Short-term Electrical Load Forecasting Based On Deep Learning

In power systems,power load forecasting is an important prerequisite for power system scheduling,planning,and equipment overhaul.With the advent of the energy Internet,a large number of electric vehicle charging piles,wind power and photovoltaic power generation equipment will be connected to the power grid,and the fluctuation of the load will be more complicated and elusive.Accurate power load forecasting is one of the indicators of the energy Internet.Accurate power load forecasting is conducive to the optimal allocation of various energy sources in the power system,thereby saving energy consumption.The neural network can theoretically describe the complex relationship between arbitrary nonlinear time series through the mapping relationship between input layer,hidden layer and output layer.Therefore,neural network is widely used in power load forecasting.At present,most static prediction models such as BP neural network are used for load prediction.However,the change of load sequence is closely related to the change of historical load.The static model can only describe the relationship between several input variables and output values.Most of the existing load forecasting techniques are point predictions.Probabilistic predictions are in the initial stage of research,and probabilistic predictions often provide more detailed information to relevant personnel.In view of the above research status,this paper,two dynamic prediction models based on long short-term memory networks are proposed to make deterministic and probabilistic predictions of load.The first model is a short-term load forecasting model based on convolutional neural networks and long short-term memory networks.When using long short-term memory networks for load forecasting,considering the powerful ability of convolutional neural networks(CNN)to extract data features,In this paper,the convolutional neural network is used to extract the load and its related meteorological data,a nd then the feature vector data is used as the input data of LSTM,which uses its internal memory unit to dynamically describe the load sequence changes.In order to further improve the accuracy of load forecasting,the adaptive moment estimation(ADAM)algorithm is used to optimize the weight and threshold of long short-term memory networks to obtain load forecasting results.It is verified that the proposed CNN-LSTM model is better than the basic LSTM model and DNN model in prediction accuracy.The second model is a probabilistic prediction model based on convolution-long short-term memory network quantile regression(CNN-LSTMQR)and kernel density estimation(KDE).The CNN-LSTMQR model can fit the load forecasting results under different quantile points,and thus describe the possible fluctuation range of future load forecasting.By introducing the kernel density estimation,a complete probability density function can be provided for each forecasting point.Experiments show that the proposed CNN-LSTMQR-KDE can provide a reliable prediction interval for the load.