The invention discloses an annual maximum load prediction method based on a temperature reduction model used in the power grid control field. The method comprises the following steps of: a historical data processing step comprising obtaining an annual maximum load of each year in the history and a maximum temperature on a maximum load day, carrying out accumulative temperature correction on the maximum temperature on the maximum load day of each year in the history, calculating an annual basic load of each year in the history, and calculating a temperature sensitive coefficient at each temperature; a load reduction step comprising determining a maximum reference temperature, solving an adjustment coefficient, and calculating an annual reduction maximum load of each year in the history; a load extrapolation prediction and result adjustment step comprising establishing a regression model, carrying out extrapolation prediction on an annual reduction maximum load of a target year by using the annual reduction maximum load, predicting an annual maximum temperature of the target year, determining a prediction interval of the annual maximum temperature of the target year, and determining an annual maximum load prediction interval according to the prediction interval of the annual maximum temperature of the target year.