = (a  NN Kp) + b,
where a and b are derived empirically from training data
sets. We note that this linear correction would only work if
NN Kp is already very good. It would not work (e.g., the
scatter would be too large) if the model output Kp was
terrible to begin with. The final predicted Kp is the output of
this function, which acts as a postprocessor for the NN.
[16] For the purpose of training and testing NNs, the solar
wind/IMF and Kp data set from 1975 – 2001 was randomly
Figure 1. Fifteen-minute interpolated Kp distribution in selected to form two equal subsets: (1) a training set and
the data set that spans over two solar cycles, 1975 –2001. (2) a test set. To estimate the arrival time of the solar wind
The test data set (dashed line) has distribution similar to the and IMF in the Earth space environment, the solar wind/
original full data set (solid line). The y-axis denotes the IMF were propagated to Earth (X = 0) with the assumption
number of events. that the planar solar wind phase front is perpendicular to the
Sun-Earth line (ballistic propagation). Then, solar wind
density (n), and velocity (V) and IMF were hourly averaged
[13] The monthly average International Sunspot Number at 15 min time intervals, which is the time granularity of the
is available at a NASA Web site (http://science.nasa.gov/ssl/ model outputs. In order to match the time granularity of Kp
pad/solar/sunspots.htm). (3 hours) to the model (15 min), the Kps are interpolated to
15 min resolution. This is done by time-tagging the Kp with
the center of the 3 hour interval, e.g., at 0130, 0430,
3. Method 0730 UT, etc., and then linearly interpolating between those
[14] This study used an NN paradigm to construct the Kp points. For operational considerations, 15 min is preferred
models. An NN is a mathematical model that attempts to over the traditional 3 hour time granularity. With 15 min
mimic the computation performed by the neurons in the time granularity, the models can warn the users of the
human brain. The advantage of this computational paradigm impending change in the space weather more quickly. For
is that it can be easily adapted to solve a wide range of example, using 15 min time granularity, if a satellite at L1,
problems that involve linear and/or nonlinear mappings e.g., ACE, detected a sudden change in the solar wind that
between a set of input to a set of output. NNs have been would cause a huge increase in geomagnetic activity, the
increasingly used to solve many problems in space physics, model would be able to warn the users with a new and
particularly as predictors of time series data. In addition to higher Kp in the next 15 min, before the actual arrival of the
the two NN Kp prediction models mentioned above, solar wind. On the other hand, using the traditional Kp
NNs have been used to predict relativistic electron flux 3 hour time resolution, the model would not be able to
at geosynchronous orbit [Koons and Gorney, 1991], ener- broadcast the impending catastrophic condition for up to the
getic storm particle (ESP) events impinging on the Earth next 3 hours, which might be too late for many operations.
[Vandegriff et al., 2005], geomagnetic storm [Wu and For practical purpose, the model 15 min Kp value can be
Lundstedt, 1997], etc. NNs have also shown great promises treated as the traditional 3 hour Kp value, as is presently
as signal/data classifiers. For example, an NN was used to done in Costello model.
classify geospace physical boundaries in the plasma data [17] In Figure 1, the solid line shows the distribution of
[Newell et al., 1990, 1991]. More recently, an NN was used the interpolated 15 min resolution Kp data points that have
to classify high-frequency (HF) radar backscattered signals corresponding good simultaneous solar wind and IMF
[Wing et al., 2003]. observations. In other words, the solid line essentially
[15] In this study we experimented with two types of shows the total number of data points that can be used for
general NN architectures: (1) the standard multilayered our study. The dashed line shows the Kp data in the test set.
feedforward – backpropagation network [Rumelhart and Figure 1 basically shows that the test data set has roughly
McClelland, 1987] and (2) the recurrent network in which the same distribution as the original full data set and hence
the outputs of the hidden nodes are fed back to inputs has no bias.
[Gershenfeld, 1999; Fernandez et al., 1990; Lo and Bassu,
1999]. We compared the best models in each architecture
and found that there is no significant difference in their 4. Previous Kp Models
performance. However, in general, the recurrent networks [18] For the purpose of model comparisons, three existing
can be trained more quickly than the feedforward – back- operational Kp prediction models are considered: (1) Cost-

3 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

by Balikhin et al. [2001] and Boaghe et al. [2001]. The

model inputs solar wind speed and dynamic pressure, IMF
jBj, Bz, and optionally a previous Kp (Kp at time = t

