Reservoir ecological flow release control system based on climate change uncertainty
Patent Information
- Application Number
- CN202610890867.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-18
- Publication Date
- 2026-09-15
Smart Images
Figure CN122755439A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of water resources and ecological protection technology, specifically to a reservoir ecological flow release control system based on the uncertainty of climate change. Background Technology
[0002] Traditional reservoir ecological regulation systems largely rely on historical hydrological statistics, using fixed ecological flow thresholds or deterministic optimization models to generate release rules. These technical solutions implicitly assume "climate stability" and cannot cope with the spatiotemporal evolution of runoff sequence mean, variance, and extreme value characteristics under global warming. When future climate deviates significantly from historical patterns, fixed rules will lead to continuous damage to ecological flow or a substantial decline in water resource utilization efficiency.
[0003] Furthermore, climate change projections themselves involve multiple uncertainties, including emission scenarios, climate model structures, downscaling techniques, and hydrological parameters. Existing technical solutions lack a complete quantitative transmission and closed-loop response mechanism for uncertainties from their source to decision-making, and cannot ensure the robustness and adaptability of reservoir ecological scheduling decisions under strongly non-stationary conditions.
[0004] Therefore, there is an urgent need for a reservoir ecological flow release control system that can incorporate climate uncertainty throughout the entire process and make adaptive decisions based on real-time feedback. Summary of the Invention
[0005] To address this issue, the present invention provides a reservoir ecological flow release control system based on the uncertainty of climate change, in order to solve the problems in the prior art.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A reservoir ecological flow release control system based on climate change uncertainty includes a data acquisition layer, a model calculation layer, a decision control layer, and an execution feedback layer.
[0008] The data acquisition layer is used to acquire climate reanalysis data, global climate model output, real-time hydrological and meteorological monitoring data of the basin, and downstream ecological response monitoring data.
[0009] The model calculation layer receives data from the data acquisition layer, which is used to transform the output of the global climate model into a high-resolution meteorological field set, simulate the corresponding runoff set, dynamically assess the time-varying ecological water demand range, quantify the multiple uncertainties of climate-hydrology-ecology, and generate risk probability and robustness descriptions.
[0010] The decision control layer receives data collected by the data acquisition layer, data output by the model calculation layer, and data uploaded by the execution feedback layer. It performs adaptive discharge decision fusion on the received data and outputs discharge control commands.
[0011] The execution feedback layer receives and executes the discharge control command, while simultaneously collecting downstream actual ecological response data and transmitting the deviation information in the response data back to the decision control layer.
[0012] Furthermore: the data acquisition layer includes:
[0013] Meteorological reanalysis and GCM data interface: Obtain daily-scale climate variables from multiple global climate models under different emission scenarios from the Global Climate Model Data Center, and provide historical reanalysis data;
[0014] The watershed hydrological and water quality remote sensing station network collects precipitation, air temperature, soil moisture content, snow cover status and water quality parameters in the reservoir-controlled watershed in real time and transmits them to the model calculation layer;
[0015] Downstream ecological monitoring sensors collect flow velocity, flow rate, water temperature, dissolved oxygen, and acoustic signals of indicator species downstream of the discharge section, and send the data to the ecological water demand assessment unit in the model calculation layer and the adaptive discharge decision fusion unit in the decision control layer, respectively.
[0016] Furthermore: the model computation layer includes:
[0017] The climate scenario ensemble generation unit is connected to the meteorological reanalysis and GCM data interface; it is used to transform the outputs of multiple global climate models under different emission scenarios into a spatially consistent and bias-corrected multi-member meteorological forcing field ensemble.
[0018] The watershed hydrological model unit is connected to the climate scenario ensemble generation unit and the watershed hydrological-water quality telemetry station network; it is used to receive meteorological forcing field ensembles and real-time watershed status data, and to generate the ensemble and probability distribution of reservoir inflow runoff by using a distributed hydrological model combined with multi-parameter ensemble simulation.
[0019] The dynamic assessment unit for ecological water demand is connected to the downstream ecological monitoring sensors and the watershed hydrological model unit; it is used to generate the time-varying lower and upper limits of ecological water demand based on future hydrological and water temperature change trends and in combination with the indicator species habitat model, with an accompanying range of boundary uncertainty.
[0020] The uncertainty quantification engine connects to the watershed hydrological model unit and the ecological water demand dynamic assessment unit; it is used to receive the inflow runoff set and the ecological water demand interval, as well as the uncertainty of the inflow runoff set and the ecological water demand interval, calculate the ecological water shortage risk probability curve under the candidate release flow, and output the results to the decision control layer.
[0021] Furthermore: the decision control layer includes:
[0022] A reservoir multi-objective stochastic optimization scheduling unit is connected to a watershed hydrological model unit and an ecological water demand dynamic assessment unit; it is used to generate a set of release strategies under multiple scenarios based on runoff sets and ecological water demand intervals, using stochastic programming or scenario tree optimization frameworks.
[0023] The climate state identifier is connected to the watershed hydrological and water quality telemetry network and external climate telemetry data sources to identify the current and future climate state and its transition probability in real time.
