Cascade reservoir pre-flood joint hydro-fluctuation sequence and hydro-fluctuation scheme construction method

Abstract

The invention discloses a cascade reservoir pre-flood joint hydro-fluctuation sequence and hydro-fluctuation scheme construction method, which comprises the following steps: based on a pre-flood state parameter set of a cascade reservoir, carrying out conditional disturbance on forecast deviation, generating a multi-scene hydrological operation input set, and quantifying non-hydro-fluctuation risk distribution parameter data representing a water level over-limit possibility; constructing a pre-flood joint hydro-fluctuation optimization model in which a hydro-fluctuation sequence decision variable is introduced, and converting non-hydro-fluctuation risk distribution parameter data into a risk constraint condition therein; solving a pre-flood joint hydro-fluctuation optimization model according to the multi-scene hydrological operation input set to obtain hydro-fluctuation sequence optimization result data meeting a safety threshold; and performing pattern recognition on the hydro-fluctuation sequence optimization result data, extracting a hydro-fluctuation sequence pattern and generating corresponding pre-flood hydro-fluctuation scheme rule base data. According to the method, the hydro-fluctuation sequence is brought into the optimization decision, and the non-hydro-fluctuation risk is explicitly constrained, so that the cooperative improvement of flood control safety and power generation benefits of the cascade reservoir in a strong uncertain environment is realized.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy engineering, and in particular, it is a method for constructing a pre-flood joint drawdown sequence and drawdown scheme for cascade reservoirs. Background Technology

[0002] Pre-flood drawdown of cascade reservoirs is a crucial scheduling phase that balances flood control safety and power generation efficiency. Before the flood season, the cascade reservoir group must lower its water level to the flood control limit to create flood control capacity, while simultaneously maintaining high water head operation to maximize power generation efficiency and avoid a surge in downstream flood control pressure or water wastage due to concentrated drawdown. How to scientifically formulate drawdown plans to achieve safe and efficient coordinated drawdown of the cascade reservoir group under variable hydrological and meteorological conditions has significant engineering value and economic implications.

[0003] Currently, research on cascade reservoir drawdown scheduling mainly focuses on optimizing the drawdown trajectory of a single reservoir or on joint cascade scheduling based on deterministic forecasts. Existing scheduling practices typically follow fixed rules based on experience, such as pre-setting a fixed drawdown order of upstream followed by downstream or synchronous upstream and downstream, or setting fixed initiation and drawdown levels and rates. In terms of optimization methods, most adopt deterministic optimization models based on historical typical years or deterministic inflow forecasts, focusing on finding the optimal water level process under a given order, or simply verifying the risk indicators of a given scheme through stochastic simulations.

[0004] However, existing technologies still have limitations when dealing with cascade combined drawdown under strong uncertainty. These limitations mainly manifest in the rigidity of drawdown order decisions and the lack of control over the risk of failure to draw down, making them unable to cope with extreme hydrological variations. Specifically, existing schemes typically treat the drawdown order as a fixed boundary condition rather than a decision variable, severing the coupling relationship between the inter-reservoir order and the drawdown trajectory. This prevents the model from automatically optimizing between different strategies, such as prioritizing upstream and downstream drawdown, making it difficult to find the optimal coordinated mode adapted to the current specific hydrological and rainfall conditions. Furthermore, traditional models often lack explicit mathematical modeling of the specific risk of failing to draw down to a safe water level on time, relying more on methods such as water release penalties or post-event verification to handle the risk. This results in the inability to rigidly constrain tail risks when facing significant forecast deviations or extreme inflow scenarios, easily leading to untimely emptying of flood control reservoirs or unnecessary power generation losses due to overly conservative approaches. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing a joint pre-flood drawdown sequence and drawdown scheme for cascade reservoirs, in order to solve the aforementioned problems existing in the prior art.

[0006] Technical solutions, including the construction method for the pre-flood joint drawdown sequence and drawdown scheme of cascade reservoirs, are as follows:

[0007] Based on the pre-flood state parameter set of cascade reservoirs, conditional perturbation is applied to the forecast deviation to generate a multi-scenario hydrological operation input set, and the distribution parameter data of the undrawn risk that characterizes the possibility of water level exceeding the limit are quantified.

[0008] A pre-flood joint drawdown optimization model is constructed by introducing drawdown order decision variables. The data of undrawdown risk distribution parameters are transformed into risk constraints in the pre-flood joint drawdown optimization model, and the drawdown order decision variables are coupled with the pre-stored pre-flood drawdown trajectory data.

[0009] Under risk constraints, a joint pre-flood drawdown optimization model is solved for multi-scenario hydrological operation input sets to obtain drawdown sequence optimization results data that meet safety thresholds;

[0010] Pattern recognition is performed on the drawdown sequence optimization results data to extract the drawdown sequence patterns and generate corresponding pre-flood drawdown scheme rule base data.

[0011] Beneficial effects: By incorporating the drawdown order into the optimization decision and explicitly constraining the risk of undrawdown, this invention achieves a synergistic improvement in flood control safety and power generation efficiency of cascade reservoirs under highly uncertain environments. Attached Figure Description

[0012] Figure 1 A flowchart illustrating the steps of constructing a pre-flood joint drawdown sequence and drawdown scheme for cascade reservoirs, as provided in this application embodiment.

[0013] Figure 2 A flowchart illustrating the steps involved in constructing a pre-flood joint drawdown optimization model that incorporates drawdown order decision variables, as provided in this application embodiment.

[0014] Figure 3 A flowchart illustrating the steps for establishing a coupling relationship between drawdown order decision variables and pre-flood drawdown trajectory data, as provided in this application embodiment.

[0015] Figure 4 A flowchart illustrating the steps for obtaining the optimized result data of the dropout sequence provided in this application embodiment. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms include and have, and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0018] like Figure 1 As shown, a method for constructing a pre-flood joint drawdown sequence and drawdown scheme for cascade reservoirs includes the following steps:

[0019] Obtain the pre-flood state parameter set of cascade reservoirs.

[0020] Alternatively, one can obtain the pre-flood basic dataset and pre-flood state parameter set of the cascade reservoirs, perform preprocessing, and obtain the preprocessed pre-flood basic dataset and pre-flood state parameter set.

[0021] Specifically, the process of acquiring the pre-flood basic dataset and pre-flood state parameter set for cascade reservoirs encompasses a complete data preprocessing stage, from data acquisition and cleaning to feature extraction. The pre-flood basic dataset is a comprehensive dataset containing historical, real-time, and forecast dimensions, including but not limited to: historical inflow sequences, historical rainfall, historical loads, and historical cascade operation records retrieved from reservoir scheduling and hydrological databases; current reservoir water levels, output, discharge flow, and maintenance status collected from real-time monitoring systems; and future forecasted inflow and rainfall data accessed from meteorological and hydrological systems. The pre-flood state parameter set is a structured index further processed from the aforementioned basic data, used to quantitatively characterize the hydrological, load, and safety situation faced by the reservoir group at the current moment.

[0022] In a preferred embodiment, to generate a structurally unified pre-flood baseline dataset, spatiotemporal alignment of raw data from different sources is required. For example, real-time data with non-uniform time steps can be resampled into a unified scheduling period, such as a day or hour, and the basin surface rainfall data can be mapped to the control sections of each cascade reservoir using spatial interpolation or area weighting. For missing or outlier values ​​in historical data, linear interpolation or correlation-based interpolation methods based on nearby stations can be used for repair. Based on this, the construction of the pre-flood state parameter set is achieved by extracting key feature indicators, such as calculating indicators reflecting the deviation of current water storage from historical levels for the same period, and load constraint indicators reflecting the grid's demand for peak hydropower capacity. These parameters provide necessary boundary conditions and probabilistic correction bases for subsequent scenario generation.

[0023] Based on the pre-flood state parameter set, the forecast deviation is conditionally perturbed to generate a multi-scenario hydrological operation input set, and the distribution parameter data of the undrawn risk that characterizes the possibility of water level exceeding the limit are quantitatively represented.

[0024] In this embodiment, deterministic forecast information is transformed into a set of probabilistic scenarios that reflect uncertainty, and a risk assessment benchmark is established accordingly. Conditional perturbation refers to adjusting the distribution characteristics of random errors based on the current pre-flood state parameter set, such as the indicators of water level abundance or scarcity. For example, when the state parameters indicate a significantly abundant year, the sampling weight of the positive deviation is artificially increased, i.e., the sampling weight of actual inflow exceeding the forecast inflow, generating a scenario set that better covers the risk of extremely high inflow. The multi-scenario hydrological operation input set consists of several hydrological, load, and constraint sequences with different probabilities of occurrence, each sequence representing a possible future evolution path. Optionally, based on this, the distribution parameter data of the unreceded risk, which quantifies the possibility of water level exceeding limits, is achieved through benchmark extrapolation. Specifically, preset empirical scheduling rules, such as synchronous water level receding according to the current reservoir capacity utilization rate, can be used to quickly simulate each scenario and obtain the initial water level process under each scenario. By comparing the difference between the water level during this process and the pre-flood safety control water level, the frequency, magnitude, and tail-end risk indicators of water level exceeding limits were statistically analyzed. These quantitative indicators constitute the distribution parameter data of unreceded risk, providing clear risk constraint boundaries for subsequent optimization models. This ensures that the optimization process is no longer blindly pursuing maximum power generation, but rather seeking the best within the known risk baseline.

[0025] A pre-flood joint drawdown optimization model is constructed by introducing drawdown order decision variables. The data of undrawdown risk distribution parameters are transformed into risk constraints in the pre-flood joint drawdown optimization model, and the drawdown order decision variables are coupled with the pre-stored pre-flood drawdown operation trajectory data. Under the risk constraints, the pre-flood joint drawdown optimization model is solved for multi-scenario hydrological operation input sets to obtain the drawdown sequence optimization results data that meet the safety threshold.

[0026] In other words, a joint pre-flood drawdown optimization model is constructed, in which drawdown order decision variables are introduced, and the data of non-drawdown risk distribution parameters are transformed into risk constraints. Under the risk constraints, the drawdown order decision variables and the pre-flood drawdown trajectory data under multiple scenarios are jointly optimized to obtain the drawdown sequence optimization results data.

[0027] In this embodiment, the main feature of the pre-flood joint drawdown optimization model is the explicit introduction of a drawdown order decision variable. This variable can be an integer variable, such as 1 representing the priority drawdown stage and 2 representing the follower stage; it can also be a 0-1 type variable describing the order relationship between reservoirs. The purpose of introducing this variable is to couple discrete order logic with continuous hydraulic variables within the same model framework. Transforming the undrawdown risk distribution parameter data into risk constraints means adding restrictive inequalities to the model, such as requiring that the average over-limit loss or extreme over-limit probability of the optimized trajectory under all scenarios must not exceed the threshold calculated based on the aforementioned parameters. For example, in the specific solution process, since this model mixes integer and continuous variables and faces the challenge of large-scale computation under multiple scenarios, a strategy combining decomposition and iteration is usually adopted. For example, the outer algorithm is responsible for searching different drawdown order combinations, while the inner algorithm, under the premise of a fixed order, solves for the optimal water level and output trajectory under each scenario. Through repeated interaction between the inner and outer layers, the optimal solution that satisfies strict risk constraints and maximizes power generation benefits is found. The final optimized drawdown sequence data not only includes the optimal drawdown order scheme, but also a set of detailed operational trajectories of the scheme under different hydrological scenarios, providing rich data samples for subsequent rule extraction.