3 hours, where t is the solar wind arrival time on Earth). The
previous Kp could be its own previous prediction (feedback
mode) or a Kp estimate, i.e., a nowcast Kp. Over time, the
model performs more poorly if it inputs its own prediction
rather than nowcast Kp. Figure 3 shows the NARMAX Kp
performance using the same test data set and a nowcast Kp
(see section 5.1 for more discussion on nowcast Kp) as an
input parameter. From Figures 2 and 3, it can be seen that
NARMAX model predicts high Kps slightly better than the
Costello NN model, although the predictions of the former
have larger scatter. The correlation coefficient r = 0.77,
which is comparable to or just slightly higher than that of
the Costello model.
4.3. Boberg et al. NN Kp Model
[21] Boberg et al. [2000] developed a Kp forecast model
that is also based on an NN. Because we do not have access
Figure 2. The performance of the Costello NN model, to this model, we cannot evaluate this model using the same
which predicts Kp 1 hour ahead. NOAA routinely provides procedure and test data set used for the other models
this model’s predictions. Official (Postdam) Kp is plotted on evaluated in this study. Instead, we rely on their own
the x-axis and the model prediction is plotted on the y-axis. evaluation of their model as presented in the work of
Perfect predictions would lie on the line with a slope of one. Boberg et al. [2000]. Fortunately, they also present a
The error bar indicates one standard deviation. predicted Kp versus official Kp plot (see Figure 4 in the
work of Boberg et al. [2000]), which can be compared with
similar plots in this paper, for example, Figures 2 and 3 with
ello NN, (2) NARMAX, (3) Boberg et al. NN. All these the above-mentioned caveat. They report correlation coef-
models predict Kp roughly 1 hour ahead using a solar wind ficient r = 0.77, which is roughly comparable to those
monitor located at L1. In order to subject the models to the obtained by the NARMAX and Costello models. Like these
same evaluation procedure and test data set, the Costello two models, their result indicates that their model does not
NN and NARMAX models were ported to our computer. predict active times as well as quiet times. Their model
We have no access to the Boberg et al. model and hence the inputs solar wind V, n, and IMF Bz.
comparison with this model is more limited.
4.1. Costello NN Kp Model 5. APL Kp Models
[19] Presently, NOAA routinely provides real-time short- [22] We developed three different Kp models: (1) a model
term Kp forecasts from the most popular Kp prediction that inputs nowcast Kp and solar wind parameters and
model, the Costello NN Kp model [Costello, 1997] (http:// predicts Kp 1 hour ahead; (2) a model with the same input
www.sec.noaa.gov/rpc/costello/index.html). This model
inputs ACE solar wind speed and IMF Bz and jBj and
predicts Kp roughly 1 hour ahead every 15 min. The actual
prediction lead time may vary, depending on the solar wind
V. In order to facilitate careful and systematic comparisons
with other models, this model was ported to our computer at
JHU/APL. Figure 2 shows the results of running the model
on the test data set shown in Figure 1. Figure 2 shows that
the model predictions are fairly accurate during quiet times
(Kp < 5) but are less accurate during moderate or active
times (Kp > 5), a typical problem for many models [Joselyn,
1995] and a feature that likely results from the low number
of high Kp events captured in the training data set [Costello,
1997]. This behavior is also consistent with the previous
evaluation of Costello model [Detman and Joselyn, 1999].
The correlation coefficient (r) between the forecast and
official Kp is 0.75.
4.2. NARMAX Kp Model
[20] We also ported a NARMAX Kp model that was
developed at the University of Sheffield, UK. NARMAX
is an acronym of Nonlinear Auto Regressive Moving Figure 3. NARMAX 1 hour ahead Kp performance
Average Models With Exogenous Inputs and is described shown in the same format as Figure 2.

