To improve the accuracy of building air-conditioning load prediction, we proposed an adaptive weighted least squares support vector machine (AWLS-SVM) algorithm based on analyzing the main related factors. This algorithm uses least squares support vector machine regression to develop the model and obtain the sample data fitting error, and calculates the initial weight according to the statistical characteristics of the prediction errors. The particle swarm optimization (PSO) is applied to obtain the optimal parameters of the AWLS-SVM, so as to improve its forecasting accuracy. Based on the simulated data from the DeST platform, we used the AWLS-SVM model to predict the hourly air-conditioning load of an office building in South China. The simulated results show that the AWLS-SVM has better accuracy compared with RBFNN model, LS-SVM model and WLS-SVM model, with a mean absolute error reduced by 51.84%, 13.95% and 3.24%, respectively. We also used the AWLS-SVM method to build a prediction model for the air-conditioning load of another office building based on real measurements. The results show that AWLS-SVM method outperform the other three models in terms of accumulated error, demonstrating that AWLS-SVM is effective for building air-conditioning load prediction.