Decision support for dam release during floods using a distributed biosphere hydrological model driven by quantitative precipitation forecasts

Transcription

1 WATER RESOURCES RESEARCH, VOL. 46,, doi: /2010wr009502, 2010 Decision support for dam release during floods using a distributed biosphere hydrological model driven by quantitative precipitation forecasts Oliver C. Saavedra Valeriano, 1,2,3 Toshio Koike, 1 Kun Yang, 4 Tobias Graf, 5 Xin Li, 6 Lei Wang, 1 and Xujun Han 6 Received 29 April 2010; accepted 24 May 2010; published 30 October [1] This study proposes a decision support system for real time dam operation during heavy rainfall. It uses an operational mesoscale quantitative precipitation forecast (QPF) to force a hydrological model and considers the forecast error from the previous time step, which is introduced as a perturbation range applied to the most recent QPF. A weighting module accounts for the location, intensity, and extent of the error. Missing precipitation intensities within contributing areas and information from surrounding areas can both be considered. Forecast error is defined as the ratio of QPF to the observed precipitation within an evaluation zone (sub basin, basin, buffer, or total domain). An objective function is established to minimize the flood volume at control points downstream and to maximize reservoir storage. The decision variables are the dam releases, which are constrained to the ensemble streamflow s information. A prototype was applied to one of the most important river basins in Japan, the Tone reservoir system. The efficiency of the approach was evident in reduced flood peaks downstream and increased water storage. The results from three events indicate that the developed decision support system is feasible for real life dam operation. Citation: Saavedra Valeriano, O. C., T. Koike, K. Yang, T. Graf, X. Li, L. Wang, and X. Han (2010), Decision support for dam release during floods using a distributed biosphere hydrological model driven by quantitative precipitation forecasts, Water Resour. Res., 46,, doi: /2010wr Introduction [2] Heavy precipitation events are very likely to continue and become more frequent as a result of global warming [Intergovernmental Panel on Climate Change (IPCC), 2007]. These events can be attributed to large variations of the water cycle at global [e.g., Huntington, 2006] and regional scales [e.g., Kyselý and Beranová, 2009], while flood damage occurs at the basin scale [Kojiri et al., 2008]. Actually, the frequency of these events has increased over most land areas, which is consistent with increases in land surface temperature. This is particularly evident in humid regions affected by tropical cyclones, such as Southeast Asia. These typhoons often bring heavy rainfall but can cause severe flooding. To 1 Department of Civil Engineering, University of Tokyo, Tokyo, Japan. 2 Now at Department of Civil Engineering, Tokyo Institute of Technology, Tokyo, Japan. 3 Also at Energy Resources and Environmental Engineering Program, Egypt-Japan University of Science and Technology, New Borg El-Arab, Egypt. 4 TEL, Institute of Tibetan Plateau Research, Chinese Academy of Sciences, Beijing, China. 5 GAF, AG, Arnu, Munich, Germany. 6 Cold and Arid Regions Environmental and Engineering Research Institute, Chinese Academy of Sciences, Gansu, China. Copyright 2010 by the American Geophysical Union /10/2010WR reduce flood damage caused by heavy rainfall, a system is needed to support reservoir operators in flood management using real time observations and weather forecast data. [3] The accuracy of weather forecasting at the basin scale has improved in the last few years as a result of more reliable numerical weather prediction models. Precipitation is one of the most difficult weather variables to predict because the atmosphere is highly unstable. However, advanced techniques have enabled reasonable predictions at the regional scale [e.g., Golding, 2000; Krzysztofowicz and Collier, 2004; Honda et al., 2005]. Because precipitation makes up the main input data for hydrological models, the accuracy of a quantitative precipitation forecast (QPF) is reflected in the streamflow forecast. Recent works [Jasper et al., 2002; Collischonn et al., 2005; Saavedra et al., 2009] have shown the feasibility of forcing distributed hydrological models (DHMs) with mesoscale QPFs. [4] A flood peak can be reduced by appropriate dam operation, assuming the dam has sufficient storage capacity. The application of preestablished rule curves is limited during extreme flood events [Chang and Chang, 2001]. Optimal release systems using hydrological models to assist dam operators have been reviewed [e.g., Yeh, 1985; Labadie, 2004]. It is noted that, because of the increase in computational power and real time data availability, simulation approaches have become feasible and attractive [e.g., Wurbs, 1993]. Some of these studies focused on the optimization of operating rules for multireservoir systems taking advantage 1of13

