Intelligent dispatching decision support method and system for cascade reservoir power station

Abstract

The application discloses a kind of cascade reservoir power station intelligent scheduling decision support method and system.The method is first based on the correlation characteristics of historical meteorological elements and runoff prediction deviation under different meteorological combination quadrant, deviation correction is carried out to future predicted runoff and a set of random runoff scenarios is generated;Then, based on the current scheduling rules, multi-scenario parallel simulation is carried out, the power generation index and the power generation loss conditional risk value are calculated to determine the rule adaptability;When there is a deterioration risk, the rule construction parameter adjustment feasible region is analyzed and the main control parameter is identified, and the optimized scheduling parameter suggestion is generated with the dual guidance of maximizing the cascade comprehensive benefit and minimizing the parameter adjustment amplitude.The application realizes the risk identification and adaptive optimization of scheduling rules under hydrological change environment.

Description

Technical Field

[0001] This invention belongs to the field of water conservancy and hydropower engineering, and in particular to a method and system for intelligent scheduling decision support of cascade reservoir power stations. Background Technology

[0002] As core projects for watershed hydropower utilization, the adaptability of the scheduling rules of cascade reservoir power stations directly determines their power generation efficiency and operational safety. Given the significant changes in watershed hydrological conditions driven by climate factors, a key technical challenge for maintaining the efficient and stable operation of cascade reservoir power station systems is how to conduct forward-looking adaptive assessments and fine-tuned parameter adjustments of existing scheduling rules based on precise runoff input.

[0003] Currently, cascade reservoir operation typically utilizes climate model outputs such as temperature and precipitation data to drive hydrological models to generate predicted runoff, and generally employs globally unified statistical formulas to correct prediction biases. In maintaining operation rules, long-term adjustment calculations are primarily based on historical measured hydrological sequences, and the rationality of the rules is verified by comparing design indicators with actual operational indicators. For adjustments to key parameters such as post-flood water levels and timing, experience-based trial-and-error algorithms or single-objective optimization methods are often used to determine parameter values ​​while meeting safety constraints.

[0004] However, existing technologies suffer from limitations in the accuracy of prediction bias correction, insufficient risk quantification, and low efficiency in parameter optimization when dealing with inconsistent hydrological inputs and multi-objective risk decision-making. Specifically, globally uniform bias correction ignores the heterogeneity of the impact of factors such as temperature and precipitation on runoff deviation under different combination quadrants (e.g., high temperature and abundant rainfall or low temperature and little rainfall), leading to larger prediction errors in specific scenarios. Furthermore, parameter adjustment methods based on historical experience or simple optimization fail to effectively quantify the risk of generation deficit (CVaR) under future extreme scenarios and lack sensitivity identification of adjustable parameters. This makes it difficult to accurately identify the controlling factors in the multi-dimensional parameter space, and fails to simultaneously maximize overall benefits while minimizing the adjustment range of dispatch rules. Summary of the Invention

[0005] Objective of the invention: To provide a method and system for intelligent scheduling decision support of cascade hydropower stations, in order to solve the aforementioned problems existing in the prior art.

[0006] Technical solution: A method for intelligent scheduling and decision support of cascade hydropower stations, comprising the following steps:

[0007] Historical meteorological and hydrological data and future meteorological forecast data of cascade reservoir basins are obtained. Based on the correlation characteristics between pre-stored historical meteorological elements and runoff prediction deviations under different meteorological combination quadrants, quadrant deviation correction is performed on future predicted runoff to generate runoff input sequences containing deterministic predicted runoff and random runoff scenario sets.

[0008] Based on the preset existing cascade scheduling rules, multi-scenario parallel hydraulic simulation is performed on the random runoff scenario set to calculate the cascade power generation index and the risk value of power generation deficit conditions, and the adaptability of the existing cascade scheduling rules is determined based on the time-varying trend of the cascade power generation index.

[0009] When it is determined that the adaptability of the current cascade scheduling rules is at risk of deterioration, the current cascade scheduling rules are analyzed to extract adjustable parameters to construct the feasible domain for parameter adjustment. The main control parameters in the adjustable parameters are identified by combining the results of multi-scenario simulation. With the dual guidance of maximizing the comprehensive benefits of the cascade and minimizing the parameter adjustment range, optimized scheduling parameter suggestions are generated.

[0010] According to one aspect of this application, generating deterministic runoff predictions specifically includes:

[0011] Extract monthly meteorological elements from historical meteorological and hydrological data, including at least temperature, precipitation, and evaporation;

[0012] Calculate the 50th percentile of the historical long-term series of monthly meteorological elements as the benchmark for distinguishing the level of each element;

[0013] Based on the relationship between the monthly meteorological element values ​​and the high / low discrimination benchmark, a three-dimensional combination classification is carried out, and historical months are classified into 8 meteorological combination quadrants defined by the three-dimensional combination of temperature-precipitation-evaporation.

[0014] For each meteorological combination quadrant, establish a deviation correction formula that links the meteorological elements of each month with the runoff forecast deviation;

[0015] The prediction bias is calculated by substituting future meteorological forecast data into the deviation correction formula corresponding to the corresponding meteorological combination quadrant, and then superimposing the prediction bias onto the pre-stored climate model predicted runoff to obtain the deterministic predicted runoff.

[0016] According to one aspect of this application, a set of stochastic runoff scenarios is generated, specifically including:

[0017] Collect historical residual sequences of corrected predicted runoff and measured runoff for historical periods, and calculate the first-order autocorrelation coefficient ρ1 of the historical residual sequences;

[0018] A random residual path is constructed by sampling residual values ​​from the historical residual sequence using random sampling with replacement, and a first-order autoregressive filter is applied to the random residual path.

[0019] ε j ''=ρ1ε j-1 ''+sqrt(1-ρ1 2 )ε j ';

[0020] Where, ε j '' represents the residual for the j-th time period after filtering, ε j ' is the original random residual obtained from the sampling of the j-th time period;

[0021] The filtered, predetermined random residual paths are superimposed onto the deterministic predicted runoff to form a set of random runoff scenarios.

[0022] According to one aspect of this application, calculating the value at risk of power generation deficit conditions includes:

[0023] Based on the pre-defined time-period recursive water balance state transition equations for cascade reservoirs, parallel hydraulic simulations under multiple scenarios are conducted. For a set of stochastic runoff scenarios, the average annual cascade power generation (HE) is calculated. m As an indicator of cascaded power generation; calculate HE m The average annual power generation is HE*.

[0024] Annual power generation based on pre-stored cascade design HE design Calculate the power generation deficit depth L for each simulated year. m :

[0025] L m = max(0, HE) design - HE m );

[0026] Arrange the power generation deficit depths in ascending order for M simulated years, set a confidence level α, and calculate the value at risk (CVaR) of the power generation deficit condition. α :

[0027] CVaR α =(1 / (Mm α ))∑ j=mα+1 M L (j) ;

[0028] Where m α For the critical order determined based on the confidence level, L (j) The j-th power generation deficit depth after sorting.

[0029] According to one aspect of this application, the adaptability of the current cascade dispatching rules is determined based on the time-varying trend of cascade power generation indicators, specifically including:

[0030] The simulation period is divided into predetermined statistical stages, and the average power generation of each stage is calculated to form a statistical time series {x}. i};

[0031] Calculate the first-order autocorrelation coefficient ρ1 of the statistical time series. When |ρ1| exceeds a preset autocorrelation threshold, apply pre-whitening processing to the statistical time series to generate the residual series y. i =x i -ρ1x i-1 ;

[0032] The Mann-Kendall trend test method is used to calculate the rank statistic S and variance Var(S) based on the residual series, and a standardized trend test statistic Z is constructed:

[0033] When S>0, Z=(S-1) / sqrt(Var(S)); when S<0, Z=(S+1) / sqrt(Var(S));

[0034] If the standardized trend test statistic Z is less than the preset negative threshold, it is determined that the cascade power generation index has a downward trend and the adaptability of the current cascade dispatch rules is at risk of deterioration.

[0035] According to one aspect of this application, parsing existing cascade scheduling rules to extract adjustable parameters to construct a feasible domain for parameter adjustment specifically includes:

[0036] Extract adjustable parameters that affect the power generation of the cascade from the existing cascade dispatch rules, including at least the start date of post-flood water storage, the completion date of post-flood water storage, the target water storage level after the flood season, and the minimum control water level during the dry season.

[0037] Based on the preset reservoir safety operation constraints, the range of adjustable parameters is defined, and the feasible domain Ω for parameter adjustment is constructed.

[0038] The Latin hypercube sampling method is used to generate a parameter sample set within the feasible region of parameter adjustment: for the i-th parameter, its value range is divided into N... s The sampling formula for a subinterval with equal probability in the j-th layer is:

[0039] X i，j =X i，min +(j-1)δ i +u i，j δ i ;

[0040] Where X i，min δ is the lower limit of the parameter. i u is the layer width. i，j The sample points are uniformly random numbers within the interval [0, 1]. By randomly permuting and combining the sampling points of each parameter, a sample of N is formed. s The parameter sample set of the group parameter vector.

[0041] According to one aspect of this application, identifying the master control parameter among the adjustable parameters includes:

[0042] Using a parameter sample set of adjustable parameters as input, parallel hydraulic simulations are performed on a set of random runoff scenarios to calculate the cascade comprehensive benefit function value U(X) corresponding to each set of parameter vectors. j Construct a scheduling parameter-target response dataset;

[0043] The first-order sensitivity index S of each adjustable parameter is calculated using the variance decomposition method based on conditional expectation. i :

[0044] S i =(1 / V(U))∑ b=1 nbin (n i，b / N s (E) i，b -U*) 2 ;

[0045] Where V(U) is the total variance of the comprehensive benefit function, U* is the sample mean, and n bin To divide the interval number, E i，b Let n be the conditional mean of the comprehensive benefit function when the i-th parameter falls into the b-th interval; i，b This represents the number of samples where the i-th parameter falls into the b-th interval;

[0046] The first-order sensitivity index S i Adjustable parameters exceeding the preset sensitivity threshold are identified as main control parameters, and the remaining parameters are fixed to the values ​​in the current cascade scheduling rules.

[0047] According to one aspect of this application, it also includes comprehensively combining the discrete levels of the main control parameters and constructing a cascade comprehensive benefit and power generation deficit risk response matrix, specifically:

[0048] The value ranges of the main control parameters are discretized to obtain the discrete levels of each main control parameter; the discrete levels of all main control parameters are then comprehensively combined to generate N. comb A comprehensive parameter vector is generated; using this comprehensive parameter vector as input, parallel hydraulic simulations are performed on a set of random runoff scenarios to obtain the cascade comprehensive benefit function value U(X) under each combination. c Value at Risk (CVaR) for Power Generation Deficit Conditions α，c Using the discrete level of the main control parameters as the index dimension, the cascade comprehensive benefit response matrix U and the power generation deficit risk response matrix CVaR are calculated.

