Research On Short-term Power Load Forecasting Based On Hyperparameter Optimization Neural Network Combined Model

As the scale of the power grid expands,how to set reasonable intra-day generation plans has become a major problem.Short-term power load forecasting predicts future loads by processing historical data,providing decision support for power companies to formulate generation plans and implement power dispatching schemes.Therefore,a good forecasting model can save generation costs and ensure the smooth operation of the power system.To improve the accuracy of short-term power load forecasting models,this paper proposes a combined prediction model called CNN-BiGRU-AT,which integrates Convolutional Neural Network(CNN),Bidirectional Gating Recurrent Unit(BiGRU),and Attention Mechanism(AT).Furthermore,an improved sparrow search algorithm is employed to optimize the hyperparameters of this combined model,further enhancing the predictive accuracy of the model.(1)A CNN-BiGRU-AT short-term power load forecasting model is constructed to solve the problem that the characteristic of different time series have inconsistent influence on power load.This model uses CNN to extract effective feature vectors from the historical load sequence,while BiGRU learns the deep relationship between the feature vectors extracted by CNN and the load,and introduces the AT mechanism to assign different probability weights to BiGRU’s hidden states,further enhancing the impact of important information in the feature sequence.(2)On the one hand,in response to the problem of low efficiency of traditional manual parameter tuning optimization methods in the CNN-BiGRU-AT combined model,it is proposed to use the Sparrow Search Algorithm(SSA)to explore the hyperparameter space through global search,find better hyperparameter combinations for the combined model,reduce parameter tuning time,and improve the prediction accuracy of the combined model.On the other hand,the SSA may fall into local optima when optimizing hyperparameters.To address this issue,an improved Sparrow Search Algorithm(ISSA)is further proposed,Firstly,a nonlinear adjustment factor is introduced in the original SSA to update the discoverer’s position;Secondly,the Levy flight and T-distribution perturbation algorithms are combined to update the joiner’s position;Finally,the elite backward learning method is introduced to enhance the optimization performance during algorithm iteration.Using ISSA for hyperparameter optimization of the CNN-BiGRU-AT combined model,two sets of simulation experiments have verified that the CNN-BiGRU-AT combined model optimized by ISSA achieves the highest prediction accuracy.