Corresponding author: G. Millward, Cooperative Institute for Research in Environmental Sciences (CIRES), University of Colorado, Boulder, CO 80305, USA. (firstname.lastname@example.org)
 Coronal mass ejections (CMEs)—massive explosions of dense plasma that originate in the lower solar atmosphere and propagate outward into the solar wind—are the leading cause of significant space weather effects within Earth's environment. Computational models of the heliosphere such as WSA-Enlil offer the possibility of predicting whether a given CME will become geo-effective and, if so, the likely time of arrival at Earth. To be meaningful, such a forecast model is dependent upon accurately characterizing key parameters for the CME, notably its speed and direction of propagation, and its angular width. Studies by Zhao et al. (2002) and Xie et al. (2004) suggest that these key CME parameters can be deduced from geometric analysis of the elliptical “halo” forms observed in coronagraph images on spacecraft such as the Solar and Heliospheric Observatory (SOHO) and which result from a CME whose propagation is roughly toward or away from the observer. Both studies assume that the CME presents a circular cross-section and maintains a constant angular width during its radial expansion, the so called “cone model.” Development work at the NOAA Space Weather Prediction Center (SWPC) has been concerned with building and testing software tools to allow forecasters to determine these CME parameters routinely within an operational context, a key aspect of transitioning the WSA-Enlil heliospheric model into operations at the National Weather Service. We find “single viewpoint” cone analysis, while a useful start, to be highly problematic in many real-world situations. In particular, it is extremely difficult to establish objectively the correct ellipse that should be applied to a given halo form and that small changes in the exact ellipse chosen can lead to large differences in the deduced CME parameters. The inaccuracies in the technique are particularly evident for analysis of the “nearly circular” elliptical forms which result from CMEs that are propagating directly toward the observer and are therefore the most likely to be geo-effective. In working to resolve this issue we have developed a new three-dimensional (3-D) graphics-based analysis system which seeks to reduce inaccuracies by analyzing a CME using coronagraph images taken concurrently by SOHO and also by the two Solar TErrestrial RElations Observatory (STEREO) spacecraft, which provide additional viewing locations well away from the Sun-Earth line. The resulting “three view” technique has led to the development of the CME Analysis Tool (CAT), an operational software system in routine use at the SWPC as the primary means to determine CME parameters for input into the WSA-Enlil model. Results from the operational WSA-Enlil system are presented: utilizing CAT to provide CME input parameters, we show that, during the first year of operations at SWPC, the WSA-Enlil model has forecasted the arrival of CMEs at Earth with an average error 7.5 h.
 Space weather is known to have societal and economic impacts across a broad range of industry sectors [e.g., National Research Council, 2008]. From the power grid, to airlines, to global positioning, and uses including surveying, drilling, and farming, space weather now touches the lives of everyone in this age of technological reliance. However, it is recognized that while forecasting space weather has been an ongoing concern for decades, it lags far behind the capabilities of terrestrial weather forecasting [Siscoe, 2006]. In the 1950s, terrestrial weather prediction launched the modern era of numerical weather prediction with computers. Fifty years later, space weather prediction finally entered the era of numerical weather prediction and in 2011 the NOAA Space Weather Prediction Center transitioned the first large scale, physics based model into operations (Pizzo et al., 2011).
 Geomagnetic storms are perturbations in the Earth's own magnetic field that are driven by solar wind flows impinging on the Earth's magnetosphere. These storms usually result from corotating high speed wind streams [Coleman et al., 1966] or transient flows associated with interplanetary coronal mass ejections [e.g., Zurbuchen and Richardson, 2006]. The largest geomagnetic storms are driven by coronal mass ejections (CMEs) [cf. Richardson and Cane, 2012] and CMEs are therefore of the greatest geospace forecasting concern. While the existence of CMEs had been known for a long time [Tousey, 1973; cf. Howard, 2006], it was not until 1993 with the publication of “The Solar Flare Myth” [Gosling, 1993] that the community gave them their proper due in the space weather enterprise.