4 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

estimates have improved significantly, having correlation

coefficient between nowcast and official Kps of r > 0.9
[Takahashi et al., 2001]. Therefore it is expected that the
difference between sources 1 and 2 would be small. Figure 4
presents an example of the performance of Takahashi et al.
[2001] Kp estimates for days 70– 120 in 1989 (from Figure 13
of Takahashi et al. [2001]). The Kp estimates were calculated
from only three magnetometer stations: Fredericksburg,
Newport, and Sitka. The accuracy would improve with
additional magnetometers. The Takahashi et al. [2001] algo-
rithm can produce a nowcast Kp every 3 hours or at higher
time resolution, e.g., every 15 min.
[24] In order to show how well the model works, Figure 5g
shows the model prediction over a randomly selected 30 day
interval in the test data set. Figures 5a – 5f display the time
shifted solar wind and IMF during this 30 day interval.
[25] The statistical model evaluation on the test data set
(from the official Kp record), as displayed in Figure 6a,
shows that model 1 is a significant improvement over
previous models. This can be seen by comparing Figure 6a
with Figures 2 and 3 of the present paper and with Figure 4
of Boberg et al. [2000]. Figure 6a shows that the model
Figure 4. Performance of Takahashi et al. [2001] nowcast predicts Kp well not only during quiet times but also during
Kp estimate algorithm for days 70– 120 in 1989, a 50-day active times. The correlation coefficient r = 0.92 is far
period that includes an extremely strong magnetic storm. higher than those from the other models.
The x-axis displays the official (Postdam) Kp. The [26] Model 1 has also been evaluated using Takahashi et
magnetometer data for Kp estimates come from only al. [2001] nowcast Kp as an input parameter, instead of the
three stations: Fredericksburg, Newport, and Sitka (from historical official Kp. For this evaluation, we processed
Figure 13 of Takahashi et al. [2001]). magnetometer data from three stations (Fredericksburg,
Newport, and Sitka) for the year 1998 to produce Kp
as model 1 and predicts Kp 4 hours ahead; and (3) a model estimates at 15 min resolution. Figure 6b shows the results
that inputs only solar wind parameters and predicts Kp of the evaluation with these Kp estimates, which are very
1 hour ahead. similar to Figure 6a. The scatter in Figure 6b is perhaps
slightly larger for high Kps, which may be attributed partly
5.1. Kp Model That Inputs Solar Wind and Nowcast to fewer data points and partly to inaccuracies in the
Kp and Predicts Kp 1 Hour Ahead (APL Model 1) nowcast Kps. Overall, Figures 6a and 6b suggest that
[23] We developed a model (hereinafter referred to as running the model in real-time with Takahashi et al.
APL model 1) that inputs solar wind jVxj, n, IMF jBj, Bz, [2001] nowcast Kp as an input parameter would result in
and nowcast Kp. Using solar wind measurements at L1, performance that does not differ appreciably from that
APL model 1 predicts Kp up to approximately 1 hour ahead obtained with 15 min interpolated official Kp.
(the actual prediction lead time may vary, depending on the [27] The superior performance of APL model 1 over the
solar wind V). The nowcast Kp algorithm [e.g., Takahashi previous models could perhaps be attributed to several
et al., 2001] at time t can estimate Kp for (t
1.5 hours). factors (the square brackets indicate the particular model(s)
For simplicity in building the model, the model inputs a Kp that lack the corresponding attribute): (1) the model was
estimate for time (t
1.25
1.5 hours) or (t
2.75 hours), trained with much larger data set spanning over two solar
where t is the solar wind/IMF arrival time on Earth. This cycles [Costello, NARMAX]; (2) the model inputs not only
assumes that most of the times the solar wind monitor at L1 solar wind V, but also n (hence dynamic pressure) [Costello];
would provide solar wind measurements <1.25 hours in and (3) the model inputs nowcast Kp, which is highly
advance. Since the nowcast Kp algorithm has only recently correlated with near-future Kp (see section 6) [Costello,
become available, for model development, the input Kp Boberg et al.].
comes from the official Kp record that spans over two solar [28] For operational consideration, the Takahashi et al.
cycles. However, for model evaluation, the input Kp comes [2001] model will soon be operational at JHU/APL and its
from two sources: (1) the official Kp record and (2) the Kp nowcast Kps will be made publicly accessible through the
estimates from Takahashi et al. [2001] algorithm. Of course, Web site given above. Alternatively, APL model 1 can use
the official Kps are used to determine the accuracy of the the nowcast Kps provided at the NOAA Web site given
forecast Kps. The advantage of source 1 is that it covers a above. The real-time ACE measurements are already
large time interval, 1975 –2001, but the disadvantage is that routinely publicly accessible at the same NOAA Web site.
the evaluation gives only a sense of the upper bound
performance. The advantage of source 2 is that it gives a 5.2. Kp Model That Inputs Solar Wind and Nowcast
realistic evaluation for real-time operations, but since the Kp and Predicts Kp 4 Hours Ahead (APL Model 2)
nowcast Kp algorithm was only recently developed, it has a [29] The second model (hereinafter referred to as APL
much more limited time coverage. However, the nowcast Kp model 2) inputs the same parameters as APL model 1 but

5 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

Figure 5. Example of the APL models Kp forecasts on a randomly selected 30 day period in the test
data set. (a) – (f) IMF Bx, By, Bz, solar wind density, velocity, and dynamic pressure, respectively. (g) – (i)
official Kp (black) and forecast Kp (red [or gray in the black and white version]) for APL models 1, 2,
and 3, respectively. The solar wind and IMF have been hourly averaged and time shifted.

predicts Kp 4 hours ahead based on solar wind measure- imately 1 hour ahead, but the model predicts Kp 4 hours
ments at L1. Thus unlike APL model 1, this model predicts ahead. Figure 5h shows the model Kp forecasts over a
Kp without knowing the entire history of the solar wind 30-day interval, which was randomly selected. Figure 7a
parameter leading up to the forecast time. Specifically, a presents the statistical model performance on the test data
satellite at L1 provides solar wind parameters only approx- set (with input Kp obtained from the official Kp record).

6 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

good as those obtained using interpolated official Kp (test

data set).
5.3. Kp Model That Inputs Only Solar Wind and
Predicts Kp 1 Hour Ahead (APL Model 3)
[ 31 ] There may be times (locations) when (where)
nowcast Kps are simply not available or reliable. For this
reason, we developed a model (hereinafter referred to as
APL model 3) that is driven solely by solar wind and IMF.
This model inputs solar wind n and jVxj and IMF jBj and
Bz. Using solar wind measurements at L1, e.g., ACE, APL
model 3 outputs 1 hour ahead Kp forecast. Figure 5i shows

Figure 6. APL NN 1 hour ahead Kp (APL model 1)

performance shown in the same format as in Figure 2. This
model inputs nowcast Kp from (a) the test data set from the
official Kp record 1975 –2001 and (b) the Takahashi et al.
[2001] Kp estimates for 1998. The model is significantly
better than Costello NN (Figure 2) and NARMAX
(Figure 3).

Figure 7b shows the evaluation using the Takahashi et al.

