不会错的方法,就是参考data : yarn
最后加一column "train"
train : Logical vector with TRUE for the training samples
         and FALSE for the test samples

TRUE  TRUE  TRUE  TRUE  TRUE  TRUE  TRUE  
TRUE  TRUE  TRUE  TRUE  TRUE  TRUE  TRUE  
TRUE  TRUE  TRUE  TRUE  TRUE  TRUE  TRUE
FALSE FALSE FALSE FALSE FALSE FALSE FALSE

21 for training(data),
7 for testing(newdata).

library(pls)
data(yarn)
nir.mvr <- mvr(density ~ NIR, ncomp = 5, data = yarn[yarn$train,])
predict(nir.mvr, comps = 1:4, newdata = yarn[!yarn$train,])


要不然predictdata,若要分开模拟的话.
ncol(predictdata) = length(nir.mvr$Xmeans)

#package example
library(pls)
data(yarn)

nir.mvr <- mvr(density ~ NIR, ncomp = 5, data = yarn[yarn$train,])
predict(nir.mvr, comps = 1:4, newdata = yarn[!yarn$train,])


density
110 51.04992
22  50.72019
31  32.01454
41  34.29076
51  30.35994
61  20.57832
71  19.07786

#自行输入数据,利用function predict
predictdata=matrix(scan(file='predictdata.txt'),7)
predict(nir.mvr, comps = 1:4, newdata = predictdata)

density
[1,] 51.04992
[2,] 50.72019
[3,] 32.01454
[4,] 34.29076
[5,] 30.35994
[6,] 20.57832
[7,] 19.07786

#自行输入数据,自行运算.
nobs=dim(predictdata)[1]

B <- rowSums(coef(nir.mvr, comps = 1:4), dims = 2)
B0 <- nir.mvr$Ymeans - nir.mvr$Xmeans %*% B
pred <- predictdata %*% B + rep(B0, each = nobs)
pred

density
[1,] 51.04992
[2,] 50.72019
[3,] 32.01454
[4,] 34.29076
[5,] 30.35994
[6,] 20.57832
[7,] 19.07786

predictdata.txt
大小:(13.7 KB)

 马上下载

首先谢谢大师帮忙，不过有一点我还是不理解。在偏最小二乘做回归时，是数据Y对提取成分T的回归。用来做预测的数据是X，是不能直接根据线性回归系数估计式代入就能得到预测值的呀。
所以#自行输入数据,自行运算.
nobs=dim(predictdata)[1]

B <- rowSums(coef(nir.mvr, comps = 1:4), dims = 2)
B0 <- nir.mvr$Ymeans - nir.mvr$Xmeans %*% B
pred <- predictdata %*% B + rep(B0, each = nobs)
pred

density
[1,] 51.04992
[2,] 50.72019
[3,] 32.01454
[4,] 34.29076
[5,] 30.35994
[6,] 20.57832
[7,] 19.07786
这部分好像不能直接用啊。
问题就转化为，R软件中有没有关于pls回归式直接转化为Y关于原数据X的回归的呢？
还有就是，comps = 1:4，这个是什么意思？

你有疑问的地方,我取自:
R package "pls" source code
predict.mvr.R  
line 42 - line45

42  ## Predict with a model containing the components `comps'
43  B <- rowSums(coef(object, comps = comps), dims = 2)
44  B0 <- object$Ymeans - object$Xmeans %*% B
45  pred <- newX %*% B + rep(B0, each = nobs)

comps = 1:4，是什么意思？
## Predicted responses for a single model with components 1, 2, 3, 4
comps = 1:4

## Predicted responses for models with 1, 2, 3 and 4 components
ncomp = 1:4

另补充:
建议你可参考matlab function plsregress.m
plsregress.m source code也许比较容易理解.
%matlab'example
load spectra
X = NIR;
y = octane;
[XL,yl,XS,YS,beta,PCTVAR] = plsregress(X,y,10);
yfit = [ones(size(X,1),1) X]*beta;
residuals = y-yfit;

%%%%
data(yarn)
在matalb执行的结果是:
Y1 = [ones(7,1) X1]*beta

Y1

51.0499
50.7202
32.0145
34.2908
30.3599
20.5783
19.0779


ps:
建议你看一下我发的这本书,有助观念分清.
Introduction to Multivariate Statistical Analysis in Chemometrics
4.7 Partial Least-Squares Regression

4.7.2 MATHEMATICAL ASPECTS

page 165/326 FIGURE 4.24

怎么得到方程呢

请问下，关于Warning message:
'newdata' had 6 rows but variables found have 12 rows
类似问题解决了吗？