Research On Power Load Forecasting Based On SSA-SVR-SAGRU

Being one of the foundational industries in the country,the electric power sector plays a crucial role for all other industries in modern society.Accurate power load forecasting is of paramount importance for the economic dispatch and safe operation of the power grid.However,using a single traditional model for forecasting may result in a relatively poor fitting effect due to the instability and volatility of the power load data.To address this issue and improve the accuracy of power load forecasting,this paper proposes a parallel heterogeneous integrated network that combines data decomposition technology for forecasting.This paper introduces the related concepts of power load forecasting,describes the relevant theories of power load forecasting,discusses the selection of power load data,and explains the reasons for errors in power load forecasting while providing the forecasting evaluation index.After studying the singular model of power load forecasting,a power load forecasting model is proposed based on singular spectrum analysis.This model integrates the support vector regression model and the gated recurrent unit neural network based on the self-attention mechanism.In previous research,singular spectrum analysis has been used to handle power load modeling,with a focus on its noise reduction ability.In these approaches,the data is decomposed into a noise sequence which is then discarded.However,in this study,there is no clear method for defining the noise composition in power load data,as it is often determined by researchers based on factors such as the contribution rate.This manual determination of noise and subsequent discarding operations can negatively impact prediction accuracy.To address this,this paper emphasizes the feature extraction capability of the singular spectrum analysis algorithm.Specifically,the algorithm is used to extract features,decompose them into several sub-matrices including singular values based on their characteristics,and then reconstruct them into sub-sequences with trend,volatility,oscillation,and noise-containing aperiodicity.The obtained subsequences are then predicted using a parallel heterogeneous SVR-SAGRU network model,with the exception of the last subsequence which is predicted using the SAGRU network.Subsequences including noise are predicted using the SVR model.Finally,the prediction results of each subsequence are superimposed to obtain the final predicted value.This study utilizes a household power load sequence from the UK as a sample dataset for predicting power load data in advance.In order to analyze the prediction results and errors,several classic models are compared.The comparative analysis reveals the following findings:(1)the prediction accuracy of the model is influenced by the noise reduction processing of the original data;(2)each component of the proposed model contributes to the prediction;(3)the singular spectrum analysis method outperforms other decomposition methods in improving the prediction accuracy;and(4)generalization experiments are conducted using the regional electric load data set provided by ISO New England,and the proposed model demonstrates favorable generalization performance.