 Coronal mass ejections erupt from the Sun with a frequency that varies with the solar sunspot cycle, although CMEs come not only from active region sources but also filament channel eruptions. Thus, they can occur at any time in the solar cycle. With speeds that are typically 300–500 km/s [Gopalswamy et al., 2004], they can erupt with plane-of-the-sky speeds as slow as a few tens of km/s or as fast as >3000 km/s. The first recognition that Earth-directed CMEs, referred to as “halo CMEs,” could be observed near the Sun was made by Howard et al. . Early attempts to predict the arrival of CMEs at Earth came with the observation of halo CMEs [Plunkett et al., 1998] observed with the Large Angle and Spectrometric COronagraph Experiment (LASCO) instruments on SOHO. It quickly became apparent that ballistic propagation, e.g., no interaction with the background solar wind, was not a useful prediction scheme [Gopalswamy et al., 2000]. Throughout the last solar cycle, empirical prediction algorithms for predicting the arrival of CMEs at Earth abounded, but statistically, none could do better than arrival errors of >±10 h [Gopalswamy et al., 2001].
 However, efforts to model the propagation of CMEs paralleled this work and by the end of the last solar cycle, it was clear that numerical prediction models outperformed the empirical prediction techniques, reducing the average absolute error in predicting the arrival time (for a select sample of CMEs) to ±6.1 h [Taktakishvili et al., 2009]. With this evidence in hand, the NOAA Space Weather Prediction Center embarked on the journey of transitioning the first large-scale, physics-based space weather prediction algorithm, Wang-Sheely-Arge Enlil (WSA-Enlil) [Odstrcil, 2003].
 Whether one uses an empirical or a numerical prediction scheme, forecasting entails accurate characterization of at least the gross properties of a CME and the ambient solar wind. First, one needs to make an assessment of where the CME is headed to know if there is a concern. Second, if a concern is noted, then the specifics of the size, speed, and direction are needed to determine when the CME will arrive. In the last solar cycle, it became standard for SWPC to assess CMEs on the basis of their morphology in white light coronagraphs. Full halo CMEs were taken as directed either toward or away from Earth, and, if headed toward Earth, would certainly result in some level of geomagnetic storming. Partial halo CMEs, if on the Earth-facing side of the Sun, would result in glancing blows, where the Earth would still experience geomagnetic storming, but it would, on average, be weaker and of shorter duration, than for an equivalent full halo. Finally, limb (and far backside) CMEs are events that would be directed well away from Earth and thus not a concern for geomagnetic storm forecasting.
 Initially, predictions of CME arrival at Earth used plane-of-sky speed, even though that was recognized as not being the most appropriate component. Attempts to correct for propagation direction did not improve predictions [Gopalswamy et al., 2001]. Additional techniques, such as “CME expansion speed,” also attempted to derive more appropriate radial velocities, but all failed to improve upon prediction of arrival time [cf. Forsyth et al., 2006; Siscoe and Schwenn, 2006].
 To provide the boundary conditions for numerical prediction techniques, the CME direction, width, and radial velocity are needed. In 2002, it was shown that the geometric and kinematic properties of Earth-directed CMEs could be reproduced using a simple “cone model” [Zhao et al., 2002]. This cone model and subsequent variations became the standard method of fitting CMEs to determine their properties and were assumed to be the method by which the NOAA Space Weather Prediction Center would initiate the WSA-Enlil model.
 However, this method alone does not provide the kind of consistent and accurate inputs needed for operational space weather forecasting, and the use of supplemental information, as described below, is the only way to obtain reliable estimates of CME speed, direction, and angular width.
 The concept of a cone model as a framework for analyzing halo CMEs was originally suggested by Zhao et al.  and then treated fully analytically by Xie et al. . In both papers, the authors consider a CME as a literal flat-faced cone having a circular cross-section and constant angular width, expanding uniformly in the radial direction. For a CME aimed directly toward an observer, this geometry yields an expanding circular halo. For the more general case, the direction of propagation is offset from the observer and thus the circular halo is seen as an ellipse in which the direction of the minor axis passes through the origin, the center of the Sun. Xie et al.  present analytical formulae in which the CME propagation direction (i.e., heliospheric latitude and longitude) and cone angle can be derived in terms of the major and minor axes of the ellipse, its distance from the origin, and the angle the minor axis makes with the horizontal. A second self-similar ellipse, from either a later or earlier time, can then be analyzed to provide CME velocity.
 We have used the analytical formulae of Xie et al.  to determine CME parameters by analyzing images from the LASCO C3 coronagraph onboard SOHO [Brueckner et al., 1995; Domingo et al., 1995]. Our results reveal, in certain cases, an extreme sensitivity to the form of ellipse chosen to bind the CME structure. The problem is particularly acute for situations in which the displacement of the CME centerline from the origin is small and the elliptical form tends to a circle. The authors note that their formulation is not applicable to the situation of a perfectly circular halo with zero displacement. Our tests extend this to show that, in practice, as the CME halo tends toward a circular geometry, the derived parameters become extremely sensitive to the exact form of the ellipse chosen and the technique as a whole cannot be considered reliable.
