Research On Multi-node Power Load Forecasting Based On Deep Multi-task Learning

single-node loads cannot fully characterize the overall characteristics of a complex distribution system,thus precise multi-node load prediction is the basis for accurate load control in distribution systems.However,accurate multi-node load forecasting is challenging due to the dynamic stochastic character of demand-side loads and the intricate interaction of nodes in the distribution network.Therefore,based on the analysis of multi-node load characteristics,this paper comprehensively uses deep gated recurrent neural networks,time convolutional networks,and soft sharing multi-task learning to research the multi-node load forecasting problem of the distribution system.First,the correlation between total load and node load is analyzed based on the multi-node load characteristics to provide a basis for building a soft shared multi-task learning model with shared total load characteristics.A gated recurrent neural network is used as a prediction network for the nodes to efficiently learn the stochastic fluctuation characteristics of multiple node load data.Second,to improve the mining ability of gated recurrent neural networks for nonlinear coupled features,a modal feature extraction network is designed using the Inception strategy and gated temporal convolution.On this basis,soft shared multi-task learning is used to optimize the prediction tasks of all nodes simultaneously and improve the model’s accuracy and generalization,to share the learning knowledge of each task network.Final,using two publicly available datasets,CDS and AEMO,the proposed model is compared against many classic and sophisticated single-task,multitask,and multi-node prediction models.The results show that the model can effectively explore coupling nature in multi-node load data.And except for accuracy measurements such as MAPE,MAE,and RMSLE,the model’s predictions show significant and stable prediction performance in overall prediction metrics such as WMA and DM.Under the CDS dataset,the MAPE of this model are reduced by 3.92% and 17.04%,respectively,in the aggregated and non-aggregated node load circumstances,compared to FANNRE.Meanwhile,the model of WMA is 98.64% compared to other models.In the AEMO dataset,our model’s MAPE decreases by 2.45%,1.87%,and 1.96%,respectively.The multi-task prediction model is also estimated multi-node demand using real-world data from steel production,with promising results in the domain of precision load control.