Week 14

Until now, linear models were considered. However, flood is highly complex event and may contain non- linearities. In order to handle them, a neural network model was used.

Model Architecture used (Initially we choose to use this assuming this model will perform well. But later we found out this model overfit and does not generalize well for a given new scenario)

Model consist of 3 hidden layers with 50 neurons ( later found out this is too much )

Relu activation function is used in hidden layers.

RMSProp algorithm was used as the optimizer

At the time of initial creation, we only had 10 days worth data. They were split into three as,

training = 60 % ( training and validation sets are split randomly)

validation = 20 %

test = 20 % ( last 20 % is chosen as test data)

Same neural network architecture trained using two different data sets also.

using only sub rivers as input variables
using same features used in regression model

From the learning curve, it is visible that number of epochs can be reduced. Also, suspicion of overfitting arise after seeing this.

Since RMSE of test set and entire data set is almost similar, for this occasion model works well.

From two models, model which used correlated data which used in the regression analysis provide best result

Page updated

Google Sites

Report abuse