[2001] Kp estimate as an input parameter for the year 1998.
As expected, the longer prediction lead time results in the
reduction in performance. The reduction in performance is
manifested in the form of larger scatter compared with that of
APL model 1. However, as shown in Figure 7, despite
forecasting further ahead in time, the 4 hours ahead Kp
forecasts for active times (Kp > 5) are more accurate than the Figure 7. APL Kp NN 4 hours ahead (APL model 2) Kp
1 hour ahead forecasts from the Costello, Boberg et al., and performance shown in the same format as in Figure 2.
NARMAX models. The correlation coefficient, r = 0.79, is (a) The input Kp comes from the test data set (official Kp
slightly higher than that of the other 1 hour ahead Kp record), and (b) the input Kp comes from the Takahashi et
forecast models (Costello, NARMAX, and Boberg et al.). al. [2001] Kp estimate for 1998. Note that the performance
[30] Consistent with the APL model 1 evaluation, the for active times, Kp > 5, is better than that of the Costello
APL model 2 evaluation also demonstrates that running the and NARMAX models, even though its forecast range is
model in real time with the Takahashi et al. [2001] nowcast 4 hours ahead, whereas those two models predict only
Kp as an input parameter produces results that are almost as 1 hour ahead.

7 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

bottom of Figures 9c – 9e. In the present paper, in order to

improve the visualization, the sunspot numbers are not
normalized to one. Examination of the skill scores and the
sunspot numbers in Figures 9c – 9e reveals that the Costello
model predicts Kp more accurately near solar maximum
than minimum.
[33] It is not clear from Figure 9 whether this solar cycle
variation is physically significant or whether it results from
having a limited training set. Therefore we calculated the
same skill scores for different models using a different and
larger solar wind data set, namely IMP-8 (1975 – 1994),
Wind (1994 – 2000), and ACE (2000 – 2001) solar wind
data. An example is shown in Figure 10, which shows the
skill scores for APL model 3 in a format similar to Figure 9.
This model, which was trained with a much larger data set
spanning over two solar cycles, inputs only solar wind/IMF
as in Costello model. Training with larger data sets helps
improve the accuracy of the predictions and significantly
reduces the solar cycle variation, but it does not appear to
Figure 8. APL Kp NN 1 hour ahead (APL model 3) Kp completely eliminate the solar cycle effect. In Figures 10a–
performance shown in the same format as in Figure 2. This 10d, the skill scores appear to exhibit mild solar cycle
model inputs only solar wind and IMF and no nowcast Kp. variations, but the maxima and minima of the sunspot
This can be directly compared with the Costello NN model number do not always appear at the same time as those of
results shown in Figure 2, since both models input only the skill scores. The maxima of the skill scores often appear
solar wind and IMF parameters and no nowcast Kp. during the declining phase of the solar maximum. Figures 9
and 10 would suggest that at least a component of the solar
cycle variation might result from a more fundamental way
the model outputs over a 30-day period. Figure 8 presents in which solar wind interacts with the geospace.
the model performance using the same test data set used [34] Recently, we applied a cumulant-based method to
throughout this study, except for the Boberg et al. model. It analyze the statistical informational dynamics of Kp time
has a correlation coefficient of r = 0.84. Figure 8 shows that series and found that there is a stronger nonlinear relation-
without the benefit of the nowcast Kp, the accuracy is lower. ship between past Kp and future Kp around solar minimum
Since this model inputs only solar wind and IMF parameters than around maximum [Johnson and Wing, 2004, 2005].
and no nowcast Kp, it can be most directly compared with This relationship is depicted in Figure 11, which plots the
the Costello and Boberg et al. models, which do not input normalized linear (SL) and nonlinear (SNL) significance as
nowcast Kp either. The comparisons show that despite a function of time delay, t, in hours for Kp data for two
having the same kind of inputs, APL model 3 still outper- solar minima (Figures 11a and 11c) and two solar maxima
forms these two models, albeit the performance improve- (Figures 11b and 11d). For time T, SL is roughly proportional
ment is less than for APL model 1. APL model 3 also to the linear correlation between Kp(T) and Kp(T
t),
outperforms the NARMAX model, despite the latter having whereas SNL gives a measure of the nonlinear correlations.
the advantage of nowcast Kp as an input parameter. The details of the calculations are given by Johnson and
Wing [2004, 2005]. The figure shows that while the linear
6. Kp (Magnetospheric) Predictability as a response is roughly the same around solar minimum and
maximum, the nonlinear response is stronger for solar
Function of Solar Cycle minimum than maximum, with a peak around t  40 hours.
[32] Papitashvili et al. [2000] show that there is a solar [35] It turns out that the nonlinear response (SNL) anti-
cycle variation in the average Kp. The accuracies of correlates with sunspot number in every solar cycle since the
nowcast Kps calculated from ground magnetometer data Kp index has been recorded, 1932 to the present (2003).
exhibit no discernable correlation with sunspot numbers Figure 12 shows clearly this anticorrelation in two
[Takahashi et al., 2001]. It would be interesting to deter- different ways: Figure 12a shows the maximum SNL, and
mine whether the accuracies of the Kp (a proxy for the state Figures 12b– 12c show the integrated SNL for >95% and
of the magnetosphere) forecast based partly or entirely on >99.5% confidence, respectively, for years 1932 – 2003.
solar wind/IMF have solar cycle dependence. In order to Figure 12 was obtained from the yearly SNL curves such
do this, we calculated the skill scores for the Costello as the ones shown in Figure 11. For easy reference, the
model predictions over two solar cycles, 1975 – 1999, using normalized sunspot number is plotted as lighter (gray) lines at
IMP-8 solar wind data. NOAA issues alerts when Kp is the bottom of each panel. In Figure 12a, the probability that
expected to exceed certain threshold, e.g., 4, 5, 6, etc. a linear model could produce significance of 10 is 1:1023. The
Figure 9 plots True Skill Statistics (TSS) and Gilbert Score integrated significance sums over all correlation times up to
(GS) for these Kp thresholds [Detman and Joselyn, 1999]. t = 200 hours and may give a better indicator of the presence
In both of these scoring systems, a random prediction = 0 of nonlinearity over long times than just the maximum
and a perfect prediction = 1. The normalized monthly value. It is only appropriate to compare the relative value
average International Sunspot Number is plotted at the of the integrated significance within each plot. The typical

