COMPARATIVE EVALUATION OF STANDARD AND OPTI-MIZED BACKPROPAGATION NEURAL NETWORKS IN RAINFALL PREDICTION

Rainfall needs to be predicted with accuracy so that there can be good water resource management, agricultural planning and mitigation of the disasters etc. However, with the messy nature of meteorological data, the simplistic methods of forecasting usually used are not able to replicate the non-linear and complex patterns that these tend to carry. This paper examined optimized BPNNs in rainfall forecasting and compare them with as well as improved BPNN models. Other methods like, Momentum, adaptive learning, learning rate tuning were used to improve the performance of the BPNNs. To guarantee robustness and cross-conditions generality, the models used in the following bigger study were trained and tested on a wide term historical weather data, over a course of 30 years (1990-2023) of a Mahanadi River Basin that is located in the state of Chhattisgarh in India. The training was done with 10 neural network architectures and a model with 10 optimization strategies. They were compared with performances of a baseline (standard BPNN) measured in such terms as Mean Squared Error (MSE), root mean square error (RMSE), and correlation coefficients. The result of the study was proved by the dominating predictive accuracy and speedy computation time using all optimized BPNNs, that is, by available Stochastic Gradient Descent (SGD) with DyNAM and momentum in comparison with the baseline model. These are the improvements in the speed of convergence and reduction in rates of prediction error. The research shows a promising use of confined use of the neural network in rainfall forecasting dragon architecture. The results we obtained promote the adoption of advanced optimisation techniques to train neural networks and therefore do better in terms of reliability and on-time rainfall forecasts. In a further research, we are going to pay attention both to adding more variables of meteorological character and experiment with alternative architecture of neural network to enhance a greater improvement on the forecasts.