Using particle flters for short-term groundwater level predictions

Abstract

Particle fiter is a Bayesian inference method and provides a way to make predictions in dynamic systems where no assumption of linearity or a Gaussian distribution of uncertainty can be made. For these reasons, it was considered a suitable method for making a short-term prediction of groundwater levels. Groundwater level predictions allow farmers to preemptively water or drain their land and thus improve crop production. In the process of developing the particle lter six steps were elaborated: 1) developing a weather model; 2) de ning the initial distribution; 3) calculating the prediction; 4) updating the prediction; 5) resampling for the next prediction and 6) calculating the outcome. Three methods were developed for the prediction step to account for applying the particle lter to a 14-days prediction. These methods were compared with a conventional hydrological model to assess prediction performance for a 48-days period. Upon testing these three methods, the characteristics of each method are described. Overall, there is not one method that outperforms the hydrological model that was taken as a baseline. Combining the methods and in this way complementing the individual strengths of each method is of interest for future research.