3d-rainfall-generator

In this project we investigate if it is possible to combine the [mulGETS][doi:10.1016/j.jhydrol.2007.06.035] rainfall generator, [Stochastic Storm Transposition][http://dx.doi.org/10.1016/j.jhydrol.2013.03.003] and general data science to create long term continuous rainfall time series at multiple site locations.

Things to do:

• Implement the mulGETS model in python.
• Compare correlation of mulGETS and observation data.
• Compare annual statistics of mulGETS model to observation data.
  • Also compare low frequency variability of the model output ([This paper][https://doi.org/10.1016/j.jhydrol.2019.05.047] suggest that this are modelled poorly by this model).
  • If this is the case, investigate if a time varying Markov chain can fix this problem.
• Autoencode data from a [RainyDay][https://github.com/danielbwright/RainyDay2] output with Principal Component Analysis (PCA).
  • The autoencoding is used to determine if certain days from the RainyDay can be substituted into the mulGETS model output.
• Autoencode data from RainyDay using:
  • Densely connected neural network.
  • Deeply connected neural network.
• Develop a "plugin" to RainyDay to allow long term simulation of rainfall in time (subdaily) and space (data dependent resolution)

The mulGETS model

The general procedure and methodology is presented in [this][doi:10.1016/j.jhydrol.2007.06.035] paper, but the general outline is:

• Fit a Markov chain to every rainfall station of interest. An example of 23 years of daily rainfall is available in this github.
• Determine the correlation matrix needed, to simulate the observed correlation of either rainfall occurrence and precipitation amount.
• Establish link between occurrence index (number of "wet" station at once, relative to the interstation correlation) and average seasonal precipitation. Use this information to construct a multi-exponential distribution for each station.

Results

Figure 1 displays the observed and simulated correlation of both daily rainfall occurrences and daily rainfall amounts.

A	B
Figure 1: (A) Interstation correlation. (B) Precipitation amount correlation.

The mulGETS model is quite succesfull at simulating these two parameters. For some reason there is a small error on the right-hand figure (the precipitation amounts). Our initial suspension is that it could be caused by either: the fact that the occurence correlation matrix converges on a somewhat high error (0.446) where the mulGETS paper reports on an error in the 1e-3. The paper also mentions making adjustment to the multi-exponential distribution, which we have not done.

Overall the results presented on figure shows that the mulGETS model retains the spatial coherence of the rainfall field.