Due to the fact that it is difficult to measure the chemical oxygen demand (COD) of sewage water environment parameters, a kind of gray level theoretical prediction model based on radial basis function neural network (GM-RBF) is proposed. The proposed GM-RBF can predict the chemical oxygen demand. The gray theory is used to predict the development and change of the system behavior, and the precision of the prediction model can be improved by combining the high precision approximation ability of the radial basis function neural network. The modeling and prediction of the key water quality parameters in the process of wastewater treatment is studied. The results show that the model can predict the COD with high accuracy, and the prediction is close to the actual value.

In order to determine the number of hidden layers in continuous deep belief network (CDBN), a determination method for the depth of CDBN based on the reconstruction error is proposed. CDBN is composed of multiple continuous restricted Boltzmann machines (CRBM). By analyzing the relevance between the reconstruction error and the network energy, and setting the threshold of the reconstruction error, the decision mechanism for the depth is designed to realize self-organizing adjustment for the depth of CDBN. The experiments show that the determination method for the depth of CDBN based on the reconstruction error can determine the optimal depth of CDBN and improve the efficiency of decision depth for CDBN.

The traditional mutation strategy of differential evolution algorithm can not reach a good balance between the global search
and the local search and the operators are constant． The differential evolution algorithm leads to premature convergence and the low
search efficiency． Based on analysis of the performance of the optimization strategies，a hybrid mutation strategy is proposed in this paper．
The scheme attempts to balance the exploration and exploitation abilities． In this way，emphasis is laid on the global search at the
beginning，which results in maintaining the diversity of population． Later，contribution from the local search increases in order to converge
to the optimal faster． Meanwhile，the random normal scaling factor F and the time － varying crossover probability factor CR are
used synchronously to improve the performance of DE． Finally，the modified differential evolution algorithm is tested on benchmark
functions． The simulation results show that the modified algorithm can effectively avoid the premature convergence，as well as modified
the global convergence ability and the search efficiency remarkably．