Neural Network-based Short-term Load Forecasting Model

With the rapid development of technology and social economy,energy consumption,especially power consumption,is increasing day by day.This is a huge opportunity and challenge for the current power system.Fast and accurate short-term load forecasting(STLF)is the key to solving these problems.Short-term load forecasting is the main content in the economic dispatch of the power system and an integral module of the energy management system(EMS).Through accurate and reliable short-term load forecasting,the safety,dispatching efficiency and economic benefits of the power system will be significantly improved,which enables people to cope with the challenges posed by the diversified power demand and the complex power system in the future.Researchers have proposed many methods to meet the increasing requirements for reliability and accuracy in load forecasting.The artificial neural network(ANN)is one of the mainstream methods in load forecasting.Its powerful nonlinear mapping learning ability enables us to fully use the hidden relationship in historical load data to accurately forecast future load.Based on ordinary neural networks,several neural network structures have been proposed,such as the convolutional neural network(CNN)with strong feature extraction capabilities and recurrent neural network(RNN)with time series input processing capabilities have been proposed.These network structures are widely used in computer vision(CV),natural language processing(NLP),and nonlinear regression.In this thesis,the characteristics,mainstream methods and implementation steps of short-term load forecasting are deeply analyzed,and the structure and characteristics of neural networks of the current mainstream short-term load forecasting methods are studied.Subsequently,the neural network is applied to short-term load forecasting.Based on CNN and RNN,two high-performance short-term load forecasting models,convolutional residual network and residual Long Short-Term Memory(LSTM)network are proposed,and the model structure and characteristics are introduced.According to the different properties of these two models,multiple sets of experiments are designed to verify their accuracy,which proves the excellent performance of the two neural network models we proposed in short-term load forecasting.This thesis mainly contains the following contents:(1)First,the background,development history and common methods of short-term load forecasting are introduced.The basic types,daily load characteristics,influencing factors and implementation steps of short-term load forecasting are analyzed.The structure and principle of neural network are also studied,which will provide guidance for establishing short-term load forecasting models.(2)We present a model based on convolutional residual network and apply it to the short-term load forecasting problem,lay emphasis on the effects of deep residual networks with different depths,widths and block structures in dealing with nonlinear regression problems.Through multiple sets of controlled experiments,we obtain the best network architecture and the corresponding hyperparameters for short-term load forecasting.The model has higher forecasting accuracy than the existing models,which proves that our proposed convolutional residual network can handle load forecasting problems and still achieve state-of-the-art results.(3)Based on another widely used neural network RNN,a residual LSTM network combining deep residual network(DRN)and long short-term memory(LSTM)is proposed for short-term load forecasting.The proposed model not only inherits the DRN’s excellent characteristic to avoid vanishing gradient for training deeper neural networks,but also continues the LSTM’s strong ability to capture nonlinear patterns for time series forecasting.Moreover,through the dimension weighted units based on attention mechanism,the dimension-wise feature response is adaptively recalibrated by explicitly modeling the interdependencies between dimensions,so that we can jointly improve the performance of the proposed model from three aspects: depth,time and feature dimension.The snapshot ensemble method has also been applied to improve the accuracy and robustness of the proposed model.By implementing multiple sets of experiments on two public datasets,we demonstrate that the proposed model has high accuracy,robustness and generalization capability,and can perform STLF better than the existing mainstream models.