With the increasing scale of distribution network, more diversity of load types, larger volume and more types of data, forming an effective up-and-down forecast network, which means to combine small-scale and characteristically obvious load forecast with large-scale and integrated load forecast, is important to distribution network planning and operation. Considering the limitation of the data quantity for the lack of distribution network measurement equipment, this paper proposes a multi-node load forecasting method of “from the whole to the nodes”. Because AR is not able to accept multiple source data and the outcome of BP neural network algorithm is severely affected by the similar day, this paper proposes a novel algorithm, AR-ANN, with complementary advantages of the previous two. Finally, the strength of AR-ANN in data processing speed and prediction error is validated respectively by tests of an ordinary single node load forecast, a traditional multi-node load forecast by the method of “from the nodes to the whole” and the new way of multi-node load forecast. This work is supported by National Natural Science Foundation of China (No. 51407116 and No. 51477098).