[0024] The adaptive release decision fusion unit is connected to the uncertainty quantification engine, the reservoir multi-objective stochastic optimization scheduling unit, the climate state identifier, and the downstream ecological monitoring sensor, respectively. It is used to select or fine-tune the release strategy set to generate the final release flow in each scheduling period, combining the current water storage status, runoff forecast, climate state probability, risk curve, and ecological response deviation in the previous period, and output it to the execution feedback layer.
[0025] The risk warning and manual intervention interface is connected to the adaptive release decision fusion unit. It is used to issue a warning signal when the ecological water shortage risk in any scenario exceeds the preset threshold, and allows operators to intervene and make adjustments.
[0026] Furthermore: the execution feedback layer includes:
[0027] The gate / valve actuator controller is connected to the adaptive discharge decision fusion unit to receive discharge flow commands and drive the gate or valve opening adjustment.
[0028] The downstream actual ecological response monitoring feedback module is connected to the downstream ecological monitoring sensor and the adaptive release decision fusion unit. It is used to transmit the actual ecological monitoring indicators after release as the feedback input of the decision fusion unit.
[0029] The scheduling log and model deviation recording module is used to record each release decision, the deviation between actual runoff and forecast values, and the status of ecological indicators for later model parameter verification.
[0030] Furthermore, the climate scenario ensemble generation unit includes a spatial downscaling module, a physical consistency reconstruction module, and a trend deviation correction module.
[0031] Furthermore: the spatial downscaling module uses quantile mapping combined with wet day frequency correction to process precipitation variables, uses normal quantile mapping to process temperature variables, and reconstructs the interdependence structure among multiple variables through Gaussian connection functions;
[0032] The trend deviation correction module employs quantile increment mapping to eliminate climate-state mean bias while preserving future change signals simulated by global climate models.
[0033] Furthermore: the watershed hydrological model unit includes a distributed hydrological model engine, a parameter uncertainty sampler, and an ensemble simulation scheduler;
[0034] The distributed hydrological model engine uses a variable infiltration capacity model combined with a river confluence scheme to simulate runoff generation and confluence.
[0035] The parameter uncertainty sampler uses Latin hypercube sampling to generate multiple sets of sensitive parameters.
[0036] The ensemble simulation scheduler cross-runs each set of meteorological forcing fields with all sensitive parameter sets to generate an inflow runoff set containing multiple members.
[0037] Furthermore: the uncertainty quantification engine includes an ensemble preprocessing unit, a Bayesian model average probability prediction unit, an ecological threshold randomization processing unit, and a risk probability and robustness calculation unit;
[0038] The set preprocessing unit performs deviation correction and outlier removal on each member of the inflow runoff set.
[0039] The Bayesian model average probability prediction unit trains the weights and variances of each member using measured flow data within a sliding window, and integrates multi-point forecasts into a probability density distribution of inflow runoff.
[0040] The ecological threshold randomization processing unit processes the lower and upper limits of ecological water demand into random variables centered on the nominal value.
[0041] The risk probability and robustness calculation unit performs Monte Carlo sampling on the candidate release flow sequence based on the inflow runoff probability density distribution and the randomized ecological water demand boundary to calculate the ecological flow damage probability and form the relationship curve between release flow and damage probability. It is also used to calculate the uncertainty level that each release decision can withstand under the assumption of no probability.
[0042] Furthermore: the reservoir multi-objective stochastic optimization scheduling unit adopts a deep reinforcement learning agent configuration;
[0043] The state space of the deep reinforcement learning agent includes the reservoir water level, statistics of the inflow runoff ensemble forecast, climate state indicators, and the frequency of downstream indicator species occurrence. The action space is the discharge flow. The reward function simultaneously includes a power generation revenue term, a water supply satisfaction term, and an ecological flow damage penalty term.
[0044] The deep reinforcement learning agent is trained offline using a virtual annual hydrological sequence generated by perturbation expansion of historical meteorological sequences to obtain a policy network, and then fine-tuned based on actual feedback during online operation.
[0045] This invention has the following advantages: First, it sets up an uncertainty quantification engine to quantify and aggregate multiple uncertainties, such as climate models, downscaling, hydrological parameters, and ecological thresholds, into a unified ecological water shortage risk curve, solving the problem of unquantifiable risk. Second, by leveraging a climate state identifier and an adaptive release decision fusion unit, it can autonomously perceive large-scale climate mode changes and dynamically switch scheduling strategies, solving the problem of unadaptive response. Third, it utilizes real-time feedback from downstream ecological responses to form a rolling verification closed loop, continuously correcting model biases and significantly improving adaptability to future deep uncertainties and long-term operational robustness.
[0046] Other features and advantages of the present invention will be set forth in the following description. Attached Figure Description
[0047] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).
[0048] Figure 1 A system block diagram of a reservoir ecological flow release control system based on climate change uncertainty provided in an embodiment of this application.
[0049] Figure 2 This is a schematic diagram illustrating the joint implementation process of the climate scenario set generation unit and the watershed hydrological model unit in the model calculation layer of this invention.
