The generalized autoregressive conditional heteroscedasticity（GARCH） model has the ability to describe the volatility of time series. A feasible approach to short-term load forecasting based on GARCH methodology is given. By analyzing the random disturbance series of the classical time series model， the existence of autoregressive conditional heteroscedasticity（ARCH） effect is demonstrated by the Lagrange multiplier（LM） test， and the non-satisfaction of homoscedasticity of the classical model is discussed. Furthermore， an improved autoregressive moving average（ARMA）-GARCH model is built， in which the conditional variance equation is included and the estimation of model parameters is obtained by maximum likelihood estimation（MLE）. A comparison between the improved model and the classical ARMA model is made by taking into account the actual forecasting effect. The results show that the new model has more practical preconditions and better forecasting performance.