Abstract: In order to improve the accuracy and noise-immunity of short-term load forecasting in the direct current (DC) distribution network, this paper proposes a short-term load forecasting model for the DC distribution network based on light gradient boosting machine (LightGBM) with hyperparameter optimization (Hyperopt). Firstly, based on the ring-shaped medium voltage DC distribution network, the basic characteristics of four kinds of loads and their correlation with factors are analyzed. To improve the training efficiency of the forecasting model and avoid over-fitting, the factors with higher correlation are selected as inputs. Secondly, a short-term load forecasting model is constructed based on the Hyperopt-LightGBM, which can enhance the tolerance of the model to sample noise by training a strong learner and further improve the accuracy of load forecasting. To reduce the burden of manual parameter tuning and improve the adaptability of the model, the Hyperopt is used to obtain the optimal LightGBM model. Finally, four kinds of load simulation data of the DC distribution network are used to verify the effectiveness of the proposed model. The average prediction errors of different prediction models under the four kinds of loads are less than 1.6% (the proposed model), less than 2.1%(eXtreme gradient boosting based model), less than 3.2%(random forest based model), and less than 4.1%(gradient boosting decision tree based model), respectively. In addition, the prediction accuracy of the proposed model is always higher than 95%, and it is also higher than other models under different noise ratios. The results verify that the proposed prediction model has better performance in accuracy, noise immunity, and adaptability. Hence, the proposed model can be applied for high-accuracy and noise-resistant load forecasting of short-term loads in the DC distribution network.