2 Figure 1. Steps of the dam decision support system DRESS for flood control under extreme events. of real coded genetic algorithms [e.g., Oliveira and Loucks, 1997; Chen, 2003; Chang, 2008]. Works on areas affected by typhoons in Southeast Asia, such as Hoa Binh dam in Vietnam [Ngo et al., 2007], have focused on the optimal rule curves. A customized geographical information system was created to support dam release decisions in Korea [Shim et al., 2002]. Studies in Taiwan targeted real time forecasting for flood control in Taiwan [e.g., Hsu and Wei, 2007; Chang and Chang, 2009, Wei and Hsu, 2009]. However, no recent studies have spatially evaluated QPFs at previous steps in real time or included them in near future dam release decisions. [5] The present system employs a distributed biosphere hydrological model (DBiHM) with an embedded dam operation module coupled to a heuristic algorithm for supporting release instructions. This comprehensive model solves the energy budget explicitly and can simulate the initial soil moisture properly, which might be critical for flood forecasting [Wang et al., 2009a]. In normal runs, the model is forced by real time observed radar products calibrated with a high density rain gauge telemetry network. The radar calibration is based on a dynamic window, yielding a highquality product almost free of shadows at high spatial and temporal resolutions. During heavy rainfall the model is driven by an operational nonhydrostatic mesoscale QPF. The study has four distinctive features: (1) forecast error weighting using a relatively wide domain according to location, intensity, and extent; (2) generation of ensemble flood forecasts using the evaluation of previous forecast errors; (3) use of a DBiHM forced by output from a nonhydrostatic mesoscale model; and (4) use of a heuristic algorithm for multireservoir operations with a multiobjective approach. 2. System Description [6] The prototype dam release support system (DRESS) is summarized in Figure 1. The DBiHM runs in normal mode forced by observed precipitation data until a heavy precipitation event is detected. At that point, the forecast error is evaluated. Recent observed precipitation and QPFs are compared over the previous time step window in terms of location, intensity, and extent. Next, a weighting method is used to calculate the weights in each zone. These weights define the amplitude of QPF perturbation to generate ensemble precipitation members over the total lead time, as shown in Figure 2. Then the DBiHM model is run to obtain the ensemble streamflow forecast. For special dam operations, a priori dam release is calculated for each ensemble member. Once the suggested optimal dam release schedule is determined using the QPF signal, the efficiency is evaluated using the observed precipitation. 2of13

3 weighting scheme based on that concept but we express it quantitatively. The forecast error is defined as the averaged value of the high and mean intensity ratio, FE i; j ¼ 0:5 HI i;qpf þ MI i;qpf ; ð1þ HI i;obs MI i;obs Figure 2. Error evaluation window t 0 when an extreme event is detected using recent issued QPF; otherwise, the system continues running with observed precipitation Evaluation of the Forecast Error Over Previous Time Steps [7] To account for QPF errors in terms of intensity and displacement of the main pattern, Ebert and McBride [2000] proposed contiguous rain areas. However, it would also be helpful to locate these areas with respect to the targeted basin. Actually, underestimated or overestimated intensities of the forecast precipitation could occur in areas surrounding the basin. DRESS applies an evaluation module to different zones. The inputs are QPF and real time observed precipitation, as illustrated in Figure 2. The outputs are the perturbation weights to be used over the total lead time forecast to generate the ensemble precipitation members. [8] Because the forecast error bias might fluctuate considerably, the error is defined as the ratio of the forecast to the observed precipitation within the proposed zones. At the first level, the zones are defined by the drainage areas of the dams as sub basins. At the second level, the total basin area is delineated down to the control point. At lower levels, a number of circular buffers are proposed on the basis of the watershed s center of mass. The total domain is defined as a rectangle that covers the basin and buffers Weighting According to Location, Intensity, and Extent [9] Ebert and McBride [2000] proposed a contingency table for rain events to evaluate the QPF signal. They expressed the forecast intensity versus the displacement of the rain pattern qualitatively. We propose a perturbation where HI represents high intensity and MI represents mean intensity. The subscripts i and j denote the zone and the evaluated time step window, respectively; QPF indicates forecast values, and OBS indicates observations. The mean intensity ratio term accounts for the error extent because it is averaged over the zones (sub basins, basin, buffers, and total domain). Thus, the intensity and extent are considered in one term, allowing two dimensionality in further analysis against the zones. If the forecast error is close to 1, a very accurate forecast is assumed. Values higher than 1 indicate overestimation in the QPF; values between 0 and 1 indicate underestimation. [10] This study assigns weights using a lookup table. For example, a very good forecast in terms of intensity, location, and extent would receive the minimum