Research On Short-Term Power Load Forecasting Based On CNN And BiGRU Combination Model

Electricity is the power base of modern technological development,the security and stability of the power system controls the economic lifeblood of a nation or region,it has a profound impact on all the sectors of industrial and agricultural production and people’s lives.Accurate and reliable short-term power load forecasting will provide a scientific foundation for the planning and dispatch of the power system,power generation arrangements,unit maintenance and other plans,and is of great importance in guaranteeing the quality of power supply and making plans for the development and construction of the power system.This paper addresses the problem of low accuracy of the traditional single forecasting model and considers the characteristics of load data,combining deep learning therapy and cluster of intelligent optimization algorithms for short-term electricity load forecasting,with the following main research elements:(1)To address the problems of missing data and data anomalies in the original data set,the missing items were processed using the mean fill method;the outliers were detected by the standard deviation method,and then the outliers were corrected using the horizontal and vertical methods.In addition,considering the complex factors influencing load prediction,the input variables of the prediction model were screened using the Pearson correlation coefficient method,and finally the maximum temperature,minimum temperature,average temperature,date type and historical load values were identified as the input features of the prediction model.(2)According to the nonlinear and complex characteristics of load data,a combined forecasting model combining CNN,BiGRU and Attention is proposed,which is called CNNBiGRU-Attention model for short.Firstly,CNN is used to perform the preliminary information extraction of the feature variables of the input data,then the feature vector is trained by BiGRU in two-way timing to further explore the deeper features of the load data,after which the Attention mechanism is introduced to allocate various weights to the output states of BiGRU to highlight the key features and weaken the non-important features,and finally the prediction results are output.The experiments prove that the evaluation indexes of the proposed model are better than those of other single and combined models and have better prediction performance.(3)To address the problem that the Sparrow Search Algorithm(SSA)is prone to partial optima and bad performance of convergence in the later stages of the search,an improvement method is proposed to introduce non-linear inertia weights and vertical and horizontal crossover strategies.The sparrow population was first initialised by the Cubic method to enhance population diversity,then non-linear inertia weights are used to improve convergence speed.Then,horizontal crossover was introduced to improve the global search capability of SSA,and vertical crossover was introduced to maintain community diversity and improve the local exploitation capability of SSA.Finally,the effectiveness of the improved strategy is verified by testing on five benchmark functions.Meanwhile,to address the problem of difficult parameter selection for the CNN-BiGRU-Attention model,the improved sparrow search algorithm is proposed to optimize the model parameters.The improved sparrow search algorithm is shown to have a stronger generalization ability and the prediction accuracy is further improved.