Scientific and accurate temperature prediction is a key link in realizing meteorological disaster early warning and efficient use of climate resources, and it is also an important prerequisite for promoting smart meteorological construction and refined meteorological services. Due to the complex characteristics of temperature series such as non-linearity, time-varying, and seasonality, and the comprehensive influence of various uncertain factors such as atmospheric circulation, topography, humidity, and wind speed, the data obtained through various data collection methods are often uncertain. These uncertainties may affect the information presented by the quantitative results, which limits the accuracy of the prediction model to a certain extent. In this paper, the monthly average minimum temperature, monthly average temperature and monthly average maximum temperature in Chengdu from 2008 to 2018 are used as three boundary points to construct a triangular fuzzy number sequence, and to convert the triangular fuzzy number sequence into three index number sequences with equal information for the sake of data integrity. For the three index number sequences, the time series Autoregressive Integrated Moving Average (ARIMA) model and the gated recurrent unit (GRU) neural network are constructed respectively. Finally, on the basis of the prediction results of a single model, the induced order weighting (IOWA) operator are introduced. A triangular fuzzy combination prediction model of ARIMA-GRU integrating IOWA operators is established, and the temperature prediction value is obtained by de-fuzzification. The new model is applied to the temperature series prediction of Chengdu from 2008 to 2018, and compared with ARIMA model and GRU neural network, the results show that the performance of the new model in terms of mean square error (MSE) and average accuracy is better than that of other models, significantly improving the prediction accuracy of air temperature.
