Abstract: Accurate prediction of load can guide thermal power plants to formulate power generation plans and scheduling arrangements, which is conducive to their reduction of energy costs and pollution emissions, and is of great significance to the economy and environmental protection of power plants. Therefore, a load forecasting method of thermal power plant based on BiLSTM-Attention is proposed in this paper. Firstly, the key characteristic variables are screened by Pearson coefficient. Secondly, BiLSTM was used to extract the long-term dependence relationship and short-term change characteristics among key variables, and finally, the Attention mechanism was integrated to further highlight the key timing information, so as to achieve accurate load prediction. A 600 MW supercritical unit in service was used for validation. Compared to LSTM, BiLSTM, LSTM-Attention, the results of BiLSTM-Attention show that the coefficient of determination R~2, root mean square error S_RMSE and mean absolute error S_MAE are optimal,they are 0.956 6、16.315 9、13.504 3,which can more accurately capture the trend of rapid load fluctuation, it can take the accurate prediction of load for thermal power plants.