Design and evaluation of a dual-spectrum low-cost sky imager for enhanced solar prediction

Solar energy is one of the largest sources of renewable power. However, efficiently integrating solar power into energy grids remains difficult due to its highly intermittent nature. This challenge can be mitigated through accurate solar irradiance estimation and forecasting, which are essential for reliable photovoltaic (PV) and microgrid operation. Although considerable research has been conducted on solar data acquisition systems, most existing platforms rely on a single spectral modality, such as visible-light (RGB) or infrared imaging, and many remain relatively expensive. In this work, we design and evaluate a low-cost dual-spectrum data acquisition system for synchronized capture of RGB, near-infrared (NIR), and environmental data to support solar irradiance estimation. The proposed platform integrates an RGB camera and an NIR imaging module with Raspberry Pi 4B units, together with a BH1750 light sensor and an SHT40 temperature–humidity sensor, all housed in a custom IP56 weatherproof enclosure for outdoor operation. The system records data at one-minute intervals with synchronized multimodal acquisition. To assess the predictive utility of the collected data, we employ a lightweight convolutional neural network (CNN) as a proof-of-concept irradiance estimation model trained separately on RGB and NIR imagery, followed by post-training fusion. Among the evaluated ensemble strategies, the harmonic ensemble achieved the best performance, reducing the mean squared error (MSE) by 832 absolute points ( ≈ 30 % ) relative to the RGB-only baseline and by 1052 absolute points ( ≈ 35 % ) relative to the NIR-only model. These results demonstrate that the proposed dual-spectrum platform provides a compact, affordable, and effective foundation for multimodal solar monitoring, while highlighting the complementary value of RGB and NIR sensing for irradiance estimation.