weight for a particular zone. In that case, the QPF requires almost no perturbation, indicating low uncertainty at the previous time step. However, overestimation or underestimation is penalized with higher weights according to the total ratio and location. A good pattern match within a sub basin is rewarded with a weight close to zero. However, for a good match in wider areas, the minimum weight increases gradually Ensemble Precipitation Forecast [11] Direct compensation for forecast errors in the previous step might not always be correct, because shifts from overestimation to underestimation and vice versa are possible. Therefore, a range is applied to the current QPF according to its performance over the previous time step. The QPF s signal can then be perturbed to generate ensemble precipitation members by underestimation and overestimation. The amplitude of perturbation is defined by the weights reflecting previous QPF performance and a noise perturbation with a Gaussian or normal distribution [Goovaerts, 1997]. An ensemble precipitation forecast is generated without assimilation, which is a kind of Monte Carlo simulation [Heuvelink, 1998]. [12] We follow the approach suggested by Turner et al. [2008] to attempt an appropriate ensemble spread. Semirandom numbers are generated assuming a normal distribution with zero mean and a standard deviation of unity, N (0, 1). Precipitation is considered a semirestricted data field because the lower boundary is constrained to zero. The generated precipitation at each grid (x, y) for each lead time step k is then defined as GPðx; y Þ k ¼ maxfqpfðx; yþf1 þ A" Nð0; 1Þwi sub þ ½B" Nð0; 1 Šg; 0g; ð2þ Þwi tot where " is 0.33, meaning an exceedence probability of [Turner et al., 2008]. The terms wi sub and wi tot denote the perturbation weights directly within the dam s subbasins and the average for the entire basin, buffers, and wider domain ( total ), respectively. Different weights are proposed for each time step k within the total QPF lead time. Finally, 3of13

4 Figure 3. Potential flood volume (PFV) reduced by multireservoir operation at control point downstream. PFV is the integrated simulated stream flow Q sim that exceeds the threshold Q thre. A and B are the priority weights, which sum to unity. For example, values of 0.6 and 0.4 indicate that 60% priority is given to dam sub basins and 40% to wider zones Distributed Biosphere Hydrological Model [13] DHMs can assimilate rainfall patterns and estimate fluxes such as a streamflow at selected points. Moreover, they provide the spatial distributions of soil moisture content, evaporation, and transpiration. However, detailed energy exchange between the land surface and lower atmosphere is neglected in most DHMs [e.g., Yang et al., 2001]. Thus, combining hydrological and land surface models (LSMs) improves the representation of the water and energy cycles within a basin. The simple biosphere model (SiB2) [Sellers et al., 1996], a well known LSM, has been embedded into a geomorphology based hydrological model (GBHM) [Yang et al., 2002] to produce a water and energy budget based DHM (WEB DHM) [Wang et al., 2009a]. [14] DRESS employs WEB DHM validated in humid areas [Wang et al., 2009b], which can predict the evolution of land surface temperature and soil moisture. The SiB2 submodel calculates turbulent fluxes between the lower atmosphere and land surface independently. It then updates the surface runoff and subsurface flow, which become the lateral inflow to the main river channel. Next, the GBHM submodel simulates water flow within the river network using one dimensional kinematic wave equations solved by finite difference approximation. This simulation sequence is performed for each sub basin using the Pfafstetter scheme [Verdin and Verdin, 1999]. Moreover, every sub basin is divided into a number of flow intervals based on the flow distance to the outlet. The flow intervals are linked by the river network according to the geomorphology. The flow downstream might be affected by the operation of upstream dams. A dam function based on the mass balance approach was used to express the change in volume. Inflows to the dams are simulated by the GBHM submodel A Priori Dam Release for Selected Members [15] The ensemble streamflow is calculated using WEB DHM at a selected control point downstream. The system is activated only when the simulated streamflow with unregulated releases surpasses the threshold flow. This flow is defined as the average of the observed values exceeding the mean annual discharge at the control point [Saavedra et al., 2010]. The simulated streamflow with closed gates is then calculated to check the effect downstream. The integrated value between the two simulations over time is the potential flood volume (PFV) for a specific dam, as shown in Figure 3. This procedure shows the effect of inflow at the control point according to dam capacity. The system defines a priori release Q from n dams based on the PFV at the control point over the present time step. Thus, the release implicitly becomes a function of inflow. [16] For the rest of the total lead time forecast, the gates are assumed to be totally closed to replenish the reservoirs to levels comparable to initial levels [Saavedra et al., 2010]. Essentially, the