Abstract: The dynamic fluid level (DFL) of an oil well is usually measured onsite by using the acoustic method. This method， however， has its limitation in determining the real-time DFL. Considering that Gaussian process regression (GPR) with single kernel function cannot significantly improve the prediction accuracy and generalization ability， a dynamic GPR for DFL with the combined kernel function optimized by artificial fish-swarm algorithm (AFSA) was proposed. The polynomial function， liner function and radial basis function were used to construct the combined kernel function of GPR in order to improve the generalization ability. The AFSA was used to optimize parameters of the combined kernel function in order to improve the prediction accuracy. The fast Fourier transform (FFT) and kernel principal analysis (KPCA) were used to extract nonlinear features of data in the time and frequency domain as the input variables of the model. The oil field application shows the validity of the proposed method.

Key words: oil well, dynamic fluid level, AFSA (artificial fish-swarm algorithm), combined kernel function, GPR (Gaussian process regression)