Short-term integrated forecasting method for wind power, solar power, and system load based on variable attention mechanism and multi-task learning

Improving the forecasting accuracy of wind power, solar power, and system load to support the source-load cooperative dispatch is an important direction to reduce the uncertainty at source and load sides. The current research mainly focuses on a single object, ignoring the interactive coupling relationship among them, which limits the improvement of forecasting accuracy. Therefore, this paper proposes a short-term integrated forecasting method of wind-solar-load. Firstly, a feature extraction module of linkage characteristics of wind-solar-load is built based on variable attention mechanism. Secondly, a multi-task learning model that can automatically calculate the optimal loss weights for different forecasting tasks is constructed to simultaneously accomplish the wind and solar power forecasting tasks through Fully Connected Neural Network. Finally, a load forecasting model which fuses historical load and power forecasting information is established based on Long Short-Term Memory. The operation data of 8 wind farms and 6 solar plants, and the load data of a nearby city are used for instance analysis. The results show that the power forecasting error (root mean square error) of each wind farm, solar plant, and system load is reduced by 4.84 %, 1.86 %, and 3.02 % on average, respectively, compared with the corresponding traditional methods.