Forecasting Solar Irradiance using Hybrid Stationary Wavelet Transform- Quaternion Valued Neural Network with a Softplus AMSGrad Learning Algorithm

Efficient use of solar energy requires reliable forecasting values. In this paper, we propose a forecasting system combining the stationary wavelet transform and quaternion-valued neural networks (SWT-QVNN) and use it to forecast solar irradiance. In addition, a quaternion variant of the softplus AMSGrad learning algorithm is developed and used in optimizing the developed neural network. The proposed system was tested using irradiance data from two cities Abu Dhabi, UAE, and Tamanrasset, Algeria. It did reduce Root Mean Squared Error (RMSE) by more than 60% compared to the baseline model that uses a real-valued neural network.