[0028] Pattern recognition is performed on the drawdown sequence optimization results data to extract the drawdown sequence patterns and generate corresponding pre-flood drawdown scheme rule base data.

[0029] In other words, pattern recognition is performed on the data of the drawdown sequence optimization to extract representative drawdown sequence patterns, and a rule base data for pre-flood drawdown schemes is constructed based on the typical operating trajectories under each pattern.

[0030] In this embodiment, to transform the complex optimization results into easily executable operational guidelines for dispatchers, pattern recognition and rule-based processing are required. Specifically, features can be extracted from a large number of drawdown sequence samples obtained from the optimization, constructing feature vectors containing temporal and performance characteristics. Clustering algorithms are then used to classify these vectors into several typical drawdown sequence patterns, such as an upstream-priority rapid drawdown pattern or a full-cascade synchronous slow drawdown pattern. For each pattern, by statistically analyzing the distribution characteristics of its corresponding multiple scenario trajectories at various times, water level control zones and power output control zones with upper and lower boundaries are extracted. These control zones and their corresponding regulation logic are encapsulated into a pre-flood drawdown scheme rule base data, forming a mapping from black-box optimization solutions to white-box scheduling rules.

[0031] In one possible implementation, a multi-scenario hydrological operation input set is generated, including:

[0032] The current water storage index is extracted from the pre-flood state parameters, and the forecast error disturbance weights used to correct the random error distribution are determined based on the current water storage index.

[0033] In this embodiment, the current water storage index I is considered to be either too high or too low. wet It is a dimensionless index used to quantitatively describe the current water storage status of a cascade reservoir group relative to historical levels for the same period. The specific calculation method can employ a standardized anomaly formula, which uses the current total water storage V of the cascade reservoirs. curr Subtract the historical average water storage V hist_avg Then divide by the standard deviation σ of the historical water storage for the same period. hist The formula is expressed as:

[0034] I wet = (V curr - V hist_avg ) / σ hist ;

[0035] The current water storage index, with a positive value indicating abundant water storage and a negative value indicating scarce water storage, can be used to determine the forecast error disturbance weight, denoted as α. To more accurately reflect the bias of forecast uncertainty under different hydrological years, asymmetric adjustments to the random error distribution are needed. Specifically, when the current water storage index I... wet A value greater than zero indicates a higher flood risk, requiring close attention to the risk of an underestimation of the forecast, i.e., an overestimation of the actual inflow. Therefore, α can be set to a value greater than 1 in the positive error range, such as 1.2 to 1.5, and a value of 1 or less than 1 in the negative error range; conversely, when the current water storage is above or below the threshold of α, the risk is lower. wet When α is less than zero, it takes a value greater than 1 in the negative error interval. This makes the subsequently generated random error sequence more statistically inclined to expose the direction of risk of greatest concern in the current state.

[0036] Based on the pre-stored bias-corrected forecast inflow data, random disturbances adjusted by forecast error perturbation weights are superimposed to generate a set of conditional scenario inflow sequences that conform to the characteristics of the current hydrological state.

[0037] Specifically, the forecast inflow data Q is obtained after calculation using a conventional meteorological and hydrological model and correction for historical biases. forecast As a deterministic baseline sequence, the original random error sequence ε is generated using Monte Carlo simulation or Latin hypercube sampling. This sequence typically follows a standard normal distribution or a t-distribution with a mean of 0. The original error sequence is then transformed by applying a forecast error perturbation weight α to obtain the adjusted error sequence ε. adj =α*ε. The adjusted error sequence is then superimposed onto the baseline forecast sequence, i.e.:

[0038] Q scenario = Q forecast + ε adj ;

[0039] Generate conditional inflow sequence Q scenario Repeating the above process N times, for example, with N ranging from 100 to 1000, yields a set of conditional scenario inflow sequences containing N sequences. This set retains both the predicted trend information and incorporates asymmetric uncertainties related to the current water storage status in its statistical characteristics, forming the core hydrological component of the multi-scenario hydrological operation input set.

[0040] By combining the inflow sequence set under various conditions with the corresponding unit operation constraints, a structured multi-scenario hydrological operation input set is constructed. This serves as the data foundation for subsequent risk assessment and joint optimization.

[0041] Specifically, the system reads a set of inflow sequences for multiple conditional scenarios, covering the entire pre-flood season scheduling period. Simultaneously, it retrieves unit operation constraint parameters from the power plant equipment database. These parameters include the upper and lower limits of each unit's output, vibration zone flow thresholds, ramp rate limits, and maintenance schedules. Data combination operations are performed to associate and bind each scenario inflow sequence with the unit operation constraint parameters. Structurally, an independent data object is created for each scenario. This object encapsulates not only the inflow and rainfall time series for that scenario but also the corresponding time-varying constraints, such as the maintenance status of a unit on a specific date. The previously scattered hydrological and equipment data are integrated into a structured multi-scenario hydrological operation input set. This input set uses a unified index and storage format, serving as the standard data interface for subsequent undrawn flow risk assessment and joint optimization model solving, enabling the model to simultaneously acquire hydraulic and equipment boundaries when reading different scenarios.

[0042] In one exemplary embodiment, quantifying the distribution parameter data of unsettled risk includes:

[0043] By using a pre-set empirical scheduling strategy, each scenario in the multi-scenario hydrological operation input set is simulated to obtain initial water level process data reflecting the baseline scheduling state.

[0044] In this embodiment, a reference benchmark needs to be established to calculate risk parameters. A preset empirical scheduling strategy, such as an equal-capacity drawdown rule, can be adopted. This rule requires that the capacity utilization rates of each reservoir in the cascade system remain synchronized during the drawdown process, where the capacity utilization rate is the ratio of the current capacity to the regulating capacity. Specifically, for the j-th scenario, at each time step t, the outflow that ensures equal capacity utilization rates for all reservoirs is calculated, and the reservoir water level is updated accordingly. By executing this simulation process for all N scenarios, N water level change trajectories reflecting the conventional scheduling mode can be obtained, i.e., initial water level process data, representing the potential water level evolution of the reservoir group without optimization intervention.

[0045] For each scenario, key time points before the flood season are selected, and the difference between the initial water level process data and the preset safety control water level is calculated. The part that is higher than the safety control water level is recorded as the unreceded excess height.

[0046] Specifically, for each reservoir i and each scenario j, examine its critical pre-flood time point T. key For example, the initial water level Z on the last day before the start of the flood season. init_i，j (T key The preset safety control water level Z safe_i This refers to the flood control limit level or the target water level after reserving flood control storage capacity. Calculate the difference between the two and define the unreceded excess height L. i，j For: When Z init_i，j (T key () greater than Z safe_i At that time, L i，j It equals the difference; otherwise, L i，j It equals 0. That is:

[0047] L i，j = max(0, Z) init_i，j (T key ) - Z safe_i ).

[0048] The unreleased water level exceeding the limit quantifies the potential degree of violation or flood risk exposure caused by the failure to release the water level below the safety line on time in a specific scenario.

[0049] Based on the unreceded height exceeding the limit under all scenarios, the probability of successful unreceding and the conditional risk value at a preset confidence level are statistically analyzed, and the statistical results are combined into unreceded risk distribution parameter data.

[0050] In this embodiment, based on the set of unresolved over-limit height samples {L} under N scenarios i，1 L i，2 , ..., L i，NThis allows us to calculate key parameters describing the risk distribution. Specifically, by counting the percentage of samples with values ​​greater than 0, we obtain the probability P of successful risk elimination. fail_i To measure the extreme risk at the tail, the conditional value of risk (CVaR) is calculated. For example: determine a confidence level β, such as 95%, and find the β quantile VaR in the sample set. β Calculate all VaR values ​​greater than or equal to β. β The arithmetic mean of the samples exceeding the limit height is used to obtain the conditional value of risk (CVaR) of the i-th reservoir at confidence level β. β_i This reflects the average extent of over-limit loss under the most unfavorable scenarios, such as the worst-case 5% scenario. The probability of unsuccessful elimination, P... fail_i Conditional Value at Risk (CVaR) β_i The indicators are organized by reservoir and time period, thus forming structured undissipated risk distribution parameter data. These parameters will serve as constraints in the subsequent optimization model, ensuring that the generated optimization scheme has clearly controllable risks. For example, the constraint boundary can be the required conditional value of risk (CVaR). β_i Less than a certain threshold.

[0051] like Figure 2 As shown, according to one aspect of this application, a pre-flood joint drawdown optimization model is constructed by introducing drawdown order decision variables, including:

[0052] Construct a data structure for drawdown order decision variables. The data structure includes an integer stage number variable for identifying the drawdown stage of a reservoir, or a 0-1 type order relationship variable for identifying the order of drawdown between reservoirs.

[0053] In this embodiment, a decision variable for the order of elimination is introduced to express the logic of which stage disappears first in the mathematical model. As a preferred implementation, an integer stage number variable y can be defined. i , where i represents the reservoir number. Variable y i The value range is {1, 2, ..., K}, where K is the preset total number of drawdown stages, for example, K=3. A value of 1 represents the reservoir belonging to the priority drawdown stage, a value of 2 represents the coordinated drawdown stage, and a value of 3 represents the following drawdown stage. Different values ​​correspond to different physical control strategies. As another optional implementation, a 0-1 type ordinal relation variable x can also be defined. ij When reservoir i enters the main drawdown period before reservoir j, x ij The value is 1 if the variable is selected, and 0 otherwise. These two methods of defining variables are mathematically equivalent. Both aim to parameterize discrete combinations of orders, making them a decision space that the optimization model can directly search, rather than a fixed input condition as in traditional methods.

[0054] Based on the data structure, a comprehensive objective function is constructed with the goals of maximizing total cascade power generation and minimizing unreduced risk. This comprehensive objective function is composed of a weighted average of expected power generation under multiple scenarios and an unreduced risk penalty term. The unreduced risk penalty term is constructed based on the conditional value-of-risk index from the unreduced risk distribution parameter data. The unreduced risk penalty term is used to quantify tail-end over-limit losses at the confidence level.

[0055] Specifically, the objective function J of the optimization model can balance economic benefits and safety risks. Its specific form is to maximize J:

[0056] J = ∑(w j *E j ) -λ* ∑(CVaR β_i (L));

[0057] Among them, the first term ∑(w) j * E j E represents the total power generation of all cascades under all scenarios j. j The weighted expected value, w j Let λ be the probability weight of scenario j. The second term is λ*∑(CVaR). β_i (L) represents the penalty for unresolved risk, CVaR β_i (L) is the conditional risk value for reservoir i regarding the undrawdown height L. In the optimization model, L is a function of the decision variables, including water level and flow rate; λ is the risk penalty coefficient, used to adjust the degree of risk aversion. By introducing this penalty term, the model actively avoids drawdown paths that, although generating high amounts of electricity, lead to a surge in CVaR (Conductivity, Capacity, and Rate of Return), i.e., those with high tail risk, during the optimization process, thus internalizing the principle of safety first at the mathematical level.