 But before passing on to the analyses at hand, it is important to emphasize two procedural considerations. First, it is critical that the projection of the minor axis of the fitted ellipse passes through the center of the Sun. Although the formulae in Xie et al.  presume this condition, they do not directly impose any such constraint, and it is therefore perfectly possible to analyze a halo CME in a way in which the assumption is violated. Such an incorrect analysis necessarily produces unrealistic results, such as CMEs that are systematically far too wide. Second, it is to be recognized that there is a great deal of subjectivity in fitting the outlines of any given CME, even when multi-view data are available. Operational experience reveals clear and measurable variations among observers, whether fitting running- or fixed-difference images. Moreover, these variations are compounded by the wide disparity in the intrinsic brightness of CMEs, the prior level of activity in the corona (residual effects of multiple CMEs in the field of view), and the energetics of the CME (whether the surrounding corona has been so profoundly affected by the launch of the CME that it is difficult to identify just what is being injected into the interplanetary medium).
 That being said, a demonstration of applying the technique is shown for two halo CMEs, that of 23 January 2012, whose centerline is offset 20–30° to the northwest, and that of the CME of 9 August 2000, which is less distinct but presents a more circular geometry indicative of a nearly head-on event. Both events are quite representative of halo CME appearance and are selected only to illustrate specific aspects of CME analysis, as follows.
 Analyses of the CME of 23 January 2012 are shown in Figure 1. Figure 1a shows a difference image of the event at 04:42 UT. Figures 1b–1e include four separate elliptical overlays consistent with the halo image and which are analyzed with the technique of Xie et al. . Each overlay shows the outline of the chosen ellipse and the minor and major axes (yellow line, blue line, and dotted green line, respectively). Note that the requirement that the projected minor axis of the ellipse passes through the center of the Sun (as shown by the small yellow dot) is met.
 Results for this event are given in Table 1. The eccentricity of the fitted ellipses increases from Figures 1b–1e, respectively. The results show a corresponding increase in the cone half angle and a decrease in the inferred radial distance, a correlation to be discussed at length in the next section. The four values of radial distance yield a relative standard deviation of 6.5%. The parameter most affected is the latitude, giving greater uncertainty as to the direction of the CME. Events that are more nearly head-on arrive sooner and with less statistical error in arrival time than glancing blows, so accurate determination of the absolute direction of events is a critical parameter in assuring system performance.
Table 1. Cone Parameters Resulting from the Ellipses Chosen in Figure 1 and Analyzed Using the Formulae of Xie et al. a
Cone Half Angle (deg)
Radial Distance (RS)
CME Axis Offset from S-E Line (deg)
aThe last column shows the angle between the CME axis and the Sun-Earth line for reference; it is not an input to the cone tool to be introduced later.
 Figure 2 shows a full halo CME event recorded by LASCO C3 on 9 August 2000. Figure 2a shows a difference image of the event at 19:42 UT, whilst Figures 2b–2f present overlays of five separate analyses.
 Results from the analyses are given in Table 2. Figures 2b, 2c, and 2e show fitted ellipses which all have similar forms but which lead to a large spread in the resulting values of cone half angle and radial distance. These figures demonstrate again an inherent correlation between cone half angle and radial distance. Figure 2d is similar but additionally shows the special case in which the ellipse is very nearly a perfect circle and the results are clearly unrealistic (i.e., a radial distance of 132 RS—a point noted by Xie et al. ). Figure 2f presents a much more elliptical solution, although it remains consistent with the form of the halo image. Fitting such an eccentric ellipse close to the Sun center yields results which are physical only if one accepts that the CME ejecta (the dense plasma blob normally supposed to represent the ICME driver gas) fill nearly an entire hemisphere about the Sun (note the cone half angle of 83°). This conflicts with the results of Burkepile et al. , which showed on the basis of Solar Maximum Mission limb events that a typical CME is only 52° wide (full width), with an increasingly sparse distribution of full widths extending out to 110°. Thus, discarding result (Figure 2d) as an “outlier,” the relative standard deviation in the values of radial distance is 38.8%. To summarize, we find the technique and formulae of Xie et al.  to be effective in characterizing only a certain class of halo CME, namely events that are very well defined, i.e., those which meet the circular cross-section criterion, and have a direction of propagation that is moderately “off-axis.” For CMEs propagating closer to the Sun-Earth line and where the boundaries of the event are poorly defined, the derived parameters cannot be considered reliable.
