Investigations on Impact of Feature Normalization Techniques for Prediction of Hydro-Climatology Data Using Neural Network Backpropagation with Three Layer Hidden

Data normalization techniques are a very important initial stage to be carried out in order to obtain a good predictive data approach. Many researchers get different prediction and error results in each use of these data normalization techniques. Thereby, in this article discusses the accuracy rate of the seven normalization techniques at the preprocessing stage in the Neural Network Backpropagation (NNBP) architecture including decimal scaling, Z-score, min-max (there are 6 types), sigmoid, tanh estimators, mean-MAD, and median-MAD. We used two data patterns: seasonal data (rainfall) and stationary data (air humidity) that taken over the past 10 years (at 10-day intervals). We use accuracy rate parameters including number of epochs, MAE, and MSE when conducting training, testing, and predictions. The results showed that the Z-score technique was very good for the normalization of rainfall data with epochs of 10, MAE of 0.051, and MSE of 0.004. In the case of air humidity data, mean-MAD and Z-score techniques can be recommended with the number of mean-MAD technique epochs of 8, MAE of 0.013, MSE of 0.0004, while the number of epochs of Z-score techniques of 7, MAE of 0.018, and MSE of 0.0006. Thus, we conclude that when other researchers predict seasonal data or stationary data can use the Z-score technique for data normalization.