Time-series Analysis Using a Transfer Function Noise Model Interpolation of Groundwater Levels from File Hajdar, Gotland

Abstract

In File Hajdar, an area in Northern Gotland, the bedrock consist mainly of limestone and the groundwater levels exhibit large fluctuations between seasons. For over 40 years groundwater levels have been recorded with varying frequency resulting in heterogeneous time-series containing considerable intervals with no recorded data. The purpose of this thesis was to apply an appropriate method that could simulate the time-series to accurately depict the complex fluctuation patterns exhibited by the groundwater. A Transfer Function Noise (TFN) model was selected on the account of it being a prediction method used in the hydrogeological field. The chosen method was evaluated based on its sensitivity to the quality and quantity of data. The model was systematically evaluated in conjunction with site characteristics. This was done for the purpose of establishing an optimal simulation scenario that could potentially be applied to other areas with similar geological composition. The model acquired moderately high goodness of fit values with the majority of the adjusted R2-values being over 0.5. Groundwater pumping was also considered and marginal improvements could be observed in the simulations, with more representative pumping data further improvements can be achieved. Overall, the fit of the simulated time-series compared to the recorded time-series varied between bores, the fitted model consistently overestimated and/or underestimated the measured maximum and minimum groundwater levels. This was assumed to be a consequence of the model not accounting for all of the hydrological processes in the area, such as the considerable surface run-off, which provides scope for further research and improvements. The sensitivity analysis concluded that the TFN model had the ability to simulate time-series using shorter time-span while still producing high goodness-of-fit measures, but was determined not function optimally for heterogeneous time-series’. Furthermore, the TFN model was compared with common interpolation methods i.e. Nearest Neighbor and Cubic Spline. These methods were con to be more beneficial to use for time-series that contain smaller intervals of missing data.