Reservoir group scheduling decision-making method considering multi-source uncertainty propagation and evolution tracking

Abstract

The invention discloses a reservoir group scheduling decision-making method considering multi-source uncertainty propagation and evolution tracking, relates to the technical field of reservoir scheduling, and is used for solving the problem of reservoir scheduling multi-target risk decision-making in a random environment. According to the method, a space-time two-dimensional risk deduction model is constructed, refined risk diagnosis and dynamic deduction are performed on a reservoir group scheduling process, an uncertainty propagation and evolution tracking mechanism is obtained, and the robustness and reliability of a reservoir scheduling decision in an uncertainty environment are effectively improved; and a preference demodulator FAHP-OWA which comprehensively considers subjective weight and preference conflicts is adopted to obtain a high-dimensional weight vector in the aspect of obtaining group preference information. Meanwhile, multiple rounds of interactive optimization and a weight distribution mechanism are combined, preference information of decision makers can be introduced step by step, the scheduling process is more flexible, decision groups are more harmonious, the decision efficiency and accuracy of the decision groups in an uncertain environment are improved, and accurate judgment of the advantages and disadvantages of the scheme is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of reservoir scheduling, and more specifically, to a reservoir group scheduling decision-making method considering multi-source uncertainty propagation and evolution tracking. Background Art

[0002] Against the backdrop of population growth, climate warming, and growing energy demand, reservoirs, as an important source of water supply, control the hydrological situation in the basin by formulating dispatching plans to achieve water resource utilization and allocation. Real-time reservoir dispatching is based on the meteorological and runoff forecasting process, taking into account dispatching objectives such as flood control, power generation, and ecology, to seek the optimal water storage or discharge strategy for the reservoir. The effectiveness of reservoir dispatching depends not only on the design of the dispatching plan, but also on the accuracy and robustness of the forecasting system. In particular, uncertain factors such as meteorological data, flood forecasts, and decision makers' preference information make the decision-making process in reservoir dispatching highly complex. These uncertainties not only directly interfere with the dispatching plan, but also pass through the "meteorological-forecast-dispatching-decision-making" chain, further amplifying the potential risks in the implementation process.

[0003] Reservoir operation is a complex task that involves optimization decisions in multiple aspects such as flood control, water supply, and power generation. However, the above objectives are often difficult to balance, which leads to considerable risks and uncertainties. In order to solve such complex problems, it is necessary to develop a multi-objective optimization model that can take into account different objectives at the same time. Existing reservoir operation methods fail to fully consider the impact of uncertainty on operation decisions. Usually, the operation model relies on a single meteorological forecast data or historical hydrological data, and fails to comprehensively consider the propagation effect of multi-source uncertainty factors. The systematic forecast error and uncertainty accumulation may lead to deviations between the operation results and the actual situation, thereby increasing the possibility of decision failure and reservoir group system risks.

[0004] In addition, in the process of reservoir operation decision-making, the multi-attribute decision-making (MADM) model also faces limitations. The traditional MADM method relies on a certain decision matrix and indicator weights, which makes it difficult to deal with the uncertainty of indicator values ​​in the decision matrix caused by forecast uncertainty in the actual operation process, as well as the uncertainty of weights caused by the ambiguity of decision makers' subjective preferences and the conflict of interests among multiple subjects. It is necessary to establish a random multi-attribute decision-making model that can deal with dual uncertainties.

[0005] As the uncertainty of the basin hydrological system continues to increase, scheduling decisions not only require the optimization model to provide reasonable solutions in the scheduling stage, but also to conduct risk assessment and dynamic adjustment after the implementation of the solution. The existing reservoir scheduling technology lacks a systematic spatiotemporal dimension deduction tool, and fails to conduct refined diagnosis and dynamic deduction of the risks of each hydrological characteristic node in the reservoir group. This deficiency limits the application of scheduling decisions under complex spatiotemporal coupling conditions, especially the interaction between upstream and downstream reservoir systems and the uncertainty of variables such as flood flow and rainfall intensity. The actual impact of scheduling schemes is difficult to be effectively quantified and evaluated. Existing technologies mostly rely on empirical evaluation methods and lack tools for comprehensive quantitative evaluation of uncertainty risks after scheduling. Therefore, it is very promising to develop a spatiotemporal two-dimensional deduction model for reservoir scheduling risks.

[0006] In view of the above problems, the present invention proposes a solution. Summary of the invention

[0007] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a reservoir group scheduling decision method considering multi-source uncertainty propagation and evolution tracking to solve the problems raised in the above-mentioned background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions: The decision-making method for reservoir group operation considering multi-source uncertainty propagation and evolution tracking includes the following steps: In the hydrological forecasting stage, a hydrological forecast uncertainty evolution model based on the time series fusion transformer (TFT) is constructed to simulate the dynamic characteristics of hydrological variables and generate the probability distribution of hydrological forecasts through deviation calibration; In the dispatching stage, a multi-objective distributed robust optimization model is constructed that comprehensively considers the uncertainty of meteorological variables, the uncertainty of runoff random scenarios, and the conflict of dispatching objectives, and the moss growth optimization algorithm is used to solve the non-inferior solution set; In the operation risk deduction stage, the spatiotemporal two-dimensional deduction of the flood control operation risk of the reservoir group is carried out based on the spatiotemporal topological relationship constructed by the directed acyclic graph, and the risk level of the reservoir operation plan is calculated; In the decision-making stage, the preference mediator is used to coordinate the preferences of the decision-making group, and the random decision matrix and weight samples are generated based on the hypercube simulation engine. The random average solution distance model is constructed through the weighted matrix fusion module, and the comprehensive evaluation value is calculated to select the uncertainty solution.

