Prediction Model of Converter Oxygen Consumption Based on Recursive …
101
Fig. 1 The recursive division process of feature space. (Color figure online)
feature space as an example, as shown in Fig. 2. The first division is based on x
(3)
=
s 3 , and the second division is based on x
(2)
= s 2,1 and x
(2)
= s 2,2 , respectively. Finally,
four feature subspaces S 1,1 , S 1,2 and S 2,1 , S 2,2 are obtained.
In summary, the MSEP can make the MSEs of the corresponding output variables on the divided feature subspaces as small as possible, thereby simplifying the
distribution of oxygen consumption variables in the entire feature space. Usually, the
maximum recursive division depth and the minimum subset sample size are selected
as the stop division condition. The deeper the recursion depth or the smaller the
minimum subset sample size, the more refined the division of the feature space, the
smaller the MSE of the output variable in the subspace, and the simpler the oxygen
consumption distribution. However, the too refined division can easily lead to too
few samples, which in turn leads to over-fitting of the prediction model.
Fig. 2 Visual display of
recursive division of 3D
feature space. (Color figure
online)