Application of PCA method to weather prediction task

A method of short-term weather forecasting based on artificial neural networks is presented. Each training sample consists of a date information combined with meteorological data from the last three days gathered at the meteorological station in Miami, USA. The prediction goal is the next day´s temperature. Prediction system is built based on multilayer perceptron network trained with backpropagation algorithm with momentum. The average prediction error of the network on the test set equals 1.12°C. The average percentage prediction error is equal to 5.72%. The results are very encouraging and provide a promise for further exploration of the issue. The so-called correlation ratio δ between predicted and real changes (trends) is equal to 0.7136. Relatively high value of δ additionally confirms good quality of presented results. Experimental results of application of the principal component analysis method at the stage of pre-processing of the input data are also presented. In that case the average prediction error and the average percentage prediction error are equal to 1.41°C and 7.93%, respectively. In order to explain the reasons of the poorer results obtained with the PCA method a closer look at the principal components defined by the network is presented. Possible reasons of the PCA failure are pointed out.