Extraction method of flood control scheduling rules for hierarchical reservoirs based on improved XGBoost-LSTM

Abstract

The present invention discloses a hierarchical reservoir flood control scheduling rule extraction method based on an improved XGBoost-LSTM, comprising: collecting hydrological data and engineering data; constructing a multi-objective optimization scheduling model, and using a multi-objective optimization method that couples time-correlated weight coefficients, dual-population collaboration, and information interaction mechanisms to obtain a solution set for the multi-objective optimization problem; screening the solution set, and based on the reservoir optimized outflows in the screened solution set, using a K-means clustering algorithm coupled with an adaptive sliding time window to classify the reservoir optimized outflows; based on the classification results, using the improved XGBoost-LSTM hierarchical reservoir flood control scheduling rule extraction method to extract reservoir flood control scheduling rules, obtain reservoir rule outflows, and verify the reservoir rule outflows using evaluation indicators. The present invention can quickly provide an effective scheduling reference plan for reservoir flood control scheduling.

Description

Technical Field

[0001] The invention of the present application relates to the technical field of reservoir flood control scheduling rule extraction, and in particular to a hierarchical reservoir flood control scheduling rule extraction method based on an improved XGBoost-LSTM. Background Art

[0002] As the scale of reservoirs and other projects continues to expand, the complexity of real-time flood control joint scheduling models for reservoir systems is increasing. Traditional joint scheduling of river basin systems, from modeling to solution, is facing increasing challenges. Even with the rapid advancement of computational technology, the contradiction between the difficulty of solving high-dimensional, nonlinear mathematical models for real-time flood control scheduling of reservoir clusters and the high timeliness required for flood control decision-making remains acute. In particular, joint scheduling of flood control systems must balance the safety of various projects and protected objects within the system. Ultimately, it often faces the decision-making challenge of selecting a scheduling solution from a multitude of mutually non-inferior options, constituting a large-scale, multi-objective, nonlinear optimization problem. Modeling studies on similar problems have mostly employed the construction and solution of multi-objective optimization models. However, multi-objective optimization suffers from low computational efficiency, a large number of options, and difficulty in discerning flood and waterlogging coordination relationships, limiting its practical application in river basin flood control scheduling.

[0003] In the actual operation of reservoirs, the implementation of reservoir scheduling is usually based on the established reservoir scheduling rules. Reservoir scheduling rules are based on the design tasks and operation constraints. Reservoir scheduling operations are carried out based on a large number of sample series, and then the optimized operation rules of the reservoir are summarized. In recent years, computer technologies such as machine learning and data mining have gradually emerged. XGBoost and LSTM models have achieved many research results in the extraction of nonlinear, multi-dimensional, and large-scale system scheduling rules. However, they are still in their infancy and still have the following shortcomings: (1) XGBoost cannot directly establish long-term dependencies of time series. Reservoir outflows are related in time and space. If XGBoost is used to extract scheduling rules for the entire scheduling period, it may affect the accuracy of simulation predictions; (2) The black box characteristics of artificial intelligence models such as LSTM fail to directly consider the complex physical mechanisms involved in the actual scheduling process, have low interpretability, high computational complexity, and long training time, which reduces the practicality of the extracted scheduling rules in actual operation. If we can improve the XGBoost and LSTM models based on the characteristics of reservoir outflow and organically couple them, we can build an intelligent scheduling rule extraction model that takes into account both computational efficiency and simulation accuracy, and quickly provide effective and physically interpretable scheduling reference plans for reservoir flood control scheduling decisions. Summary of the Invention

[0004] The purpose of the invention is to provide a hierarchical reservoir flood control scheduling rule extraction method based on improved XGBoost-LSTM to solve the above-mentioned problems existing in the prior art.

