Extreme weather events can cause sudden increases or drops in electrical loads

posing challenges to power grids' stability and power supply capacity. However

the existing ultra-short-term load forecasting methods cannot predict the load characteristics of nonlinear and dynamic changes under extreme weather. To cope with the problem of the reduction of prediction accuracy caused by strong load abruptness and severe fluctuation under extreme weather

this paper proposes an ultra-short-term power load prediction method based on quadratic reconstruction decomposition and denoising and Bidirectional Long Short-Term Memory (BiLSTM) which considering extreme weather. Firstly

the maximum information coefficient was used to select the climate characteristics strongly correlated with load. Then

the features of multiple frequency bands are extracted by the quadratic reconstruction decomposition and denoising method

which reduces the data complexity and provides a cleaner and clearer input sequence for the BiLSTM model to improve the model's training effect and prediction ability. Finally

based on the historical datasets of Belgium

Fujian Province

and Tétouan

the Mean Absolute Percentage Error in different cases decreased to 1.024%

0.875%

1.270%

and 1.009%

respectively. Experimental results verify that the proposed method performs well and has broad application prospects in ultra-short-term power load prediction during extreme weather.