Rapid changes in meteorological conditions can lead to frequent switching of wind turbine operating states, causing wind power sequences to exhibit pronounced non-stationarity and multimodal characteristics. As a result, conventional single prediction models often struggle to simultaneously maintain forecasting accuracy and stability under different operating conditions. To address this issue, this paper proposes a wind power forecasting method based on the Convolutional Normalized Transformer Encoder and Multi-Task Learning (CNTE-MTL). First, operating samples of wind turbines are divided into different operating conditions according to typical meteorological variables, such as wind speed, wind direction, and ambient temperature, to characterize differences in meteorology-driven operating patterns. Then, wind power forecasting under different meteorological conditions is formulated as multiple related subtasks, and a multi-task learning framework consisting of a shared feature extraction network and condition-specific prediction heads is constructed. In this framework, the shared feature extraction network employs one-dimensional convolution to extract local temporal fluctuation information and combines it with a Transformer encoder to capture long-term dependency features. The condition-specific prediction heads further characterize the differentiated power evolution patterns under different meteorological conditions, thereby enabling the sharing of common cross-condition information and differentiated modeling. Short-term forecasting, long-term forecasting, supplementary comparative experiments, and ablation experiments are conducted based on SCADA data from an actual wind farm. The results show that the proposed CNTE-MTL model achieves an RMSE of 0.0165 and an R2 of 0.9689 in the one-month short-term forecasting experiment, and an RMSE of 0.0072 and an R2 of 0.9980 in the three-month long-term forecasting experiment, outperforming comparative models such as CNTE, Informer, Transformer, TCN, and LSTM. The ablation experiments further verify the effectiveness of meteorology-driven operating condition division, the shared feature extraction network, and the condition-specific prediction heads in improving forecasting performance.
Zhao et al. (Tue,) studied this question.