8 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

Figure 9. The Costello model predicts Kp more accurately near solar maximum than near solar
minimum for Kp > 3. (a) – (e) Skill scores for the Costello NN Kp model over two solar cycles for Kp =
2 – 6, respectively. The black solid line = True Skill Statistic (TSS) and the dashed line = Gilbert score
(GS). The lighter (gray) solid lines at the bottom of Figures 9c – 9e indicate the normalized monthly
average International Sunspot Number, which exhibits trends similar to the skill scores.

correlation time for high nonlinear significance events lies sphere is dominated more by internal dynamics than it is
around 40 hours (see also Johnson and Wing [2004, 2005]). around solar maximum, when it is more directly driven by
The timescale over which the nonlinear significance is the external inputs, namely solar wind and IMF. The
appreciable is on the order of a week. These timescales are presence of the internal dynamics would complicate the
similar to timescales associated with ring current Kp predictions that rely entirely or partly on the external
relaxation (hours to days) and storm relaxation (1 week). drivers.
[36] Taken together, these nonlinearities and the skill [37] The statistical information dynamical analysis of Kp
scores suggest that around solar minimum, the magneto- shown in Figure 11 indicates that Kp highly correlates with

9 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

Figure 10. Skill scores over two solar cycles for a Kp model that inputs solar wind parameters and
predicts Kp 1 hour ahead (APL model 3). (a)– (e) Skill scores for APL model 3 for Kp = 2 – 6,
respectively. The black solid line = True Skill Statistic (TSS) and the dashed line = Gilbert score (GS).
The lighter (gray) solid lines at the bottom of Figures 10a– 10d plot the normalized monthly average
International Sunspot Number. In Figures 10a– 10d, while the skill scores are higher than those for the
Costello model, they still exhibit yearly variations that roughly follow the same general pattern as the
sunspot number.

the past Kps within the last few hours, linearly and non- [38] Near solar minimum, as shown in Figure 11, there
linearly (the Kp has been interpolated to 15 min resolution). are significant nonlinear correlations between Kp and past
Therefore including the nowcast Kp can improve the 1 hour Kps that may be attributed to internal dynamics. Thus in
ahead Kp forecasting. However, both the linear and nonlin- theory the nowcast and all past Kps may help improve the
ear correlations drop off sharply with increasing t, which Kp forecasts, since they contain some information that
contribute to the rapid decrease in the performances of APL pertains to the internal dynamics, including persistence.
model 2 compared with APL model 1. This information is not available in the solar wind/IMF.

10 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

Figure 11. Normalized linear (SL) and nonlinear (SNL) significance as a function of time delay t, for
Kp data from near (a) solar minimum 1975, (b) solar maximum 1982, (c) solar minimum 1987, and
(d) solar maximum 1992. SL is proportional to the linear correlation coefficient between Kp(T) and
Kp(T
t), whereas SNL gives a measure of the nonlinear correlation. SL = solid line and SNL = dashed
line. SL drops off quickly with increasing t for both solar maxima and minima. Near solar minima, there
is a strong nonlinear response with a peak around t = 40 hours, whereas near solar maxima, SNL tapers
off quickly.

The recurrent NN has some capacity to memorize the past on the solar wind V and the location of the solar wind
Kps (for APL models 1 and 2), but it is doubtful that it can monitor. All these models are based on NNs and were
effectively capture the complex correlations or dynamics developed with 27 years of data, 1975– 2001.
beyond the last few hours (the nonlinear correlations can [40] Our extensive evaluation based on data spanning
extend beyond several hours). more than two solar cycles shows that (1) our models give
significantly more accurate predictions than previous
models, with the most dramatic improvements occurring
7. Discussion and Summary during moderate and active times, Kp > 4, and (2) Kp is
[39] In order to satisfy different needs and operational slightly more predictable near solar maximum than it is near
constraints, we developed three types of Kp forecast solar minimum. Information dynamics analysis of Kp
models, two of which require nowcast Kp as an input suggests that the magnetosphere is more externally driven
parameter. The focus of the model development was near solar maximum (or the declining phase of the solar
improving the forecasts for magnetically moderate and maximum) than near solar minimum. Around solar mini-
disturbed times, which are notoriously difficult to predict mum, the internal dynamics such as loading and unloading
[e.g., Joselyn, 1995]. The three models are (1) APL model 1 of the energy in the magnetotail, ring current decays, storm
that inputs nowcast Kp and solar wind parameters and relaxation, etc., may play a more important role in the
predicts Kp 1 hour ahead; (2) APL model 2 that has the magnetospheric dynamics and may introduce some diffi-
same input as APL model 1 and predicts Kp 4 hours ahead; culties to Kp models that rely entirely or partly on solar
and (3) APL model 3 that inputs only solar wind parameters wind parameters.
and predicts Kp 1 hour ahead. The 1 and 4 hour prediction [41] The performances of various models can be summa-
lead times are just rough estimates for a solar wind monitor rized in Figure 13, which plots the TSS scores of the Kp
at L1. The actual prediction lead times may vary, depending threshold [Detman and Joselyn, 1999]. Figure 14 summa-