[0058] like Figure 3 As shown, in a further embodiment, a coupling relationship is established between the drawdown order decision variable and the pre-flood drawdown trajectory data, including:

[0059] We construct order relation constraints for the drawdown order decision variables to ensure that the drawdown order among cascade reservoirs satisfies logical transitivity and that there are no illegal stage jumps.

[0060] In this embodiment, to ensure the physical rationality of the elimination order, order relation constraints need to be applied. For the integer stage number variable y... i The following constraints need to be constructed: Spatial topological constraints. If reservoir A is located upstream of reservoir B and the hydraulic connection is close, it is generally required that the upstream drawdown is no later than the downstream drawdown to avoid the superposition of downstream flood control pressure, i.e., y A ≤y BPhase capacity constraints: To avoid all reservoirs crowding into the same drawdown phase and causing excessive pressure on the power grid's peak shaving, the number of reservoirs in each phase can be limited, for example, ∑(bool(y i == k))≤N max_k Where bool is the Boolean function, k is the dropout stage number, and N is the number of the dropout stage. max_k This represents the maximum allowed number of reservoirs in the k-th drawdown phase. Furthermore, logical constraints can be added to ensure that the phase numbers are consecutive and do not jump; for example, reservoirs may exist in phases 1 and 3, but phase 2 may be empty. These constraints define the feasible region of the drawdown order, excluding combinations of orders that are not engineering-feasible.

[0061] Based on order relation constraints, differentiated physical boundary parameters are set for cascade reservoirs with different drawdown priorities. The physical boundary parameters are used to limit the upper limit of discharge flow and the rate of water level decline in the pre-flood drawdown operation trajectory data, so that reservoirs that enter the drawdown stage first have higher discharge authority than reservoirs in the follow-up stage.

[0062] In this embodiment, the discrete variable y is implemented. i Physical coupling of continuous variables, including flow rate Q out_i (t), water level Z i (t). Specifically, according to y i The value of is used to divide the pre-flood time axis into corresponding control periods. For the priority drawdown phase, i.e., y... i For a reservoir with a discharge rate of 1, the upper limit of its discharge flow Q is released within its main drawdown time window. max_i (t) is set to a larger value, such as the unit's full-capacity flow rate or even including some of its water discharge capacity, allowing it to drop at full speed; at the same time, its water level drop rate constraint is relaxed. Conversely, for the stage of following the drop, i.e., y i For reservoirs with a drawdown rate of 3, within the same time window, the upper limit of their outflow is limited to a smaller value, such as only meeting the ecological base flow and minimum output, or a strict limit is imposed on their water level drop rate, such as a drop of no more than 0.5 meters per day, forcing them to maintain a high water level while awaiting subsequent instructions. The drawdown order decision variable directly changes the boundary shape of the hydraulic constraints, substantially controlling the drawdown rate of each reservoir.

[0063] like Figure 4 As shown, in one optional embodiment, obtaining the dropout sequence optimization result data includes:

[0064] In the outer layer of the pre-flood joint drawdown optimization model, a sequence search algorithm is used to perform a combined search on the drawdown order decision variables to generate a set of candidate drawdown orders. In the inner layer of the pre-flood joint drawdown optimization model, for each candidate drawdown order in the candidate drawdown order set, the order relationship between reservoirs is fixed, and for each scenario in the multi-scenario hydrological operation input set, the pre-flood drawdown operation trajectory data that satisfies water balance is calculated. The comprehensive objective function value under each scenario is calculated and the risk constraints are verified. The candidate drawdown orders that satisfy the risk constraints and their pre-flood drawdown operation trajectory data that satisfy water balance are summarized into drawdown sequence optimization result data.

[0065] Alternatively, it can be said that the comprehensive objective function value under each scenario is calculated and the risk constraints are verified, the candidate order that does not meet the risk constraints is eliminated, and the order that meets the conditions and its trajectory data are summarized into the result data of the dropout sequence optimization.

[0066] For example, given the complexity of the model, a two-level decomposition algorithm can be used to solve it. In the outer layer, a genetic algorithm, particle swarm optimization algorithm, or sequence search heuristic algorithm is used to search for y within the feasible region that satisfies the order relation constraints. i The combination of these factors is used. For example, a candidate solution vector Y = [1, 2, 2, 3] is generated, representing the order of the four reservoirs. This vector is then passed to the inner layer model. In the inner layer, since the order variable Y is fixed, the model degenerates into a standard multi-scenario reservoir scheduling problem with specific physical boundary parameters. At this point, dynamic programming (DP), successive approximation dynamic programming (DPSA), or linear programming (LP) can be used to efficiently solve the optimal water level process Z under each scenario. i，j (t) and the output process P i，j (t) represents the pre-flood drawdown trajectory data. After the inner layer solution is completed, the corresponding objective function value J and risk index are calculated, and this information is fed back to the outer layer algorithm to evaluate the fitness of the candidate solution Y, guiding the outer layer to generate a better next-generation candidate order until the convergence condition is met. The final output drawdown sequence optimization result data is the optimal order Y* and its corresponding set of all scenario trajectories.

[0067] In a preferred implementation, a two-layer nested decomposition and iteration strategy is employed. In the outer layer of the model, the system primarily handles discrete order variables. Sequence search algorithms, such as genetic algorithms or discrete particle swarm optimization, can be used to perform combined searches on the drawdown order decision variables within the feasible region. Each search iteration generates one or more sets of candidate drawdown orders composed of different reservoir priority rankings. For each candidate drawdown order in the set, the system uses it as a fixed parameter, thereby locking in the drawdown order relationship between the cascade reservoirs, i.e., the reservoir priority relationship. Based on this, the system traverses each scenario in the multi-scenario hydrological operation input set. Under the premise of satisfying the water balance equation, reservoir capacity curve constraints, and unit output characteristics, dynamic programming or linear programming algorithms are used to calculate the optimal water level and flow process under that scenario, i.e., generating pre-flood drawdown operation trajectory data. After completing the inner-layer calculation, the calculation results of each scenario are collected, the comprehensive objective function value is calculated, and it is strictly verified whether the preset risk constraints are met; the comprehensive objective function value includes power generation benefits and risk penalties, and the risk constraint can be that the conditional risk value is less than a threshold. For sequences that fail the risk verification, a penalty value is imposed or they are directly eliminated. Candidate drawdown sequences that meet both the risk constraints and the physical constraints such as water balance in the inner-layer calculations are summarized and stored together with their corresponding pre-flood drawdown trajectory data as drawdown sequence optimization result data.

[0068] This embodiment achieves deep coupling optimization of sequential logic and physical processes by constructing a mixed integer nonlinear programming model and solving it using a decomposition algorithm.

[0069] In one embodiment of this application, a dropout sequence pattern is provided, including:

[0070] Information describing the order of drawdown between reservoirs is extracted from the drawdown sequence optimization results data as drawdown time series feature components, and information describing the total power generation of the cascade and the risk indicators of non-drawdown is extracted as risk benefit feature components. The drawdown time series feature components and risk benefit feature components are concatenated to construct drawdown sequence feature vector data.

[0071] In other words, drawdown sequence feature vector data is constructed, which includes drawdown time series feature components describing the order of drawdown between reservoirs, and risk-benefit feature components describing the risk indicators of non-drawdown and the total power generation of the cascade.

[0072] In this embodiment, to classify the large number of dropout schemes generated by optimization, it is necessary to define a mathematical vector that can comprehensively characterize the features of the schemes. Specifically, for each dropout sequence sample k obtained by optimization, a high-dimensional feature vector V is constructed. kThe vector consists of two parts: the first part is the drawdown time-series characteristic component, specifically including the start time T of each reservoir entering the main drawdown phase. start_i Duration of fall D draw_i and the time difference between warehouse start-up Δ T_ij For example, for a cascade system containing two reservoirs, the time-series components can be represented as [T start_up T start_down T start_up -T start_down This reflects the temporal structure of which series declines first, where T... start_up T represents the start time of the main drawdown phase of the upstream reservoir. start_down This refers to the start time of the main drawdown phase of the downstream reservoir. The second part is the risk-benefit characteristic component, specifically including the average total cascade power generation E under multiple scenarios. avg_k The probability P of not successfully eliminating the target fail_k Conditional Value at Risk (CVaR) k These two parts are then concatenated and standardized to eliminate the influence of different physical dimensions, forming the final dropout sequence feature vector data used for clustering.

[0073] Clustering algorithms are used to group the feature vector data of the dropout sequence. Each cluster obtained by grouping is defined as a dropout sequence pattern, and a mapping relationship between each dropout sequence pattern and the corresponding optimized solution set is established.

[0074] Specifically, after obtaining the standardized feature vector set, an unsupervised clustering algorithm is used to extract typical patterns. As a preferred implementation, the K-means clustering algorithm is employed. To determine the optimal number of clusters K, the silhouette coefficient method or the inflection point method based on the sum of squared deviations within each cluster can be used. For example, the average silhouette coefficient is calculated when K ranges from 2 to 8, and the K corresponding to the maximum coefficient is selected as the optimal number of clusters. Assume the clustering results divide the samples into three categories: the first category shows that the upstream reservoir's start-up time is significantly earlier than the downstream, labeled as the upstream-priority drawdown pattern; the second category shows that the start-up times of the upstream and downstream reservoirs are similar, labeled as the synchronous drawdown pattern; and the third category shows that the downstream reservoir drains before the upstream, labeled as the downstream-priority emptying pattern. Each cluster not only contains a center vector but also associates all the original optimized trajectory samples belonging to that cluster.

[0075] In some alternative implementations, hierarchical clustering or density-based clustering algorithms can also be used to accommodate non-convex sample distributions. For the clustered patterns, their frequency of occurrence can be further calculated, and sparse patterns with frequencies below a certain threshold can be removed, retaining the statistically representative mainstream patterns.

[0076] In a further embodiment, generating pre-flood drawdown scheme rule base data includes:

[0077] For each drawdown sequence pattern, the pre-flood drawdown trajectory data corresponding to that pattern are collected and aligned on a unified time axis. The median, upper quantile, and lower quantile of the water level are calculated at each time step of the aligned pre-flood drawdown trajectory data. The target water level curve is constructed using the median water level, and the water level control zone enveloping the target water level curve is constructed using the upper and lower quantiles. The target water level curve and the water level control zone are encapsulated as normal drawdown control rule data and stored in the pre-flood drawdown scheme rule base data.

