Model Compression for Very Short-Term Solar Irradiance Forecasting Using Infrared Sky Images
Ifran Rahman Nijhum, Arnav Jaitly, David Lillis and Soumyabrata Dev
In Proceedings of the 28th Irish Machine Vision and Image Processing Conference (IMVIP 2026), pages 42--49, Maynooth, Ireland, Aug. 2026.
Abstract
Accurate short-term solar irradiance forecasting is essential for reliable photovoltaic power integration. While convolutional neural networks (CNNs) have demonstrated strong performance for sky-image-based forecasting, deploying deep learning models on resource-constrained edge devices remains challenging due to their computational and memory requirements. However, the impact of model compression on solar forecasting accuracy and deployment efficiency has not been systematically investigated. In this work, we present a comprehensive benchmark of structured and unstructured pruning for infrared sky image-based solar irradiance forecasting using infrared sky images from the Girasol dataset. We evaluate four widely used CNN architectures ResNet-18, EfficientNet-B0, MobileNetV2, and MobileNetV3-Small along with a lightweight custom baseline model (TinyCNN) across multiple pruning levels ranging from 20\% to 80\%. Models are assessed using forecasting accuracy, model size, parameter reduction, and inference latency. Experimental results show that MobileNetV2 achieves the strongest baseline forecasting performance, with an RMSE of 37.11 W/m2. Structured pruning yields substantial parameter and latency reductions but is architecture-dependent, with some models degrading sharply even at the lowest pruning level tested. Unstructured pruning preserves accuracy more consistently across architectures and substantially reduces compressed checkpoint size, but provides little to no inference speed improvement on conventional hardware. TinyCNN tolerates the highest compression of any architecture under both strategies, though from a much higher baseline error. Overall, the two strategies offer complementary trade-offs structured pruning favoring latency and parameter reduction, unstructured pruning favoring accuracy preservation demonstrating that meaningful compression is achievable for CNN-based solar irradiance forecasting, with implications for edge deployment.