1 点估计

1.1 归一化均方根误差

${\rm{nRMSE = 100} }\; \times \;\sqrt {\frac{ {\sum\nolimits_{i\; = \;1}^{ {N_t} } {\sum\nolimits_{t\; = {t_0}\;}^T | } {x_{i,\;t} } {\rm{ - } } {\mu _{i,\;t} } { {\rm{|} }^2} } } { { {N_t}(T - {t_0})} } } \;\%$

1.2 权重均值绝对百分比误差

${\rm{WMAPE = 100} }\; \times \;\frac{ {\sum\nolimits_{i = 1}^{ {N_t} } {\sum\nolimits_{t = {t_0} }^T | } {x_{i,\;t} } {\rm{ - } } {\mu _{i,t} } {\rm{|} } } } { {\sum\nolimits_{i = 1}^{ {N_t} } {\sum\nolimits_{t = {t_0}\;}^T | } {x_{i,t} } {\rm{|} } } }\%$

2 概率估计

2.1 平均覆盖误差（ACE）

首先求出预测区间覆盖概率：
${\rm{PICP} } = \frac{ {\sum\nolimits_{i\; = 1}^{ {N_t} } {\sum\nolimits_{t\; = {t_0}\;}^T { {c_{i,\;t} } } } } } { { {N_t}(T - {t_0})} }$,{c_i} = \left\{ {\begin{array} {*{20} {c} } {1,{x_{i,\;t} } \in [{\rm{U} }_{i,\;t}^{\;\alpha \;},\;{\rm{L} }_{i,\;t}^{\;\alpha \;}]\;}\\ {0,\;{x_{i,\;t} }\; \notin [{\rm{U} }_{i,\;t}^{\;\alpha \;},\;{\rm{L} }_{i,\;t}^{\;\alpha \;}]\;} \end{array} } \right.
得平均覆盖误差：${\rm{ACE = |} }\;{\rm{PICP - 100} }({\rm{1 - } }\alpha )\% \;{\rm{|} }$

2.2 综合锐度与准确率分数

预测区间归一化平均宽度：${\rm{PINAW} } = {\rm{100} } \times \frac{ {\sum\nolimits_{i = 1}^{ {N_t} } {\sum\nolimits_{t = {t_0};}^T {({\rm{U} }{i,;t}^{;\alpha ;} } } - {\rm{L} }{i,;t}^{;\alpha ;})} } { { {N_t} \times (T - {t_0}) \times {\rm{R} } } }% $

综合锐度与准确率分数：{\rm{I} } { {\rm{S} }^\alpha }({x_{i,\;t} }) = \left\{ {\begin{array} {*{20} {l} } { - 2\alpha ({U^\alpha }({x_{i,{\kern 1pt} t} }) - {L^\alpha }({x_{i,{\kern 1pt} t} })) - 4({L^\alpha }({x_{i,{\kern 1pt} t} }) - {x_{i,{\kern 1pt} t} }),{\mkern 1mu} if{\mkern 1mu} {x_{i,{\kern 1pt} t} } < {L^\alpha }({x_{i,{\kern 1pt} t} })}\\ { - 2\alpha ({U^\alpha }({x_{i,{\kern 1pt} t} }) - {L^\alpha }({x_{i,{\kern 1pt} t} })),\;\quad \qquad \qquad if{\mkern 1mu} {x_{i,{\kern 1pt} t} } \in P{I^\alpha }({x_{i,{\kern 1pt} t} })}\\ { - 2\alpha ({U^\alpha }({x_{i,{\kern 1pt} t} }) - {L^\alpha }({x_{i,{\kern 1pt} t} })) - 4({x_{i,{\kern 1pt} t} } - {U^\alpha }({x_{i,{\kern 1pt} t} })),{\mkern 1mu} if{\mkern 1mu} {x_{i,{\kern 1pt} t} } > {U^\alpha }({x_{i,{\kern 1pt} t} })} \end{array} } \right.

其中${ {U^\alpha }({x_{i,{\kern 1pt} t} })}$为置信上限，${ {L^\alpha }({x_{i,{\kern 1pt} t} })}$为置信下限。

3 部分概率预测方法评估

在文章An Integrated Missing-Data Tolerant Model for Probabilistic PV Power Generation Forecasting中，给出了LSTM、ANN（人工神经网络）、GP（高斯过程回归）的概率预测效果表，观察到，LSTM类的综合得分最高，表现优良。