[0009] In a preferred embodiment, the uncertainty of the hydrological forecast is quantified by the following steps; Based on the TFT model, the meteorological forecast errors are simulated and corrected to generate the probability distribution of meteorological variables; A composite probability distribution model is used to describe runoff uncertainty, including a mixture of normal and skewed distributions: ; In the formula, is the standard normal distribution function, is the skewed distribution function, μ is the expected value of runoff, σ is the standard deviation of runoff, γ is the skewness coefficient, and α is the weight parameter; The conditional probability distribution of runoff output disturbed by meteorological errors is obtained using Monte Carlo simulation: ; In the formula, Given the error E and parameters The conditional probability of runoff outcome when Given the error E and parameters The posterior probability distribution when , is the parameter of meteorological variables.

[0010] In a preferred embodiment, the objective function of the multi-objective distributed robust optimization model includes: Minimize flood control storage capacity occupancy rate: ; In the formula, is the storage capacity of the i-th reservoir at time t under the l-th random scenario; is the storage capacity corresponding to the design flood limit water level of the i-th reservoir; is the maximum flood control storage capacity of the i-th reservoir; Maximize the safety of downstream control nodes: ; In the formula, is the outflow of the ith reservoir at time t under the lth random scenario; is the maximum outflow allowed by the i-th reservoir.

[0011] In a preferred embodiment, the optimization process of the moss growth optimization algorithm includes: The initial solution is generated by random perturbations of meteorological variables and the initial population of solutions is formed by combining hydrological uncertainties; Dynamically adjust the expansion path of the solution according to the wind direction mechanism: ; In the formula, represents the similarity of individual i, is the directional weight of wind direction; The dual reproduction mechanism of moss is simulated to expand or adjust the local search range along the optimal direction in the solution space.

[0012] In a preferred embodiment, the two-dimensional space-time deduction is achieved by the following steps: Construct a directed acyclic graph based on the spatial relationship between upstream and downstream, main and tributary rivers, and left and right banks to represent the spatiotemporal topology of reservoir operation; Use the graph attention network to learn the features of each node and dynamically adjust the information weights between nodes: ; ; In the formula, and They represent the linear transformation of the feature vectors of hydrological variable nodes i and j after the weight matrix W; a is a learnable parameter used to learn the interaction between the features of hydrological variable nodes; LeakyReLU is a nonlinear activation function used to introduce nonlinearity and prevent the gradient vanishing problem; is the set of neighbor nodes of hydrological variable node i; The spatiotemporal propagation path is deduced to generate a risk distribution map, which characterizes the evolution of scheduling risk at each node and time step.

[0013] In a preferred embodiment, the decision-making stage specifically includes the following steps: Extract the solution eigenvalues ​​from the non-inferior solution set and construct a random decision matrix; A preference mediator that comprehensively considers subjective weights and preference conflicts is used to obtain a high-dimensional weight vector, and a comprehensive evaluation value is calculated based on the weighted forward distance and reverse distance. The random average solution distance model ranks the priorities of various scheduling schemes in different decision-making scenarios through the comprehensive evaluation value.

[0014] In a preferred embodiment, the risk-based multi-attribute decision-making model simultaneously considers the following uncertainties: Uncertainty in the values ​​of decision indicators, including maximum water level, reservoir overtopping risk, maximum outflow, downstream flood risk, and final water level; The uncertainty of weights is addressed by using the fuzzy analytic hierarchy process combined with the ordered weighted average operator to coordinate group preference conflicts, and the randomness of weight distribution is quantified through the feasible weight space.

[0015] The technical effects and advantages of the reservoir group dispatching decision-making method considering multi-source uncertainty propagation and evolution tracking of the present invention are as follows: The present invention comprehensively and systematically considers the multi-source uncertainties in the meteorological, forecasting, dispatching and decision-making chains. By constructing a two-dimensional risk deduction model in time and space, the invention conducts refined risk diagnosis and dynamic deduction of the reservoir group dispatching process, obtains the uncertainty propagation and evolution tracking mechanism, and effectively improves the robustness and reliability of reservoir dispatching decisions in an uncertain environment. The random average solution distance method for multiple uncertainties created by the present invention provides a new method for quantitatively evaluating the pros and cons of scheduling schemes. In the acquisition of group preference information, the preference mediator FAHP-OWA that comprehensively considers subjective weights and preference conflicts is used to obtain high-dimensional weight vectors. At the same time, it combines multi-round interactive optimization and weight distribution mechanisms, and can gradually introduce decision makers' preference information, making the scheduling process more flexible and the decision group more coordinated, improving the decision-making efficiency and accuracy of the decision group in an uncertain environment, and realizing accurate judgment of the pros and cons of the scheme. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic flow chart of a reservoir group dispatching decision method considering multi-source uncertainty propagation and evolution tracking according to the present invention; Figure 2 This is a flow chart of the numerical solution algorithm of the MC-EDAS model of the present invention; Figure 3 A two-stage decision-making process for risk-type group decision-making in the present invention; Figure 4 A three-dimensional bar graph of the probability of scheme ranking under four different types of weights of the MC-EDAS model of the present invention; Figure 5 It is the central weight vector of each flood control scheduling scheme of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] The present invention effectively improves the robustness and reliability of reservoir scheduling decisions in an uncertain environment by integrating runoff random simulation, multi-objective random optimization scheduling, spatiotemporal two-dimensional deduction of reservoir scheduling risks, and risk-based multi-attribute decision-making modules.

