|
 |
|
IBM Research
| |
Visual Data Fusion for
Applications of High-Resolution Numerical Weather Prediction
Lloyd A. Treinish
IBM Thomas J. Watson Research Center
Yorktown Heights, NY
lloydt@watson.ibm.com
Introduction
Visualization is a method of computing
by which the enormous bandwidth and processing power of the human visual
(eye-brain) system becomes an integral part of extracting knowledge from
complex data. In that regard, our previous work has discussed methods
of appropriate mapping of user goals to the design of pictorial content
by considering both the underlying data characteristics and the (human)
perception of the visualization [10].
The introduction of new applications further challenges these ideas.
In particular, we have
extended our earlier work [e.g., 9] for situations where high-resolution
models can be utilized in variety of weather-sensitive decision-making
efforts such as emergency planning, energy production, airline operations,
risk assessment, agricultural activities, commodity trading, etc.
For each of these applications, information is assessed and decisions are
made based upon a variety of static and dynamic data sets, a subset of
which are weather-related. The utilization of these data and the
complexity of the decision-making process changes when high-resolution
predictive data are incorporated. These applications imply the coupling
of weather simulations with other models, analyses and data. Visualization
is a critical component to such integration. To enable effective
assessment and appropriate decisions, focused visualizations must be designed
to integrate these distinct data sources, yet still be driven by user goals.
Resultant visualizations which represent a fusion of weather and non-weather
data may not even illustrate forecasts of weather phenomena directly.
In these cases, the relevant information is in the impact of weather via
derived properties, which are influenced by weather, not weather variables
produced by a simulation. The problem is illustrated schematically
in Figure 1. Two traditional data generators are shown on the top
and the bottom (weather and non-weather, respectively). Although
visualization is applicable to both, typically this is mutually independent.
We propose an approach of visual data fusion to address the visualization
design problem in such applications.
 Figure 1. Visual
Data Fusion for Weather Model Applications.
Data Fusion
Data fusion
is simply the integration of multiple data sets. This notion is derived
from the fact that understanding of phenomena from a scientific basis,
creating an engineering design, or assessment for sound decision making
requires the utilization of data from many distinct sources. Traditionally
such tasks have utilized a single data set, but as a result is often incomplete
for larger-scale problems that are becoming more prevalent today.
In parallel with the growth in problem complexity are additional factors
that make the need for data fusion more practical and thus, more pervasive.
The relative availability of relevant data enables a comparison study for
a data generator as much as it does an independent analysis. Secondly,
data generators have become more capable and accessible. Digital
data acquisition is easier and cheaper. Computational simulations
are gaining fidelity and detail while becoming more practical to compute.
From verification of computational and experimental models to steering
simulations with real-world observations, bringing data from multiple sources
together is much more powerful than using each source separately.
Visualization is critical to this integration, without which the beneficiaries
of such data would be overwhelmed by volume or complexity [13].
Data from multiple sources require
care in their presentation so that artifacts due to the visualization process
are not introduced by data fusion and erroneously interpreted as features
in the data. For example, the data may not be uniformly available
for the spatial domains being examined. Each of the data sets to
be "fused" are generally not geographically co-registered and are defined
on differing geometric structures. Further, the coordinate system
for visualization and interaction may need to differ from those native
to the data sets of interest.
These issues have been considered
by others in a variety of applications including earth science, physics,
astronomy and medical imaging [13]. In the majority of these cases,
the user goals focused on analysis or verification as opposed to data assessment
as illustrated in efforts to compare computational fluid dynamics results
with experimental data from wind tunnels [4]. More recent work has
considered decision support [2] but from a human factors perspective.
Approach
To enable visual data fusion, a
perspective of data management must be adopted by introducing an uniform
data model that is matched to the structure of the data as well how such
data are used. This implies a generalized mechanism to classify and
access data as well as efficiently map data to operations. The implementation
of such a data model effectively decouples the management of and access
to the data from the actual application. This encapsulates the variety
of sampling and representations for diverse data and provides uniform access.