[0050] Figure 3 This is a schematic diagram illustrating the implementation process of the uncertainty quantification engine in an embodiment of the present invention. Detailed Implementation
[0051] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Those skilled in the art can make some non-essential improvements and adjustments to the present invention based on the above-described content.
[0052] Please see Figures 1-3 The reservoir ecological flow release control system based on the uncertainty of climate change includes four layers: data acquisition layer, model calculation layer, decision control layer, and execution feedback layer.
[0053] The data acquisition layer is used to acquire climate reanalysis data, global climate model outputs, real-time hydro-meteorological monitoring data of the basin, and downstream ecological response monitoring data.
[0054] The model computation layer, connected to the data acquisition layer, is used to transform the output of global climate models into a high-resolution meteorological field set, simulate the corresponding runoff set, dynamically assess the time-varying ecological water demand range, quantify multiple uncertainties in climate-hydrology-ecology, and generate risk probability and robustness descriptions.
[0055] The decision control layer, connected to the model calculation layer, is used to perform adaptive release decision fusion in a multi-objective stochastic optimization strategy set based on uncertainty quantification results, current reservoir status and climate status identification information, and output release control commands.
[0056] The execution feedback layer, connected to the decision control layer, is used to execute release commands and collect actual downstream ecological responses, and to send deviation information back to the decision control layer to form a closed-loop verification.
[0057] The data acquisition layer includes meteorological reanalysis and GCM data interfaces, a watershed hydrological and water quality telemetry network, and downstream ecological monitoring sensors.
[0058] Among them, the meteorological reanalysis and GCM data interface obtains future daily meteorological sequences of at least 5 GCMs and 3 SSP scenarios from the Global Climate Model Data Center (CMIP6, etc.); in this embodiment, the 3 SSP scenarios are SSP1-2.6, SSP2-4.5, and SSP5-8.5.
[0059] The watershed hydrological and water quality remote sensing station network includes rain gauges, water level stations, soil moisture stations, and automatic water quality monitoring stations, which upload data in real time via the Internet of Things;
[0060] The downstream ecological monitoring sensors, consisting of an acoustic Doppler current profiler, a multi-parameter water quality meter, and a fish echo detector, are deployed at ecologically sensitive sections to continuously monitor the flow field and ecological response after the release.
[0061] The model computation layer includes a climate scenario set generation unit, a watershed hydrological model unit, an ecological water demand dynamic assessment unit, and an uncertainty quantification engine.
[0062] The climate scenario ensemble generation unit includes a GCM data acquisition interface, a spatial downscaling module, a physical consistency reconstruction module, and a trend bias correction module. It is used to generate a high-resolution, multi-member meteorological forcing field ensemble by spatially downscaling, physically consistent processing, and trend bias correction of multiple scenario outputs from different global climate models.
[0063] The GCM data acquisition interface pulls at least five daily-scale outputs of GCMs (such as EC-Earth3, MRI-ESM2-0, GFDL-ESM4, etc.) from the CMIP6 archive node via standard protocols (such as OpenDAP) under three shared socioeconomic paths (SSP1-2.6, SSP2-4.5, SSP5-8.5). Variables include precipitation, maximum temperature, minimum temperature, surface downward shortwave radiation, wind speed, specific humidity, etc. The GCM data acquisition interface is equipped with data integrity verification and time slice alignment functions.
[0064] Spatial downscaling and physical consistency processing employs quantile mapping in statistical downscaling combined with local intensity scaling techniques. Targeting a historical 0.1° grid observation dataset of the watershed (such as CN05.1), transfer functions are established for daily precipitation and daily temperature, respectively. Specifically, precipitation is mapped using gamma distribution quantiles with added wet day frequency correction, while temperature is mapped using normal distribution quantiles.
[0065] To ensure the physical correlation among multiple variables, Gaussian Copula dependency reconstruction was applied to variables such as radiation and wind speed after marginal distribution correction, so that the downscaled meteorological field statistically retains the interdependence structure among variables; the downscaling result is a daily meteorological series with a spatial resolution of 0.1°×0.1°.
[0066] Trend bias correction addresses the long-term trend bias of GCM by calculating the mean rate of change of each variable between the simulation period and the reference period (e.g., 1995-2014) within a sliding window (e.g., 30 years). Then, the quantile increment mapping method is used to both preserve the future change signals simulated by GCM (e.g., warming, increased precipitation variability) and eliminate its climatological mean bias.
[0067] The meteorological forcing field set is packaged, and finally a set of meteorological driving files (NetCDF format) covering the next 30-50 years with a spatial resolution of 0.1° and daily coverage is generated for each group (GCM, SSP). Uncertainty source tags (model ID, SSP, implementation member) are marked in the metadata. These files constitute a multidimensional set of climate scenario × emission path, which can be directly read by the watershed hydrological model unit.
