According to the chaotic features of load series, a new forecasting method combining phase space reconstruction and Gaussian process regression is proposed. Firstly, two parameters of time series (delay time and delay window) are earned at the same time by means of the C-C method. Secondly, the reconstructed series of the separate load as well as the multi-variable model considering load and other influence factors are established. Then, the load sample is trained by GPR models using both single and composite kernel function and the optimal hyper-parameters are calculated, with which the 24-hour daily loads are predicted. Finally, the forecasting consequence of the single load model is contrasted with SVM model and the multi-variable GP model. Prediction results indicate that the model using multi-variable and composite kernel function achieves better effects and the new method is not only feasible but also satisfies the requirements of the engineering precision.