system releases only a water volume that could be re stored after the event according to the QPF. This preventive measure reduces flood volumes directly and reduces flood peaks as a consequence. The mean and standard deviation of dam release for the participating members are then calculated to express the uncertainty. They are used as the initial guess and boundary condition for the decision variables in the optimization, respectively Optimization of Dam Release [17] Defining a proper objective function is crucial to optimal dam operation. We propose to extend the application of shuffled complex evolution (SCE) [Duan et al., 1992] in multireservoir operation using the QPF during heavy rainfall Objective [18] The flood peak should be minimized at control points downstream and the available free volume at dams should be reduced (i.e., water storage should be encouraged) Strategy [19] A trade off is made between water allocation at the control point, which might result in flood damage, and stored water at the reservoirs for potential water use Objective Function [20] Minimize Z, which is defined as Z ¼ Xn i¼1 PFV i þ Xn i¼1 RFV i ; where PFV i and RFV i are the potential flood and reservoir free volumes for n reservoirs, represented on the left and right, respectively, in Figure 4. PFV i was calculated as shown in equation (3); RFV i can be estimated as X n i¼1 RFV i ¼ Xn i¼1 ð3þ ðmxv i SV i Þ; ð4þ where MXV i is the maximum storage and SV i is the simulated volume Optimization [21] The decision variables are the dam releases. Their initial values are the mean a priori dam release obtained previously. Their upper and lower boundaries are the mean plus and minus one standard deviation, respectively. The heuristic algorithm SCE drives WEB DHM, seeding different combinations of decision variables until the objective function is fulfilled. 3. Application of the System to the Upper Tone River [22] To validate DRESS, it was applied to the upper Tone reservoir system in Japan. The northern headwaters of the 4of13

5 Figure 4. Objective function to reduce PFV at (left) the control point while attempting to reduce the (right) reservoir free volumes (RFVs) at the end the time step. Tone River basin are about 130 km northwest of Tokyo. The river supplies water for irrigation and power generation to the Tokyo metropolitan area and surroundings. Therefore, its management is crucial for the region. The watershed and drainage sub basins are shown on the right in Figure 5. The elevation varies from 100 m to about 2500 m with a mean of about 1020 m. The long term average precipitation is about 1500 mm per year. Heavy rainfall, commonly associated with typhoons and seasonal front activities, occurs from July to October. Typhoons bring the heaviest rainfall. After a catastrophic flood in 1947 [Takahasi and Okuma, 1979], development of the upper reservoir system, as well as embankment improvements in the middle and lower reaches, were emphasized Spatial Data Preparation [23] The geomorphology was obtained using a digital elevation model with 50 m resolution and aggregated into 500 m grids. Land use, soil type, and geological maps were also prepared for the study area. Land use data were reclassified according to the predefined SiB2 categories, and broadleaf deciduous trees were found to be dominant in the basin. Thematic maps, such as those for surface terrain slope, topsoil depth, and hillslope length, were derived from the basic data described above using a geographic information system. Dynamic vegetation parameters, such as the leaf area index and fraction of photosynthetically active radiation, were obtained from satellite data. The air temperature, wind speed, and sunshine duration were obtained from available meteorological sites managed by the automated meteorological data acquisition system (AMeDAS). Relative humidity and air pressure were obtained from three radiation stations maintained by the Japan Meteorological Agency (JMA). Downward solar radiation was estimated from sunshine duration, temperature, and humidity as measured by Yang et al. [2006]. Longwave radiation was then estimated from the temperature, relative humidity, pressure, and solar radiation as measured by Crawford and Duchon [1999]. High quality observed radar precipitation, calibrated using a dense rain gauge network based on a dynamic window, was also obtained from AMeDAS. The temporal resolution was 1 h at 2.5 km until 2002, 30 min at 1 km from 2003 to 2004, and 10 min at 1 km later. The AMeDAS data are transmitted through advanced telemetry almost in real time. Figure 5. Location of upper Tone River system, northwest of Tokyo, which encompasses 3304 km 2. The map on the left shows the proposed zones around the basin to evaluate the forecast error. The drawing on the right depicts the dams arrangement in squares and the streamflow gauging stations in triangles. 5of13

