Abstract: In order to improve the accuracy of short-term load forecasting, focusing on the construction of feature combinations, a feature combination based on Holt-Winters exponential smoothing（FCHW） is proposed. And combined with the temporal convolutional network（TCN）, the FCHW-TCN load forecasting framework is established. Firstly, the Holt-Winters exponential smoothing is applied to the load sequence forecasting, and the level component and seasonal component related to the load sequence are obtained. By using the above components as input features and combining them with conventional features（historical load, date）, the FCHW is constructed. Secondly, the TCN is chosen as the forecasting model, and the FCHW acts as the input of the TCN to build the FCHW-TCN forecasting framework. Finally, two different load data sets and multiple forecasting models are used to verify the FCHW and the FCHW-TCN. The results show that the FCHW contributes to the improvement of the forecasting accuracy for models. Compared with other forecasting models, FCHW-TCN forecasting framework has the highest forecasting accuracy and superior forecasting ability.