Optimizing the deep learning parameters using metaheuristic algorithms for PV-battery-based DC nanogrids

Abstract

This paper presents an implementation of metaheuristic algorithms to optimise deep learning neural networks (DNNs) for sensor-less control of photovoltaic (PV) Converters in DC nanogrids. Using a Fixed Forward Neural Network (FFNN), it estimates PV output current (IPV ) based on three days of real data. Given the vulnerability of current sensors in DC system measurements, accurately replicating current sensor data is vital. The data exhibits dynamic nonlinear relationships with inputs like solar irradiance, temperature, and voltage. The study assesses the effectiveness of Evolutionary Mating Algorithm (EMA) and Sand Cat Swarm Optimisation (SCSO), comparing them with Adaptive Moment Estimation (ADAM), Genetic Algorithm (GA), and Particle Swarm Optimisation (PSO). The research leverages these metaheuristic algorithms to optimise machine learning integration, addressing regression estimation problems, and enhancing system reliability by eliminating sensors. Key findings indicate that the EMA-DNN combination achieved remarkable performance, with the lowest Mean Squared Error (MSE) of 0.5906, the lowest Mean Absolute Error (MAE) of 0.4680, and the lowest Mean Absolute Percentage Error (MAPE) of 12.1780%. These results offer valuable insights into how the metaheuristic-DNN approach can solve regression estimation problems and reduce the number of sensors in PV battery-based nanogrids’ control layers.