It it then a prerequisite to building applications that utilize the data
sets to be integrated [ 11].
One consequence of such a data (model)-centric approach is that the same
operation(s) can be applied to data sets that need to be visually fused
or correlated (i.e., displayed and interacted together) without introducing
superfluous interpolation or resampling to a common mesh. The latter
process implies a modification to the data, whose impact could be hidden
in subsequent visualization. Further, if a specific visualization
task requires a cartographic projection, then these data sets can be independently
warped by the prerequisite transformation. Any geometric distortion
that is introduced is due only to the actual projection since the data
and topology remain invariant through such a transformation. It is
also independent of the choice of realization or rendering technique or
cartographic projection, and hence, provides a framework for experimenting
with different visualization strategies. As a result, the fidelity
of the original data sets is preserved in a coordinate system suitable
for dynamic interaction. It implies that correlative visualization
for visual fusion can be approached from four perspectives. In all
cases, the specific choices are dictated by the goal of the visualization
task(s) as defined by the individuals or applications utilizing the data.
-
Image Level. The
capability to look at multiple sets of data in exactly the same fashion
(i.e., visual comparison within a common framework). This can be
achieved with multiple visualizations in adjacent windows or mosaiced together
for qualitative comparison. These visualizations are usually static,
but might be accompanied by synchronized animation sequences or geometric
transformations in which the representations are linked. Outside
of the latter, interaction is typically indirect.
-
Common View. The
capability to utilize a variety of visualization strategies within a chosen
coordinate system dictated by one of the data sets or independently by
user task. This represents a visual fusion which can support both
direct and indirect interaction, including numerical querying. All
of the relevant data are registered within this common viewing framework.
Qualitative comparisons are clearly supported, but direct quantitative
comparisons are defined by interaction.
-
Data Level. The
capability to numerically compare distinct data sets using either of the
two previous approaches for visualization. This does require the
transformation (e.g., interpolation) of one or more data sets to a common
basis (mesh, coordinate system, etc.) from which derived quantities can
be calculated (e.g., point-wise operations). The visualizations may
involve the original data and/or the derived data. From the discussion
earlier this can violate the principle of preserving fidelity at the cost
of supporting numerical comparisons.
-
Multiple Views.
The capability to numerically and visually compare multiple data sets,
particularly when some of the data sets do not have a common basis for
visual fusion. In this case, the utilization of a variety of different
strategies is required, some of which must be in separate instead of a
common framework. Interaction may be complex because separate metaphors
for direct interaction are required for each framework, although common
methods for indirect interaction are feasible. Unlike case 1, quantitative
access is supported such that linked displays would indicate related numerical
values or "regions" of commonality that are queried.
Applications
The aforementioned approach to
data fusion is applied to problems that relate to economic and societal
impacts of weather. In some cases, additional sources of weather
information, both historical and predictive may have further benefit, when
integrated into the decision-making process.
Emergency Planning
Weather-related catastrophes have
led to over $48B in property insurance claims from 1989 to 1993 in the
US. In North Carolina alone, ten major hurricanes from 1983 to 1996
resulted in about $50B worth of damage, almost $30B of which led to losses
by insurance companies [6]. Hence, disaster planning or hedging for
underwriting risk-related insurance can benefit from improved weather predictions.
In both cases, the impact of weather is relevant in visualization but not
the weather data directly. Although geo-referenced visualizations
are required, the illustration of time-dependent factors related to property
loss due to severe weather are needed, not merely a visualization of predicted
wind velocity, for example. Usually, an Image Level (case
1) approach is applied as shown in Figure 2. Each image contains
a simple two-dimensional map of a set of glyphs colored by a different
parameters. The glyphs are located at the centroid of the area associated
with zip codes. An example animation is also
available for viewing.
Figure 2. Image Level
data fusion of a weather forecast with demographic data over an 800 x 800
km domain at 8 km resolution centered over Dallas. Colored glyphs
at zip code locations illustrate a subset of demographic and derived data.
