In order to solve the problem that the time series features cannot be fully extracted due to the characteristics of power load, such as nonlinearity, high volatility and strong randomness, a short-term power load prediction method based on improved complete ensemble empirical mode decomposition with adaptive noise and a deep error correction multi-scale bidirectional temporal convolution network-bidirectional long short-term memory network is proposed. First, the maximum information coefficient is used to select a highly relevant feature set to reduce feature redundancy. Then, improved complete ensemble empirical mode decomposition with adaptive noise is employed to decompose the highly fluctuating load into intrinsic mode functions with different frequencies and residual, while sample entropy is introduced to reconstruct new subsequences with similar complexities to reduce computational load. Next, using the optimal feature set, each subsequence is forecasted with a multi-scale bidirectional temporal convolution network-bidirectional long short-term memory network model optimized by the sparrow search algorithm to obtain initial load sequence predictions. Finally, the error sequence is corrected using the error correction module to refine the initial predictions. Experimental verification shows that compared with other prediction models, RMSE is reduced by 51.42% at most and 34.26% at least, which verifies the accuracy and effectiveness of the model.