6 Table 1. Characteristics of Upper Tone Reservoir System Name of Dam Complete (year) Drainage (km 2 ) Capacity (Mm 3 ) Fujiwara Aimata Sonohara Yamba Not yet Mesoscale QPF Data [24] QPF was obtained from the operational nonhydrostatic JMA model [Saito et al., 2007], which has a 10 km horizontal resolution and was issued four times a day until February 2006; the data are available at the University of Tokyo s website ( tokyo.ac.jp/gpv/). The mesoscale model grid point values (MSM GPV) data set was downloaded and clipped to the study area, which covers all of Japan (2900 km 3600 km); the most relevant heavy rainfall from 2002 to 2004 was targeted. The data set has a total lead time of up to 18 h updated every 6 h. Thus, for each hour of observed precipitation, three different precipitation forecast values are available: those issued 1 6, 7 12, and h previously, as illustrated in Figure Reservoir Characteristics [25] Three dams, labeled 1, 2, and 3 in Figure 5, were selected for flood control to protect downstream plains. Their main characteristics and capacities are summarized in Table 1. The location of the embankments was placed within the river network in the grid based model. The reservoirs are normally operated using mutually independent rule curves. A release is decided according to the actual water level of the reservoir, inflow to the reservoir, and water demand [Ministry of Construction, 1995]. However, during extreme events, reservoir release is determined according to the judgment of the director of the reservoir operation office, according to the reservoir s status and downstream flood conditions. Normal operation for each reservoir was simulated by Yang et al. [2004], and optimal operation was simulated using radar data by Saavedra et al. [2010]. 4. Results [26] To validate the efficiency of DRESS, the highest yearly flood peaks between 2002 and 2004 were selected. WEB DHM s soil and vegetation associated parameters were calibrated and validated by Wang et al. [2009b] using the data observed between 2001 and The simulated river discharge and land surface temperature were compared with values recorded in situ and by satellite, respectively. The WEB DHM model had a 1 h time step, whereas the optimization horizon had an 18 h time step according to the QPF lead time. The initial conditions were set to those several hours before the event using observed data. Three multipurpose dams arranged in parallel, Fujiwara, Aimata, and Sonohara, were selected to examine flood reduction at the Maebashi gauging station (see Figure 5). Once an extreme event was identified, the system verified the situation downstream. The spatial distribution of the 7 12 h MSM GPV series was compared to the AMeDAS radar observations, as shown in Figures 6a and 6b (2002 event), Figures 6c and 6d (2003 event), and Figures 6e and 6f (2004 event). In these particular examples, a close match between the 2002 and 2003 events is evident. However, the results for 2004 indicate that this might not always be the case. The spatial distribution of precipitation shows a high resolution and quality of AMeDAS and MSM GPV data Forecast Error Evaluation [27] After the forecast error was evaluated within different zones, the perturbation weights for the three MSM GPV series were estimated using Table 2. The six proposed zones are displayed in Figure 6. The weights oscillate from 0 to 5. The approximate forecasts were defined for ratios between 0.9 and 1.1. The intensity and extent are included in the magnitude of the forecast error (columns). The location is expressed by the evaluated zones (rows). [28] Three cases must be carefully considered when using ratios to express the forecast error: [29] 1. If only the observed rainfall tends to zero, the ratio tends to infinity, giving an abnormal overestimation. In this case, an eligible weight from a larger zone should be assigned until the ratio no longer tends to infinity. [30] 2. If only the forecast rainfall tends to zero, the ratio tends to zero, giving an abnormal underestimation. This should be distinguished from cases where the weight is zero (indicating a very good forecast); thus, it is penalized by being assigned the highest weight in the evaluation zone. [31] 3. If the observed and forecast precipitations are both zero, which is a good match, the ratio is mathematically undetermined, and the minimum weight should be assigned. [32] The averaged ratios of the MSM GPV data to the radar AMeDAS observations over the sub basins are presented in Figure 7, which shows the temporal and spatial variability of the forecast error. The ratio can shift from overestimation to underestimation and vice versa. This confirms the need to use the ensemble approach. In general, the MSM GPV signal tended to increase as the climax of a heavy rainfall event approached, as seen in the last time steps of the graphs in Figure Ensemble Streamflow Forecast [33] By applying the perturbation weights from Table 2, an ensemble of 50 MSM GPV members (i.e., recently issued QPFs) were generated using equation (2). They were given 60% priority in the contributing areas (i.e., sub basins) and 40% in wider zones. The ensemble precipitation forced the WEB DHM model, generating an ensemble streamflow. Then, the system verified whether the members exceeded the threshold flow. An ensemble size of 50 members was found suitable for the three events considered here and might also be appropriate in similar studies. The threshold flow that triggers special operation was calculated as 500 m 3 s 1 for the Maebashi gauging station using historical data following the approach of Saavedra et al. [2010] A Priori Release [34] Most ensemble members during the three events suggested activating the system for a priori release, except during the first step of the October 2004 event. At this step, 96% (48 of the 50 members) suggested that the system be activated. This indicates that the ensemble amplitude using the perturbation weights was efficient for those events. The a priori release was obtained using the PFVs at the Maebashi gauging station for each member for the three dams. 6of13

