A Comparative Deep Learning Approach for Short-Term Wind

Power Generation Prediction

Mona Ahmed Yassen 1,2,* Mohamed Gamal Abdel-Fattah 1,2 Islam Ismael 3

Hossam El-Din Moustafa 1,2

1Department of Electronics and Communications Engineering, Faculty of Engineering, Mansoura University, Mansoura 35516, Egypt

2Faculty of Artificial Intelligence, Horus University, Egypt

3Department of Electrical Engineering, Faculty of Engineering, Mansoura University, Mansoura 35516, Egypt

Received: December 24, 2025 Revised: February 12, 2026 Accepted: April 16, 2026 ⋆ Corresponding author

ABSTRACT

Accurate wind power forecasting is essential for reliable renewable energy integration, grid stability, reserve scheduling,

and wind farm operation because turbine output is highly variable and strongly influenced by meteorological

conditions. However, forecasting wind power remains challenging due to the nonlinear relationship between weather

variables and power generation, the temporal dependency of hourly observations, and the circular nature of wind

direction data. This study aims to develop and compare deep learning models for predicting normalized wind turbine

power output using a field-based hourly dataset collected from an operational wind energy site starting from January

2, 2017. The dataset includes temperature, relative humidity, dew point, wind speed at 10 m and 100 m, wind

direction at 10 m and 100 m, wind gusts, and normalized turbine output. Five predictive models, namely LSTM, RNN,

GRU, CNN, and Dense neural networks, were trained and evaluated after applying data preprocessing procedures,

including missing-value handling, feature scaling, temporal alignment, and wind-direction transformation. Model

performance was assessed using MSE, RMSE, MAE, MBE, correlation coefficient (r), coefficient of determination

(R2), RRMSE, NSE, and WI. The empirical results showed that recurrent architectures outperformed the CNN and

Dense models, confirming the importance of temporal learning in hourly wind power forecasting. Among all models,

LSTM achieved the best overall performance, with MSE = 0.0008, RMSE = 0.0282, MAE = 0.0106, MBE = -0.0006,

r = 0.9940, R2 = 0.9880, RRMSE = 0.0861, NSE = 0.9880, and WI = 0.9970. These findings demonstrate that LSTM

can effectively capture nonlinear and sequential relationships between meteorological variables and turbine power

generation, providing a reliable forecasting approach for operational wind energy management and supporting more

stable integration of wind power into modern electricity systems.

Keywords: Wind power forecasting Deep learning Long Short-Term Memory (LSTM) Renewable energy prediction

Time-series forecasting

1. INTRODUCTION

Wind energy is one of the most rapidly expanding renewable

energy sources and plays an essential role in the transition

toward low-carbon electricity systems [1, 2, 3]. As wind penetration

increases in modern power systems, accurate wind

power forecasting becomes increasingly important for grid