Work

A preliminary LSTM-IDW model for spatiotemporal hourly solar radiation estimation in tropical regions

Year: 2025

Type: article

Abstract: • A novel LSTM-IDW approach addresses extreme solar radiation fluctuations in tropical regions. • IDW spatializes GHI estimates, addressing the limitation of LSTM, which is point-based. • Using satellite data as input enhances the accuracy of machine learning models for solar radiation estimation. • Pearson correlation serves as a tool to identify optimal input parameters for machine learning models. • Optimal LSTM inputs to estimate hourly solar radiation include air temperature, relative humidity, wind speed, and solar zenith angle. The use of renewable energy, such as solar power, has the potential to mitigate the negative impacts of fossil fuel consumption. West Java Province holds significant potential for solar-based electricity development. This study aims to estimate hourly solar radiation, addressing extreme fluctuations in intensity within the study area. Solar radiation estimation is performed using a Long Short-Term Memory machine learning model. The model uses data from eight measurement stations operated by the Badan Meteorologi, Klimatologi, dan Geofisika, recorded from 2022 to 2023, along with satellite imagery from the Geo-KOMPSAT-2A satellite to improve accuracy. Spatial interpolation using the Inverse Distance Weighting method is applied to estimate the spatial distribution of solar radiation, addressing gaps in previous studies that overlooked spatial aspects. The results indicate that input selection based on Pearson correlation analysis plays a role in influencing model accuracy. The best-performing model, which incorporates Air temperature, Relative humidity, Wind speed, Solar zenith angle, and Raw satellite pixel value as input variables, achieves an RMSE of 149.46 W/m² and an rRMSE of 39.99%, with overall rRMSE ranging from 39.99% to 44.05% and rMBE between − 0.44% and 10.33%. Inverse Distance Weighting transforms point-based Global horizontal irradiance estimates into continuous spatial data, but accuracy variations across stations, particularly in high-altitude areas, limit its effectiveness. These findings suggest that hybrid machine learning models or advanced spatialized techniques should be considered for future research. Despite its limitations, this study contributes to improving solar radiation estimation and spatial analysis, supporting renewable energy development in West Java.

Source: Renewable and Sustainable Energy Transition

DOI: 10.1016/j.rset.2025.100105