7 Figure 6. Spatial distribution of the accumulated precipitation over 6 h using forecast QPF and observed radar: July 2002 event issued at 10.09z (local time) using (a) MSM GPV and (b) AMeDAS radar; August 2003 event issued at 09.09z (local time) using (c) MSM GPV and (d) AMeDAS radar; and October 2004 event issued at 20.09z (local time) using (e) MSM GPV and (f) AMeDAS radar. 7of13

8 Table 2. Perturbation Weights According to the Forecast Error Magnitude and Zone a Approx. Underestimation Correct Overestimation >10 Total ratio Dam drainage Basin First Buffer Second Buffer Third Buffer Total Domain a From left to right, column values indicate underestimation through overestimation of the forecast error magnitude; rows indicate zone. [35] The mean and standard deviation of the dam release were then obtained, as shown in Figure 8. The lengths of the bars denote the amplitudes of the perturbation (i.e., uncertainty) according to the QPF at the previous step. Comparing Figures 7 and 8 for the same steps, we find that the performance Figure 7. Area precipitation ratio MSM GPV over AMeDAS at local time: (a) 9 10 July 2002, (b) 8 9 August 2003, and (c) 20 October of13

9 of the MSM GPV data is consistent with the a priori release. According to Figure 8, for those steps in which the MSM GPV was overestimated, the mean release tended to be higher. For example, consider the partial overestimation for the Aimata sub basin in steps 09.03z (2003) and 20.15z (2004). These overestimations might lead to water loss, as seen in Table 3, if we evaluate the system s performance only for this sub basin. However, decision makers could consider the uncertainty of the dam release using the information in Figure Optimization of Multireservoir Operation [36] The upper and lower boundaries of the bars in Figure 8 were used to optimize dam network operation. The initial guess was the mean dam release (red squares in Figure 8). After the iteration of the SCE algorithm was completed, the suggested optimized variables were obtained. The dam release tended to be closer to the lower boundary in most cases because water storage was preferred. However, this would not be the case if the initial water levels before the event were higher (i.e., more volume were available). Then, higher dam release magnitudes would be expected. Figure 8. Dam release uncertainties for the events (a) 9 10 July 2002, (b) 8 9 August 2003, and (c) 20 October Real Time Operation [37] The process described above was applied every 6 h with the issue of new MSM GPV data. Thus, the simulation used only the most recent available data as would be done in a real time basis. Integrating the suggested dam release over the evaluated steps yielded a release schedule per event. Then the WEB DHM model was run in evaluation mode using the observed AMeDAS precipitation. The flood reduction at the Maebashi gauging station using the suggested release schedule for the three dams is shown in Figure 9 (2002 event), Figure 10 (2003 event), and Figure 11 (2004 event). The efficiencies in flood peak reduction were 18%, 28%, and 7%, respectively, for the three events. Early release from the three dams is evident for the 2002 and 2004 events. During the 2003 event, the total suggested release was lower than the observed release (except for the Aimata dam, due to MSM GPV overestimation for this sub basin, as discussed previously). Underestimation in the precipitation forecast was found to increase water storage but to decrease flood control efficiency. Such cases should be examined carefully when attempting to reduce flood damage. The total gain and loss of water at each reservoir for each event is summarized in Table 3. Overall, the results indicate gains in water storage, demonstrating the system s efficiency. 5. Discussion [38] The effects of overestimation or underestimation in QPFs were crucial to the system s responses, leading to water loss or less effective flood reduction, respectively. As QPF Table 3. Evaluation of the DRESS System s Performance in Tone River, Japan Event Flood Peak Reduction at Control Point (%) Gain or Loss at Fujiwara Reservoir (MCM) Gain or Loss at Aimata Reservoir (MCM) Gain or Loss at Sonohara Reservoir (MCM) Total Gain or Loss, Reservoirs (MCM) of13

10 Figure 9. (a) Flood reduction at Maebashi gauging station using suggested release from (b) Fujiwara, (c) Aimata, and (d) Sonohara dams for July 2002, 10.04z 11.24z. 10 of 13

11 Figure 10. (a) Flood reduction at Maebashi gauging station using suggested release from (b) Fujiwara, (c) Aimata, and (d) Sonohara dams for August 2003, 09.04z 10.24z. 11 of 13