However, the glyph locations are
only marked on the map when a set of conditions on house value, population
and estimated damage due to wind are met. Therefore, a Common
View (case 2) approach is more efficient by leveraging user interaction
as illustrated in Figure 3. The user is free to interactively set
the conditions and animate in time corresponding to the weather simulation
in hourly steps. This enables the determination of areas of greatest
impact due to severe weather. Essentially, it represents a simple
method to specify a query against various data sets, which are then used
to constrain a visual integration for display and interaction. This
approach becomes Data Level (case 3) because the forecast data are
interpolated to zip code locations in order to support the query constraints.
These thresholds can also be augmented to include other relevant demographic,
customer or property data. The demographic data shown are derived
from available census information (http://tiger.census.gov).
An example
animation
is also available for viewing.
Figure 3. Common view
data fusion showing the relationship between demographic data and a weather
forecast in a screen capture of an interactive session. This also
is a data view data fusion because the locations of estimated damage are
calculated from the weather model data.
In this example, the conditions
for display are enhanced to include a simple computational model.
The level of wind-induced damage is based upon analysis of effects on typical
residential buildings from severe weather [12]. This approach to
data fusion may be useful for planning purposes by an insurance company
or deployment of repair crews by a utility or local highway department.
Electricity Demand Forecasting
Another application of a predictive
weather model is to forecast load on a power-generation facility or transmission
lines
for efficient running of the facility or for power trading. In both
cases, meteorological information is an important input as weather is a
primary driver for electricity demand. It has been estimated that
the annual cost of under or over predicting electricity demand due to poor
temperature forecasts is several hundred million dollars in the US alone.
Erroneous weather data associated with startup-shutdown of generation units
can be worth $500K per day during peak load periods or conservatively $8M
annually to a regional power authority. In addition, improved severe
storm predictions to reduce outage time can save a few hundred thousand
dollars a year for a typical utility [5].
Decisions in this industry are driven by diverse non-weather data and processes
including load forecasting and econometric models, customer demographics,
geography of power facilities, etc. that are not well integrated.
The weather information currently used is relatively coarse leading to
poor and costly decisions. Typically, hourly forecast surface temperature
and dew point values averaged over a large geographic region are used.
Alternatively, more accurate data at greater frequency which are distinct
for different loads by geographic location and altitude can be applied
coupled with other factors that influence load (e.g., storm and cloud predictions).
Since there is a relationship between accuracy in load prediction vs. economic
efficiency (i.e., an under prediction implies having to buy power at a
premium and over prediction means resources are wasted), coupling of weather
forecasts with econometric models is also feasible.
These ideas are illustrated
in Figure 4 using Data Level fusion. It shows a map of Georgia
with forecasted heat indices at 8 km resolution. Major cities and
locations of the generators owned and operated by Georgia Power, the local
electric utility, are shown by name. Each power plant location is
also marked with a pin. whose height and color indicate a predicted electricity
demand. A dual encoding is used because the capacities of the power
plants range over five orders of magnitude. Hence, height is a linear
mapping while color bands are scaled logarithmically. An animation
of this visualization for a 24-hour simulation at 10-minute time steps
illustrates the temporal and geographic variation of predicted load.
Figure 4. Data level data
fusion of a weather forecast at 8 km resolution centered over Atlanta with
a prediction of electricity demand at power plants operated by Georgia
Power. The demand is calculated from a model whose input is derived
from the numerical weather prediction.
The load is computed interactively
as a function of temperature, humidity and time of day from a simple model.
The temperature dependence is based upon a polynomial approximation of
the relationship between historical data of power demand and weather observations,
shown in Figure 5 [7]. Regression on the data from summer weekdays
in the southeastern United States after outliers are removed yields,
W = 1.146 - 0.0225T
- 0.000240T2 + 0.0000397T3<
(1)
Figure 5. Weather-dependent
component of energy load, W(T).
The temporal variation is based
upon a spline fit of hourly electricity requirements for mid-week days
in urban and suburban tropical environments, which is consistent with other
results in the literature [3]. That component is shown in Figure
6, which is then normalized for this application.
