Related Links | Frequently Asked Questions | Community Feedback | Downloads | Sitemap
CCMC Home | CCMC Stakeholders | Our Team | Publications | Meetings and Workshops | Concept of Operations
Models At A Glance | ModelWeb Catalog and Archive
Request Procedures | Generate Input Data Files & Parameters | Movies on Request | About the Run Process | Publications Policy
Search run database | Request run output | Special events | Kameleon Software | Space Weather Explorer | Publications policy
Instant Model Run
Forecasting Support tools | iSWA | DONKI | Mission Support | Experimental Real Time Simulations | Operational Geospace Model Validations
Intl Forum | GEM Challenge | CEDAR ETI Challenge | GEM-CEDAR Challenge | SHINE Challenge | CME Arrival Time Scoreboard | Flare Scoreboard | SEP Scoreboard | IMF Bz Scoreboard
Educational materials &activities | Space Weather REDI Initiative | SW REDI Bootcamp | Student Research Contest | Tutorial at CEDAR | Forecaster Tools
Missions near Earth/in Earth-orbit | MMS | Van Allen Probes | THEMIS | MESSENGER | STEREO | Spitzer | MAVEN | MSL | Dawn | Kepler | EPOXI | Juno | CASSINI | Voyager | New Horizons | Sounding Rockets | International
Research Community Support | CCMC Workshops | NASA Robotic Mission Operator Workshops | LWS Support | Exo-CCMC | DREAM2 Support | HELCATS Support
iSWA | DONKI | Kameleon | StereoCat | EEGGL | CME Scoreboard | SEP Scoreboard | FLR Scoreboard | SEA5

The Space Weather Forecasting Testbed


Model Developer(s)
Chunming Wang, Univerity of Southern California
Anthony Mannucci , JPL

Model Description
The Space Weather Forecast Testbed (SWFT) was developed in a joint effort between the Jet Propulsion Laboratory and the University of Southern California as a part of a NASA/NSF supported project for medium-term forecasting of thermosphere and ionosphere anomalies. Our objective is to identify promising approaches to produce 1 to 4-day advance forecast for significant disturbances of thermosphere and ionosphere using available solar, interplanetary magnetic field and other space weather observations.

Model Input
In order to best leverage the available machine-learning algorithms and human talents for the development of forecast models for space weather, we found that the construction of a testbed would be beneficial to the community. The key elements of a testbed should include:

  1. A rich database of quality controlled historical space weather data. In addition to raw observations, the database should also contain important features extracted from raw observations.
  2. Utilities for data preparation for forecast experiments. Unlike common machine learning problems in which a collection of "random samples" is randomly sampled for learning and validation, in forecasting, samples are inherently ordered in time. Concepts of forecasting must take into practical constraints in forecast lead time, data latency.
  3. An ensemble of machine-learning algorithms and data analysis tools that allow deep and wide exploration of possible techniques for space weather forecast.
  4. A platform where researchers can easily share their experiences through data and algorithm exchanges.
Our attempt in achieving the above goals is represented by the Space Weather Forecast Testbed (SWFT) which is developed in Matlab. The current version of SWFT uses Matlab's Deep Learning and Statistics Machine-Learning Toolboxes. The SWFT package currently comprises a database of historic space weather observation and features, a stand-alone set of Matlab scripts that provides users with graphic interface designing forecast experiments and a collection of examples we have used in creating forecast models.
The current SWFT database contains the following 4 main categories of historical data:
  1. Common space weather observations (suah as solar wind and interplantarey amgnetic fields), indices (such ss Dst, Kp, AE).
  2. Anomaly flags constructed from the data in category 1.
  3. Ionosphere features extracted from Global Ionospheric Maps (GIM).
  4. Anomaly flags constructed from data in category 2.
For more detailed descriptions of data in SWFT, see [1], [2]. The current SWFT database covers 12 years of historical data from 2003 through 2014. The main user interface for SWFT is to guide users in defining key parameters for a forecast experiment. In a such experiment, a user designates a "current epoch" which is typically a time cover by the historical database. For an example, a possible current epoch may be May 1st, 2011. Then a user must specify the desired forecast lead time. If the lead time is one-day, then the user intents to produce forecast for May 2nd, 2011 using all data available on May 1st, 2011 and forecast for May 3rd, 2011 using all available data on May 2nd, 2011, etc. The data available on a given day depends on data latency. Indeed, for most of observation data, there are necessary data processing and distribution delays that a forecaster must take into consideration. For an example, if data latency is one day, then the most recent data available on May 1st, 2011 is from April 30th, 2011. Even though SWFT requires the user to specify a global data latency value, in practice, each data source has its specific data latency. SWFT addresses this issue by allowing the user to specify a detailed time-stamp value relative to the current epoch for each variable used in producing the forecast. These parameters help to partition data in the SWFT database into sets available for training a forecast model and for validating a model. The following diagram, which is a part of the SWFT user interface, illustrates the relationship between training and validation data for a forecast model.

SWFT Time Intervals described

A user also selects a variable to forecast and a list of variables used to produce the forecast. For example, a user may select the global vertical Total Electron Content (TEC) as the quantity to forecast based on F10.7, Ap index, Sun spot numbers, etc. For each of these variables, a user can specify a subset of available historical values to be used. Based on these inputs, SWFT prepares and packages a set of training and a set of validation data consisting of collections of matched pairs {(Xk,Yk), k=1...n} where Yk represents the value for the variable to forecast and Xk represents the vector of all historical values to be used to predict Yk.

Following the preparation of dataset, a user can train a forecast model using a set of tools in SWFT. The currently available tools in SWFT serve as examples for a wider collection of machine-learning techniques that can be implemented for developing forecast models. Through collaboration with CCMC, we intend to gradually expand this collection of algorithms and analysis tools. SWFT source code is available to the community. We are currently developing a mechanism to integrate user contributed algorithms into SWFT.

Model Output
One attractive feature of SWFT hosted at CCMC is the user data repository. The model training and validation data package prepared by SWFT can be deposited into this repository to be shared with other users. It is our hope that as the user community expands, this data repository can serve to foster collaborations between scientists of different backgrounds.

Limitations and Caveats
The forecast model is completely user-defined. The parameter space is large and instructive examples are provided to guide the user for selected applications.

References and relevant publications
[1] Chunming Wang, I. Gary Rosen, Bruce T. Tsurutani, Olga P. Verkhoglyadova, Xing Meng and Anthony J. Mannucci, Statistical characterization of ionosphere anomalies and their relationship to space weather events, J. Space Weather Space Clim., 6, A5 (2016)
[2] Chunming Wang, I. Gary Rosen, Bruce T. Tsurutani, Olga P. Verkhoglyadova, Xing Meng and Anthony J. Mannucci, Medium Range Forecasting of Solar-Wind: A Case Study of Building Regression Model with Space Weather Forecast Testbed (SWFT), Accepted, Journal of Space Weather (2020).

Relevant links
Getting Started instructions to come
Download 2020/06/04 snapshot of full package (Matlab source code , data and documentation)
Download 2020/06/04 snapshot of Matlab source code and documentation
Download 2020/06/04 snapshot of training and validation data (covering years 2003 through 2014)
We are working on a process to generate more years of data. New data releases will be posted when available.

National Aeronautics and Space Administration Air Force Materiel Command Air Force Office of Scientific Research Air Force Research Laboratory Air Force Weather Agency NOAA Space Environment Center National Science Foundation Office of Naval Research

| | Privacy, Security Notices

CCMC logo designed by artist Nana Bagdavadze