Eliminating Meteorological Dependencies in Solar Power Forecasting: A Deep Learning Solution With NeuralProphet and Real-World Data

Abstract

Forecasting solar power generation is essential for efficient energy management and grid stability. However, existing predictive models often rely on external datasets, such as meteorological and sensor data, to make accurate predictions. This dependency introduces complexities and limits their application in data-sparse scenarios. In this study, we propose a novel forecasting approach based on the NeuralProphet algorithm, a deep learning model that predicts solar power generation solely from its historical data, eliminating reliance on additional input data. To evaluate the proposed approach, we conducted two case studies. The first utilized a 10-month dataset from a 1.2 kW small-scale solar power unit at Bursa Technical University's Smart Grids laboratory, recorded at 15-minute intervals. Despite the limited dataset, the model achieved an R-squared value exceeding 0.74, demonstrating promising predictive capability. The second case study applied the NeuralProphet-based model to a large-scale dataset of nationwide solar power generation in Germany, spanning five years and collected at 15-minute intervals. Models trained on this dataset achieved R-squared values exceeding 0.99, highlighting the algorithm's capacity to effectively capture seasonal and temporal patterns at a national scale. Our results indicate that the NeuralProphet-based forecasting approach offers a viable and efficient alternative for solar power prediction, achieving high accuracy without external data dependencies.