[0005] The technical solution provides a method for extracting flood control scheduling rules for hierarchical reservoirs based on an improved XGBoost-LSTM, which includes the following steps:

[0006] Collect hydrological data and engineering data as input, use multi-objective optimization method to solve the reservoir flood control operation model and obtain the solution set;

[0007] The solution set is screened, and based on the reservoir optimized outflow obtained from the screened solution set, a clustering algorithm is used to divide it into low-flow process and high-flow process;

[0008] The GRU-XGBoost model is used to extract scheduling rules for low-flow processes and obtain regular outflows for low-flow processes. The LSTM model coupled with physical constraints is used to extract scheduling rules for high-flow processes and obtain regular outflows for high-flow processes. The regular outflows of high and low-flow processes are combined into regular outflows of reservoirs.

[0009] Preferably, building a model and obtaining a solution set for a multi-objective optimization problem includes:

[0010] Construct a multi-objective optimization scheduling model, set the objective function as minimizing the maximum flood control storage capacity utilization rate of the reservoir, minimizing the maximum average excess flood volume of the system, minimizing the maximum flow at the flood control point, and set constraints;

[0011] Based on the objective function, a time-related weight coefficient is introduced, the priority of the objective function is adjusted according to the reservoir flood control scheduling stage, and the objective function is normalized to obtain the objective function based on the dynamic multi-objective decomposition mechanism; and the global exploration-local development dual population collaboration and information interaction mechanism is adopted to solve the problem and obtain the solution set of the multi-objective optimization problem.

[0012] Preferably, constructing an objective function based on a dynamic multi-objective decomposition mechanism includes:

[0013] Set the time-related weight coefficients w1, w2, and w3, including:

[0014] The weight coefficient of reservoir flood control capacity utilization rate is w1(t) = -sin(πT / t)+1, t∈[0,T];

[0015] System average excess flood weight coefficient w2(t) = sin(πT / t), t∈[0,T];

[0016] The flood control point flow weight coefficient w3(t) = 0.5, t∈[0,T]; where t represents the current time and T represents the total scheduling duration.

[0017] The time-related weight coefficients w1, w2, and w3 are multiplied by the corresponding objective function, and the maximum and minimum values of the objective function are normalized using the range method to obtain the objective function based on the dynamic multi-objective decomposition mechanism.

[0018] Preferably, a global exploration-local development dual population collaboration and information interaction mechanism is used to solve multi-objective optimization problems, including:

[0019] Construct a global exploration population, determine the number of splits H and the number of objective functions M, use the Das-Dennis method to generate reference points N and normalize them, where N = (H + M-1) + (M-1), and dynamically adjust the reference point distribution according to the current solution set distribution density;

[0020] Construct a local development population and design a mutation factor U = 0.5 × (1 + rand), U∈[0.5,1]. Use the DE / current-to-pbest / 1 strategy to generate the mutation vector Vi = Xi + U(Xpbest-Xi) + U(Xr1-Xr2); where Xi is the i-th individual in the current population, Xpbest is a randomly selected individual from the top p% of elite individuals in the population, and Xr1 and Xr2 are two randomly selected individuals.

[0021] Select the top a% of solutions from the global exploration population as candidate solutions A, and select the top a% of solutions from the local development population as candidate solutions B to form a candidate pool;

[0022] Calculate the target space cosine similarity between the candidate solutions in the candidate pool and the target population individuals, and select the individual with the lowest similarity for replacement based on the information interaction mechanism;

[0023] When the solution sets of the two populations change less than a threshold for consecutive K generations, convergence is determined and the solution set of the multi-objective optimization problem is obtained.

[0024] Preferably, the solution set is screened to obtain the optimized outflow of the reservoir and cluster it, including:

[0025] The solution set is divided into 7 categories, including the minimum point of objective 1, the minimum point of objective 2, the minimum point of objective 3, the three-objective equilibrium solution, the solution that is beneficial to objective 1, the solution that is beneficial to objective 2, and the solution that is beneficial to objective 3;

[0026] A hierarchical clustering method based on Euclidean distance is used to select a representative solution set for the solutions that are beneficial to each goal, and a solution with the shortest Euclidean distance to other solutions is selected as the representative solution in each cluster;

[0027] Calculate the average value of the objective function of each solution in the equilibrium solution set, sort the average values, and select the solutions whose frequency difference from 50% is less than 0.01 as the representative solution set of the equilibrium solution;

[0028] Integrate the representative solution sets of all categories to obtain the filtered solution set;

[0029] The objective function value, reservoir storage and optimized outflow of the reservoir are extracted from the filtered solution set. Based on the optimized outflow of the reservoir, the K-means clustering algorithm coupled with an adaptive sliding time window is used to divide the optimized outflow of the reservoir into low-flow process and high-flow process, which are used as sample data for subsequent models.

