Short-term Probabilistic Load Forecasting Based On Convolution Neural Network

In future smart grids,distributed renewable energy sources,active loads,power markets and other factors affect the load,and the complexity and uncertainty of load will increase greatly,which results in a serious challenge to power system security and stability.Short-term probabilistic load forecasting(STPLF)providing the load variant interval and its occurrence probability,is one of the most effective and practical methods for addressing this serious situation,and plays a key role in optimizing resource allocation,reducing power system operating cost and ensuring the safety and stability of power system operation.The forecasting method based on artificial intelligence(AI)is the recently research hotspot of STPLF.Convolutional neural network(CNN)effectively overcomes the overfitting problems of traditional AI methods owing to its outstanding performance on feature extraction and generalization.In this paper,STPLF methods based on CNN are studied systematically,whose main innovative achievements are as follows:(1)The improve architecture of CNN for STPLF is designed.The STPLF model based on the traditional CNN has some problems to be solved,such as low efficiency of feature extraction,matching confusion between influence factors and weight coefficients in convolution calculation.With the consideration of these problems,the paper improves the architecture of CNN based on the characteristics of STPLF data and the convolution computer theory.The improvement measures involve the adaptive designs of input data,convolution kernel and activation function.Case study shows that the improved architecture of CNN enhances the STPLF precision.(2)The stochastic-batch gradient descent(S-BGD)method and gradient pile(GP)method for training CNN network are proposed.CNN bears a low speed of training the parameter by the batch gradient descent(BGD)method.Aiming at the problem,the S-BGD method is proposed.Based on the use of the stochastic gradient descent(SGD)method with the rapid training,the proposed algorithm possesses the advantages of both SGD and BGD.Furthermore,BGD method searches for the optimal solution along the direction of the negative gradient of the parameters,which easily lead to the training result falling into the local optimum and saddle point.To address the problem,the improved GP method is put forward under the reference of kinematic theory.This method searches for the optimal solution along the direction of the negative cumulative gradient,which enhances the global search ability,and then improves the training effect of CNN.(3)A parametric STPLF method based on CNN and Bootstrap is proposed.Parametric STPLF methods usually estimate the load probability distribution(LPD)based on the existing distribution function,which are easy to calculate and have advantages in application scenarios with few samples.As one of the most effective parametric methods,the Bootstrap method can flexibly and exactly estimate the forecasting errors caused by the forecasting data and model,so as to more accurately estimate the forecasting error distribution.In this paper,a parametric STPLF method based on CNN and Bootstrap is proposed.This method generates LPD by combining the expected value of forecasting load and the forecasting error distribution.The case results show that the proposed parameter STPLF method possess a superior prediction effect than several other classical parametric STPLF methods.(4)A nonparametric STPLF method based on CNN and quantile regression(QR)is proposed.Nonparametric STPLF methods avoiding presupposing the probability distribution function,are beneficial to improve the prediction accuracy.As a representative nonparametric probability distribution estimation method,QR method has outstanding performance in application scenarios with multi samples.In this paper,a nonparametric STPLF method based on quantile CNN(QCNN)is proposed to improve the prediction accuracy of STPLF.The QCNN first employs CNN to mine the deep features of the nonlinear relationship between influencing factors and load,and then conveys the extracted features to a QR model.The QR model is used for regression analysis and output load quantiles.However,the training of QCNN model is seriously affected by the nondifferentiable training objective.To solve the training problem of the QCNN,a two-stage training strategy is proposed.In the first training stage,the feature extraction network based on CNN is extended to a load point-forecasting model.The parameters of feature extraction network are obtained by training the load point-forecasting model.In the next training stage,the training optimization problem of QR model is transformed into a dual linear programming problem,which can obtain the parameters using analytical methods.Case study shows that the prediction precision of the improved QCNN method is evidently better than other classical nonparametric STPLF methods.(5)A novel STPLF method based on CNN and load range discretization(LRD)strategy is proposed.To further promote the prediction accuracy,a novel STPLF method based on CNN and LRD is proposed.The LRD method constructs the initial discrete LPDs(including the load value and the corresponding probability)of training samples by discretizing the load range.Then,the optimal estimation is employed to further optimize the discrete LPDs for training samples.As a result,the training samples containing probability information can be utilized to train the CNN forecasting model,so that the model can forecast LPD directly.The case results show that the proposed STPLF method possess a superior prediction effect than several other classical parametric and nonparametric STPLF methods.