To solve the problem of on-line power load forecasting, a comprehensive optimization method for short-term power load forecasting based on sliding window strategy and improved artificial fish swarm algorithm was proposed, which combined multiple variable phase space reconstruction and least squares support vector machine. Firstly, the multiple variable phase space reconstruction was used to restore the dynamic characteristics of the real power system, and then the kernels were arranged and combined to transform the construction of the combined kernels into the optimization of the weight coefficients. Furthermore, the delay time, embedding dimension, LS-SVM parameters and the weight of the kernel function were taken as the whole parameter vectors, and the adaptive artificial fish swarm algorithm was used to optimize the fitness function of the prediction accuracy of the training data, so as to obtain the optimal parameters of the prediction model. Finally, the short-term power load is predicted on-line by sliding time window strategy, and the results prove the effectiveness of the proposed method.