Abstract: To ensure the normal power supply of the power plant, a short-term heavy rainfall weather wind farm power generation prediction model is designed to improve the power generation prediction effect. Expanding small samples of short-term heavy rainfall weather through the principles of Euclidean distance and angle; using the improved depth separable convolution algorithm, the spatiotemporal characteristics of weather power are extracted from the normal weather samples, and input into the long-term and short-term memory network to establish the prediction model of normal weather wind farm power generation reference value, and obtain the power generation reference value; in the Transformer algorithm, input an expanded sample to establish a prediction model for power generation loss value under short-term heavy rainfall weather; using a Sequence to Sequence network based on attention mechanism, combined with expanding samples, a prediction model for the time point of power generation loss is constructed. Combined with a loss value prediction model, the final power generation loss value is obtained; subtract the loss value from the reference value to obtain the prediction results of wind farm power generation under short-term heavy rainfall weather. The experiment proves that the model can effectively expand the small samples of short-term heavy rainfall weather; this method can accurately predict the time point of power generation loss, obtain the value of power generation loss, and complete power generation prediction; under different wind speeds, the key error index and deviation degree of the model’s power generation prediction are relatively low, indicating a higher accuracy of power generation prediction.