Robust scheduling method of water resources considering supply and demand uncertainties

Abstract

The application discloses a kind of water resource collaborative robust scheduling method considering supply and demand double uncertainty, comprising: the risk exposure contribution rate of risk of system risk of inflow and water demand is evaluated, and asymmetric inflow and water demand robust uncertainty set is constructed accordingly;Supply and demand risk coupling amplification index is calculated, based on the index, the joint uncertainty space is screened and reconstructed, and a joint scenario tree is generated;Scenario tree is input into water resource hierarchical collaborative scheduling model, inter-period optimization is carried out using global resource allocation layer, and stage risk quantification parameter representing risk constraint strength of each stage is determined;The parameter is used as cross-layer risk constraint by using local scheduling decision layer, and a dynamic scheduling strategy that meets the requirements of global risk control is generated.The application effectively solves the problems of over-conservatism and extreme risk defense failure in the scheduling process through risk scenario construction driven by physical mechanism and cross-layer risk explicit conduction.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of water conservancy engineering and operations research optimization, and in particular to a robust collaborative scheduling method for water resources that considers both supply and demand uncertainties. Background Technology

[0002] In watershed water resource systems, the dual uncertainties of inflow (rainfall runoff) and demand (industrial and agricultural water use) are core factors affecting the safety and economy of scheduling decisions. Constructing robust scheduling models that can adapt to random fluctuations in supply and demand, accurately quantify system risks, and coordinate long-term and short-term decisions has significant engineering value for improving water resource allocation efficiency and responding to hydrological events such as extreme droughts.

[0003] Existing water resource scheduling methods often employ stochastic programming or robust optimization to handle uncertainties. Robust optimization typically constructs a box-shaped or ellipsoidal set of uncertainties, requiring the solution to satisfy constraints under any perturbation within the set; scenario tree generation technology uses probabilistic distances (such as Euclidean distance) of historical data for clustering and reduction to generate representative scenarios; and hierarchical scheduling models typically employ a unidirectional loosely coupled mechanism where the upper level plans the total target and the lower level is responsible for real-time allocation.

[0004] However, the aforementioned existing technologies suffer from the following technical problems in practical applications: Traditional robust assemblies often employ symmetric or static structures, failing to distinguish the heterogeneous contributions of incoming and demanding water to system failure risk. This leads to unnecessary margins reserved in non-critical disturbance directions, resulting in overly conservative and wasteful resource allocation. Existing scenario construction methods rely solely on the probabilistic statistical characteristics of data, ignoring the physical shrinkage effect of joint supply and demand disturbances on the scheduling solution space. This results in extreme risk scenarios with low probability of occurrence but causing a sharp reduction in the scheduling feasible region being easily eliminated during the reduction process, creating potential safety hazards. Furthermore, there is a lack of explicit risk transmission mechanisms between long-scale resource planning and short-scale real-time scheduling. The risk boundary at the macro level cannot be transformed into hard mathematical constraints at the micro-operational level, causing local strategies to easily exceed the global safety baseline when pursuing real-time benefits. Summary of the Invention

[0005] The purpose of this invention is to provide a robust collaborative scheduling method for water resources that considers both supply and demand uncertainties, in order to solve at least one of the aforementioned problems in the existing technology.

[0006] According to one aspect of this application, a robust collaborative scheduling method for water resources that considers both supply and demand uncertainties includes:

[0007] Based on pre-stored historical water inflow data and historical water demand data, the risk exposure contribution rate of water inflow uncertainty and water demand uncertainty to system risk is assessed, and an asymmetric water inflow robust uncertainty set and an asymmetric water demand robust uncertainty set are constructed accordingly.

[0008] Based on the asymmetric inflow robust uncertainty set and the asymmetric demand robust uncertainty set, the supply and demand risk coupling amplification index is calculated, and the joint uncertainty space is screened and reconstructed accordingly to generate a joint scenario tree;

[0009] Input the joint scenario tree into the pre-built hierarchical collaborative scheduling model for water resources;

[0010] By utilizing the global resource allocation layer of the water resource hierarchical collaborative scheduling model, cross-period resource allocation optimization is performed based on the joint scenario tree to determine the phased risk quantification parameters that characterize the risk constraint intensity at each stage;

[0011] By utilizing the local scheduling decision layer of the water resources hierarchical collaborative scheduling model, the phased risk quantification parameters are used as cross-layer risk constraints to generate a dynamic scheduling strategy that meets the requirements of global risk control.

[0012] According to another aspect of this application, a robust collaborative scheduling system for water resources that considers both supply and demand uncertainties includes:

[0013] The asymmetric uncertainty modeling module is used to assess the risk exposure contribution rate of inflow uncertainty and demand uncertainty to system risk based on pre-stored historical inflow and demand data, and to construct asymmetric inflow robust uncertainty set and asymmetric demand robust uncertainty set accordingly.

[0014] The supply and demand risk coupling scenario construction module is used to calculate the supply and demand risk coupling amplification index based on the asymmetric inflow water robust uncertainty set and the asymmetric demand water robust uncertainty set, thereby filtering and reconstructing the joint uncertainty space and generating a joint scenario tree.

[0015] The global resource allocation module is used to optimize intertemporal resource allocation based on the global resource allocation layer of the pre-built water resource hierarchical collaborative scheduling model and the joint scenario tree to determine the phased risk quantification parameters that characterize the risk constraint intensity of each stage.

[0016] The local collaborative scheduling decision module is used to utilize the local scheduling decision layer of the water resource hierarchical collaborative scheduling model, and use the phased risk quantification parameters as cross-layer risk constraints to generate a dynamic scheduling strategy that meets the requirements of global risk control.

[0017] Beneficial effects: This invention effectively solves the problems of overly conservative and extreme risk defense failure in the scheduling process by constructing risk scenarios driven by physical mechanisms and explicitly transmitting risks across layers. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the water resource collaborative robust scheduling method that considers the dual uncertainties of supply and demand provided in the embodiments of this application.

[0019] Figure 2 This is a schematic diagram of the process for assessing the contribution rate of water inflow uncertainty and water demand uncertainty to system risk, provided in an embodiment of this application.

[0020] Figure 3 This is a schematic diagram illustrating the process of constructing the asymmetric inflow robust uncertainty set and the asymmetric demand robust uncertainty set provided in the embodiments of this application.

[0021] Figure 4 This is a schematic diagram of the process of generating a dynamic scheduling strategy that meets the global risk control requirements using the local scheduling decision layer of the water resources hierarchical collaborative scheduling model, as provided in the embodiments of this application. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] Example 1 details the overall process of a robust water resource scheduling method that considers both supply and demand uncertainties, such as... Figure 1 As shown, a complete technical method is provided, encompassing data input, risk modeling, scenario construction, hierarchical scheduling, and parameter feedback. This method effectively addresses the scheduling failure problem caused by neglecting the coupling of supply and demand risks and the transmission of risks across layers in traditional scheduling methods. In this embodiment, the Hongze Lake water resource scheduling system in the Huai River Basin is used as an example. This system includes rainfall and runoff input, lake regulation and storage, gate control, and multi-user water supply output.

[0024] Step 101: Based on the pre-stored historical water inflow data and historical water demand data, assess the risk exposure contribution rate of water inflow uncertainty and water demand uncertainty to system risk, and construct an asymmetric water inflow robust uncertainty set and an asymmetric water demand robust uncertainty set based on the risk exposure contribution rate.

[0025] Specifically, historical water inflow data includes, but is not limited to, daily rainfall within the basin, inflow processes at main stream control stations, and outflow data from upstream reservoirs; historical water demand data covers daily water demand records for various users, including agricultural irrigation, industrial production, urban living, and the ecological environment. In this embodiment, the collected raw data undergoes standardization processing, for example, using Z-score (standard score method) to eliminate dimensional differences and removing outliers caused by sensor malfunctions or human input errors to ensure the quality of the input data.

[0026] Based on this, the operation of the scheduling system was simulated to statistically analyze the frequency of violations of key system constraints (such as minimum water level and minimum ecological baseflow) and the corresponding water supply gaps under scenarios of simple inflow and simple demand disturbances. Based on the statistical results, the degree to which inflow and demand uncertainties each cause system risk was quantified, i.e., the risk exposure contribution rate. This contribution rate was used as a weighting factor to differentiate the boundary range of the uncertainty set. For example, if the assessment results show that inflow fluctuations contribute significantly more to the system's water shortage risk than demand fluctuations, the constructed asymmetric inflow robust uncertainty set will have a larger boundary expansion amplitude than the asymmetric demand robust uncertainty set. This asymmetric design avoids the conservative treatment of all uncertainty sources uniformly in traditional symmetric robust methods, improving resource utilization efficiency while ensuring safety.

[0027] Step 102: Based on the asymmetric inflow robust uncertainty set and the asymmetric demand robust uncertainty set, calculate the supply and demand risk coupling amplification index, and based on the supply and demand risk coupling amplification index, screen and reconstruct the joint uncertainty space to generate a joint scenario tree.

[0028] Specifically, using the two sets constructed in step 101 as boundaries, a joint uncertainty space containing a large number of combined water inflow and demand scenarios is generated through discretization sampling. For each scenario node in the space, its corresponding supply-demand risk coupling amplification index (RCCI) is calculated. This index not only considers the probability of a specific water inflow and demand state occurring simultaneously, but also introduces a physical mechanism-level measure, namely, the degree to which the combined state leads to the contraction of the scheduling feasible region, and the dynamic correlation of supply-demand deviations over time.

[0029] Based on the calculated RCCI values, a massive number of scenarios are screened and reconstructed. A preset risk threshold is set, and scenarios with RCCI values ​​higher than this threshold are identified as high-risk scenarios and retained in their entirety. These scenarios typically correspond to low-probability but highly destructive supply-demand mismatches, such as a dry year coupled with high water demand. For low-risk scenarios below the threshold, clustering algorithms (such as K-means clustering) are used for compression to extract representative scenario nodes. The retained high-risk scenarios are merged with the compressed representative scenarios to generate a joint scenario tree with manageable scale and focused risk. This process significantly reduces computational dimensionality while ensuring that the scheduling model can perceive and defend against extreme risks.

[0030] Step 103: Input the joint scenario tree into the pre-constructed water resource hierarchical collaborative scheduling model. The water resource hierarchical collaborative scheduling model includes a global resource allocation layer and a local scheduling decision layer.

[0031] Specifically, the hierarchical collaborative water resource scheduling model is a mathematical model architecture comprising two decision-making levels. The input interface is designed to receive structured joint scenario tree data, including hydrological element values, occurrence probabilities, and their evolution over time for each scenario node. The model internally includes pre-set physical constraint equations for Hongze Lake scheduling (such as water balance equations and reservoir capacity curve relationships) and multi-objective evaluation functions.