12 Figure 11. (a) Flood reduction at Maebashi gauging station using suggested release from (b) Fujiwara, (c) Aimata, and (d) Sonohara dams for October 2004, 20.10z 21.24z. accuracy increases, dam release operations using decision support systems such as DRESS are expected to become more efficient and reliable. [39] In real life operation, DRESS should be able to suggest water release before the next QPF is issued. The total processing time should be less than 6 h, which is the QPF update interval employed. The present application took around 17 h using a 2 GHz personal computer; the run time can be reduced by as much as 10 times by using a parallel computing system. The time could also be reduced as QPFs become more accurate, because the perturbation range would narrow, decreasing the ensemble size. [40] Several modifications are possible for future applications. For example, the objective function can be adjusted depending on local conditions. A flushing procedure for flood control is needed in humid mountains and highly vegetated basins. In drier basins, the objective function can give priority to maintaining a higher water level in reservoirs as long as possible. Also, only QPFs were used as real time forcing data; however, other atmospheric model outputs, such as air humidity and wind profiles, could be included. Since a DBiHM is used, updates of soil moisture conditions could eventually interact with land data assimilation systems [e.g., Boussetta et al., 2008; Yang et al., 2009]. [41] QPFs in Japan are reasonably useful. However, this might not be true in countries with limited infrastructure, technology, and resources. In such cases, satellite derived precipitation could be used in near real time depending on the time period [e.g., Saavedra et al., 2009]. Priority could be given to recent satellite based measurements (i.e., those calibrated with available rain gauges) to improve system reliability. 6. Conclusion [42] A prototype DRESS was developed for flood control during heavy rainfall. It employs a DBiHM forced by an operating mesoscale QPF in Japan, the MSM GPV. To reflect recent forecast errors in the near future dam release schedule, an ensemble forecast approach and optimization algorithm are employed. An objective function was formulated to reduce flood effects downstream and increase the total volume in the reservoirs. [43] DRESS was validated in the upper Tone River reservoir network using the worst heavy rainfall in 2002, 2003, and The efficiency of the system was demonstrated for all events, with some differences depending on the accuracy of the MSM GPV. The objective function successfully minimized the flood volume at a control point and the free volume of the reservoirs, taking into account different release scenarios. The results indicate that DRESS is feasible for reallife dam operation. [44] Acknowledgments. We would like to thank the River Section Bureau of the Ministry of Land Infrastructure, Transportation, and Tourism of Japan for providing valuable data from operating dams. This study was supported by the Science and Technology National Project Japan: Data Integration and Analysis System. Additionally, we would like to thank the three anonymous reviewers who provided constructive suggestions for the final version. 12 of 13

13 References Boussetta, S., T. Koike, K. Yang, T. Graf, and M. Pathmathevan (2008), Development of a coupled land atmosphere satellite data assimilation system for improved local atmospheric simulations, Remote Sens. Environ., 112, Chang, L. C. (2008), Guiding rational reservoir flood operation using penalty type genetic algorithm, J. Hydrol., 354, Chang, L. C., and F. J. Chang (2001), Intelligent control for modeling of real time reservoir operation, Hydrol. Processes, 15, Chang, L. C., and F. J. Chang (2009), Multi objective evolutionary algorithm for operating parallel reservoir system, J. Hydrol., 377, Chen, L. (2003), Real coded genetic algorithm optimization of long term reservoir operation, J. Am. Water Resour. Assoc., 39(5), Collischonn, W., R. Hass, I. Andreolli, and C. E. Morelli Tucci (2005), Forecasting River Uruguay flow using rainfall forecasts from a regional weather prediction model, J. Hydrol., 305, Crawford, T. M., and C. E. Duchon (1999), An improved parameterization for estimating effective atmospheric emissivity for use in calculating daytime downwelling long wave radiation, J. Appl. Meteorol., 38, Duan, Q., S. Sorooshian, and V. K. Gupta (1992), Effective and efficient global optimization for conceptual rainfall runoff models, Water Resour. Res., 28(4), Ebert, E. E., and L. L. McBride (2000), Verification of precipitation in weather systems: Determination of systematic errors, J. Hydrol., 239, Golding, B. W. (2000), Quantitative precipitation forecasting in the UK, J. Hydrol., 239, Goovaerts, P. (1997), Geostatistics for Natural Resources Evaluation, 483 pp., Oxford Univ. Press, New York. Heuvelink, G. B. M. (1998), Error Propagation in Environmental Modelling with GIS, 127 pp., Taylor & Francis, London. Honda, Y., M. Nishijima, K. Koizumi, Y. Ohta, K. Tamiya, T. Kawabata, and T. Tsuyuki (2005), A pre operational variational data assimilation system for a non hydrostatic model at the Japan Meteorological Agency: Formulation and preliminary results, Q. J. R. Meteorol. Soc., 131, Hsu, N. S., and C. C. Wei (2007), A multipurpose reservoir real time operation model for flood control during typhoon invasion, J. Hydrol., 