[0019] Embodiment The present invention provides a multi-objective scheduling decision method for a reservoir group considering multi-source uncertainty propagation and evolution tracking, comprising the following steps: Figure 1 : S1: Collect the meteorological, engineering, watershed, underlying surface and other data required for modeling, perform data cleaning and preprocessing operations on the data set, and obtain the data set of the target watershed; S2: Build a meteorological statistical post-processing module and use the Time Series Fusion Transformer (TFT) to simulate and quantify the uncertainty of various meteorological variables such as precipitation; S3: Construct a universal basin rainfall-runoff model, take the meteorological forecast results after statistical post-processing as input, integrate the universal probability distribution line type, and generate a runoff probability forecast that can comprehensively consider normal and skewed (positive / negative skewness); S4: Establish a reservoir real-time dispatching distributed robust optimization model with minimum coupled flood control storage capacity occupancy and maximum downstream control node safety; S5: The moss growth optimization algorithm (MGO) is used to solve the reservoir real-time dispatch distributed robust optimization model with the minimum coupled flood control storage capacity occupancy rate and the highest safety of downstream control nodes, and a random non-inferior solution set is obtained; S6: Based on the obtained random non-inferior solution set samples, the cumulative rainfall forecast error, reservoir and interval flood peak flow forecast error, reservoir and interval flood volume forecast error, cross-section combined flood peak flow, and flood control maximum water level are used as network nodes. According to the spatial topological relationship between upstream and downstream, main and tributary rivers, and left and right banks, a directed acyclic graph neural network is drawn in accordance with the principle of "cause nodes in front and result nodes in the back", and the graph attention network (GAT) is used to perform spatiotemporal two-dimensional deduction of reservoir operation risks; S7: Based on the average solution distance method, a random average solution distance method suitable for multiple uncertainty environments was created. The random decision matrix and the high-dimensional weight vector gathered from the preferences of many decision makers were extracted and input into the random average solution distance model. The solutions on the random non-inferior solution set were quantitatively evaluated to obtain a scheduling plan that meets the decision maker's preferences.

[0020] Furthermore, the specific steps of step S1 are: S11: For each hydrological node, obtain and integrate relevant variable data, including key elements such as precipitation intensity field, aerodynamic temperature field and fluid motion velocity field. Extract physical constraint data such as reservoir capacity constraints and spillway design thresholds, and calculate and determine engineering parameters. Obtain the geographic spatial information of the basin by analyzing the complex topological structure of the river channel, extracting terrain elevation data, and determining the spatiotemporal distribution of hydrological observation stations. The underlying surface characteristics are quantified using indicators such as soil permeability coefficient and vegetation coverage. All data are collected, cleaned and standardized based on a multi-source monitoring technology platform, and integrated into a high-dimensional spatiotemporal data set to provide accurate basic support for subsequent modeling processes.

[0021] S12: The construction of the dataset needs to be based on the spatiotemporal distribution characteristics of multi-source data, perform time series rearrangement operations, and perform spatial cropping according to the range of each target hydrological node to ensure the time series consistency of the data.

[0022] S13: After the dataset is cleaned, it is stored in the data management system, which dynamically provides data access support through the API interface to ensure that the subsequent model modules can directly obtain the spatiotemporal data of the required target watershed and realize the transmission of cross-module data flow.

[0023] Furthermore, the specific formulas for the timing rearrangement and spatial cropping in step S12 are as follows: (1) Rearrange the data of hydrological variable node i in time sequence to ensure data consistency: ; In the formula, It is the time offset calculated according to the specific requirements of the hydrological node.

[0024] (2) Perform spatial cropping according to the range of the target hydrological node to obtain the cropped target watershed dataset: ; In the formula, It is a spatially clipped dataset. It represents the data value at a specific time t and spatial location i.

[0025] Furthermore, the specific formula for data set cleaning and storage in step S13 is as follows: (1) Clean the trimmed data set, remove outliers, and obtain the cleaned data set ; In the formula, is in the time window In the dataset for spatial position i The median is calculated without being affected by extreme values ​​(outliers).

[0026] (2) Store the cleaned data set in the data management system and provide data access support through the API interface: ; In the formula, parameters include time series, spatial range and other parameters.

[0027] Furthermore, the specific steps of step S2 are: S21: Construct a time series fusion transformer (TFT) to describe the spatiotemporal variation characteristics of complex hydrological variables, especially for key hydrological elements such as rainfall intensity, river flow and wind speed field, to provide data support for the reservoir runoff forecasting process.

[0028] S22: Perform error correction and bias calibration on meteorological forecast data, compare historical hydrological observation data with forecast data, generate meteorological forecast uncertainty assessment results based on the TFT model, and output the probability distribution of prediction errors, corrected meteorological data and confidence intervals.

