Session CD5 - The Ensemble Method in Space Weather Forecasting: bridging the gap between expectation and reality
Siegfried Gonzi (UK Met Office), Vic Pizzo (SWPC Boulder, USA), Eric Adamson (SWPC Boulder, USA), Emiliya Yordanova, onsite (Swedish Institute of Space Physics), Rachel Bailey, onsite
Ensemble techniques in terrestrial weather forecasting have come a long way and it is fair to say that modern weather forecasts are unthinkable without the aid of ensembles. At the risk of not reinventing the wheel the space weather community should learn from the terrestrial weather community. But it is less clear how much of that already existing knowledge can easily be applied to space weather forecasting. The community lacks a clear strategy of how to manage this problem. We are faced with a possible practical limitation of ensemble techniques in space weather forecasting due to a lack of observations in the heliosphere. The idea underpinning ensemble techniques is to draw uncertainties from possible prior states that show some semblance to a real state which is often derived from observations. Ensembles can help with gaining insights into how model and observation uncertainties unfold and manifest themselves in the forecasts. This will benefit model developers and forecasters alike. This is the first ever ESWW session that deals with ensemble methods and user needs. We invite contributions from colleagues working in all fields relating to space weather forecasting, modelling and observations. This includes Sun to Earth, radiation belt, magnetosphere modelling and forecasting techniques. We also invite colleagues to submit presentations that demonstrate that ensemble methods would only add little value to their work.
Poster ViewingThursday October 27, 08:30 - 13:30, Poster Area Talks Wednesday October 26, 14:15 - 15:15, Fire Hall Click here to toggle abstract display in the schedule
Talks : Time scheduleWednesday October 26, 14:15 - 15:15, Fire Hall14:15 | OSPREI: A Coupled Ensemble Approach to Modeling CME-Driven Space Weather With Automatically Generated, User-Friendly Outputs | Kay, C et al. | Oral | | Christina Kay [1,2], M. L. Mays [1], Y. M. Collado-Vega [1] | | [1] NASA GSFC, [2] CUA | | Coronal mass ejections (CMEs) drive space weather activity at Earth and throughout the solar system. Current CME-related space weather predictions rely on information reconstructed from coronagraphs, sometimes from only a single viewpoint, to drive a simple interplanetary propagation model, which only gives the arrival time or limited additional information. We present the coupling of three established models into OSPREI (Open Solar Physics Rapid Ensemble Information), a new tool that describes Sun-to-Earth CME behavior, including the location, orientation, size, shape, speed, arrival time, and internal thermal and magnetic properties, on the timescale needed for forecasts. First, Forecasting a CME's Altered Trajectory (ForeCAT) describes the trajectory that a CME takes through the solar corona. Second, ANother Type of Ensemble Arrival Time Results simulates the propagation, including expansion and deformation, of a CME in interplanetary space and determines the evolution of internal properties via conservation laws. Finally, ForeCAT In situ Data Observer produces in situ profiles for a CME's interaction with a synthetic spacecraft. OSPREI includes ensemble modeling by varying each input parameter to probe any uncertainty in their values, yielding probabilities for all outputs. Standardized visualizations are automatically generated, providing easily accessible, essential information for space weather forecasting, including a range of probabilities. We show OSPREI results for a CMEs observed in the corona on 22 April and 09 May 2021. We approach these CME as a forecasting proof-of-concept, using information analogous to what would be available in real time rather than fine-tuning input parameters to achieve a best fit for a detailed scientific study. The OSPREI “prediction” shows good agreement with the arrival time and in