Figure 6. Diurnal, mid-week
component of energy load, N(t).
The temperature and
temporal components are combined for a total estimated load, L, such that
L = C[(0.2768809N(t)
+ 0.7231191)(W(THI)/2.9175)]
(2)
The function is scaled
by the rated power plant capacity, C, using published data ( http://www.georgiapower.com/newsroom/plants.asp).
Heat index, THI, is employed
as a more accurate measure of demand than simply temperature. It
is an apparent temperature derived from both temperature and humidity as
an indicator of personal comfort during the summer [8]. Therefore,
it is directly related to air conditioning usage and thus, electricity
demand. The weather model results are interpolated at each time step
to the location of each of the power plants. An example for the specific
24-hour period of the forecast is shown in Figure 7 for the largest power
plant operated by Georgia Power (Bowen).
Figure 6. Predicted power demand
at a specific generator site derived from a weather-model-driven load forecast.
All of these capabilities are illustrated
in Figure 8, which is a screen capture of a prototype of an interactive
application for detailed load forecasting. The user has the ability
to select the type of power plant (fossil, hydroelectric and/or nuclear),
what data to show on the map (e.g., weather, geographic or other customer/demographic)
and to query individual power plants (i.e., by visual selection).
The results of the query include the predicted load at each time step (as
fine as every 10 minutes) as well as a plot of predicted load over 24 hours
with weather data at that location. The interactive application is
then a Multiple View (case 4) fusion.
Figure 8. Data Level
(3d window) and
multiple View (with 2d plot) data fusion for weather-model-driven
energy load forecasting in a screen capture of an interactive session.
The multiple views are linked in time sequence and by interactive selection
in the 3d window.
The visual fusion techniques of
Figures 3 and 4 are combined in Figure 9, which shows the load forecast
at the power plants that use fossil fuels with a population map.
The population data are shown as colored contours on a logarithmic scale
to segment urbanized areas (red) and their location with respect to the
power generators under the heaviest demand. Although these data are
derived from static census sources, the same techniques would apply to
similar but proprietary customer data owned by a regional electric utility.
Figure 9. Data Level
and Common View data fusion to illustrate the correlation of a load
forecast with demographic data. The proximity of fossil fuel power
plants with high predicted load to major population centers (red) is easily
seen.
The applications shown in Figures
3 and 8 present a user interface based upon XWindow/Motif for indirect
interaction and OpenGL for direct three-dimensional interaction in cartographic
coordinates native to the weather simulation. They have been implemented
with Data Explorer (DX) [ 1].
DX is a portable, open source, general-purpose software package for visualization
and analysis ( http://www.research.ibm.com/dx
and http://www.opendx.org). A generic
toolkit was used to avoid having to implement a graphics and computational
infrastructure. Unlike traditional meteorological graphics or geographic
information systems, DX is parallelized for multiprocessor workstations
and can utilize three-dimensional graphics accelerators. DX is built
upon an unified data model that enables these applications to operate directly
on the native gridded weather data without transformation or compression.
Conclusions and Future Work
The visualization of applications
of high-resolution weather modelling have benefited from a focus on specialized
interfaces and tools matched to user goals and underlying visualization
tasks. Since the underlying toolkit is extensible tools can be reused
between applications with similar user interface components. Although
these applications and associated user goals are different, underlying
data fusion requirements and visualization tasks are the same. Further,
the need to employ a relatively simple user interface is desirable to reduce
the effort for training of users in time-critical activities such as decision
support. It also reduces the cost of development and maintenance,
and enables more rapid iterative refinement with or adaptation to new users.
Therefore, within any given application, incorporation of additional and
more complex data sets can also be addressed. But the goal remains
the same -- to develop simple interfaces and useful visual fusion.