[0049] According to one aspect of this application, optimized scheduling parameter recommendations are generated with the dual orientation of maximizing the comprehensive benefits of the cascade and minimizing the parameter adjustment range, including:

[0050] Set an acceptable threshold for power generation deficit risk r thBased on the power generation deficit risk response matrix CVaR, we screen for those that meet the CVaR. α，c ≤r th The combinations constitute a feasible set of combinations Φ;

[0051] Find the combination c* within the feasible combination set Φ that maximizes the tiered comprehensive benefit function:

[0052] The recommended values ​​for the master control parameters corresponding to the combination c* are the scheduling parameters that maximize the overall benefits of the cascade.

[0053] Calculate the standardized adjustment distance D of each combination in the feasible combination set Φ relative to the current cascade scheduling rule values. c :

[0054] Set a threshold for improving efficiency ΔU min Within the feasible combination set Φ, filter those satisfying U(X) c )-U(X c0 )≥ΔU min The combinations constitute the candidate combination set Ψ, where U(X) c0 This represents the comprehensive benefits of the current cascade scheduling rules;

[0055] Find the combination c** with the smallest standardized adjustment distance in the candidate combination set Ψ, and take the master control parameter value corresponding to combination c** as the scheduling parameter suggestion with the smallest parameter adjustment range.

[0056] According to one aspect of this application, a smart dispatching decision support system for cascade reservoir power stations includes:

[0057] The predicted runoff correction module is used to acquire historical meteorological and hydrological data and future meteorological prediction data of the cascade reservoir basin. Based on the correlation characteristics between pre-stored historical meteorological elements and runoff prediction deviations under different meteorological combination quadrants, a quadrant deviation correction model is constructed. Based on this, the predicted runoff is corrected for quadrant deviations, and a runoff input sequence containing deterministic predicted runoff and random runoff scenario sets is generated.

[0058] The scheduling rule response evaluation module is used to perform multi-scenario parallel hydraulic simulation on a random runoff scenario set based on the preset existing scheduling rules of the cascade, calculate the cascade power generation index and the risk value of power generation deficit conditions, and determine the adaptability of the existing scheduling rules of the cascade based on the time-varying trend of the cascade power generation index.

[0059] The parameter adaptation and adjustment module is used to respond to the risk of degradation in the adaptability of the current scheduling rules of the cascade. It analyzes the current scheduling rules of the cascade to extract adjustable parameters to construct the feasible domain of parameter adjustment. It identifies the main control parameters among the adjustable parameters by combining the simulation results of multiple scenarios. With the dual guidance of maximizing the comprehensive benefits of the cascade and minimizing the parameter adjustment range, it generates optimized scheduling parameter suggestions.

[0060] Beneficial effects: This invention improves runoff prediction accuracy through meteorological element classification and multinomial bias correction, and generates a set of random scenarios considering temporal correlation, providing a reliable hydrological basis for scheduling decisions under hydrological changes; it quantifies the exposure of power generation deficit risk under extreme scenarios using the Conditional Value at Risk (CVaR) method, and combines trend testing to determine the risk of adaptive degradation of advance scheduling rules, achieving forward-looking risk identification and assessment; based on sensitivity analysis, it identifies cascade master control scheduling parameters and constructs a parameter-target response matrix, maximizing comprehensive benefits or minimizing parameter adjustment amplitude under the constraint of power generation deficit risk, providing decision support for adaptive adjustment of scheduling rules; and it realizes risk identification and adaptive optimization of scheduling rules under hydrological change environments. Attached Figure Description

[0061] Figure 1 A flowchart illustrating the steps of an intelligent scheduling decision support method for cascade reservoir power stations provided in this application embodiment.

[0062] Figure 2 A comparison chart of the meteorological quadrant deviation correction effect provided in the embodiments of this application.

[0063] Figure 3 A comparison chart of the overall benefits before and after parameter adjustment provided in the embodiments of this application.

[0064] Figure 4 A comparison chart of power generation deficit risk before and after parameter adjustment provided in the embodiments of this application. Detailed Implementation

[0065] Solving the above problems requires a three-pronged approach. First, improve the accuracy of predicted runoff. Based on the relationship between meteorological elements such as temperature, precipitation, and evaporation and quantile benchmarks, historical months are categorized into meteorological quadrants, and polynomial deviation correction formulas are established for each quadrant. Quadrant-specific corrections are applied to future predicted runoff to generate deterministic predicted runoff and a set of stochastic runoff scenarios. Second, establish a dispatch rule response evaluation mechanism. Establish a cascade reservoir water balance state transition equation, use parallel hydraulic simulation technology to handle multi-scenario dispatch simulations, and output the cascade power generation response sequence based on cascade dispatch rules and adjustable parameters. The Conditional Value at Risk (CVaR) method is used to quantify the risk exposure of power generation deficits under extreme scenarios. The Mann-Kendall trend test method is used to analyze the time-varying characteristics of cascade power generation and determine the risk of degradation in the adaptability of the current dispatch rules. Third, establish methods for adapting and adjusting operating parameters. The system analyzes the scheduling rules to extract adjustable parameters and constructs a feasible region for parameter adjustment. It uses Latin hypercube sampling to generate a parameter sample set. Combined with parallel hydraulic simulation, it constructs a parameter-target response dataset and uses variance decomposition to identify the main control parameters. It comprehensively combines the discrete levels of the main control parameters to construct a cascade comprehensive benefit and power generation deficit risk response matrix. Under the constraint of power generation deficit risk, it outputs two types of scheduling parameter adjustment suggestions: one with the maximum comprehensive benefit and one with the minimum adjustment range.

[0066] Current research has made progress in three areas: prediction correction, rule evaluation, and parameter optimization, but each has its shortcomings. Prediction correction lacks quadrant bias correction based on three-dimensional combination classification of meteorological elements; rule evaluation lacks Mann-Kendall trend test and conditional value at risk (CVaR) quantification methods; and parameter optimization lacks methods for identifying key control parameters, comprehensive combination analysis, and multi-objective decision-making. More importantly, there is currently no mature solution for linking the three modules of predicted runoff correction, dispatch rule response evaluation, and operational parameter adaptive adjustment in a data information flow to form a complete decision support system. Therefore, it is necessary to establish an adaptive optimization system for dispatch parameters of cascade reservoir hydropower stations to provide technical support for improving the scientific, forward-looking, and adaptive nature of dispatch decisions.

[0067] like Figure 1 As shown, a method for intelligent scheduling decision support of cascade reservoir power stations includes the following steps:

[0068] Historical meteorological and hydrological data and future meteorological forecast data of cascade reservoir basins are obtained. Based on the correlation characteristics between pre-stored historical meteorological elements and runoff prediction deviations under different meteorological combination quadrants, quadrant deviation correction is performed on future predicted runoff to generate runoff input sequences containing deterministic predicted runoff and random runoff scenario sets.

[0069] Based on the preset existing cascade scheduling rules, multi-scenario parallel hydraulic simulation is performed on the random runoff scenario set to calculate the cascade power generation index and the risk value of power generation deficit conditions, and the adaptability of the existing cascade scheduling rules is determined based on the time-varying trend of the cascade power generation index.

[0070] When it is determined that the adaptability of the current cascade scheduling rules is at risk of deterioration, the current cascade scheduling rules are analyzed to extract adjustable parameters to construct the feasible domain for parameter adjustment. The main control parameters in the adjustable parameters are identified by combining the results of multi-scenario simulation. With the dual guidance of maximizing the comprehensive benefits of the cascade and minimizing the parameter adjustment range, optimized scheduling parameter suggestions are generated.

[0071] The present invention will be further described below with reference to specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of this application can be combined with each other.

[0072] This embodiment focuses on a cascade hydropower system in a river basin, consisting of an upstream reservoir and a downstream reservoir connected in series. The upstream reservoir controls a basin area of ​​156,000 km². 2 Total reservoir capacity: 6.25 billion cubic meters 3 The normal water level is 1158.0m, the dead water level is 1145.0m, the flood control limit water level is 1147.0m, the installed capacity is 2800MW, and the designed annual power generation is 12.28 billion kWh. The downstream reservoir is located 86km downstream of the upstream reservoir, with a total storage capacity of 1.83 billion m³. 3 The normal water level is 1068.5m, the dead water level is 1063.0m, the flood control limit water level is 1065.5m, the installed capacity is 1200MW, and the designed annual power generation is 5.42 billion kWh. The cascade design annual power generation is 17.7 billion kWh. The drainage area between the two reservoirs is 8200 km². 2 .

[0073] The upstream reservoir is an annual regulating reservoir, primarily responsible for power generation, flood control, and water supply. The downstream reservoir is a daily regulating reservoir, primarily responsible for power generation and counter-regulation. The current cascade operation rules are as follows: during the flood season (June 1st to September 30th), water levels are controlled; during the post-flood storage period (October 1st to October 31st), water levels are gradually raised to the target storage level; and during the drawdown period (November 1st to May 31st of the following year), water levels are controlled to decrease. The adjustable parameters in the current cascade operation rules are: for the upstream reservoir, the post-flood storage start date is October 1st, the completion date is October 31st, the target storage level is 1157.0m, and the minimum control level during the dry season is 1152.0m; for the downstream reservoir, the post-flood storage start date is October 8th, the completion date is October 31st, the target storage level is 1068.0m, and the minimum control level during the dry season is 1065.5m.

[0074] I. Implementation of the predicted runoff correction module.

[0075] Extract 360 months of historical meteorological and hydrological data from the upstream reservoir: temperature (T), precipitation (P), evaporation (E), and measured monthly average inflow (Qobs). Calculate the 50th quantile of the data: T. 50 =15.2℃, P 50 =98mm, E 50 =106mm.

[0076] A three-dimensional combination classification is performed based on the magnitude relationship between meteorological elements and the baseline. Temperature T ≥ 15.2℃ is classified as high temperature (H), otherwise as low temperature (L); precipitation P ≥ 98mm is classified as rainy (W), otherwise as dry (D); evaporation E ≥ 106mm is classified as high evaporation (H), otherwise as low evaporation (L). The 360 ​​historical months are categorized into 8 meteorological combination quadrants: Quadrant 1 (HWH) 48 months, Quadrant 2 (HWL) 42 months, Quadrant 3 (HDH) 38 months, Quadrant 4 (HDL) 46 months, Quadrant 5 (LWH) 41 months, Quadrant 6 (LWL) 49 months, Quadrant 7 (LDH) 37 months, and Quadrant 8 (LDL) 59 months.

[0077] A bias correction formula was established for each quadrant. Since the sample size is less than 100, a simplified first-order polynomial ΔQ = a0 + a1T + a2P + a3E was used. Leave-one-out cross-validation was employed to determine the coefficients. Figure 2 As shown, taking quadrant 1 as an example, after 48 rounds of cross-validation, the Nash efficiency coefficient NSE is calculated to be 0.67, which meets the criterion of NSE ≥ 0.5. The cross-validation NSEs for the other quadrants are as follows: Quadrant 2: 0.63, Quadrant 3: 0.53, Quadrant 4: 0.66, Quadrant 5: 0.47, Quadrant 6: 0.72, Quadrant 7: 0.49, and Quadrant 8: 0.74. Among them, Quadrant 5 (NSE = 0.47, sample size 41) and Quadrant 7 (NSE = 0.49, sample size 37) did not reach the validity criterion threshold of 0.5. The bias correction relationship for these two small sample quadrants is set to 0, and the original climate model predicted runoff is used directly. The remaining 6 quadrants all meet the validity criteria. The statistical significance test of NSE showed that the NSE values ​​of quadrants 1, 2, 4, 6, and 8 were significantly better than the no-skill prediction (NSE=0) at the 95% significance level. However, although the NSE value of quadrant 3 (NSE=0.53, sample size 38) exceeded the 0.5 threshold numerically, its statistical significance test p-value was 0.08, which was marginally significant, indicating that there was some uncertainty in the bias correction effect of this quadrant.

