BOD Soft-sensing Based on SONNRW

Aim  ing at the problem that biochemical oxygen demand (BOD) soft-sensing is difficult to be forecasted accurately and in real time, a self-organizing neural network with random weights(SONNRW) is proposed by using the sensitivity analysis method. Firstly, we select the original auxiliary variables using mechanism analysis, then we use PCA method to select variables after the data preprocessing. The selected variables are the input SONNRW to forecast the key water quality parameter BOD. The residual error’s sensitivities to the hidden nodes are defined by their outputs and weight vectors connecting to the out layer using the sensitivity analysis method. First, we calculate the sensitivities by using the hidden layer outputs and the responding output layer weight vectors. Then the orders are sorted based on the sensitivity of each hidden node. Then those nodes which have lower sensitivities will be pruned by using the Leave-one-out method. By using this method in BOD soft-sensing, experiments show that the pruning NNRW has high prediction accuracy, more streamlined network size and better generalization performance.