Abstract
The urbanization of central Florida has progressed substantially in recent decades, and the total population in Lake, Orange, Osceola, Polk, and Seminole Counties more than quadrupled from 1960 to 2010. The Floridan aquifer system is the primary source of water for potable, industrial, and agricultural purposes in central Florida. Despite increases in groundwater withdrawals to meet the demand of population growth, recharge derived by infiltration of rainfall in the well-drained karst terrain of central Florida is the largest component of the long-term water balance of the Floridan aquifer system. To complement existing physics-based groundwater flow models, artificial neural networks and other data-mining techniques were used to simulate historical lake water level, groundwater level, and spring flow at sites throughout the area.
Historical data were examined using descriptive statistics, cluster analysis, and other exploratory analysis techniques to assess their suitability for more intensive data-mining analysis. Linear trend analyses of meteorological data collected by the National Oceanic and Atmospheric Administration at 21 sites indicate 67 percent of sites exhibited upward trends in air temperature over at least a 45-year period of record, whereas 76 percent exhibited downward trends in rainfall over at least a 95-year period of record. Likewise, linear trend analyses of hydrologic response data, which have varied periods of record ranging in length from 10 to 79 years, indicate that water levels in lakes (307 sites) were about evenly split between upward and downward trends, whereas water levels in 69 percent of wells (out of 455 sites) and flows in 68 percent of springs (out of 19 sites) exhibited downward trends. Total groundwater use in the study area increased from about 250 million gallons per day (Mgal/d) in 1958 to about 590 Mgal/d in 1980 and remained relatively stable from 1981 to 2008, with a minimum of 559 Mgal/d in 1994 and a maximum of 773 Mgal/d in 2000. The change in groundwater-use trend in the early 1980s and the following period of relatively slight trend is attributable to the concomitant effects of increasing public-supply withdrawals and decreasing use of water by the phosphate industry and agriculture.
On the basis of available historical data and exploratory analyses, empirical lake water-level, groundwater-level, and spring-flow models were developed for 22 lakes, 23 wells, and 6 springs. Input time series consisting of various frequencies and frequency-band components of daily rainfall (1942 to 2008) and monthly total groundwater use (1957 to 2008) resulted in hybrid signal-decomposition artificial neural network models. The final models explained much of the variability in observed hydrologic data, with 43 of the 51 sites having coefficients of determination exceeding 0.6, and the models matched the magnitude of the observed data reasonably well, such that models for 32 of the 51 sites had root-mean-square errors less than 10 percent of the measured range of the data. The Central Florida Artificial Neural Network Decision Support System was developed to integrate historical databases and the 102 site-specific artificial neural network models, model controls, and model output into a spreadsheet application with a graphical user interface that allows the user to simulate scenarios of interest.
Overall, the data-mining analyses indicate that the Floridan aquifer system in central Florida is a highly conductive, dynamic, open system that is strongly influenced by external forcing. The most important external forcing appears to be rainfall, which explains much of the multiyear cyclic variability and long-term downward trends observed in lake water levels, groundwater levels, and spring flows. For most sites, groundwater use explains less of the observed variability in water levels and flows than rainfall. Relative groundwater-use impacts are greater during droughts, however, and long-term trends in water levels and flows were identified that are consistent with historical groundwater-use patterns. The sensitivity of the hydrologic system to rainfall is expected, owing to the well-drained karst terrain and relatively thin confinement of the Floridan aquifer system in much of central Florida. These characteristics facilitate the relatively rapid transmission of infiltrating water from rainfall to the water table and contribute to downward leakage of water to the Floridan aquifer system. The areally distributed nature of rainfall, as opposed to the site-specific nature of groundwater use, and the generally high transmissivity and low storativity properties of the semiconfined Floridan aquifer system contribute to the prevalence of water-level and flow patterns that mimic rainfall patterns. In general, the data-mining analyses demonstrate that the hydrologic system in central Florida is affected by groundwater use differently during wet periods, when little or no system storage is available (high water levels), compared to dry periods, when there is excess system storage (low water levels). Thus, by driving the overall behavior of the system, rainfall indirectly influences the degree to which groundwater use will effect persistent trends in water levels and flows, with groundwater-use impacts more prevalent during periods of low water levels and spring flows caused by low rainfall and less prevalent during periods of high water levels and spring flows caused by high rainfall. Differences in the magnitudes of rainfall and groundwater use during wet and dry periods also are important determinants of hydrologic response.
An important implication of the data-mining analyses is that rainfall variability at subannual to multidecadal timescales must be considered in combination with groundwater use to provide robust system-response predictions that enhance sustainable resource management in an open karst aquifer system. The data-driven approach was limited, however, by the confounding effects of correlation between rainfall and groundwater use, the quality and completeness of the historical databases, and the spatial variations in groundwater use. The data-mining analyses indicate that available historical data when used alone do not contain sufficient information to definitively quantify the related individual effects of rainfall and groundwater use on hydrologic response. The knowledge gained from data-driven modeling and the results from physics-based modeling, when compared and used in combination, can yield a more comprehensive assessment and a more robust understanding of the hydrologic system than either of the approaches used separately.