[0029] S23: Conduct sensitivity analysis on different meteorological variables (such as rainfall, wind speed, etc.) to evaluate the impact of their errors on runoff and water level prediction results. Monte Carlo simulation is used to perturb the meteorological variable errors to observe how these errors propagate and affect the final runoff output. The specific formula is as follows: ; In the formula, For a given error E and parameter The conditional probability of runoff outcome when Given the error E and parameters The posterior probability distribution when , are the parameters of meteorological variables, such as precipitation, wind speed, etc., and E is the error term of meteorological forecast.

[0030] Furthermore, the initialization timing fusion converter (TFT) architecture used in step S21 includes the following modules: GRN, VSN, SCE, TFD, and the specific operations are as follows: (1) Perform GRN processing on time series data (X_t) to ensure the effective flow of meteorological and hydrological information in the multi-layer network: ; (2) A variable selection network (VSN) is used to assign weights and select the time series inputs of various hydrological variables. VSN is used to capture the most relevant features of the input variables and generate an adaptive variable selection weight matrix: ; Where Wv and bv are trainable weight and bias parameters respectively.

[0031] (3) The temporal encoding module SCE maps time information into a low-dimensional space representation through a periodic function: ; In the formula, t is the time step, T is the period length of the time series, and the time series encoding result et will be input into the model together with the input hydrological variables.

[0032] (4) The time series fusion decoder module TFD aggregates the time series features of meteorological variables through the attention mechanism to generate the final meteorological forecast data. The calculation formula is as follows: ; Where Q, K, V are query, key and value matrices respectively, and dk is the dimension of the key vector.

[0033] Furthermore, the specific steps of performing error correction and deviation calibration on the meteorological forecast data in step S22 are as follows: (1) Generate weather forecast error: ; In the formula, is the residual between the weather forecast data and the actual value at time t, is the actual observed value at time t, is the actual observed value at time t.

[0034] (2) The meteorological data forecast value is corrected by the residual, and the corrected forecast value is calculated as: ; In the formula, is the revised forecast value at time t+1, is the initial forecast value at time t+1 (uncorrected), is the expected value of the residual at time t.

[0035] (3) Calculate the confidence interval as: ; In the formula, is the revised forecast value at time t+1, is the preliminary forecast value at time t+1 (uncorrected), is the expected value of the residual at time t, is the standard deviation of the residual.

[0036] Furthermore, the specific steps of step S3 are: S31: Construct a dynamic characterization framework for a general rainfall-runoff model. The model covers the nonlinear dynamic response mechanisms of rainfall, surface runoff, soil infiltration, and groundwater recharge, and uses the multi-scale, multi-factor coupling mechanism under complex basin conditions to model hydrological variables as a whole. The model input is provided by the time series forecast of meteorological data, and multi-dimensional simulation is performed on key variables such as rainfall intensity and wind speed field. The formula framework can be expressed as: ; Where Q represents the total runoff volume per unit time in the basin, P represents the time series rainfall input, A represents the geographical unit area of ​​the basin, C represents the complex dynamic coefficient in the hydrological cycle, which depends on the underlying surface characteristics, vegetation coverage and soil saturation, and Φ represents the dynamic response function of the runoff coefficient, which is adjusted based on the basin characteristics.

[0037] S32: Integration and standardization of multi-dimensional time series input data,The meteorological data set includes multi-dimensional elements such as rainfall intensity, wind speed field and temperature change.,Standardization and data preprocessing strategies are used to perform high-level integration of the statistical post-processed meteorological data generated by the Time Series Fusion Transformer (TFT) to ensure the spatiotemporal consistency and continuity of the input data.

[0038] S33: A composite distribution model is used to simulate runoff uncertainty. The model characterizes the asymmetry and discreteness of runoff through a mixture of two probability distributions (normal distribution and skewed distribution). It is particularly suitable for runoff processes under nonlinear rainfall events. The formula is as follows: ; In the formula, is the standard normal distribution function, is the skewed distribution function, μ is the expected value of runoff, σ is the standard deviation of runoff, γ is the skewness coefficient, and α is the weight parameter that controls the contribution of normal and skewed distributions.

[0039] S34: Calibrate the model with high precision and use the Bayesian inference method to optimize the key parameters μ, σ, and γ in the composite distribution model to maximize the model fit and prediction accuracy. Dynamically adjust the model parameters by inverting and extrapolating historical runoff observation data. The Bayesian inference formula is as follows: ; In the formula, is the model parameter under given observation data D The posterior distribution of is the likelihood function, is the prior distribution, representing the initial estimate of the parameter.

[0040] S35: Generate high-confidence runoff probability forecast results. Through the calibrated model, the runoff probability distribution of multiple time steps is output, and the confidence interval is combined to characterize the uncertainty of runoff. The forecast results can quantitatively reflect the distribution of runoff under future rainfall scenarios, especially the basin response under extreme meteorological conditions. ; In the formula, represents the runoff forecast at time t+1, is the prediction interval, and α is the confidence level, which represents the uncertainty.

[0041] Furthermore, the specific steps of step S4 are: S41: Multidimensional modeling of objective functions The setting of objective functions is centered on minimizing the flood control reservoir capacity occupancy rate and maximizing the safety of downstream nodes, ensuring that the dispatching strategy takes both flood control and safety into account. The model transforms reservoir capacity restrictions, outflow restrictions, and downstream node safety into an optimizable objective function, which is specifically expressed as: The first objective function is to minimize the flood control storage capacity occupancy rate of each reservoir under different random scenarios. The objective formula is as follows: ; In the formula, is the storage capacity of the i-th reservoir at time t under the l-th random scenario; is the storage capacity corresponding to the design flood limit water level of the i-th reservoir; is the maximum flood control capacity of the ith reservoir.

