TRNet: A trend and residual network utilizing novel hilly attention mechanism for wind speed prediction in complex scenario

Accurate wind speed prediction plays an important role in scheduling optimization of wind power generation systems. In complex scenarios, wind speed features are difficult to be captured effectively, leading to unsatisfactory prediction results. To overcome this problem, a trend and residual network (TRNet) based on a novel hilly attention mechanism is proposed for wind speed forecasting, which consists of a trend component (HAttn-AR), a residual component (CHAttn) and an entangle component. Specifically, HAttn-AR composed of hilly attention layers and a recursive layer is constructed, which can separately extract parallel and inference features. In the CHAttn, a cross hilly attention layer is built to capture dynamic residual features. Considering the complexity of trend and residual features, an entangle component is constructed to integrate these features. To validate the performance of TRNet, two offshore wind speed data sets from farms in China and Norway have been used. The results show that TRNet is superior to other existing advanced models. In particular, in the China data, the MAE value of TRNet is improved to 5.26 %–77.32 %, the MSE value is reduced to 7.14 %–91.13 %, and the SMPAE value is optimized to 12.50 %–74.07 %. https://github.com/linsonM/TRNet.