Short-Term Load Rolling Forecasting Method Based On Attention Mechanism And Bi-Directional Long Short-Term Memory Network

Short-term load forecasting plays an important role in the planning and operation dispatching of power systems.In the context of development of smart grid,the wide use of distributed generations with high volatility brings new challenges to the accuracy,quickness and intelligence of short-term load forecasting.At the same time,the extensive use of smart measuring devices such as smart electricity meters also provides new opportunities and means for short-term load forecasting based on big data and artificial intelligence methods.Therefore,this thesis studies on short-term load forecasting based on artificial intelligence,and the main work is as follows.A novel bi-directional long short-term memory(Bi-LSTM)network model based on attention mechanism for short-term load rolling forecasting is proposed.Firstly,rolling mechanism is utilized to transmit the latest known data information continuously to prediction model in real time,making the input data of the model more effective.Secondly,influence weights are assigned through attention mechanism to highlight the effective characteristics of the input variables.Thirdly,a Bi-LSTM network is used for model training.At the same time,due to the influence of rolling mechanism,the Bi-LSTM network also updates the model parameters periodically to obtain a new prediction model that can better reflect the operating state of system.Finally,the data trained by Bi-LSTM network is output through the linear transformation layer and softmax layer to obtain the final prediction results of loads.Common load data set from the New South Wales(NSW)in Australia is employed to verify the validity of the proposed prediction model.The experiment is designed to predict the loads for whole week at the end of four different quarters respectively.Experimental results show that the introduction of attention mechanism and rolling mechanism can improve the load prediction accuracy of model,whether it is the overall forecasting of whole week or its daily forecasting.Compared with other prediction models,the mean absolute percentage error(MAPE)and the root mean square error(RMSE)of the proposed model are both the smallest in terms of the overall forecasting of whole week or the forecasting of each day in different quarters.In conclusion,the test set verifies that the proposed model can better process and analyze the input data of model,reflect the latest operating state characteristics of system,and thus show higher prediction accuracy and better generalization ability.