I am new to neural networks and want to apply multilayer feedforward neural network with error back propagation algorithm for the rainfall prediction. I have just monthly rainfall data of 30 years. Is this enough?? Also should I use the monthly values and enter in the model or I calculate the yearly rainfall values and use it in the model??
Please guide me that how I arrange my data, monthly values or yearly values.
I am following this paper " Prediction of Indian summer monsoon rainfall using Niño indices: A neuralnetwork approach".
Thanking in advance. Looking forward for positive responses.
Dear Asfandiyar Khan,
You have basically dataset of 360 months. Though the dataset is not enough, you can apply Bayesian Regularisation learning algorithm for training. The BR algorithm works well with less number of datasets comparative to other learning algorithms (viz., LM, SCG etc.).
Moreover, you can divide the dataset into 80:20 (training:testing) as there is no need of validation phase in BR algorithm separately.
While training with different number of layers and neurons, you will have to check for the minimum-converging values of MSE and maximum-converging values of effective parameters (gamma). This way you can decide the optimal number of layers and neurons.
Regards
Jyotindra Narayan
Hi
Use Random forest algorithm followed by feature extraction.
Use Machine learning algorithms and later go for the Fully connected layer.
Regards'
I recommend both previous answers, I just want to add one more thing, I think to test your model´s ability it could be better if you try to apply k-fold cross validation technique instead of manually data devision for your data resampling, for your samples let k=10.
Dear Bazhdar Noureddin Mohammed,
Thanks for recommending the answer. You are correct by mentioning the 'k-fold cross validation' for dividing the dataset instead of simple holdout method.
Moreover, one can use stratified k-fold cross validation where mean response value nearly equal in all k partitions.
Regards
Jyotindra Narayan
- Badges
- Science topic