An Improved Autoencoder and Partial Least Squares Regression-Based Extreme Learning Machine Model for Pump Turbine Characteristics

Chu Zhang,Tian Peng,Jianzhong Zhou,Jie Ji,Xiaolu Wang,Zhang,Peng,Zhou,Ji,Wang

DOI: https://doi.org/10.3390/app9193987

2019-09-23

Applied Sciences

Abstract:Complete characteristic curves of a pump turbine are fundamental for improving the modeling accuracy of the pump turbine in a pump turbine governing system. In view of the difficulty in modeling the "S" characteristic region of the complete characteristic curves in the pump turbine, a novel Autoencoder and partial least squares regression based extreme learning machine model (AE-PLS-ELM) was proposed to describe the pump turbine characteristics. First, a mathematical model was formulated to describe the flow and moment characteristic curves. The improved Suter transformation was employed to transfer the original curves into WH and WM curves. Second, the ELM-Autoencoder technique and the partial least squares regression (PLSR) method were introduced to the architecture of the original ELM network. The ELM-Autoencoder technique was employed to obtain the initial weights of the Autoencoder based extreme learning machine (AE-ELM) model. The PLS method was exploited to avoid the multicollinearity problem of the Moore-Penrose generalized inverse. Lastly, the effectiveness of the proposed AE-PLS-ELM model has been verified using real data from a pumped storage unit in China. The results demonstrated that the AE-PLS-ELM model can obtain better modeling accuracy and generalization performance than the traditional models and, thus, can be exploited as an effective and sufficient approach for the modeling of pump turbine characteristics.