[0078] We obtained the upstream reservoir forecast meteorological elements for 1200 months over the next 100 years, averaged from five models in the CMIP6 climate model ensemble. For each of the next 1200 months, we determined the quadrant based on the relationship between the meteorological element values ​​and historical baselines, and calculated the runoff bias using the corresponding quadrant's correction formula. For months in quadrants 5 and 7 (approximately 22% of the total months), no bias correction was applied, and the original climate model runoff forecasts were used directly. For months in quadrant 3 (approximately 15%), bias correction was applied, but a larger uncertainty range was retained. This yielded the deterministic runoff forecasts.

[0079] The residual sequences of corrected predicted and measured runoff were collected from six significant and effective quadrants (quadrants 1, 2, 4, 6, and 8, totaling 244 months) over a historical period of 360 months. The first-order autocorrelation coefficient of the residuals was calculated to be 0.44. 1200 residual values ​​were extracted from the historical residual sequences using random sampling with replacement, and a first-order autoregressive filter was applied to the random residuals. This sampling process was repeated 500 times to obtain 500 random residual paths. These paths were then superimposed onto the deterministic predicted runoff to form a set of 500 random runoff scenarios.

[0080] II. Implementation of the scheduling rule response evaluation module.

[0081] The system receives 500 sets of random runoff scenarios, each containing runoff sequences over 1200 months. The upstream reservoir's Δt is calculated using a daily step, while the downstream reservoir's Δt is calculated using a 6-hour step. A recursive water balance state transition equation V is established for each reservoir for each time period. t+1 = V t + (Q in，t - Q gen，t - Q sup，t - Q sp，t )Δt, V t Q represents the initial reservoir capacity during the time period. in，t Q represents the inbound flow during time period t. gen，t Q sup，t Q sp，t These represent the components of power generation water diversion flow, water supply flow, and overflow discharge flow during time period t, where Δt is the calculation time step; ecological base current Q eco，t The reservoir discharge flow rate is the rigid lower limit for time period t, and the river discharge flow rate is the power generation water diversion flow rate Q. gen，t With the overflow discharge Q sp，t The sum. The ecological base flow of the upstream reservoir is taken as 2600m. 3 / s, the reservoir's annual water supply is 1.82 billion m³. 3 The downstream reservoir's ecological base flow is taken from 3200m. 3 / s.

[0082] Parallel hydraulic simulation technology was employed, with sequential calculations over time, upstream-downstream reservoir calculations in the cascade topology dimension, and simultaneous parallel calculations for 500 scenarios. The average annual power generation of the cascade under these 500 scenarios was calculated over the next 100 years, yielding a future average annual power generation (HE*) of 17.46 billion kWh, a decrease of 240 million kWh (-1.4%) compared to the designed annual power generation of 17.7 billion kWh. This overall decrease in power generation is primarily due to the predicted overall decrease in runoff under future climate scenarios and the increased water wastage losses caused by the intensified uneven spatial and temporal distribution of inflows.

[0083] The Conditional Value at Risk (CVaR) method was used to assess the generation deficit risk. Based on the cascade design annual generation of 17.7 billion kWh, the generation deficit depth for each year was calculated. The deficit depths over 100 years were sorted in ascending order, and with a confidence level of α=0.9, the Conditional Value at Risk (CvaR) for generation deficit was calculated. 0.9 =880 million kWh, representing the average power generation deficit depth of 880 million kWh in the worst 10% scenario (i.e., the worst 10 years) at a 90% confidence level, accounting for 5.0% of the designed annual power generation. The power generation deficit depth distribution exhibits a right-skewed characteristic (skewness coefficient 0.48), indicating that the deficit depth in extreme drought years is significantly higher than in generally unfavorable years, and the CVaR index effectively captures the tail extreme risks. Further analysis shows that in years with a deficit depth exceeding 1 billion kWh, 70% are accompanied by extreme weather events (such as severe drought lasting more than 3 consecutive months).

[0084] The Mann-Kendall trend test method was used to evaluate the time-varying characteristics of cascade power generation. The next 1200 months were divided into 40 statistical periods (30 months per period), and the mean of cascade power generation in each period across 500 scenarios was calculated to form a statistical time series. The first-order autocorrelation coefficient of the series was 0.44, exceeding a preset autocorrelation threshold, indicating significant autocorrelation. Pre-whitening was applied to the series to generate residual sequences. Based on the residual sequences, the rank statistic S = -217, variance Var(S) = 7333.33, and standardized trend test statistic Z = -2.52 were calculated. A significance level of 0.05 and a critical value Z were set. critical=±1.96. Since Z=-2.52<-1.96 and Z<0, a significant downward trend in the cascade power generation sequence is determined (significance level p=0.012, two-tailed test), indicating that the current cascade dispatching rules face the risk of adaptive degradation. Robustness analysis of the trend test results shows that using different statistical stage division methods (e.g., 30 stages, 50 stages), a consistent conclusion of significant decline is obtained, indicating that the trend determination result is robust and reliable. The temporal evolution characteristics of power generation show a pattern of decline amid fluctuations, reflecting the superimposed effect of the long-term cumulative impact of climate change on the water resource system and the short-term drastic fluctuations caused by the increased frequency of extreme events.

[0085] III. Implementation of the operating parameter adaptation and adjustment module.

[0086] The current cascade scheduling rules were analyzed to extract adjustable parameters affecting cascade power generation. Adjustable parameters for upstream reservoirs include: post-flood storage start date, completion date, target post-flood storage level, and minimum control water level during the dry season. Adjustable parameters for downstream reservoirs include: post-flood storage start date, completion date, target post-flood storage level, and minimum control water level during the dry season. A total of eight adjustable parameters are included.

[0087] The physical constraints for parameter adjustment are defined based on reservoir design specifications and safe operation requirements. An improved Latin hypercube sampling method is used to generate a uniformly distributed parameter sample set within the feasible region. The sampling size Ns = 500 is set, and the maximum-minimum distance criterion is employed to optimize the sample space distribution, ensuring sufficient and uniform coverage of the parameter space. The value intervals of each parameter within the feasible region are uniformly divided into 500 equally probable sub-intervals, and sample points are randomly selected within each sub-interval. Independent random permutations are performed on the 500 representative values ​​of each parameter. The elements at the same position in the permuted sequences of the eight parameters are combined one by one to form 500 sets of 8-dimensional scheduling parameter vectors. This sampling scheme achieves a uniform coverage density of 2.37 layers / dimensional in the 8-dimensional parameter space. Combined with spatial optimization using the maximum-minimum distance criterion, the minimum Euclidean distance between sample points is not less than 0.15 normalized units, ensuring the sample representativeness and reliability of the global sensitivity analysis results.

[0088] Using 500 sets of scheduling parameter vectors as adjustable parameters, parallel computation is performed on 500 sets of random runoff. For the j-th set of parameter vectors, the future average annual power generation HE* of the cascade is calculated. j Value at Risk (CVaR) for Power Generation Deficit Conditions α，j The average annual power generation of the cascade project is HE*. j The distribution ranges from 17.18 billion kWh to 17.83 billion kWh; CVaR at a confidence level of α=0.9 α，j The distribution ranges from 610 million kWh to 1.19 billion kWh.

[0089] Establish a tiered comprehensive benefit function U(X) j )=ω1×M×HE* j - ω2×(M-(α×M)) ×CVaR α，j =0.7·100·HE* j -0.3(100-90) CVaR α，j =70HE* j -3CVaR α，j The decision weight coefficients are ω1=0.70 and ω2=0.30. This objective function comprehensively considers the expected power generation benefit (the first term, which is dominant) and the penalty for extreme risks (the second term, reflecting risk aversion). The weight coefficient ω1=0.70 reflects the primary pursuit of maximizing power generation benefit, and ω2=0.30 reflects the risk management requirements for a 10% extremely unfavorable scenario. The ratio of the two is approximately 7:3, reflecting a balanced consideration of benefit and risk. The cascade comprehensive benefit function was calculated for 500 sets of scheduling parameter vectors, forming a scheduling parameter-target response dataset.

[0090] The main control parameters are identified using a variance decomposition method based on conditional expectation. For the i-th scheduling parameter, its value range is evenly divided into 10 intervals. The conditional mean of the comprehensive benefit function in each interval is calculated, and then the first-order sensitivity index S is calculated. i A sensitivity threshold of Sth = 0.08 was set. Sensitivity indices for eight adjustable parameters were calculated, with S3 ≥ 0.08, S4 ≥ 0.08, and S7 ≥ 0.08. Three of these were determined to be the main control parameters: the target water level after the flood season in the upstream reservoir, the minimum control water level during the dry season in the upstream reservoir, and the target water level after the flood season in the downstream reservoir. The remaining five parameters were fixed to the values ​​set by the current scheduling rules.

[0091] The discrete levels of the main control parameters are comprehensively combined. The target water level after the flood season in the upstream reservoir is discrete into 5 levels (0.5m interval), the minimum control water level during the dry season is discrete into 4 levels, and the target water level after the flood season in the downstream reservoir is discrete into 3 levels. The total number of comprehensive combinations is N. comb =5×4×3=60. The current dispatching rule corresponds to combination number c0=45, with the corresponding parameter value being: the post-flood target water level Z of the upstream reservoir. target， 上游 = 1157.0m, the lowest control water level Z during the dry season of the upstream reservoir low， 上游 =1152.0m, the target water level Z of the downstream reservoir after the flood season. target， 下游 = 1068.0m.

[0092] Using 60 sets of comprehensive combined parameter vectors as adjustable parameters, calculations were performed for 500 sets of stochastic runoff scenarios. The future average annual power generation and the risk value of power generation deficit conditions for each cascade under each combination were calculated, and then the comprehensive benefit function of the cascade was calculated. Using the discrete level of the main control parameters as the index dimension, a comprehensive benefit response matrix U∈R for the cascade was constructed. 5×4×3 and the power generation deficit risk response matrix CVaR∈R 5×4×3 Interaction analysis of the response matrix shows a significant synergistic effect between the two main control parameters of the upstream reservoir: when the target water level decreases after the flood season, the reduction of the minimum control water level during the dry season can further enhance the risk avoidance effect while having a relatively small impact on power generation.

[0093] This module utilizes a 128-core CPU cluster for parallel computing, achieving a parallel efficiency of approximately 72% and a total computation time of approximately 68 minutes. The average computation time per month is approximately 1.1 milliseconds, encompassing the complete hydraulic simulation process, including solving the water balance equation, water level and reservoir capacity interpolation, power generation calculation, and constraint checks.

[0094] Two calculation results were obtained based on the response matrix:

[0095] (1) The scheduling parameters that maximize the overall benefits of the cascade.

