Hydro-NODE

Torch version for Hydro-NODE [https://github.com/marv-in/HydroNODE]

Environment requirement

• python=3.7
• torch=1.13.1
• torchdiffeq=0.2.3
• pytorch-lightning=1.8.6

How to run the experiment

• step 0, run M0_optimization to get the best parameters for the model
• step 1, run M0_main.py to get the train data, save in the data
• step 2, run M50_main.py for pretraining and training the M50 model
• step 3, run M100_main.py for pretraining and training the M100 model

Note

• utils provides some training methods, such as train, forecast and base learner object BaseLearner

My own idea

A more neural training method for M50 and M100

• Batch Training

  • How does the batch size affect the accuracy ?

    • incorrect nse evaluate for each batch training (is in consistent with the total series nse)

  • How to determine the S0 for each batch

    • constant S0

    • optimize S0

    • warm up method

    • Neural ODE

• Neural hydrology parameter

  • hyperparameter (hyperparameter searching)

  • inner parameter (gradient optimization)

• impact factor

  • interpolate methods