Solar Irradiance Forecasting Performance Enhancement Using Hybrid Fuzzy-Based CNN-BiLSTM-Transformer Model

Abstract

Accurate forecasting of solar irradiance plays a vital role in optimizing solar energy utilization, but it remains a challenging task due to high variability and uncertainty caused by fluctuating atmospheric conditions. Traditional forecasting techniques often fail to capture nonlinear patterns and long-term dependencies effectively, leading to reduced prediction accuracy. Although recent advancements in deep learning have shown superior performance in time series forecasting, their integration with fuzzy time series (FTS) methods has been relatively unexplored. To bridge this gap, this article introduces an innovative FTS-based forecasting framework that integrates deep learning with fuzzy modeling to overcome these limitations. The proposed model combines Convolutional Neural Networks, Bidirectional Long Short-Term Memory, and Transformer architecture (CNN-BiLSTM-Transformer) with a fuzzy model employing Gaussian membership functions to process historical solar irradiance data. This approach enables the model to generate accurate forecasts while managing both first-order and high-order fuzzy relations. Additionally, the Sine Cosine Optimization algorithm is used to fine-tune the model’s hyperparameters, further enhancing its performance. The effectiveness of the model is validated through experiments using real-world solar irradiance datasets collected from three different websites for Mangalore location. The results demonstrate that the proposed model achieves a Mean Absolute Error (MAE) of 21.805 W/m2, a Root Mean Square Error (RMSE) of 93.089 W/m2, and an R2 score of 0.981 for one-step-ahead forecasting using NREL data, outperforming the performance of state-of-the-art methods and highlighting its effectiveness in solar irradiance forecasting. © 2013 IEEE.