Abstract: Due to the diversity of data variation, the fluctuations characteristic of the medium-term power load is different from short-term and long-term power load. Based on the view of electric power system complexity, the mid-term power load forecasting problem is discussed, including the forecasting model uncertainty, parameter time-varying characteristics and the periodic law of load fluctuation. According to the features of mid-term power load, a semi-parametric model based on nonparametric smoothing is built, and the division of the function interval is defined. After that, a new dynamic prediction method is put forward based on variable interval. Combined with the ensemble empirical mode decomposition algorithm and wave energy test, the noise sequence analysis and separation method is presented. The study shows that, climatic factor have the greatest impact on the electricity consumption, while economic factor has less impact on it. In different forecast periods, the explanatory of factors to the forecasting model varies over time. As the proposed semi-parametric model can be used for accurate multi-dimensional and multi-granularity analysis of electricity load, then grasp the variation, it can be efficiently used for mid-term load forecasting.

Key words: semi-parametric model, time-variant interval weights, dynamic prediction, EEMD, mid-term electricity consumption prediction