Deep learning and optimization algorithms based PV power forecast for an effective hybrid system energy management

Abstract

Economic and demographic development has led to energy consumption increment around the world. The utilization of renewable energies is the best solution to offset this increase. The photovoltaic energy is widely used around the word through grid connection or standalone systems. Climatic changes can influence the generated power and the operating management strategy. Thus, photovoltaic power forecasting is very crucial to ensure stability. Reliable prediction accuracy provides information to ensure an efficient energy management of a PV/Battery/Diesel hybrid system. This paper presents a comparative study among various photovoltaic power prediction methods based on deep learning and optimization algorithms. Three topologies are outlined: the feed forward neural network with Particle- Swarm-Optimization tool (FFNN-PSO), the long short-term memory recurrent neural network (LSTM) and the bidirectional LSTM network with the Bayesian Optimization Algorithm (BiLSTM-BOA). The predictors’ accuracy evaluation is done via statistical metrics. The simulation analysis show the performance of the BiLSTM-BOA on photovoltaic power forecasting. The application of the management algorithm using the forecasted PV power proved a high level of efficiency for both clear and disturb days. It maximizes the contribution of the renewable resource, minimizes the utilization of the batteries and the diesel generators and ensures load supply continuity.