Journal of Hydrologic Engineering

This paper examines the potential of the support vector machine (SVM) in long-term prediction of lake water levels. Lake Erie mean monthly water levels from 1918 to 2001 are used to predict future water levels up to $12\text{months}$ ahead. The results are compared with a widely used neural network model called a multilayer perceptron (MLP) and with a conventional multiplicative seasonal autoregressive model (SAR). Overall, the SVM showed good performance and is proved to be competitive with the MLP and SAR models. For a 3- to 12-month-ahead prediction, the SVM model outperforms the two other models based on root-mean square error and correlation coefficient performance criteria. Furthermore, the SVM exhibits inherent advantages due to its use of the structural risk minimization principle in formulating cost functions and of quadratic programming during model optimization. These advantages lead to a unique optimal and global solution compared to conventional neural network models.

This paper summarizes methods and results for a recent precipitation frequency study of the Kansas City metropolitan area. The study was motivated by prevalent concern that frequency estimates used for hydrologic design and analysis are based on studies published decades ago. Data currently used for design were extracted from National Weather Service (NWS) rainfall atlases TP-40 and HYDRO-35, published in 1961 and 1977, respectively. Results of this study are compared to those two atlases and a third regional study published by Huff and Angel in 1992. In general, the rainfall depths predicted by this study are higher than TP-40 and HYDRO-35 for return periods longer than $10\text{years}$ , yet for short durations and low return periods, the converse is true. For the most part, rainfall depths from this study are within a few percent of NWS estimates.

Simulation of whether precipitation arrives at the land surface in a liquid or frozen state can have a pronounced effect on the peak flows produced by a hydrologic model. While many hydrologic models rely on surface air temperature to discriminate precipitation type, both past studies and the current effort indicate that there can be substantial shortcomings in that approach. As an alternative, the $0°\mathrm{C}$ level detected by a vertically pointing radar is used in two modes, one by ingesting the $0°\mathrm{C}$ level directly into a hydrologic model to determine precipitation type at the land surface, and a second by deriving average surface air temperature thresholds using the radar data. Where snowfall and snow melt are important, using radar-detected $0°\mathrm{C}$ levels can provide a measurable improvement in simulated flow response. The derived surface air temperature thresholds recovered most of the benefit for streamflow prediction obtained by using the $0°\mathrm{C}$ level directly in the hydrologic model for the period concurrent with the radar observations, but these temperature thresholds were not generally transferable to other regions or periods.

Monthly rainfall data are needed in the simulation of water resources systems, and in the estimation of water yield from large catchments. Models to generate monthly streamflow data can be applied to generate monthly rainfall data, but this presents problems for most regions, which have significant months of no rainfall. This paper compares two established approaches for generation of monthly hydrological variables. These approaches are (1) the method of fragments modified so as to ensure accurate representation of over-year variability and persistence between the last month of the year and the first month of the next year, and (2) the nonparametric order-1 simulation model with long-term dependence, that considers aggregate variables representing the previous 12 months to impart long-term persistence in addition to the representation of a short-term order-1 Markovian dependence. The first of the two methods, while simpler to implement, has the limitation that it represents a disaggregation of an annual aggregate variable that is generated using a separate stochastic model. The second method, while more mathematically complex, introduces over-year or longer-term persistence through the use of an internally accounted aggregate variable, thereby removing the need to generate aggregate values separately. In this study both the methods are applied to generate rainfall data from ten rainfall stations located in various parts of Australia, and results compared to evaluate performance at both monthly and annual time scales. In the comparison performed both the models were found to preserve the annual and monthly characteristics adequately.

The transport and retention of a soybean oil-in-water emulsion was evaluated in laboratory columns packed with a medium to fine clayey sand amended with varying amounts of kaolinite. Results from these experiments demonstrated that appropriately prepared soybean oil-in-water emulsions can be distributed in clayey sand at least $80\mathrm{cm}$ away from injection point. Kaolinite addition to the clayey sand resulted in an increase in the maximum oil retention. However, the empty bed collision efficiencies in columns packed with clayey sand amended with kaolinite were lower than in columns packed with only clayey sand, suggesting that kaolinite is a less efficient collector of oil droplets than natural clayey sand. A standard colloidal transport model provided an adequate description of effluent breakthrough and the final oil distribution in the laboratory columns. This transport model was implemented as a user defined module within RT3D. Model parameters determined in replicate columns and at varying velocities were reasonably reproducible.