[0032] Step 104: Utilize the global resource allocation layer of the water resource hierarchical collaborative scheduling model, perform inter-period resource allocation optimization based on the joint scenario tree, and determine the stage-specific risk quantification parameters that characterize the risk constraint intensity at each stage.

[0033] Specifically, the global resource allocation layer focuses on long-scale (e.g., monthly or decadal) resource allocation problems. At this layer, a Distributed Optimization Route (DRO) model based on Wasserstein distance is established to find the optimal resource allocation strategy under the worst-case probability distribution. By solving this optimization model, not only are the target water storage levels and macro-level water supply plans obtained for each time period, but also Lagrange multipliers associated with risk constraints (such as water supply reliability constraints) are extracted through duality theory. These multipliers serve as the stage-specific risk quantification parameter, numerically representing the marginal cost or consequence of violating risk constraints in the current time period. When a period faces a high risk of supply-demand mismatch, the corresponding risk quantification parameter value will increase significantly, signaling a tightening of constraints to the lower layers.

[0034] Step 105: Using the local scheduling decision layer of the water resources hierarchical collaborative scheduling model, the phased risk quantification parameters are used as cross-layer risk constraints to generate a dynamic scheduling strategy that meets the requirements of global risk control.

[0035] Specifically, the local scheduling decision layer focuses on real-time operational issues at short scales (such as daily or hourly scales). Deep reinforcement learning algorithms (such as Soft Actor-Critic, SAC) are employed at this level. To achieve collaboration with the global layer, the stage-based risk quantification parameter output in step 104 is directly used as an extended dimension of the reinforcement learning state space, allowing the agent to explicitly perceive the current macro-level risk during decision-making. Simultaneously, a penalty term based on this parameter is introduced into the reward function and transformed into a hard constraint for action selection. The resulting dynamic scheduling strategy can output specific gate opening degrees and diversion quotas based on real-time water levels, inflow, and risk signals, satisfying both the flexibility of real-time scheduling and strictly adhering to the risk control boundaries of the global layer.

[0036] The boundary conditions are handled as follows: when historical water inflow or water demand data is missing, linear interpolation or time series prediction models are used to complete it; when extreme water inflow or water demand inputs that exceed the historical extreme value range occur, the system automatically triggers an early warning mechanism and extends the uncertainty set boundary to cover the extreme value.

[0037] The exception handling mechanism is as follows: when real-time monitoring data transmission is interrupted, the system switches to a conservative scheduling mode based on historical statistical values; when the optimization solver fails to converge within a preset time limit, the system adopts the scheduling strategy of the previous period as an alternative and records the exception log for subsequent analysis.

[0038] Example 2 elaborates on how to construct an asymmetric robust set by quantifying the risk contribution of different uncertainty sources, such as... Figure 2 , Figure 3 As shown, a modeling method based on a combination of data-driven and model simulation is provided to solve the conservatism problem caused by subjective or symmetrical boundary setting in traditional robust optimization.

[0039] Step 201: Construct a disturbance space based on historical water inflow data and historical water demand data, and generate multiple sets of uncertain disturbance samples.

[0040] Specifically, based on the statistical characteristics of historical data (such as mean, variance, and distribution type), the basic fluctuation range of water inflow and water demand is determined, and an initial disturbance space is constructed. In this embodiment, it is assumed that the water inflow Q follows a log-normal distribution and the water demand D follows a normal distribution. The Monte Carlo sampling method is used to generate N sets (e.g., N=1000) of uncertain disturbance samples {(Q... s D s )} s=1 N (Q) s Let D represent the water inflow sequence in the s-th sample group. s(This represents the water demand sequence in the s-th sample group). Each sample group represents a possible supply and demand combination scenario.

[0041] Step 202: Input the uncertainty disturbance samples into the pre-built scheduling system for simulation operation, count the trigger frequency of key constraints under different disturbance scenarios, and calculate the water supply failure risk index.

[0042] Specifically, each generated sample set is used as the input boundary condition for the Hongze Lake scheduling simulation model, and long-series adjustment calculations are performed according to established conventional scheduling rules. During the simulation, the execution of key system constraints is monitored in real time. Key constraints include, but are not limited to: minimum dead water level constraint (Z≥Z0). min ), maximum flood control limit water level constraint (Z≤Z) max ) and minimum ecological discharge flow constraint (Q eco ≥Q eco,min Where Z is the reservoir water level, Z min Dead water level, Z max To control floodwater levels, Q eco Q represents the actual ecological outflow. eco,min This is the minimum ecological flow threshold.

[0043] For any scenario s, count the total number of times various key constraints are triggered within the scheduling period, and calculate the normalized key constraint trigger frequency f. s The calculation formula is as follows:

[0044] ;

[0045] Among them, f s Let be the normalized critical constraint trigger frequency under scenario s; k be the constraint type index; K be the total number of constraint types; m be the index of the scheduling period; M be the total number of scheduling periods; I s (k,m) is an indicator function, which takes the value 1 when scenario s triggers the k-th type constraint in time period m, and 0 otherwise; w k represents the importance weight of the k-th type of constraint.

[0046] At the same time, calculate the water supply failure risk index L under this scenario. s This indicator comprehensively reflects the likelihood of water shortages and the severity of their consequences; its calculation formula is as follows:

[0047] L s =P s *E s ;

[0048] Among them, L s P is the water supply failure risk indicator under scenario s. s Let E be the probability of water supply disruption (i.e., water supply less than water demand) occurring under scenario s.s This represents the expected amount of water shortage or the maximum water shortage depth.

[0049] Step 203: Based on the trigger frequency of key constraints and the risk index of water supply failure, quantify the risk exposure intensity under the scenarios dominated by inflow disturbance and demand disturbance, respectively, and calculate the contribution rate of inflow risk exposure and the contribution rate of demand risk exposure based on the risk exposure intensity.

[0050] Specifically, the scenario set is first divided based on the perturbation amplitude of the samples. A perturbation threshold θ for the incoming water is then set. Q and water demand disturbance threshold θ D For example, take 1.5 times the historical standard deviation of each. Define the set of dominant inflow disturbance scenarios S. Q :

[0051] ;

[0052] and the dominant scenario set S of water demand disturbance D :

[0053] ;

[0054] in, This is the average of historical water inflow data (baseline water inflow). The mean of historical water demand data (baseline water demand) is N, and the total number of samples is N.

[0055] For each of the two sets, calculate the cumulative risk exposure intensity R. Q With R D :

[0056] ;

[0057] ;

[0058] Among them, R Q R represents the cumulative risk exposure intensity under the dominant scenario of inflow disturbance. D S represents the cumulative risk exposure intensity under the dominant water demand disturbance scenario; s is the scenario index; S Q The set of dominant scenarios for incoming water disturbance; f s L represents the trigger frequency of key constraints in scenario s; s This is a risk indicator for water supply failure under scenario s.

[0059] Calculate the normalized contribution rate of inflow risk exposure α Q Contribution rate α to water demand risk exposure D :

[0060] α Q =R Q / (RQ +R D );

[0061] Where, α Q Contribution rate to inflow water risk exposure; R Q R represents the risk exposure intensity under the dominant scenario of inflow disturbance. D This represents the risk exposure intensity under the dominant water demand disturbance scenario.

[0062] α D =R D / (R Q +R D ).

[0063] This contribution rate directly reflects which source of uncertainty poses a greater threat to the safe operation of the system.

[0064] Step 204: Perform statistical analysis on historical water inflow data and historical water demand data, extract statistical features, and determine the uncertain set of water inflow benchmark and the uncertain set of water demand benchmark based on the statistical features.

[0065] Specifically, the inflow reference uncertainty set Ω Q,0 With the water demand baseline uncertainty set Ω D,0 It is typically defined as an interval range covering a certain confidence level (such as 90% or 95%). For example, for inflow water, its baseline set can be represented as:

[0066] ;

[0067] in σ is the mean. Q The standard deviation is denoted as .

[0068] Step 205: Using the contribution rate of inflow risk exposure and the contribution rate of demand risk exposure as weighting factors, calculate the inflow boundary expansion coefficient and the demand boundary expansion coefficient respectively.

[0069] Specifically, a basic expansion parameter η is introduced, for example, 0.2, and the boundary expansion coefficient k is calculated based on the contribution rate. Q With k D :

[0070] k Q =1+η*α Q ;

[0071] Where, k Q Let be the boundary expansion coefficient of the uncertain set of incoming water; η be the basic expansion parameter (e.g., take 0.2); α be the boundary expansion coefficient of the uncertain set of incoming water; η be the boundary expansion coefficient of the uncertain set of incoming water; ... Q Contribution rate to risk exposure of incoming water.

[0072] k D =1+η*αD .

[0073] Therefore, if α Q >α D Then k Q >k D This indicates that the boundary of the uncertain set of incoming water will be expanded to accommodate its higher risk contribution.

[0074] Step 206: Based on the inflow boundary expansion coefficient and the demand boundary expansion coefficient, the inflow reference uncertainty set and the demand reference uncertainty set are expanded outward in a directional manner to obtain the asymmetric inflow robust uncertainty set and the asymmetric demand robust uncertainty set.

[0075] Specifically, the upper and lower limits of the reference set are multiplied by the corresponding expansion coefficients (for the lower limit, this is usually divided by the coefficient or shifted outward) to obtain the final asymmetric robust uncertainty set Ω. Q With Ω D For example, the expanded inflow set is Ω. Q =k Q *Ω Q,0 The aforementioned targeted scaling mechanism ensures sufficient robustness margins in higher-risk dimensions while avoiding unnecessary conservatism in lower-risk dimensions, thus achieving a balance between risk control and resource efficiency.

[0076] Example 3 elaborates on the construction of the Supply-Demand Risk Coupling Amplification Index (RCCI) and its application in scenario tree generation. It introduces feasible domain shrinkage and dynamic correlation indicators at the physical mechanism level, and provides a scenario screening method that can accurately identify low-probability, high-damage risks.

[0077] Step 301: Using the asymmetric inflow robust uncertainty set and the asymmetric demand robust uncertainty set as boundaries, construct a discretized joint uncertainty input space.

[0078] Specifically, utilizing the asymmetric boundary Ω Q and Ω D Define a two-dimensional (or high-dimensional, if considering multiple time periods and multiple sites) rectangular or hyperrectangular region. Within this region, generate a series of discrete scene nodes (Q nodes) using grid partitioning or random sampling methods. i D j ), which constitute a joint uncertainty input space.

[0079] Step 302: For each scenario node in the joint uncertainty input space, calculate the corresponding supply and demand risk coupling amplification index. The supply and demand risk coupling amplification index is used to characterize the comprehensive amplification effect of the joint supply and demand disturbance on the system risk under that scenario node.

