Application of ConvLSTM Network in Numerical Temperature Prediction Interpretation

Abstract

The application research of machine learning methods has attracted attention in the meteorological field. In this paper, we establish a spatiotemporal temperature deviation prediction model (PredTemp) based on convolution and long short-term memory network (ConvLSTM). The model is trained with numerical weatherprediction (NWP), and the prediction results are used to correct the temperature prediction in NWP. Exploring the influence of the weather elements added to the model on the prediction results is also the focus of this paper. Two datasets are constructed for this purpose: the temperature forecast deviationdataset (dataset1) is constructed by using the temperature forecastsand the analysis field in NWP, a precipitation forecast dataset (dataset2) was constructed using the precipitation forecasts in NWP. The experimental results show that the model is effective. Using dataset1 as dataset for training, the accuracy rate of temperaturecorrected by PredTempwas increased by 3%compared to NWP; using dataset1 and dataset2 as dataset for training, the accuracy rate of temperature corrected by PredTemp was increased by 4%. The addition of precipitation elements has played a positive role in improving the accuracy of the model prediction.