11 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

Figure 12. The nonlinear response anticorrelates with the sunspot number in every solar cycle from the
first Kp record to present, 1932 – 2003. (a) The black line shows the maximum nonlinear significance
(SNL); (b) and (c) the black line shows the integrated SNL at >95% confidence and at >99.5%
confidence, respectively. The lighter (gray) lines at the bottom in Figures 12a– 12c show the normalized
sunspot number.

rizes the correlation coefficients for all the models, which is

fairly consistent with the skill scores. For comparisons
with other published model evaluations, the correlation
coefficient r is calculated here for all Kp ranges. Therefore
Figure 14 understates the dramatic improvements in the
active times (Kp > 5) forecast range achieved by APL Kp
models. We also note that r, by itself, does not indicate how
well the model performs. For example, it would be possible
to have a model that systematically and consistently under-
predicts Kp but does so with little scatter and high r. Thus
comprehensive model evaluations should include scatter-
plots (e.g., Figures 2, 3, 6, 7, 8), skill scores, and r.
[42] The advantage of having nowcast Kp can be seen, at
least qualitatively, by comparing model 1 and model 3, e.g.,
comparing Figures 6 and 8. The advantage of larger data
sets plus inclusion of solar wind density can be seen by
comparing APL model 3 with the Costello model, e.g.,
Figures 2 versus 8. However, the quantifications of these
effects would require substantially more work and should be
addressed in the future work.
[43] For operational consideration, nowcast Kps, based Figure 13. APL Kp models outperform the existing Kp
on an Air Force algorithm, are now routinely made publicly forecast models. True Skill Statistic (TSS) as a function of
available by NOAA. Furthermore, Takahashi et al. [2001] Kp threshold for APL model 1 (red), APL model 2 (blue),
have developed a very accurate nowcast Kp algorithm. APL model 3 (purple), the NARMAX model (green), the
These nowcast Kps will soon be publicly available at a Costello model (cyan).

12 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

JHU/APL Web site once the live magnetometer data streams

become available. Our study demonstrates that these now-
cast Kps can be used to improve the forecast Kps. Thus
APL models 1 and 2 can be used whenever nowcasts Kps
are available, while APL model 3 can be used when only
solar wind/IMF measurements are available.
[44] For practical and operational considerations, the APL
models predict a new Kp every 15 min instead of every
3 hours. The higher time resolution was chosen to enable
the models to warn users of the impending change in space
weather in a more timely manner. For example, using 15 min
time granularity, if ACE detected a sudden change in the
solar wind that would cause a huge increase in geomagnetic
activity, the model would be able to warn the users in the
next 15 min, before the actual arrival of the solar wind,
whereas using the traditional Kp 3 hour time granularity, the
model would not be able to do so. All model evaluations
presented so far are done by comparing the model output
Kp at 15 min granularity with the 15 min interpolated
historical official Kp. This is deemed more accurate than
the comparisons with the historical 3 hour Kp because the
magnetospheric state changes in a continuous manner rather
than in 3 hour step function. However, we have also
statistically evaluated the model output Kp against the
traditional 3 hour official Kp in a similar manner as in
Figures 6 – 8, 13, and 14. The results do not show appre-
ciable differences with those evaluations against the inter-
polated 15 min Kp presented above. Figure 15 shows two
examples: for APL model 2 (Figure 15a) and for APL
model 3 (Figure 15b). Thus for practical purpose, the
model output 15 min Kp could be treated as the traditional
3 hour Kp. Evaluated in this manner, APL Kp models still
outperform the existing Kp models.
[45] The APL Kp models can provide a short-term
forecast, up to a few hours ahead using solar wind/IMF
Figure 15. Evaluations with official 3 hour Kp in the same
format as Figures 7 and 8 for (a) APL model 2 and (b) APL
model 3. The evaluations with official 3 hour Kp and with
15 min interpolated Kp are very similar, as comparisons of
the above figures with Figures 7 and 8 reveal.

measurements at L1. These short-term forecasts would be

valuable for space weather users, but for many applications
longer-term forecasts are more desirable. We would like to
note that there are several heliospheric models that are
currently operational and/or under various stages of devel-
opment. These models input solar observations and predict
solar wind/IMF days in advance, e.g., the Hakamada-
Akasofu-Fry (HAF) code [Fry et al., 2001; Hakamada
and Akasofu, 1982] and the Wang-Sheeley model [Wang
and Sheely, 1995; Arge and Pizzo, 2000]. Also, the NASA
Living With a Star (LWS) roadmap calls for the launch of a
multispacecraft space weather network (Solar Dynamics
Observatory and Sentinels) with a main goal of providing
the solar wind parameters with longer lead time (http://
lws.gsfc.nasa.gov). Thus the APL Kp model 3, which is
driven solely by solar wind/IMF, can be linked with these
models/products to provide longer-term forecasts.
Figure 14. Correlation coefficients for the three APL [46] Acknowledgments. We thank NASA CDAWeb for supplying
models, and the Costello, NARMAX, and Boberg et al. IMP-8, Wind, and ACE Key Parameter data and NOAA for providing ACE
models. data. We also thank MIT for supplying IMP-8 plasma data and GFZ

