Wind power forecasting: A hybrid forecasting model and multi-task learning-based framework

Accurate forecasting of wind power is of significance for scheduling the grid system when wind power is integrated. However, the deficiency of the training data restricts the models’ forecasting performance and modeling efficiency. In this study, we propose a hybrid forecasting model that is composed of a dual dilated convolution-based self-attention sub-model and an autoregressive sub-model. The dual-branch sub-model utilizes a dual convolution architecture to extract both global and local temporal patterns before capturing attention-based dependencies between multivariate inputs to reflect non-linear correlations. The autoregressive sub-model learns linear correlations to provide supplementary information that compensates for the insensitivity of model response. Furthermore, a multi-task learning-based framework is designed to address insufficient training data of a new turbine cluster. The framework can be divided into one task-shared linear component and multiple task-specific non-linear components. By weighting multiple forecasting tasks, the proposed framework utilizes the collaborative relationships between tasks to improve accuracy on the target turbines. Experiment results show that the proposed forecasting model presents the better forecasting accuracy on actual datasets, and the framework has a significant improvement of 20.08% in accuracy while further reducing dependence on training data, especially for source domain data in transfer learning.