Certain marine water quality or ecosystem model parameters vary in space and time because of different plankton taxonomic compositions over a large domain. The same parameter vectors can result in suboptimal calibration. In the present paper, a data-driven model based on an artificial neural network is developed to inverse the values of model parameters dynamically. All training data used are calculated using numerical water quality models from the results of multiparameter matching design cases such that physical properties are not disturbed. The aim is to determine the relationship between the model parameters and the pollution concentration values of interior stations. Field data are used in the analysis of the relationship for inversing optimal parameters. The temporal and spatial variations of sensitive parameters are considered using four inversion methods, namely, temporalspatial, spatial, temporal and non-temporal, and non-spatial, to enhance the model accuracy. In water quality models, an integrated element method is simultaneously applied using grids for spatial variation. Case studies in the Bohai Sea, China, and an identical experiment using dissolved inorganic nitrogen are conducted to validate the aforementioned methods. The average maximum of absolute error is reduced from 0.0435 to 0.00756, with a reduction rate of 82.62%. The results show that the temporal-spatial inversion method improves the accuracy of the water quality model.

Included in

Engineering Commons