[0068] The watershed hydrological model unit includes a distributed hydrological model engine, a parameter uncertainty sampler, and an ensemble simulation scheduler;
[0069] In this embodiment, the distributed hydrological model engine uses the large-scale distributed hydrological model VIC (Variable Infiltration Capacity Model) to divide the watershed into 0.1° grids. The VIC model independently simulates the energy and water balance of each grid and can output the runoff depth grid by grid. Then, it is calculated to the reservoir dam site section by a river confluence model (such as Lohmann routing) to form the inflow sequence.
[0070] The parameter uncertainty sampler identifies parameters in the VIC that are sensitive to and poorly understood in relation to runoff generation and confluence processes, such as soil layer thickness (d1, d2, d3), variable infiltration curve parameters (binfil), and baseflow nonlinear regression coefficients (Ds, Dsmax, Ws), totaling 6-8 parameters. Latin hypercube sampling is used to generate 100-200 parameter sets from the prior distribution of the parameters (based on soil texture and range determined by remote sensing inversion), with each parameter set representing a possible realization of watershed characteristics.
[0071] The ensemble simulation scheduler sequentially performs hydrological simulations using 100-200 parameter sets for each meteorological forcing field in the climate scenario set; that is, the number of simulations = number of GCMs × number of SSPs × number of parameter sets, forming a complete cross set of climate × parameters; the ensemble simulation scheduler adopts a parallel computing architecture and can be deployed on high-performance computing clusters or cloud computing platforms, with each parameter set-climate pair running as an independent task.
[0072] Post-processing and probability output: The daily inflow sequence obtained from the simulation at the dam site is statistically analyzed using quantiles for the predicted time period. Specifically:
[0073] The flow distribution (parameter uncertainty) resulting from the set of calculation parameters within each scenario (same GCM+SSP);
[0074] By mixing the traffic of all members across scenarios, a large-scale probability forecast is generated, which gives the 5th, 25th, 50th, 75th and 95th percentiles of the daily inflow traffic, as well as the probability that the traffic is below a certain ecological threshold.
[0075] The probability distribution is sent via data bus to the reservoir multi-objective stochastic optimization scheduling unit and uncertainty quantification engine in the decision control layer for subsequent robust optimization.
[0076] In addition, to improve the accuracy of near-term forecasts, the watershed hydrological model unit can reserve an assimilation interface to receive real-time soil moisture and snow water equivalent data from the telemetry network. The VIC state variables of each parameter set are updated through ensemble Kalman filtering, so that the runoff ensemble is closer to the actual measurement in the 7 days before the forecast period.
[0077] The ecological water demand dynamic assessment unit embeds a habitat suitability model (such as the PHABSIM framework). It uses predictions of future flow, water temperature, dissolved oxygen, etc., to calculate the suitable flow range for different life stages of target species (such as schizothorax and sturgeon), outputs time-varying lower and upper limits, and includes the range perturbation amplitude caused by water temperature and water quality threshold drift, as input to the uncertainty quantification engine.
[0078] like Figure 2The schematic diagram shown illustrates the joint implementation process of the climate scenario ensemble generation unit and the watershed hydrological model unit, intuitively describing the complete information chain from climate model data input to inflow runoff ensemble output; the specific processing flow is as follows:
[0079] a. First, the climate scenario ensemble generation unit obtains the global daily-scale output of multiple gcm and multiple emission scenarios through the interface, and then undergoes spatial downscaling, physical consistency reconstruction and trend bias correction to obtain a high-resolution multi-scenario meteorological ensemble that can be directly input into the hydrological model.
[0080] b. Subsequently, the watershed hydrological model unit reads these meteorological fields, combines them with hundreds of model parameter sets generated by Latin hypercube sampling, and uses the distributed VIC model to run the overall cross simulation. Finally, at the reservoir dam site section, it produces a set of inflow runoff that integrates multiple uncertainties of climate models, emission paths, downscaling processes and hydrological parameters, and provides it to the downstream decision-making process in the form of a probability distribution.
[0081] The uncertainty quantification engine receives the inflow runoff set output by the watershed hydrological model unit and the time-varying ecological water demand interval and its uncertainty description output by the ecological water demand dynamic assessment unit. For the candidate release flow from the decision control layer, it quantifies the probability that the release behavior will lead to ecological water shortage and generates the release flow-ecological water shortage risk curve. At the same time, it can output the robust protection boundary under the worst scenario to deal with the deep uncertainty when the model is unreliable.
[0082] The uncertainty quantification engine includes an ensemble preprocessing unit, a Bayesian model average probability prediction unit (BMA probability prediction unit), an ecological threshold randomization processing unit, and a risk probability and robustness analysis unit.
[0083] The preprocessing unit performs outlier removal, bias correction, and assimilation period marking on the runoff ensemble, generating a formatted forecast membership matrix to ensure the purity of subsequent probabilistic modeling.
[0084] The Bayesian model average probability forecasting unit uses a sliding window Bayesian model averaging method to merge runoff forecasting members from different GCM, SSP and parameter combinations into a single probability density function, outputting the normal mixture distribution parameters of daily inflow within the forecast period.
