Effects of Driving Parameter Filtering on SWPSNN Model Performance

The SWPSNN (Solar Wind Plasma Sheet Neural Network) model recently developed by Swiger et al. (SpWe, 2022) predicts differential electron flux in the near-Earth plasma sheet. The model was trained using OMNI solar wind parameter data as inputs and THEMIS electron flux data as outputs. The objective of this study was to determine if the performance of the model depends on the subset of input data used.

The frequency distribution of each OMNI driving parameter was split into quartiles. Each timestamp in the output dataset was matched with a timestamp ten minutes prior in the input dataset to account for travel time of parameter variations from the bow shock to the plasma sheet. The THEMIS timestamps which corresponded to input parameter values within the desired quartile were used to filter both the input and output datasets. After running the model with the filtered datasets, model accuracy (MSA), correlation of observed and model flux, model bias (SSPB), and the model's prediction efficiency (PE) were calculated for eight logarithmically-spaced energy channels between 0.5 keV and 41 keV and graphed as a function of quartile.

The model was observed to perform best for energy channels in the one to ten keV range. Additionally, for quartiles based on solar wind velocity and solar wind number density, MSA, SSPB, and PE were much better for moderate (quartiles two and three) conditions than for extreme conditions. For all other parameters, the performance of the model did not change significantly based on the subset used. The latter result was somewhat unexpected, as subsets based on convective electric field and the z-component of the interplanetary magnetic field- both of which were ranked, during preliminary analysis, as more important in model prediction than NSW- did not alter the model's performance. When the above process was repeated for driving parameter deciles, the same trend was observed.