[0080] The supply-demand risk coupling amplification index comprehensively considers the following three dimensions:

[0081] The joint probability of abnormal water inflow and water demand deviation under scenario nodes;

[0082] The combined supply and demand disturbances cause the solution space for the system to satisfy the operating constraints to shrink by the extent of the shrinkage of the scheduling feasible region relative to the preset baseline scenario;

[0083] And the dynamic correlation between the relative deviation sequence of water inflow and the relative deviation sequence of water demand, constructed based on historical water inflow and historical water demand data, and the supply and demand deviation in the current period.

[0084] Specifically, the formula for calculating the Supply-Demand Risk Coupling Index (RCCI) is as follows:

[0085] RCCI S =P(Q S D S )*(1+λ'*ΔΩ S )*(1+μ*ρ Q,D );

[0086] Among them, RCCI S P(Q) is the supply and demand risk coupling amplification index for scenario node S; S represents a specific scenario node; P(Q) S D S ) represents the joint probability of this scenario occurring; ΔΩ S To determine the contraction magnitude of the feasible region for scheduling; ρ Q,D The dynamic correlation between supply and demand deviations; Q S D represents the incoming water state under scenario S; S Let S represent the water demand state under scenario S; λ' and μ are adjustment coefficients, for example, both taken as 1.0, used to balance the influence weights of various physical factors.

[0087] Next, we will elaborate on ΔΩ. S and ρ Q,D The specific calculation process.

[0088] Step 303: Under the baseline scenario and the current scenario, perform stratified sampling within the physical feasible interval of the preset scheduling decision variables, and count the proportion of feasible samples that meet the system operation constraints.

[0089] This step and step 304 describe in detail the calculation of the feasible region shrinkage magnitude. Since the scheduling feasible region is usually a complex polyhedron in a high-dimensional space, it is difficult to directly calculate the volume. Therefore, a sampling-based surrogate index method is adopted.

[0090] Specifically, the Stratified Latin Hypercube Sampling (SLHS) method is employed. First, a bounding box containing the physical upper and lower limits of all decision variables (such as the flow rate of each gate and the reservoir water level at each time period) is constructed. Within this bounding box, a large number of uniformly distributed sampling points x are generated. k For example, N=10000.

[0091] The sampling formula is:

[0092] ;

[0093] Where, x i k Let x be the value of the k-th sampling point on the i-th dimension of the decision variable; i,min Let k be the physical lower bound of the i-th decision variable; k is the index of the sampling point (from 1 to N); u k The number is a random number in the interval (0,1); N is the total number of sampling points; x i,max Let be the physical upper bound of the i-th decision variable.

[0094] For each sampling point x k Substitute it into the system constraint equations g j (x k Q S D S The test is performed if the condition is less than or equal to 0. The number of feasible samples N that satisfy all constraints is then counted. feasible And calculate the feasible sample proportion V. S =N feasible / N.

[0095] Step 304: Calculate the rate of decrease of the proportion of feasible samples in the current scenario relative to the proportion of feasible samples in the baseline scenario, and define the rate of decrease as the magnitude of the contraction of the scheduling feasible region to characterize the degree of geometric compression of the solution space.

[0096] Specifically, let V0 be the feasible sample proportion under the baseline scenario (usually the multi-year average inflow and demand), and V be the feasible sample proportion under the current scenario S. S The shrinkage magnitude ΔΩ of the scheduling feasible region. S The calculation is as follows:

[0097] ΔΩ S =max(0,(V0-V S ) / V0);

[0098] Where, ΔΩ S V represents the shrinkage magnitude of the scheduling feasible region under scenario S; V0 represents the proportion of feasible samples under the baseline scenario; V S This represents the proportion of feasible samples under the current scenario S.

[0099] This indicator intuitively reflects how much the operational space left to dispatchers is compressed when water inflow decreases or water demand increases. For example, if 80% of the solutions are feasible under the baseline scenario, but only 20% of the solutions are feasible under an extreme drought scenario, the contraction is (0.8-0.2) / 0.8=0.75, indicating that the dispatching difficulty increases sharply.

[0100] Step 305: Construct the relative deviation sequence of incoming water and the relative deviation sequence of demand water, and set the sliding time window.

[0101] This step and subsequent steps describe in detail the calculation of dynamic correlation.

[0102] Specifically, the relative deviation sequence is defined as follows:

[0103] ;

[0104] ;

[0105] Among them, Q t and D t These represent the actual inflow and actual water demand at time t, respectively. and These are average inflow and average demand, respectively.

[0106] Set the length W of the sliding time window, for example, W=30 days, to capture short-term variation features of the correlation.

[0107] Step 306: Within the sliding time window, calculate the Spearman rank correlation coefficient between the relative deviation sequence of inflow water and the relative deviation sequence of demand water.

[0108] Specifically, for a window centered at time t, the subsequences within the window are extracted. Since hydrological and meteorological data often do not follow a strict normal distribution, the Spearman rank correlation coefficient can more robustly measure the monotonic correlation between two sequences. The calculation formula is:

[0109] ;

[0110] Where, ρ t d is the Spearman rank correlation coefficient within a sliding window centered at time t; i is the difference between the rank of the inflow deviation and the rank of the demand deviation for the i-th sample within the window; W is the length of the sliding time window (number of samples).

[0111] Step 307: Based on the time decay weighting mechanism, the Spearman rank correlation coefficients of different time windows are fused to obtain the dynamic correlation of supply and demand deviation.

[0112] Specifically, to reflect that recent correlations have a greater impact on current scheduling than long-term correlations, a time decay weight w is introduced. t :

[0113] w t =exp(-(t curr -t) / τ);

[0114] Among them, w t The time decay weight corresponding to the correlation coefficient at time t; exp is the exponential function; t curr t represents the current decision-making time; t represents the historical time; τ represents the decay constant, such as 90 days.

[0115] The final dynamic correlation ρ Q,D This is the weighted average of the correlation coefficients across all windows. When ρ Q,D When the negative value is large, that is, when the water inflow is small and the water demand is large, the (1+μ*ρ) in the RCCI formula... Q,D Although the value of the term becomes smaller, considering that negative correlation is usually accompanied by extremely low joint probability P, μ can take a negative value here or the formula form can be adjusted to ensure that negative correlation (adverse working conditions) can amplify the risk index. Alternatively, the absolute value of the correlation or a specific mapping can be taken in the RCCI definition to ensure that the physical fact that negative correlation exacerbates risk is reflected. In the preferred embodiment, the correction term (1+μ*|min(0,ρ)) is used. Q,D The risk index is amplified only when a negative correlation is observed.

[0116] Step 308: Set a preset risk threshold, identify scenario nodes with a supply-demand risk coupling amplification index greater than or equal to the preset risk threshold as high-risk scenario nodes, and retain them completely.

[0117] Specifically, a threshold θ is determined based on the calculated distribution (e.g., quantiles) of the RCCI values ​​for all scenario nodes. S Nodes with a value ≥ θ are marked as high-risk nodes and directly included in the final scenario set.

[0118] Step 309: Identify scenario nodes with supply and demand risk coupling amplification index less than a preset risk threshold as low-risk scenario nodes, and perform clustering and compression on the low-risk scenario nodes to extract representative scenario nodes.

[0119] Specifically, for the remaining low-risk nodes, the K-means clustering algorithm is used to divide them into K clusters. In each cluster, the node closest to the cluster center or the node with the largest RCCI value is selected as the representative scenario node, and the probabilities of all nodes in the cluster are summed to obtain the probability of the representative node.

[0120] Step 310: Merge high-risk scenario nodes with representative scenario nodes to generate a joint scenario tree.

[0121] Specifically, the two types of node sets are merged, and the probabilities are renormalized to construct a joint scenario tree with several branches. This scenario tree has dense and detailed scenarios in high-risk areas (such as areas with high water consumption during droughts), and sparse scenarios in low-risk areas (such as areas with low water consumption during abundant water periods), resulting in high computational efficiency. This provides a high-quality input foundation for subsequent hierarchical scheduling models.

[0122] Example 4 details the construction and solution process of the global resource allocation layer in the hierarchical collaborative scheduling model of water resources, and provides a two-stage distributed bar optimization (DRO) method based on Wasserstein distance. This method can not only effectively deal with the probability distribution deviation in the joint scenario tree, but also extract the risk shadow quantity with physical meaning through the mathematical duality mechanism, providing accurate risk boundary signals for local real-time scheduling.

[0123] Step 401: Construct a two-stage sub-Brubar optimization model in the global resource configuration layer, and use the sub-Brubar uncertainty set based on Wasserstein distance to characterize the distribution bias in the joint scenario tree.

[0124] Specifically, the global resource allocation layer primarily addresses medium- to long-term (e.g., monthly or ten-day) water resource allocation issues. In this layer, decision variables are divided into two stages:

[0125] The first-stage decision variable x represents the pre-scheduling instructions before the uncertainty is revealed, such as the macro-level water supply quota for each period and the target water level control trajectory of the reservoir.

[0126] Second-stage decision variable y s This refers to remedial dispatching actions that occur after a specific uncertainty scenario is revealed, such as adjusting the amount of emergency water diversion to deal with a sudden drought.

[0127] The uncertain set D of the Brønsted bar is based on the empirical distribution P. * A spherical set centered at a point with a Wasserstein distance δ as its radius, i.e., D = {P|W1(P,P...} * )≤δ}, where W1 is the Wasserstein distance (distance between distributions), P * It is a discrete empirical distribution constructed based on joint scenario tree samples, where δ is the robust radius parameter used to cover the deviation between the true distribution and the empirical distribution. This set contains all potential probability distributions that are statistically close enough to the empirical distribution, ensuring that the model remains feasible even under the worst distribution shift.

[0128] Step 402: With the objective of minimizing the expected risk-benefit composite cost under the worst-case distribution scenario, the two-stage sub-Bruker optimization model is solved to obtain the intertemporal resource allocation decision, which is used to determine the phased risk quantification parameters.

[0129] Specifically, the objective function of the two-stage sub-Bruker optimization model is constructed as follows:

[0130] min x {c T x+sup P∈D E P [Q(x,ξ)]};

[0131] Among them, c T x represents the deterministic cost of the first phase; sup represents the supremum; E P The expectation operator is Q(x,ξ); Q(x,ξ) is the second-stage optimal cost function under random parameter ξ (i.e., water inflow and water demand states), which typically includes water shortage penalty cost, water abandonment loss cost, and operation and scheduling cost.

