The effect of some normalization methods on neural networks and robust methods with the presence of outliers

This research aims to use some statistical transformations such as Z-Score and MIN_MAX to see how these transformations affect the performance of some robust methods and neural networks when there are outliers in the data and to compare the robust methods (LTS, MCD, and MM) and some neural networks including (RNN, GRU, and LSTM) with different activation functions represented by (Relu, Elu, and Selu).The research sample included 3000 private sector electricity generators in Iraq for the year 2021 taken from the Central Statistical Organization.The comparison was made using the mean square error (MSE) and using the statistical programs Python, R, and Excel.The results showed that both normalization techniques significantly improved the performance of the models, especially Min-Max normalization is the best for all (robust methods and neural networks) and especially the superiority of neural networks RNN, especially with deeper structures, showing good performance across different activation functions. As for the robust methods, the MM method was consistently the best, giving the lowest mean square error across all normalization techniques.