Power load forecasting plays an important role in balancing energy distribution, economy, safe and reliable operation of power system. Accurate load forecasting can reduce the cost and risk of power operation and improve the environmental and economic revenues of power grid. Firstly, the weighted grey relational projection algorithm was used to process the data, then the attention mechanism was applied to improve the bi-directional long short-term memory (Bi-LSTM) model. By combining with eXtreme gradient boosting (XGBoost) model, a power load combination forecasting model with the weight determined by the error reciprocal method was developed, so as to obtain a new short-term power load forecasting method. Finally, based on the Singapore power market data set, the results of the proposed prediction method show that our prediction results are closer to the real data and have less error than the single prediction methods, which reflect the effectiveness, accuracy and practicability of proposed method.