[0085] The ecological threshold randomization processing unit accepts the time-varying lower limit and time-varying upper limit given by the ecological water demand dynamic assessment unit, and characterizes the interval disturbances caused by uncertainties in water quality, water temperature and phenology as random variables that follow a uniform distribution or a Beta distribution, forming a joint probability description of the ecological water demand boundary.
[0086] The risk probability and robustness calculation unit performs Monte Carlo simulation or numerical convolution on a given release decision set to calculate the probability of ecological flow damage and plot the risk curve; at the same time, it embeds an information gap robustness model to estimate the maximum level of uncertainty that the system can tolerate and the corresponding guaranteed release domain under no probability assumptions.
[0087] The specific implementation method of the uncertainty quantization engine is as follows:
[0088] (1) Ensemble preprocessing: The daily inflow ensemble obtained from the watershed hydrological model unit contains N members (N = number of climate scenarios × number of parameter sets); First, the ratio or difference between the simulated sequence and the measured flow of each member in the historical period (e.g., the past 10 years) is statistically analyzed, the linear deviation correction factor of each member is calculated, and it is applied to the future forecast sequence; Second, members that show significant abnormalities in the historical period (MAE exceeding 3 times the median of the ensemble) are removed to obtain the effective ensemble S;
[0089] (2) Bayesian model average probability forecast, using a sliding training window (e.g., a window width of 45 days, scrolling forward each day), using the measured flow of the most recent 45 days and the forecast values of each member in the same period to train the BMA model;
[0090] The prediction probability density function is:
[0091] ;
[0092] Where K is the total number of valid forecast members; Let be the posterior weight of the k-th member; For For the mean, Let be the normal distribution density function of the variance;
[0093] Weights and variances are estimated using the expectation-maximization algorithm, and the likelihood function adopts the logarithmic scoring rule to ensure the sharpness and calibration of the probability forecast.
[0094] After training, for each day of the forecast period, BMA outputs a normal mixture distribution, which can be simplified to be described by the mean, quantiles and probability density curves.
[0095] To ensure adaptability under non-stationary climate conditions, BMA weights are allowed to change slowly over time: when a climate state transition is detected (triggered by the climate state identifier) or a significant trend appears in the flow sequence within the training window (Mann-Kendall test), the weights are updated with a decay factor, gradually favoring members with better recent performance.
[0096] (3) Ecological threshold randomization: The ecological water demand dynamic assessment unit outputs the lower boundary based on future changes in water temperature and habitat area. and upper boundary The nominal value and fluctuation range;
[0097] Uncertainty quantification engine will Modeled as a uniform distribution ,
[0098] , ;
[0099] Given by the uncertainty of the habitat model;
[0100] Similarly, Modeled as a uniform distribution When ecological water demand is given only as a lower limit with no upper limit, only the destruction of the lower limit is considered. The two boundaries can be assumed to be independent or related by a copula, but in engineering practice, independence is more conservatively assumed.
[0101] (4) Generation of the risk curve for releasing decision-making, based on the set of candidate release flows transmitted from the decision-making level. (For example, using a step size of 5 m³ / s from 0 to the maximum discharge capacity), perform the following calculation for each candidate value Q:
[0102] 1) From the inflow probability density generated by the BMA, M random samples (M=10) are drawn based on the reservoir water balance relationship (if the discharge is directly controlled, the inflow-outflow delay can be ignored, and it is conservatively assumed that the discharge equals the river flow). 5 ), denoted as ;
[0103] 2) At the same time from and M samples are drawn from each of the distributions, denoted as L and U;
[0104] 3) Statistics on sabotage incidents: or The sample proportion (if an upper limit exists) represents the probability of ecological water shortage risk under this release decision. ;
[0105] 4) Iterate through all candidate Qs to obtain the curve. This curve is directly passed to the adaptive discharge decision fusion unit to select the operating point in the cost-risk trade-off.
[0106] (5) Information gap robustness enhancement (for deep uncertainty): When decision-makers have doubts about the reliability of probabilistic assumptions, the engine additionally performs information gap analysis, as follows:
[0107] 1) Define uncertainty parameters Allowing actual runoff to be within the BMA forecast range Outside the range of fluctuations, meaning that the actual runoff may belong to the set. (i.e., the uncertainty interval); where, For BMA forecasts quantiles; For BMA forecasts quantiles;
[0108] 2) The ecological water demand boundary is also scaled to the most unfavorable direction; where L is the upper bound of the interval and U is the lower bound of the interval;
[0109] 3) For a candidate drain Q, solve for... The maximum value that ensures ecological flow remains intact (i.e., Q always falls within [L,U]) under a level of uncertainty. value ;
[0110] 4) Output robustness curve This indicates the degree of uncertainty that the decision to release the risk can tolerate, assuming the probability of loss is forgone. Decision-makers can choose to guarantee [the risk] based on their own level of uncertainty aversion. The discharge flow.
