A FRAMEWORK TO ASSESS EFFECTIVENESS AND RISKS OF INTEGRATED RESERVOIR OPERATION FOR FLOOD MANAGEMENT CONSIDERING ENSEMBLE HYDROLOGICAL PREDICTION XVI World Water Congress, IWRA Cancun, Mexico, June 2nd, 2017 Daisuke NOHARA1, Hiroki SAITO2 and Tomoharu HORI1 1Water Resources Research Center, Disaster Prevention Research Institute, Kyoto University, Japan 2Graduate School of Engineering, Kyoto University, Japan Thank you for visiting us. I am glad to exchange perspectives related to reservoir management practices in Taiwan and Japan. I would like to talk about …

INTRODUCTION Real-time reservoir operation Playing a significant role for effective water resources management including flood management by controlling fluctuation in river water flow More integrated reservoir operation considering the latest information on hydrological conditions in the target river basin is needed for more effective water resources management Japan Sea Reservoir operation plays a significant role for effective water resources management including flood management by controlling fluctuation in river water flow.

Prediction of flood occurrence INTRODUCTION Preliminary release operation for flood management Keeping water level as high as possible, at the same time as safely decreasing water level in advance of flood events so as to secure storage capacity for flood control (enables integrated management) Prediction of flood occurrence Storing flood water As one of integrated reservoir operations, preliminary release operation is introduced in multi-purpose reservoirs in Japan. The preliminary release operation is an operation method for flood management, in which water stored in the reservoir is released just before the flood water arrives to the reservoir to secure empty storage for flood control. This operation is expected to bring out enhanced capability of a multi-purpose reservoir for integrated water resources management because the reservoir can keep its water level as high as possible for water utilization in no flood situation at the same time as safely decreasing water level in advance of flood events for flood management. As it is needed to consider future hydrological condition to release water in advance of flood, consideration of hydrological prediction is important. Normal period Before a flood event After the flood event Consideration of hydrological prediction is important.

INTRODUCTION Effect of prediction uncertainty Handling uncertainty contained in the predictions has been issues Preliminary release Insufficient water recovery Inflow less than predicted Overestimated prediction Japan Sea Before a flood event After the flood event Risk of overflow No preliminary release Inflow more than predicted However, when we consider some hydrological prediction, we also need to be careful of its uncertainty at the same time. For example, let us think when we conduct the preliminary release operation considering some imperfect inflow prediction. There will not be any problem in conducting preliminary release operation if the prediction is well enough. But if we conduct preliminary release based on an overestimated prediction, storage water would not be recovered after the flood event as we have less amount of inflow to store than predicted. On the other hand, if we conduct preliminary release based on an underestimated prediction, we would not be able to complete necessary preliminary release of water before a flood, and may increase a risk of overflow as we have too much water actually to be stored in the empty volume secured by the incomplete preliminary release. So, considering uncertainty contained in the prediction has been issues. Normal period Underestimated prediction Before a flood event After the flood event

INTRODUCTION Operational ensemble hydrological forecasts Various operational hydrological forecasts are provided by the authorities, including ensemble hydrological predictions (EHPs) EHP consists of multiple predicted sequences with different initial conditions to consider the effect of growth of initial errors in time horizon Including information on uncertainty, which can be estimated by considering spread of the predictions at each predicted time cross section Considered to bring more robust decision making in the reservoir operation by taking them into account Japan Sea As a method to mitigate the effect of uncertainty contained in the predictions, various operational ensemble hydrological forecasts have been provided by the authorities. consists of multiple sequence of predictions with different initial conditions to consider the effect of growth of errors in time horizon Ensemble hydrological predictions include information on uncertainty, which can be estimated by considering spread of the predictions at each predicted time cross section. And it is considered to bring more robust decision making in the reservoir operation. Ensemble forecast by JMA ( 850hPa temperature in 7 days MA, JMA)

INTRODUCTION Analysis on effects of prediction uncertainty is important for designing an effective way to conduct prior release operation of reservoirs Stochastic approach is needed to thoroughly investigate effects of prediction uncertainty on prior release operation However, amount of actual prediction data provided under flood conditions is not much enough to carry on statistical analysis Simulated generation of prediction with various degrees of uncertainty is considered to be effective for such an analysis

