Short-term load forecasting with dense average network

As an important part of the power system, load forecasting directly affects national economy. Small improvements in forecasts can save millions dollars for industry. Therefore, improving accuracy has always been pursuing goal a system. Based on this goal, paper proposes novel connection, dense average which outputs all preceding layers are averaged as input next layer feed-forward fashion. Dense connection alleviate problem gradient explosion without introducing new parameters. we construct network (DaNet) forecasting. On two public datasets (ISO-NE dataset and NAU dataset), use MAPE, MAE RMSE to evaluate performance DaNet. The predictions DaNet better than those existing benchmarks. basis, uses ensemble method reduce peak value prediction bias, helps dispatching caused by unexpected loads. To verify reliability model predictions, robustness is analyzed verified adding disturbances. experimental results show that proposed effective robust