This module is designed to use Long-short Term Memory (LSTM) Nerural Netowrks - for feature identification and gapfilling of flux time series
The general workflow is as follows:
-
Bayseian optimisation of a "full model" with all potential factors using gaussian process regression to identify the intial "optimal" number of nodes and timesteps
-
Itterative construction of models for feature identification from a pool of p potential features (f) using the N & T values identified in step 1)
A) Start with 1 factor model, and loop through factores f1 ... fp, select the factor
$f_{min}$ that yield the model with the losest MSE B) Increase the model size to 2 and train iteratively on feature$f_{min}$ and factors f1 ... fp-1, selecte factor$f_{min2}$ C) Repeate B until MSE stops decreasing, select the smallest model where MSE is lower than previous models by some score ... 95% CI?? -
Repeate step 1 to optimize final model. Probably benificial?? A) Iteratively deconstruct final model??
-
Use optimized model to fill time series.