Accurate temperature prediction is an important support for meteorological business automation, agricultural disaster defense and urban fine operation. In this paper, a temperature prediction model based on convolutional neural network and gated recurrent unit is proposed. The multivariate time series samples are constructed by using the hourly meteorological observation data of multiple stations in Jilin Province. The model takes temperature, relative humidity, air pressure, wind speed, wind direction, precipitation, sunshine duration and time period encoding as input, extracts the local meteorological combination features in a short time window through CNN, and then captures the time series dependence in day and night variation, cold and warm conversion, and continuous temperature evolution by GRU. The experimental results show that the proposed model achieves 0.56℃ MAE, 0.82℃ RMSE, 4.31% MAPE and 0.962 , which is better than ARIMA, SVR, LSTM, CNN-LSTM and Transformer methods. The model has both prediction accuracy and computational efficiency, and can provide data-driven support for hourly temperature correction, short-term forecast product generation and regional meteorological services.