situ properties. | 14:30 | Title: Deep Learning models in confronting ADAPT and satellite observations. | Zhou, Y et al. | Oral | | Y. Zhou [1], S. Gonzi[3], D. Jackson[3], C. Budd[1], T. Fincham-Haines[2] | | [1] University of Bath, Department of Mathematical Science, [2] University of Bath, Department of Computer Science, [3] UK Met Office, | | The UK Met Office uses the WSA-Enlil model to simulate the solar wind. WSA is initialised using Earth-based magnetogram measurements from GONG. There are issues with this initialisation due to missing observations of the far side and polar regions of the Sun. Such errors can be improved to a certain extent by applying an assimilation model called ADAPT which gives an ensemble of 12 equally probable realisations of solar magnetic field. However, predictions based on these realisations contains different uncertainties and the solar wind simulation behaviour is difficult to understand. Our aim in this paper is to analyse the accuracy of the 12 outputs from the ADAPT model and to rank them to select the best initial condition for WSA-Enlil. The paper addresses the problem of selecting the most realistic data based upon this ranking. To do this ranking we compare hourly low resolution WSA output data which is in the Carrington cylindrical equal area projection with high resolution satellite observations which are in the Helioprojective Gnomonic projection. Making such a comparison directly is complicated because the data is presented in these different coordinate systems and has different resolutions. As a consequence, a coordinate transformation must be applied to allow like-for-like comparison. Additionally, the satellite observations must be downsampled from high to low resolution to match WSA output. The satellite observations are pre-processed by the CHIMERA segmentation to give the coronal hole masks. These masks are compared with the WSA coronal hole ensembles (derived from the ADAPT magnetic field maps) by using a convolutional neural network (CNN). The CNN learns (by looking at a year's worth of training data) to assimilate different ensembles with the satellite observed CHs. The similarity to the assimilated target is then calculated for each realisation. This then gives the ranking of each output which then produces different rankings for each of the 12 realisations. | 14:45 | Reduced-physics solar wind models for large ensemble forecasting | Owens, M et al. | Oral | | Mathew J. Owens [1], Luke A. Barnard [1], Huw Morgan [2], Anthony Yeates [3],Shaun Bloomfield [4] | | [1] University of Reading, UK, [2] Aberystwyth University, [3] Durham University, [4] Northumbria University | | Prediction of near-Earth solar wind conditions is essential for space-weather forecasting with lead times longer than tens of minutes. This is achieved using heliospheric models initialised with near-Sun solar wind conditions, which are in turn are typically provided by a combination of magnetogram-constrained coronal models and coronagraph observations of CMEs. Uncertainty in these near-Sun boundary conditions and the subsequent effect on the solar wind forecast at Earth can be quantified by perturbing the near-Sun conditions and generating an ensemble of heliospheric model solutions. Unfortunately, the dimensionality of the parameter space (uncertainty in the ambient solar wind and multiple properties of each CME) means this requires very large ensembles, O($10^3 -- 10^5$), which is computationally prohibitive with 3-D MHD models of the solar wind. We show how reduced-physics heliospheric models can provide this capability, augmenting the forecasts provided by more sophisticated numerical models. | 15:00 | Solar Predict: a tool to forecast the solar activity cycle: a Cycle 25 update | Brun, A et al. | Oral | | A.S. Brun [1], C.P. Hung [1], L. Jouve [2], A. Strugarek [1] | | [1] DAp-AIM, CEA Paris-Saclay, France, [2] IRAP, France | | We present our Solar Predict services that forecast the solar activity cycle and is now being adapted within the SWESNET project to the ESA SWE portal.
