Forecasting wavelet denoised global horizontal irradiance using attention-based long short term memory network: A case study of South Africa

Abstract

Microgrids are becoming a crucial component of the electricity grid in dependability, economics, and environmental sustainability. Microgrids rely heavily on renewable energy sources. From an engineering standpoint, anticipating short-term solar generation is a critical challenge in microgrid planning and design. Anticipating solar power is heavily reliant on forecasting sun radiation. Short-term solar radiation forecasting may also be used to estimate the energy potentials of photovoltaic (PV) panels impacted by degradation rates. A comparison of multiple models, namely the Autoregressive Integrated Moving Average (ARIMA), Long Short Term Memory (LSTM), Attention-based LSTM and a hybrid Attention-based LSTM-ARIMA for forecasting 5-day ahead 1-minute solar radiation is performed in this study. The best model for forecasting Global Solar Radiation(GHI) from Richtersveld station is ARIMA with MAE = 0.782 and RMSE = 1.271, followed by hybrid model with MAE = 4.120 and RMSE = 4.987. For Stellenbosch University station, attention LSTM was the best with MAE = 1.512 and RMSE = 1.640, followed by hybrid with MAE = 2.011 and RMSE = 2.511. The hybrid attention-based LSTM-ARIMA model on the USAid Venda station was the best fitting model with RMSE = 7.383 and MAE = 14.1293, followed by LSTM with MAE = 7.817 and RMSE = 8.444. Comparing the results on nonwavelet denoised and wavelet denoised, models performed better on wavelet denoised data. ARIMA model was the best with MAE = 0.194 and RMSE = 0.542, followed by hybrid with MAE = 2.176 and RMSE = 2.308.