The increase of stochastic factors leads to increasing complexity of power load data components. This makes short-term load forecasting progressively more difficult. Thus a combined forecasting model fusing a temporal convolutional network with multiple linear regression on the time-frequency scale is proposed. The complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is used to decompose the load data into multiple intrinsic mode functions with different frequency features in the time-frequency domain, and the intrinsic mode functions are clustered into random and trend terms under the fuzzy entropy criterion. The Pearson correlation coefficient is used to pick out features that are highly relevant to the power load from many influential factors. The analysis of a small time scale makes it easier to determine local detailed features, and the fine granularity feature set of the random and trend terms are constructed respectively. The temporal convolutional network (TCN) with strong nonlinear processing ability is used to predict the random term, and the multiple linear regression (MLR) with simple structure and good linear fitting effect is used to predict the trend term. The final predicted value is obtained by superposing and reconstructing both predicted results. Experimental results on two datasets including for Singapore and Belgium prove that the proposed model has high prediction accuracy, good generalizability and robustness.