M2TNet: Multi-modal multi-task Transformer network for ultra-short-term wind power multi-step forecasting

Abstract

Accurate wind power forecasting is crucial for the safe, stable, and economical operation of green power systems. Multi-step forecasting of wind power has received increasing attention in recent years. From the perspective of machine learning, multi-step prediction can be transformed into a multi-task learning problem, i.e., multiple single-step prediction tasks. From a data-driven perspective, it is often difficult to respect the differences and relationships among multi-source heterogeneous data that are typically used in multi-step forecasting research, e.g., wind power and numerical weather prediction (NWP) data. This work proposes a multi-modal multi-task transformer network (M2TNet) model that can achieve ultra-short-term wind power multi-step forecasting based on multi-source heterogeneous data. The M2TNet is a unified framework that integrates multiple feature extractors, including a Transformer, a feature fusion layer-based fully-connected network, and a prediction terminal layer. The developed model applies multi-modal and multi-task learning strategies to effectively fuse multi-modal information and enable knowledge sharing among multiple single-step prediction tasks. In addition, the Transformer’s computing efficiency and ability to mine complex dependent data is exploited for joint learning of multi-source heterogeneous data. Data, including NWP and wind power, from a wind farm in Northeast China were used to validate M2TNet. The correlations between the input variables were analyzed using maximal information coefficient method to control the scale of the model. For prediction accuracy, the experimental results showed that, compared with the existing model, M2TNet reduced the root mean square error by 0.19%, 0.99%, 1.05%, and 1.53% in 4-, 8-, 12-, and 16-step ahead predictions, respectively. Furthermore, for computational efficiency, the training time of the existing model at a wind farm is 1.66 times that of M2TNet. This confirmed that the M2TNet model performs better in terms of prediction accuracy and computational efficiency. Our work illustrates the potential of M2TNet for large-scale wind farm applications.