Enhancing ultra short-term wind power prediction accuracy with feature-based CNN-LSTM models: A comparative study

Abstract

This study explores the enhancement of ultra-short-term wind power prediction accuracy through a novel integration of Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) models, enriched with an innovative feature-based approach. Utilizing data from Turkish wind farm with 10 minutes interval and prediction horizon is 1 hour (t + 1, t + 2, t + 3, t + 4,t + 5, t + 6), the research undertakes a comprehensive comparative analysis of feature engineering methods, including Recursive Feature Elimination with Cross-Validation (RFECV), Genetic Algorithm (GA), and Consensus Feature Analysis (CFA). These methods are employed to refine the selection of features that significantly influence the prediction accuracy of the hybrid CNN-LSTM model. The findings underscore the effectiveness of the feature-based methodology in optimizing the performance of wind power prediction algorithms. Notably, the CFA method emerges as a superior technique for feature selection, demonstrating its potential in improving model accuracy and computational efficiency. The results show that the proposed approach significantly improves prediction accuracy. In particular, the CFA-based model achieved approximately 0.529 MAE, 0.5351 MSE, 0.7315 RMSE, and 3.6002 MAPE for the t + 1 prediction step, resulting in a lower error rate and higher computational efficiency compared to other methods. This work contributes to the renewable energy sector by providing a robust framework for the accurate prediction of wind power, thereby facilitating the integration of wind energy into the energy grid and enhancing the reliability of renewable energy systems. The study lays a foundational basis for future research on the application of hybrid artificial intelligence models in wind energy forecasting and the broader field of renewable energy optimization.