A Recurrent Neural Network for Disruption Prediction

This repository contains a neural network to predict disruptions from bolometer data in two different ways:

• Time to disruption (ttd): predicts the remaining time towards an impending disruption. In this case, the model is trained on disruptive pulses only.

• Probability of disruption (prd): predicts whether a pulse is disruptive or not. In this case, the model is trained on both disruptive and non-disruptive pulses.

Requirements

• Keras 2.1.2, TensorFlow 1.4.1 (minimum)

• Configure ~/.keras/keras.json as follows:

{
    "epsilon": 1e-07,
    "floatx": "float32",
    "backend": "tensorflow"
    "image_data_format": "channels_last",
}

Instructions

• Run dst_bolo.py to get the disruption time and bolometer data from every pulse.

  • This script will only run on a JET computing cluster (e.g. Freia).

  • The disruption time is is obtained from the JET disruption database.

  • The bolometer data is down-sampled from 5 kHz to 200 Hz (1 sample every 5 ms).

  • An output file dst_bolo.hdf will be created.

• Run model_train.py on each folder (prd and ttd) to train the corresponding model.

  • Training will finish automatically once the validation loss no longer improves.

  • The model and its weights will be saved to model.hdf.

  • A log file with the loss and validation loss will be saved to train.log.

• During training, run plot_train.py to see how the loss and validation loss are evolving.

  • The script will also indicate the epoch where the minimum validation loss was achieved.

• After training both models, run model_predict.py to run the models on the test set.

  • This script will plot the time to disruption and the probability of disruption for each test pulse.

  • Each plot will be saved to an images/ folder as separate PNG file.

  • An output file dst_pred.hdf will be created with the output of both networks for each validation pulse.

References

• D.R. Ferreira, P.J. Carvalho, H. Fernandes, Deep Learning for Plasma Tomography and Disruption Prediction from Bolometer Data, IEEE Transactions on Plasma Science, 2019 (to appear)