[0111] See Figure 3 The implementation process of the uncertainty quantification engine is as follows:
[0112] The multi-member inflow runoff set first enters the preprocessing stage, where systematic biases and abnormal members are removed by sliding correction of historical measured data to form a purified forecast matrix;
[0113] Meanwhile, recent measured flow rates are fed into the BMA unit as a training target. The weights and forecast variances of each member are estimated using a 45-day rolling window. The deterministic multipoint forecasts are fused into a normal mixture probability density, giving a complete probability description of the inflow flow for each day in the future.
[0114] The lower and upper bound nominal values and their disturbance amplitudes from the ecological water demand assessment unit are randomized into uniformly distributed variables to cover threshold drift caused by water temperature, phenology, etc.
[0115] After receiving the candidate release flow sequence provided by the decision-making level, the risk probability calculation unit simultaneously performs large-scale Monte Carlo sampling on the runoff probability distribution and the ecological threshold distribution, counts the frequency of ecological flow damage under each candidate release value, and generates a release flow-damage probability curve.
[0116] For cases where the reliability requirements of the probabilistic model are extremely stringent, the system performs information gap robustness calculation in parallel to solve the uncertainty level that each release decision can withstand without relying on probabilistic assumptions, and plots the robustness curve.
[0117] Ultimately, both are provided to the adaptive discharge decision fusion unit as a direct basis for risk-cost trade-offs.
[0118] The decision control layer includes a reservoir multi-objective stochastic optimization scheduling unit, a climate state identifier, an adaptive release decision fusion unit, and a risk warning and manual intervention interface.
[0119] The reservoir multi-objective stochastic optimization scheduling unit can be configured into two forms depending on the application scenario;
[0120] Form 1 can adopt multi-stage scenario tree stochastic programming, that is, the scenario tree nodes correspond to the climate state, the branches represent the water inflow scenarios of different GCM-SSP combinations, and the maximum expected power generation under the most unfavorable scenario with an ecological damage rate of no more than 5% can be solved offline to form a robust release rule table.
[0121] Form 2 can employ a deep reinforcement learning agent. The agent's state space includes reservoir water level, runoff ensemble forecast statistics, climate status indicators, and the frequency of downstream ecological occurrences. The action space is the discharge flow. The reward function coordinates power generation, water supply, and ecological penalties. The policy network is obtained through offline training with a large number of virtual annual sequences and fine-tuned monthly during online operation.
[0122] The climate state identifier receives indicators such as sea surface temperature anomaly, Southern Oscillation Index, and Arctic Oscillation, and uses Hidden Markov Models or pattern matching algorithms to output the current climate state (such as El Niño development period, La Niña decline period, and persistent drought period) and the state transition probability distribution for the next few months, which is used by the decision fusion unit for weighted selection in the strategy set.
[0123] The adaptive release decision fusion unit obtains risk curves from the uncertainty quantification engine, strategy sets from the reservoir multi-objective stochastic optimization scheduling unit, state probabilities from the climate state identifier, and water storage status and ecological feedback deviation from the data acquisition layer. It adopts a rolling time domain control framework for multi-source information fusion.
[0124] Form 1, an adaptive release decision fusion unit, can directly look up robust release rules from tables based on real-time climate conditions and adjust boundaries in conjunction with rolling runoff forecasts;
[0125] Form 2, the adaptive release decision fusion unit, can input the current state into the online policy network to obtain the basic release action, and then make mandatory constraint corrections based on the extreme scenario risks given by the risk curve. In both forms, the unit can apply a small bias to the release amount based on the actual downstream ecological response (for example, real-time feedback of ecological indicators that can respond quickly after release (such as changes in dissolved oxygen, water temperature, and flow velocity field), and use long-term ecological indicators such as fish spawning frequency as the basis for periodic verification), so as to achieve closed-loop adaptation.
[0126] The execution feedback layer includes a gate / valve execution controller, a downstream actual ecological response monitoring and feedback module, and a scheduling log and model deviation recording module.
[0127] Among them, the discharge gate / valve actuator controller receives the flow command output by the fusion unit, converts it into gate opening degree or valve position signal, drives the actuator through the PID controller, and sends back the actual opening degree and flow monitoring value.
[0128] The downstream actual ecological response monitoring and feedback module packages and sends dissolved oxygen, water temperature, fish sonar signals, etc. after the release to the adaptive release decision fusion unit as feedback items for the next decision.
[0129] The scheduling log and model deviation recording module stores the entire chain of data, including forecasted inbound traffic, actual inbound traffic, release instructions, and ecological indicators, into the database, providing historical samples for the deviation correction of the uncertainty quantification engine and the online adjustment of the decision fusion unit.
[0130] In the actual operation of this invention, the data acquisition layer continuously acquires multi-GCM climate output and real-time watershed and ecological monitoring data; the model calculation layer performs daily rolling operations, the climate scenario set generation unit updates the meteorological set, the watershed hydrological model unit rolls the prediction of the inflow runoff set for the next 14-30 days, the ecological water demand assessment unit updates the ecological water demand range, and the uncertainty quantification engine integrates the above uncertainties into a risk curve.
