Volume 51

Prediction of Daily Solar Irradiance for Solar Energy Approximation using LSTM Neural Network Roslin Mahmud Joy, Md. Faruk Abdullah Al Sohan, Afroza Nahar, and Nusrat Jahan Anannya

Abstract

Solar energy systems suffer from the unstable nature of solar irradiance. This paper proposes a novel approach to predict solar irradiance using a sequential model with a Long Short-Term Memory (LSTM) network. The model uses the mean squared error (MSE) loss function and the Adam optimization algorithm to provide robustness to noise and efficient computation of gradients. The model aims to forecast solar irradiance for 2023 in Dhaka, Bangladesh, using weather data and sunlight duration data from NASA Power and Data. GISS NASA. The dataset includes daily records from 2000-2023. The proposed LSTM model achieves 5.47% more accuracy than RandomForestRegressor in terms of root mean squared error (RMSE) to predict 1 year of irradiance data, indicating a substantial degree of precision in forecasting the objective variable. The results corroborate previous research highlighting the advantages of utilizing temporal dependencies and historical data to achieve accurate solar energy forecasts.

Keywords solar energy, deep learning, daily irradiance prediction, LSTM, Adam optimization