A simple yet effective multivariate long sequence wind speed prediction model for urban blocks with spatio-temporal feature embedding

Accurate and reliable wind field prediction in urban neighborhoods can help with building design and urban planning to mitigate environmental issues such as the heat island effect and air pollution. Recently, spatio-temporal graphical neural networks have been widely used in this area due to the advantages of combining graph convolutional networks and sequence models. However, recent studies have shown limited improvements in performance despite the increasing complexity of approaches employed. To overcome this limitation, this paper proposes a simple yet effective multivariate long sequence wind speed prediction model for urban blocks and a method to improve it's accuracy. Specifically, the model uses a one-dimensional convolutional neural network (CNN) to efficiently extract short-term fluctuation features from the wind speed time series. Spatial relationships are represented in lower dimensions by embedding additional spatial features, preserving their physical meaning. Meanwhile, long-term trends and periodic patterns in wind speed variations are captured by adding embedding layers for temporal features. This design facilitates complementary joint modeling of time series data and static background information. Furthermore, aiming to address the issue of significant performance degradation during long-term predictions, a novel accuracy enhancement method based on prediction data is proposed. The accuracy of the proposed method is verified using both measured and open weather datasets. The spatial feature-embedded model accurately predicts wind speed under significant fluctuating conditions. The mean absolute error is only 0.051 m/s, which is 27 % lower than that without spatial feature embedding. Embedding timestamps can significantly enhance the accuracy of wind speed prediction. Compared with the model without embedded timestamps, the mean absolute error is reduced by 0.145 m/s, resulting in a 14.5 % improvement in accuracy. Finally, we implement the proposed accuracy enhancement method based on the predicted data. The results demonstrate that the declining accuracy of long sequence predictions can be effectively mitigated without additional inputs. Overall, this paper proposes new methods for predicting wind speed that have significant potential for application in long-term urban neighborhood wind speed prediction.