[0096] Set an acceptable CVaR risk threshold r for power generation deficit th =850 million kWh. Within the feasible combination set (36 feasible combinations, accounting for 60% of all combinations), the combination number c*=27 that maximizes the comprehensive benefit function corresponds to the main control parameter value: Z. target， 上游 = 1158.0m, Z low， 上游 = 1152.0m, Z target， 下游 = 1068.5m. For example... Figure 3 and Figure 4 As shown, the optimal combination corresponds to an average annual power generation of 17.62 billion kWh, and the power generation deficit risk CVaR* 0.9 =800 million kWh, comprehensive benefit U*=70×176.2-3×8.0=12334 - 24.0 = 1231 billion kWh.

[0097] Compared to the current dispatch rules' average annual power generation of 17.46 billion kWh, the power generation deficit risk CVaR 0.9，c0 =880 million kWh, comprehensive benefit U c0 =70×174.6 - 3×8.8 = 12222-26.4 = 12195.6 billion kWh. The optimal combination achieves: an increase of 160 million kWh in average annual power generation, a reduction of 80 million kWh in power generation deficit risk, and an increase of 114.4 billion kWh in overall benefits.

[0098] (2) The scheduling parameter with the smallest adjustment range.

[0099] Set a threshold for improving efficiency ΔU min =3 billion kWh. In the candidate combination set Ψ (there are 18 candidate combinations), find the combination number c**=33 with the smallest adjustment distance, and the corresponding master control parameter value is: Z. target， 上游 = 1157.5m, Z low， 上游 =1152.0m, Z target， 下游 = 1068.0m. The minimum adjustment range combination corresponds to an average annual power generation of 17.5 billion kWh, and a power generation deficit risk CVaR**. 0.9 =860 million kWh, comprehensive benefit U** = 70×175.0 - 3×8.6 = 12224.2 billion kWh, corresponding parameter adjustment distance D** = sqrt((0.5 / 2.0) 2 +0 2 +0 2 =0.25.

[0100] Compared to the current dispatch rules, the minimum adjustment combination achieves: an average annual increase in power generation of 0.4 billion kWh, a reduction in power generation deficit risk of 0.2 billion kWh, and an overall benefit increase of 2.86 billion kWh.

[0101] Comparative analysis of the two schemes shows that: Scheme c* aims to maximize overall benefits with a larger adjustment range, making it suitable for policymakers who are proactive in responding to future climate change; Scheme c** aims to minimize the adjustment range with a more conservative approach, making it suitable for policymakers who prefer gradual adjustments. A common feature of both schemes is the moderate alteration of the target post-flood storage level of the upstream reservoir of the main reservoir, reflecting a shift in dispatching strategies under future climate change scenarios.

[0102] According to one aspect of this application, an embodiment is provided to describe how to achieve the dual optimization objectives of maximizing comprehensive benefits and minimizing parameter adjustment range in cascade reservoir scheduling by constructing a quantitative mathematical model. This embodiment solves the technical problems of traditional experience-based scheduling, which struggles to balance long-term power generation benefits with extreme risks and lacks quantitative basis for parameter adjustment.

[0103] Step 1001: Based on the preset tiered comprehensive benefit function, calculate the tiered comprehensive benefit function value U(X) for each combination. c Value at Risk (CVaR) for Power Generation Deficit Conditions α，c .

[0104] In this embodiment, the design of the tiered comprehensive benefit function is a core link connecting the physical scheduling process and economic risk decision-making. To balance long-term economic expectations with risk aversion under extreme weather conditions, this function is constructed as a linear weighted combination model. Specifically, the tiered comprehensive benefit function U(X) j The calculation formula for ) is as follows:

[0105] U(X j )=ω1 ·M·HE* j - ω2 ·(M-┌αM┐) CVaR α，j ;

[0106] In the formula, X j HE* represents the scheduling parameter vector of the j-th group; j CVaR represents the average future annual power generation of the cascade system calculated based on K sets of stochastic runoff scenarios during the simulation period of M years. Its physical unit is 100 million kilowatt-hours, representing the expected power generation revenue of the system under average hydrological conditions. α，j This represents the value of risk of power generation deficit conditions at confidence level α, used to quantify the average depth of power generation deficit in the most unfavorable (1-α) proportion of years. This term is introduced into the objective function as a risk penalty term.

[0107] Furthermore, ω1 and ω2 in the formula are decision weight coefficients, satisfying the constraint ω1 + ω2 = 1. These two coefficients reflect the decision-maker's preference for pursuing returns versus avoiding risks. For example, in this embodiment, ω1 is set to 0.70 and ω2 to 0.30, indicating that the decision model, while pursuing maximum power generation, assigns approximately 30% weight to extreme risk control. The symbol ┌αM┐ represents rounding up αM, used to determine the critical order of risk calculation. The system can uniformly transform evaluation indicators with different dimensions and physical meanings into a single comprehensive benefit value, supporting subsequent optimal ranking.

[0108] Step 1002: Calculate the standardized adjustment distance D of each combination in the feasible combination set Φ relative to the value of the current cascade scheduling rule. c .

[0109] Specifically, after determining the optimal solution, to assess the feasibility of parameter adjustments, it is necessary to calculate the degree of deviation of the new solution from the current situation. Since different scheduling parameters have different physical dimensions—for example, the unit for the post-flood target water level is meters, while the unit for the post-flood water storage start date is days—directly calculating the Euclidean distance lacks physical meaning. Therefore, this embodiment uses the range method to perform dimensionless processing of the parameters. Standardized adjustment distance D c The calculation formula is as follows:

[0110] D c =sqrt(∑i=1 dkey ((X key，i (lc，i) -X key，i (m0，i) )⁄(X key，i ni -X key，i 1 )) 2 );

[0111] Where d key This indicates the number of main control parameters. Numerator X key，i (lc，i) -X key，i (m0，i) This represents the difference between the value of the i-th master control parameter in the c-th combination and the value of the same parameter in the current cascade scheduling rules. The denominator term X key，i ni -X key，i 1 This represents the difference between the maximum and minimum values ​​of the i-th master control parameter across all discrete levels, i.e., the range of values ​​for that parameter.

[0112] By dividing by the range, the variation range of each parameter is uniformly mapped to the interval between 0 and 1. The summation of squares and the square root operation in the outer layer of the formula constitutes the standard Euclidean distance metric in the multidimensional parameter space. Distance D c The smaller the value, the closer the optimization plan is to the existing rules, the smaller the impact of implementing the optimization plan on the existing operation and management model, and the easier it is for the scheduling and operation department to accept it.

[0113] According to one aspect of this application, in the steps of performing multi-scenario parallel hydraulic simulation of a set of random runoff scenarios, the topological calculation logic of the cascade reservoir group and the specific flow extrapolation rules for different scheduling periods are as follows:

[0114] Step S201: Based on the preset existing cascade scheduling rules, perform multi-scenario parallel hydraulic simulation on the random runoff scenario set.

[0115] In this embodiment, a parallel computing strategy based on topological layering is adopted for the hydraulic connections of a cascade reservoir group. Specifically, the system first performs topological sorting on the cascade reservoirs. Taking a certain cascade in the watershed involved in this embodiment as an example, the upstream reservoir is located upstream and is defined as layer 1; the downstream reservoir is located downstream and is defined as layer 2.

[0116] In the simulation calculation process, a strict temporal recursion is followed in the time dimension, while parallel processing is performed in the spatial and scenario dimensions. The specific process is as follows: At the beginning of any calculation time period t, the computing cluster utilizes 128 CPU cores to perform parallel calculations of the water balance equation of the first-level upstream reservoir for 500 sets of random runoff scenarios, obtaining the outflow from the upstream reservoir. Using the river evolution model, the outflow from the upstream reservoir is calculated through interval confluence and used as the inflow input for the second-level downstream reservoir. Based on this, the system again performs parallel calculations of the water balance of the second-level downstream reservoir for 500 sets of scenarios. After completing all levels of calculations for the current time period t, the system state is updated to time period t+1, and the above process is repeated. This effectively solves the computational dependency problem caused by upstream and downstream hydraulic coupling, achieving a parallel efficiency of approximately 72% in this embodiment, and controlling the total calculation time to around 68 minutes.

[0117] Step S202: Determine the values ​​of the flow components based on the tiered scheduling rules and adjustable parameters.

[0118] Specifically, in the above simulation process, after determining the boundary conditions such as inflow, water supply, and ecological base flow for each specific time period, the core difficulty lies in determining the power generation flow rate Q. gen and the discharge flow rate Q sp This embodiment uses the following logic for refined calculation based on the scheduling period to which the time segment belongs:

[0119] Regarding ecological baseflow constraints, regardless of the scheduling period, the system will always use ecological baseflow Q. eco This serves as a rigid lower limit constraint on the reservoir's outflow from the river channel at various time periods. The outflow from the river channel consists of the sum of the power generation flow and the water discharge.

[0120] During the flood season scheduling phase, when the reservoir water level is less than or equal to the flood control limit, the system uses an iterative trial calculation method to determine the power generation flow. The specific iterative logic is as follows: the initial value is determined by the surplus after deducting the water supply flow from the inflow, and is constrained by the unit's rated flow and minimum technical flow. The reservoir water level at the end of the time period is calculated based on the water balance principle. The system iteratively adjusts the power generation flow using a bisection method or a successive approximation method, based on the principle that the reservoir water level at the end of the time period is closest to but does not exceed the flood control limit. If the calculated power generation flow is less than the ecological base flow, the power generation flow is adjusted to the ecological base flow value. When the reservoir water level exceeds the flood control limit, the power generation flow is directly taken as the unit's rated flow, and the water discharge flow is calculated backwards based on the water balance equation to force the reservoir water level at the end of the time period to fall below the flood control limit.

[0121] During the post-flood storage period, which begins on the start date of storage and ends on the completion date, the system calculates the average storage flow required to complete the storage task based on the target reservoir capacity corresponding to the target storage level, the initial reservoir capacity for the current period, and the remaining storage periods. After combining the inflow flow with the subtracted supply flow, the planned discharge flow for this period is determined. When the planned discharge flow is less than the ecological base flow, the system forcibly increases the planned discharge flow to the ecological base flow level. In this case, the power generation flow is taken as the value closest to the planned discharge flow within the range of the unit's rated flow and minimum technical flow, while the water discharge is taken as the difference between the planned discharge flow and the power generation flow.

[0122] During the drawdown period, when the reservoir water level is higher than the minimum control water level for the dry season, the power generation flow is determined by the surplus after deducting the water supply flow from the inflow, and is also subject to unit capacity constraints. When the reservoir water level is equal to or lower than the minimum control water level for the dry season, the system determines to stop power generation and releases ecological base flow by opening the flood discharge facilities, while prioritizing the water supply flow demand.

