Abstract: Accurate load forecasting is the important guarantee for the safe and stable operation of a power system. When multiple factors are fully considered, the preprocessing of massive input data and variables selection are essential to improve the accuracy of load forecasting. For the problems like unclear information and dimension redundancy in the traditional time variables, the k-means clustering analysis technology based on cosine similarity is introduced to realize the classification of daily load curves, and the classification labels are used to replace the traditional time variables; Considering the relationship between load and multiple factors at hourly granularity, a feature selection strategy based on the feature set reconstruction and the Maximal Information Coefficient is proposed to achieve a refined feature selection with hourly granularity; The time convolutional network with great learning ability is introduced to achieve high-precision short-term load forecasting. The results show that after applying the former two proposed strategies, the quality of input data are optimized by the replacement of the inefficient time variables and the optimal feature sets with hourly granularity, and the prediction performance of multiple classic models are significantly improve, among which, the time convolutional network model combined with two strategies has a higher accuracy. And also, the proposed strategies are suitable for load forecasting throughout the year, with good portability and robustness.