[0030] Preferably, a K-means clustering algorithm coupled with an adaptive sliding time window is used, including:

[0031] The optimized outflow sequence of the reservoir is plotted as the optimized outflow process curve Q(t), and the scheduling period of the reservoir is set as T;

[0032] Calculate the standard deviation σ of the slope S(t) of the reservoir's optimized outflow process curve at each moment to reflect the magnitude of local fluctuations;

[0033] The local average slope is calculated using the continuous difference method. A dynamic sliding time window W is set based on the comparison between |S(t)| and σ. When |S(t)|<σ, W∈[t-α·T, t+α·T]; when |S(t)|>σ, W∈[t-β·T, t+β·T], where β<α<1. The local average slope S within the window is calculated.

[0034] Based on the comparison between the slope standard deviation σ within the window and the preset threshold ξ, as well as the value of the average slope S, the reservoir optimized outflow is divided into a low-flow process candidate set and a high-flow process candidate set;

[0035] The mean of the low-flow process candidate set and the high-flow process candidate set are used as the initial center point, and the K-means clustering algorithm is used to cluster the low-flow process and the high-flow process are obtained.

[0036] Preferably, the GRU-XGBoost model is used to extract scheduling rules for low-flow processes, including:

[0037] The GRU model is introduced and trained based on the optimized outflow of the reservoir to obtain the time series characteristics of the low flow process;

[0038] The time series characteristics of low-flow processes are added to the model samples to build a GRU-XGBoost prediction model. The second-order Taylor expansion is used to optimize the objective function, and the complexity is controlled by the regularization term.

[0039] The samples were divided into training and validation sets. The GRU-XGBoost model was trained using the measured reservoir inflow, interval water inflow, reservoir storage, reservoir remaining forecast inflow, objective function value, and low-flow process time series characteristics as independent variable factors, and the optimized reservoir outflow and reservoir remaining forecast outflow as dependent variable factors.

[0040] The independent variable factors of the validation set are input into the GRU-XGBoost model to obtain the regular outflow for predicting low flow process and calculate the fitting error;

[0041] The GRU model is trained twice based on the fitting error to obtain the predicted fitting error. The predicted regular outflow during low flow process is added to the predicted fitting error to obtain the regular outflow during low flow process.

[0042] Preferably, it includes:

[0043] Based on the independent variable factors of the GRU-XGBoost model, the gradient boosting algorithm is used to construct a decision tree, the quantile discretization method is used to generate a set of candidate split thresholds, and the multi-threading technology is used to parallelly calculate the split gains of each variable factor;

[0044] Traverse all independent variable factors and their split gains of the current node, use the greedy algorithm to select the point with the largest split gain as the split condition, and repeat the above process for the left and right child nodes until the maximum tree depth is reached or the gain is lower than the set threshold;

[0045] Set the learning rate parameter, which is multiplied by the parameter when the prediction value of each newly added tree is updated. By reducing the contribution of a single tree, the model is made more dependent on the weighted results of multiple trees. At the same time, regularization terms are combined to prevent overfitting.

[0046] Integrate the prediction results of each tree to obtain the nonlinear mapping relationship between the independent variable factors and the dependent variable factors.

[0047] Preferably, the specific steps of extracting scheduling rules for high-traffic processes include:

[0048] Based on the model samples and the water level-discharge capacity curve, a loss function was constructed that included a root mean square error penalty, a water balance constraint penalty, a high flow weighted penalty, a discharge capacity constraint penalty, and a non-negative constraint penalty. An LSTM model coupled with physical constraints was then built.