[0123] According to one aspect of this application, the first-order sensitivity index S i Adjustable parameters exceeding a preset sensitivity threshold are identified as primary control parameters, specifically:

[0124] In this embodiment, based on the generated scheduling parameters and target response dataset, the first-order sensitivity index S of eight adjustable parameters was calculated. i Based on this, the main control parameters were selected. The specific calculation results are as follows: the sensitivity index of the upstream reservoir's post-flood target water level is 0.31, the sensitivity index of the upstream reservoir's minimum control water level during the dry season is 0.18, and the sensitivity index of the downstream reservoir's post-flood target water level is 0.12. The sensitivity indices of these three parameters all exceed the preset threshold of 0.08, and therefore they are determined to be the main control parameters.

[0125] In contrast, other parameters have a smaller impact. For example, the sensitivity index for the start date of post-flood water storage in the upstream reservoir is 0.02, while the sensitivity index for the start date of post-flood water storage in the downstream reservoir is 0.01. These parameters are determined to be non-sensitive and are fixed to the values ​​in the existing scheduling rules in subsequent optimization steps, reducing the dimensionality of the optimization problem.

[0126] According to another aspect of this application, a smart dispatching decision support system for cascade reservoir hydropower stations is also provided. This system includes a predicted runoff correction module, a dispatching rule response evaluation module, and an operating parameter adaptation adjustment module, all connected in series according to data information flow. This system adopts a modular, serial architecture design, with each module transmitting information through standardized data interfaces, forming a complete decision chain of prediction, evaluation, and optimization. The serial data information flow ensures that the output of the upstream module becomes the input of the downstream module, achieving seamless information connection. Specifically, the predicted runoff correction module outputs deterministic predicted runoff and stochastic runoff scenario sets as input to the dispatching rule response evaluation module; the dispatching rule response evaluation module outputs an adaptive degradation risk assessment conclusion, triggering the operation of the parameter adaptation adjustment module; and the operating parameter adaptation adjustment module outputs optimized dispatching parameter suggestions. This serial architecture ensures both the functional independence and maintainability of each module while achieving synergy and intelligence in the overall decision-making process.

[0127] The predicted runoff correction module extracts historical monthly meteorological elements and calculates quantile benchmarks. It categorizes historical months into meteorological quadrants using a three-dimensional combination of temperature, precipitation, and evaporation. For each quadrant, it establishes polynomial bias correction formulas and employs leave-one-out cross-validation to calibrate coefficients. Quadrant-based correction is then applied to future predicted runoff to generate deterministic predicted runoff. Historical residual sequences are collected, and random sampling and autoregressive filtering are used to generate a set of random runoff scenarios. This module addresses the inconsistency between meteorological drivers and hydrological responses under different climatic conditions. Quadrant-based modeling significantly improves the runoff fit between flood and dry seasons. The generation of random scenarios maintains the temporal structure of historical residuals, enabling subsequent scheduling simulations to cover extreme scenarios.

[0128] The dispatch rule response assessment module establishes the water balance state transition equations for cascade reservoirs. Based on the cascade dispatch rules and adjustable parameters, it employs parallel hydraulic simulation technology to conduct multi-scenario dispatch simulations and outputs the cascade power generation response sequence as a benefit-generating indicator. The Conditional Value at Risk (CVaR) method is used to quantify the power generation deficit risk exposure under extreme scenarios, and the Mann-Kendall trend test method is used to analyze the time-varying characteristics of cascade power generation to determine the risk of degradation in the adaptability of the current dispatch rules. This module integrates multi-scenario simulation, extreme risk quantification, and trend stability testing into a single evaluation framework, enabling early identification of risks before the rules show significant failure, providing dispatching departments with a basis for proactive adjustments.

[0129] The parameter adaptation and adjustment module analyzes scheduling rules to extract adjustable parameters and constructs a feasible region for parameter adjustment. It uses Latin hypercube sampling to generate a parameter sample set and combines parallel hydraulic simulation to construct a parameter-target response dataset. Variance decomposition is used to identify the main control parameters. The discrete levels of the main control parameters are comprehensively combined to construct a cascade comprehensive benefit and power generation deficit risk response matrix. Under the constraint of power generation deficit risk, the module calculates and outputs scheduling parameter adjustment suggestions that maximize comprehensive benefits and minimize adjustment range. By reducing optimization dimensions through sensitivity identification, avoiding local optima through comprehensive combination, and ensuring the feasibility of the scheduling scheme through risk constraints, the module forms optimal parameter suggestions that both improve power generation efficiency and control risk.

[0130] Furthermore, in the predicted runoff correction module: monthly meteorological element sequences for historical periods of the reservoir basin are extracted, including temperature T (degrees Celsius), precipitation P (millimeters), and evaporation E (millimeters). These three factors jointly control the runoff generation and concentration process in the basin and are key meteorological driving factors determining runoff formation. The 50th percentile T of each element is calculated using historical data from the past 30 years. 50 P 50 E 50 As a benchmark for distinguishing between high and low values, the 30-year period conforms to the climate standard period recommended by the World Meteorological Organization (WMO). The 50th percentile (median) is insensitive to extreme values ​​and can robustly represent the long-term average state of meteorological elements. A three-dimensional combination classification is performed based on the magnitude relationship between monthly meteorological elements and the benchmark, when T ≥ T... 50 If the temperature is high, it is considered high; otherwise, it is considered low. Precipitation and evaporation are determined using the same rules. Historical months are thus categorized into eight meteorological quadrants, with the sample size for the q-th meteorological quadrant denoted as N. q The eight quadrants correspond to all combinations of high and low temperatures, high and low precipitation, and high and low evaporation, which can finely depict the runoff deviation characteristics under different meteorological patterns.

[0131] For the q-th meteorological quadrant, extract the climate model-predicted runoff Q within that quadrant. pre Compared with measured runoff Q obs Based on historical data, establish a bias correction formula. When the sample size N... q When the value is ≥100, a polynomial is used:

[0132] ΔQ = a0 + a1 T + a2 P + a3 E + a4 T 2 +a5 P 2 +a6 E 2 +a7 TP +a8 TE +a9 PE;

[0133] When the number of samples N q When <100, use a simplified linear polynomial:

[0134] ΔQ = a0 + a1 T + a2 P + a3 E;

[0135] Where ΔQ is the runoff prediction bias (unit: m). 3 / s), a i These are the undetermined coefficients. Simplifying the model reduces the number of undetermined parameters, adhering to the principle that model complexity should match the amount of data information, and avoiding overfitting on small samples.

[0136] Leave-one-out cross-validation is used to evaluate the generalization ability of the bias correction relation: N in the quadrant q Each month was used as a validation sample, and the remaining (N) samples were used each time. q -1) samples were used as the training set, and the undetermined coefficients a of the polynomial were calibrated using the least squares method. i Make the objective function ∑(Q obs -Q pre -ΔQ) 2 To minimize this, substitute the meteorological elements of the validation sample into the established relationship and calculate the corrected runoff Q. corr，t =Q pre，t + ΔQ t .

[0137] After traversing all months in the quadrant, summarize N. q Calculate the cross-validation Nash efficiency coefficient based on the prediction results of each validation sample:

[0138] NSE=1-(∑ t=1 Nq (Q obs，t -Q corr，t ) 2 ) / (∑ t=1 Nq (Q obs，t -Q* obs ) 2 );

[0139] Where Q* obs This represents the average measured runoff within each quadrant.

[0140] A bias-corrected model in a quadrant is considered valid when NSE ≥ 0.5. This threshold is a commonly used standard for acceptable performance in hydrological model evaluation, meaning the model explains more than 50% of the data variability. After successful cross-validation, all NSE values ​​within the quadrant are used. q The polynomial coefficients a of the sample recalibration are... i This yields the final deviation correction formula for practical applications in this quadrant. If NSE < 0.5, the deviation correction formula is set to 0. When accuracy requirements are not met, the original predicted values ​​are used directly, reflecting the principle of robustness.

[0141] For the predicted future M years (12M months), the quadrant to which it belongs is determined based on the relationship between its meteorological element values ​​and historical baselines. The deviation ΔQ is then calculated by substituting these values ​​into the correction formula for that quadrant. fut Deterministic runoff predictions were obtained:

[0142] Q det =Q pre，fut +ΔQ fut ;

[0143] Q pre，fut To predict runoff for uncorrected future climate models.

[0144] Collect the residual sequences ε of corrected predicted runoff and measured runoff for each month in historical periods. t =Q obs，t -Q corr，t Calculate the first-order autocorrelation coefficient of the residuals:

[0145] ρ1=(∑ t=1 N-1 (ε t -μ ε )(ε t+1 -μ ε )) / (∑ t=1 N (ε t -μ ε ) 2 );

[0146] Where μ ε Let be the mean of the residuals, and N be the length of the historical residual sequence. The autocorrelation coefficient measures the continuity of residuals between adjacent time periods, reflecting the memory effect of runoff processes.

[0147] Using random sampling with replacement, 12M residual values ​​are drawn from the historical residual sequence to form a random residual path {ε1', ε2', ..., ε 12M Apply a first-order autoregressive filter to the random residuals to preserve time-series correlation:

[0148] ε1''=ε1';

[0149] ε j ''=ρ1ε j-1 ''+sqrt(1-ρ1 2 )ε j '; j=2,3,...,12M;

[0150] Where ε1'' is the first value of the random residual path after filtering, ε j-1 '' represents the value of the previous time period of the filtered random residual path, ε j ' represents the original residual obtained from random sampling, ε j'' represents the j-th value of the random residual path after filtering. In the filtering formula, sqrt(1-ρ1) 2 The coefficients ensure that the variance of the filtered residual sequence is consistent with that of the original sequence, avoiding artificial amplification or reduction of the variance.

[0151] The sampling process is repeated K times to obtain K random residual paths. These random residual paths are then superimposed onto the deterministic predicted runoff to form a set of random runoff scenarios with a length of M years, or 12M months.

[0152] Q stoc，k，j =Q det，j +ε k，j '';k=1,2,...,K;j=1,2,...,12M;

[0153] Q det，j For the deterministic prediction of runoff in month j, ε k，j '' represents the j-th filtered residual of the k-th random path. The number of scenarios K is typically between 100 and 1000, which covers typical scenarios and some extreme scenarios while maintaining computational feasibility.

[0154] In the scheduling rule response evaluation module: it receives K sets of random runoff scenario sets from the predicted runoff correction module, each set containing a runoff sequence of 12M months. The time step for calculating the reservoir's water balance state transition is determined based on the reservoir's regulation performance: a ten-day or daily step is used for multi-year and seasonal regulation reservoirs, while an hourly step is used for daily regulation and runoff-type reservoirs. The time step selection follows the principle of time scale matching and is adapted to the reservoir's regulation cycle.