[0132] Since the objective function contains a supration operation with respect to the probability distribution P, a direct solution is very difficult. In this embodiment, the infinite-dimensional worst-case distribution expectation problem is reconstructed into a finite-dimensional dual problem using strong duality theory. The specific reconstruction formula is as follows:

[0133] ;

[0134] Where sup is the supremum; inf is the infremum; λ is the dual variable corresponding to the Wasserstein distance constraint; δ is the radius of the Wasserstein sphere; N is the number of samples; ξ i Let ||ξ-ξ| be the i-th sample point in the joint scenario tree; i || represents the norm distance in the context space.

[0135] This reconstruction transforms the original problem into a semi-infinite programming problem, which is then numerically solved using the Sample Average Approximation (SAA) or the tangent plane method, ultimately yielding the optimal first-stage resource allocation decision x. * .

[0136] Step 403: Construct a Lagrangian function for the stage risk constraints in the two-stage sub-Bruker optimization model, and introduce Lagrangian multipliers corresponding to the stage risk constraints.

[0137] Specifically, in the aforementioned DRO model, in addition to the objective function, a series of phased risk constraints g are also included. t (x,y sRisk constraints are ≤0, such as the water supply guarantee rate constraint or the ecological baseflow satisfaction rate constraint for a certain time period t. To extract the risk signal, a Lagrange function L is constructed. This function connects the objective function and the risk constraint through the Lagrange multiplier π. t Perform a linear combination, in the form of:

[0138] L=Obj+Σ t π t (Risk Constraint_t -ε t );

[0139] Where L is the Lagrangian function, Obj is the objective function value under the worst-case distribution, and Risk Constraint_t ε is a risk indicator for time period t (such as expected water shortage). t For the allowable limit or risk budget of this constraint, π t Let π be the Lagrange multiplier for time period t (i.e., the shadow quantity of stage risk). t The risk constraint corresponding to each time period is a non-negative real number.

[0140] Step 404: Solve for the Lagrange function and define the obtained Lagrange multiplier sequence as the stage risk shadow quantity. The stage risk shadow quantity is used to quantify the marginal cost caused by violating the risk constraint in the current stage. The stage risk shadow quantity is the stage risk quantification parameter.

[0141] Specifically, when solving the dual problem of the DRO model, the optimal Lagrange multiplier sequence {π} can be obtained simultaneously. t * From an economic perspective, π t * This indicates the risk constraint limit ε at stage t. t The increase in the overall expected cost of the system when tightening by one unit. In other words, it is the shadow price of risk. If π in a certain time period t... t A large value for π indicates a sharp supply-demand imbalance during that period, and violating the constraints will lead to significant systemic risks or a surge in costs; conversely, if π is small... t A value close to 0 indicates that the system is in a relaxed state (slack) due to abundant water inflow or low water demand during that period, resulting in significant regulatory redundancy. By calculating and outputting this sequence, the global layer transforms the complex inter-period Bruker optimization results into a set of concise marginal cost signals, providing a clear redline indication for local scheduling.

[0142] Example 5 details the construction of the local scheduling decision layer and its coupling mechanism with the global risk signal. By embedding a hard constraint mechanism into the Soft Actor-Critic (SAC) reinforcement learning framework, it addresses the problem that traditional reinforcement learning algorithms struggle to guarantee constraint satisfaction in safety-critical domains. Figure 4 As shown.

[0143] Step 501: Construct a constraint-enhanced deep reinforcement learning policy optimization module in the local scheduling decision layer, and use the Soft Actor-Critic algorithm as the core algorithm for policy generation to establish a policy network for mapping the state space of the water resource system to the action space of scheduling instructions.

[0144] Specifically, the local scheduling decision layer is primarily responsible for fine-grained operations on short scales (such as daily or real-time). Within this layer, an agent based on a Markov Decision Process (MDP) is constructed. The constrained reinforcement learning policy optimization module employs the SAC algorithm, an off-policy algorithm based on maximum entropy theory, which achieves a good balance between exploration and exploitation. This module includes a policy network (Actor) and a value network (Critic). In a specific configuration of this embodiment, both the policy network and the value network use a fully connected neural network structure, including an input layer, 2 to 3 hidden layers, and an output layer. Each hidden layer has 64 to 256 neurons, and the ReLU activation function is preferred. The output of the policy network generates continuous action instructions using a Gaussian distribution reparameterization technique.

[0145] Step 502: Using the stage risk shadow quantity (stage risk quantification parameter) as an extended dimension of the state space, construct an extended state space that includes the state of the water resource system and the intensity of risk constraints.

[0146] Specifically, traditional MDP state definitions typically only include physical states. In this embodiment, to achieve cross-layer collaboration, the state space has been significantly expanded. The original water resource system state s t Specifically, this includes: the current lake water level H t Current storage capacity V t The current inflow rate Q to the lake in_t Water demand deviation ΔD in each zone t And the current time index.

[0147] Based on this, the stage-specific risk shadow quantity π corresponding to the current time period output by the global layer will be... t As an independent feature dimension, it is added to the state vector to form the extended state s. t ext =[H t V t Qin_t ,ΔD t ,...,π t Through state extension, reinforcement learning agents, when observing their environment, can see not only how much water there is (physical state), but also how expensive / dangerous using water is (risk state). For example, when π t When the risk level suddenly rises, the agent can immediately sense that it is in a high-risk period and tends to take conservative scheduling actions.

[0148] Step 503: Embed a staged risk quantification parameter penalty term into the reward function of the Soft Actor-Critic algorithm, and transform the staged risk shadow quantity (staged risk quantification parameter) into a hard constraint through the Lagrange multiplier dynamic adjustment mechanism, so as to guide the policy network to generate a dynamic scheduling policy that satisfies the global resource allocation risk constraint in the extended state space.

[0149] Specifically, to ensure that the generated actions strictly conform to the global risk requirements, the original reward function r is modified. t (Usually calculated based on direct benefits such as power generation and water supply) is corrected to construct a reward function r' that includes risk penalties. t The corrected formula is:

[0150] r' t =r t -π t *max(0,g t (s t ,a t ))-λ Lag *max(0,C hard (s t ,a t )).

[0151] Where, r' t The corrected reward function; r t As a reward for the original operating efficiency; π t The shadow quantity of phased risks transmitted at the global layer; s t The original state of the water resource system; a t For scheduling actions; g t (s t ,a t ) represents the soft constraint indicator corresponding to global risk; λ Lag For dynamic Lagrange multipliers with hard constraints; C hard (s t ,a t ) refers to violations of physical hard constraints that must be strictly followed (such as the maximum water level limit).

[0152] This section introduces a dynamic adjustment mechanism for Lagrange multipliers, which involves setting a learnable multiplier λ for hard-constrained terms. Lag During training, λ is dynamically updated based on the degree of constraint violation (Constraint_Violation). Lag :

[0153] λ Lag <-λ Lag +α lr *(Constraint_Violation), where α lr λ is the learning rate. When the action generated by the policy violates the constraints, λ... Lag It will automatically increase, applying stronger penalties to force the policy network to update parameters to avoid violations. The resulting dynamic scheduling policy a t =Policy(s t ext Specifically, this includes the outflow of water from the lake, Q. out and water supply Q in each zone sup It can maximize cumulative operational benefits while meeting hard physical constraints and global risk constraints.

[0154] Example 6 details how to optimize asymmetric robust control parameters through a closed-loop feedback mechanism, and how to generate a multi-agent acceptable collaborative scheme.

[0155] Step 601: Introduce preset asymmetric robust control parameters and embed the asymmetric robust control parameters as adjustment factors into the boundary construction process of the asymmetric inflow robust uncertainty set and the asymmetric demand robust uncertainty set.

[0156] Specifically, to overcome the problem of over-conservatism or risk exposure that may be caused by static robustness boundaries, two adjustable scalar parameters are introduced: the inflow robustness control parameter β. Q and water demand robust control parameter β D Two parameters are embedded as adjustment factors in the boundary construction formula. The modified formula for calculating the boundary expansion coefficient is as follows:

[0157] k Q =1+β Q *η*α Q ;

[0158] k D =1+β D *η*α D ;

[0159] Among them, k' Q k' is the corrected boundary expansion coefficient of the incoming water; D β is the corrected water demand boundary expansion coefficient; QFor robust control parameters of incoming water; β D For water demand robust control parameters; η is the basic expansion parameter; α Q Contribution rate to inflow water risk exposure; α D Contribution rate to water demand risk exposure.

[0160] By adjusting β Q and β D The value of β can be dynamically scaled to cover the range of uncertainties. Initially, β can be... Q and β D All are set to 1.0.

[0161] Step 602: Run the dynamic scheduling strategy based on the joint scenario tree, and statistically analyze the strategy operation response results under the current asymmetric robust control parameters. The strategy operation response results include at least the violation frequency of key constraints and the overall system performance indicators.

[0162] Specifically, a dynamic scheduling strategy is used to conduct simulations on all branches of the joint scenario tree. The total number of violations of various key constraints (such as the minimum water level constraint) during the simulation is counted, and the constraint violation frequency F is calculated. Simultaneously, the overall system benefit indicators B, such as cumulative water supply and power generation, as well as the risk penalty value R incurred due to constraint violations, are recorded. These indicators constitute the strategy operation response results, reflecting the actual system performance under the current parameter β configuration.

[0163] Step 603: Calculate the numerical sensitivity gradient of the asymmetric robust control parameters to the violation frequency of key constraints.

[0164] This step and subsequent steps describe in detail the gradient calculation process.

[0165] Calculate the numerical sensitivity gradient of asymmetric robust control parameters to the violation frequency of key constraints, specifically including:

[0166] Construct a constraint violation measurement function, which is a weighted sum of the frequency of key constraint violations, the expected value of constraint violation magnitude obtained from the statistical analysis of the policy operation response results, and the risk penalty indicator;

[0167] The partial derivatives of the constraint violation metric function with respect to the asymmetric robust control parameters are calculated using the finite difference method to obtain the numerical sensitivity gradient.

[0168] To comprehensively consider the optimization direction, we first construct a composite constraint violation metric function V.

[0169] V=ω f *F+ω m *M+ω r *R;

[0170] Where V is the constraint violation metric function, F is the frequency of critical constraint violations, M is the expected value of the constraint violation magnitude, such as the average depth below the dead water level, and R is the risk penalty index; ω f ω m ω r These are the preset normalized weighting coefficients for frequency, amplitude, and risk, respectively.

[0171] Since the relationship between the function V and the parameter β is highly nonlinear and lacks an analytical expression, this embodiment uses the finite difference method to approximate the numerical sensitivity gradient. The specific calculation formula is as follows:

[0172] ;

[0173] in, Let the metric function V be related to the parameter β. Q The partial derivative (numerical sensitivity gradient) of δ' is a small perturbation step size, such as 0.01.

[0174] Similarly, we can calculate with respect to β D The gradient.

[0175] Step 604: Based on the numerical sensitivity gradient, the asymmetric robust control parameters are updated collaboratively and adaptively until the preset convergence condition is reached, thus forming an adaptive dual robust risk constraint system.

