1

I have a dataset of the daily temperature for a couple of years. The data is in the interval form, including daily high temp and daily low temp.

I want to do a forecasting of the data, and I recently read several paper mention that the multilayer perceptron have the advantage to do this. However, after reading the paper I still got puzzled. I know in order to conduct it, I will need to have input, hidden layer and output. But in Matlab, though I have the code already, I still don't know how to simulate it. What should I put as its input and output, should I put the interval data as the input and output? And how can I choose hidden layer?

user456649
  • 11
  • 1

1 Answers1

0

The input in an MLP network is the input feature data that you are trying to predict the outcome of. The output is what you are trying to predict. For the hidden layer that will determine how well it predicts, which you want as large as it needs to achieve reasonable prediction results. Going too large and it just memorizes the data rather than generalize on a pattern when training is run.

For example, if your input layer would be what day of the year it is (1-365), what the high was of the day, and what the low is of the day. And I assume is what the high and low temperature would be for the next day?

The more relevant input features you have the better the network will be.

Stephen Flynn
  • 105
  • 9
  • Thank you very much for your response. I think my question for MLP is more basic or naive. In the matlab code that I found, it asks for the input, hidden-layer and output. I think I have dome problems in understanding how to use the information. For example, I have the data set 365(days) by 5(years), so here, what can the MLP forecast, the next year's temperature? And if it can give a forecasting about the next year's temperature, what should I put into the function as the input data, 365 by 1 (first year's data) or 365 by 5 (all five yesr's data)? and how about the hidden layer and output? – user456649 Jul 26 '18 at 22:42
  • The input layers needs to be the size of the number of data points you are predicting off of. That would make 1 input node per data point. The hidden layer size determines the size of the network. Too small and it will not learn, and too large is not good either, just try it with a few different sizes of hidden layers and see how well it predicts. The output layer is as large as the number of data points you are trying to predict, which in this case is 1 being the temperature next year. Are you aware of how to train an MLP? – Stephen Flynn Jul 27 '18 at 13:05