Font Size: a A A

Research On Combined Forecasting Of Power Load Based On Recurrent Neural Network

Posted on:2019-02-20Degree:MasterType:Thesis
Country:ChinaCandidate:T WangFull Text:PDF
GTID:2382330563992465Subject:Computer system architecture
Abstract/Summary:PDF Full Text Request
With the rapid development of power information,new technologies such as big data and artificial intelligence are urgently needed to promote the development of the power industry.Power load forecasting plays a very important role in modern power systems.It is an important basis for grid scheduling and power grid planning,and provides basic data for many applications such as grid operation and planning.The accuracy and stability of the power load forecast not only affect the rationality of the electricity use plan,but also affect the safe operation of the grid.For the characteristics of multiple sources of power load data,different companies will produce different data sets,pointing out that the current power load forecasting model cannot cross data set prediction problems,and propose a combined forecasting model to achieve cross-data set learning and high prediction accuracy.The combined prediction model is mainly based on an improved clustering model and a tuned recurrent neural network.For the problem that the original combination model has low prediction accuracy,a method for optimizing the structural parameters of the combined forecasting network model and a method for reducing over-fitting risk in the model training process are proposed,and the prediction accuracy of the combined model is effectively improved.In view of the choice of recurrent neural network,the working principle of the recurrent neural network is analyzed and studied in detail,and the predictive performance of multiple recurrent neural network models is compared and tested.The importance of data processing in power load forecasting is pointed out with respect to the characteristics of power load data mass.Different processing methods may cause serious deviations in prediction results.In order to improve the prediction accuracy of the model,a data suitable for power load forecasting is proposed.Processing solutions: Micro-processing and improved data processing steps including data set expansion,feature quantification,data cleaning,and standardization.In addition,a power management system based on the user side is designed and implemented.The system is a multi-functional integrated power management platform integrating on-line monitoring,energy efficiency evaluation,data mining,and artificial intelligence forecasting,and it will be based on a recurrent neural network-based power load combination.The prediction model is embedded in the system and is mainly used for power information data mining and intelligent prediction.Finally,the data set consisting of data from 100 companies in a certain area of Jiangsu Province during the 20 months was processed and used for combination forecasting model training.The test results show that the combined forecasting model of power load based on recurrent neural network can be used for multi-dataset prediction,and the average prediction error on the entire dataset is as low as 3.54%.
Keywords/Search Tags:Power Load, Recurrent Neural Network, Combination Forecasting Model, Long-Short Term Memory Network, Gated Recurrent Unit Network
PDF Full Text Request
Related items