灰色预测- 灰色模型GM(1,N)
上篇博客说到了 G M ( 1 , 1 ) GM(1,1) GM(1,1),即只有单个变量,且模型是1阶的,此次我们来说说 G M ( 1 , N ) GM(1,N) GM(1,N),即 1 1 1 阶的,包含有 N N N 个变量的灰色模型。如果 G M ( 1 , 1 ) GM(1,1) GM(1,1)弄明白了,那么 G M ( 1 , N ) GM(1,N) GM(1,N)就简单了。
GM(1,N)模型
类似地,定义 N N N个变量的原始序列为:
x i ( 0 ) = ( x i ( 0 ) ( 1 ) , x i ( 0 ) ( 2 ) , . . . , , x i ( 0 ) ( n ) ) , i = 1 , 2 , . . . , n (1) x^{(0)}_i=(x^{(0)}_i(1), x^{(0)}_i(2),...,, x^{(0)}_i(n) ),i=1,2,...,n \tag{1} xi(0)=(xi(0)(1),xi(0)(2),...,,xi(0)(n)),i=1,2,...,n(1)
对应的一次累加序列 x i ( 1 ) x^{(1)}_i xi(1)为:
x i ( 1 ) = ( x i ( 1 ) ( 1 ) , x i ( 1 ) ( 2 ) , . . . , , x i ( 1 ) ( n ) ) , i = 1 , 2 , . . . , n (2) x^{(1)}_i=(x^{(1)}_i(1), x^{(1)}_i(2),...,, x^{(1)}_i(n) ),i=1,2,...,n \tag{2} xi(1)=(xi(1)(1),xi(1)(2),...,,xi(1)(n)),i=1,2,...,n(2)
其中,
x i ( 1 ) = ∑ j = α k x i ( 0 ) ( j ) , k = α , α + 1 , . . . n (3) x^{(1)}_i=\sum_{j=\alpha}^k x^{(0)}_i(j),k=\alpha,\alpha+1,...n \tag{3} xi(1)=j=α∑kxi(0)(j),k=α,α+1,...n(3)
其中 α ≤ n \alpha \le n α≤n,且为正整数, ( 3 ) (3) (3)式中,取 α = 1 \alpha = 1 α=1,称为一般累加过程,记作 1 − A G O 1-AGO 1−AGO。定义 x i ( 1 ) ( k ) x^{(1)}_i(k) xi(1)(k)的灰导数(实际上就是累减),为:
d i ( k ) = x i ( 0 ) ( k ) = x i ( 1 ) ( k ) − x i ( 1 ) ( k − 1 ) (4) d_i (k)=x^{(0)}_i(k)=x^{(1)}_i(k)- x^{(1)}_i(k-1) \tag{4} di(k)=xi(0)(k)=xi(1)(k)−xi(1)(k−1)(4)
取 x i ( 1 ) x^{(1)}_i xi(1)的等权重紧邻值,
z i ( 1 ) ( k ) = 0.5 x i ( 1 ) ( k ) + 0.5 x i ( 1 ) ( k − 1 ) , k = 2 , 3 , . . . , n (5) z^{(1)}_i(k)=0.5 x^{(1)}_i(k)+0.5 x^{(1)}_i(k-1) ,k=2,3,...,n \tag{5} zi(1)(k)=0.5xi(1)(k)+0.5xi(1)(k−1),k=2,3,...,n(5)
定义灰微分方程 G M ( 1 , N ) GM(1,N) GM(1,N):
d 1 ( k ) + a z 1 ( 1 ) ( k ) = ∑ i = 2 N b i x i ( 1 ) ( k ) , k = 2 , 3 , . . . , n (6) d_1(k)+az_1^{(1)}(k)=\sum _{i=2}^N b_i x_i^{(1)} (k), k=2,3,...,n \tag{6} d1(k)+az1(1)(k)=i=2∑Nbixi(1)(k),k=2,3,...,n(6)
即:
x 1 ( 0 ) ( k ) + a z 1 ( 1 ) ( k ) = ∑ i = 2 N b i x i ( 1 ) ( k ) , k = 2 , 3 , . . . , n (7) x^{(0)}_1(k)+az_1^{(1)}(k)=\sum _{i=2}^N b_i x_i^{(1)} (k), k=2,3,...,n \tag{7} x1(0)(k)+az1(1)(k)=i=2∑Nbixi(1)(k),k=2,3,...,n(7)
注意: 虽然我们在上面定义了
d
i
(
k
)
d_i (k)
di(k),和
z
i
(
1
)
z_i^{(1)}
zi(1)
(
i
=
1
,
2
,
.
.
.
,
n
)
( i=1,2,...,n )
(i=1,2,...,n),但是在灰微分方程中仅仅用到了
i
=
1
i=1
i=1。
同样地,将
k
=
2
,
3
,
.