[0078] In this embodiment, for each drawdown sequence pattern identified by clustering, all samples belonging to that pattern label are extracted from the optimization result database and aggregated to form a dataset of pre-flood drawdown trajectory data corresponding to that drawdown sequence pattern. Since the drawdown start time may vary slightly under different scenarios, to extract common patterns, the system aligns these trajectories on a unified time axis, typically using days or hours as the scale. After alignment, the system performs statistical analysis on the distribution characteristics of the dataset at each time step. Specifically, for any time T, the system obtains the water level values ​​of all trajectories at that time, forming a sample sequence. This sample sequence is sorted, and the value in the middle position is identified as the median water level, representing the typical operating path under that pattern. Simultaneously, the value at the higher percentile of the sequence is identified as the upper quantile, and the value at the lower percentile is identified as the lower quantile. By calculating these three statistics point by point along the entire time axis, a center curve and a control band enveloped by the upper and lower quantiles are constructed, completing the transformation from a discrete trajectory set to standardized control rules.

[0079] For example, multiple specific trajectories under each pattern are abstracted into control strips that provide guidance. Specifically, for the m-th pattern, all N contained within it are extracted. m N scenario trajectories. For each time step t, calculate these N... m The statistical distribution values ​​of each water level sample are used. The median value is selected and connected to form the target water level curve Z. target_m (t) represents the recommended operating path under this mode. The upper quantile, such as the 95th quantile, is selected as the upper limit Z of the water level control zone. upper_m (t), select the lower quantile, such as the 5th quantile as the lower limit Z of the water level control zone. lower_m (t). For example, the constructed normal drawdown control rule data specifies that, under normal operating conditions, the measured water level should be maintained at [Z]. lower_m (t), Z upper_mWithin the interval [t], if the water level deviates from the target curve but remains within the control zone, only minor adjustments to the unit output are needed; if it approaches the upper or lower boundaries, a larger-scale flow adjustment is required. Similarly, the same statistical analysis can be performed on the outflow and unit output trajectories to generate corresponding flow control zones and output control zones. Compared to the traditional single average line method, this embodiment can more robustly accommodate fluctuations caused by hydrological uncertainties, giving scheduling operations a certain degree of flexibility.

[0080] In a further embodiment, generating the pre-flood drawdown scheme rule base data also includes:

[0081] Adjust the risk penalty coefficient in the pre-flood joint drawdown optimization model, or relax the threshold parameters in the risk constraints; re-solve the pre-flood joint drawdown optimization model under the adjusted or relaxed parameter conditions to obtain the set of emergency optimization trajectories under specific working conditions; extract the emergency water level control zone based on the set of emergency optimization trajectories, generate emergency drawdown control rule data containing triggering conditions and recovery logic, and supplement it to the pre-flood drawdown scheme rule base data.

[0082] Alternatively, it can be said that the construction of emergency or delayed drawdown control rules includes adjusting and optimizing model parameters and resolving them, as well as refining emergency water level control zones.

[0083] In this embodiment, to cope with extreme abnormal situations, such as sudden changes in forecasts or rapid shifts in flood control, the rule base needs to include emergency plans. The specific construction process is as follows: Return to the joint optimization model and manually adjust the boundary conditions or parameters. For example, to generate emergency drawdown rules, the risk penalty coefficient λ can be lowered, or the upper limit of the constraint on the probability of unsuccessful drawdown can be appropriately relaxed, for example, from 1% to 5%, and the weight of power generation in the objective function can be increased. Under the new parameter settings, the optimization solution is rerun to obtain a set of emergency optimization trajectories that tend towards aggressive power generation or rapid reservoir emptying. The statistical method is repeated to refine the emergency trajectories and generate emergency water level control zones and emergency flow control zones. It is understood that the emergency control zone will be wider than the normal control zone, or its centerline will have a steeper drawdown slope. In the generated emergency drawdown control rule data, the triggering conditions and recovery logic also need to be clearly defined. For example, the triggering condition is defined as the actual inflow exceeding the forecast value by 50% for three consecutive days or the real-time water level exceeding the upper limit of the normal control zone for 24 hours. The recovery logic is defined as the risk indicator returning to below the safe threshold and remaining there for 48 hours. A closed-loop rule system of normal operation, emergency response, and recovery is established, so that the system has a basis to follow under any operating condition.

[0084] As an optional implementation, it also includes:

[0085] Based on the pre-flood state parameter set, the target drawdown order scheme data and the target pre-flood drawdown operation scheme data are matched from the pre-flood drawdown scheme rule base data.

[0086] In this embodiment, the transformation from offline rules to online decision-making is realized. During actual operation, the pre-flood state parameters monitored in real time are continuously updated over time. These parameters include the latest inflow accumulation and changes in weather forecasts. The system matches the current real-time state feature vector with the applicable conditions of each pattern in the rule base, calculates the similarity, and selects the drawdown sequence pattern most suitable for the current operating conditions. Based on the matched pattern, the corresponding drawdown order scheme and water level control zone are directly retrieved to form target drawdown order scheme data and target pre-flood drawdown operation scheme data, which serve as specific instructions to guide the current scheduling cycle. This ensures that the scheduling scheme can dynamically adapt to constantly changing hydrological and meteorological conditions, maintaining an optimal balance between safety and efficiency.

[0087] In a preferred implementation, matching the target drawdown sequence scheme data and the target pre-flood drawdown operation scheme data can also be achieved by: constructing a real-time state feature vector using pre-stored pre-flood real-time operation data and forecast information, and matching the target drawdown sequence scheme data and the target pre-flood drawdown operation scheme data from the pre-flood drawdown scheme rule base data.

[0088] Specifically, the system collects the latest real-time data every day or every period of the actual scheduling cycle. This includes: the current actual water level of each reservoir, the cumulative inflow deviation, the rolling rainfall forecast for the next 10-15 days, and the latest load demand curve released by the power grid. Based on this data, the pre-flood state parameters are updated and calculated, and a real-time state feature vector S is constructed. real The dimension of this vector should be consistent with the feature vectors of pattern applicability conditions stored in the rule base, and typically includes components such as reservoir capacity utilization, inflow trend abundance / sparseness, and forecast error dispersion. Calculate the real-time state feature vector S. real The typical feature vector S of each drop-off sequence pattern in the rule base mode_k The similarity between them. For example, a weighted Euclidean distance formula is used:

[0089] D k = sqrt(∑(w j * (S real_j - S mode_k_j ) 2 ));

[0090] Where w j The weight of the j-th feature component can be preset based on the degree of influence of this feature on the reduction decision; S real_j S represents the component value of the real-time state feature vector in the j-th dimension. mode_k_j Let be the component value of the feature vector of the k-th disappearing sequence pattern in the rule base in the j-th dimension. The weighted Euclidean distance D is selected. kThe pattern with the lowest similarity is selected as the current recommended pattern. The system directly retrieves the corresponding drawdown sequence plan and the target water level range for the day to form preliminary target drawdown sequence plan data and target pre-flood drawdown operation plan data.

[0091] The feasibility of the target drawdown sequence scheme data and the target pre-flood drawdown operation scheme data is verified using pre-stored real-time pre-flood operation data, and the scheme feasibility verification results are generated.

[0092] Alternatively, it can be said that the feasibility of the matched schemes is verified by using real-time pre-flood operation data, and the feasibility verification results of the schemes are generated.

[0093] Specifically, the preliminary plan is generated based on historical statistical patterns and needs to be rigorously verified under the current physical conditions. The verification includes two aspects: first, physical capacity verification, checking whether the required discharge flow exceeds the maximum capacity of the flood discharge facilities, or whether the required rate of water level decline may induce reservoir bank landslide risks. Second, risk compliance verification, rapidly simulating the preliminary plan using the latest multi-scenario hydrological inputs to calculate its uncollected risk indicators, such as CVaR, at future critical nodes. If the simulation results show that implementing the preliminary plan at the current actual water level would cause future risk indicators to exceed preset safety thresholds, such as CVaR > 0.5 meters, or that physical constraints would be violated, a feasibility verification result for the plan that fails verification is generated, and the specific violation period and violation indicators are marked. Conversely, if the verification passes, the preliminary plan is directly confirmed as the final directive.

[0094] When the feasibility verification results indicate that there is an excessive risk of uncontrolled drawdown or that the discharge capacity exceeds the limit, the emergency drawdown control rules pre-stored in the pre-flood drawdown scheme rule base data are called to make local corrections to the water level control zone in the target pre-flood drawdown operation scheme data, and generate an executable target scheme.

[0095] In other words, when the verification results indicate that there is an excessive risk or a failure to meet the capability limit, the emergency drawdown control rules are invoked to make partial corrections to the plan and generate the final executable target plan.

[0096] In this embodiment, if the verification fails, the system enters a correction process. The correction strategy employs a tiered processing mechanism. For minor deviations, such as only a slight exceedance of physical constraints, the normal control zone boundary in the rule base is used for truncation correction, forcibly limiting the outflow within the capacity range, and recalculating the subsequent water level process. For severe deviations or excessive risks, such as a high probability that the water level will not reach the target level in the future, the emergency drawdown control rules in the rule base are triggered. Specifically, the system switches the target orientation of the current trajectory from the normal target curve to the lower boundary of the emergency control zone, i.e., accelerating the drawdown path, or directly activating the backup emergency drawdown sequence, such as initiating the drawdown of the downstream reservoir in advance. The correction process is iterative: after generating a new corrected trajectory using emergency rules, the verification process is executed again until a solution that satisfies both physical constraints and keeps the risk within an acceptable range is found. The corrected drawdown sequence and water level / flow control instructions for the next few days are output as executable target solutions and sent to the power plant control system, realizing a safe closed loop from offline strategy to online instructions.

[0097] Although the foregoing embodiments focus on the method flow and algorithm logic, those skilled in the art should understand that the method for constructing the pre-flood joint drawdown sequence and drawdown scheme of cascade reservoirs can be executed by computer program instructions in a specific electronic device or computer system.

[0098] In one specific hardware implementation, the construction method runs on a high-performance computing device, which includes at least one processor, memory, communication interfaces, and a bus system connecting these components. The processor may be a central processing unit (CPU), a graphics processing unit (GPU), a field-programmable gate array (FPGA), or a combination thereof. When performing joint optimization model solving, due to the involvement of large-scale multi-scenario simulation and mixed-integer programming, a processor or computing cluster configured with multi-core parallel computing capabilities is preferably used.

[0099] In this embodiment, the memory is used to store computer program instructions and various types of data generated during processing. Specifically, the memory is divided into different logical areas. The program storage area stores the instruction code for functional modules such as data acquisition, scenario construction, joint optimization, rule recognition, and online verification. The data storage area is used to maintain the key datasets mentioned in the previous embodiments, including but not limited to: historical hydrological databases, pre-flood basic datasets, multi-scenario hydrological operation input sets, undrawdown risk distribution parameter data, and generated pre-flood drawdown scheme rule base data.

[0100] When the computer program instructions are loaded and executed by the processor, the electronic device performs the following operations: acquires real-time and forecast data from external SCADA systems and meteorological service interfaces through communication interfaces; cleans the data and extracts state features using the processor's built-in mathematical library; calls optimization solvers, such as those based on CPLEX, Gurobi kernels, or custom heuristic algorithm libraries, to solve the joint optimization model loaded into memory; and displays the generated drawdown sequence patterns and control rules to dispatchers through a visualization interface, or directly issues them to the reservoir's on-site control unit through an instruction interface.

