Solar Irradiation Forecast Enhancement Using Hybrid Architecture

Abstract

Power balancing at the grid is much more involved process due to the fact that solar power generation is primarily weather dependent, as it is relied on solar irradiation, which is very volatile and unpredictable. Accurate solar irradiation forecasting can significantly increase the performance of solar power plants. This research is motivated by the current advancements in deep learning (DL) models and its practical use in the green energy field. The proposed model combines two DL architectures: convolutional neural network (CNN) and long short-term memory (LSTM). The effectiveness of the same is analysed by comparing with recurrent neural network (RNN) family architectures. The RNN family models are Long Short Term Memory (LSTM), Bi-directional LSTM (Bi-LSTM), Gated Recurrent Unit (GRU), and Bi-directional GRU (Bi-GRU). The simulations are conducted on a publicly available data set from Desert Knowledge Australia Solar Centre (DKASC), Australia. A meteorological station across the Northern Territory (NT Solar resource) collects high resolution solar and climate data from Darwin location, which is used for the experiment. From the results, it is evident that each of the bidirectional model outperform its unidirectional equivalent architectures. However, the hybrid network (CNN-LSTM) outperforms all the individual models as per the error metric analysis. © 2023 IEEE.