Hybrid ultra-short term solar irradiation forecasting using resource-efficient multi-step long-short term memory

Abstract

Accurate forecasting of solar irradiance is a key tool for optimizing the efficiency and service quality of solar energy systems. In this paper, a novel approach is proposed for multi-step solar irradiation forecasting using deep learning models optimized for low computational resource environments. Traditional forecasting models often lack accuracy, and modern, deep-learning based models, while accurate, require substantial computational resources, making them impractical for real-time or resource-constrained environments. Our method uniquely combines dimensionality reduction via image processing with an LSTM-based architecture, achieving significant input data reduction by a factor of 4250 while preserving essential sky condition information, resulting in a lightweight neural network architecture that balances prediction accuracy with computational efficiency. The forecasts are generated simultaneously for multiple time steps: 1 minute, 5 minutes, 10 minutes and 20 minutes. Models are evaluated against a custom dataset, spanning across more than 3 years, containing 1 min samples encompassing both all-sky imagery and meteorological measurements. The approach is demonstrated to achieve better forecasting accuracy, namely a forecast skill of 10 % compared to persistence, and a significantly reduced computational overhead compared to benchmark ConvLSTM models. Moreover, utilizing the preprocessed image features reduces input size by a factor of 6 compared to the raw images. Our findings suggest that the proposed models are well-suited for deployment in embedded systems, remote sensors, and other scenarios where computational resources are limited.