Development of a stochastic model of RBF neural network for forecasting relative humidity rates

The multiplicity of models aimed at predicting parameters makes it difficult to find the one that performs best in terms of prediction accuracy, so it is necessary to test several models. In this sense, the present study focuses on the development of a neural network architecture based on the RBF radial basis function, with ideal performance to prediction humidity variations in the city of Tangier, and compares it with the exponential regression method, based on the two statistical indicators, the correlation coefficient and the mean square error. For this, we use a database of the city of Tangier, consisting of eight factors, recorded in the period from January 1985 to December 2022, which counts 13869 days, defining the variations of 7 input parameters: temperature, shortwave radiation, direct shortwave radiation, total precipitation, evapotranspiration, vapor pressure deficit and wind speed, and the dependent variable, which describes the relative humidity. The optimization of the learning rate, the number of neurons in the hidden layer, and the learning algorithm demonstrates that the RBF architecture [7-14-1], which utilizes $50 \%$ of the database for learning, and the Levenberg-Marquardt learning algorithm, produces a highly robust performance. The nonlinear relationship between the explanatory parameters and the dependent variable was established with an accuracy of MES =0.0039 and R= 0.977, which outperformed the exponential regression model (MES = 8.505 and R=0.942). The optimized RBF model is therefore highly effective in fitting the new data, with an extremely low probability of error in the prediction of relative humidity in Tangier.