[0131] The decision control layer triggers a decision once every scheduling period (e.g., 6 hours or 1 day): the climate state identifier first outputs the state probability, and the adaptive release decision fusion unit generates the current release instruction based on the candidate strategy, risk curve, climate state and ecological feedback of the previous period provided by the scheduling unit, and the gate is adjusted by the execution controller.
[0132] After the release, the downstream sensor data is returned to the fusion unit to form a closed loop. When the risk warning interface detects that the probability of ecological water shortage in any scenario in the next week exceeds the limit, it will prompt the operators to intervene through the human-machine interface and can automatically raise the confidence lower limit of the release volume.
[0133] In this invention, through the above system, the reservoir can automatically switch scheduling strategies based on quantitative uncertainty in the strong non-stationary hydrological sequence caused by climate change, and maximize the benefits of power generation or water supply while ensuring that the probability of downstream ecological flow damage is controllable, so as to achieve truly adaptive ecological flow control in the face of future uncertainties.
[0134] Furthermore, this invention also provides a method for controlling reservoir ecological flow release based on climate change uncertainty, comprising the following steps:
[0135] Step S1 (Data Acquisition): Acquire multi-source data;
[0136] Step S2 (Model calculation, parallel or sequential): Based on the data from S1, generate the inflow runoff set and the ecological water demand range;
[0137] Step S3 (Uncertainty Quantification): Calculate the ecological water shortage risk curve using the results of S2;
[0138] Step S4 (Decision Control): Based on the risk curve of S3, the data of S1, and the results of S2, generate a release instruction;
[0139] Step S5 (Execution Feedback): Execute the release command of S4 and send back the actual ecological response to correct the subsequent decisions of S4.
[0140] During the implementation of this technical solution, the data collection activities of all participating parties must strictly comply with the provisions of laws and regulations applicable to the data source and processing activities, including data security. Data collection is based on the principles of legality, legitimacy, and necessity, and is conducted on the premise of obtaining the explicit consent of the information subject or meeting other legal legal basis. According to actual business needs, security protection measures such as minimizing the scope and accuracy of data, de-identification, and anonymization are adopted in accordance with the risk level. At the same time, complete operation logs, source certificates, and authorization records are retained throughout the data collection process to meet the requirements of subsequent compliance audits and quality traceability.
[0141] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A reservoir ecological flow release control system based on the uncertainty of climate change, characterized in that, It includes a data acquisition layer, a model calculation layer, a decision control layer, and an execution feedback layer; The data acquisition layer is used to acquire climate reanalysis data, global climate model output, real-time hydrological and meteorological monitoring data of the basin, and downstream ecological response monitoring data. The model calculation layer receives data from the data acquisition layer, which is used to transform the output of the global climate model into a high-resolution meteorological field set, simulate the corresponding runoff set, dynamically assess the time-varying ecological water demand range, quantify the multiple uncertainties of climate-hydrology-ecology, and generate risk probability and robustness descriptions. The decision control layer receives data collected by the data acquisition layer, data output by the model calculation layer, and data uploaded by the execution feedback layer. It performs adaptive discharge decision fusion on the received data and outputs discharge control commands. The execution feedback layer receives and executes the discharge control command, while simultaneously collecting downstream actual ecological response data and transmitting the deviation information in the response data back to the decision control layer.
2. The reservoir ecological flow release control system based on climate change uncertainty as described in claim 1, characterized in that, The data acquisition layer includes: Meteorological reanalysis and GCM data interface: Obtain daily-scale climate variables from multiple global climate models under different emission scenarios from the Global Climate Model Data Center, and provide historical reanalysis data; The watershed hydrological and water quality remote sensing station network collects precipitation, air temperature, soil moisture content, snow cover status and water quality parameters in the reservoir-controlled watershed in real time and transmits them to the model calculation layer; Downstream ecological monitoring sensors collect flow velocity, flow rate, water temperature, dissolved oxygen, and acoustic signals of indicator species downstream of the discharge section, and send the data to the ecological water demand assessment unit in the model calculation layer and the adaptive discharge decision fusion unit in the decision control layer, respectively.
3. The reservoir ecological flow release control system based on climate change uncertainty as described in claim 1, characterized in that, The model computation layer includes: The climate scenario ensemble generation unit is connected to the meteorological reanalysis and GCM data interface; it is used to transform the outputs of multiple global climate models under different emission scenarios into a spatially consistent and bias-corrected multi-member meteorological forcing field ensemble. The watershed hydrological model unit is connected to the climate scenario ensemble generation unit and the watershed hydrological-water quality telemetry station network; it is used to receive meteorological forcing field ensembles and real-time watershed status data, and to generate the ensemble and probability distribution of reservoir inflow runoff by using a distributed hydrological model combined with multi-parameter ensemble simulation. The dynamic assessment unit for ecological water demand is connected to the downstream ecological monitoring sensors and the watershed hydrological model unit; it is used to generate the time-varying lower and upper limits of ecological water demand based on future hydrological and water temperature change trends and in combination with the indicator species habitat model, with an accompanying range of boundary uncertainty. The uncertainty quantification engine connects to the watershed hydrological model unit and the ecological water demand dynamic assessment unit; it is used to receive the inflow runoff set and the ecological water demand interval, as well as the uncertainty of the inflow runoff set and the ecological water demand interval, calculate the ecological water shortage risk probability curve under the candidate release flow, and output the results to the decision control layer.
