MODELLING HUMIC SUBSTANCE DERIVED COLOURATION OF SURFACE WATERS IN PEATLAND CATCHMENTS USING ARTIFICIAL NEURAL NETWORKS

In upland peat catchments the geochemical properties of the catchment basin and physico-chemical properties of the hydrological system can result in high raw water colour in surface water reservoirs. The high colour is particularly attributed to the release of humic substances originating from sedimentary organic carbon contained within the peat soils into aqueous solution. The complex problem is impractical to model deterministically. Thus, there is an inherent inability to anticpitate and effectively remove these colour causing humic substances. This results in aesthetic water quality failure and high levels of potentially harmful disinfection by-products become embedded in drinking water supplies. This paper explores a new approach to the modelling of surface water colour using the learning-based computer paradigm of Artificial Neural Network technology. Based on this technique, future levels of surface water colour are forecast by mapping input-output relationships of an historical data-set. Geochemical, hydrological (including raw water colour) and meteorological data have been collected over a twenty year period for a case study catchment area. Derived from these historical data, temperature ranges, precipitation, and potential evaporation rates were demonstrated to be the important input variables with which to train a model to learn the input-output relationships in order to predict future colour. Time was a further factor conditioning the processes that contributed to the colouration of surface water. A dynamic Artificial Neural Network was necessary for modelling time-dependency. Therefore a Time-Lagged Recurrent Network was configured. The effect of time was inferred by training the model with a window series of inputs combined with its adaptive feedback mechanisms, which form part of the network topology, in order to generate time-dependent relationships. The final methodology derived for the case study catchment was verified using data from another simple catchment. Further verification was undertaken using data from a more complex catchment which involved more than one impounding surface water source.