[0101] According to another aspect of this application, a computer-readable storage medium, such as a hard disk, optical disk, flash memory, or cloud object storage, is also provided. A computer program is non-transitory stored on this medium. When the program is read and executed by a computer, it can implement the methods described in any of the above embodiments. For example, the program includes a scenario generation module for performing Monte Carlo simulations, a risk assessment module for calculating the Conditional Value at Risk (CVaR), and a pattern recognition module for performing K-means clustering.

[0102] As a preferred system architecture variant, this computing system can adopt a cloud-edge-device collaborative architecture. The demanding multi-scenario joint optimization and pattern recognition tasks are deployed on high-performance servers in the cloud or scheduling center, utilizing powerful computing capabilities for offline or near real-time policy updates. Meanwhile, online matching and verification tasks are deployed on edge computing nodes or scheduling workstations to ensure millisecond- or second-level response capabilities to real-time water condition changes. This approach guarantees both the depth and accuracy of the algorithm while also meeting the real-time requirements of practical engineering applications.

[0103] In a detailed embodiment, reservoirs A (upstream) and B (downstream), which have an upstream-downstream relationship, are selected as examples to demonstrate the entire process from state identification to scheme generation. Specifically, it is assumed that the current time is May 1st, before the flood season. The historical average total water storage of the two reservoirs during the same period is V. hist_avg The volume is 5 billion cubic meters, and the standard deviation σ is 5 billion cubic meters. hist The current total water storage capacity is 500 million cubic meters. curr The total is 6 billion cubic meters. Calculate the current water storage index (I) for both periods of high and low water levels. wet = (60 - 50) / 5 = 2.0. This value is greater than 0, indicating that the current flood situation is severely above average, and the flood control pressure is high. Based on I wet = 2.0, determine the forecast error perturbation weight α. The basic rule is set as follows: when I wet When the value is greater than 1.0, the positive error, i.e., the weight α of the larger inflow, is... plus = 1.0 + 0.2 * I wet Calculation yields α plus= 1.4 means that when generating scenarios, the portion of the forecast error distribution greater than 0 will be stretched by a factor of 1.4. Assuming the baseline forecast inflow is 1000 m³ / s, and the original random error of a certain sampling is +100 m³ / s, after weight adjustment it becomes +140 m³ / s, resulting in an inflow of 1140 m³ / s for this scenario. Through 1000 samplings, a multi-scenario input set for enhancing the risk of high inflow is generated.

[0104] Using empirical rules, such as equal utilization decline deduction, in scenario S1, the critical moment T... key The projected water level of Reservoir A is 152 meters. The flood control limit water level, i.e., the safe water level Z, of Reservoir A is also known. safe The value is 150 meters. Therefore, the unreceded over-limit height L in this scenario is... A_S1 = max(0, 152 - 150) = 2 meters. After analyzing all 1000 scenarios, 200 scenarios exceeded the limit. The tail mean of the 95th quantile of these 200 scenarios was used to calculate the Conditional Value at Risk (CVaR). 95 = 3.5 meters. This value is used as the upper limit of risk constraints for subsequent optimization, such as requiring the optimized CVaR to be... 95 ≤1.0 meter.

[0105] Construct an optimization model and introduce the order variable y. A y B The outer algorithm searches for the candidate order Y=[y A =1, y B =2], meaning upstream drawdown is prioritized. Under this sequence, the inner model relaxes the upper limit of reservoir A's discharge in the first stage to full capacity, while simultaneously restricting the discharge of reservoir B. Calculation results show that by allowing reservoir A to drain at full speed ahead of time, flood control capacity is freed up. Although the water level in reservoir B drops more slowly in the early stages, reservoir A already has the capacity to hold back large inflows later, avoiding a cumulative effect. Ultimately, the CVaR under this scheme is... 95 The water level was lowered to 0.8 meters, meeting the constraint of less than 1.0 meters, and the total power generation increased by 5% due to the high head operation.

[0106] Statistical analysis was performed on the preferred trajectory set belonging to the upstream priority drawdown mode. At the time of May 10th, the median water level of all trajectories was 148 meters, the 95th percentile was 149 meters, and the 5th percentile was 147 meters. Therefore, the normal drawdown control zone for this time was generated as [147 meters, 149 meters], with a target water level of 148 meters. On May 15th, real-time monitoring revealed that the water level of Reservoir A reached 149.2 meters, exceeding the upper limit of the control zone (149 meters), but not reaching the flood control high water level. At this point, the system judged it as a slight deviation and did not trigger lifecycle adjustments. Instead, it issued a correction command to increase the unit output by 10% to attempt to pull the water level back within the control zone. If the water level continues to rise above 150 meters, the system will trigger the emergency drawdown rule, forcibly switching to the maximum flood discharge mode.

[0107] In one possible implementation, the pre-flood basic dataset is obtained, including:

[0108] By integrating meteorological and hydrological forecast data, a more accurate forecast inflow is generated using deviation correction methods, and forecast error statistical parameters are calculated.

[0109] In this embodiment, to ensure the baseline accuracy of the scenario construction, the original meteorological and hydrological forecast products are not used directly. Instead, the output data from the meteorological forecast system and the watershed hydrological forecast system are accessed. A method combining historical verification and real-time correction is used to process the initial forecast inflow sequence. Specifically, a continuously updated historical forecast and actual data comparison database is established to analyze systematic deviations at different scales, such as during dry seasons, flood seasons, and flood seasons. Statistical regression models or machine learning models are used to learn the deviation patterns and correct the current initial forecast inflow sequence to obtain the deviation-corrected forecast inflow data. Simultaneously, the error distribution characteristics of the forecast in the same historical period are calculated, and statistical measures such as variance and skewness are extracted to form forecast error statistical parameter data.

[0110] Based on the statistical parameters of forecast errors, time alignment and spatial allocation are performed on multi-source data to generate a spatiotemporally consistent pre-flood basic dataset.

[0111] Specifically, to address the issues of inconsistent time steps and spatial scale mismatches among multi-source data, strict spatiotemporal alignment operations are performed. Linear interpolation or spline interpolation methods are used to unify all time series to a standard scheduling step size. For spatial data, the Thiessen polygon method or grid mapping method is used to accurately distribute meteorological forecast data across the watershed to the control sections of each cascade reservoir. After consistency verification, a structured pre-flood base dataset is formed, ensuring the physical and logical consistency of the model input. The consistency verification includes checking the upstream and downstream water balance.

[0112] In one alternative implementation, the optimized result data of the dropout sequence can also be obtained as follows:

[0113] A comprehensive evaluation is conducted on the decline order decision variables and the corresponding multi-scenario operating trajectories generated during the joint optimization process to form a candidate solution set.

[0114] Specifically, during the joint optimization process, the outer search algorithm generates a large number of potential drawdown order combinations. To construct a rich and diverse rule base, it is not sufficient to retain only a single optimal solution; rather, a batch of non-dominated or near-optimal solutions must be retained. A multi-dimensional evaluation is performed on each drawdown order scheme that satisfies the basic constraints. Evaluation metrics include: average power generation efficiency under multiple scenarios, total water wastage, undrawdown risk indicators, and downstream flood safety margin. For all schemes that fall within the risk constraints and perform reasonably well in terms of efficiency indicators, their corresponding multi-scenario operational trajectory data are retained to form a candidate solution dataset for the drawdown sequence.

[0115] Based on the principle of prioritizing safety while considering overall benefits, the optimization results data of the dropout sequence with strong representativeness are selected from the candidate solution set.

[0116] In this embodiment, to provide representative samples for subsequent cluster analysis, the large candidate set needs to be reduced and screened. The screening logic follows a dual standard of safety baseline and benefit optimization. Specifically, using the unresolved risk index and conditional value of risk index as hard thresholds, all schemes with uncontrollable risks under extreme scenarios are eliminated. Among the remaining safe schemes, a multi-objective ranking method, such as Pareto front ranking, is used to select several representative schemes with different emphases on power generation benefits and water wastage control. For example, typical schemes of different types, such as those with the highest power generation benefits, the least water wastage, and the most balanced risks, are selected. These selected representative schemes and their corresponding multi-scenario operating trajectories are packaged as the final elimination sequence optimization result data. This ensures that the data subsequently input to the pattern recognition module contains sufficient diversity while eliminating inferior solutions, improving the efficiency and quality of rule construction.

[0117] As an optional implementation method, quantifying the distribution parameters of unreceded risk can also involve: reading the predicted inflow data after deviation correction and the statistical parameters of the prediction error, as well as the pre-flood state parameter set, and constructing conditional inflow probability description data under the current pre-flood state conditions for each control section and each time step. Specifically, the predicted inflow data after deviation correction is used as the conditional inflow mean, and the variance, skewness, and other information in the statistical parameters of the prediction error are used as uncertainty characterization parameters. The error distribution is corrected according to the current water storage abundance / dryness indicators in the pre-flood state parameter set. In the case of abundance, the tail weight of the high inflow side is increased, and in the case of dryness, the tail weight of the high inflow side is decreased, resulting in conditional inflow probability description data for each time step. The conditional inflow probability description data is then read, and a preliminary candidate scenario inflow sequence data is constructed using a random sampling method. Specifically, for each control section and each time step, multiple sets of inflow sample values ​​are generated using Monte Carlo sampling or Latin hypercube sampling based on the probability distribution parameters given by the conditional inflow probability description data. These sample values ​​are then spliced ​​hourly or daily over the entire pre-flood period to form multiple complete candidate inflow time series. To ensure the coverage of samples across different wet and dry seasons, the sampling results can be stratified according to preset wet, normal, and dry water probability intervals to ensure that the candidate scenarios include not only extremely high inflow sequences but also extremely low inflow sequences and typical sequences close to the median. All these time series are then uniformly organized into structured candidate scenario inflow sequence data. The candidate scenario inflow sequence data is read, and combined with the safety objectives in the pre-flood state parameter set, the candidate sequences are filtered and compressed to generate the final scenario-based inflow sequence set. Specifically, this includes: calculating the total inflow of each candidate scenario during the pre-flood period, the peak inflow within several key time windows, and the deviation from historical statistical characteristics, and eliminating extreme abnormal scenarios that are clearly inconsistent with the current pre-flood state; using clustering or similarity measures to group the candidate scenario inflow sequences, selecting several representative scenarios within each group to ensure that the final scenario-based inflow sequence set is sufficiently representative in terms of abundance / dampness, inflow concentration, and peak characteristics; and packaging the retained scenario time series into a scenario-based inflow sequence set according to a unified structure.

[0118] By combining the scenario-based inflow sequence set with the corresponding initial reservoir water level, generating unit output capacity, discharge capacity, and grid load demand information from the pre-flood baseline dataset, a unified multi-scenario hydrological operation input set is formed. In this process, each scenario inflow sequence is appended with corresponding grid load paths, generating unit maintenance status paths, and constraint parameter paths, ensuring that each scenario contains comprehensive information on hydrological, load, and operational constraints. This enables subsequent risk assessment and optimization models to fully perceive the multidimensional constraints faced by the reservoir system during the pre-flood season within each scenario.