[0176] Specifically, a variant of gradient descent is used for parameter updates. The update formula is:

[0177] ;

[0178] Where, β Q k+1 β is the robust control parameter for the incoming water in the (k+1)th iteration. Q k Let be the parameter value for the k-th iteration; γ is the update step size for gradient descent. Similarly, the formula for the water-required robust control parameters is as follows:

[0179] .

[0180] This process will continue to repeat until a convergence condition is met: for example, the change in the function value V between two consecutive iterations is less than a threshold ε, or the gradient magnitude is close to 0, or the maximum number of iterations is reached. The parameter β obtained at this point... Q * and β D * This is the optimal adaptive parameter, and the corresponding risk constraint system can better balance safety and robustness.

[0181] Step 605: Quantify the expected system benefits of the dynamic scheduling strategy obtained under the adaptive dual robust risk constraint system, the constraint failure risk under extreme supply and demand mismatch scenarios, and the sensitivity of scheduling performance to parameter changes.

[0182] Specifically, for the finalized strategy, its expected system benefit B is calculated, such as the average annual water supply revenue. Its constraint failure risk R is calculated, such as the water shortage rate in drought years. Furthermore, a sensitivity index S needs to be calculated to measure the strategy's robustness to parameter fluctuations. Sensitivity S is defined as:

[0183] ;

[0184] Where S is the sensitivity index of scheduling performance to parameter changes; B is the expected benefit of the system; and R is the risk of constraint failure. and These are the partial derivatives of benefits and risks with respect to the robustness parameter, respectively.

[0185] The smaller the S value, the less the scheduling effect is affected by the parameter settings, and the more stable the strategy is.

[0186] Step 606: Construct a three-dimensional joint mapping space that includes risk dimension, benefit dimension and sensitivity dimension, and use the three-dimensional joint mapping space as a multi-subject common decision domain to determine a collaborative scheduling scheme that meets the preset multi-subject consistent acceptance conditions.

[0187] Specifically, the (R,B,S) triplet obtained from the above calculation is mapped to a three-dimensional coordinate system. For different parameter configuration combinations, a set of points or a surface can be formed in space. Different management entities within the watershed have different preferences: flood control departments pursue lower risk R, water supply departments pursue higher benefits B, while specific implementing agencies may prefer lower sensitivity S to reduce the workload of parameter tuning.

[0188] Within this common decision domain, a set of solutions satisfying the Pareto optimality property is searched. The rules for determining the solution can be set as follows:

[0189] Meeting the minimum acceptable threshold for all subjects (e.g., R) <R max And B>B min Under the premise of [missing information], the scheme with the lowest sensitivity S is selected first, or the scheme with the closest Euclidean distance to the ideal point is selected. The final determined collaborative scheduling scheme will be output and applied to the actual water resource scheduling and control system.

[0190] Example 7: A computer system architecture for implementing the above method is described in detail.

[0191] This embodiment provides a robust collaborative scheduling system for water resources that considers both supply and demand uncertainties. Physically, the system includes one or more processors, a memory, and a network communication interface. The memory stores computer program instructions, which, when executed by the processor, implement the various steps of the robust collaborative scheduling method for water resources that considers both supply and demand uncertainties. Logically, the system includes:

[0192] The asymmetric uncertainty modeling module is used to assess the risk exposure contribution rate of inflow uncertainty and demand uncertainty to system risk based on pre-stored historical inflow and demand data, and to construct asymmetric inflow robust uncertainty set and asymmetric demand robust uncertainty set based on the risk exposure contribution rate.

[0193] This module is responsible for connecting to the hydrological and meteorological telemetry database, acquiring real-time and historical data, and embedding a statistical analysis engine to calculate the risk exposure contribution rate and generate an asymmetric robust uncertainty set.

[0194] The supply and demand risk coupling scenario construction module is used to calculate the supply and demand risk coupling amplification index based on the asymmetric inflow water robust uncertainty set and the asymmetric demand water robust uncertainty set, and to filter and reconstruct the joint uncertainty space based on the supply and demand risk coupling amplification index to generate a joint scenario tree.

[0195] This module contains a high-performance computing unit for performing hierarchical Latin hypercube sampling and correlation analysis, calculating the RCCI index, and generating a joint scenario tree using clustering algorithms.

[0196] The global resource allocation module is used to optimize intertemporal resource allocation based on the global resource allocation layer of the pre-built water resource hierarchical collaborative scheduling model and the joint scenario tree, and to determine the stage-specific risk quantification parameters that characterize the risk constraint intensity of each stage.

[0197] This module is equipped with interfaces for large-scale optimization solvers, such as commercial optimization solvers like Gurobi or CPLEX, to solve two-stage biblical bar optimization models and extract Lagrange multipliers as risk shadow quantities.

[0198] The local collaborative scheduling decision module is used to utilize the local scheduling decision layer of the water resource hierarchical collaborative scheduling model, and use the phased risk quantification parameters as cross-layer risk constraints to generate a dynamic scheduling strategy that meets the requirements of global risk control.

[0199] This module is built on a deep learning framework (such as PyTorch or TensorFlow) and loads a pre-trained SAC policy network. During actual runtime, it receives real-time water level and flow sensor data, as well as risk signals from a global module, and generates gate opening commands in milliseconds.

[0200] Through the above embodiments, the present invention constructs a complete, physically meaningful, and practically applicable water resource collaborative robust scheduling technology method, which effectively solves the scheduling problem under complex and uncertain environments.

[0201] Example 8 provides an alternative to the calculation of dynamic correlation of supply and demand deviations, and elaborates on the specific method of correcting dynamic correlation under extreme supply and demand mismatch scenarios, so as to solve the problem that conventional correlation calculation may be distorted when the sample is extremely scarce.

[0202] Step 801: Based on the relative deviation sequence of incoming water and the relative deviation sequence of demand water, identify extreme samples and calculate the correlation of extreme samples.

[0203] Specifically, in order to capture the supply and demand linkage characteristics under extreme weather conditions, an extreme threshold θ for incoming water is first set. Q and water demand extreme threshold θ D For example, we can take the 5th and 95th quantiles of the deviation sequence, respectively. For any sample point within the sliding time window, if the absolute value of its inflow deviation is greater than θ... Q Or the absolute value of the water demand deviation is greater than θ D If the extreme sample is not found, it is marked as an extreme sample. For each extreme sample within the statistical window, its Spearman rank correlation coefficient is calculated individually and denoted as the extreme sample correlation ρ. extreme .

[0204] Step 802: Using the extreme scenario weighting factor, the normal correlation and the extreme sample correlation are weighted and fused to obtain the corrected dynamic correlation of supply and demand deviation.

[0205] Specifically, in order to balance overall trends and extreme risks, a weighted fusion formula is used to calculate the final dynamic correlation:

[0206] ρ final =(1-γ)*ρ normal +γ*ρ extreme .

[0207] Where ρfinal represents the corrected final dynamic correlation of the supply-demand deviation, ρ normal For standard full-sample correlation, γ is the extreme scenario weighting factor, for example, a value between 0.2 and 0.4. If the number of extreme samples within the window is insufficient, for example, less than 10, then γ is automatically set to 0, and ρ... extreme This is the correlation calculated separately based on extreme samples.

[0208] This correction mechanism enables the final input RCCI index to more sensitively reflect the supply and demand resonance risk when extreme drought and high temperature heat waves occur simultaneously, further improving the robustness of scenario construction.

[0209] Example 9: An alternative scheme for a robust water resource scheduling method considering both supply and demand uncertainties is provided, comprising the following steps:

[0210] Step 901: Construct a disturbance space for uncertainties in inflow and demand, generate uncertainty disturbance samples within the disturbance space, propose an asymmetric robust boundary generation mechanism based on risk exposure contribution rate, simulate the operation of uncertainty disturbance samples under the asymmetric robust boundary generation mechanism, and form a robust uncertainty set for inflow and a robust uncertainty set for demand.

[0211] Step 902: Taking into account the combined probability of the occurrence of abnormal water inflow and water demand deviation, as well as the amplification effect of the combined supply and demand fluctuations on the contraction of the scheduling feasible domain, construct the supply and demand risk coupling amplification index RCCI. Based on the supply and demand risk coupling amplification index, perform risk-driven screening and scenario reconstruction on the robust uncertainty set of water inflow and the robust uncertainty set of water demand, and generate a joint scenario tree focusing on high-risk supply and demand mismatch states.

[0212] Step 903: Using the joint scenario tree as the only uncertainty input, construct a risk-decision strongly coupled hierarchical collaborative scheduling model for water resources. In this model, the global resource allocation layer adopts a two-stage sub-Bruker optimization framework to characterize the risk dependency relationship of inter-period resource allocation and obtain the risk shadow quantity of each stage. The local scheduling decision layer introduces a constraint-enhanced deep reinforcement learning strategy optimization module, and uses the Soft Actor-Critic algorithm combined with the risk shadow quantity hard constraint embedding mechanism to use the stage risk shadow quantity output by the global layer as the extended dimension of the state space to generate a dynamic scheduling strategy.

[0213] Step 904: Introduce asymmetric robust control parameter β Q With β D Constraint optimization of the dynamic scheduling strategy is performed to obtain the strategy response results. Based on the strategy response results, a sensitivity gradient collaborative adaptive correction mechanism is proposed to synchronously update β. Q With β D An adaptive dual robust risk constraint system is formed, a risk-benefit-sensitivity joint mapping mechanism is proposed, and a three-dimensional joint mapping space is constructed as a multi-subject common decision domain to obtain a collaborative scheduling scheme with minimum risk conflict and maximum policy tolerance.

[0214] According to one aspect of this application, in the process of forming the robust uncertainty set of inflow water and the robust uncertainty set of demand water, determining the baseline uncertainty sets of inflow water and demand water further includes:

[0215] Collect historical water inflow data from multiple sources, including rainfall, runoff, and outflow data from upstream reservoirs, and perform data cleaning and standardization to form a water inflow disturbance sample set;

[0216] Collect historical water demand data from multiple sources, including agricultural, industrial, domestic and ecological water demand data, and perform data cleaning and standardization to form a water demand disturbance sample set;

[0217] Based on the inflow disturbance sample set and the demand disturbance sample set, inflow disturbance space and demand disturbance space are constructed respectively, and multiple sets of corresponding uncertainty disturbance samples are generated. At the same time, the baseline uncertainty set of inflow and demand is determined in advance according to the statistical characteristics of historical data, as the initial reference range for subsequent asymmetric boundary expansion.

