Considering that it is difficult for traditional BP neural network to deal with the correlation between power load data, this paper proposes an Improved Long Short Term Memory based on Dropout(Dropout-ILSTM) for short-term power load forecasting. Firstly, an Improved Long Short Term Memory (ILSTM) network is proposed, which enhances the representation of linear components in the target system by fully connecting multiple time-step inputs of long- and short-term memory networks with output vectors. Secondly, using Dropout Optimizing the ILSTM network improves the generalization ability of the network and reduces the training time of the model. Finally, the Dropout-ILSTM power load forecasting model is constructed with the date, temperature, electricity price and power load data as inputs. This paper uses the NSW electric load data provided by AEMO as a test case. The experimental results show that the proposed Dropout-ILSTM model has higher prediction accuracy and more generalization ability than other neural network models. And the Dropout-ILSTM model is suitable for power load forecasting with different prediction widths.