[0049] The model samples were divided into a training period and a validation period. The measured inflow, interval water inflow, reservoir storage, remaining predicted inflow, and objective function value were used as independent variable factors, while the optimized outflow and remaining predicted outflow were used as dependent variable factors to train the LSTM model coupled with physical constraints.

[0050] The independent variable factors of the validation period are input into the LSTM model coupled with physical constraints to obtain the regular outflow of high-flow processes.

[0051] Preferably, building an LSTM model coupled with physical constraints includes:

[0052] Based on the reservoir storage, water level-storage capacity curve and water level-discharge curve data in the model sample, the reservoir discharge capacity is obtained;

[0053] Set parameters y1, y2, y3, y4, and y5, where parameter y1 is the predicted regular outflow value of the reservoir, parameter y2 is the reservoir discharge capacity, parameter y3 is the maximum value of parameter y1, parameter y4 is the maximum value of the optimized outflow of the reservoir, and parameter y5 is the predicted remaining forecast outflow value of the reservoir;

[0054] Based on the measured inflow of the reservoir, parameter y1, the remaining forecast inflow of the reservoir, parameter y5, the storage capacity at the end of the period and the storage capacity of the reservoir, the water balance constraint penalty term is calculated according to the water balance principle;

[0055] Based on the optimized outflow of the reservoir, parameters y3 and y4, the weight coefficient that changes according to the size of the reservoir outflow is calculated to obtain the high flow weighted penalty term;

[0056] Compare the parameters y1 and y2. If y1 is greater than y2, calculate the discharge capacity constraint penalty term. Otherwise, the penalty term is 0.

[0057] Compare the size of the parameter y1 with 0 and calculate the non-negative constraint penalty term;

[0058] The LSTM model loss function is added to the above constraint penalty term to form the loss function of the LSTM model coupled with physical constraints.

[0059] Preferably, the water balance constraint penalty term is calculated according to the following formula:

[0060] loss1 i = [1 / M×∑((Q i ,t-y1 t )×Δt+(WU i ,t-y5 t )-(V i ,T+1-V i ,t)) 2 ] 1 / 2 , t∈[1,M];

[0061] Among them, loss1 i is the water balance constraint penalty term of reservoir i, M is the number of samples extracted from the training set for each training, Q i ,t is the measured inflow of the reservoir, Δt is the unit time interval, WU i,t is the remaining forecast inflow of the reservoir, V i,T+ 1 is the storage capacity of the reservoir at the end of the period, V i,t It is the storage capacity of the reservoir;

[0062] The high-traffic weighted penalty term is calculated according to the following formula:

[0063] loss2 i = [1 / M×∑(QO i ,t / y4×(QO i ,t-y1 t )) 2 ] 1 / 2 , t∈[1,M];

[0064] Among them, loss2 i is the weighted penalty term for high flow of reservoir i, QO i,t Optimize outflow from reservoirs.

[0065] The discharge capacity constraint penalty term is calculated according to the following formula:

[0066] loss3 i = 1 / M×∑max{(y1 t -y2 t ),0} 2 , t∈[1,M];

[0067] Among them, loss3 i is the discharge capacity constraint penalty term of reservoir i;

[0068] The non-negative constraint penalty term is calculated according to the following formula:

[0069] loss4 i = 1 / M×∑min{y1 t ,0}×O,t∈[1,M];

[0070] Among them, loss4 i is the non-negative constraint penalty term for reservoir i, and O is a large constant;

[0071] The original loss function of the LSTM model is added to the water balance constraint penalty term, the high flow weighted penalty term, the discharge capacity constraint penalty term and the non-negative constraint penalty term to form the loss function of the LSTM model coupled with physical constraints.

