Sensitivity analysis on neural networks for meteorological variable forecasting

The problem that arises in a neural network with many inputs is being able to eliminate the irrelevant ones. In the particular case of short-term weather forecasting, there are variables that may have little or no impact on the forecasts. A technique of sensitivity analysis of outputs over inputs has been applied to the trained network. Thus the most relevant inputs have been determined, as have less important inputs that can be eliminated. By employing this technique, a smaller sized neural network is obtained which also has a greater capacity for generalization.<
>