Space weather forecasting is a coordinated attempt that takes data from ground- and space-based instruments and combines them with science-based numerical models to generate accurate and uninterrupted forecasts of the variable conditions in the near-Earth space environment, particularly geomagnetic storms and auroral activity. It includes issuing short-term forecasts through event-driven (e.g., solar flares, geomagnetic storms, proton events, electron events) alerts and warnings, offering daily forecasts of solar and geomagnetic activity and other long-term forecasts and warnings. Space weather research can be viewed as a cross-disciplinary scientific effort, or as a hybrid of basic space science research and applied science. Forecasters come from government, academic, or private sectors, and forecast models exist to predict actual physical parameters (or their proxies) and to predict the rate of changes over a specific time range. They provide advanced prediction, over a variety of parameters, of the state of the regions within close proximity of the Earth and of the region between the Sun and Earth, with alerts and warnings issued periodically that could extend out to the order of minutes to days, depending upon the object of forecast. They rely on real-time, in situ data measurements to augment the process of decision making. The level of confidence a forecaster places on predictive tools is critical for a user (e.g., power company, airline industry) to invest resources where such tools should be updated and augmented through improved understanding as needed, minimize false alarms, and offer better reliability and redundancy at critical times.

The high-speed coronal mass ejection (CME), ejected on 23 July 2012, observed by STEREO-A on the same day as the leading edge of the CME arrived at 1AU was unique both in respect to the observed plasma and magnetic structure and the large solar energetic particle flux that dynamically regulated the shock front. Because of its great intensity, it has been hailed as "Carrington 2" by some, warning that, had that CME been heading toward the Earth, it might have caused a major space weather event. We used the Rice Artificial Neural Network algorithms with the solar wind and interplanetary magnetic field parameters measured in situ by STEREO-A as inputs to infer what the "geoeffectiveness" of that storm might have been. We have also used an MHD model in Open Geospace General Circulation Model to understand the global magnetospheric process in time sequence. We presently show our neural network models of Kp and Dst on our real-time prediction site: http://mms.rice.edu/realtime/forecast.html. Running this event through our models showed that, in fact, this would have been an exceptional event. Our results show a prediction resulting in a Kp value of 8+, a Dst of nearly -250 nT, but when assumptions about maximum dipole angle tilt and density are made, predictions resulting in Kp of 11- and Dst dipping close to -700 nT are found. Finally, when solar energetic proton flux is included, the Kp and Dst predictions drop to 8- and approximate to-625 nT, respectively.

The Rice neural network models of Kp have been running in real time at http://mms.rice.edu/realtime/forecast.html since October 2007; Dst and AE models were added to our operations in May 2010. All these models use the Boyle index as basis functions computed from ACE real time inputs. Later, two more driving functions were included in November 2012: (a) the “Ram” functions that had dynamic pressure term added to the Boyle index and (b) the Newell functions. The Wing models are a set of neural network‐based Kp forecast models adopted by NOAA/Space Weather Prediction Center in March 2011 to supersede the Costello Kp model. This study indicates that any of the three Rice neural net predictors had a better success rate than the Wing model in predicting Kp (r=0.828 with Boyle, r=0.843 with Ram, and r=0.820 with Newell for 1 h predictions; similarly, r=0.739, 0.769, and 0.755 for 3 h predictions) in real time. In a head‐to‐head challenge using harvested real‐time outputs between April 2011 and February 2013, the Rice Boyle Kp models predicted better than the Wing models (0.771 versus 0.714 for 1 h predictions and 0.770 versus 0.744 for 3 h predictions). In addition, Wing's prediction was missing more often than the Rice prediction (≈6% versus 4.6%), meaning it had less reliability. The Rice models also predict AE(r=0.811 with Boyle; 0.806 with Ram; 0.765 with Newell, and 0.743 with Boyle; 0.747 with Ram for 1 h and 3 h predictions) and pressure‐corrected Dst (r=0.790; 0.767, and 0.704, and r=0.795; 0.797 and 0.707 for 1 h and 3 h predictions).

We have improved our space weather forecasting algorithms to now predict Dst and AE in addition to Kp for up to 6 h of forecast times. These predictions can be accessed in real time at http://mms.rice.edu/realtime/forecast.html. In addition, in the event of an ongoing or imminent activity, e‐mail “alerts” based on key discriminator levels have been going out to our subscribers since October 2003. The neural network–based algorithms utilize ACE data to generate full 1, 3, and 6 h ahead predictions of these indices from the Boyle index, an empirical approximation that estimates the Earth's polar cap potential using solar wind parameters. Our models yield correlation coefficients of over 0.88, 0.86, and 0.83 for 1 h predictions of Kp, Dst, and AE, respectively, and 0.86, 0.84, and 0.80 when predicting the same but 3 h ahead. Our 6 h ahead predictions, however, have slightly higher uncertainties. Furthermore, the paper also tests other solar wind functions—the Newell driver, the Borovsky control function, and adding solar wind pressure term to the Boyle index—for their ability to predict geomagnetic activity.

We present a new algorithm with an improvement in the accuracy and lead time in short‐term space weather predictions by coupling the Boyle Index, Φ = 10−4ν2 + 11.7Bsin3(θ/2) kV, to artificial neural networks. The algorithm takes inputs from ACE and a handful of ground‐based magnetometers to predict the next upcoming Kp in real time. The model yields a correlation coefficient of over 86% when predicting Kp with a lead time of 1 hour and over 85% for a 2 hour ahead prediction, significantly larger than the Kp persistence of 0.80. The Boyle Index, available in near‐real time from http://space.rice.edu/ISTP/wind.html, has been in use for over 5 years now to predict geomagnetic activity. The logarithm of both 3‐hour and 1‐hour averages of the Boyle Index correlates well with the following Kp: Kp = 8.93 log10 < Boyle Index> –12.55. Using the Boyle Index alone, the algorithm yields a correlation coefficient of 85% when predicting Kp with a lead time of 1 hour and over 84% for a 3 hour ahead prediction, nearly as good as when using Kp in the history but without any possibility of “persistence contamination.” Although the Boyle Index generally overestimates the polar cap potential for severe events, it does predict that severe activity will occur. Also, 1‐hour value less than 100 kV is a good indicator that the magnetosphere will be quiet. However, some storm events with Kp > 6 occur when the Boyle Index is relatively low; the new algorithm is successful in predicting those events by capturing the influence of preconditioning.