Research On Short-Term Electricity Load Forecasting Based On Deep Fusion Algorithm

Short-term load forecasting(STLF)is the key technology to make scientific power supply plan and maintain the balance between supply and demand of the smart grid.It is also the foundation of power market operation and an important part of the power planning.Improving the accuracy of the short-term load forecasting is helpful in promoting the utilization ratio of power equipment and reducing the energy consumption.The actual electricity load data is usually non-stationary,and it includes complex random patterns.Due to strong volatility of the electricity load,it creates the challenge for the load forecasting.Thus,this thesis combines the advantages of different deep learning methods to construct the forecast model for accurate STLF.The main research contents are as follows:(1)The thesis proposes a convolutional temporal load forecast model based on the attention mechanism.The model is used to relieve the problem that the existing models do not extract the crucial features from the historical load and temperature data at different time scales,and it causes the decline in load forecasting performance of the model.The model comprises a One Dimensional Convolutional Neural Network(1DCNN),a Bidirectional Long Short-term Memory Network(BiLSTM)and attention mechanism.The model is abbreviated as Attention-1DCNN-BiLSTM.This thesis constructs the input feature set and adopts three independent1 DCNNs to extract the crucial feature from different variables,such as historical load,temperature,and calendar data,respectively.Then,the extracted information is used as the input of the BiLSTM to learn the temporal feature between data.Finally,the attention mechanism can help model learn the essential information of the input data,improving the forecasting performance.The experimental results show that the Attention-1DCNN-BiLSTM model can effectively combine the advantages of the 1DCNN and the BiLSTM.Moreover,the proposed method can produce more accurate forecasted result compared to CNN-LSTM,MCL,and the PCL.(2)The thesis proposes a DCRB-GRU load forecast model based on the Densely Connected Residual Block(DCRB)and Gated Recurrent Unit(GRU).The model is adopted to overcome the problem that input data cannot satisfy the space invariance,and it causes the decline in load forecasting performance of CNN.The DCRB can efficiently extract the essential features from the above multi-scale input data than the traditional CNN-based models.In addition,the extracted feature can affect the change trend of the load series.Then,the GRU can help model effectively learn the temporal correlation between the input data.The experimental results indicate that the DCRB-GRU model can promote the load forecasting performance.(3)The thesis proposes a BasicNet-Enhanced ResNet load forecast model based on the basic network(BasicNet)and an enhanced residual network(Enhanced ResNet)for STLF.The model is used to relieve the problem that the deep neural networks often suffer from the vanishing gradients and network degradation.Firstly,the model adopts the input feature set as the input of the basic network,which can help model capture the change of the electricity load and obtain the preliminary forecasted value.Then,it is fed into an enhanced residual network to generate the final forecasted result for the model.The experimental result indicates that the BasicNetEnhanced ResNet model can produce the accurate forecasted result.