.
.
,
n
k=2,3,...,n
k=2,3,...,n的数据带入上面的
(
8
)
(8)
(8)式,得到:
{ x 1 ( 0 ) ( 2 ) + a z 1 ( 1 ) ( 2 ) = ∑ i = 2 N b i x i ( 1 ) ( 2 ) x 1 ( 0 ) ( 3 ) + a z 1 ( 1 ) ( 3 ) = ∑ i = 2 N b i x i ( 1 ) ( 3 ) . . . . x 1 ( 0 ) ( n ) + a z 1 ( 1 ) ( n ) = ∑ i = 2 N b i x i ( 1 ) ( n ) (8) \left\{ \begin{matrix} x^{(0)}_1(2)+az^{(1)}_1(2)= \sum _{i=2}^Nb_ix_i^{(1)} (2) \\ x^{(0)}_1(3)+az^{(1)}_1(3)= \sum _{i=2}^Nb_ix_i^{(1)} (3) \\ .... \\ x^{(0)}_1(n)+az^{(1)}_1(n)= \sum _{i=2}^Nb_ix_i^{(1)} (n) \\ \end{matrix} \right. \tag{8} ⎩⎪⎪⎪⎨⎪⎪⎪⎧x1(0)(2)+az1(1)(2)=∑i=2Nbixi(1)(2)x1(0)(3)+az1(1)(3)=∑i=2Nbixi(1)(3)....x1(0)(n)+az1(1)(n)=∑i=2Nbixi(1)(n)(8)
将上面写成矩阵的形式,令:
Y = ( x 1 ( 0 ) ( 2 ) , x 1 ( 0 ) ( 3 ) , . . . , x 1 ( 0 ) ( n ) ) T Y = (x^{(0)}_1(2),x^{(0)}_1(3),...,x^{(0)}_1(n))^T Y=(x1(0)(2),x1(0)(3),...,x1(0)(n))T
B = ( − z 1 ( 1 ) ( 2 ) x 2 ( 1 ) ( 2 ) … x N ( 1 ) ( 2 ) − z 1 ( 1 ) ( 3 ) x 2 ( 1 ) ( 3 ) … x N ( 1 ) ( 3 ) ⋮ ⋮ ⋮ ⋮ − z 1 ( 1 ) ( n ) x 2 ( 1 ) ( n ) … x N ( 1 ) ( n ) ) B =\begin{pmatrix} -z^{(1)}_1(2) & x^{(1)}_2(2) & \dots & x^{(1)}_N(2) \\ -z^{(1)}_1(3) & x^{(1)}_2(3) & \dots& x^{(1)}_N(3) \\ \vdots &\vdots & \vdots& \vdots \\ -z^{(1)}_1(n) & x^{(1)}_2(n) & \dots & x^{(1)}_N(n) \\ \end{pmatrix} B=⎝⎜⎜⎜⎜⎛−z1(1)(2)−z1(1)(3)⋮−z1(1)(n)x2(1)(2)x2(1)(3)⋮x2(1)(n)……⋮…xN(1)(2)xN(1)(3)⋮xN(1)(n)⎠⎟⎟⎟⎟⎞
u = ( a b 2 b 3 … b N ) T u = \begin{pmatrix} a &b_2 &b_3 &\dots&b_N \end{pmatrix}^T u=(ab2b3…bN)T
称 Y Y Y为数据向量, B B B 为数据矩阵, u u u 为参数向量,则 G M ( 1 , N ) GM(1,N) GM(1,N)模型可以表示为矩阵方程:
Y = B u (9) Y = Bu \tag{9} Y=Bu(9)
(
8
)
(8)
(8)式可以看出,参数个数为
N
N
N个,方程个数
(
n
−
1
)
(n-1)
(n−1)一般大于
N
N
N,因此利用最小二乘求解待解未知数
u
u
u。最小二乘(使得
J
(
u
^
)
=
(
Y
−
B
u
^
)
T
(
Y
−
B
u
^
)
J(\hat u)=(Y-B\hat u)^T(Y-B\hat u)
J(u^)=(Y−Bu^)T(Y−Bu^)最小)解为:
u
^
=
(
a
^
b
^
2
b
^
3
…
b
^
N
)
T
=
(
B
T
B
)
−
1
B
T
Y
(10)
\hat u = \begin{pmatrix} \hat a &\hat b_2 & \hat b_3 &\dots&\hat b_N \end{pmatrix}^T=(B^TB)^{-1}B^TY \tag{10}
u^=(a^b^2b^3…b^N)T=(BTB)−1BTY(10)