[0155] For the recursive water balance state transition equations for the establishment period of each reservoir in the cascade:

[0156] V t+1 = V t + (Q in，t - Q gen，t - Q sup，t - Q sp，t )Δt;

[0157] Where V t The initial reservoir capacity at time t (unit: m³) 3 ), Q in Inbound flow rate (unit: m³) 3 / s), Q gen Water diversion flow rate for power generation (unit: m³) 3 / s), Q sup The water supply flow rate of the reservoir to the water demanders (unit: m³) 3 / s), Q sp Overflow discharge (unit: m³) 3 / s), Δt is the calculation time step (unit: seconds); ecological base current Q eco，t The reservoir discharge flow rate is the rigid lower limit for time period t, and the river discharge flow rate is the power generation water diversion flow rate Q. gen，t With the overflow discharge Q sp，t The sum of these equations. The water balance state transition equation embodies the discretized form of the continuity equation in the reservoir system, reflecting the conservation relationship between inflow, storage, use, and discharge. Reservoir capacity and water level are converted using the water level-capacity relationship curve. The values ​​of each flow component in the equation are determined according to the cascade scheduling rules and adjustable parameters, specifically as follows:

[0158] Ecological base current Q eco The design ecological flow value is taken as the rigid lower limit constraint for the reservoir's outflow from the river channel (the sum of power generation flow and wastewater discharge) at each time period. Water supply flow Q sup Take the water demand flow rate of the water-consuming entity during this period. Power generation flow rate Q gen and the discharge flow rate Q sp The specific timeframes are determined based on the scheduling period to which they belong. During the scheduling process in each period, it is necessary to ensure that the sum of the power generation flow and the water discharge flow is not less than the ecological base flow requirement; otherwise, the difference will be made up by increasing the water discharge flow.

[0159] During the flood season, when the reservoir water level is less than or equal to the flood limit level, the power generation flow is iteratively calculated: the initial value is determined by the surplus after deducting the water supply flow from the inflow, constrained by the rated flow and minimum technical flow of the generating units, and the reservoir water level at the end of the period is calculated based on the water balance principle; the power generation flow is iteratively calculated based on the principle that the reservoir water level at the end of the period is closest to but does not exceed the flood limit level. If the calculated power generation flow is less than the ecological base flow, the power generation flow is adjusted to the ecological base flow. When the reservoir water level exceeds the flood limit level, the power generation flow is taken as the rated flow of the generating units, and the water discharge flow is determined according to the water balance to control the reservoir water level at the end of the period from not exceeding the flood limit level. At this time, the total outflow from the river channel usually automatically meets the ecological base flow requirements.

[0160] The post-flood storage period begins on the start date of storage and ends on the completion date of storage. Based on the target reservoir capacity corresponding to the target post-flood storage level, the current reservoir capacity, and the remaining storage periods, the average storage flow required to complete the storage task is calculated. This average flow is then combined with the inflow, minus the supply flow, to determine the planned discharge flow for that period. When the planned discharge flow is less than the ecological base flow, the planned discharge flow is forcibly increased to the ecological base flow. The power generation flow is taken as the value closest to the planned discharge flow within the range of the unit's rated flow and minimum technical flow. The water discharge is taken as the difference between the planned discharge flow and the power generation flow.

[0161] During the drawdown period, when the reservoir water level is higher than the minimum control water level for the dry season, the power generation flow is determined by the surplus after deducting the water supply flow from the inflow, and is constrained by the unit's rated flow and minimum technical flow. When the power generation flow is less than the ecological base flow, the water discharge flow is the difference between the two; otherwise, the water discharge flow is zero. When the reservoir water level is equal to or lower than the minimum control water level for the dry season, power generation is stopped, and the ecological base flow is released through water discharge, while simultaneously meeting the water supply flow requirements.

[0162] Cascade power generation output P at different times gen The calculation formula (unit: kW) is as follows:

[0163] P gen，t =9.81η comp Q gen，t H net，t ;

[0164] Where η comp For the overall efficiency of the hydro-generator unit, H net Clean water head (unit: m):

[0165] H net，t =(Z up，t +Z up，t+1 ) / 2-Z tail，t -H loss ;

[0166] Z up，t Z represents the initial reservoir water level at time t (unit: m). tail The tailwater level (unit: m) is determined by looking up the flow-tailwater level relationship curve based on the total discharge flow. H loss Head loss (unit: m).

[0167] The module employs parallel hydraulic simulation technology to handle multi-scenario cascade reservoir power station scheduling simulations. It maintains causal relationships through sequential recursion in the time dimension, and performs topological sorting and layering based on hydraulic connections in the cascade topology dimension, with parallel calculations within the same layer and sequential calculations across different layers. A topological sorting algorithm is used to stratify the cascade reservoirs, ensuring that inflows to reservoirs within the same layer are independent and can be calculated in parallel. Reservoirs in different layers are calculated layer by layer according to topological order, ensuring that the outflow from upstream reservoirs is calculated first as the inflow to downstream reservoirs. In the scenario dimension, K sets of scenarios are calculated synchronously and in parallel. The scenarios are completely independent and can be massively parallelized on multi-core CPUs or GPUs. For each calculation period, the following calculations are performed in parallel for the K sets of scenarios: determination of flow components based on cascade scheduling rules and adjustable parameters, water balance state transition, reservoir capacity-level conversion, flow tail-level conversion, and power generation output calculation, outputting the response sequence of cascade power generation as a benefit indicator. The average annual cascade power generation HE for the K sets of scenarios is statistically calculated year by year for the next M years. m m = 1, 2, ..., M (unit: 100 million kWh), and HE mThe average future annual power generation of the cascade is calculated as HE*.

[0168] The module employs the Conditional Value at Risk (CVaR) method to assess the risk value of cascade power generation under deficit conditions. CVaR is an advanced method for quantifying extreme risks, providing a more comprehensive characterization of tail losses compared to traditional VaR. This is based on the cascade design annual power generation (HE). design Calculate the depth of power generation deficit for each year:

[0169] L m =max(0, HE) design -HE m ), m=1,2,...,M.

[0170] The deficit depths of year M are sorted in ascending order to obtain L. (1) ≤L (2) ≤...≤L (M) Set a confidence level α and determine the critical sequence m. α =┌αM┐, where ┌·┐ is rounded up, used to calculate the Value at Risk (CVaR) for the power generation deficit condition. α :

[0171] CVaR α =1 / (Mm α )∑ j=mα+1 M L (j) ;

[0172] CVaR α Characterize the exposure of dispatch rules and adjustable parameters to power generation deficit risk under low-probability, high-impact events at confidence level α.

[0173] The module employs the Mann-Kendall trend test to evaluate the time-varying characteristics of the cascade power generation response sequence. The Mann-Kendall test requires no assumptions regarding data distribution. Adjustable parameters are extracted based on the current cascade scheduling rules, and parallel hydraulic simulation technology is invoked to perform parallel calculations on K groups of random runoff. The simulated next 12 months are divided into n equal parts. period In each statistical stage, the mean of the cascade power generation in each stage is calculated under K scenarios, forming a statistical time series {x1, x2, ..., x...}. nperiod}

[0174] Calculate the first-order autocorrelation coefficient of the sequence:

[0175] ρ1=(∑ i=1 nperiod-1 (x i -x*)(x i+1 -x*)) / (∑ i=1 nperiod (x i -x*)2 );

[0176] Where x* is the sequence mean. When |ρ1| exceeds a preset autocorrelation threshold, the autocorrelation is considered significant, and pre-whitening is applied to the sequence to generate the residual sequence y. i =x i -ρ1 x i-1 (i=2,3,...,n) period Pre-whitening treatment eliminates the interference of autocorrelation on trend testing and avoids an increase in the false positive rate.

[0177] Calculate the rank statistic:

[0178] S=∑ i=1 nperiod -1 ∑ j=i+1 nperiod sgn(xy j -xy i );

[0179] Where xy represents the original sequence elements when autocorrelation is insignificant, or the residual sequence elements when autocorrelation is significant; the sign function sgn(xy) j -xy i ) is: when xy j -xy i When xy > 0, it takes the value 1. j -xy i When xy = 0, the value is 0. j -xy i When the value is less than 0, it takes the value -1.

[0180] Calculate the variance of the statistic:

[0181] Var(S) = (n period (n period -1)(2n period +5)-∑ p=1 g t p (t p -1)(2t p +5)) / 18;

[0182] Where g is the number of repeated numerical groups in the sequence, t p g is the number of elements in the p-th group of repeated values; g = 0 when all elements in the sequence are distinct.

[0183] Constructing standardized trend test statistics:

[0184] Z={(S-1) / sqrt(Var (S)), S>0; 0, S=0; (S+1) / sqrt(Var (S)), S<0};

[0185] The standardized statistic Z asymptotically follows a standard normal distribution under the null hypothesis. When |Z| exceeds the critical value Z0... critical Furthermore, when Z < 0, it is determined that the cascade power generation sequence of the beneficial indicators shows a significant downward trend, indicating that the current cascade dispatching rules face the risk of adaptive degradation. The dispatching rule response evaluation module outputs the judgment conclusion of the adaptive degradation risk of the current cascade dispatching rules.

[0186] In the operating parameter adaptation and adjustment module: the current scheduling rules of the cascade are parsed to extract adjustable parameter values ​​that affect the beneficial indicators of cascade power generation, including the post-flood storage start date t of each reservoir. start Post-flood water storage completion date t end Post-flood target water storage level Z target Minimum control water level during the dry season Z low These four types of parameters control the seasonal regulation process of the reservoir and are key decision variables for power generation scheduling, forming a cascade adjustable parameter vector X=[X1, X2, ..., X...]. d ], where d is the number of parameters.

[0187] Based on reservoir design specifications and safe operation requirements, define the physical constraints for parameter adjustment:

[0188] Z dead <Z target ≤Z normal ;

[0189] Z dead <Z low <Z normal ;

[0190] Z normal At the normal water storage level, Z dead This is the dead water level.

[0191] Post-flood water storage time is subject to an upper limit t end，max and lower limit t start，min Constraints:

[0192] t start，min ≤t start <t end ≤t end，max ;

[0193] The constraints described above reflect the rigid requirements of flood control safety, water supply and power generation needs, and temporal logic. A multidimensional feasible region Ω is constructed by integrating all constraints for parameter adjustment. Within this feasible region, each parameter is discretized into multiple levels. A Latin hypercube sampling method is used to generate a uniformly distributed parameter sample set within the feasible region, ensuring uniform coverage of the parameter space while saving computational resources. Compared to simple random sampling, Latin hypercube sampling has better space-filling properties, allowing for reliable estimates of the parameter-target relationship with fewer samples. The sampling size N is set. s The range of values ​​for each parameter within the feasible region [X] i，min X i，max (i=1,2,...,d) are uniformly divided into N s There are 1 equally probable subintervals, with a layer width of δ. i =(X i，max -X i，min ) / N s Randomly select a sample point within the j-th sub-interval:

[0194] X i，j =X i，min +(j-1)δ i +u i，j δ i ;

[0195] Where u i，j The numbers are uniformly random numbers in the interval [0, 1], j = 1, 2, ..., N. s For each parameter N s Each of the d representative values ​​is subjected to an independent random permutation. The elements at the same position in the sequence after the d parameters are permuted are combined one by one to form N. s A group of d-dimensional scheduling parameter vectors {X1, X2, ..., X...} Ns}, where the j-th parameter vector X j =[X 1，j X 2，j , ..., X d，j It is composed of the j-th element of each parameter permutation sequence.

[0196] N s The group scheduling parameter vector is used as an adjustable parameter, and parallel hydraulic simulation technology is invoked to perform parallel calculations for K groups of random runoff. For the j-th group parameter vector X j The average annual cascade power generation HE for K scenarios is calculated year by year over the next M years. m，j m = 1, 2, ..., M (unit: 100 million kWh), and HE for M years m，j The average future annual power generation of the cascade is calculated as HE*. j Invoke the Conditional Value at Risk (CVaR) method to calculate the j-th parameter vector X. jBelow, the Value at Risk (CVaR) of Cascaded Power Generation Deficit Conditions α，j .

