Bi-LSTM, GRU and 1D-CNN models for short-term photovoltaic panel efficiency forecasting case amorphous silicon grid-connected PV system

Abstract

Photovoltaic (PV) panels stand as a prominent solution to meet the world's growing energy demands, due to their resistance to hard climate conditions, low-cost maintenance, and long lifetime. Nonetheless, the integration of PV power into electrical grids poses a significant challenge due to its inherent intermittency. This study aims to forecast future PV power based on historical records using Bidirectional Long Short-Term Memory (Bi-LSTM), One-Dimensional Convolutional Neural Network (1D-CNN), and Gated Recurrent Unit (GRU). Various performance metrics have been used to evaluate and compare the accuracy of the three models, including mean squared error, root mean squared error, mean absolute error, max error and R-squared for evaluation. The prediction of power photovoltaic based on the values registered in the last hour was carried out. Two scenarios have been investigated, with and without nighttime values to fit the models. Results reveal that the three forecasting models provide exceptional accuracy, achieving a correlation coefficient range of 96.9–97.2% for daytime PV power prediction in both scenarios, indicating a promising potential for these DNNs forecasters for optimizing energy production and improving overall PV system efficiency. The GRU, and BiLSTM-based forecasters showed identical results in terms of RMSE, MSE and MAE in both scenarios, while the 1D-CNN based forecaster showed accurate results in the second scenario, However, despite this improvement, it still falls behind forecasters based on Bi-LSTM or GRU in both scenarios.

Keywords

• Bi-LSTM
• 1D-CNN
• GRU
• Photovoltaic power forecasting
• Time series
• Machine learning