[0042] The second objective function is to maximize the safety of downstream control nodes. The objective formula is as follows: ; In the formula, is the outflow of the ith reservoir at time t under the lth random scenario; is the maximum outflow allowed by the i-th reservoir.

[0043] By introducing the weight parameter λ, the multi-objective optimization is transformed into a unified objective function: ; S42: Multidimensional extension of constraints The constraints of the model include water balance constraints of the reservoir, outflow constraints, and safety flow constraints of downstream control nodes. All constraints rely on dynamically monitored meteorological and hydrological data to ensure that real-time scheduling operates within physical boundaries: (1) Storage capacity balance constraint: ; In the formula, and are the storage capacities of the ith reservoir at time t+1 and time t under the lth random scenario respectively. The other variables are similarly applicable to the lth random scenario.

[0044] (2) Water level constraint: ; (3) Outbound flow constraints: ; (4) Constraints on the fluctuation range of outbound flow: ; (5) Maximum discharge capacity constraints: ; (6) Initial and boundary conditions: ; In the formula, is the initial water level of the reservoir under the lth random scenario; is the target water level of the reservoir at the end of the period under the lth random scenario.

[0045] S43: The model processes multi-source random disturbances (such as weather forecast errors, rainfall intensity deviations, etc.) to form a robust scheduling strategy. The model uses distributed robust optimization to handle uncertainty and quantify the potential impact of uncertainty on reservoir group scheduling, thereby obtaining the optimal solution under multi-source uncertainty. Assuming that the disturbances come from an uncertainty set and simulating the impact of these disturbances through a distribution set, the key formula in the optimization process is as follows: ; In the formula, u is the scheduling decision variable, representing the reservoir inflow and outflow decisions; p is the disturbance parameter, representing meteorological and hydrological uncertainties; It is the uncertainty set, which defines the possible distribution range of meteorological variables (such as precipitation, wind speed, etc.).

[0046] S44: Construct a multi-time scale adaptive scheduling mechanism and dynamically update the scheduling strategy according to different time steps. In the short term, based on the timeliness of hydrological forecasts, in the medium and long term, combine multi-source data to achieve robust real-time adjustment of the system. Dynamic adjustment for different time steps: In short time steps (such as hourly scales), the model is frequently updated according to real-time hydrological forecasts; in long time steps (such as daily or weekly scales), the model is combined with meteorological forecast trends and historical data for global adjustments. The specific formula is as follows: (1) Short-term step size adjustment ; (2) Long-term step size adjustment ; Where N is the length of the scheduling time step (e.g., days or hours); is the expected value of the uncertainty parameter.

[0047] Furthermore, the specific steps of step S5 are: S51: Incorporating the uncertainty factors of meteorology and flood forecasting, the initial solution of the moss growth algorithm is generated through random perturbations and heuristic strategies: ; In the formula, lb and ub are the upper and lower bounds of the variable, respectively, and rand(0,1) generates a random number between 0 and 1. After the initial solution is generated, the system applies it to the multi-objective flood control scheduling model and performs random initialization based on meteorological uncertainties such as flood forecasts and rainfall intensity, thereby ensuring that the generated solution can cope with multi-source uncertainties.

[0048] S52: Determine the wind direction and expand the growth direction of the solution. In the moss growth model, wind direction is one of the key factors for individuals to be affected by external factors (such as meteorological disturbances and flood forecasts). The introduction of the wind direction mechanism enables most solutions to develop in the direction of the optimal solution, thereby reducing the model search space. The wind direction calculation formula in this process is: ; In the formula, represents the similarity of individual i, is the directional weight of wind direction. Wind direction enables the solution to flexibly adjust its evolution path according to current meteorological and hydrological information.

[0049] S53: Simulate the growth and expansion process of moss, propagate some solutions as "seeds" along the wind direction, and other solutions readjust the path according to environmental disturbances. The propagation path of seeds can be described by the following formula: ; S54: Implement a dual reproduction mechanism to simulate the growth strategies of seed reproduction and moss reproduction. The reproduction formula is: ; ; S55: When mosses are ineffective for a long time or in a local optimum, the cryptobiotic strategy is used to simulate the mosses entering a metabolic dormant state, and they will recover and continue searching when conditions are right. The cryptobiotic mechanism ensures that the algorithm can dynamically restore its search capability under extreme meteorological and hydrological conditions, thereby enhancing the robustness of the global search.

[0050] Furthermore, the specific steps of step S6 are: S61: Processing random non-inferior solution set samples: These samples contain multi-objective optimization results, covering different reservoir operation schemes. Each sample constructs its probability distribution function according to different meteorological variables (such as cumulative rainfall forecast error, peak flow forecast error, etc.). The formula is as follows: ; In the formula, represents the uncertainty probability distribution of the i-th reservoir at time t; Eprecip, Eflow, and Estorage represent the prediction errors of cumulative rainfall, peak flow, and reservoir capacity, respectively. Each reservoir is set as a network node to construct the core node of spatiotemporal risk propagation, and the interactive relationship between meteorological variables and reservoir operation is used to describe the dynamic evolution of the entire system.

