Abstract
Accurately predicting solar power is essential for ensuring electric grid reliability and integrating renewable energy sources. This paper presents a novel approach to probabilistic solar power forecasting by combining Convolutional Neural Networks (CNN) with Gated Recurrent Units (GRU) into a hybrid Quantile-CNN-GRU model. The proposed model generates intra-day probabilistic quantile forecasts and is rigorously evaluated using datasets from geographically and climatically diverse regions and hemispheres: the Netherlands (temperate maritime climate), Alice Springs (arid desert climate), and Hebei (humid subtropical climate). These datasets cover varied temporal horizons (1-hour, 6-hour, 12-hour, and 24-hour predictions) and seasonal conditions (summer, fall, spring, and winter), highlighting the model's adaptability to different scenarios. The performance of the proposed Quantile-CNN-GRU model is benchmarked against state-of-the-art deep learning models, including standalone quantile-based architectures such as Quantile-GRU and Quantile-Long Short Term Memory (LSTM). A comprehensive evaluation framework is applied, employing probabilistic tools like the Continuous Ranked Probability Score (CRPS) for assessing forecast reliability, sharpness, and reliability diagrams with consistency bars to evaluate the calibration of the predictions. Results demonstrate that the proposed Quantile-CNN-GRU model consistently outperforms its counterparts in terms of CRPS, across varying forecast horizons and seasonal conditions. To further enhance performance, a multivariate case study incorporating exogenous inputs, specifically Numerical Weather Prediction (NWP) data, is conducted. Through sensitivity analysis, the influence of these additional inputs on forecast horizons and seasonal variability is systematically explored. The study reveals that integrating NWP data significantly improves the model's predictive skill, particularly for longer forecast horizons and during transitional seasons like spring and fall, when solar variability is higher.