The three-phase balance of power load is a significant guarantee for the safe operation of the power system, which promotes efficient transmission of power quality, reduces line losses, and provides high-quality services. This paper proposes a short-term three-phase load combination forecasting model based on Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Long Short-Term Memory Neural Network (LSTM) to address the difficulty in three-phase load forecasting caused by differences in power loads. In order to obtain better forecasting results, this method processes the load data of each phase separately by assigning different weights to the forecasting results of different neural networks, instead of processing the total load data of the three-phase, then different operating states of each phase can be obtained. Finally, the neural network combination forecasting method proposed in this paper was validated by using three-phase load data from a certain regional power grid. The experiment showed that the proposed network model has good performance and high accuracy in three-phase load forecasting, and has certain engineering application value.