[0051] S62: Based on the spatial topological relationship of the reservoir and the basin, such as upstream and downstream, main and tributary rivers, and left and right banks, a directed acyclic graph (DAG) is constructed. This structure represents the physical connection between the reservoir and its dispatching process. The edges between nodes represent the water flow transmission path and time delay, ensuring that the cause-effect relationship is clear, that is, the principle of "cause node first, result node later".

[0052] S63: Graph Attention Network (GAT) is used for two-dimensional spatiotemporal deduction. GAT deduces the changes in the state of each node at different time steps. By superimposing time steps, the propagation and evolution process of reservoir operation risk in time and space is demonstrated. Specifically, it includes: (1) Learn the weight relationship between nodes, identify the nodes that have the greatest impact on the decision-making risk of the reservoir system (such as the flow forecast error of the upstream reservoir), and dynamically adjust the weight according to the size of the influence.

[0053] ; ; In the formula, and They represent the linear transformation of the feature vectors of hydrological variable nodes i and j after the weight matrix W; a is a learnable parameter used to learn the interaction between the features of hydrological variable nodes; LeakyReLU is a nonlinear activation function used to introduce nonlinearity and prevent the gradient vanishing problem. is the set of neighbor nodes of hydrological variable node i.

[0054] (2) Dynamically adjust the information transmission weights between nodes. Based on the results of the dynamic weight adjustment, determine the optimal reservoir operation path under specific spatiotemporal conditions, ensure that the flood control storage capacity can be fully utilized, and maximize the protection of downstream area safety, and accurately capture the spatiotemporal evolution of reservoir operation risks; ; In the formula, is the updated feature of each reservoir node i, is the activation function.

[0055] (3) Generate a reservoir operation risk distribution map, which contains the risk level of each reservoir node and its adjacent nodes at different time steps, identify key nodes and high-risk areas, and then predict and infer the risk of any node in the network to improve the accuracy of reservoir operation decision-making.

[0056] ; In the formula, Represents the risk level of hydrological variable node i at time step t.

[0057] Furthermore, the specific steps of step S7 are: S71: Extract characteristic information of each variable from the set of random non-inferior solutions, use statistical analysis to obtain the distribution of each indicator value, and then construct a decision matrix under uncertainty.

[0058] S72: Aggregate the preferences of reservoir dispatch experts and use the preference mediator FAHP-OWA that comprehensively considers subjective weights and preference conflicts to obtain high-dimensional weight vectors. The preference mediator consists of three components: basic preference expresser, information aggregator, and feature extractor. The basic preference expresser is used to preliminarily obtain the high-dimensional weight information of each dispatch expert. This component supports information expression in the form of triangular fuzzy numbers; the information aggregator uses the OWA operator to fuse and aggregate the high-dimensional weight information; the feature extractor fits the distribution characteristics of the high-dimensional indicator weight set and performs spatial expansion.

[0059] Furthermore, in step S72, the information aggregator aggregates the high-dimensional weight vector into a coordination weight vector, and the output of the basic preference expresser is a high-dimensional weight information vector , k=1,2, ,L, L is the number of scheduling experts participating in group decision making, The information aggregator weight fusion formula coupled with the OWA operator is calculated by triangular fuzzy numbers and is as follows: ; ; ; The coordination weight vector that combines all experts' preferences is as follows: ; Furthermore, in step S72, since the decision needs to consider as much preference information as possible, the feature extractor performs spatial expansion on the coordination weight vector according to a certain probability distribution. The specific formula is as follows: ; In the formula, is the preference conflict level. When the decision-making group has a small preference conflict and is more satisfied with the calculated coordination weight vector, A smaller value can be taken, otherwise A larger value is taken to make the weight space cover a larger range.

[0060] S73: The random average solution distance model can handle high-dimensional weight vectors and random decision matrices. The model prioritizes uncertain solutions in a set of random non-inferior solutions.

[0061] Furthermore, the random version of the average solution distance model in step S73 is composed of the following key components: Figure 2 As shown: (1) Hypercube simulation engine: Latin hypercube sampling technology is used to generate sample sets of decision matrices and high-dimensional weight vectors. The value range of each random variable is divided into several equally spaced intervals. A sample point is randomly selected in each interval to ensure that the samples of each variable are evenly distributed.

[0062] (2) Weighted matrix fusion module: Combines the extracted decision matrix samples with the high-dimensional weight vector, assigns corresponding indicator weights to each sample, and ensures that the combination of all variables and weights is representative.

[0063] (3) Run the simulation component: perform model operations on each weighted decision matrix sample generated, evaluate the distance between each sample and the average solution, and simulate different decision scenarios. This step can be calculated using the EDAS model.

[0064] (4) Probabilistic decision-making component: collects the output results of all samples and extracts distribution features, including calculating the mean, variance, confidence interval, etc., and outputs statistical indicators such as the probability of sorting solutions and weighted comprehensive evaluation indicators.

[0065] Furthermore, the main calculation steps of the average solution distance (EDAS) model in step S73 are as follows: (1) Select evaluation indicators for reservoir operation schemes.

[0066] (2) The n-dimensional index values ​​of the m options constitute the decision matrix X.

[0067] ; Where Xij represents the j-th index value of the i-th solution.

