NWPsolarNet: A Scalable Deep Learning Framework for Medium-Term Solar Forecasting across Europe

(DNN) framework, designed for regional medium-term solar power forecasting. By utilizing a comprehensive dataset of power plants across Europe spanning January 2022 to July 2025, we introduce a physics-informed quality control pipeline that validates solar installations based on physical consistency—specifically the diurnal production cycle and production-to-irradiance ratios—rather than relying solely on unreliable metadata.

Central to this methodology is a robust normalization strategy that mitigates the impact of metadata errors and equipment degradation. Instead of using rated peak power, we derive a scaling factor based on the 10th percentile of the ratio between theoretical clear-sky irradiance and actual production, a method which effectively aligns output with the physical potential of the site. The forecasting architecture itself fuses meteorological inputs from the Icosahedral

Nonhydrostatic for Europe (ICON-EU) Numerical Weather Prediction (NWP) system with theoretical clear-sky Global Tilted Irradiance (GTI). This design employs an NWP Encoder to extract spatio-temporal features from atmospheric variables, such as cloud cover and wind components, and integrates them with geometric references to generate hourly predictions up to 120 hours ahead. Crucially, the model learns a generalized mapping from local meteorological conditions to solar output, decoupling the forecasting capability from sitespecific instrumentation and enabling spatially continuous predictions on a grid.

Experimental validation demonstrates that NWPsolarNet substantially outperforms both operational and baseline models on a seasonally diverse test set. In a 48-hour forecast horizon, the framework achieved a 28.0% reduction in Mean Absolute Error (MAE) compared to the operational Quartz Solar Forecast. Over the full 120-hour horizon, the model reduced MAE by 14.6% against a Light Gradient Boosting Machine (LightGBM) baseline and consistently outperformed adapted versions of PVNet and TimeXer, while substantially eliminating systematic bias. Furthermore, the proposed data-driven normalization strategy was shown to improve MAE by approximately 4.0% compared to standard capacity-based normalization, confirming that reliance on reported capacity introduces bias. Ultimately, this framework establishes a scalable, robust solution for managing the inherent variability of continental-scale renewable energy grids.