[0072] Beneficial effects: By coupling a K-means clustering algorithm with an adaptive sliding time window, this paper scientifically and rationally divides reservoir outflow during the reservoir operation period into low-flow and high-flow processes. This paper proposes a hierarchical reservoir flood control operation rule extraction method based on an improved XGBoost-LSTM model. By organically coupling the XGBoost and LSTM models, the improved model is used to extract reservoir flood control operation rules for low-flow and high-flow processes, respectively. This method can quickly obtain high-precision regular reservoir outflows, providing an effective scheduling reference plan for reservoir flood control operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is a flow chart of the method of the present invention.

[0074] Figure 2 A schematic diagram of adding physical constraints to the loss function of the LSTM model.

[0075] Figure 3 This is a comparison chart of the reservoir regular outflow obtained by extracting the dispatching rules of a reservoir in the study area using this method and the reservoir regular outflow obtained by a single model.

[0076] Figure 4 This is a flow chart of the present invention using the K-means clustering algorithm coupled with an adaptive sliding time window. DETAILED DESCRIPTION

[0077] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0078] like Figures 1 to 4 As shown, the following technical solutions are proposed.

[0079] According to one aspect of the present application, a hierarchical reservoir flood control scheduling rule extraction method based on an improved XGBoost-LSTM includes the following steps:

[0080] Step S1, collecting hydrological data and engineering data, the hydrological data including the measured reservoir inflow at each time step, the interval water flow from the outflow section of each reservoir to the flood control section, the remaining predicted inflow of the reservoir, the remaining predicted outflow of the reservoir, and the end-of-term storage capacity of the reservoir; the engineering data including the water level-storage capacity curve and the water level-discharge capacity curve of the reservoir;

[0081] In this embodiment, the above information can be obtained from public data, with the purpose of gaining an in-depth understanding of the basic conditions and hydrological characteristics of the study area, and providing basic information and data support for subsequent model design and construction.

[0082] Step S2, constructing a multi-objective optimization scheduling model, reading the measured inflow of the reservoir, the interval water flow from the outflow section of each reservoir to the flood control section, the water level-reservoir capacity curve, and the water level-discharge capacity curve into the model, setting the objective function and constraints, and solving the problem using a multi-objective optimization method with a coupled time-related weight coefficient, dual-population collaboration, and information interaction mechanism to obtain a solution set of the multi-objective optimization problem;

[0083] In this embodiment, a multi-objective optimization method that couples time-correlated weight coefficients, dual-population collaboration, and information interaction mechanisms can significantly improve the dynamic adaptability, solution efficiency, and decision reliability of the model.

[0084] Step S3: Filter the solution set and extract the objective function value, reservoir storage at each time step, and reservoir optimized outflow data from the filtered solution set; based on the reservoir optimized outflow in the filtered solution set, use the K-means clustering algorithm coupled with an adaptive sliding time window to divide the reservoir optimized outflow into a low-flow process and a high-flow process; and based on the low-flow process and the high-flow process, organize the extracted data into the form of input data for subsequent GRU-XGBoost and LSTM models coupled with physical constraints as model samples;

[0085] In this embodiment, the K-means clustering algorithm coupled with an adaptive sliding time window is used to divide the optimized outflow of the reservoir into low-flow processes and high-flow processes, which solves the problems of "misclassification" and "missing classification" when using the traditional K-means clustering algorithm to classify the optimized outflow of the reservoir.

[0086] Step S4, introduce the gated recurrent unit (GRU) model, train the GRU model based on the reservoir optimized outflow, obtain the time series characteristics of the low-flow process, build a GRU-XGBoost and coupled physical constraint LSTM model based on the model sample, the low-flow process time series characteristics and the water level-discharge capacity curve, use the GRU-XGBoost model to extract the reservoir flood control scheduling rules for the low-flow process, and obtain the low-flow process regular outflow, use the LSTM model with coupled physical constraints to extract the reservoir flood control scheduling rules for the high-flow process, and obtain the high-flow process regular outflow, arrange the low-flow process regular outflow and the high-flow process regular outflow in a continuous time series to obtain the reservoir regular outflow, and use the reservoir optimized outflow to verify the output of the model.