Table 2. Cone Parameters Resulting from the Ellipses Chosen in Figure 2 and Analyzed Using the Formulae of Xie et al. 
Cone Half Angle (deg)
Radial Distance (RS)
CME Axis Offset from S-E Line (deg)
3 Enforcing the Minor-Axis Assumption
 To ensure that the inferred CME parameters are consistent with the Xie et al.  assumptions, we have built a graphics system which creates a self-consistent 3-D geometric representation of the projected CME objects in the coronagraph images. The elliptical forms produced by this 3-D graphics system are thus completely consistent with those demonstrated by Zhao et al.  in their Figures 1 and 2.
 To do this, we utilize the 3-D object graphics system available as part of the Interactive Data Language (IDL; http://www.exelisvis.com). The IDL 3-D graphics system works by placing objects within a three-dimensional space, providing for a perspective projection, and then rendering the 2-D scene which results from viewing from a given “eye” location. The 3-D objects are physically located within an [X, Y, Z] space. The eye is located at X = 0, Y = 0 and at a location somewhere on the positive Z axis. A 2-D image results from rendering the 3-D objects onto the X-Y plane located at Z = 0.
 For a single (LASCO) view, we consider a 3-D space in which the center of the Sun lies at [0, 0, 0], the eye lies at a distance of roughly 215 RS (the location of SOHO, although this varies according to the orbit of the spacecraft), and the view plane in X-Y is a square of side roughly 60 RS (again variable but essentially the viewing area of the LASCO C3 coronagraph.) The CME cone is built into a model which has an axis that rotates about [0, 0, 0]. This axis is the direction of propagation of the CME. The CME cone itself is represented by a circular outline which is translated outward along the cone axis by a distance equal to the position of the CME leading edge. The radius of the circle is made such that it is consistent with the given cone angle. The result is a cone geometry, whose flat front can be adjusted for a correct perspective view accounting for the CME latitude, longitude, cone angle, and leading-edge distance. As viewed from the eye position, this presents an elliptical shape with a minor axis that passes through the Sun's center. This modeled 3-D scene is then scaled and overlaid directly on LASCO C3 coronagraph images.
4 The Essential Challenges and a Workable Solution
 However, whilst this new 3-D technique corrects the minor-axis problem, it does nothing for two other shortcomings that are critical to assuring reasonable estimates of CME properties.
As pointed out above, in practice, there is no way to distinguish between changes to an elliptical halo shape produced by adjustment of the angular width versus those induced by adjustment in the radial distance. This ambiguity is all the more compounded by uncertainties in propagation direction, which also fold into the final fit.
Limb events do not fit the standard flat-face cone paradigm—such CMEs just do not look like flat-faced cones viewed from the side. Rather, they most often appear more like ice cream cones, with a rounded front [Fisher and Munro, 1984; Xue et al., 2005].
 Neither problem can be resolved in a standard single-view analysis (e.g., from Earth, using LASCO observations) without additional information and/or assumptions concerning the CME geometry. For example, for halo CMEs one could simply assume a “typical” CME width from, say, Burkepile et al. , but that ignores the quite significant range in inherent widths that CMEs exhibit. Unless some other means (correlations with coronal dimming or net brightness, etc.) could be used to gauge the width of any given CME, large errors in predicted properties are unavoidable. Conversely, for near-limb CMEs, the (latitudinal) width and successive locations of the CME front are observable, but the distance from the limb (azimuthal angle)—needed to determine accurately the direction of the CME and to set its geometric scale and hence velocity—is lacking. Finally, assumptions as to the geometric shape of any given CME also fold into the fit, as discussed below.
 To expand upon (1) above, it is impossible in a practical sense to tell whether an “on-axis” halo of near-spherical cross-section represents a CME with a given cone angle and radial distance, or a different CME with a smaller cone angle at a larger distance. Experience with our 3-D graphics system makes it clear that this inherent cone angle/distance ambiguity is not just constrained to the “on-axis” case; the same problem applies equally to moderately “off-axis” CMEs which present an elliptical shape. This is exacerbated by the practical limitation that an operator is almost never presented with a perfectly elliptical shape, but rather a less well-defined “elliptical-like” blob which could be adequately well represented by any number of different cone realizations.