[0119] By reading the pre-flood baseline dataset and historical hydrological datasets, and analyzing the actual dispatching behavior characteristics under different inflow conditions during the pre-flood period in previous years, empirical baseline dispatching rules are extracted, and structured baseline dispatching rule parameter data is formed. For example, by statistically analyzing the pre-flood water level change curves of each cascade reservoir in different years of abundant and scarce inflow, the target reservoir utilization rate interval and adjustment rate parameter in the equal reservoir utilization rate drawdown rule are determined; by analyzing the commonly used start and drawdown dates and drawdown rhythms in actual dispatching, the start time, end time, and corresponding water level target of the fixed start and drawdown date rule are determined. After the above parameters are extracted, the obtained parameters are packaged into baseline dispatching rule parameter data, which serves as a unified rule input for rapid extrapolation under different scenarios. By reading the multi-scenario hydrological operation input set and baseline dispatching rule parameter data, water balance calculations and baseline dispatching simulations are performed for each scenario to obtain the initial water level process data under each scenario. Specifically, in each scenario, starting from the current reservoir water level, the outflow and power generation are calculated step-by-step according to the start and drawdown times, target reservoir capacity utilization, and drawdown rate in the baseline dispatching rule parameters. Based on the scenario inflow, outflow, and water loss at each time step, the reservoir water level is updated to obtain the water level time series for the entire pre-flood period. During this process, for different baseline dispatching rules, such as the equal reservoir capacity utilization rule and the fixed start and drawdown date rule, the corresponding water level processes can be calculated separately under the same scenario, and a single baseline water level process can be selected or weighted and synthesized according to the dispatcher's preference. The baseline water level processes under all scenarios are then uniformly compiled into initial water level process data.

[0120] The system reads initial water level process data and pre-set safe water level control parameters, and performs reservoir-by-reservoir non-discharge exceedance analysis for each scenario, generating single-scenario non-discharge exceedance data. The safe water level control parameters include the safe control water level and safety margin for each reservoir at key pre-flood time points. Specifically, in each scenario, a predetermined key pre-flood time point is selected, and the difference between the reservoir water level at that time and the safe control water level minus the safety margin is compared. If the difference is positive, it is considered that the water level failed to drop on time, and the difference is the exceedance height for that scenario; if the difference is negative or zero, it is considered that the water level dropped successfully, and the exceedance height is recorded as zero. The exceedance height is calculated for each reservoir and each scenario, and the results are summarized to form single-scenario non-discharge exceedance data. This data records in detail whether the water level drop was successful and the degree of exceedance under each scenario. The single-scenario non-discharge exceedance data and the scenario weight information contained in the multi-scenario hydrological operation input set are read to calculate the non-discharge success risk index data for each reservoir. Specifically, this includes: statistically analyzing the proportion of scenarios where the excess height is greater than zero across all scenarios to obtain the probability of successful non-discharge; calculating the average excess height when the excess height is greater than zero to obtain the average degree of non-discharge; and, if necessary, calculating the maximum excess height as the extreme degree of non-discharge. These multiple risk indicators are organized into a unified structure to form non-discharge success risk indicator data. Single-scenario non-discharge data and scenario probability or scenario weight information are read to construct non-discharge risk distribution parameter data characterizing the distribution pattern of non-discharge risk at each reservoir dimension. Specifically, a cumulative distribution function is constructed for the excess height under all scenarios, and the quantile excess height at a specified confidence level is calculated to depict the possible loss under extreme scenarios; further, the conditional value of risk (VHR) index is calculated, for example, the VHR can be defined as the average excess height when the excess height exceeds a certain threshold; simultaneously, the excess probability threshold used to construct opportunity constraints is recorded. This yields non-discharge risk distribution parameter data including the quantile of the non-discharge excess height, the conditional value of risk, and the excess probability threshold.

[0121] In one possible implementation, obtaining the drawdown sequence optimization results data can also involve: reading the multi-scenario hydrological operation input set and the undrawdown risk distribution parameter data, and combining them with the pre-flood season scheduling benefit and safety objectives to construct a complete comprehensive objective function definition. The comprehensive objective function can take the following form:

[0122] J total = Σ scenario (weight) scenario × E total_scenario ) - λ risk × CVaR β (L risk );

[0123] Among them, the comprehensive objective function Jtotal For the objective that needs to be maximized, weight scenario E is the weight of each scenario in the multi-scenario hydrological operation input set. total_scenario CVaR represents the total power generation of the cascade reservoirs during the pre-flood season under this scenario. β (L risk For the loss L of undiminished risk at confidence level β risk Conditional Value at Risk (VaR) and Risk Penalty Coefficient λ risk To weigh the power generation benefits against the risk of unabated emissions, Σ scenario Summing over all hydrological scenarios. Undissipated risk loss L risk This can be derived from the distribution parameters of undrawn drawdown risk, for example, by taking the weighted average excess height of a reservoir when the excess height is greater than zero as a measure of loss. Power generation revenue and undrawn drawdown risk are jointly incorporated into a unified objective function, forming structured objective function definition data, which is then written into the pre-flood joint drawdown optimization model data structure. The distribution parameters of undrawn drawdown risk are read, and the safety requirements regarding unsuccessful drawdown on time are transformed into mathematical constraints, forming risk constraint definition data. Specifically, based on the pre-flood safety plan, a maximum permissible probability threshold for successful drawdown and a corresponding conditional risk value threshold are set for each reservoir, using the probability of successful drawdown and the conditional risk value index from the undrawdown success risk index data for constraint. For example, the following opportunity constraint can be constructed: probability of successful drawdown ≤ P max ; and the following risk constraints: CVaR β (L risk )≤L max The probability of failure to achieve a settlement rate is derived from the risk indicator data of failure to achieve a settlement rate, specifically the Conditional Value at Risk (CVaR) indicator. β (L risk (This is derived from the unresolved risk distribution parameter data, P) max and L maxA preset safety threshold is used. Risk constraint definition data is generated for optimization, and this risk constraint, along with water balance constraints and water level constraints, is written into the pre-flood joint drawdown optimization model data structure. The multi-scenario hydrological operation input set, objective function definition data, and risk constraint definition data are read. Simultaneously, water balance constraints, water level upper and lower limits constraints, downstream flow constraints, unit output constraints, and flood control safety constraints are combined to generate complete constraint condition definition data. The objective function definition data and constraint condition definition data are then uniformly encapsulated to form the pre-flood joint drawdown optimization model data structure. In the constraint condition definition data, water balance constraints ensure consistency between reservoir water level updates and inflow and outflow at each time step; water level upper and lower limits constraints ensure that the reservoir water level remains within the allowable range; downstream flow constraints and unit output constraints ensure the safe operation of hydraulic facilities and power grid equipment; and flood control safety constraints ensure that the safe water level and flow at downstream control sections are not exceeded during the pre-flood stage and in cases of potential early inflow. This results in a pre-flood joint drawdown optimization model data structure containing complete objectives and constraints.

[0124] The pre-flood joint drawdown optimization model data structure is read, and drawdown order decision variables are introduced for each cascade reservoir, with the variable types and physical meanings clearly defined. The default implementation uses integer drawdown stage number variables, where each reservoir is defined with an integer drawdown stage number representing the stage number of the main drawdown process. The stage number ranges from one to a preset number of stages, such as one to three, corresponding to the priority drawdown stage, coordinated drawdown stage, and following drawdown stage, respectively. Optionally, a 0-1 type precedence relationship variable, the inter-reservoir drawdown precedence variable, can be introduced to the model to represent the relationship where one reservoir's drawdown precedes another. The definitions and value ranges of the above variables are uniformly organized into drawdown order decision variable data, which is written back to the pre-flood joint drawdown optimization model data structure as the updated set of decision variables. The drawdown order decision variable data is read, and corresponding order relationship constraint definition data is constructed for the drawdown stage number variables. For example, when using drawdown stage numbering variables, inequality constraints stipulate that the priority stage number is less than the coordination stage number, and the coordination stage number is less than the following stage number. Simultaneously, the number of reservoirs within each stage is limited to a preset capacity. Constraints also ensure that there are no illegal stage jumps. When using 0-1 type inter-reservoir drawdown order variables, transitivity and antisymmetry constraints can be constructed. For example, for any three reservoirs, if Reservoir 1 precedes Reservoir 2 and Reservoir 2 precedes Reservoir 3, Reservoir 1 must precede Reservoir 3. For any two reservoirs, a relationship where Reservoir 1 precedes Reservoir 2 and Reservoir 2 precedes Reservoir 1 is not allowed simultaneously. All these order relationship constraints are written into the order relationship constraint definition data in a unified form and integrated with the original water balance constraints and water level constraints to ensure that the drawdown order always meets the requirements of physical rationality and engineering experience during the optimization solution process. By reading the order relation constraint definition data and the multi-scenario hydrological operation input set, the impact of drawdown order on the water level changes and discharge capacity of each reservoir is explicitly coupled to the water balance and flood control safety constraints, forming an updated pre-flood joint drawdown optimization model data structure. Specifically, based on the drawdown stage number variable for each reservoir, the pre-flood time axis is divided into several stages, and different water level decline rates and discharge flow limits are applied to each stage. This ensures that reservoirs in the priority drawdown stage have greater discharge space in the early stages, while reservoirs in the follow-up stages maintain relatively slow water level changes in the early stages. Simultaneously, considering flood control time and downstream reservoir capacity constraints, when upstream reservoirs enter the rapid drawdown stage, additional constraints are applied to limit the water level of downstream reservoirs from approaching the safe water level limit within the corresponding time window, preventing downstream safety risks caused by large upstream flows arriving prematurely. The order relation constraint definition data is no longer just an independent logical constraint, but forms a coupled relationship with the water balance and flood control safety constraints, constituting a new pre-flood joint drawdown optimization model data structure.

