Enhancing relative humidity modelling using L2 regularization updates

Abstract

This study explores L2 regularization to mitigate overfitting in artificial neural networks (ANNs), focusing on the regularization coefficient, Lambda, and its effect on data distribution and multi-layer perceptron (MLP) performance. Meteorological data from Tangier (1985-2022) with eight variables influencing relative humidity were analyzed using principal component analysis (PCA) and self-organizing maps (SOM). PCA identifies key correlations, such as between total precipitation and relative humidity or vapor pressure and temperature, but struggles with non-linear relationships. SOM complements PCA by highlighting data structure nuances and detect complex correlations. L2 regularization, particularly with Lambda = 0.01, effectively reduces data complexity and dispersion, preventing overfitting while enhancing prediction accuracy. Adjusting Lambda during training optimizes weight biases in Kohonen and MLP networks, improving model performance and enabling precise relative humidity prediction. This strategy demonstrates the value of combining PCA, SOM, and L2 regularization for meteorological modelling.