基 于 分 位 数 插 值 和 深 度 自 回 归 网 络 的 光 伏 出 力 概 率 预 测

This paper proposes a novel probability prediction method of photovoltaic output combining the probabilistic model of quantile linear interpolation with the deep autoregressive recurrent neural network, which can make up for the shortcomings of traditional methods in probability modeling and describing complex nonlinear relationships. First, a probability model for photovoltaic output is built based on the quantile linear interpolation, which can comprehensively and accurately depict probability distribution of photovoltaic output under various circumstances. Then, continuous ranked probability score (CRPS) is used as the loss function for the training prediction model, and the closed analytic expression of the CRPS integral is derived for model training to ensure the feasibility and efficiency of training. Finally, a deep autoregressive recurrent neural network is adopted to model time-series of photovoltaic output. Combined with the proposed probability model of photovoltaic output, a new photovoltaic output probability prediction method is formed. Numerical results of the case study demonstrate that the proposed method can provide reliable and high-quality probability prediction for photovoltaic output.