The load of industrial customers usually contains various types. This leads to a complicated load structure and an inevitable composition of large impact loads. Traditional load forecasting methods find it difficult to accurately forecast these sudden changes in load patterns, resulting in low forecast accuracy. To improve load forecasting, an improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) based method is proposed for industrial customers. It decomposes the load of industrial customers into different components in terms of frequencies and forecasts them separately. First, the ICEEMDAN algorithm decomposes the load into high and low-frequency modal components. The local mean is introduced to replace the modal estimation. This avoids the influence of Gaussian noise on modal decomposition, and improves on mode mixing in the traditional modal decomposition method. Secondly, long short-term memory (LSTM) and least squares support vector regression (LSSVR) algorithms are adopted to establish the forecasting models of high and low-frequency modes. Finally, the forecasting results of each component are superimposed and reconstructed to obtain the final load forecasting. Compared with multiple traditional methods such as the single forecasting and other combined forecasting methods, the mean absolute percentage error (MAPE) of the proposed method is decreased by 26.35% and 12.75% respectively, thus it has the highest forecast accuracy among them. This work is supported by the Science and Technology Project of State Grid Corporation of China (No. 522730191002).