This repository is based on a project that modelg the rainflow of a watershed. For that, we proposed the application of 4 Neural Networks to predict the final outflow based on temporal rain (rainfall) and on the resulting outflow of the Watersheds analized.
As the aim of this repository, we studied a Watershed located in the south of Brazil, in the state of Rio Grande do Sul. That covered several municipalities, such as: Getúlio Vargas, Erechim, Gaurama and Tapejara.
As inputs, we used the data from 3 Pluviometric Stations and 1 Fluviometric Station, in the total of 4 inputs.
As outputs, we predicted the final Outflow of the Watershed, based on the inputs.
The study was divided and predicted based on 4 different architectures, as the following: ANN, GRU, LSTM and RNN.
We considered lags to set the best parameter to be chosen for the forecasts. As that, the lags 1, 3, 5, 10 and 30 were selected.
In order to have a robust prediction and a reasonable comparison, we selected several forecasts for the output. So, we considered: t + 0, t + 1, t + 3, t + 5 and t + 10. These results were an accurate method to evalueate the models and their capacity to forecast by considering the metrics for every step.
The best rainfall model for us were set based on the evaluation of the metrics. For that, we used some classical modeling metrics, such as: RMSE, MAE and the timing of the training and fitting stage of each model. All the process could checked on the Fluxogram. to better understand. the workflow.
The workflow of this repository is shown in the following Fluxogram.
As we used several lags steps and forecasts, each Neural Network was optimized based on the best performance for the forecast 't + 0', by reaching the best proposed metrics. After that, the forecasts were conduceted: t + 0, ... , t + 10. The following table summurized the parameters for each combination.
| Neural Network | Selected Lag | Forecasts (t) |
|---|---|---|
| ANN | 10 | t + 0, ..., t + 10 |
| GRU | 10 | t + 0, ..., t + 10 |
| LSTM | 30 | t + 0, ..., t + 10 |
| RNN | 10 | t + 0, ..., t + 10 |
The results and discussion of this repository are under develpment by the authors, wich, in any case will be published and discussed under a scientific paper.