Lag-Based Feature Engineering for One-Hour-Ahead Wind Speed Prediction Using Artificial Neural Networks

Abstract

Accurate one-hour-ahead wind speed prediction is critical for the efficient use of wind energy, energy generation planning, and the reliable integration of wind energy systems. However, the time-dependent variability and nonlinear behavior of wind speed make this prediction problem challenging. In addition, abrupt wind-speed peaks and temporal dependence in hourly data require models that can effectively capture short-term wind-speed dynamics without temporal data leakage. In this study, an artificial neural network (ANN)-based model for one-hour-ahead wind speed prediction was developed using hourly meteorological data. The model inputs included temperature, humidity, radiation, pressure, the sine and cosine components of wind direction, lagged wind speed variables, and the most recent wind speed observation. The output variable was defined as the wind speed one hour ahead. The modeling procedure consisted of constructing lag-based input variables, aligning the input–output dataset for the one-hour-ahead prediction task, applying a chronological training–validation–test split, and selecting the final ANN model through validation-based tuning. The model was trained and tested using a total of 8757 hourly data points from 2024 obtained from a meteorological station in Elazig, Türkiye. A chronological training, validation, and test split was used to prevent temporal data leakage, and the proposed model was compared with persistence-based, linear, tree-based, and meteorological-only ANN benchmark models. On the independent test dataset, the proposed tuned lag-based ANN achieved an RMSE of 0.8904 m/s, an MAE of 0.6394 m/s, and an R of 0.8897. The proposed model obtained the lowest RMSE among the evaluated models, although the persistence model achieved slightly better MAE and R values. The results show that using lagged variables allows the model to learn the temporal dependencies of wind speed more effectively and produce competitive RMSE-based prediction performance, while the persistence model remains competitive according to MAE and R. Therefore, the proposed lag-based ANN approach provides a competitive RMSE-based framework for one-hour-ahead station-level wind speed prediction under similar onshore meteorological station conditions.