[0218] According to one aspect of this application, during the process of forming the robust uncertainty set of inflow and the robust uncertainty set of demand, the contribution rate of inflow uncertainty and demand uncertainty to the risk exposure of the scheduling system is evaluated, further as follows:

[0219] Using multiple sets of uncertain disturbance samples as input, the operation of the scheduling system under the conditions of water inflow disturbance-dominated, water demand disturbance-dominated, and combined supply and demand disturbance is simulated respectively, and the triggering frequency of key system constraints under various disturbance scenarios is statistically analyzed;

[0220] Calculate water supply failure risk indicators based on the trigger frequency of key system constraints, including the probability of water shortage and the expected value of water shortage;

[0221] By combining the trigger frequency of key constraints with water supply failure risk indicators, the contribution rate of water inflow uncertainty and water demand uncertainty to the overall risk exposure of the system is quantified.

[0222] According to one aspect of this application, in the process of forming the robust uncertainty set of incoming water and the robust uncertainty set of demand water, an asymmetric robust uncertainty set of incoming water and robust uncertainty set of demand water with risk-sensitive differences are formed, further as follows:

[0223] The risk exposure contribution rate is introduced as a weighting factor, and the inflow boundary expansion coefficient and the demand boundary expansion coefficient are calculated separately.

[0224] Based on the baseline uncertainty set, the inflow uncertainty boundary and the demand uncertainty boundary are expanded outward in a directional manner according to the expansion coefficient, and the inflow robust uncertainty set and the demand robust uncertainty set are generated independently.

[0225] In the scheduling model, the robust uncertainty set of inflow water and the robust uncertainty set of demand water are input respectively. The independent regulation effect is verified through simulation operation to ensure that the adjustment of the uncertainty boundary of inflow water does not interfere with the uncertainty boundary of demand water.

[0226] According to one aspect of this application, a joint scenario tree focusing on high-risk supply-demand mismatch states is generated, further comprising:

[0227] Construct a supply and demand risk amplification index (RCCI), while also considering abnormal water inflow conditions (Q).S Deviation from water demand state D S The joint occurrence probability P(Q) S D S The sensitive amplification effect of scheduling feasible region shrinkage and the dynamic correlation ρ between supply and demand deviations. Q,D ;

[0228] RCCI S =P(Q S D S )·(1+λ'·ΔΩ S )·(1+μ·ρ Q,D );

[0229] Where λ' and μ are the coupling amplification adjustment coefficients, ΔΩ S =(Ω0-Ω S ) / Ω0, Ω0 and Ω S Let S represent the scheduling feasible domain volume under the baseline scenario and scenario S, respectively.

[0230] In this embodiment, under conditions where both inflow and demand are uncertain, supply and demand disturbances often amplify system risks through mechanisms such as shrinking of the feasible scheduling region and triggering multiple constraints. To characterize this amplification mechanism, based on the robust uncertainty sets of inflow and demand, the abnormal state Q of inflow is comprehensively considered. S Deviation from water demand state D S The joint occurrence probability P(Q) S D S The shrinkage ratio ΔΩ of the feasible scheduling domain volume relative to the baseline scenario under combined supply and demand disturbances. S And the dynamic correlation ρ between water inflow and water demand disturbances. Q,D Construct a supply and demand risk coupling amplification index RCCI S .

[0231] The feasible region of scheduling refers to the set of scheduling decision variables that satisfy system operation constraints under given inflow and demand scenarios. It is jointly limited by water balance constraints, water level constraints, water conveyance capacity constraints, and ecological flow constraints, forming a high-dimensional feasible solution space. Since this feasible region typically manifests as a high-dimensional convex polyhedron or approximately convex set, its volume... Since analytical solutions are difficult to find in practical engineering, this invention uses a surrogate index-based approach for calculation.

[0232] Without directly calculating the high-dimensional volume, surrogate indices such as the proportion of feasible scheduling solutions, the weighted sum of constraint violation degrees, or the tightening magnitude of the feasible region boundary can be used to characterize the relative size of the feasible region. For example, using the proportion of feasible solutions as a normalized approximation of the feasible region volume, then:

[0233] ΔΩS =1-N feasible,s / N feasible,0 ;

[0234] Where, N feasible,s N represents the number of feasible solutions. feasible,0 This represents the number of feasible solutions in the baseline scenario. This surrogate form significantly reduces computational complexity while maintaining consistency in physical meaning, making it suitable for high-dimensional scheduling problems.

[0235] Furthermore, using the robust uncertainty set of water supply and the robust uncertainty set of water demand as a joint uncertainty input space, the supply and demand risk coupling amplification index (RCCI) value is calculated for each scenario node in the space.

[0236] In this embodiment, the robust uncertainty set of water supply and the robust uncertainty set of water demand are used as the joint uncertainty input space. The corresponding supply and demand risk coupling amplification index (RCCI) value is calculated for each joint scenario node in the space. The RCCI value is used as a unified indicator to measure the impact of different supply and demand disturbance combinations on the scheduling risk of Hongze Lake, thereby realizing the quantitative ranking of the risk level of each scenario node in the joint uncertainty space.

[0237] Based on this, a risk threshold θ is set, and risk-driven screening and scenario reconstruction are performed on the joint uncertainty space according to the RCCI value. High-risk scenario nodes with RCCI≥θ are fully retained, and low-risk scenarios with RCCI<θ are clustered, compressed, and representatively merged to generate a joint scenario tree with significantly reduced dimensionality but enhanced risk focusing ability.

[0238] In this embodiment, after calculating the RCCI of the joint scenario nodes, a risk threshold θ is set. Based on the RCCI value, risk-driven screening and scenario reconstruction are performed on the joint uncertainty space. High-risk scenario nodes with RCCI not less than the risk threshold are fully retained to ensure that low-probability but highly destructive supply-demand mismatch states are fully considered in subsequent scheduling decisions. For low-risk scenario nodes with RCCI less than the risk threshold, clustering compression is performed based on the distance metric between scenario feature vectors. In each cluster, representative scenario nodes with the largest RCCI value are retained. This effectively controls the scale of scenarios while maintaining an accurate characterization of the system risk structure, ultimately generating a joint scenario tree with significantly reduced dimensionality and enhanced risk focusing capability.

[0239] According to one aspect of this application, generating a dynamic scheduling strategy further comprises:

[0240] A hierarchical collaborative scheduling model framework is constructed based on a joint scenario tree, which maps inter-period resource allocation decisions and phased real-time scheduling decisions to a global resource allocation layer and a local scheduling decision layer, respectively.

[0241] In this embodiment, a joint scenario tree is used as the sole uncertainty input. The Hongze Lake scheduling problem is divided into two coupled decision-making levels: intertemporal resource allocation and phased real-time scheduling. These are mapped to a global resource allocation layer and a local scheduling decision layer, respectively. The global resource allocation layer focuses on water allocation, storage and release strategies, and intertemporal regulation and storage trade-offs at each stage within the planning period. The local scheduling decision layer focuses on the selection of actions such as real-time water release, water diversion, and water supply allocation at each node of the joint scenario tree. By establishing a unified risk constraint transmission interface between the two layers, the phased risk intensity parameters output by the global layer can serve as exogenous constraints for the generation of local layer strategies, structurally ensuring the consistency of intertemporal allocation and real-time execution in terms of risk scale.

[0242] A two-stage stub-bar optimization model is constructed in the global resource allocation layer. The stub-bar uncertainty set based on Wasserstein distance is used to characterize the distribution bias in the joint scenario tree. The decision variable in the first stage is denoted as x (corresponding to intertemporal resource allocation and phased goal decomposition), and the decision variable in the second stage is denoted as y. s (Corresponding to remedial scheduling and corrective actions under scenario node s), and defining the Bruker uncertainty set D using Wasserstein distance, the objective function is constructed to minimize the expected risk-benefit composite cost under the worst-case distribution:

[0243] ;

[0244] Among them, c T x is the deterministic cost of the first phase, sup is the supremum, and E P For the expectation operator, q s T y s This is the second-stage cost item (also known as recourse cost or compensation cost). It satisfies the system constraint set g of each node in the joint scenario tree, which includes at least water balance constraints, reservoir capacity upper and lower limits constraints, flood control water level constraints, ecological baseflow constraints, and water conveyance capacity constraints, denoted as g. t (x,y s )≤0.

[0245] In this embodiment, the partial Bruker uncertainty set D is constructed using the 1-Wasserstein distance (W1 distance), which is defined as the distance from the empirical distribution P. * A Wasserstein sphere centered at a radius of δ:

[0246] D={P|W1(P,P * )≤δ};

[0247] Among them, P *The empirical distribution is constructed based on historical water inflow-water demand joint scenario samples (joint scenario tree samples); δ is the Wasserstein sphere radius (robust radius parameter), used to characterize the maximum deviation of the true distribution from the empirical distribution. In this embodiment, the radius δ can be estimated based on the statistical bounds based on the historical sample size and confidence level.

[0248] Since the objective function has a nested structure, in order to improve solvability, this invention adopts the dual reconstruction method of Wasserstein DRO to transform the worst-case distribution expectation term into an equivalent dual optimization problem.

[0249] Under the 1-Wasserstein distance condition, the worst-case expectation term of the inner layer distribution can be equivalently expressed as:

[0250] ;

[0251] Where C(x,y) s The total cost function (x) represents the second-stage cost (such as water shortage penalties or scheduling adjustment costs), and is the sum of the first-stage decision x and the second-stage response y. s The function, denoted as C(x,y(ξ)) under context dependence, is where ξ is the uncertainty parameter (e.g., the joint vector of incoming and demanded water); sup is the supremum; inf is the infremum; λ is the dual variable corresponding to the Wasserstein distance constraint; δ is the radius of the Wasserstein sphere; N is the number of samples; ξ i Let ||ξ-ξ| be the i-th sample point in the joint scenario tree; i || represents the norm distance in the context space.

[0252] In this embodiment, the two-stage Wasserstein bipolar optimization problem is solved using the sample average approximation (SAA) and dual reconstruction method: an empirical distribution is constructed using finite historical samples to approximate the Wasserstein dual problem, transforming it into a deterministic two-stage programming problem, and the solution is obtained by iteratively updating the dual variables.

[0253] To obtain a quantitative index of staged risk intensity that can be used for cross-layer transmission, a Lagrangian function is constructed when solving the above two-stage sub-Bruker optimization problem:

[0254] L=sup P∈D E P [C(x,y s )]+∑ t π t (sup P∈D E P [g t (x,y s )]-ε t );

[0255] Where ε t This represents the slack or risk budget of the risk constraint at stage t; it is related to the Lagrange multiplier π corresponding to the key risk constraint. t Defined as a phased risk shadow quantity, and its marginal risk meaning is characterized by the partial derivative with respect to the slack quantity:

[0256] .

[0257] The π obtained by solving t A phased risk shadow quantity sequence is formed to reflect the marginal impact of risk constraints at each stage on the global objective function, and serves as the input for subsequent local scheduling strategy training and hard constraint embedding.