需要注意的是, ( 10 ) (10) (10)式有解,必须满足 ( B T B ) (B^TB) (BTB)是非奇异矩阵,否则, ( B T B ) − 1 (B^TB)^{-1} (BTB)−1就不存在。但注意到 u ^ \hat u u^ 的元素实际上是各子因素对主因素影响大小的反映,因此,我们可以引入加权矩阵 W = d i a g ( w 1 , w 2 , . . . , w N ) W = diag(w_1,w_2 ,...,w_N ) W=diag(w1,w2,...,wN),使对各因素的未来发展趋势进行调整控制。对于未来发展减弱趋势的因素赋予较大的权值,而对于未来增强趋势的因素赋予较小的权值,使之更好地反映未来的实际情况。此时,向量 u ^ \hat u u^ 可以表示为:
u ^ = ( a ^ b ^ 2 b ^ 3 … b ^ N ) T = W − 1 B T ( B W − 1 B T ) − 1 Y (11) \hat u = \begin{pmatrix} \hat a &\hat b_2 &\hat b_3 &\dots&\hat b_N \end{pmatrix}^T=W^{-1}B^T(BW^{-1}B^T)^{-1}Y \tag{11} u^=(a^b^2b^3…b^N)T=W−1BT(BW−1BT)−1Y(11)
GM(1,N)的白化型
对于 G M ( 1 , N ) GM(1,N) GM(1,N)的灰微分方程 ( 7 ) (7) (7),如果将 x i ( 0 ) ( k ) x^{(0)}_i (k) xi(0)(k) 的时刻 k = 2 , 3 , . . . n k = 2,3,... n k=2,3,...n视为连续连续的变量 t ,则数列 x i ( 1 ) x^{(1)}_i xi(1)就可以视为时间 t t t 的函数,记为 x i ( 1 ) = x i ( 1 ) ( t ) , i = 2 , 3 , . . . N x^{(1)}_i = x^{(1)}_i (t),i=2,3,...N xi(1)=xi(1)(t),i=2,3,...N ,并让灰导数 x 1 ( 0 ) ( k ) x^{(0)}_1 (k) x1(0)(k) 对应于导数 d 1 x ( 1 ) d t \frac{d^{x(1)}_1}{dt} dtd1x(1),背景值 z 1 ( 1 ) ( k ) z^{(1)}_1 (k) z1(1)(k) 对应于 x 1 ( 1 ) ( t ) x^{(1)} _1(t) x1(1)(t) 。于是得到 G M ( 1 , 1 ) GM(1,1) GM(1,1)的灰微分方程对应的白微分方程为( G M ( 1 , N ) GM(1,N) GM(1,N)的白化型):
d 1 x ( 1 ) d t + a x 1 ( 1 ) ( t ) = ∑ i = 2 N b i x i ( 1 ) ( t ) (12) \frac{d^{x(1)}_1}{dt} +ax^{(1)}_1(t)=\sum _{i=2}^Nb_ix_i^{(1)} \tag{12} (t) dtd1x(1)+ax1(1)(t)=i=2∑Nbixi(1)(t)(12)
需要注意的是, G M ( 1 , N ) GM(1,N) GM(1,N)的白化型并不是 G M ( 1 , N ) GM(1,N) GM(1,N)的灰微分方程离散得到的,仅仅是一种类推。 G M ( 1 , N ) GM(1,N) GM(1,N)的白化型是一个真正的微分方程,如果白化型模型精度高,则表明所用数列建立的模型 G M ( 1 , N ) GM(1,N) GM(1,N)与真正的微分方程模型吻合较好,反之亦然。 G M ( 1 , N ) GM(1,N) GM(1,N)的白化型为一阶 N 个变量的微分方程。
GM(1,N)预测
通过将( G M ( 1 , N ) GM(1,N) GM(1,N)的白化型方程离散,得到 x 1 ( 1 ) x^{(1)}_1 x1(1)的预测值 x ^ 1 ( 1 ) \hat x^{(1)}_1 x^1(1), x 1 ( 1 ) x^{(1)}_1 x1(1)是由 x 1 ( 0 ) x^{(0)}_1 x1(0)累加得到的,因此可以通过累减 x ^ 1 ( 1 ) \hat x^{(1)}_1 x^1(1)得到 x 1 ( 0 ) x^{(0)}_1 x1(0)的预测值 x ^ 1 ( 0 ) \hat x^{(0)}_1 x^1(0),即:
x ^ 1 ( 0 ) ( k + 1 ) = x ^ 1 ( 1 ) ( k + 1 ) − x ^ 1 ( 1 ) ( k ) (13) \hat x^{(0)} _1(k+1)= \hat x^{(1)}_1 (k+1)-\hat x^{(1)}_1 (k) \tag{13} x^1(0)(k+1)=x^1(1)(k+1)−x^1(1)(k)(13)