Research On Day Ahead Load Forecasting Based On Per-unit Curve Rotated Decoupling Strategy

In order to reduce the safety hazards and economic losses caused by load randomness to the power system,improving the accuracy of load forecasting has become one of the most important solutions.Currently,in the day-ahead load forecasting,the prediction results of existing decoupling strategies are prone to deflection under sudden load changes.In this paper,a new decoupling strategy,i.e.,the scalar curve rotation decoupling strategy,is proposed inspired by the existing decoupling strategies.Under this decoupling strategy,the day-ahead load prediction is decomposed into a load curve shape prediction and two single-step prediction problems,and these problems are investigated separately.To address the problem that the similarity day algorithm cannot take full advantage of large time span data sets and relies too much on a priori knowledge,this paper proposes the GraphSAGE-Conv(GSC)load scaling curve prediction algorithm based on the existing graph sample and aggregate(GraphSAGE)neural network.In this paper,each date is treated as a node,and the algorithm can automatically learn the scalene curve of neighboring nodes so as to learn the aggregation method between nodes,and then generate the scalene curve of the target node by the learned aggregation method.Finally,the effectiveness of the algorithm proposed in this paper for load curve shape prediction is verified by a comparative analysis with the back propagation neural network based on similar-day and the convolutional neural network based on similarday.Compared with the two comparison models,the MPAE of the GSC model decreased by 0.58%and 0.4%.For the existing recurrent neural network-like algorithms cannot achieve better prediction performance at large time steps,the hybrid temporal convolutional network(HTCN)single-step load prediction algorithm is proposed in this paper.The algorithm combines the advantages of a temporal convolutional neural network model with a great perceptual field for time series prediction and a hybrid inflated convolutional model that avoids grid effects to improve the learning ability of the model for detailed features at long time steps.Experimental results show that the HTCN model has more effective prediction performance compared to the temporal convolutional neural network,long and short-term memory network,and gated recurrent cell network for larger time steps.Finally,the HTCN model is used as a daily average load forecasting model in the framework of the existing static decoupling of the scalar curve and combined with the GSC load scalar curve forecasting algorithm in an early fusion manner for day-ahead load forecasting.The experimental results show that the GSC-HTCN model with the decoupling strategy further improves the prediction accuracy compared to the temporal convolutional neural network,long and short-term memory network with the recursive strategy,and multiple-input and multiple-output strategy.Compared with the CNN-TCN model,which is also based on the static decoupling of the scalar curve,the MAPE of the GSC-HTCN model is reduced by 1.43%.In response to the existing per-unit curve static decoupling strategy that deflects the predicted values of the load curve at some dates,this paper proposed a per-unit curve rotated decoupling method.The load curve is further decoupled into the rotated per-unit load curve,the load value at 0 a.m.and the daily average load value.The experimental results show that compared with the GSC-HTCN day-ahead load prediction algorithm based on the static decoupling strategy of the scalar curve,the MAPE of the proposed algorithm is reduced by 0.33%on average on the test set of normal days and 3.32%on average on the test set of sudden load change days.In summary,the proposed algorithm of day-ahead load forecasting based on the per-unit curve rotated decoupling method effectively improves the stability of the dayahead load forecasting for the extraction of daily load curve shape features,alleviates the deflection problem of load curve forecasting results,and provides a new idea for the day-ahead load forecasting research.