Load Forecasting Method Based On Highway Neural Network And Attention Mechanism

In the context of grid energy interconnection transformation and intelligent construction,the factors affecting short-term power load are increasingly diversified.In addition to the influence of external weather,the role of distributed energy sources such as wind power,photovoltaics and electric vehicles is also superimposed,which brings great challenges to short-term load forecasting based on massive data.High-precision load prediction can promote the safety and stability of the power grid and the rational use of energy,and can also provide a strong guarantee for the planning,construction and dispatching and operation of the system.In view of the problems existing in the current method,such as insufficient data feature mining,unbalanced sample datas and weak model learning ability,based on the clustering of running days and data enhancement,a new short-term load forecasting method based on highway neural network and attention mechanism that considers multiple influencing factors is proposed in this paper,and the main research contents are as follows:(1)Aiming at the problem of insufficient data feature mining,the load feature analysis and influencing factors screening are proposed from qualitative and quantitative perspectives respectively.Firstly,in order to improve the quality of the model input data,a reasonable preprocessing scheme is developed for the sample set.Next,the characteristics of load change are qualitatively analyzed from different time dimensions,which provides guidance and reference for the classification of subsequent running days.Moreover,considering that external weather is the key factor affecting load change,the correlation law between temperature,humidity,rainfall,wind speed and load is first qualitatively analyzed,and then the correlation coefficient between each factor and the load is quantitatively calculated based on Pearson correlation theory,which provides a basis for the screening of input variables in the subsequent model.(2)Aiming at the imbalance of each training dataset,the data enhancement processing is completed based on k-means running-day clustering.Firstly,in order to avoid the burden of data mining caused by high-dimensional matrices,principal component analysis(PCA)is used to reduce the dimension of feature matrices,which creates conditions for visualization of clustering.Then,based on the k-means clustering technique,the running days are classified based on the dimensionality reduction matrix as input,and the classification modeling prediction scheme of the load is proposed according to the clustering results.Finally,an improved time GAN data enhancement method is proposed to generate time-series enhancement samples that conform to the distribution of real samples,and effectively expand the samples of each training set,laying a foundation for subsequent model training.(3)Aiming at the power load samples with complex nonlinear characteristics,a refined short-term load forecasting method based on Attention-HNN is proposed in this paper.Firstly,in order to solve the problems existing in the current deep network facing massive input samples,a new idea of applying the HNN to the field of short-term power load prediction is proposed.Then,considering the problem that long-term sequence input may lead to information loss,the introduction of Attention that highlights key features makes improvements to deep HNNs.Moreover,The Attention-HNN prediction models are constructed for different load clusterings.The influence of different parameters on the prediction effect are further explored.Then,according to the actual example of a city in Shandong Province,a variety of comparison methods are used to complete the prediction at the same time.The effectiveness and superiority of the proposed method is verified by testing.Finally,the generalization performance and applicability of the proposed method are also analyzed.