This work presents a novel transformer-based deep neural network architecture integrated with wavelet transform for forecasting wind speed and wind energy (power) generation for the next 6 h ahead, using multiple meteorological variables as input for multivariate time series forecasting. To evaluate the performance of the proposed model, different case studies were investigated, using data collected from anemometers installed in three different regions in Bahia, Brazil. The performance of the proposed transformer-based model with wavelet transform was compared with an LSTM (Long Short Term Memory) model as a baseline, since it has been successfully used for time series processing in deep learning, as well as with previous state-of-the-art (SOTA) similar works. Results of the forecasting performance were evaluated using statistical metrics, along with the time for training and performing inferences, both using quantitative and qualitative analysis. They showed that the proposed method is effective for forecasting wind speed and power generation, with superior performance than the baseline model and comparable performance to previous similar SOTA works, presenting potential suitability for being extended for the general purpose of multivariate time series forecasting. Furthermore, results demonstrated that the integration of the transformer model with wavelet decomposition improved the forecast accuracy.

Highlights:

• A new transformer-based model for multivariate forecasting of wind speed and energy.•Integration with wavelet transform for feature augmentation to improve predictions.

• Full assessment of the novel model comparing with LSTM, persistence and SOTA works.

• Lower time to train than the fine-tuned LSTM, being more energy-efficient.

• Evidenced superiority over LSTM and performance comparable to similar SOTA works.