2026, Vol. 7, Issue 1, Part B

This study addressed the challenge of accurate wind power forecasting by developing a hybrid big data time-series model that integrated ARIMA, LSTM, GRU, and CNN-LSTM architectures with pattern-based sequence clustering. Using a real-world dataset containing 34,080 time-stamped observations recorded at 15-minute intervals, the research utilized ten meteorological and operational variables, including wind speed, preliminary power output, wind direction, temperature, humidity, atmospheric pressure, and rounded turbine measurements, along with YD15, a 15-minute-ahead power target used for supervised learning. A comprehensive preprocessing workflow—comprising outlier removal, missing-value interpolation, normalization, and feature engineering—was applied to ensure data quality. Dynamic Time Warping (DTW) clustering was employed to group similar temporal sequences, enabling localized model training across diverse wind regimes. The hybrid architecture was deployed in a distributed environment using Apache Spark, ensuring scalability and high processing throughput. Experimental evaluation on the dataset demonstrated that the hybrid model consistently outperformed standalone approaches, achieving lower MAE and RMSE and higher Accuracy, Precision, Recall, and F1-scores. Overall, the study provided a robust, scalable, and data-driven forecasting solution capable of capturing both linear and nonlinear wind power dynamics, supporting more reliable smart-grid operations and sustainable energy management.