Injection of edible oils into the subsurface can provide an effective, low-cost alternative for stimulating anaerobic bioremediation processes. However, concerns have been raised about the effects of oil buoyancy and variations in aquifer permeability on the final distribution of oil in the subsurface. Three-dimensional (3D) sandbox experiments $(1.2\mathrm{m}\times 0.98\mathrm{m}\times 0.98\mathrm{m})$ were conducted to study the distribution of edible oil emulsions under homogeneous and heterogeneous conditions. A fine emulsion was first injected followed by chase water to distribute the emulsion throughout the sandbox. This approach was very effective, resulting in a reasonably uniform volatile solids distribution in the top, middle, and bottom layers, measured 5 and $7\text{weeks}$ after the completion of emulsion injection. A standard colloidal transport model that includes a Langmuirian blocking function was used to simulate emulsion transport and retention in the 3D sandbox. All parameters for the emulsion transport model were measured independently. Simulations results generally matched observed values for both the homogeneous and heterogeneous injection tests demonstrating that this approach can be used to describe the transport and distribution of emulsified oil in sandy sediments.

A hybrid, seasonal, Markov chain-based model is formulated for daily streamflow generation at multiple sites of a watershed. Diurnal increments of the rising limb of the main channel hydrograph were stochastically generated using fitted, seasonally varying distributions in combination with an additive noise term, the standard deviation of which depended linearly on the actual value of the generated increment. Increments of the ascension hydrograph values at the tributary sites were related by third- or second-order polynomials to the main channel ones, together with an additive noise term, the standard deviation of which depended nonlinearly on the main channel’s actual increment value. The recession flow rates of the tributaries, as well as of the main channel, were allowed to decay deterministically in a nonlinear way. The model-generated daily values retain the short-term characteristics of the original measured time series (i.e., the general shape of the hydrograph) as well as the probability distributions and basic long-term statistics (mean, variance, skewness, autocorrelation structure, and zero-lag cross correlations) of the measured values. Probability distributions of the annual maxima, means, and minima of the measured daily values were also well replicated.

Surface area specific runoff coefficients were measured for nonclogged Portland cement pervious concrete systems according to the rational method. The systems were simulated with pervious concrete blocks with porosities ranging from 16 to 27% placed over sand subbases. Rainfall was simulated in a flume setup with surface slopes ranging from 2 to 10%. There was negligible runoff for typical rainfall events under $100\text{years}$ frequency in South Carolina. Runoff rates were also simulated for a combination of direct rainfall and additional stormwater runoff from adjacent areas. The results obtained are important for the design of pervious concrete runoff management systems.

A newly developed watershed environmental hydrology (WEHY) model is presented as a state-of-the-art nonpoint source (NPS) model. The model consists of hydrologic and environmental modules, and describes environmentally relevant hydrologic processes based upon physically based governing equations to model the fate of pollutants such as sediment and phosphorus in the watershed. Unlike other physically based NPS models, the WEHY model is unique in its upscaling approach to the governing equations of hydrologic and environmental processes, which results in the governing equations that are compatible with the computational grid resolution while accounting for subgrid heterogeneities through upscaled model parameters. Upscaling was performed by means of a technique called ensemble averaging. The model was tested at the Ward Creek Watershed in Lake Tahoe Basin for its performance in a subalpine watershed setting. Comparisons of predicted and observed values were in good agreement and showed good promise of the approach used in the development of the model. Because of the physical basis of the WEHY model and its use of upscaled conservation equations, the model has the advantage of being applicable to ungauged basins and to large watersheds.

The determination of aquifer diffusivity from water level observations in various wells close to a stream, in response to an arbitrary variation in the stream stage, is discussed. The stream stage has been approximated by the sum of a series of basis functions, which are products of a linear and an exponential term. The aquifer response to individual basis function is investigated in detail. Approximate but very accurate expressions are proposed for explicit determination of diffusivity from data on peak attenuation and lag for cases where the stream stage may be represented by a single basis function. For other cases, a linear combination of 3–7 basis functions was found to represent a typical stage variation quite well. Application to a field data set demonstrates the suitability of the proposed methodology and the predicted aquifer diffusivity agrees with that reported in the literature.

Best management practices (BMPs) are often implemented as part of drainage policies using design methods generated from uncalibrated computer models even though their in-place effectiveness has not been validated using measured hydrologic data. A procedure for determining the data required to validate the hydrologic performance of a BMP is presented. The proposed validation procedure includes eight steps: (1) develop a conceptual model of the BMP; (2) identify important hydrologic performance criteria (e.g., peak reduction, trap efficiency); (3) perform a sensitivity analysis of the model; (4) identify storm characteristics that are important to BMP performance; (5) assess the uncertainty in performance; (6) identify the type of storms needed to be monitored; (7) use statistical theory to identify the total number of storms to monitor; and (8) based on hydrologic performance uncertainty, distribute the number of storms among the important storm characteristics. Analyses showed that proper BMP validation depends on the type of BMP, the hydrologic performance criterion, the number of criteria, and storm characteristics. It is not just the total number of storms that needs to be monitored for BMP validation, but the specific characteristics of the storms. For peak discharge control with a cistern, the barrel diameter and the distance between the low-stage and high-stage orifices are the important design variables, with approximately 92 storm events monitored.