An improved convolutional neural network with load range discretization for probabilistic load forecasting

Electricity load forecasting plays a vital role in power system planning and operations. Probabilistic forecasting is expected to become a popular load prediction form due to providing more uncertainty information for the decision of smart grid. As one of the promising forecasting methods, the convolutional neural network has an outstanding advantage in feature extraction. However, there is a critical problem that needs to be solved when using a convolutional neural network for probabilistic load forecasting. The classical parametric and nonparametric techniques for generating probability distribution suffer from the predetermined load probability distribution types or the nondifferentiable training function, which might affect the prediction accuracy of the convolutional neural network. In this paper, a load range discretization method is proposed to generate load probability distributions. The method constructs discrete load probability distributions by segmenting the load range. Then, the optimal estimation is employed to optimize the load probability distributions for training samples. As a result, the samples can be utilized to train the convolutional neural network, so that the network can forecast load probability distributions directly. There is no probability distribution assumption and nondifferentiable training function in the proposed method. Based on the data of independent system operators in New England, the superiority of the proposed method is verified by comparing with 7 well-established benchmarks. The proposed method acquires more reliable and sharper load probability distributions, which can be beneficial to various decision-making activities in power systems. (C) 2020 Elsevier Ltd. All rights reserved.