Combining Transformer-Based Models for Wind Data Prediction

This study explores the use of combined Transformer-based models for time series forecasting, evaluating the Informer, Autoformer, Transformer, and CALF architectures. models were trained on wind data under two configurations, with and without denoising, and evaluated across different forecasting horizons. Two combination strategies were developed: a Multi-Objective Optimization (MOO) approach using NSGA-III to balance MAPE and RMSE, and another one through the SAW multicriteria technique, which dynamically selects the most suitable model at each time step based on sliding windows. Results show that the SAW model outperformed the MOO-based approach by adapting to the evolving behavior of the data, while the MOO model consistently provided stable performance compared to the based model. Moreover, denoising did not have a clear impact on accuracy but significantly reduced training time. It is concluded that combined approaches can offer more robust forecasts than individual models, especially in scenarios with pronounced temporal variability