[0197] Establish a cascade comprehensive benefit function U(·), which comprehensively considers power generation benefits and power generation deficit penalties:

[0198] U(X j )=ω1 M·HE* j -ω2 (M-┌αM┐) CVaR α，j ;

[0199] This function uses a linear weighted form to normalize and combine power generation benefits and risk penalties. Here, ω1 and ω2 are decision weight coefficients, satisfying ∑ i=1 2 ω i =1.

[0200] For N s The group scheduling parameter vector completes the calculation of the tiered comprehensive benefit function, forming a scheduling parameter-target response dataset {(X j ,U(X j) |j=1,2,...,N s The variance decomposition method based on conditional expectation is used to identify the main control parameters among the adjustable parameters of the tiered system. The variance decomposition method originates from global sensitivity analysis theory, decomposing the variance of the objective function into the contributions of each input parameter. Based on the scheduling parameter-target response dataset, the sample mean U* and total variance V(U) of the tiered comprehensive benefit function are calculated:

[0201] U*=1 / N s ∑ j=1 Ns U(X j V(U) = 1 / N s ∑ j=1 Ns (U(X j )-U*) 2 ;

[0202] For the i-th scheduling parameter X i (i=1, 2, ..., d), let its value range be [X i，min X i，max Divide evenly into n bin There are n intervals, and the b-th interval is denoted as I. i，b (b=1, 2, ..., n) bin ). Statistics N s The value of the i-th parameter in the group parameter vector sample is X. i，j The number of samples n falling into the b-th interval i，b Calculate the conditional mean of the tiered comprehensive benefit function within this interval:

[0203] E i，b =(1 / n i，b )∑ j:Xi，j∈Ii，b U(X j );

[0204] Where X i，j Let X be the parameter vector of the j-th group. j The value of the i-th scheduling parameter.

[0205] Based on the degree of dispersion of the conditional mean, the quantization parameter X i The contribution of the first-order sensitivity index to the variance of the tiered comprehensive benefit function is calculated as follows:

[0206] S i =1 / V(U)∑ b=1 nbin (n i，b / N s (E) i，b -U*) 2 ;

[0207] S i The larger the value, the more likely the parameter X is to be larger. i The more significant the impact on the objective function, the greater the sensitivity threshold S. th When S i ≥S th Time determination of adjustable parameter X i The main control parameter, when S i th Parameters that are deemed non-sensitive are fixed to their values ​​according to the current scheduling rules. A subset of the main control parameters is obtained after filtering, and the number of main control parameters is denoted as d. key Let them be denoted as X. key，i (i=1,2,...,d) key The discrete level number of each master control parameter is denoted as n. i .

[0208] A comprehensive combination of the discrete levels of the master control parameters is performed. Let the i-th master control parameter X... key，i The discrete levels are arranged in ascending order as X key，i (1) X key，i (2) , ..., X key，i (ni) The current scheduling rule takes the value of the m-th node. 0，i There are several levels, denoted as X. key，i (m0，i) The Cartesian product method is used to generate a full range of combinations of the master control parameters, with a total number of combinations N. comb for:

[0209] N comb =∏ i=1 ​dkey n i ;

[0210] Where ∏ is the Cartesian product of the discrete levels of the main control parameters. The complete set of combined parameter vectors is denoted as {X}. c |c=1,2,...,N comb}, where the horizontal index vector of the main control parameters of the c-th combination is (l c，1 , l c，2 , ..., l c，dkey The non-sensitive parameters are taken from the current scheduling rule values. The combination number corresponding to the current scheduling rule is denoted as c0. N... comb Using a comprehensive combination parameter vector as adjustable parameters, parallel hydraulic simulation technology is employed to perform calculations for K sets of stochastic runoff scenarios, yielding the future average annual power generation HE*c (c=1, 2, ..., N) of each cascade under each combination. comb The Conditional Value at Risk (CVaR) method is used to calculate the Conditional Value at Risk (CVaR) for each combination of power generation deficits. α，c Calculate the tiered comprehensive benefit function for each combination:

[0211] U(X c )=ω1 M·HE*c-ω2 (M-┌αM┐) CVaR α，c ;

[0212] Where ω1 and ω2 are decision weight coefficients, satisfying ∑ i=1 2 ω i =1.

[0213] After completing all combined calculations, a tiered comprehensive benefit response matrix is ​​constructed using the discrete level of the main control parameter as the index dimension: U∈R n1×n2×...×ndkey and the power generation deficit risk response matrix CVaR∈R n1×n2×...×ndkey .

[0214] Two calculation results were obtained based on the response matrix U and CVaR:

[0215] (1) The scheduling parameters that maximize the overall benefits of the cascade system:

[0216] Define a feasible set of combinations with the constraint that the risk of power generation deficit is acceptable:

[0217] Φ={c∣CVaR α，c ≤r th c = 1, 2, ..., N comb};

[0218] Where r th The acceptable CVaR risk threshold for power generation deficit.

[0219] Find the combination number within the feasible combination set Φ that yields the maximum value of the tiered comprehensive benefit function:

[0220] c*=argmax c∈Φ U(X c );

[0221] The optimal combination corresponds to the following main control parameter values, denoted as X*=(X key，1 (lc*，1) X key，2 (lc*，2) , ..., X key，dkey (lc*，dkey) The corresponding comprehensive benefit is denoted as U*=U(X). c* The corresponding power generation deficit risk is denoted as CVaR. α *=CVaR α，c* .

[0222] (2) The scheduling parameter with the smallest adjustment range:

[0223] The range method is used to standardize each master control parameter, and the standardized adjustment distance of each combination relative to the current scheduling rule value is calculated:

[0224] D c =sqrt(∑ i=1 dkey ((X key，i (lc，i) -X key，i (m0，i) )⁄(X key，i ni -X key，i 1 )) 2 );

[0225] Standardized adjustment distances eliminate the influence of different parameter units and numerical ranges. A benefit improvement threshold ΔU is set. min Within the feasible set of combinations Φ, further filter candidate combinations that offer improved overall benefits compared to the current situation:

[0226] Ψ={c∣c∈Φ,U(X c )-U(X c0 )≥ΔU min};

[0227] Find the combination number with the smallest adjustment distance in the candidate combination set Ψ:

[0228] c**=argmin c∈Ψ D c ;

[0229] The values ​​of the main control parameters corresponding to the minimum adjustment range combination are denoted as:

[0230] X**=(X key，1 (lc**，1) X key，2 (lc**，2) , ..., X key，dkey (lc**，dkey) The corresponding comprehensive benefit is denoted as U**=U(X). c** The corresponding power generation deficit risk is denoted as CVaR. α **=CVaR α，c** The corresponding parameter adjustment distance is denoted as D**=D c** .

[0231] The operating parameter adaptation module outputs two calculation results: the scheduling parameter value X* that maximizes the cascade comprehensive benefit function, and the corresponding comprehensive benefit U* and power generation deficit risk CVaR. α *; The minimum adjustment value of the dispatch parameter X** and its corresponding comprehensive benefit U**, and the risk of power generation deficit CVAR. α **Scheduling parameter adjustment distance D**. The two types of suggestions respectively satisfy the decision-making needs of maximizing benefits and minimizing adjustments, providing decision-makers with different risk preferences with choice.

[0232] According to one aspect of this application, a multi-scenario parallel hydraulic simulation is performed on a set of random runoff scenarios. Specifically, in the time dimension, the causal relationship is maintained by sequential recursion. In the cascade topology dimension, topological sorting and layering are performed according to hydraulic connections, and parallel calculations are performed in the same layer but sequential calculations in different layers to ensure that the outflow from the upstream reservoir is calculated first as the inflow from the downstream reservoir. In the scenario dimension, each group of scenarios is calculated synchronously and in parallel. For each calculation period, the following calculations are performed in parallel for each group of scenarios: determination of flow components based on cascade scheduling rules and adjustable parameters, water balance state transition, reservoir capacity level conversion, flow tail level conversion, and power generation output calculation, outputting the response sequence of cascade power generation as a beneficial indicator.

[0233] According to one aspect of this application, determining flow components based on cascade scheduling rules and adjustable parameters specifically includes:

[0234] Ecological baseflow constraint: Ecological baseflow Q eco，t The reservoir discharge flow rate is the rigid lower limit for time period t, and the river discharge flow rate is the power generation water diversion flow rate Q. gen，t With the overflow discharge Q sp，t The sum of the power generation water diversion flow and the flood discharge flow is less than the ecological base flow; when the sum of the power generation water diversion flow and the flood discharge flow is less than the ecological base flow, the difference is made up by increasing the flood discharge flow.

[0235] Flood season dispatching rules: During the flood season, when the reservoir water level is less than or equal to the flood limit water level, the power generation water diversion flow is iteratively calculated based on the principle that the reservoir water level at the end of the period is closest to and does not exceed the flood limit water level. The power generation water diversion flow is constrained by the upper limit of the rated flow of the generating unit and the lower limit of the minimum technical flow. When the calculated power generation water diversion flow is less than the ecological base flow, the power generation water diversion flow is adjusted to the ecological base flow. When the reservoir water level exceeds the flood limit water level, the power generation water diversion flow is taken as the rated flow of the generating unit, and the overflow water discharge flow is determined according to the water balance to control the reservoir water level at the end of the period from not exceeding the flood limit water level.

[0236] Post-flood water storage period scheduling rules: The post-flood water storage period begins on the post-flood water storage start date and ends on the post-flood water storage completion date; based on the target reservoir capacity corresponding to the post-flood target water level, the current reservoir capacity, and the number of remaining water storage periods, the average water storage flow required to complete the water storage task is calculated; the planned discharge flow for this period is determined by deducting the water supply flow from the inflow flow; when the planned discharge flow is less than the ecological base flow, the planned discharge flow will be forcibly increased to the ecological base flow; the power generation water diversion flow is taken as the value closest to the planned discharge flow within the range of the unit's rated flow and minimum technical flow; the flood discharge discharge flow is taken as the difference between the planned discharge flow and the power generation water diversion flow.

[0237] Drawdown period scheduling rules: During the drawdown period, when the reservoir water level is higher than the minimum control water level during the dry season, the power generation water diversion flow is determined by the surplus after deducting the water supply flow from the inflow, and is subject to the rated flow and minimum technical flow of the generating unit; when the power generation water diversion flow is less than the ecological base flow, the overflow water discharge flow is the difference between the two, otherwise the overflow water discharge flow is zero; when the reservoir water level is equal to or lower than the minimum control water level during the dry season, power generation is stopped, and the ecological base flow is released through overflow water discharge to meet the water supply flow requirements.

