Spatio-Temporal Hybrid Network with TCN and Attention for City Temperature Forecasting

Meteorological data are complex and diverse, with nonlinear characteristics. Temperature changes are affected by many factors, so accurate temperature prediction is a major challenge. A hybrid prediction model based on time series and spatial interaction characteristics is proposed for the mutual influence and temporal characteristics of various meteorological indicators. First, the time convolutional network (TCN) model is used to extract the temporal characteristics of each meteorological element, including long-term dependence and short-term dependence characteristics. Secondly, the attention mechanism (AM) is used to extract the interaction characteristics between meteorological elements, and the temporal and spatial interaction characteristics of meteorological elements are respectively input to the gated recurrence unit (GRU). Through empirical research on six major cities with different distributions in China, the study shows that compared with other existing methods, the proposed mixed model can simultaneously capture the temporal and spatial characteristics between various factors. Compared with the LSTM model, MAE, MAPE and RMSE have an average increase of 28.74%, 29.71%, and 27.97%, and are significantly better than other models. The research results provide new ideas for feature extraction in the field of meteorological prediction.