The support vector machine (SVM) has recently been successfully used for classification and regression problems. As the core of SVM, the kernel function is used to map the input vectors into a higher dimensional feature space to realize the non-linear algorithm for original space. The selection of kernel function and corresponding parameters directly affects the learning ability and generalization ability of the support vector regression (SVR). This paper compares the performance of different kernel functions on the energy efficiency data obtained from the UCI repository. The kernel functions used are as follows: Linear, Gaussian, Polynomial, Sigmoid, Inv. Multiquadric and the Semi Local kernel. The influence of parameters on the mean square error of cross validation (MSECV) for each kernel is shown in the paper. The combination of SVR parameters and kernel parameter with minimum MSECV obtained by grid search method based on cross-validation is chosen for SVR model building. The experiment results based on the dataset showed that the Semi Local kernel has the best predictive capability. The mean-squared error of prediction (MSEP) of the Semi Local kernel was 97.13%, 8.26%, 0.96%, 99.95%, 27.36% lower than the value of the Linear, Gaussian, Polynomial, Sigmoid and Inv. Multiquadric kernel.

Motivated by a global economy that is shaped by the landscape changes in energy and manufacturing, research on intelligent manufacturing in which data mining is of great importance is studied. In this paper, an energy monitoring platform for small and medium-sized enterprises (SMEs) is first introduced, which monitors and records the energy consumption of manufacturing processes at various levels of granularity. In processing the collected data, some incompleteness in the data due to various factors needs to be addressed first otherwise it may lead to the inaccurate portrayal of the system and poor generalisation of the resultant model. Hence, a novel OLS (orthogonal least square)-based autoencoder is proposed for the imputation of missing values. This approach is to learn the representative code from the original samples by constructing an improved encoder network in which the hidden neurons are orthogonal with each other. The new samples are then generated through the associated decoder network. The proposed approach selects the hidden neurons one by one based on the OLS estimation until an adequate network is built. The classical techniques are compared to verify the effectiveness of the proposed algorithm. In the experiment, the two real industrial data sets from a baking process and a polymer extrusion process are adopted and the percentage of missing values varies from 0.02 to 0.25. The experimental results confirm that the proposed method offers stable performance in the presence of different missing ratios, and it outperforms significantly alternative approaches while the missing ratio is greater than 0.05.