DTTM: A deep temporal transfer model for ultra-short-term online wind power forecasting

Accurate wind power forecasting (WPF) is vital for grid stability. Most existing studies rely on the combination methods, and the multi-source information (MSI) related to the wind power, but rarely consider the model performances in online prediction scenarios, where an efficient model structure and the timeliness of MSI are worth considering. In this study, a deep temporal transfer model (DTTM) and a synchronous training strategy (STS) are initially introduced. Firstly, temporal synchronous data is mapped to external temporal data (ETD), not only avoiding the asynchronous obtaining time of the MSI, but also making the past and future spatial-temporal information relevant. Secondly, to expand the forecasting model structure for adapting the ETD, a lightweight model named parallel network (PN) is developed as a forecasting cell. Based on deep parallel network (DPN), to fully train ETD for acquiring superior online prediction effect, a highway convolutional neural network (HCNN) is utilized to link the ETD to DPN, mitigating the degradation caused by insufficient information transmission. Through applying DTTM structure adopted STS, the Normalized Mean Absolute Error (NMAE) reduces by up to 74.42 % and at least 38.11 % compared with other models, achieving better online WPF performance.