With the increasing proportion of wind power connected to the grid year by year, the power system has higher requirements for the accuracy and stability of wind power forecasting. For the same wind farm, to avoid different feature subsets of wind farms selected by different feature selection methods, this paper proposes a short-term wind power forecasting method based on a causal regularization extreme learning machine from the perspective of causality. First, the extreme learning machine (ELM) model is modeled as a structural causal model, and then the average causal effect vector between hidden layer neurons and output layer neurons is calculated. Then the average causal effect vector is combined with the weight of the output layer to form a causal regularization term. This minimizes the training error and maximizes the causal relationship of the network, further improving the forecasting accuracy and stability of the model. Finally, taking the data of a wind farm in Mengxi, China as an example, and comparing with the forecasting model with or without the feature selection method, the effectiveness and applicability of the proposed method are verified.