Research On End-to-end Probabilistic Forecasting Of Short-term Electricity Price

At this stage,China is in a critical period of power system reform.Electricity price reform is an important part of power system reform and the core of marketization.In a market-based electricity market,electricity prices are not only a signal carrier that regulates the relationship between supply and demand,but also an important factor that affects the economic interests of market participants.Precise electricity price prediction will help market participants to take an advantage in the fierce market competition.However,there are many factors influencing electricity prices.In addition to electricity demand,electricity prices are also seriously affected by many other external factors.As a result,electricity price is highly uncertain and difficult to predict.Therefore,electricity price prediction has become a research hotspot in power systems.Existing electricity price forecasting methods are mainly focused on point forecasting,that is,deterministic forecasting,and research on probability forecasting methods for electricity prices is still scarce.The point forecasts can only provide the possible values of future electricity prices,and cannot quantify the uncertainty in the electricity price series.Therefore,this paper proposes a short-term electricity price probability forecasting,algorithm based on end-to-end learning,which realizes the prediction of electricity price probability density distribution at a certain time in the future.This algorithm is different from previous probability prediction algorithms.It is modeled by convolutional neural network and label distribution learning forest,and implements joint training of probability prediction models in an end-to-end manner.In this study,the electricity price and its influencing factor sequence were first converted into a two-dimensional form and input into a convolutional neural network.The abstract features extracted by the convolutional neural network are then assigned to each split node of the label distribution learning forest.In the end,the final prediction is produced by the forest.The training process of the short-term electricity price probability density forecasting model proposed in this paper belongs to the category of supervised learning,and the true probability density distribution of electricity prices is difficult to obtain directly in reality.So on the basis of the above research,this article further carried out research on the true probability density distribution of electricity prices.Experiments on multiple public data sets show that our proposed short-term electricity price probability density prediction method has good generalization performance.The main innovations of this article are as follows:(1)In this paper,a convolutional neural network and label distribution learning forest are used to construct an end-to-end probabilistic prediction model,which combines feature extraction and probabilistic prediction modeling into a network model.In the model training process,joint optimization of tree learning and representation learning is realized.(2)Most of the existing probabilistic prediction methods use shallow models.This paper makes full use of the advantages of convolutional neural networks to extract features to extract abstract features in electricity price sequences.Compared with the features extracted manually,the electricity price data information contained in the features is more complex and comprehensive,which is more conducive to the expression of prediction results.(3)Based on the idea of neighbors,this paper uses the kernel density estimation technology to construct the probability density distribution of electricity price that best matches the characteristics of electricity price data.This distribution is used as a data label to participate in model training.