The specific work discussed herein
is on-going. To date, most of this work has been of a prototyping
nature as both generic proof of concept as well as illustration of specific
decision support problems. One aspect of continued efforts will be
to incorporate more sophisticated models or processing as the consumer
of weather forecast data. For example, the simple load forecasting
model can be enhanced to include wind speed and sunshine duration effects
and also, adjusted for more realistic temporal variation based upon day
of week and season. To aid in the decision making applications, migration
to a probabilistic representation will also be advantageous. In addition,
it is believed that these ideas can be extended to other application areas
such as agriculture, aviation, finance, etc.
The need for this
type of data fusion occurs in other disciplines when an integrated view
of the problem domain is adopted. Hence, the applicability of these
methods to other problem areas will be investigated. This includes
large-scale engineering design when computer-aided (mechanical) designs
have to integrated with computational fluid dynamics, structural analysis,
styling design, wind-tunnel measurements with physical models and actual
testing with full-scale prototypes. Other applications may include
multi-modal imaging such as in astronomy, earth sciences or medicine.
Although challenging they have less data diversity because conceptually
similar measurements are taken of the same physical space with varying
sampling schemes that yield different information that have to be registered
into a common coordinate system.
References
-
Abram,
G. and L. Treinish.
An Extended Data-Flow Architecture for Data Analysis
and Visualization.
Proceedings of the IEEE Visualization 1995 Conference,
October 1995, Atlanta, pp. 263-270.
-
Bisantz, A. M., R. Finger, Y. Seong
and J. Llinas. Human Performance and Data Fusion Based Decision Aids.
Proceedings
of the FUSION '99 Conference, July 1999, Sunnyvale, Volume 2, pp. 918-925.
-
Chang, C. S. and M. Yi.
Real-Time
Pricing Related Short-Term Load Forecasting. Proceedings of the
Energy Management and Power Delivery 1998 Conference, March 1998, Singapore,
pp. 411-416.
-
Keely, L. and S. Uselton.
Development
of a Multi-Source Visualization Prototype. Proceedings of the IEEE
Visualization 1998 Conference, October 1998, Raleigh, pp. 411-414.
-
Keener, R. N. The Estimated Impact
of Weather on Daily Electric Utility Operations. Proceedings of
the Workshop on the Social and Economic Impacts of Weather, April 1997,
Boulder (http://www.esig.ucar.edu/socasp/weather1/keener.html)
-
Kunkel, K. E., R. A Pielke, Jr., S.
A. Changnon.
Temporal Flunctuations and Climate Extremes that Cause
Economic and Human Health Impacts: A Review. Bulletin of the American
Meteorolofical Society,
80, n. 6, June 1999, pp. 1077-1098.
-
Robinson, P. J. Modeling Utility
Load and Temperature Relationships for Use with Long-Lead Forecasts.
Journal
of Applied Meteorology, 36, n. 5, May 1997, pp. 591-598.
-
Rothfusz, L. P. The Heat Index Equation
(or more than you ever wanted to know about heat index). National
Weather Service Southern Region Technical Attachment, SR/SSD 90-23,
Fort Worth, 1990.
-
Snook, J. S., P. A. Stamus, J. Edwards,
Z. Christidis, J. A. McGinley. Local-Domain Mesoscale Analysis and Forecast
Model Support for the 1996 Centennial Olympic Games. Weather and
Forecasting,
13, n. 1, 1998, pp. 138-150.
-
Treinish,
L. Task-Specific Visualization Design.
IEEE Computer Graphics
and Applications, 19, n. 5, September/October 1999, pp. 72-77.
-
Treinish,
L. A Function-Based Data Model for Visualization.
Proceedings
of the IEEE Visualization 1999 Conference Late Breaking Hot Topics,
October 1999, pp. 73-76.
-
Unanwa, C. O., J. R. McDonald, K. C.
Mehta and D. A. Smith. The Development of Wind Damage Bands for Buildings.
Journal
of Wind Engineering and Industrial Aerodynamics, 84, n. 1, January
2000, pp. 119-149.
-
Uselton, S., J. Ahrens, W. Bethel,
L. Treinish and A. State. Multi-Source Data Analysis Challenges.
Proceedings
of the IEEE Visualization 1998 Conference, October 1998, Raleigh, pp.
501-504.
|
|
|
|
|
