Artificial neural networks and time series modelling of TP concentration in boreal streams: a comparative approach

Abstract

Increased nutrient concentrations in fresh waters may lead to depressed dissolved oxygen concentrations, increased cyanobacterial biomass, and potentially high levels of cyanobacterial toxin production. Phosphorus loading, during snow melt and storm events, is the main source of nutrient enrichment to water bodies on the Canadian Boreal Plain. This study compared two approaches for modelling total phosphorus (TP) concentration: autoregressive moving average with exogenous input (ARMAX) and artificial neural network (ANN) models. Derived models were applied to a small forested watershed on the Canadian Boreal Plain. Results showed that ANN outperformed ARMAX based on four measures of goodness-of-fit statistics. This study confirmed that the ANN modelling approach is superior to ARMAX in modelling time-correlated gapped data, provided step-by-step guidelines for modelling time-correlated variables, and presented a feasible alternative for modelling diffuse pollutants in small forested watersheds. Key words: watershed, snow melt, diffuse pollutants, phosphorus, artificial neural networks (ANN), time series (TS), multi-slab hidden layer, ARMAX, boreal forest.