[0258] Based on this, a constraint-enhanced Soft Actor-Critic reinforcement learning module is constructed in the local scheduling decision layer. The phased risk shadow quantity sequence is used as an extended dimension of the state space, and a shadow quantity penalty term is embedded in the reward function. The risk shadow quantity is transformed into a hard constraint through a Lagrange multiplier dynamic adjustment mechanism.

[0259] In this embodiment, a constraint-enhanced Soft Actor-Critic (SAC) reinforcement learning module is constructed in the local scheduling decision layer to abstract the real-time scheduling process of Hongze Lake into a Markov decision process <S,A,P,R>.

[0260] Here, the state space S is defined as a continuous vector space, and the state vector s t It should include at least the following components:

[0261] s t =[H t V t Q in_t ,ΔD t ,κ t ];

[0262] Among them, H t V represents the lake's water level. t Q represents the storage capacity. in_t Indicates the inflow rate to the lake, ΔD t Indicates the water demand deviation of a zone, κ t This indicates the node identifier or stage index of the current stage in the joint scenario tree.

[0263] To achieve cross-layer risk consistency constraints, the staged risk shadow quantity π output by the global resource allocation layer is used. t Embedded as a state extension dimension in the local scheduling layer, forming the extended state s t ext =[s t ,πt This enables local strategies to explicitly perceive the strength of global risk constraints at the current stage when making decisions.

[0264] Action space A is defined as a continuous action space, used to represent the real-time scheduling decisions of Hongze Lake in time period t. The action vector is represented as: a t =[Q out_t Q sup_t,1 ,…,Q sup_t,m ]. Among them, Q out_t Q represents the amount of water released from the lake. sup_t,i Let represent the water supply volume of the i-th water supply zone. All action components are constrained by the physically feasible interval.

[0265] The reward function in the original running profit r t Based on this, a constraint penalty term consistent with the risk shadow quantity is introduced to construct the risk perception reward: r t '=r t -π t ·max(0,g t (s t ,a t Among them, g t (s t ,a t ) indicates a local action a t This measures the violation of key constraints (including upper water level limits, ecological baseflow, supply and demand balance, etc.) at stage t. When the system is in a high-risk stage, the penalty for constraint violations in the reward function will be automatically increased.

[0266] Meanwhile, to further strengthen the soft penalty into a hard constraint, a dynamic adjustment mechanism of Lagrange multipliers is introduced in the parameter update process of SAC. The multipliers corresponding to the constraint violation terms are adaptively updated, so that the samples that violate the constraint are continuously suppressed in the policy gradient update. This guides the learned policy to achieve an adaptive balance between maximum entropy exploration and constraint feasibility, and ensures that the local policy does not deviate from the global risk control objective at the execution level.

[0267] Furthermore, by transmitting risk shadow quantities between the global resource allocation layer and the local scheduling decision layer, cross-layer risk consistency constraints are achieved, forming a hierarchical collaborative scheduling model with strong risk-decision coupling.

[0268] In this embodiment, the risk shadow quantity π t Serving as a unified risk interface between the global and local layers, it permeates all stages of the joint scenario tree and updates dynamically with each stage. The global layer utilizes π... t The strength of the phased risk constraint is output in a computable form; the local layer extends the state s. t ext With reward modifier r tThe risk intensity is explicitly embedded into the policy generation and policy update process.

[0269] Furthermore, the reinforcement learning policy network is trained on each node of the joint scenario tree to generate a dynamic scheduling decision sequence that satisfies the global risk shadow quantity constraint.

[0270] In this embodiment, training sample sequences are constructed at each node of the joint scenario tree, and state-action-reward-next state sample quadruples are generated through scenario node transitions and environmental simulation: (s t ext ,a t ,r t ',s t+1 ext The local policy is iteratively trained using a SoftActor-Critic dual-Q network and policy network architecture.

[0271] The SAC network structure is implemented using existing technologies:

[0272] Both the policy network (Actor) and the value network (Critic) are multi-layer feedforward neural networks;

[0273] The network consists of an input layer, several hidden layers, and an output layer;

[0274] The number of hidden layers is 2-3, and the number of neurons in each layer is 64-256.

[0275] The activation function uses ReLU or its equivalent variants;

[0276] Critic employs a dual-Q network structure to reduce the overvaluation bias;

[0277] The policy network outputs actions to construct a continuous action distribution and achieves differentiability through reparameterization techniques.

[0278] Key hyperparameters used during training include, but are not limited to: learning rates of the policy network and value network, experience replay batch size, discount factor, entropy regularization coefficient, and soft update coefficient of the target network. Their specific values ​​can be set according to the actual scheduling system scale and stability requirements, and are optional implementations of this invention.

[0279] After training, the output is a dynamic scheduling decision sequence {α} that satisfies the global risk shadow quantity constraint. t This sequence can adaptively generate real-time scheduling actions for Hongze Lake on the uncertain evolution path given by the joint scenario tree, and achieve a balance between risk control and operational efficiency.

[0280] In actual scheduling operations, the level of uncertainty and the state of system risk are dynamically changing. If the robust control parameters are set fixedly, it is easy to cause a decline in resource utilization efficiency or the accumulation and exposure of risks in long-term operation. To address the above problems, this invention introduces a robust control parameter β for incoming water based on the hierarchical collaborative scheduling model. Q With water demand robust control parameter β D Asymmetric robust constraint optimization is performed on the dynamic scheduling strategy, and based on the strategy's operational response under the joint scenario tree, a sensitivity gradient collaborative adaptive correction mechanism is proposed to synchronously update β. Q With β D This forms an adaptive dual robust risk constraint system. Furthermore, by constructing a three-dimensional joint mapping space of risk-benefit-sensitivity, the complex scheduling results are mapped into a common decision domain that can be jointly understood and negotiated by multiple subjects. Under different risk tolerance and benefit preferences, a collaborative scheduling scheme with minimum risk conflict and maximum strategy tolerance is determined.

[0281] According to one aspect of this application, a cooperative scheduling scheme with minimum risk conflict and maximum strategy tolerance is obtained, further comprising:

[0282] Future water robust control parameter β Q The boundary adjustment factor of the embedded water robust uncertainty set will be used to adjust the water demand robust control parameter β. D By embedding boundary adjustment factors into the water-demand-robust-uncertain set, a dual-robust-risk constraint system is formed.

[0283] In this embodiment, a robust control parameter β for incoming water is introduced. Q With water demand robust control parameter β D These are used as boundary adjustment factors for the robust uncertainty sets of inflow and demand water, respectively, enabling the robust boundary to be adjustable according to the risk state during operation; specifically, β... Q The adjustment function embedded in the boundary of the uncertain inflow set will β D An adjustment function is embedded in the boundary of the uncertain water demand set to form a dual robust risk constraint system with two adjustable parameters. This system can enhance robust protection when the risk of supply and demand mismatch increases, and weaken robust conservatism to improve resource utilization efficiency when the risk decreases or the system redundancy is large.

[0284] Based on the running response results of the dynamic scheduling strategy at each node of the joint scenario tree, the constraint violation frequency, redundancy and overall performance indicators are calculated.

[0285] In this embodiment, a dynamic scheduling strategy is run on each node of the joint scenario tree to record the key constraint responses of the Hongze Lake scheduling system under different stages and scenarios. The number and magnitude of key constraints violations within a given time window are statistically analyzed and normalized into a constraint violation frequency index. At the same time, the redundancy index of the system under the constraint conditions is calculated, such as reservoir capacity safety margin, ecological baseflow satisfaction margin, and water supply guarantee margin. The overall performance index, including water supply benefits, storage benefits, ecological benefits, and risk penalties, is obtained as the basis for subsequent robust control parameter sensitivity assessment and collaborative updates.

[0286] β is calculated based on the runtime response results. Q With β D Numerical sensitivity gradient for constraint violation frequency, and based on the sensitivity gradient, β Q With β D Perform collaborative adaptive updates to form an adaptive dual robust risk constraint system;

[0287] In this embodiment, a constraint violation metric function V(β) is constructed regarding the robust control parameters. Q ,β D ),

[0288] V(β Q ,β D )=ω f ·F(β Q ,β D )+ω m ·M(β Q ,β D )+ω r ·R(β Q ,β D );

[0289] Wherein, F(β) Q ,β D M(β) represents the weighted sum of the frequencies of critical constraint violations. Q ,β D R(β) represents the expected value of the constraint violation magnitude or over-limit degree. Q ,β D ω represents the risk penalty indicator caused by constraint violation. f ω m ω r This is a weighting factor used to balance the importance of frequency, amplitude, and risk.

[0290] To achieve adaptive adjustment of parameters, the constraint violation metric function V is calculated with respect to β. Q With β D The numerical sensitivity gradient is calculated, and parameters are updated synchronously using a cooperative gradient descent approach:

[0291] ;

[0292] ;

[0293] Where k is the iteration round and γ is the update step size.

[0294] In implementation, the gradient can be approximated using finite difference, for example:

[0295] ;

[0296] Among them, the disturbance δ' is a positive decimal, and its typical value is β. Q The current value can be 1%-5%, or an adaptive perturbation that gradually decreases with iteration can be used. The update step size γ can be fixed or adaptively adjusted according to the gradient magnitude. For example, γ can be increased appropriately when V decreases monotonically and decreased when V oscillates, in order to ensure the stability of the update process.

[0297] The iteration termination condition includes at least one of the following conditions:

[0298] 1) The change in V between two consecutive iterations is less than the preset threshold ε V ;

[0299] 2) β Q With β D The update magnitudes are all less than the preset threshold ε β ;

[0300] 3) The number of iterations reaches the maximum allowed number K. max .

[0301] When one of the above conditions is met, stop updating the parameters and obtain the final adaptive dual robust risk constraint parameters.

[0302] Based on the final scheduling strategy results obtained under the updated adaptive dual robust risk constraint system, the expected system benefit index B, the constraint failure risk index R under extreme supply-demand mismatch scenarios, and the sensitivity index S of scheduling performance to changes in robust parameters are quantified respectively. These three are then mapped into a three-dimensional joint mapping space Ψ={(R,B,S)}, where the sensitivity index can be taken as... , , , Combinatorial functions:

[0303] ,

[0304] It is used to characterize the stability and tolerance of the strategy to parameter disturbances, and to transform complex scheduling results into a common evaluation coordinate system that can be understood by multiple subjects.

[0305] The three-dimensional joint mapping space is used as the common decision domain for generating multi-agent collaborative scheduling strategies. By restricting strategy search to be conducted only within the common decision domain, a collaborative scheduling scheme that satisfies the condition of unanimous acceptance by multiple agents is determined.