 The essential issues contributing to the cone angle/distance ambiguity are illustrated in Figure 3. Here we depict a cross-sectional view of two hypothetical head-on (to Earth) CMEs with flat-faced cone geometries. In either case (short and wide or long and narrow assumed cone), what we actually observe is an expanding circular disk of enhanced brightness having, presumably, something like a definable edge. In both cases, as viewed from Earth and projected on the plane of sky, the apparent angular width χ subtended by the edge of the propagating CME front in one image increases by Δχ in the time Δt that elapses, when a second image is taken. Suppose the “real” cone half width, ω2, is 25° (i.e., the narrower option in Figure 3) and also the measured Δχ/Δt implies 1000 km/s speed for that assumed ω2 (Δr2 can be derived trivially from χ, Δχ, and ω2). Then, had we mistakenly assumed the “real” half angle to be ω1 (=60°), we would infer a speed of only 269 km/s [it can be shown easily that (V2/V1) = tan(w1)/tan(w2)], a disparity of a factor of 3.72 in speed ratio. Thus, to the extent to which the applicable value of ω can—or cannot—be pinned down, the quality of the CME speed estimate can vary wildly between “good” to “useless.”
 Compounding this dilemma is the fact that the CME arrival time at R0 = 21.5 RS (the Enlil inner boundary) determines the point where the CME is injected into the corotating solar wind ambient flow in the propagation model. In this example, the time for the ω2 cone face to reach R0 is 2.4 h, whereas for ω1 it is 13.7 h. Thus the azimuth of the CME entry at R0 differs by roughly a half day rotation, or about 6° in longitude around the Sun. Depending upon the location of ambient stream fronts, that can be enough to affect the propagation of weak to moderate strength CMEs appreciably.
 Finally, in a companion paper [de Koning et al., 2012], we demonstrate how the unrealistic assumption of a flat-faced cone also leads to significant systematic underestimation of CME speeds, with the error increasing toward the limb. That is, taking it as a fact that the projected boundary of a CME top is rounded—which is consistent with limb observations—affects the inferred radial scale of the CME fronts and results in faster inferred propagation speeds near the Sun.
 Given the shortcomings of the familiar single-view cone analysis, it was realized that a more viable solution could be obtained by viewing a given CME event concurrently from another vantage point. Coronagraph images from the two STEREO spacecraft [Biesecker et al., 2008] can be combined with that from SOHO to provide a much improved assessment of CME properties. Since the launch of STEREO in October 2006, the two spacecraft have been traveling in heliocentric orbits at a distance of roughly 1 AU but are gradually moving relative to the Sun-Earth line: STEREO-A is leading, steadily progressing ahead of the Earth, whereas STEREO-B is lagging, moving in the opposite direction. Both spacecraft are increasing their angular separation from Earth at a rate of 22° per year [Driesman et al., 2008].
 At the time we were undergoing our cone tool investigations in 2011, both spacecraft had moved to positions roughly 90° around their orbits relative to Earth. Thus for a CME event aimed directly at Earth and producing a full halo as viewed from LASCO, both STEREOs would have “side-views”, as limb CMEs from which a reliable estimate of at least the north-south cone angle could be readily obtained.
 An example of just such a situation, which nicely illustrates the value of additional viewpoints, occurred on 15 February 2011. Figure 4a shows the emerging limb CME as viewed by STEREO-B COR2 [Howard et al., 2008; Sibley et al., 2012] while Figure 4b shows the view from SOHO LASCO C3, a full halo. Without a view from STEREO-B, analysis would be hampered by the cone angle/distance ambiguity as described above. However, with the addition of even just a single image from STEREO, a reasonable estimation of the cone angle could be had, providing a constraint on the solution afforded by the halo view from SOHO. (It should be noted, however, that the STEREO side view in this instance affords little to no diagnostic information in terms of the east-west orientation of the CME centerline).
5 Basics of the Three-View Concept
 In practice, the orbits of both STEREO spacecraft are always evolving relative to Earth and, in general, a CME will have an arbitrary location. However, what is equally clear is that analysis of the CME as seen concurrently from three widely spaced sun-pointing coronagraphs has the potential to provide far greater accuracy than anything that can be attained by analyzing images from a single telescope.
