Historical electricity load data of industrial enterprises has the characteristics of strong volatility and complex sequence, which brings challenges for accurately predicting electricity load. In order to solve these problems, a short-term load forecasting method for industry based on multi scale weight adaptation and bidirectional gated recurrent unit（MSWA BiGRU）is proposed. The proposed model is composed by a weight adaptation layer, a BiGRU layer, a feature embedding layer, and a fully connected prediction layer. Firstly, the weight adaptation layer adaptively generates the dependent thermal coefficients for different time scale load data, and then the BiGRU layer simultaneously learnes the transient fluctuation characteristics and steady-state periodic characteristics of the historical load series on multiple scales. Then, other features is embedded in the feature embedding layer. Finally, the temporal features are fused with other features to obtain the final load prediction result through the fully connected prediction layer. Experimental results on real data of electricity load in industrial and commercial enterprises show that the prediction performance of the proposed method is superior to other methods,thus the effectiveness and feasibility of this method are verified.