Accurate power load forecasting is the basis for ensuring the stable operation of the power grid. There are still problems in long-term load forecasting in China. The interaction ability of various indicators of load forecasting is weak, and the screening ability is insufficient, resulting in low accuracy. Therefore, according to the above background, a Gold Subtractive Optimization Extreme Learning Machine (GSABO-ELM) algorithm was proposed. The golden sine algorithm improves the subtraction average optimizer and its optimization ability. Then, the improved subtraction average optimizer is used to find the best adaptive state of particles, and the optimal initial threshold and weight are given to the extreme learning machine for training and prediction. To analyze the forecast results quantitatively, this paper combines three error evaluation standards: RMSE, RPD, and MAPE. The prediction model obtained by GSABO-ELM by comparing the experimental study of ELM, SABO-ELM, and BWO-ELM has more robust generalization and stability in long-term power load prediction.