[0068] (3) Calculate the average solution. ; (4) According to the different indicator types, calculate the distance between each alternative plan and the average plan, and determine the forward distance matrix and the inverse distance matrix .in and The calculation formula is: ; ; (5) Determine the weighted sum of the forward distances of each solution and inverse distance weighted sum .

[0069] (6) Weighted sum of the forward distances of each solution and inverse distance weighted sum Perform standardization.

[0070] (7) Calculate the comprehensive evaluation value ASi of each scheme and rank the schemes accordingly. The larger the comprehensive evaluation value, the better the scheme. ; Furthermore, in step S73, three statistical indicators, namely, the scheme ranking probability, the weighted comprehensive evaluation index, and the center weight vector, are defined, as shown below: (1) Probability of sorting options ( ) represents the proportion of the number of times the reservoir operation scheme Ai obtains the ranking priority r in the Latin hypercube sampling, and the calculation formula is: ; In the formula, It means that the scheme Ai obtains the ranking r ( ) corresponds to the weight space.

[0071] (2) Weighted comprehensive evaluation index ( ) represents the probability that the reservoir dispatching scheme Ai obtains all rankings, and the calculation formula is: ; In the formula, is the secondary weight, reflecting all the The degree of contribution to the weighted comprehensive evaluation index. It is a comprehensive evaluation of plan Ai.

[0072] (3) Center weight vector ( ) is defined as the center of the weight expansion space when the solution Ai obtains the optimal ranking: ; In the formula, is the weight when the sorting is optimal. Characterize the corresponding relationship between different weight combinations and optimal reservoir operation decisions.

[0073] The following takes the reservoir system in the Dadu River Basin as an example, including the Pubugou Reservoir and the downstream protection target Leshan City, to illustrate the effectiveness and rationality of the method.

[0074] The Pubugou Reservoir plays an important role in flood control, power generation, and water supply. The normal water level of the reservoir is 850m, and the total storage capacity is 5.39 billion cubic meters, of which the flood control storage capacity is 1.056 billion cubic meters. Leshan City is an important flood control control point downstream. In this numerical experiment, a general basin rainfall-runoff model was used to randomly generate 40 runoff scenarios. The generated runoff scenario set was used as the input of the reservoir real-time scheduling distributed robust optimization model. The moss growth optimization algorithm (MGO) was used to solve the model and obtain a set of random non-inferior solutions under uncertainty. The resulting reservoir multi-objective optimization scheduling random decision matrix consists of 30 random variables and 20 constants, as shown in Table 1.

[0075] Table 1 Random decision matrix for real-time multi-objective optimal operation of reservoirs

[0076] In the actual scheduling operation period, the decision is divided into two stages. The specific steps are as follows: Figure 3 As shown in Figure 2, in the first stage, decision makers are not required to express any preferences, and the MC-EDAS model is run when the indicator weights are completely unknown. Figure 4 The three-dimensional bar chart of the probability of ranking the schemes when the weight information is completely unknown is shown. It can be seen from the figure that schemes A1, A2, A9, and A10 have a higher probability of being ranked at the bottom. Among them, A1 and A2 have a very high probability of obtaining a poor ranking (the probability of ranking in the top five is 0), and the sum of the probabilities of A9~A10 ranking in the top five is less than 10%. The above results show that these four schemes can be judged as obviously inferior schemes and eliminated from the scheme set. Schemes A3~A5 all have a high probability of being ranked at the top or in the middle, and can be identified as compromise schemes. In addition, schemes A6~A8 all have a certain probability of obtaining the best ranking, among which scheme A7 is 47.0%, scheme A6 is 24.6%, and scheme A8 is 15.3%. These three schemes can be identified as obviously superior schemes and should be focused on in the second stage.

[0077] Figure 5 The central weight vectors of the flood control scheduling schemes when they are optimally ranked are shown. It can be seen from the figure that the central weight vector of the A3 scheme output by the MC-EDAS model gives smaller weights to Qmax and PQ, which are 0.1774 and 0.0499 respectively; and larger weights to Zmax, PZ and Ze, which are 0.3317, 0.246 and 0.195 respectively.

[0078] In the second stage of the decision-making process, based on the preliminary optimization results of the first stage and the central weight vector, four reservoir operation experts were invited to express their preference information in the form of triangular fuzzy numbers. The fuzzy judgment matrix given by the expert group is shown in Table 2.

[0079] Table 2 Fuzzy judgment matrix given by four reservoir operation experts

[0080] The FAHP-OWA coordination weighting method is used to weight each indicator and obtain the coordination weight. As shown in Table 3, the fuzzy judgment matrix of the expert group is first converted into a quantitative weight value according to the FAHP method. Secondly, the OWA operator is used to perform an orderly weighted aggregation of the conflicts of the preferences of the four experts to obtain the coordination weight. Based on the coordination weight, the feasible weight space is further used to quantify the uncertainty of the weight information, as shown in Table 4.

[0081] Table 3 Weight quantification based on fuzzy analytic hierarchy process and OWA operator weighting results

[0082] Table 4 Three different forms of weight information

[0083] The MC-EDAS model was rerun again with the random decision matrix and the three types of weights as input. Figure 4 (b), (c), (d) and Figure 4 (a) Compared with the control group, the output of the MC-EDAS model becomes clearer after adding the preference information of the decision-making group, and can better distinguish between obviously superior and inferior options.

