Recurrent Self-Organizing RBF Neural Network for Effluent TN Prediction

The patent introduces auxiliary variables (NH4+-N, NO3--N, effluent SS, BOD, TP) as intermediaries to indirectly determine effluent TN concentration. Instead of directly measuring TN offline, the system uses these easily measurable parameters as inputs to the RSORBFNN model, which computes the TN concentration in real-time. This intermediary approach enables online prediction while maintaining acceptable accuracy.

The patent replaces the traditional chemical-mechanical measurement system (alkaline potassium persulfate UV spectrophotometry) with a computational system (RSORBFNN). The neural network model substitutes the physical-chemical measurement process, using adaptive second-order training algorithms and growing-pruning structure optimization to achieve real-time predictions without the time-consuming offline analysis procedures.

The patent creates a virtual copy of the measurement process through the RSORBFNN model. Instead of developing physical sensors to directly measure TN, the system copies the relationship between auxiliary variables and TN concentration from historical data, then uses this computational model to predict TN in real-time. This virtual measurement approach achieves real-time detection capability without the need for expensive new hardware sensors.

The patent makes the existing measurement system multi-functional by using the same infrastructure to measure multiple parameters (NH4+-N, NO3--N, SS, BOD, TP) and then combining these measurements through the neural network to also provide TN concentration. This universal approach allows one system to perform both traditional measurements and real-time TN prediction without requiring dedicated new sensors.

The patent implements a dynamic neural network structure through the growing-pruning algorithm. The RSORBFNN model adapts its structure during operation, adding hidden neurons when prediction accuracy needs improvement and removing them when they are no longer necessary. This dynamic adjustment allows the model to maintain high prediction accuracy while automatically optimizing its complexity to avoid unnecessary intricacy.

The patent enables the model to self-optimize through automatic structure design and parameter training. The growing-pruning algorithm and adaptive second-order training method allow the RSORBFNN to automatically adjust its own structure and parameters without requiring manual intervention or complex external optimization systems. This self-service capability simplifies implementation while maintaining high prediction accuracy.