Physically based hydrological models
cannot replicate every detail of the true system, thus deviations between model
forecasts and observed values are inevitable. In many groundwater applications
these deviations are commonly addressed by adding additional complexity and
parameters to the model. However, due to the many sources of uncertainty, it
is possible that model-to-measurement deviations are not unbiased but rather
have both temporal and spatial structure that may be able to be captured using
data-driven models. Data-driven models can be trained to map the prediction
errors as a function of the measured data, model inputs, outputs and other variables.
The forecasted error is then simply superimposed upon the prediction from the
physically based model to get a more accurate result that can be extended over
a long-term prediction period and to unsampled locations. It is expected that
this approach may become more widely considered in the future, particularly
with increasing use of real-time sensors that will give temporally and spatially
richer datasets for assessment of model errors.
This paper illustrates the application of three data-driven models, namely, artificial neural network (ANN), K-nearest neighbors (KNN) and instance based weighting (IBW), for forecasting MODFLOW head prediction errors and subsequently updating the predictions of head at existing observation wells. A synthetic dataset, generated based on a phytoremediation site at the Argonne National Laboratory, was used to evaluate the performance. Compared to the standard calibration approach that adds additional model parameters to improve the future forecast, the proposed approach shows a significant enhancement to model prediction accuracy even for a considerably long forecast lead-time.