 Our three-view technique takes this approach. It works by constructing a three-dimensional geometric representation of the inner heliosphere using computer graphics techniques similar to those outlined in section 'Enforcing the Minor-Axis Assumption'. In domain of the graphics tool, the Sun is placed at the origin and has a radius of 1. An emerging CME is represented as a solid polygon object, again with its origin at the center. This polygon has a classic “teardrop” form derived from the lemniscate of Bernouli [Lawrence, 1972]. In two dimensions, the lemniscate is given by the formula
where a is a constant controlling size. This form is then extended into three dimensions as a surface of revolution, such that the “teardrop” has a circular cross-section. The lemniscate object has much in common with the simpler cone: It can be described completely in terms of a cone angle and radial size, but more accurately reproduces the curved nature, particularly of the leading edge, typical of observed CMEs. As with the 3-D-based cone tool outlined in section 'Enforcing the Minor-Axis Assumption', changes in CME propagation direction (i.e., latitude and longitude) are achieved by rotation of the lemniscate object about the origin.
 Having created our modeled CME polygon, three separate scenes are then rendered to provide perspective views from the locations of SOHO and STEREO-A and -B, respectively, correctly accounting for spacecraft position and rotation and the respective coronagraph fields of view.
 In 3-D computer graphics, scenes are usually viewed with appropriate lighting, giving objects placed in the scene a realistic shaded appearance. Figure 5a shows our 3-D lemniscate with such shading. For our purposes, we instead switch off the lighting/shading (whilst maintaining the exact same 3-D perspective view) rendering the object as “flat” (Figure 5b). From here we simply compute the boundary of the shape (Figure 5c). The “three-view” technique then works by overlaying the three modeled 3-D scenes containing the CME boundary on top of the actual coronagraph images from LASCO (C3 and also C2) and STEREO-A/B COR2.
 Figure 6 shows cropped images of nine matches made in the analysis of the 10 September 2011 CME event. The left column shows matches to STEREO-B COR2 for times of 04:24, 05:39, and 07:24 UT down the column, respectively. The center column shows matches to LASCO C2 for a time of 04:24 UT and then LASCO C3 for times of 05:30 and 06:18 UT. The right column shows matches to STEREO-A for times of 04:54, 05:24, and 05:54 UT. For this event, the results obtained were a latitude and longitude (in heliospheric coordinates) of 43° and 26° respectively, a cone half-angle of 38° and a radial velocity of 710 km/s.
6 The CME Analysis Tool (CAT)
 Our “three-view” concept has been realized as a practical software tool for routine operational use within the SWPC forecast center in the form of CAT, the NOAA/SWPC CME Analysis Tool. The prime philosophy followed in putting the tool together was to take maximum advantage of the available data while keeping the functional complexity of the tool within reason. Out of necessity, the tool relies upon near-real-time data streams from SOHO and STEREO, which are subject to significant data gaps; in addition, only beacon-quality images are available from STEREO in time for forecast purposes. Given the oftentimes ambiguous structure exhibited by CMEs in the corona (by energetic ones, in particular), we have intentionally constrained the cone form in the CAT to have a circular cross-section, i.e., elliptically shaped CME ejecta are not considered, nor is any internal structure assumed (e.g., croissant-shaped mass and/or magnetic distributions [Thernisien et al., 2006])—these enhancements are consigned for now to the research realm. (It should also be remarked that any elliptically shaped CME form input to Enlil will quickly develop a near-spherical disturbance front, as the pressure buildup at the leading edge of the CME responds to the largest-scale component of the input mass/momentum distribution).
 To provide a feel for the workings of the operational tool, a screenshot of its graphical interface is shown in Figure 7. Figures 7a–7j are the relevant functional panels within the interface.
 To analyze a given CME event, the forecaster loads the relevant images using the image loader (Figure 7e). The CME form and shape, as seen in the coronagraph images, are enhanced by using image differencing techniques and additionally with controls that adjust image saturation and contrast (Figure 7g). Both running difference and base difference techniques are available, although base differencing is to be preferred, as this best allows the operator to distinguish the actual CME ejecta material from other aspects, such as associated shockwave phenomena or the deflection of background streamers.
