A deep learning laboratory for Total Electron Content prediction experiments, using Tensorflow 2.
Models can be found at ./models/. Some reusable layers are on ./models/custom_layers.py.
The c111 and c333 models were inspired by Boulch (2018).
BOULCH, A.; CHERRIER, N.; CASTAINGS, T. Ionospheric activity prediction using convolutional recurrent neural networks. arXiv:1810.13273 [cs], 6 nov. 2018. https://github.com/aboulch/tec_prediction/
The models on ./models/dev/gps_solutions.py were described on the article:
de Paulo MCM, Marques HA, Feitosa RQ, Ferreira MP (2023) New encoder–decoder convolutional LSTM neural network architectures for next-day global ionosphere maps forecast. GPS Solut 27(2):95. https://doi.org/10.1007/s10291-023-01442-4
python3 downloader.py Or download from: https://drive.google.com/file/d/1Sm_PiVUIabaew_3Y7sT0NWBqu7xsdHvi/view?usp=share_link
python3 ionex_samples.py
The configuration file "config.csv" is used to setup many hyperparameters for each experiment, such as chosen model, input window, prediction window, train and test datasets, among others.
The columns "batch_train" and "batch_test" on "config.csv" can be used to perform batch testing. Set them as True on the line that describes the experiment and run
python3 batch_run.py
The results will be created on the "output" folder, under a subfolder with the experiment's name.
python3 train.py
The "parameters.py" file is created during training. If you retrain the network, please remove it.
python3 evaluate.py
python3 plotresults.py
https://github.com/mauriciodev/tec_forecast/blob/main/examples/tec_forecast.ipynb