Enhancing water productivity prediction in solar stills using a hybrid feedforward neural network and leader gradient-based optimizer

Accurately predicting water productivity from solar stills is vital for optimizing their performance and promoting sustainable water management. This study introduces effective hybrid approach that combines a Feedforward Neural Network (FNN) with a Leader Gradient-Based Optimizer (LGBO) to enhance the predictive accuracy of solar still water production. The Gradient-Based Optimizer (GBO) is a recently advanced metaheuristic with population-based characteristics, inspired by the gradient-based Newton’s method. It employs two main operators: the Gradient Search Rule and the Local Escape Operator, along with a set of vectors to discover the search space for solving optimization problems. Therefore, it has a strong ability in global search. However, it suffers from dealing with local search problems and experiencing premature convergence. The proposed LGBO algorithm presents several advantages over the GBO algorithm including enhanced exploration capability and enabling a more comprehensive search for potentially superior solutions. The FNN, configured with 200 hidden neurons, effectively models complex nonlinear relationships, while the LGBO algorithm optimizes its weights and biases to minimize prediction errors. The optimization process employs an objective function based on the Mean Squared Error of the FNN, with LGBO efficiently navigating the parameter space to achieve optimal solutions. Key input parameters, such as solar radiation and ambient temperature, are used to train the FNN. The study focuses on conventional and stepped solar stills. The performance of the developed LGBO-FNN model is compared with other metaheuristic algorithms, including the Artificial Hummingbird Algorithm (AHA), Grey Wolf Optimizer (GWO), One-to-One-Based Optimizer (OOBO), Sand Cat Swarm Optimization (SCSO), and Gradient-Based Optimizer (GBO). Results reveal that LGBO-FNN surpasses these methods in terms of prediction accuracy, convergence speed, and robustness. This research underscores the potential of integrating advanced metaheuristic algorithms with neural networks to enhance predictive modeling in renewable energy applications. The proposed approach provides a reliable tool for predicting solar still performance, supporting the design and operation of efficient desalination systems.