| With the rapid development of power Io T,the power system is transforming into a smarter and more flexible interactive system.Because of the characteristics of electrical energy is not easy to store,to ensure the stable operation of the power grid,it is necessary to maintain the real-time power balance between the power generation side and the load side of the grid,which requires accurate prediction of future power demand,so as to provide data support for the development of power generation plans.However,with the massive connection of clean energy to the grid and the yearly increase of electric vehicle ownership,the increasing complexity and uncertainty of the grid has undoubtedly challenged the traditional power load forecasting models.Deep learning,as an end-to-end feature learning model,can model complex data by multi-layer nonlinear transformation.By using deep learning methods to learn features of electric load data,it is expected to improve the accuracy of short term load forecasting(STLF).Therefore,this paper mainly focuses on how to improve the accuracy of short term load forecasting based on deep learning theory.In order to accurately predict short-term electric loads,this paper considers both internal and external aspects,namely the intrinsic cyclical pattern of historical loads and their external influencing factors.First,the correlation analysis of input features is performed by calculating the Pearson correlation coefficients between historical load and external influencing factors to filter out strongly correlated features;then,an attention mechanism is introduced to highlight the input features that play a key role in load,and a BIGRU-Attention forecasting model is constructed that can consider both internal and external influencing factors and can focus on key variables.Based on the publicly available dataset,it is verified to have better prediction performance by designing comparison experiments.Second,in order to fully exploit the intrinsic information of historical load,the adaptive noise complete empirical modal decomposition technique is introduced,and subsequently,in order to solve the problems of complex model and long training time due to the large number of input dimensions and the introduction of attention mechanism,a CEEMDAN-CNN-BIGRU-Attention based model is proposed in combination with the powerful capability of convolutional neural network in feature extraction.Attention network-based short-term load prediction model.The experimental results show that the prediction method based on this hybrid network outperforms the Chapter 3 model in terms of both prediction accuracy and computationally consumed resources.Finally,the prediction method in this paper is applied to engineering projects to design and develop a short-term load prediction application system for power systems. |