 Once the difference images are suitably enhanced, the operator manipulates the four free parameters (latitude, longitude, cone angle, and leading edge distance) which control the form of the 3-D lemniscate, so that the modeled CME outline exhibits as close a visual match as possible to the CME as imaged in the three coronagraphs and at different times (Figure 7h). The essential aim is to arrive at suitable values such that the model CME outline yields consistent matches to a variety of coronagraph images with the values for latitude, longitude, and cone angle remaining fixed and only the leading edge distance being adjusted. A minimum of two matches is required, at different times and leading edge distances, in order for a radial velocity to be calculated. The matches can be made from any combination of the four available telescopes.
7 Preliminary Results
 The CAT has been used routinely in near-real-time to analyze potentially geo-effective CMEs since the WSA-Enlil forecast model was introduced officially into operations at the National Weather Service on 1 October 2011. Table 3 shows results from the system for the first full year of operations (October 2011 to October 2012). The Table shows 25 events in which a model forecast predicted a CME impact at Earth and the CME was subsequently observed (by the ACE spacecraft) at Earth/L1.
Table 3. CME Arrival Time Predictions for 25 CME Events During the First Year of WSA-Enlil Operations (October 2011 to October 2012)
Arrival at ACE
Lead Time (days)
1 Oct 2011
5 Oct 16:00
5 Oct 06:47
02 Oct 02:00
22 Oct 2011
25 Oct 02:00
24 Oct 17:48
23 Oct 02:00
26 Oct 2011
30 Oct 10:00
30 Oct 08:55
27 Oct 22:00
9 Nov 2011
12 Nov 02:00
12 Nov 05:30
10 Nov 08:00
26 Nov 2011
29 Nov 12:00
28 Nov 21:15
27 Nov 02:00
25 Dec 2011
27 Dec 20:00
28 Dec 09:56
26 Dec 06:00
16 Jan 2012
20 Jan 04:00
21 Jan 05:02
16 Jan 20:00
23 Jan 2012
24 Jan 14:00
24 Jan 14:31
23 Jan 08:00
10 Feb 2012
14 Feb 07:00
14 Feb 07:00
11 Feb 02:00
24 Feb 2012
26 Feb 20:00
26 Feb 20:58
22 Feb 20:00
7 Mar 2012
08 Mar 10:00
08 Mar 10:45
07 Mar 08:00
10 Mar 2012
12 Mar 23:00
12 Mar 08:42
11 Mar 10:00
13 Mar 2012
16 Mar 07:00
15 Mar 12:40
14 Mar 18:00
2 Apr 2012
05 Apr 18:00
05 Apr 20:03
02 Apr 18:00
18 Apr 2012
21 Apr 16:00
21 Apr 09:25
20 Apr 06:00
19 Apr 2012
22 Apr 23:00
23 Apr 02:27
20 Apr 06:00
17 May 2012
19 May 20:00
20 May 01:36
18 May 00:00
18 May 2012
21 May 03:00
21 May 18:44
19 May 22:00
13 Jun 2012
16 Jun 19:00
16 Jun 09:01
14 Jun 20:00
14 Jun 2012
16 Jun 19:00
16 Jun 19:31
14 Jun 20:00
12 Jul 2012
14 Jul 12:00
14 Jul 17:28
13 Jul 06:00
28 Jul 2012
02 Aug 08:00
02 Aug 09:22
29 Jul 06:00
31 Aug 2012
03 Sep 17:00
03 Sep 11:23
01 Sep 08:00
2 Sep 2012
05 Sep 12:00
04 Sep 22:03
03 Sep 00:00
28 Sep 2012
30 Sep 15:00
30 Sep 22:13
28 Sep 20:00
Mean forecast accuracy
 For these 25 events, the approximate event date in the corona and the predicted arrival time at Earth are given in the first two columns, followed by the actual arrival time measured at ACE. The difference between the two (in hours) is a negative number if the model prediction was ahead of the observed arrival and positive if it lagged behind. The next column shows the time at which the model prediction was available to the SWPC forecasters and was published on the Internet. The last column presents the “lead time” for the model prediction, which we have defined to be the difference between the arrival of the CME at ACE and the time of the prediction, in days.
 Ignoring the sign of the difference in column 4, the 25 model predictions have an average difference of 7.5 h and an RMS of 10.0 h. It is important to appreciate that these results are from an operational system making routine predictions well in advance of the CME arrival at Earth and they are all the actual published model results. In some cases, we undertake multiple model runs, whereupon it is the job of the forecaster to make a decision as to which model run is to be published as our official model simulation. (This is based upon an assessment of input data quality, consistency of the cone fits, and how well the predicted solar wind speed ahead of the anticipated CME front matches ACE near-real-time observations.) The results in Table 3 are the published predictions. Note that because the radial grid spacing in the model (512 points spread uniformly between 0.1 and 1.7 AU) is coarse relative to the speed and linear scale of sharp interplanetary CME fronts, the practical accuracy of our forecasts is limited to about an hour.
