Short-Term Power Load Forecasting Based On ISSA And GRU-Transformer

The new energy system,which primarily relies on electricity,has gradually improved in recent years as the "carbon neutral" strategy has been promoted.Forecasting power load is a crucial task in the planning and management of the public power system and is essential to the secure functioning of the power system.For the power system’s decision-making,accurate short-term power load forecasting can offer a significant level of assurance.Thus,it is crucial to utilize more practical and effective techniques to increase the short-term power load forecasting’s accuracy.At the moment,machine learning and deep learning are the most widely used techniques in research on short-term power load prediction.There are still some issues,though,including a lack of input features caused by an incomplete selection of the power load’s influential factors,low load prediction accuracy brought on by the power load prediction model’s simplistic structure,and locally optimal selection of the model’s super parameters based on experience.Based on the aforementioned issues,the primary focus of this essay is as follows:1.Power load data analysis and preprocessing.In order to make the data entered into the prediction model as valid data,preprocessing methods are studied in light of the issues with missing values,outliers,and data irregularities in power load data.In addition,in light of the issue with incomplete influencing factors in power load forecasting,a preliminary analysis of influencing factors based on experience is carried out to determine the influencing factors related to power load.A Pearson correlation coefficient study was conducted to ascertain the correlation between each influencing component and the power load in light of the uncertainty surrounding the correlation of the input model’s influencing factors.2.Create a model for predicting power load that consists of an input fully connected layer,an output fully connected layer,and a Gate Recurrent Unit(GRU).The recurrent input fully connected layer integrates complex factors that affect power load and uses two-layer GRUs to improve learning capability.The output complete connection layer combines distinguishing elements with gathered historical load data.To uncover the intricate mapping relationship between input characteristics and load and to address the issue of information loss in lengthy sequences of short-term power load prediction,the encoder structure of Transformer is employed as a feature extractor.To successfully dig high-order properties of influencing elements and compensate for the limitations of GRU network learning,a prediction model of GRU-Transformer is constructed.3.Create the GRU-Transformer model based on the Improved Sparrow Search Algorithm(ISSA)to address the issue of challenging super parameter selection.introduction of adaptive learning method to enhance global search capacity of discovery,reverse learning strategy to initialize the sparrow number,and Levy flight technique to easily avoid local optimum in iteration populations.ISSA has been shown to have a faster convergence time and accuracy than Sparrow Search Algorithm(SSA)and Particle Swarm Optimization(PSO)on the identical kernel function4.A power load dataset is used as an example for an analysis to determine the daily and overall forecast errors for the dataset.The experimental findings demonstrate that the proposed ISSA-GRU-Transformer prediction method is more accurate than the LSTM,GRU,Transformer and other models,and the values of various error indicators are lower,further demonstrating the viability and superiority of the model proposed in this article.