336, Huntington, T. G. (2006), Evidence for intensification of the global water cycle: Review and synthesis, J. Hydrol., 319, Intergovernmental Panel on Climate Change (IPCC) (2007), Climate Change 2007: The Physical Science Basis, 996 pp., Cambridge Univ. Press, Cambridge, U. K. Jasper, K., J. Gurtz, and H. Lang (2002), Advanced flood forecasting in Alpine watersheds by coupling meteorological observations and forecasts with a distributed hydrological model, J. Hydrol., 267, Kojiri, T., T. Hamaguchi, and M. Ode (2008), Assessment of global warming impacts on water resources and ecology of a river basin in Japan, J. Hydro Environ. Res., 1, Krzysztofowicz, R., and C. G. Collier (Eds.) (2004), Quantitative precipitation forecasting II, J. Hydrol., 288(1 2). Kyselý, J., and R. Beranová (2009), Climate change effects on extreme precipitation in central Europe: Uncertainties of scenarios based on regional climate models, Theor. Appl. Climatol., 95, Labadie, J. W. (2004), Optimal operation of multireservoir systems stateof the art review, J. Water Resour. Plann. Manage., 130(2), Ministry of Construction (1995), Reservoir operation manuals for the upper Tone River basin (in Japanese), Reservoir Manage. Off. of the Upper Tone River Basin, Kanto Bur., Minist. of Construction, Japan. Ngo, L. L., H. Madsen, D. Rosbjerg (2007), Simulation and optimisation modelling approach for operation of the Hoa Binh reservoir, Vietnam, J. Hydrol., 336, Oliveira, R., and D. Loucks (1997), Operating rules for multireservoir systems, Water Resour. Res., 33(4), Saavedra, O., T. Koike, K. Yang, D. Yang, C. T. Nyunt, D. V. Khanh, and L.C. Chau (2009), Flood simulation using different sources of rainfall in the Huong River, Vietnam, Hydrol. Sci. J., 54(5), Saavedra, O., T. Koike, K. Yang, and D. Yang (2010), Optimal dam operation during flood season using a distributed hydrological model and a heuristic algorithm, J. Hydrol. Eng., 15(7), , doi: / (ASCE)HE Saito, K., J. I. Ishida, K. Aranami, T. Hara, T. Segawa, M. Narita, and Y. Honda (2007), Nonhydrostatic atmospheric models and operational development at JMA, J. Meteorol. Soc. Jpn., 85B, Sellers, P. J., D. A. Randall, G. J. Collatz, J. A. Berry, C. B. Field, D. A. Dazlich, C. Zhang, G. D. Collelo, and L. Bounoua (1996), A revised land surface parameterization (SiB2) for atmospheric GCMs, Part I: Model formulation, J. Clim., 9, Shim, K. C., D. G. Fontane, and J. W. Labadie (2002), Spatial decision support system for integrated river basin flood control, J. Water Resour. Plann. Manage., 128(3), Takahasi, Y., and T. Okuma (1979), Historical review of floods and their damage in the Tone River basin, J. Nat. Disaster Sci., 1, Turner, M. R. J., J. P. Walker, and P. R. Oke (2008), Ensemble member generation for sequential data assimilation, Remote Sens. Environ., 112, Verdin, K. L., and J. P. Verdin (1999), A topological system for delineation and codification of the Earth s river basins, J. Hydrol., 218, Wang, L., T. Koike, K. Yang, T. K. Jackson, R. Bindlish, and D. Yang (2009a), Development of a distributed biosphere hydrological model and its evaluation with the Southern Great Plains Experiments (SGP97 and SGP99), J. Geophys. Res., 114, D08107, doi: / 2008JD Wang, L., T. Koike, K. Yang, and P. J. F. Yeh (2009b), Assessment of a distributed biosphere hydrological model against streamflow and MODIS land surface temperature in the upper Tone River Basin, J. Hydrol., 377, Wei, C. C., and N. S. Hsu (2009), Optimal tree based release rules for real time flood control operations on a multipurpose multireservoir system, J. Hydrol., 365(3 4), Wurbs, R. A. (1993), Reservoir system simulation and optimization models, J. Water Res. Plann. Manage., 119(4), Yang, D., S. Herath, T. Oki, and K. Musiake (2001), Application of a distributed hydrological model in the Asian monsoon tropic region with a perspective of coupling with atmospheric models, J. Meteorol. Soc. Jpn., 79(1B), Yang, D., S. Herath, and K. Musiake (2002), Hillslope based hydrological model using catchment area and width functions, Hydrol. Sci. J., 47, Yang, D., T. Koike, and H. Tanizawa (2004), Application of a distributed hydrological model and weather radar observations for flood management, Hydrol. Processes, 18, Yang, K., T. Koike, and B. Ye (2006) Improving estimation of hourly, daily, and monthly downward shortwave radiation by importing global data sets, Agric. For. Meteorol., 137, Yang, K., T. Koike, I. Kaihotsu, and Q. Jun (2009), Validation of a dualpass microwave land data assimilation system for estimating surface soil moisture in semiarid regions, J. Hydrometeorol., 10(3), Yeh, W. W. G. (1985), Reservoir management and operation models: A state of the art review, Water Resour. Res., 21(12), T. Graf, GAF, AG, Arnulfstr. 197, D Munich, Germany. X. Han and X. Li, Cold and Arid Regions Environmental and Engineering Research Institute, Chinese Academy of Sciences, Lanzhou, Gansu , China. T. Koike and L. Wang, Department of Civil Engineering, University of Tokyo, Bunkyo ku, Tokyo , Japan. O. C. Saavedra Valeriano, Department of Civil Engineering, Tokyo Institute of Technology, Meguro ku, Tokyo , Japan. (saavedra. o.aa@m.titech.ac.jp) K. Yang, TEL, Institute of Tibetan Plateau Research, Chinese Academy of Sciences, Beijing , China. 13 of 13