 It is essential here to draw a distinction between prediction accuracy in near-real-time applications versus that achievable in retrospective research studies. The latter has the advantage of minimal data gaps, access to science-grade images, more latitude in event selection, and—most unfortunately—foreknowledge of the actual arrival times of the subject CMEs. Near-real-time operations, on the other hand, make use of the as-is STEREO beacon imagery and SOHO data streams, with data gaps posing a serious obstacle in many instances. Moreover, at SWPC, the event selection is driven by the forecaster assessment as to whether a given CME poses a threat to the near-Earth environment, meaning that many glancing blow events, which are inherently more difficult to predict, are folded into the statistical evaluation. All these issues and more need to be kept in mind in comparing performance measures between operational and research domains.
 Finally, it is worth noting that prior to the introduction of the WSA-Enlil model at SWPC, there was no fully operational model and forecasters relied instead upon flare location and intensity, type-II sweep information, and sporadic experimental model runs [e.g., Smith et al., 2000]. Before the day WSA-Enlil went into operations, SWPC recognized a timing accuracy of no better than a 12 h window on each side of a forecast arrival (i.e., somewhere within a 24 h interval).
 A key aspect of transitioning WSA-Enlil to the National Weather Service has been the development of an operational tool to provide accurate analysis of CME parameters for model input. Initial studies focused on applying flat-faced cone geometries to LASCO C3 coronagraph images of halo CMEs. We found the technique to only be of limited use, particularly for cases where the CME presents a full, relatively circular, halo. In this situation, uncertainties in the projected elliptical face of the cone were seen to lead to large errors in calculations of cone angle and radial velocity. In attempting to address this issue, we moved to full three-dimensional geometric modeling of the CME, first using the cone concept but then progressing to a more realistic rendition, a three-dimensional lemniscate. Extension of this system to model the CME from the three positions of SOHO and STEREO-A and -B has led to the current form of our operational CAT, which makes best use of the near-real-time imagery available from these non-operational spacecraft, with all the limitations and degradation of performance these impose.
 Clearly the CAT is most effective when images from all three spacecraft contribute to the analysis. Our experience however is that CAT remains effective when just two of the three views are available and even when the two views are significantly compromised with data dropouts: In essence, the second view provides enough of a constraint on the first for the system to work. Of course the relative positions of the STEREO spacecraft are constantly evolving and it is anticipated that certain future orientations will not be as useful as others. In addition, ultimately there are no guarantees of being able to leverage multiple coronagraph views to assist with the operational determination of CME parameters. Our work cautions that parameters derived from a single viewpoint may be unreliable unless studies to establish a statistical relationship for CME cone angle with other properties or to make use of potential sources of supplemental information (e.g., coronal dimming associations, STEREO polarization sets, HI-1 imagery, etc.) are fruitful.
 Finally, it needs be kept in mind that the model prediction accuracy of the overall WSA-Enlil model (or any such model) depends only in part upon its near-Sun CME inputs. Many other factors, such as reliable characterization of the ambient flow, the assumed structure of the ejecta entering the interplanetary medium, the occurrence of multiple CMEs in the system, whether a direct hit or a glancing blow is anticipated, etc.—all these and more affect the quality of the prediction in any given instance. Most importantly, while the model output constitutes a significant contribution to the official forecast, the efficacy of the system ultimately hinges upon the judgment of properly trained staff made in light of experience and a well-functioning set of practical tools, of which the CAT is but one. Thus, the issue of forecast accuracy is not at all straightforward, as will be addressed in subsequent papers.
 Source code for the CAT will be released to the solar physics community, not only to allow for general usage, but also to encourage further developments as both a research and operational tool. We plan for the release to be via SolarSoft (http://www.lmsal.com/solarsoft) in early 2013.
 The authors would like to thank the NASA Community Coordinated Modeling Center (CCMC) for providing initial cone model algorithms and Hong Xie for helpful discussions. We would also like to thank Dusan Odstrcil, Nick Arge, and colleagues at the Air Force Research Laboratory (AFRL) for their assistance with the transition of WSA-Enlil into operations at the National Weather Service.