Regional Short-term Load Forecasting Model Combining Attention Mechanism And Deep Neural Network

The rise of the smart grid has enabled the connection of large-scale renewable energy power producing equipment and electric vehicle charging stations to the electricity grid.The power load is influenced by a range of complicated factors and exhibits randomness,nonlinearity,and time sequence characteristics,among others.Improving STLF’s accuracy has become a difficult undertaking.STLF’s research focus in recent years has shifted to intelligent forecasting approaches based on deep neural networks.Convolutional neural networks(CNNs)can mine complicated data features deeply,whereas long shortterm memory(LSTM)networks are capable of learning time series.The hybrid STLF model based on CNN-LSTM is examined in this thesis.The following are the major works:To begin,the 1DCNN-LSTM prediction model based on sequence feature extraction is developed to address the temporal and nonlinear aspects of load data.The historical load and temperature data are organized into feature series with the same time dimension as the load data.The coupling features of load and temperature of nearby nodes are extracted from feature series time dimension using the convolution kernel structure of one dimensional convolutional neural network(1DCNN),and then the extracted feature vectors are rebuilt into the form of time series and used as the input of LSTM for load prediction.The calculation results reveal that 1DCNN-LSTM combines the benefits of 1DCNN and LSTM and achieves a greater prediction accuracy than single models of 1DCNN,RNN,and LSTM.Second,in order to fully exploit the temporal peculiarities of load data,a2DCNN-LSTM prediction model based on time sequence module feature extraction is presented(SFE-2DCNN-LSTM).The time series feature module is established using the autocorrelation of load sequences and the correlation between load and temperature.Two separate two dimensional convolutional neural networks(2DCNNs)extract the time series feature of the load module and the nonlinear characteristic of the load-temperature module.The LSTM algorithm is in charge of mapping the retrieved feature to the load prediction value.The calculation results reveal that SFE-2DCNN-LSTM outperforms 1DCNN-LSTM in terms of ability to extract long sequence features and prediction accuracy.Simultaneously,in order to address the problem of multi-step load forecasting,daily load forecasting models based on SFE-2DCNN-LSTM are constructed for rolling strategy and multi-output strategy.The results indicate that a multi-output technique is preferable to a rolling strategy for multi-step daily load forecasting.Finally,an enhanced SFE-2DCNN-LSTM prediction model based on the attention mechanism is developed to further increase the accuracy of load forecasting.Aiming to address the issue that typical LSTMs have difficulty extracting high-dimensional characteristics from sequences and are prone to losing essential information when the sequence is too long.The mapping weighting of attention mechanism is used to assign different weights to the implicit state of the LSTM,highlighting the critical qualities that affect the load and assisting the model in determining the interdependence of long sequences.The results demonstrate that adding an attention mechanism improves the daily and weekly load prediction performance of SFE-2DCNN-LSTM,which has a greater accuracy and resilience than 1DCNN-LSTM,TCN,and Attention-LSTM.Simultaneously,it is applicable to both big and small areas with daily maximum loads approaching 5,000 MW and 2,000 MW.