In response to China's "dual carbon" goals, the proportion of new energy sources, represented by wind power, in the power output for power grids continues to increase. Effective wind turbine output prediction is particularly important for formulating grid scheduling and power generation plans ahead of time. Due to the strong irregularity and seasonality of wind power data, a single model prediction method cannot solve the problem of wind power intermittency while ensuring prediction accuracy. To address this, a combined model using the autoregressive integrated moving average (ARIMA) time series, long- and short-term memory (LSTM) network, and radial basis function (RBF) neural network is proposed for short-term prediction of wind turbine output in a certain region. First, data preprocessing and sequence stationarity analysis are performed to obtain a stationary sequence and predict it through ARIMA. Secondly, irregular data that do not meet the criteria of residual white noise analysis are predicted through LSTM. Then, the RBF neural network is used to learn and simulate the predicted values to improve accuracy. Finally, simulations are conducted based on data from a wind power station. Compared with other single model prediction methods, the results show that the proposed combined model prediction method can predict wind power data with strong seasonality and irregularity and has better prediction accuracy, providing a reference for the operation and scheduling of corresponding equipment and enhancing power supply reliability.