Solar Predict is based on advanced variational + sequential data assimilation pipeline coupling a mean field dynamo models to solar data such as the sunspot number (SSN) and the Blos magnetic field (from magnetograms). We also include an ensemble forecasting at then end of the assimilation window of the previous magnetic states. We find that the tool is able to forecast cycles 22 to 24 efficiently, with an ensemble reproducing well the observed activity. We also find that the quality of the polar field goodness of fit is key for accurate forecasting and not just the SSN. For cycle 25, we predict a maximum late in 2024/early 2025 with 9 months of uncertainty and a SSN number slightly larger than the maximum of the previous cycle 24, SSN25_max ~ 120 +/- 15. We do not expect a grand minimum phase in the coming years. |
Posters1 | To Ensemble or Not Ensemble | Camporeale, E et al. | Poster | | Enrico Camporeale | | University of Colorado & NOAA Space Weather Prediction Center | | Why Ensemble methods are not a good idea for space weather, and possible alternatives. | 2 | Daily ensemble forecasting from the Sun to 1 AU - The PAGER EU project. | Arber, T et al. | Poster | | Tony Arber, Keith Bennett, Andrew Angus, Bart van der Holst | | University of Warwick | | Ensemble predictions of SW at 1 AU are made possible by using a 1D field-line tracing model for the solar corona to drive a 3D inner heliospheric model. This is a work-package in the larger EU funded PAGER project. Using this combination of 1D coronal and 3D heliospheric models it is possible to run ensembles with 20 members every few hours. The ensemble members are sampled by Latin hypercube methods across the range of known uncertainties in the key physics inputs. The optimal value from this ensemble is found through Gaussian Process regression. This overall method is therefore capable of predictive ensembles of SW for the steady SW every few hours. To compliment this a CME injection model is used. These are triggered by alerts from the CCMC DONKI service and involves injecting a Gibson-Low flux rope model CME into the simulations at 0.1 AU. The model CME is selected from a pre-calculated database of Gibson-Low CME's generated by the SWMF framework. Results will be presented across a range of SW conditions both with and without CMEs. | 3 | Over 20-year global magnetohydrodynamic simulation of Earth's magnetosphere | Honkonen, I et al. | Poster | | Ilja Honkonen, Max van de Kamp, Theresa Hoppe, Kirsti Kauristie | | Finnish Meteorological Institute | | We present our approach to modeling over 20 years of the solar wind-magnetosphere-ionosphere system using version 5 of the Grand Unified Magnetosphere-Ionosphere Coupling Simulation (GUMICS-5). As input we use 16 s magnetic field and 1 min plasma measurements from ACE satellite starting from 1998. We have parallelized the magnetosphere of GUMICS-5 using the Message Passing Interface and have made several improvements which have e.g. decreased its numerical diffusion. The first version of the 20 year run used a maximum magnetospheric resolution of 0.5 Earth radii (Re) while currently we have simulated over 2 years using the standard GUMICS maximum magnetospheric resolution of 0.25 Re. At this resolution, 1 day of size-optimized results requires approximately 20 GB of space giving a total required disk space estimate of over 150 TB for the period 1998-2020. We describe some of the challenges of working with such a large data set, which is also relevant for working on large model datasets obtained from large scale and/or long term ensemble forecasting. We investigate how increased magnetospheric resolution improves forecasts of magnetosphere, ionosphere and geomagnetic indices on solar cycle time scales. We also investigate how such a large simulation dataset could be used to reduce the number of ensemble runs required for statistically relevant forecasts. | 4 | How ensemble modelling can be easily employed to a simple Drag-Based Model: Drag-Based Ensemble Model (DBEM) | Čalogović, J et al. | Poster | | Jaša Čalogović [1], Manuela Temmer [2], Mateja Dumbović [1], Bojan Vršnak [1], Astrid Veronig [2] | | [1] Hvar Observatory, Faculty of Geodesy, Kačićeva 26, HR-10000 Zagreb, Croatia, [2] Institute of Physics, University of Graz, Universitätsplatz 5, A-8010 Graz, Austria | | The Drag-Based Ensemble Model (DBEM) is a probabilistic model for simulating the heliospheric propagation of Coronal Mass Ejections (CMEs). The output of the model covers a CME hit chance for chosen targets, most probable arrival times and speeds, and quantifies the prediction uncertainties and calculates confidence intervals. It is based on the 2D analytical Drag-based Model (DBM) with very short computational time. That enables a user to run thousands of DBM runs for n different input parameters within seconds to obtain the distribution and significance of results. The DBM/DBEM web application is integrated as an important operational tool for space weather forecasters into the Heliospheric Weather Expert Services in the frame of the ESA Space Safety programme (https://swe.ssa.esa.int/current-space-weather). As example of the successful employment of ensemble method in space weather forecasting, the most recent version of DBEM (DBEMv4) will be presented. It uses as input the dynamic solar wind data in real-time taken from simple persistence modeling, thus improving the CME propagation forecast in more complex heliospheric conditions when solar wind can’t be assumed as constant value in the heliosphere. |
|
|