[0084] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0085] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0086] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application of the technical solution and the invention constraints. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0087] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0088] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0089] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A reservoir group scheduling decision-making method considering multi-source uncertainty propagation and evolution tracking, characterized in that: The steps include: In the hydrological forecasting stage, a hydrological forecast uncertainty evolution model based on the time series fusion transformer (TFT) is constructed to simulate the dynamic characteristics of hydrological variables and generate the probability distribution of hydrological forecasts through deviation calibration; In the dispatching stage, a multi-objective distributed robust optimization model is constructed that comprehensively considers the uncertainty of meteorological variables, the uncertainty of runoff random scenarios, and the conflict of dispatching objectives, and the moss growth optimization algorithm is used to solve the non-inferior solution set; In the operation risk deduction stage, the spatiotemporal two-dimensional deduction of the flood control operation risk of the reservoir group is carried out based on the spatiotemporal topological relationship constructed by the directed acyclic graph, and the risk level of the reservoir operation plan is calculated; In the decision-making stage, the preference mediator is used to coordinate the preferences of the decision-making group, and the random decision matrix and weight samples are generated based on the hypercube simulation engine. The random average solution distance model is constructed through the weighted matrix fusion module. The comprehensive evaluation value is calculated and the priority of each scheduling scheme in different decision-making scenarios is ranked.

2. The reservoir group scheduling decision-making method considering multi-source uncertainty propagation and evolution tracking according to claim 1 is characterized in that: The uncertainty of hydrological forecasts is quantified through the following steps; Based on the TFT model, the meteorological forecast errors are simulated and corrected to generate the probability distribution of meteorological variables; A composite probability distribution model is used to describe runoff uncertainty, including a mixture of normal and skewed distributions: ; In the formula, is the standard normal distribution function, is the skewed distribution function, μ is the expected value of runoff, σ is the standard deviation of runoff, γ is the skewness coefficient, and α is the weight parameter; The conditional probability distribution of runoff output disturbed by meteorological errors is obtained using Monte Carlo simulation: ； In the formula, For a given error E and parameter The conditional probability of runoff outcome when Given the error E and parameters The posterior probability distribution when , is the parameter of meteorological variables.

3. The reservoir group scheduling decision-making method considering multi-source uncertainty propagation and evolution tracking according to claim 1 is characterized in that: The objective function of the multi-objective distributed robust optimization model includes: Minimize flood control storage capacity occupancy rate: ; In the formula, is the storage capacity of the i-th reservoir at time t under the l-th random scenario; is the storage capacity corresponding to the design flood limit water level of the i-th reservoir; is the maximum flood control storage capacity of the i-th reservoir; Maximize the safety of downstream control nodes: ; In the formula, is the outflow of the ith reservoir at time t under the lth random scenario; is the maximum outflow allowed by the i-th reservoir.

4. The reservoir group scheduling decision-making method considering multi-source uncertainty propagation and evolution tracking according to claim 1 is characterized in that: The optimization process of the moss growth optimization algorithm includes: The initial solution is generated by random perturbations of meteorological variables and the initial population of solutions is formed by combining hydrological uncertainties; Dynamically adjust the expansion path of the solution according to the wind direction mechanism: ; In the formula, represents the similarity of individual i, is the directional weight of wind direction; The dual reproduction mechanism of moss is simulated to expand or adjust the local search range along the optimal direction in the solution space.

5. The reservoir group scheduling decision-making method considering multi-source uncertainty propagation and evolution tracking according to claim 1 is characterized in that: The two-dimensional space-time deduction is achieved by the following steps: Construct a directed acyclic graph based on the spatial relationship between upstream and downstream, main and tributary rivers, and left and right banks to represent the spatiotemporal topology of reservoir operation; Use the graph attention network to learn the features of each node and dynamically adjust the information weights between nodes: ; ; In the formula, and They represent the linear transformation of the feature vectors of hydrological variable nodes i and j after the weight matrix W; a is a learnable parameter used to learn the interaction between the features of hydrological variable nodes; LeakyReLU is a nonlinear activation function used to introduce nonlinearity and prevent the gradient vanishing problem; is the set of neighbor nodes of hydrological variable node i; The spatiotemporal propagation path is deduced to generate a risk distribution map, which characterizes the evolution of scheduling risk at each node and time step.

6. The reservoir group dispatching decision-making method considering multi-source uncertainty propagation and evolution tracking according to claim 5 is characterized in that: The decision-making stage includes the following steps: Extract the solution eigenvalues ​​from the non-inferior solution set and construct a random decision matrix; A preference mediator that comprehensively considers subjective weights and preference conflicts is used to obtain a high-dimensional weight vector, and a comprehensive evaluation value is calculated based on the weighted forward distance and reverse distance. The random average solution distance model ranks the priorities of various scheduling schemes in different decision-making scenarios through comprehensive evaluation values.

7. The reservoir group dispatching decision-making method considering multi-source uncertainty propagation and evolution tracking according to claim 6 is characterized in that: The risk-based multi-attribute decision-making model considers the following uncertainties simultaneously: Uncertainty in the values ​​of decision indicators, including maximum water level, reservoir overtopping risk, maximum outflow, downstream flood risk, and final water level; The uncertainty of weights is addressed by using the fuzzy analytic hierarchy process combined with the ordered weighted average operator to coordinate group preference conflicts, and the randomness of weight distribution is quantified through the feasible weight space.

Images