OBJECTIVE Analysis on effects of real-time ensemble hydrological prediction and its uncertainty on preliminary release operation of a multi-purpose reservoir For hourly operation of a single multi-purpose reservoir in Japan before and under flood conditions Coupled with a simulated generation process of ensemble inflow prediction (EIP) with various degrees of errors Monte Carlo simulation (MCS) of preliminary release operation with generated EIPs to analyze effects of prediction uncertainty on preliminary release operation in terms of flood control and water recovery Considering these circumstances, this study aims at developing a method to analyze the effects of real-time hydrological prediction and its uncertainty on prior release operation of a multi-purpose reservoir. We considered hourly operation of a single multi-purpose reservoir in Japan under flood condition, coupling with a simulated generation process of inflow prediction with various degrees of errors. Monte Carlo simulation of prior release operation with the generated inflow predictions was conducted to analyze effects of prediction uncertainty on both the flood control during a flood and water utilization after the flood event.

NAGAYASUGUCHI RESERVOIR A multi-purpose reservoir for flood control, water supply and power generation in the Naka River Basin, Japan Non flood situation Flood situation Flood control capacity 11.0 MCM Water use capacity Secured by the preliminary release operation 43.5 MCM The Naka River Basin (874 km2) Nagayasuguchi Reservoir This is a summary of the target reservoir and the river basin. The Nagayasuguchi Reservoir in the Naka River Basin, located in the south west part of Japan, is considered as a target reservoir. The Nagayasuguchi Reservoir is a multi-purpose reservoir operated for flood control, water supply and power generation. These figures show the allocation of storage volumes for the purposes. The reservoir has no empty volume for flood control in non-flood situation, and all the storage is allocated for water use purposes. The reservoir is designed to prepare empty volume for flood control (called as flood control capacity) by preliminary water release in advance of the arrival of flood waters if flood situation is predicted. The Naka River basin area: 874 km2 Nagasuguchi Reservoir catchment area: 539km2 Catchment area of the dam: 539 km2

FRAMEWORK OF ANALYSIS … … True hydrograph of a flood event Prediction 1 Simulation 1 Prediction 2 Simulation 2 … … True hydrograph of a flood event Simulated generation of 1000 predictions with errors MCS of preliminary release & flood control operations with each prediction Assessment of effectiveness / impacts on flood control Simulation 1 This shows the schematic flow diagram of the proposed Monte Carlo simulation of reservoir’s prior release operation for a flood event. At first, we derive the true hydrograph of a historical flood event, then generate 1000 inflow predictions based on the true hydrograph and artificially generated prediction errors, and conduct simulation of prior release and flood control operations with each generated prediction as a Monte Carlo simulation. We can assess the effectiveness and impacts of prior release operation with imperfect prediction on flood control in an aggregated manner during this process. As a result of flood control simulation, now we have water level after the flood control operation, which is also initial storage volumes for water utilization operation after the flood event. We conduct simulation on long-term reservoir operation for water supply and power generation with these initial storages as a Monte Carlo simulation to analyze long-term impact on these purposes. Simulation 2 … Analysis of long-term impact on water supply & power generation Water storage after the flood event

GENERATION OF ENSEMBLE INFLOW PREDICTION Synthetic generation of ensemble inflow predictions for the coming 8 days Generation of predictions so that they have a certain degree of uncertainty by changing values for the following indices: Ensemble mean error: Averaged error (or goodness) of predictions Spread: Degree of uncertainty contained in an ensemble prediction M: Number of ensemble members x*m: Prediction by member m xo: True value of inflow ：Average of predictions (ensemble mean) True Value Ensemble Mean Ensemble Mean Error eM Spread sD

GENERATION OF ENSEMBLE INFLOW PREDICTION Generation of prediction errors x*[prediction] = xo[true value] + e[error] Assuming that prediction errors follows a normal distribution N(μe, σe2) Values of prediction therefore follows the normal distribution N(μe+xo, σe2), where μe and σe2 respectively correspond to the ensemble mean error and spread. If one considers σe2 = Ce(l) σo2 and assume an error growth function like Ce(l)= αl+β, it can be modelled that error variance becomes greater as lead time of prediction becomes long (where σo2 is the variance of true values of inflow). Normal distribution that predicted values follow Inflow N(μe+io, σe2) True value … True value xo Ensemble mean error μe σe2 Spread (variation) of prediction l=1 l=2 l=3 l=4 Lead time (l)

GENERATION OF A SERIES OF ENSEMBLE INFLOW PREDICTION Randomly sample m values from a normal distribution which prediction errors follow and respectively add them to a true value to get an ensemble prediction with m members for each time step Generation of a series of prediction errors for each ensemble member using the following AR(1) model ：Standardized prediction errors ：Serial correlation ： Inflow A random value following True value Prediction (Member 1) Prediction (Member 2) Prediction (Member 3) … l=1 l=2 l=3 l=4 Lead time (l)

