| CPC H02J 3/004 (2020.01) [G06N 3/0442 (2023.01); G06N 3/045 (2023.01); G06N 3/0464 (2023.01); H02J 2203/20 (2020.01)] | 9 Claims |
|
1. A method for predicting distributed regional generated power based on a stacked integrated model, comprising the following steps:
step 1, by a processor, dividing a predicted region into a plurality of sub-regions; using historical data of power stations in the predicted region as a data set; the historical data comprises a generated power x of each of the power stations, a total generated power of the sub-regions, a total generated power of the predicted region and a plurality of influence features H;
step 2, according to a degree of correlation with power output, selecting one or more key features h from the influence features H, by the processor;
step 3, by the processor, performing a normalization processing on the generated power x of the power stations, the total generated power of the sub-regions, the total generated power of the predicted region and the one or more key features h; dividing the data set after the normalization processing into a training set and a test set;
step 4, building the stacked integrated model by the processor;
the stacked integrated model comprises a basic model and a meta-model; the basic model takes key features of a representative power station as input, extracts features through Convolutional Neural Network, (CNN) network and Long Short-Term Memory, (LSTM) network respectively, and weights and fuses output features of the CNN network and the LSTM network by using a multi-head-attention mechanism; advanced features extracted by the multi-head-attention mechanism are transformed into a predicted total power of the sub-regions of a final predicted output through a fully connected layer; the meta-model is a Gated Recurrent Unit, (GRU) network; the meta-model takes a predicted total power of each of the sub-regions output by the basic model as input, and a predicted value of a predicted total power of the predicted region as output;
step 5, by the processor, training the basic model by using the training set, and inputting the test set into a trained basic model to output and obtain a predicted total generated power of the sub-regions; combined jointly training the meta-model using the training set and the test set;
step 6, by the processor, selecting a distinct representative power station for each of the corresponding sub-regions and obtaining corresponding historical generated power and key parameters; inputting the historical generated power and the key parameters of into the basic model to predict a future total generated power of the sub-regions, and inputting a predicted future total generated power of all the sub-regions into the meta-model to predict a future total generated power of the predicted region;
wherein the selecting a distinct representative power station for each of the corresponding sub-regions comprises: constructing a graph attention network, (GAT) for each of the sub-regions, and predicting generated power of the power stations of the sub-regions in corresponding time period by inputting the key parameters; comparing the predicted generated power with an actual generated power, and taking a corresponding power station with a smallest error as the representative power station in the sub-region.
|