[0125] The updated pre-flood joint drawdown optimization model data structure is read, and an outer-layer combinatorial search strategy is designed for the drawdown order decision variables to form a candidate drawdown order set. Specifically, an evolutionary algorithm suitable for integers and 0-1 variables, such as a genetic algorithm, is used to encode the drawdown stage number variable, constructing an initial population of candidate solutions for the drawdown order. New candidate solutions are generated iteratively through operations such as crossover, mutation, and selection. In each generation of the population, the candidate solutions are sorted according to the comprehensive objective function value and risk satisfaction obtained from the previous inner-layer solution, and the best-performing drawdown order scheme is selected to enter the next iteration. This generates a continuously updated candidate drawdown order set. The candidate drawdown order set data and the multi-scenario hydrological operation input set are read, and combined with the comprehensive objective function definition data and constraint condition definition data, the operation trajectory is optimized for each scenario under the premise of a given candidate drawdown order, obtaining the corresponding pre-flood drawdown water level process data, pre-flood drawdown power output process data, and pre-flood drawdown discharge process data. In practical implementation, successive optimization algorithms or piecewise dynamic programming methods can be used to divide the overall pre-flood scheduling period into multiple sub-periods. Each sub-period is optimized sequentially, and consistency between sub-periods is ensured through state propagation. In each sub-period solution, constraints on drawdown order, water balance, upper and lower water level limits, discharge flow, and flood control safety must be satisfied. By analyzing the trajectory data obtained from solving all scenarios, a set of scenario operation trajectories corresponding to a candidate drawdown order is formed, and the objective function value of this set under multiple scenarios is calculated. The scenario operation trajectory set data and the corresponding comprehensive objective function value are read, and combined with the unspent risk distribution parameter data, the performance of the outer-layer candidate drawdown orders is evaluated and iteratively updated until the convergence criterion is met. Specifically: for each candidate drawdown order, it is checked whether the corresponding scenario operation trajectory meets the risk constraints and other constraints. If not, the fitness of the candidate solution is reduced in the outer-layer evolutionary algorithm; for candidate solutions that meet the constraints, the comprehensive objective function J is used to further refine the algorithm. total The size of the target value maps its fitness to a selection probability, prioritizing the retention of candidate solutions with larger target values ​​in the next generation; this process is repeated until the change in the optimal target value is below a preset threshold or the number of iterations reaches the upper limit. Upon convergence, all candidate solutions that satisfy the risk constraints and are at the forefront of non-dominated solutions, along with their corresponding scenario trajectories, are included in a aggregated set of pre-flood drawdown water level process data, pre-flood drawdown power output process data, and pre-flood drawdown discharge process data. This provides input for further screening to form a candidate solution dataset for the drawdown sequence and a multi-scenario trajectory dataset.

[0126] A comprehensive multi-scenario evaluation was conducted on the decision variables for each group of drawdown sequences, along with their corresponding pre-flood drawdown water level, power output, and discharge flow data. This resulted in a candidate drawdown sequence dataset and a corresponding multi-scenario operational trajectory dataset. Specific evaluation indicators included: non-drawdown risk indicators, conditional risk value indicators, total water wastage, total cascade power generation, and flood control safety margin at key sections under different scenarios. Schemes that did not meet the preset risk constraints or performed significantly worse than other drawdown sequences under multiple scenarios were eliminated. The selected schemes and their corresponding multi-scenario trajectories were then included in the candidate drawdown sequence dataset and the multi-scenario operational trajectory dataset. This not only generated several candidate drawdown sequences with superior safety and efficiency but also provided a rich data foundation for subsequent pattern recognition and rule extraction.

[0127] Based on the candidate solution dataset of drawdown sequences and the multi-scenario operation trajectory dataset, and adhering to the principles of prioritizing safety and balancing comprehensive benefits, several representative drawdown sequences are selected as the optimization results data. Specifically, the selection criteria are based on undrawn risk indicators and conditional risk value indicators, retaining schemes with risk levels within a set threshold. Furthermore, based on the total cascade power generation and water wastage, several representative schemes in terms of risk-benefit trade-offs are selected, and the corresponding multi-scenario operation trajectories are aggregated into a multi-scenario operation trajectory dataset. Unlike the traditional approach of selecting only a single optimal solution, the output drawdown sequence optimization results data exists in the form of a set of multiple schemes, providing diverse samples for subsequent drawdown sequence pattern recognition and scheme rule construction, and can accommodate scheduling needs with different safety and benefit preferences in practical applications.

[0128] According to one aspect of this application, generating a pre-flood drawdown scheme rule base data can also involve: reading the drawdown sequence optimization result data and the corresponding multi-scenario operation trajectory dataset, extracting information such as the drawdown start time, drawdown end time, and drawdown phase duration for each reservoir, and constructing drawdown time series feature data describing the drawdown time series characteristics. Specifically, for each optimization result, the main water level decline phase of each reservoir is identified, and the start time, end time, and phase length of this phase are used as basic features; the difference in drawdown start time, the degree of overlap of drawdown phases, and the phase order between reservoirs are calculated to obtain time series features such as whether upstream or downstream drawdown occurs first, and whether the entire cascade drawdown is synchronous or phased. By normalizing the above time series parameters, drawdown time series feature data suitable for clustering algorithms is generated. The multi-scenario operation trajectory dataset and undrawdown risk distribution parameter data are read, and representative indicators of risk and benefit for each drawdown sequence are extracted to constitute risk-benefit feature data. Specifically, this includes calculating the average total power generation of each cascade, the total amount of water wasted, the probability of failure to draw down, the conditional risk value, and the flood safety margin for each drawdown sequence, and then concatenating these indicators with the drawdown time series characteristic data to form a high-dimensional drawdown sequence feature vector data.

[0129] Read the feature vector data of the dropout sequences, preprocess and standardize the feature components of different dimensions and magnitudes to form standardized feature vector data suitable for cluster analysis. Specifically, zero-mean unit variance standardization or interval mapping is performed on time-series features and risk-benefit features respectively, so that each dimension of features has a similar scale in the feature space; for indicators that may have large skewness, such as total water wastage or conditional risk value, logarithmic transformation or piecewise stretching can be performed to reduce the impact of extreme values ​​on the clustering results. After standardization, the feature vectors of all dropout sequences are uniformly organized into standardized feature vector data. Read the standardized feature vector data, select an appropriate clustering algorithm and determine the number of clusters, perform cluster analysis, and generate preliminary dropout sequence pattern label data. Specifically, the process involves: Firstly, the K-means clustering algorithm is preferred, using Euclidean distance as the similarity measure to cluster standardized feature vectors. To determine the number of clusters, K-means clustering can be performed repeatedly with different candidate cluster numbers, and the silhouette coefficient or sum of squared intra-cluster deviations corresponding to each cluster number can be calculated. A reasonable number of clusters K is selected based on the method of maximizing the silhouette coefficient or identifying a significant inflection point in the intra-cluster deviations. After determining the number of clusters, K-means clustering is executed, assigning each dropout sequence to a specific cluster and outputting the pattern number to which each dropout sequence belongs, forming dropout sequence pattern label data. Then, the dropout sequence pattern label data and dropout sequence feature vector data are read, and feature statistics and interpretation are performed on each cluster to generate dropout sequence pattern data, which is then compiled into a dropout sequence pattern library. Specifically, for each cluster, the average value and distribution range of the drawdown time series characteristics within that cluster are calculated to determine whether the model belongs to upstream priority drawdown, downstream priority drawdown, or synchronous drawdown across all cascades. Simultaneously, statistical values ​​of the risk-benefit characteristics within the cluster are calculated, recording the typical undrawdown risk level, total power generation, and water wastage for that model. Based on these statistical results, each cluster is assigned a model name, model number, and model characteristic description, and this information is encapsulated into drawdown sequence model data. All model data are then aggregated to form a drawdown sequence model library.

[0130] The system reads and aligns the water level, power output, and discharge flow trajectories within each drawdown sequence mode, based on the drawdown sequence mode library, drawdown sequence optimization results, and multi-scenario operational trajectory dataset. Statistical analysis is then performed to generate trajectory statistical description data. Specifically, for any mode, all drawdown sequences and their corresponding multi-scenario operational trajectories belonging to that mode's label are selected. The time axis is unified to the same pre-flood period, and each trajectory is interpolated and aligned according to a fixed time step. At each time step, the median, upper quantile, and lower quantile of water level, as well as similar statistical values ​​for power output and discharge flow, are calculated for all trajectories within that mode. This yields recommended values ​​and upper and lower control boundaries at each time step. The above statistical results are summarized to form trajectory statistical description data at the mode dimension, providing a foundation for constructing typical control zones. The trajectory statistical description data is then read, and the statistical results within each mode are transformed into typical water level process data, typical power output process data, and typical discharge flow process data that can be used for control. Specifically, for each model, at each time step, the median water level is used as the typical control value, and the upper and lower quantiles are used as the upper and lower limits of the water level control zone. Similarly, the median and quantiles of power output and discharge flow are used to construct the power output control zone and discharge control zone, respectively. The typical process data formed in this way not only reflects the average behavior under this model, but also reflects the reasonable fluctuation range under multi-scenario uncertainty through the form of control zones. Typical water level process data, typical power output process data, and typical discharge flow process data for all models are output uniformly.

[0131] Read the drawdown sequence pattern library data, typical water level process data, typical power output process data, and typical discharge flow process data. Divide the entire pre-flood period into several control stages according to pre-flood scheduling requirements, generating control stage division data. Specifically: based on the water level change trends in the typical water level process data, identify obvious drawdown acceleration, stabilization, and termination stages, and define the corresponding time periods as control stages. When dividing control stages, consider the drawdown stage numbers of different reservoirs in the drawdown sequence pattern to ensure that priority drawdown reservoirs have independent water level control targets in the early control stages, while coordinated and following stage reservoirs have separate regulation space in the later control stages. Organize the start and end times and stage numbers of the control stages for each pattern into control stage division data. Read the control stage division data, typical water level process data, typical power output process data, and typical discharge flow process data. For each drawdown sequence pattern, generate normal drawdown control rule data for each control stage. Specifically, within a certain control phase, the median value of typical water level process data is used as the target water level curve, and the upper and lower quantiles are used as the upper and lower limits of water level control, forming the water level control zone for that phase. Similarly, a power output control zone is formed using typical power output process data, and a discharge control zone is formed using typical discharge flow process data. Simultaneously, considering water balance and equipment capacity limitations, the maximum allowable adjustment amplitude and adjustment rate when the water level deviates from the control zone are defined for each phase. Typical process data for each mode are converted into phased water level, power output, and discharge control zones, comprehensively forming structured normal drawdown control rule data, which is then written into the pre-flood drawdown scheme rule base data. The normal drawdown control rule data and multi-scenario hydrological operation input sets are read, and the processing logic when the measured water level, power output, or discharge flow deviates from the normal control zone during actual operation is defined. This logic is also written into the pre-flood drawdown scheme rule base data. Specifically, several deviation levels are set, such as slight deviation and severe deviation. When the measured water level exceeds the water level control zone but does not approach the upper limit of the safe water level, the slight deviation handling logic is triggered, and the water level is brought back to the control zone by slightly adjusting the output or discharge flow. When the measured water level approaches the upper limit of the safe water level or the discharge flow approaches the upper limit of the capacity, the severe deviation handling logic is triggered, which can temporarily increase the discharge ratio of the priority drawdown reservoir or start the drawdown of the coordinated phase reservoir in advance to quickly reduce the risk. The above deviation handling logic, together with the normal control zone, constitutes a complete normal drawdown control rule data, so that the rule base not only contains the ideal trajectory, but also the automatic adjustment strategy for deviation.