[0306] In this embodiment, the multiple stakeholders include at least a flood control management stakeholder, a water supply management stakeholder, and an ecological protection stakeholder. Each stakeholder focuses on different aspects of the risk indicator R, benefit indicator B, and stability indicator S, respectively. The consensus condition for multiple stakeholders is defined as: selecting a scheduling scheme in the three-dimensional joint mapping space that satisfies the following conditions:

[0307] 1) The risk indicator R does not exceed the maximum acceptable risk threshold for each entity;

[0308] 2) The efficiency index B shall not be lower than the minimum acceptable efficiency level for each entity;

[0309] 3) The sensitivity index S does not exceed the preset stability threshold;

[0310] 4) Among the candidate solutions that satisfy the above constraints, the selected solution is a Pareto non-dominated solution, that is, there is no alternative solution that can further improve other indicators without reducing any principal indicator.

[0311] In terms of implementation, the preferred approach is to adopt a multi-objective selection criterion that prioritizes risk minimization and benefits maximization while simultaneously constraining sensitivity to no more than a threshold. This approach determines a collaborative scheduling scheme that satisfies the condition of unanimous acceptance by multiple stakeholders, enabling the scheme to achieve collaborative and robust scheduling under conditions of minimal risk conflict and maximum policy tolerance in the actual operation of Hongze Lake.

[0312] This application employs an asymmetric modeling mechanism based on risk exposure contribution rates. By simulating the heterogeneous contribution weights of inflow and demand to system failure risk, it differentiates the boundary expansion coefficient of the uncertainty set, reserving sufficient margin in the direction of higher-risk disturbances and shrinking the boundary in the direction of lower-risk disturbances, thus achieving an accurate balance between safety and economy. This solves the problem of overly conservative or wasteful resource allocation caused by the symmetric setting of traditional robust sets.

[0313] This application constructs a Supply-Demand Risk Coupling Amplification Index (RCCI). By introducing physical geometric indicators of the contraction magnitude of the scheduling feasible region and the dynamic correlation of supply and demand deviations, it accurately identifies and fully preserves highly destructive scenarios that, although low in probability, lead to a sharp compression of the solution space, effectively avoiding vulnerabilities in extreme risk defense. It addresses the problem in existing scenario construction methods that rely solely on probabilistic statistics while ignoring the contraction effect of physical constraints, resulting in the erroneous exclusion of extreme risk scenarios.

[0314] This application establishes a hierarchical collaborative scheduling model, utilizing the risk shadow quantity (Lagrange multiplier) extracted by global layer split-bar optimization. This risk shadow quantity is used as the state extension dimension and hard penalty constraint for local layer reinforcement learning, forcing the micro-level decision-making agent to perceive and abide by the macro-level risk boundary, thus achieving cross-level risk closed-loop control. This solves the problem of the lack of an explicit risk transmission mechanism between long-scale planning and short-scale real-time scheduling.

[0315] 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 robust collaborative scheduling method for water resources considering both supply and demand uncertainties, characterized in that, include: Based on pre-stored historical water inflow data and historical water demand data, the risk exposure contribution rate of water inflow uncertainty and water demand uncertainty to system risk is assessed, and an asymmetric water inflow robust uncertainty set and an asymmetric water demand robust uncertainty set are constructed accordingly. Based on the asymmetric inflow robust uncertainty set and the asymmetric demand robust uncertainty set, the supply and demand risk coupling amplification index is calculated, and the joint uncertainty space is screened and reconstructed accordingly to generate a joint scenario tree; Input the joint scenario tree into the pre-built hierarchical collaborative scheduling model for water resources; By utilizing the global resource allocation layer of the water resource hierarchical collaborative scheduling model, cross-period resource allocation optimization is performed based on the joint scenario tree to determine the stage-specific risk quantification parameters that characterize the risk constraint intensity at each stage; By utilizing the local scheduling decision layer of the water resources hierarchical collaborative scheduling model, the phased risk quantification parameters are used as cross-layer risk constraints to generate a dynamic scheduling strategy that meets the requirements of global risk control. Assess the contribution of inflow uncertainty and demand uncertainty to the risk exposure of system risk, including: A disturbance space is constructed based on historical water inflow and water demand data, and uncertain disturbance samples are generated. Uncertainty disturbance samples are input into the scheduling system for simulation, the trigger frequency of key constraints under different disturbance scenarios is statistically analyzed, and water supply failure risk indicators are calculated. Based on the triggering frequency of key constraints and the risk index of water supply failure, the risk exposure intensity under the scenarios dominated by inflow disturbance and demand disturbance is quantified respectively, and the contribution rate of inflow risk exposure and the contribution rate of demand risk exposure are calculated accordingly. The calculation of the supply-demand risk coupling amplification index includes: Using the asymmetric inflow robust uncertainty set and the asymmetric demand robust uncertainty set as boundaries, a discretized joint uncertainty input space is constructed. For each scenario node in the joint uncertainty input space, the corresponding supply and demand risk coupling amplification index is calculated to characterize the comprehensive amplification effect of the joint supply and demand disturbance on the system risk under that scenario node; The supply-demand risk coupling amplification index comprehensively considers the following three dimensions: The joint probability of abnormal water inflow and water demand deviation under scenario nodes; The combined supply and demand disturbances cause the scheduling feasible region to shrink by a smaller margin relative to the baseline scenario, resulting in a reduction in the solution space for the system to satisfy operational constraints. Dynamic correlation between the relative deviation series of inflow water and the relative deviation series of demand water in the current period; It also includes adaptive optimization of dynamic scheduling strategies: An asymmetric robust control parameter is introduced and embedded as an adjustment factor in the boundary construction process of the asymmetric inflow robust uncertainty set and the asymmetric demand robust uncertainty set; Based on the joint scenario tree, the dynamic scheduling strategy is run, and the strategy operation response results under the current asymmetric robust control parameters are statistically analyzed, including the violation frequency of key constraints and the overall system performance indicators.

2. The water resource collaborative robust scheduling method considering both supply and demand uncertainties as described in claim 1, characterized in that, Constructing asymmetric inflow robust uncertainty sets and asymmetric demand robust uncertainty sets, including: Based on the statistical characteristics of historical water inflow data and historical water demand data, the uncertain sets of water inflow benchmark and water demand benchmark are determined. Using the contribution rate of inflow risk exposure and the contribution rate of demand risk exposure as weighting factors, the inflow boundary expansion coefficient and the demand boundary expansion coefficient are calculated respectively. Based on the inflow boundary expansion coefficient and the demand boundary expansion coefficient, the inflow reference uncertainty set and the demand reference uncertainty set are expanded outward in a directional manner to obtain the asymmetric inflow robust uncertainty set and the asymmetric demand robust uncertainty set.

3. The robust collaborative scheduling method for water resources considering both supply and demand uncertainties as described in claim 1, characterized in that, Based on the supply-demand risk coupling amplification index, the joint uncertainty space is screened and reconstructed to generate a joint scenario tree, including: Set a preset risk threshold, identify scenario nodes with a supply and demand risk coupling amplification index greater than or equal to the preset risk threshold as high-risk scenario nodes, and retain them completely; Scenario nodes with a supply-demand risk coupling amplification index less than a preset risk threshold are identified as low-risk scenario nodes, and they are clustered and compressed to extract representative scenario nodes. High-risk scenario nodes are merged with representative scenario nodes to generate a joint scenario tree.

4. A robust water resource scheduling method considering both supply and demand uncertainties as described in claim 1, characterized in that, Utilizing the global resource allocation layer of a hierarchical collaborative water resource scheduling model, intertemporal resource allocation optimization is performed based on a joint scenario tree, including: A two-stage sub-Brussels bar optimization model is constructed in the global resource allocation layer, and the distribution bias in the joint scenario tree is characterized by the sub-Brussels bar uncertainty set based on Wasserstein distance. With the objective of minimizing the expected risk-benefit composite cost under the worst-case distribution scenario, a two-stage sub-Bruker optimization model is solved to obtain intertemporal resource allocation decisions, which are used to determine the phased risk quantification parameters.

5. A robust water resource scheduling method considering both supply and demand uncertainties according to claim 4, characterized in that, Determine the phased risk quantification parameters that characterize the intensity of risk constraints at each stage, including: A Lagrangian function is constructed for the stage-specific risk constraints in the two-stage sub-Bruker optimization model, and Lagrangian multipliers corresponding to the stage-specific risk constraints are introduced. Solve for the Lagrange function and define the resulting sequence of Lagrange multipliers as a phased risk quantification parameter, which is used to quantify the marginal cost of violating risk constraints in the current phase.

6. A robust water resource scheduling method considering both supply and demand uncertainties according to claim 1, characterized in that, By utilizing the local scheduling decision layer of the water resource hierarchical collaborative scheduling model, a dynamic scheduling strategy that meets the requirements of global risk control is generated, including: In the local scheduling decision layer, a constraint-enhanced deep reinforcement learning policy optimization module is constructed. The Soft Actor-Critic algorithm is used to generate policies and establish a policy network for mapping the state space of the water resource system to the action space of scheduling instructions. By using the phased risk quantification parameters as an extended dimension of the state space, an extended state space is constructed that includes the state of the water resource system and the intensity of risk constraints. The reward function of the Soft Actor-Critic algorithm is embedded with a stage-based risk quantification parameter penalty term, and the stage-based risk quantification parameter is transformed into a hard constraint through a Lagrange multiplier dynamic adjustment mechanism, which guides the policy network to generate a dynamic scheduling policy that satisfies the global resource allocation risk constraint in the extended state space.

7. A robust collaborative scheduling system for water resources that considers both supply and demand uncertainties, characterized in that, include: The asymmetric uncertainty modeling module is used to assess the risk exposure contribution rate of inflow uncertainty and demand uncertainty to system risk based on pre-stored historical inflow and demand data, and to construct asymmetric inflow robust uncertainty set and asymmetric demand robust uncertainty set accordingly. The supply and demand risk coupling scenario construction module is used to calculate the supply and demand risk coupling amplification index based on the asymmetric inflow water robust uncertainty set and the asymmetric demand water robust uncertainty set, thereby filtering and reconstructing the joint uncertainty space and generating a joint scenario tree. The global resource allocation module is used to optimize intertemporal resource allocation based on the global resource allocation layer of the pre-built water resource hierarchical collaborative scheduling model and the joint scenario tree to determine the phased risk quantification parameters that characterize the risk constraint intensity of each stage. The local collaborative scheduling decision module is used to utilize the local scheduling decision layer of the water resource hierarchical collaborative scheduling model, taking the phased risk quantification parameters as cross-layer risk constraints, and generating a dynamic scheduling strategy that meets the requirements of global risk control. The system implements a robust collaborative scheduling method for water resources that considers both supply and demand uncertainties during operation, as described in any one of claims 1-6.

Images