13 of 14
A04203 WING ET AL.: Kp FORECAST MODELS A04203

Postdam for supplying Kp records. S. Wing was supported by NSF Space Lo, J., and D. Bassu (1999), Mathematical justification of recurrent multi-
Weather grant ATM-9819705, NASA grant NAG5-10971 and DOD con- layer perceptrons with long- and short-term memories, in Proceedings of
tract N00024-03-D-6606. J. R. Johnson was supported by NSF grant ATM- International Joint Conference on Neural Networks 1999, vol. 1, pp.
0218847 and NASA grant W-19880. We thank J. Vandegriff for assisting 364 – 369, IEEE Press, Piscataway, N. J.
with the coding of the recurrent network. Mauk, B. H., and C. E. McIlwain (1974), Correlation of Kp with the sub-
[47] Arthur Richmond thanks Harry C. Koons and another reviewer for storm-injected plasma boundary, J. Geophys. Res., 79, 3193 – 3196.
their assistance in evaluating this paper. Newell, P. T., S. Wing, C. I. Meng, and V. Sigillito (1990), A neural
network based system for monitoring the aurora, Johns Hopkins APL
References Tech. Dig., 11(3, 4), 291 – 299.
Newell, P. T., S. Wing, C. I. Meng, and V. Sigillito (1991), The auroral oval
Arge, C. N., and V. J. Pizzo (2000), Improvement in the prediction of solar position, structure and intensity of precipitation from 1984 onwards: An
wind conditions using near-real time solar magnetic field updates, automated on-line data base, J. Geophys. Res., 96, 5877 – 5882.
J. Geophys. Res., 105, 10,465 – 10,479. Newell, P. T., V. A. Sergeev, G. R. Bikkuzina, and S. Wing (1998),
Balikhin, M. A., O. M. Boaghe, S. A. Billings, and H. St C. K. Alleyne Characterizing the state of the magnetosphere: Testing the ion precipita-
(2001), Terrestrial magnetosphere as a nonlinear resonator, Geophys. Res. tion maxima latitude (b2i) and the ion isotropy boundary, J. Geophys.
Lett., 28, 1123 – 1126. Res., 103, 4739 – 4745.
Bartels, J. (1949), The standardized index, Ks, and the planetary index, Kp, Newell, P. T., T. Sotirelis, J. F. Carbary, K. Liou, J. P. Skura, C.-I. Meng,
IATME Bull., 12b, 97. C. Deehr, D. Wilkinson, and F. J. Rich (2002), OVATION: Oval
Boaghe, O. M., M. A. Balikhin, S. A. Billings, and H. Alleyne (2001), Variation, Assessment, Tracking, Intensity, and Online Nowcasting,
Identification of nonlinear processes in the magnetospheric dynamics and Ann. Geophys., 20, 1039 – 1047.
forecasting of Dst index, J. Geophys. Lett., 106, 30,047 – 30,066. Papitashvili, V. O., N. E. Papitashvili, and J. H. King (2000), Solar cycle
Boberg, F., P. Wintoft, and H. Lundstedt (2000), Real time Kp prediction effects in planetary geomagnetic activity: Analysis of 36-year long OMNI
from solar wind data using neural networks, Phys. Chem. Earth, 25, dataset, Geophys. Res. Lett., 27, 2797 – 2800.
275 – 280. Rostoker, G. (1972), Geomagnetic indices, Rev. Geophys., 10, 935 – 950.
Costello, K. A. (1997), Moving the Rice MSFM into a real-time forecast Rumelhart, D. E., and J. L. McClelland (Eds.) (1987), Parallel Distributed
mode using solar wind driven forecast models, Ph.D. dissertation, Rice Processing, vol. I, MIT Press, Cambridge, Mass.
Univ., Houston, Texas. Sergeev, V. A., M. Malkov, and K. Mursula (1993), Testing the isotropic
Crooker, N. U., and K. I. Gringauz (1993), On the low correlation between boundary algorithm method to evaluate the magnetic field configuration
long-term averages of solar wind speed and geomagnetic activity after in the tail, J. Geophys. Res., 98, 7609 – 7620.
1976, J. Geophys. Res., 98, 59. Takahashi, K., B. A. Toth, and J. V. Olson (2001), An automated proce-
Detman, T., and J. A. Joselyn (1999), Real-time Kp predictions from ACE dure for near-real-time Kp estimates, J. Geophys. Res., 106, 21,017 –
real time solar wind, in Solar Wind Nine, edited by S. R. Habbal et al., 21,032.
AIP Conf. Proc., 471, 729 – 732. Tsyganenko, N. A. (1989), A magnetospheric magnetic field model with a
Fernandez, B., A. G. Parlos, and W. K. Tsai (1990), Nonlinear dynamic warped tail current sheet, Planet. Space Sci., 37, 5 – 20.
system identification using artificial neural networks (ANNs), in Pro- Vandegriff, J., K. Wagstaff, G. Ho, and J. Plauger (2005), Forecasting space
ceedings of International Joint Conference on Neural Networks 1990, weather: Predicting energetic storm particle events using neural networks,
edited by M. Caudill, pp. 133 – 141, IEEE Press, Piscataway, N. J. Adv. Space Sci., in press.
Fok, M.-C., T. E. Moore, and W. N. Spjeldvik (2001), Rapid enhancement Wang, Y.-M., and N. R. Sheely (1995), Solar implications of ULYSSES
of radiation belt electron fluxes due to substorm dipolarization of the interplanetary field measurements, Astrophys. J., 447, L143 – L146.
geomagnetic field, J. Geophys. Res., 106, 3873 – 3881. Wing, S., and P. Newell (2003), LLBL contribution to the plasma sheet
Fry, C. D., W. Sun, C. S. Deehr, M. Dryer, Z. Smith, S.-I. Akasofu, ions, in Earth’s Low-Latitude Boundary Layer, Geophys. Monogr. Ser.,
M. Tokumaru, and M. Kojima (2001), Improvements to the HAF solar vol. 133, edited by P. T. Newell and T. Onsager, pp. 273 – 282, AGU,
wind model for space weather predictions, J. Geophys. Res., 106, Washington D. C.
20,985 – 21,001. Wing, S., P. T. Newell, D. G. Sibeck, and K. B. Baker (1995), Large
Garrett, H. B., A. J. Dessler, and T. W. Hill (1974), Influence of solar wind statistical study of the entry of interplanetary magnetic field Y-component
variability on geomagnetic activity, J. Geophys. Res., 79, 4603 – 4610. into the magnetosphere, Geophys. Res. Lett., 22, 2083 – 2086.
Gehred, P. A., W. Cliffswallow, and J. D. Schroeder III (1995), A compar- Wing, S., R. A. Greenwald, C.-I. Meng, V. G. Sigillito, and L. V. Hutton
ison of USAF Ap and Kp indices to Gottingen indices, Tech. Memo. ERL (2003), Neural Networks for automated classification of ionospheric
SEL-88, NOAA, Silver Spring, Md. irregularities from HF radar backscattered signals, Radio Sci., 38(4),
Gershenfeld, N. (1999), The Nature of Mathematical Modeling, pp. 150 – 1063, doi:10.1029/2003RS002869.
153, Cambridge Univ. Press, New York. Wu, J., and H. Lundstedt (1997), Geomagnetic storm predictions from solar
Hakamada, K., and S.-I. Akasofu (1982), Simulation of three-dimensional wind data with the use of dynamic neural networks, J. Geophys. Res.,
solar wind disturbances and resulting geomagnetic storms, Space Sci. 102, 14,255 – 14,268.
Rev., 31, 3 – 70.
Hardy, D. A., M. S. Gussenhoven, R. Raistrick, and W. J. McNeil (1987),
Statistical and functional representations of the pattern of auroral energy
flux, number flux, and conductivity, J. Geophys. Res., 92, 12,275 – 12,294.
Hedin, A. E. (1987), MSIS-86 thermospheric model, J. Geophys. Res., 92,






