ERROR PARAMETERS ESTIMATED FROM OPERATIONAL ENSEMBLE PREDICTION Ensemble inflow prediction estimated from JMA’s One-week Ensemble Forecast (JMA-EPSW) of precipitation with temporal range of 192 hours (8 days) Averaged Ensemble Mean Errors of Ensemble Inflow Predictions Regression: μe = -0.1343l-136.44 Japan Meteorological Agency JMA-EPSW This is the specs of JMA’s One-week Ensemble Forecast of precipitation. It has leadtime of 192 hours, that is 8 days, the forecast is provided to Japan Area, as a form of grid point value (GPV) with spatial resolution of 1.25 degrees grid. Temporal resolution is 6 hours, and updated every day, number of ensemble members is 51. Hydro-BEAM (Rainfall-runoff Model)

PRELIMINARY RELEASE OPERATION Preliminary release operation rules of the Nagayasuguchi Reservoir 1st stage: Water storage is deceased to 38.1 MCM when inflow exceeds 70 m3/s. 2nd stage: Water storage is decreased to 32.5 MCM when inflow is expected to exceed 500 m3/s. Keep high water level for water use (Policy 1) Preliminary release of 1st stage (Policy 2) Preliminary release of 2nd stage (Policy 3) Qin > 500 m3/s Qin > 70 m3/s Considering the ensemble inflow prediction, reservoir states are estimated in a form of ensemble prediction. Here I am showing the actual operation rule for the preliminary release by the Nagayasuguchi Reservoir. They have two stages, water storage is decreased to 38.1 MCM as the first step of the preliminary release operation when inflow exceeds 70 m3/s, and to be decreased to 32.5 MCM as the second step when inflow is expected to exceed 500 m3/s . Considering the actual operation rules, operation policies for preliminary release of the Nagayasuguchi Reservoir can be summarized into three policies: keeping water level for water use operation without conducting preliminary release operation as a Policy 1, conducting only 1st stage of preliminary release operation for Policy 2, and conducting preliminary release operation up to 2nd stage as Policy 3. 43.5 MCM 38.1 MCM 32.5 MCM

ASSESSMENT ON PRELIMINARY RELEASE OPERATION CONSIDERING ENSEMBLE INFLOW PREDICTION Results of Monte Carlo simulation of reservoir operation with 1000 simulations to analyze impacts and risks of preliminary release operation considering ensemble inflow prediction Results of simulations on preliminary release operation when ensemble mean predictions were considered Events Rate of simulations where preliminary release was conducted (%) Rate of simulations where water storage recovered (%) Storage rate after flood averaged over the simulations (%) Flood 1 0.0 100.0 Flood 2 Results of simulations on preliminary release operation when the ensemble member with the maximum value of prediction was considered Events Rate of simulations where preliminary release was conducted (%) Rate of simulations where water storage recovered (%) Storage rate after flood averaged over the simulations (%) Flood 1 99.7 96.7 99.9 Flood 2 100.0 On the other hand, these figures show the examples for estimated ensemble reservoir states when we take a specific release policy with each ensemble member of inflow prediction. The situations if water release was conducted according to Policy 2 and Policy 3 can be compared in these figures. For example, we can find risks to water use brought down by the preliminary release operation by seeing the number of ensemble members with which storage was not recovered to the full capacity, or the averaged storage after the flood event. At the same time, we can also estimate risks to flood management by seeing the number of ensemble members with which empty storage volume secured by the preliminary release operation was not enough. These kinds of information can be considered useful for reservoir managers to make a decision for preliminary release considering potential effects of each release policy on both the flood management and drought management in an integrated manner.

CONCLUDING REMARKS A method to quantitatively analyze effects and risks of preliminary release operation of a multi-purpose reservoir considering ensemble inflow prediction was developed. A method of synthetic generation of ensemble hydrological predictions with a certain error characteristics was applied. Potential effects of preliminary release operation considering ensemble hydrological predictions on flood management or risk on water utilization can be quantified by using the proposed method employing error parameters derived from operational predictions Future tasks include: Estimation of more realistic probabilistic distribution of prediction errors; More case studies changing flood situations, target areas, or error parameter settings; Providing information how good predictions should be to be successfully applied to preliminary release operation Discussions as a concluding remarks,

THANK YOU FOR YOUR ATTENTION ! ANY QUESTIONS?