[0238] According to another aspect of this application, generating deterministic runoff predictions can also be:

[0239] Count the number of historical samples N within the q-th meteorological combination quadrant. q ;

[0240] When N q When the value is ≥100, a polynomial containing quadratic and interaction terms is used to establish the deviation correction relationship, and the formula is as follows:

[0241] ΔQ = a0 + a1T + a2P + a3E + a4T 2 +a5P 2 +a6E 2 +a7TP+a8TE+a9PE;

[0242] When N q When <100, a simplified first-order polynomial is used to establish the deviation correction relationship, and the formula is:

[0243] ΔQ = a0 + a1T + a2P + a3E;

[0244] Where ΔQ is the runoff prediction bias, T, P, and E are temperature, precipitation, and evaporation, respectively, and a i These are coefficients to be determined;

[0245] The undetermined coefficients are calibrated using the leave-one-out method for cross-validation. The cross-validation Nash efficiency coefficient NSE is calculated based on the calibration results. When NSE ≥ 0.5, the deviation correction formula for the meteorological combination quadrant is deemed valid. When NSE < 0.5, the deviation correction formula is set to 0.

[0246] The prediction bias is calculated by substituting future meteorological forecast data into the deviation correction formula corresponding to the corresponding meteorological combination quadrant, and then superimposing the prediction bias onto the pre-stored climate model predicted runoff to obtain the deterministic predicted runoff.

[0247] 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 intelligent scheduling and decision support of cascade reservoir power stations, characterized in that, Includes the following steps: Historical meteorological and hydrological data and future meteorological forecast data of cascade reservoir basins are obtained. Based on the correlation characteristics between pre-stored historical meteorological elements and runoff prediction deviations under different meteorological combination quadrants, quadrant deviation correction is performed on future predicted runoff to generate runoff input sequences containing deterministic predicted runoff and random runoff scenario sets. Based on the preset existing cascade scheduling rules, multi-scenario parallel hydraulic simulation is performed on the random runoff scenario set to calculate the cascade power generation index and the risk value of power generation deficit conditions, and the adaptability of the existing cascade scheduling rules is determined based on the time-varying trend of the cascade power generation index. When it is determined that the adaptability of the current cascade scheduling rules is at risk of deterioration, the current cascade scheduling rules are analyzed to extract adjustable parameters to construct the feasible domain for parameter adjustment. The main control parameters in the adjustable parameters are identified by combining the results of multi-scenario simulation. With the dual guidance of maximizing the comprehensive benefits of the cascade and minimizing the parameter adjustment range, optimized scheduling parameter suggestions are generated.

2. The method according to claim 1, characterized in that, Generating deterministic runoff predictions specifically includes: Extract monthly meteorological elements from historical meteorological and hydrological data, including at least temperature, precipitation, and evaporation; Calculate the 50th percentile of the historical long-term series of monthly meteorological elements as the benchmark for distinguishing the level of each element; Based on the relationship between the monthly meteorological element values ​​and the high / low discrimination benchmark, a three-dimensional combination classification is carried out, and historical months are classified into 8 meteorological combination quadrants defined by the three-dimensional combination of temperature-precipitation-evaporation. For each meteorological combination quadrant, establish a deviation correction formula that links the meteorological elements of each month with the runoff forecast deviation; The prediction bias is calculated by substituting future meteorological forecast data into the deviation correction formula corresponding to the corresponding meteorological combination quadrant, and then superimposing the prediction bias onto the pre-stored climate model predicted runoff to obtain the deterministic predicted runoff.

3. The method according to claim 2, characterized in that, Generate a set of random runoff scenarios, specifically including: Collect historical residual sequences of corrected predicted runoff and measured runoff for historical periods, and calculate the first-order autocorrelation coefficient ρ1 of the historical residual sequences; A random residual path is constructed by sampling residual values ​​from the historical residual sequence using random sampling with replacement, and a first-order autoregressive filter is applied to the random residual path. e j ''=p1e j-1 ''+sqrt(1-ρ1 2 )e j '; Where, ε j '' represents the residual for the j-th time period after filtering, ε j ' is the original random residual obtained from the sampling of the j-th time period; The filtered, predetermined random residual paths are superimposed onto the deterministic predicted runoff to form a set of random runoff scenarios.

4. The method according to claim 1, characterized in that, Calculating the value at risk of a power generation deficit condition includes: Based on the pre-defined time-period recursive water balance state transition equations for cascade reservoirs, parallel hydraulic simulations under multiple scenarios are conducted. For a set of stochastic runoff scenarios, the average annual cascade power generation (HE) is calculated. m As an indicator of cascaded power generation; calculate HE m The average annual power generation HE*; based on pre-stored cascade design, the annual power generation HE design Calculate the power generation deficit depth L for each simulated year. m : L m = max(0,HE design - HE m ); Arrange the power generation deficit depths in ascending order for M simulated years, set a confidence level α, and calculate the value at risk (CVaR) of the power generation deficit condition. α : CVARs α =(1 / (Mm α ))∑ j=mα+1 M IT (j) ; Where m α For the critical order determined based on the confidence level, L (j) The j-th power generation deficit depth after sorting.

5. The method according to claim 1, characterized in that, The adaptability of the current dispatch rules for cascade power generation is determined based on the time-varying trend of cascade power generation indicators, specifically including: The simulation period is divided into predetermined statistical stages, and the average power generation of each stage is calculated to form a statistical time series {x}. i }; Calculate the first-order autocorrelation coefficient ρ1 of the statistical time series. When |ρ1| exceeds a preset autocorrelation threshold, apply pre-whitening processing to the statistical time series to generate the residual series y. i =x i -ρ1x i-1 ; The Mann-Kendall trend test method is used to calculate the rank statistic S and variance Var(S) based on the residual series, and a standardized trend test statistic Z is constructed: When S>0, Z=(S-1) / sqrt(Var(S)); when S<0, Z=(S+1) / sqrt(Var(S)); If the standardized trend test statistic Z is less than the preset negative threshold, it is determined that the cascade power generation index has a downward trend and the adaptability of the current cascade dispatch rules is at risk of deterioration.

6. The method according to claim 1, characterized in that, The process involves parsing the existing scheduling rules at each level to extract adjustable parameters in order to construct a feasible domain for parameter adjustment. Specifically, this includes: Extract adjustable parameters that affect the power generation of the cascade from the existing cascade dispatch rules, including at least the start date of post-flood water storage, the completion date of post-flood water storage, the target water storage level after the flood season, and the minimum control water level during the dry season. Based on the preset reservoir safety operation constraints, the range of adjustable parameters is defined, and the feasible domain Ω for parameter adjustment is constructed. The Latin hypercube sampling method is used to generate a parameter sample set within the feasible region of parameter adjustment: for the i-th parameter, its value range is divided into N... s The sampling formula for a subinterval with equal probability in the j-th layer is: X i，j =X i，min +(j-1)δ i +u i，j d i ; Where X i，min δ is the lower limit of the parameter. i u is the layer width. i，j The sample points are uniformly random numbers within the interval [0, 1]. By randomly permuting and combining the sampling points of each parameter, a sample of N is formed. s The parameter sample set of the group parameter vector.

7. The method according to claim 6, characterized in that, Identify the master control parameters among the adjustable parameters, including: Using a parameter sample set of adjustable parameters as input, parallel hydraulic simulations are performed on a set of random runoff scenarios to calculate the cascade comprehensive benefit function value U(X) corresponding to each set of parameter vectors. j Construct a scheduling parameter-target response dataset; The first-order sensitivity index S of each adjustable parameter is calculated using the variance decomposition method based on conditional expectation. i : S i =(1 / V(U))∑ b=1 nbin (n i，b / N s )(E i，b -U*) 2 ; Where V(U) is the total variance of the comprehensive benefit function, U* is the sample mean, and n bin To divide the interval number, E i，b Let n be the conditional mean of the comprehensive benefit function when the i-th parameter falls into the b-th interval; i，b This represents the number of samples where the i-th parameter falls into the b-th interval; The first-order sensitivity index S i Adjustable parameters exceeding the preset sensitivity threshold are identified as main control parameters, and the remaining parameters are fixed to the values ​​in the current cascade scheduling rules.

8. The method according to claim 7, characterized in that, This also includes a comprehensive combination of the discrete levels of the main control parameters and the construction of a cascade comprehensive benefit and power generation deficit risk response matrix, specifically: The value ranges of the main control parameters are discretized to obtain the discrete levels of each main control parameter; the discrete levels of all main control parameters are then comprehensively combined to generate N. comb A comprehensive parameter vector is generated; using this comprehensive parameter vector as input, parallel hydraulic simulations are performed on a set of random runoff scenarios to obtain the cascade comprehensive benefit function value U(X) under each combination. c Value at Risk (CVaR) for Power Generation Deficit Conditions α，c Using the discrete level of the main control parameters as the index dimension, the cascade comprehensive benefit response matrix U and the power generation deficit risk response matrix CVaR are calculated.

9. The method according to claim 8, characterized in that, Guided by the dual objectives of maximizing the overall benefits of the cascade system and minimizing the magnitude of parameter adjustments, optimized scheduling parameter recommendations are generated, including: Set an acceptable threshold for power generation deficit risk r th Based on the power generation deficit risk response matrix CVaR, we screen for those that meet the CVaR. α，c ≤r th The combinations constitute a feasible set of combinations Φ; Find the combination c* within the feasible combination set Φ that maximizes the tiered comprehensive benefit function: The recommended values ​​for the master control parameters corresponding to the combination c* are the scheduling parameters that maximize the overall benefits of the cascade. Calculate the standardized adjustment distance D of each combination in the feasible combination set Φ relative to the current cascade scheduling rule values. c : Set a threshold for improving efficiency ΔU min Within the feasible combination set Φ, filter those satisfying U(X) c )-U(X c0 )≥ΔU min The combinations constitute the candidate combination set Ψ, where U(X) c0 This represents the comprehensive benefits of the current cascade scheduling rules; Find the combination c** with the smallest standardized adjustment distance in the candidate combination set Ψ, and take the master control parameter value corresponding to combination c** as the scheduling parameter suggestion with the smallest parameter adjustment range.

10. A smart scheduling decision support system for cascade reservoir power stations, characterized in that, include: The predicted runoff correction module is used to acquire historical meteorological and hydrological data and future meteorological prediction data of the cascade reservoir basin. Based on the correlation characteristics between pre-stored historical meteorological elements and runoff prediction deviations under different meteorological combination quadrants, a quadrant deviation correction model is constructed. Based on this, the predicted runoff is corrected for quadrant deviations, and a runoff input sequence containing deterministic predicted runoff and random runoff scenario sets is generated. The scheduling rule response evaluation module is used to perform multi-scenario parallel hydraulic simulation on a random runoff scenario set based on the preset existing scheduling rules of the cascade, calculate the cascade power generation index and the risk value of power generation deficit conditions, and determine the adaptability of the existing scheduling rules of the cascade based on the time-varying trend of the cascade power generation index. The parameter adaptation and adjustment module is used to respond to the risk of degradation in the adaptability of the current scheduling rules of the cascade. It analyzes the current scheduling rules of the cascade to extract adjustable parameters to construct the feasible domain of parameter adjustment. It identifies the main control parameters among the adjustable parameters by combining the simulation results of multiple scenarios. With the dual guidance of maximizing the comprehensive benefits of the cascade and minimizing the parameter adjustment range, it generates optimized scheduling parameter suggestions.

Images