Speaker
Description
The first DTM_nrt (near real time), developed in 2013, used exospheric temperature corrections inferred from observed densities essentially for debiasing. We have started the development of an updated DTM_nrt, which uses the more precise DTM2020 as background model, and more precise density data (CHAMP, GRACE, GOCE, Swarm-A, and GRACE-FO) calculated by TU Delft, as well as updated Stella daily-mean densities. This presentation shows the significant progress since the last ESWW in 2024.
The new DTM_nrt model is not directly trained on temperature values, but on temperature variations relative to an 81-day trailing mean. This approach significantly reduces systematic biases across temperature measurements derived from different satellites. As a result, the model achieves consistent predictive skill whether validated on the training satellite or on independent satellites. Several deep neural network architectures were compared in this study; the best-performing architecture, which was ultimately retained, is a Gated Recurrent Unit-based autoencoder.
A key result for 24-hour temperature variation forecasting is that the 10.7 cm solar flux was found to be an irrelevant input for this neural network, whereas the inclusion of predicted daily-mean ap (those actually used in CNES operations) emerged as a critical driver of forecast accuracy. However, this dependence also points to current limitations, since historical predictions of 3-hourly ap values could not be obtained, leaving opportunities for further improvement. The choice of training and validation periods leverages the periodicity of the solar cycle as well as prior knowledge of temperature amplitudes. Although only two complete solar cycles were available, this setup, focused on training speed, enabled a validation dataset representative of the full range of available conditions, while mitigating common issues of temporal causality in forecasting tasks.
Overall, the new DTM_nrt model outperforms DTM2020 by more than 50% for a 24-hour forecast when predicted drivers are used in both models.
