Application Of Deep Learning-based Recurrent Neural Network In Short-term Load Forecasting

Short-term load forecasting is an important work of power system dispatching operation department,especially with the establishment and development of power market,short-term load forecasting will play an increasingly important role.Its prediction accuracy will have a great influence on the safety and economy of power system operation.In the traditional load prediction model,the cyclic neural network has shown remarkable advantages and potentials in load prediction due to its special treatment of time series.With the advent of the era of big data and the development of deep learning,the accuracy of traditional load prediction model has more room for improvement.The main problem studied in this paper is the application of deep learning-based cyclic network in the field of short-term load prediction,which mainly includes the following contents:First to present various predictive models of load forecast,and their respective advantages and disadvantages and applicable scope of generalizations,and combined with the research status at home and abroad,analyzes the deep learning and its related algorithm in the application and development prospect in the field of load forecasting,determines the loop based on deep learning neural network as the main method of load forecasting.Then learn to summarize a number of studies in the selection of input character types,to deal with load data,autocorrelation analysis method is used to determine the maximum Lag factor of the relevant historical data Lag = 13,considering load cyclical and spatial similarity,will determine the load sequence,the day before the load sequence,weeks as input characteristic greatly similar load sequence.Then,the deep learning model is used to mine the characteristics of historical load sequence,and the deep learning network architecture is built.Under this framework,LSTM,GRU and CNN are used as the kernel for modeling,and the model is optimized by means of grid enumeration search.Finally,when the hidden layer is 1 and the hidden node of the fully connected layer is 5,the prediction set error MAPE of LSTM is 2.35% when the cycle depth is 2,the prediction set error MAPE of GRU is 2.03% when the cycle depth is 4,and the prediction set error MAPE of CNN is 2.87% when the convolution layer depth is 4.In view of the problems such as the difficulty in training and the long training time of traditional circular network,the difficulty in selecting the super parameters and the low optimization efficiency of deep learning network,the improvement of network efficiency and super parameter optimization efficiency are the further research directions.Inspired by the latest achievements in the field of deep learning,Res Net and Dense Net added residual structure in the simple cyclic neural network.Through experiments and selection,the residual structure(Res RNN)was constructed by adding a full connection layer to the residual structure,supplemented by a standardized layer.It is proved by an example that Res RNN has advantages in training cost and network performance compared with LSTM and GRU.Finally,considering that currently in the field of load prediction,researchers usually rely on subjective experience or make a lot of attempts by enumerating parameter combinations in the super-parameter optimization of deep learning models.This paper proposes a residual loop algorithm based on bayesian superparameter optimization: the algorithm adopts a more efficient residual loop network,establishes the proxy function by gaussian process,and then optimizes the collection function step by step,and finally determines the optimal superparameter combination.In the example analysis,after 20 times of super parameter optimization,the model error MAPE obtained by the residual cycle algorithm based on bayes super parameter optimization was reduced by 0.28%,saving 138 min of optimization time,and providing a new idea for automatic and efficient optimization of the load prediction model based on deep learning.