4. The reservoir ecological flow release control system based on climate change uncertainty as described in claim 1, characterized in that, The decision control layer includes: A reservoir multi-objective stochastic optimization scheduling unit is connected to a watershed hydrological model unit and an ecological water demand dynamic assessment unit; it is used to generate a set of release strategies under multiple scenarios based on runoff sets and ecological water demand intervals, using stochastic programming or scenario tree optimization frameworks. The climate state identifier is connected to the watershed hydrological and water quality telemetry network and external climate telemetry data sources to identify the current and future climate state and its transition probability in real time. The adaptive release decision fusion unit is connected to the uncertainty quantification engine, the reservoir multi-objective stochastic optimization scheduling unit, the climate state identifier, and the downstream ecological monitoring sensor, respectively. It is used to select or fine-tune the release strategy set to generate the final release flow in each scheduling period, combining the current water storage status, runoff forecast, climate state probability, risk curve, and ecological response deviation in the previous period, and output it to the execution feedback layer. The risk warning and manual intervention interface is connected to the adaptive release decision fusion unit. It is used to issue a warning signal when the ecological water shortage risk in any scenario exceeds the preset threshold, and allows operators to intervene and make adjustments.
5. The reservoir ecological flow release control system based on climate change uncertainty according to claim 1, characterized in that, The execution feedback layer includes: The gate / valve actuator controller is connected to the adaptive discharge decision fusion unit to receive discharge flow commands and drive the gate or valve opening adjustment. The downstream actual ecological response monitoring feedback module is connected to the downstream ecological monitoring sensor and the adaptive release decision fusion unit. It is used to transmit the actual ecological monitoring indicators after release as the feedback input of the decision fusion unit. The scheduling log and model deviation recording module is used to record each release decision, the deviation between actual runoff and forecast values, and the status of ecological indicators for later model parameter verification.
6. The reservoir ecological flow release control system based on climate change uncertainty according to claim 3, characterized in that, The climate scenario ensemble generation unit includes a spatial downscaling module, a physical consistency reconstruction module, and a trend deviation correction module.
7. The reservoir ecological flow release control system based on climate change uncertainty according to claim 6, characterized in that, The spatial downscaling module uses quantile mapping combined with wet day frequency correction to process precipitation variables, uses normal quantile mapping to process temperature variables, and reconstructs the interdependence structure among multiple variables through Gaussian link function. The trend deviation correction module employs quantile increment mapping to eliminate climate-state mean bias while preserving future change signals simulated by global climate models.
8. The reservoir ecological flow release control system based on climate change uncertainty according to claim 3, characterized in that, The watershed hydrological model unit includes a distributed hydrological model engine, a parameter uncertainty sampler, and an ensemble simulation scheduler. The distributed hydrological model engine uses a variable infiltration capacity model combined with a river confluence scheme to simulate runoff generation and confluence. The parameter uncertainty sampler uses Latin hypercube sampling to generate multiple sets of sensitive parameters. The ensemble simulation scheduler cross-runs each set of meteorological forcing fields with all sensitive parameter sets to generate an inflow runoff set containing multiple members.
9. The reservoir ecological flow release control system based on climate change uncertainty according to claim 3, characterized in that, The uncertainty quantification engine includes an ensemble preprocessing unit, a Bayesian model average probability prediction unit, an ecological threshold randomization processing unit, and a risk probability and robustness calculation unit. The set preprocessing unit performs deviation correction and outlier removal on each member of the inflow runoff set. The Bayesian model average probability prediction unit trains the weights and variances of each member using measured flow data within a sliding window, and integrates multi-point forecasts into a probability density distribution of inflow runoff. The ecological threshold randomization processing unit processes the lower and upper limits of ecological water demand into random variables centered on the nominal value. The risk probability and robustness calculation unit performs Monte Carlo sampling on the candidate release flow sequence based on the inflow runoff probability density distribution and the randomized ecological water demand boundary to calculate the ecological flow damage probability and form the relationship curve between release flow and damage probability. It is also used to calculate the uncertainty level that each release decision can withstand under the assumption of no probability.
10. The reservoir ecological flow release control system based on climate change uncertainty according to claim 4, characterized in that, The reservoir multi-objective stochastic optimization scheduling unit adopts a deep reinforcement learning agent configuration. The state space of the deep reinforcement learning agent includes the reservoir water level, statistics of the inflow runoff ensemble forecast, climate state indicators, and the frequency of downstream indicator species occurrence. The action space is the discharge flow. The reward function simultaneously includes a power generation revenue term, a water supply satisfaction term, and an ecological flow damage penalty term. The deep reinforcement learning agent is trained offline using a virtual annual hydrological sequence generated by perturbation expansion of historical meteorological sequences to obtain a policy network, and then fine-tuned based on actual feedback during online operation.