4649. M. Balikhin, Department of Automatic Control and System Engineering,
Johnson, J. R., and S. Wing (2004), A cumulant-based analysis of nonlinear University of Sheffield, Mappin Street, Sheffield S1 3JD, UK.
magnetospheric dynamics, Publ. PPPL-3919, Princeton Plasma Physics K. Bechtold, J. Jen, C.-I. Meng, K. Takahashi, and S. Wing, Johns
Laboratory, Princeton, N. J. (Available at http://www.pppl.gov/pub_ Hopkins University Applied Physics Laboratory, 11100 Johns Hopkins
report//2004/PPPL-3919.pdf) Road, Laurel, MD 20723-6099, USA. (simon.wing@jhuapl.edu)
Johnson, J. R., and S. Wing (2005), A solar cycle dependence of nonlinear- K. Costello, NASA Lyndon B. Johnson Space Flight Center, 2101 NASA
ity in magnetospheric activity, J. Geophys. Res., doi:10.1029/ Parkway, Houston, TX 77058, USA.
2004JA010638, in press. J. Freeman, Physics and Astronomy Department, Rice University, P. O.
Joselyn, J. (1995), Geomagnetic activity forecasting: The state of the art, Box 1892, Houston, TX 77005, USA.
Rev. Geophys., 33, 383 – 401. J. R. Johnson, Princeton Plasma Physics Laboratory, Princeton
Koons, H. C., and D. J. Gorney (1991), A neural network model of the University, P. O. Box 451, MS 28, Princeton, NJ 08543-0000, USA.
relativistic electron flux at geosynchronous orbit, J. Geophys. Res., 96, D. G. Sibeck, NASA Goddard Space Flight Center, LEP Code 696, 8800
5549 – 5556. Greenbelt Road, Greenbelt, MD 20771, USA.

14 of 14

View publication stats