Abstract: Daily temperature forecast is an important part in weather forecasting. New methods are explored to improve the prediction accuracy since the atmosphere system is a complex nonlinear system. A model based on multivariate time series and local support vector regression was proposed. After constructing multivariate time-series with C-C method and minimum prediction error method, a way to extract the nearest neighbor from each predictor's sequences was used to build the 1-day ahead local forecasting model for daily maximum and minimum temperature. To demonstrate the effectiveness, the model was applied and tested in data from 753 stations data package of China. Simulation results show that the model can improve prediction accuracy effectively, and has a better application value in short-term daily temperature forecast compared with univariate time series.

Key words: daily temperature, multivariate time series, segmentation, nearest neighbor, local support vector regression