LSTM based Model for Weather-based Solar Irradiance Prediction for Long-Term PV Energy Planning

Authors: Anushka Bandara, Mahela Pandukabhaya, Keshawa Ratnayake, Roshan Godaliyadda, Parakrama Ekanayake, Janaka Ekanayake.

Abstract: The increasing global emphasis on sustainable energy solutions has elevated the role of solar power as a prominent renewable energy source. Photovoltaic (PV) systems face the challenge of the presence of an uncertainty in long-term energy production due to the stochastic and intermittent nature of solar irradiance. Accurate forecasting of solar irradiance is essential for planning and optimizing PV plant performance. Traditionally, long-term solar irradiance forecasting relies on historical irradiance data. However, such data is often limited to short durations (2–5 years) at prospective PV plant sites, whereas reliable forecasts may require 15 or more years of data. To address this limitation, we propose a novel deep learning-based methodology that leverages Long Short-Term Memory (LSTM) networks enhanced with attention mechanisms to reconstruct historical solar irradiance using available long-term meteorological data. This approach exploits the known correlation between weather parameters and solar irradiance, filling in historical data gaps and enabling long-term forecasting. The proposed model predicts missing irradiance values and also facilitates extended solar energy generation forecasts essential for site planning and investment decisions. A case study is presented to demonstrate the effectiveness of the model, and the proposed model achieved an RMSE of 66.9 W/m^2 and R^2 = 0.9578 for Colombo and 56.7 W/m^2, R^2 = 0.9683 for Hotevilla, outperforming the conventional LSTM by approximately 3-5\%. Hence, this study contributes to the growing body of research in solar forecasting by offering a practical and accurate solution for long-term irradiance prediction using limited observed data.

Keywords: solar irradiance prediction, long-short term memory (LSTM), photovoltaic plant, attention, deep learning.