基于条件变分自编码器和贝叶斯神经网络的短期电力负荷概率预测

Probabilistic Short-term Load Forecasting Based on Conditional Variational Autoencoder and Bayesian Neural Network

考虑到输入信息和预测模型的不确定性对负荷预测结果的影响,本文提出一种基于条件变分自编码器和贝叶斯神经网络的短期电力负荷概率预测方法.通过条件变分自编码器生成指定天气因素预测值和日历特征条件下实际天气因素可能的多个随机样本,以模拟天气预测信息的不确定性;构建GRU-S2S贝叶斯神经网络学习模型参数的分布特征,以反映预测模型的不确定性,并结合MC dropout技术获得多个可能的负荷预测值;遍历天气因素全部模拟样本,将预测模型输出的负荷预测值构成集合,并通过核密度估计获得预测时段内各时刻预测负荷服从的概率分布.实际算例结果表明,该方法在短期负荷概率预测中具有更高的分位数预测精度和更可靠稳定的区间预测结果.

Considering the influence of the uncertainty of inputs information and prediction model on the results of load prediction,a novel method of probabilistic short-term load forecasting based on conditional variational autoencoder and Bayesian neural network is proposed in this paper.Firstly,a conditional variational autoencoder is adopted to generate the multiple possible samples of actual weather factors under the specified calendar and prediction weather factors to simulate the uncertainty of weather prediction values.By learning the probabilistic distribution characteristics of model parameters,a GRU-S2S Bayesian neural network is built to characterize the uncertainty of the prediction model.Therefore,The values of multiple possible load prediction using the MC dropout technique are output.Traversing all the simulated samples of weather factors,the values of possible load prediction at each time point in the forecast period are output and the corresponding set is formed.Finally,the probability distribution of the prediction load at each forecasting time point is obtained by kernel density estimation.The practical examples show that the proposed method has higher accuracy of quantile prediction and more reliable and stable results of interval prediction in probabilistic short-term load prediction.