Neural Network-Based Estimation of Onshore and Offshore Wind Power Generation in Germany

Abstract

Wind power generation has become increasingly important in Germany, where the installed capacity of wind sources currently reaches 75.75 GW. Accurate wind power generation forecasting is essential for predicting spot prices and grid management. Deep learning approaches have emerged as promising alternatives for renewable energy forecasting due to their ability to capture complex nonlinear relationships in time-series data. This study develops and compares two neural network architectures for estimating hourly wind power generation using Copernicus ERA5 reanalysis data in Germany. Models in this study are simple regression model with separate onshore/offshore branches and Convolutional Neural Network (CNN). For training, we used the wind components (u₁₀₀, v₁₀₀) from ERA5 reanalysis, representing the eastward and northward wind speeds, matched with hourly generation data and monthly data of installed capacity for wind offshore and onshore from Energy-Charts transparency platform. Training and validation data span from 2019 to 2024, testing is performed on first 5 months in 2025 to evaluate performance. Data from Copernicus are standardized, while data of installed capacity and generation are normalized. The regression model achieved validation/test Mean Square Error (MSE)/Mean Absolute Error (MAE) of 0.0765/0.0937; the CNN achieved 0.0497/0.0517, outperforming the simple model and demonstrating the effectiveness of convolutional layers for capturing spatial wind patterns. Estimation of power generation was better in onshore case than offshore case in both models. The results highlight the potential of CNNs for wind power forecasting and suggest further improvements by incorporating additional meteorological variables such as pressure, temperature, and humidity and exploring other architectures, such as CNN Long Short-Term Memory Networks.