Multi-step ahead wind power forecasting based on dual-attention mechanism

Accurate forecasting is essential for the economic benefits and efficient operation of intermittent wind power systems. Multi-step ahead wind power forecasting provides multiple benefits in the planning and operation of the power systems. This study proposes a deep learning model based on a dual-attention mechanism for multi-step ahead wind power forecasting. Both the attention mechanisms are applied over the encoder–decoder based sequence-to-sequence model consisting of long short-term memory (LSTM) blocks. The Bayesian optimization algorithm is applied to the proposed model to obtain the optimal combination of hyper-parameters. To evaluate its effectiveness, the proposed model has been compared with the persistence model, and state-of-the-art models such as simple LSTM, LSTM-attention, ensemble model, and neural basis expansion analysis for time series (N-BEATS) model. Furthermore, the performance of the attention mechanism with respect to impactful input features, such as wind speed and air pressure, was analyzed. The forecasting skill score of the proposed model was the highest among all other models in comparison, which indicates the effectiveness of the proposed model. Similarly, the proposed model outperformed the traditional methods in terms of other evaluating criteria, including mean absolute error (MAE), and root mean square error (RMSE) among others, hence, proving its efficacy. The average RMSE score of the proposed model for multi horizon forecasting was 0.04995, whereas that of N-BEATS was 0.0876, ensemble method was 0.1132 and LSTM-attention was around 0.101375. Similarly, the average forecasting skill score of proposed method was 0.6625, whereas that of N-BEATS was 0.3975 and LSTM-attention achieved 0.3775 skill score. •Multi-step ahead wind power forecasting by Dual attention mechanism.•Hyper-parameter tuning of dual-attention model by Bayesian optimization.•Comparative study with state-of-the-art models.•Analyzing attention mechanism against influential features and temporal sequences.