[0132] The data structure of the pre-flood joint drawdown optimization model and the distribution parameters of undrawn risk are read. Under emergency scenarios, the risk constraint parameters and objective function weights are adjusted, and the drawdown order and trajectory optimization under risk constraints are re-executed to obtain emergency water level process data, emergency power output process data, and emergency discharge flow process data under emergency conditions. Specifically, when the flood control situation is predicted to be relatively relaxed or the power grid has a higher demand for pre-flood power generation, the upper limit of the probability of successful undrawn drawdown or the upper limit of conditional risk value is appropriately relaxed, or the risk penalty coefficient λ is reduced. risk This allows the model to accept a slightly higher risk of uncollected water level within the control range in exchange for higher power generation efficiency. Under the new parameters, the search for combinations of drawdown order and the solution of multi-scenario operating trajectories are repeatedly executed to form an emergency optimized trajectory set different from the normal scenario, which is then uniformly organized into emergency water level process data, emergency power output process data, and emergency discharge flow process data. Reading the emergency water level process data, emergency power output process data, and emergency discharge flow process data, and according to trajectory statistics and stage division methods, control zones and control rules under the emergency scenario are constructed, forming emergency drawdown control rule data and delayed drawdown control rule data. Specifically, the emergency trajectories are aligned and statistically analyzed within the model to calculate the median and quantiles of the emergency water level process, constructing an emergency water level control zone; similarly, emergency power output control zones and emergency discharge control zones are constructed; based on these control zones, water level and power output regulation strategies are defined for each stage under emergency conditions, such as allowing delayed drawdown in the early stage to increase power generation, but requiring accelerated discharge in the later stage to complete drawdown before the main flood season. The aforementioned control zones and regulation strategies are uniformly organized into emergency drawdown control rule data and delayed drawdown control rule data, and added to the pre-flood drawdown scheme rule base data. Real-time state feature vector data and undrawdown risk distribution parameter data are read, and combined with the previously generated emergency drawdown control rule data and delayed drawdown control rule data, the triggering conditions for emergency control rules and the switching logic for reverting to normal control rules are defined, and the relevant logic is written into the pre-flood drawdown scheme rule base data. Specifically: several triggering conditions are set, such as triggering the delayed drawdown control rule when the real-time state feature vector shows that the recent forecast inflow is low and the actual water storage is significantly below average; triggering the enhanced emergency drawdown rule when the forecast inflow shows extremely heavy concentrated rainfall and cannot quickly recede in the short term; after triggering the emergency rule, the real-time update of the undrawdown risk distribution parameter data is monitored, and when the risk index recovers to below the normal safety threshold for a certain period of time, the system gradually switches back to the normal drawdown control rule through preset recovery logic to avoid frequent switching causing operational oscillations. Emergency drawdown control rule data and delayed drawdown control rule data have been fully incorporated into the pre-flood drawdown scheme rule base data, forming a closed-loop control system of normal-emergency-recovery.

[0133] In summary, a method for constructing a pre-flood joint drawdown sequence and drawdown scheme for cascade reservoirs is proposed. This method involves conditionally perturbing the predicted inflow based on a pre-flood state parameter set, constructing a multi-scenario hydrological operation input set, and quantifying the distribution parameters of undrawdown risk that characterize the possibility of water level exceeding limits. A pre-flood joint drawdown optimization model is constructed, incorporating drawdown sequence decision variables. The undrawdown risk distribution parameters are transformed into rigid risk constraints within the model. A strategy combining outer-layer sequence search and inner-layer trajectory optimization is used to jointly solve for the drawdown sequence and operational trajectory. The optimization results are clustered to identify drawdown sequence patterns and generate a normal and emergency drawdown control rule base. Finally, the target scheme is matched and corrected online based on real-time status.

[0134] This invention addresses the problem of rigid drawdown order decision-making by constructing a mixed-integer nonlinear programming model. It introduces drawdown order decision variables, such as stage number variables or 0-1 order relation variables, and establishes coupling constraints between the order variables and the physical hydraulic boundary. This transforms the order from a fixed input condition into an optimizable decision variable, enabling automatic searching and matching of the optimal inter-reservoir coordination sequence under multiple scenarios, thus solving the problem of poor adaptability of scheduling strategies caused by fixed orders. To address the lack of control over undrawdown risk and insufficient ability to cope with extreme variability, this invention employs a current-state-based conditional scenario generation technique and a conditional risk value constraint mechanism. By adjusting the forecast error disturbance weights according to the water level fluctuations, more targeted input scenarios are generated. Explicitly incorporating conditional risk value constraints and penalty terms regarding the undrawdown exceedance height into the optimization model achieves quantitative control of tail risks. This ensures that the generated scheme meets the safety baseline even under extreme forecast deviations, solving the problem that traditional models relying solely on water abandonment penalties cannot effectively mitigate safety risks. Furthermore, by employing pattern recognition and rule base construction technologies, complex optimization results are transformed into visualized control zones and emergency rules, solving the problem that complex optimization theories are difficult to directly apply in actual scheduling and achieving effective integration between offline optimization and online decision-making.

[0135] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for constructing the pre-flood joint drawdown sequence and drawdown scheme for cascade reservoirs, characterized in that, include: Based on the pre-flood state parameter set of cascade reservoirs, conditional perturbation is applied to the forecast deviation to generate a multi-scenario hydrological operation input set, and the distribution parameter data of the undrawn risk that characterizes the possibility of water level exceeding the limit is quantified. A pre-flood joint drawdown optimization model is constructed by introducing drawdown order decision variables. The data of undrawdown risk distribution parameters are transformed into risk constraints in the pre-flood joint drawdown optimization model, and the drawdown order decision variables are coupled with the pre-stored pre-flood drawdown trajectory data. Under risk constraints, a joint pre-flood drawdown optimization model is solved for multi-scenario hydrological operation input sets to obtain drawdown sequence optimization results data that meet safety thresholds; Pattern recognition is performed on the drawdown sequence optimization results data to extract the drawdown sequence patterns and generate corresponding pre-flood drawdown scheme rule base data.

2. The method according to claim 1, characterized in that, Construct a pre-flood joint drawdown optimization model that incorporates drawdown order decision variables, including: Construct a data structure for drawdown order decision variables, the data structure including an integer stage number variable for identifying the drawdown stage of a reservoir, or a 0-1 type order relationship variable for identifying the drawdown order between reservoirs; Based on the data structure, a comprehensive objective function is constructed with the goal of maximizing the total power generation of the cascade and minimizing the undissipated risk. The comprehensive objective function is composed of a weighted sum of the expected power generation under multiple scenarios and the undissipated risk penalty term. Among them, the penalty for unresolved risk is constructed based on the conditional value of risk index in the unresolved risk distribution parameter data.

3. The method according to claim 2, characterized in that, The coupling relationship between the drawdown order decision variables and the pre-flood drawdown trajectory data is established, including: Construct order relation constraints for the drawdown order decision variables to ensure that the drawdown order among cascade reservoirs satisfies logical transitivity and that there are no illegal stage jumps; Based on order relation constraints, differentiated physical boundary parameters are set for cascade reservoirs with different drawdown priorities; By using physical boundary parameters to limit the upper limit of discharge flow and the rate of water level decline in the pre-flood drawdown trajectory data, reservoirs that enter the drawdown phase first have higher discharge authority than reservoirs that follow the drawdown phase.

4. The method according to claim 1, characterized in that, Obtain the optimization results data of the dropout sequence, including: In the outer layer of the pre-flood joint drawdown optimization model, a sequence search algorithm is used to perform a combined search on the drawdown order decision variables to generate a candidate drawdown order set. In the inner layer of the pre-flood joint drawdown optimization model, for each candidate drawdown order in the candidate drawdown order set, the order relationship between reservoirs is fixed, and for each scenario in the multi-scenario hydrological operation input set, the pre-flood drawdown operation trajectory data that satisfies water balance is calculated. The comprehensive objective function value under each scenario is calculated and the risk constraints are verified. The candidate drawdown order that meets the risk constraints and the pre-flood drawdown trajectory data that meets the water balance are summarized into drawdown sequence optimization result data.

5. The method according to claim 1, characterized in that, Quantitative data on the distribution parameters of unresolved risk include: By using a pre-set empirical scheduling strategy, each scenario in the multi-scenario hydrological operation input set is simulated to obtain initial water level process data reflecting the baseline scheduling state. For each scenario, select key time points before the flood season, calculate the difference between the initial water level process data and the preset safety control water level, and record the part above the safety control water level as the un-dropped excess height; Based on the unreceded height exceeding the limit under all scenarios, the probability of successful unreceding and the conditional risk value at a preset confidence level are statistically analyzed, and the statistical results are combined into unreceded risk distribution parameter data.

6. The method according to claim 1, characterized in that, The proposed drop-off sequence pattern includes: Information describing the order of drawdown between reservoirs is extracted from the drawdown sequence optimization results data as drawdown time series feature components, and information describing the total power generation of the cascade and the risk indicators of undrawdown are extracted as risk-benefit feature components. The drawdown time-series feature components and the risk-benefit feature components are concatenated to construct the drawdown sequence feature vector data; Clustering algorithms are used to group the feature vector data of the dropout sequence. Each cluster obtained by grouping is defined as a dropout sequence pattern, and a mapping relationship between each dropout sequence pattern and the corresponding optimized solution set is established.

7. The method according to claim 6, characterized in that, Generate a rule base data for the pre-flood drawdown scheme, including: For each drawdown sequence pattern, the pre-flood drawdown trajectory data corresponding to that drawdown sequence pattern is collected and aligned on a unified time axis; Calculate the median, upper quantile, and lower quantile of water level at each time step of the aligned pre-flood drawdown trajectory data; The target water level curve is constructed using the median water level, and the water level control zone that encloses the target water level curve is constructed using the upper and lower water level quantiles. The target water level curve and water level control zone are encapsulated as normal drawdown control rule data and stored in the pre-flood drawdown scheme rule database.

8. The method according to claim 7, characterized in that, The generation of the pre-flood drawdown scheme rule base data also includes: Adjust the risk penalty coefficient in the pre-flood joint drawdown optimization model, or relax the threshold parameters in the risk constraints; Under the adjusted or relaxed parameter conditions, the pre-flood joint drawdown optimization model is re-solved to obtain the set of emergency optimization trajectories under specific working conditions; Emergency water level control zones are extracted from the emergency optimized trajectory set, and emergency drawdown control rule data containing triggering conditions and recovery logic is generated and added to the pre-flood drawdown scheme rule base data.

9. The method according to claim 1, characterized in that, Generate a multi-scenario hydrological operation input set, including: The current water storage index is extracted from the pre-flood state parameter set, and the forecast error disturbance weights used to correct the random error distribution are determined accordingly. Based on the pre-stored bias-corrected forecast inflow data, random disturbances adjusted by forecast error disturbance weights are superimposed to generate a set of conditional scenario inflow sequences that conform to the characteristics of the current hydrological state. By combining the set of inflow sequences under conditional scenarios with the corresponding unit operation constraint parameters, a structured multi-scenario hydrological operation input set is constructed.

10. The method according to claim 1, characterized in that, Also includes: Based on the pre-flood state parameter set, match the target drawdown order scheme data and the target pre-flood drawdown operation scheme data from the pre-flood drawdown scheme rule base data; The feasibility of the target drawdown sequence scheme data and the target pre-flood drawdown operation scheme data is verified using pre-stored real-time pre-flood operation data, and the scheme feasibility verification results are generated. When the feasibility verification results indicate that there is an excessive risk of uncontrolled drawdown or that the discharge capacity exceeds the limit, the emergency drawdown control rules pre-stored in the pre-flood drawdown scheme rule base data are called to make local corrections to the water level control zone in the target pre-flood drawdown operation scheme data, and generate an executable target scheme.

Images