New methods based on a genetic algorithm back propagation (GABP) neural network and general regression neural network (GRNN) for predicting the occurrence of trihalomethanes in tap water

Monitoring trihalomethanes (THMs) levels in water supply systems is of great significance in ensuring drinking water safety. However, THMs detection is a time-consuming task. Developing predictive THMs models using parameters that are easier to obtain is an alternative. To date, there is still no application of optimization algorithms and general regression neural networks in predicting disinfection by-products levels. This study was to explore the feasibility of back propagation neural network (BPNN), genetic algorithm back propagation (GABP) neural network and general regression neural network (GRNN) for predicting THMs occurrence in real water supply systems. The results showed that the BPNN models' prediction ability was limited (test rp = 0.571–0.857, N25 = 61.5 %–91.5 %). Optimized by the genetic algorithm (GA), GABP models were generated and exhibited better prediction performance (test rp = 0.573 and 0.696–0.863, N25 = 68.2 %–93.6 %). However, GABP models took a lot of time and their prediction performance was unstable. A GRNN was then used to generate simpler neural network models, and the resulting prediction performance was excellent (total trihalomethanes and bromodichloromethane: test rp = 0.657–0.824, N25 = 81.8 %–100 %). In general, GRNN was the best at predicting THMs concentrations among the three models. However, it is worth noting that the prediction accuracy of bromodichloromethane (BDCM) was not high, which may be due to the absence of key parameters affecting BDCM formation. Accurate predictions of THMs by GRNN with these nine water parameters made THMs monitoring in real water supply systems possible and practical. [Display omitted] •Prediction ability of BPNN was limited (test rp = 0.571–0.857, N25 = 61.5 %–91.5 %).•GABP exhibited better prediction performance (test rp = 0.573 and 0.696–0.863).•The performance of GABP was unstable due to the randomness of population.•GRNN well predicted T-THMs and DBCM levels (test rp = 0.657–0.824, N25 = 81.8 %–100 %).•The convergence time of GRNN was the shortest (t = 11.047 s).