TempCast: A Multi-modal Transformer for Short-Term Temperature Forecasting

Accurate weather forecasting benefits us in a variety of ways, from scheduling flights to agricultural harvests. However, existing temperature forecasting models have two problems, one is the accumulation of errors caused by autoregressive models, and the other is difficult to predict complex and varying temperatures such as those in highlands. In this paper, we proposed TempCast, a multi-modal Transformer model for short-term temperature prediction. The model has two features: (i) Modeling the features entirely by self-attention, which can effectively capture the exact long-term dependent coupling between output and input. And multiple predictions are obtained at once using a generative decoder, (ii) The modeling of multi-source data through a decoupled multi-modal fusion mechanism can effectively come to cope with the drastic changes of weather in highlands and mountains, etc. The experimental results show that the method can well achieve short-term temperature prediction and significantly outperforms all traditional methods in several indicators. The method also provides a new solution idea for multi-modal temperature prediction. Our code and data are available at https://github.com/Adam618/Temp Cast.