[0087] In this embodiment, a hierarchical method of extracting scheduling rules is adopted to organically couple the two models, thereby fully improving the efficiency and accuracy of the models.

[0088] According to one aspect of the present application, step S2 is further:

[0089] Step S21, constructing a multi-objective optimization scheduling model, reading the measured inflow of the reservoir, the interval water flow from the outflow section of each reservoir to the flood control section, the water level-reservoir capacity curve, and the water level-discharge capacity curve into the model, setting the objective function as follows: minimizing the maximum flood control storage capacity utilization rate of the reservoir, minimizing the maximum average excess flood volume of the system, and minimizing the maximum flow at the flood control point. The constraints are water balance constraint, water level constraint, scheduling period initial condition constraint, scheduling period boundary condition constraint, outflow discharge constraint, discharge amplitude constraint, and non-negative variable constraint;

[0090] Step S22: Based on the objective function, a time-related weight coefficient is introduced, the priority of the objective function is adjusted according to the reservoir flood control operation stage, and the objective function is normalized to obtain an objective function based on a dynamic multi-objective decomposition mechanism;

[0091] Step S23: According to the objective function based on the dynamic multi-objective decomposition mechanism, a global exploration-local exploitation dual population collaboration and information interaction mechanism is adopted to perform multi-objective optimization and solve the problem, thereby obtaining a solution set of the multi-objective optimization problem.

[0092] According to one aspect of the present application, step S22 is further as follows:

[0093] Step S22a: Setting time-related weight coefficient w 1. w 2. w 3, among which w 1 is the weight coefficient of reservoir flood control capacity utilization rate, w 2 is the system average excess flood weight coefficient, w 3 is the flood control point flow weight coefficient, the formula is as follows:

[0094] w 1( t ) = -sin(π T / t )+1 t ∈[0, T ];

[0095] w 2( t ) = sin(π T / t ) t ∈[0, T ];

[0096] w 3( t ) = 0.5 t ∈[0, T ];

[0097] Step S22b: Time-related weight coefficient w 1. w 2. w 3 is multiplied by the corresponding objective function, and the objective function is normalized using the range method to obtain the objective function based on the dynamic multi-objective decomposition mechanism;

[0098] F i = ( F i - F i，min ) / ( Fi，max - F i，min );

[0099] In this embodiment, a time-related weight coefficient is introduced according to the stage characteristics of reservoir flood control scheduling. In the early stage of scheduling, pre-discharge is given priority, focusing on minimizing the utilization rate of flood control storage capacity. In the middle stage of scheduling, the flood peak needs to be reduced to reduce the risk of downstream floods, focusing on minimizing the average excess flood volume of the system. In the late stage of scheduling, the reservoir water volume needs to be discharged as much as possible, focusing on minimizing the utilization rate of flood control storage capacity, and preparing for the next flood. This is to better adapt to the needs of different scheduling stages. In addition, the use of the range method to normalize the objective function can effectively eliminate the dimensional differences between different objectives, making multi-objective optimization more effective.

[0100] According to one aspect of the present application, step S23 is further:

[0101] Step S23a: Construct a global exploration population and determine the number of splits according to the objective function based on the dynamic multi-objective decomposition mechanism. H and the number of objective functions M , using the Das-Dennis method to generate reference points N , N = ( H + M -1) + ( M -1), generate all the targets that satisfy x 1+ x 2+…+ x M = H The non-negative integer combination is normalized and mapped to the target space. After that, the objective function based on the dynamic multi-objective decomposition mechanism is associated with the nearest reference point using the vertical distance. Then, according to the current solution set distribution density, the reference points in the sparse area are increased and the reference points in the dense area are reduced. The individual fitness evaluation is adjusted to give priority to the solution that meets all constraints.

[0102] Step S23b: Construct local development population and design variation factor U =0.5×(1+rand), U ∈[0.5,1], adopting DE / current-to-pbest / 1 strategy to balance convergence speed and diversity;

[0103] V i = X i + U ( X pbest - X i )+UX r1 - X r2 );

[0104] Where: V i is the generated mutation vector; X i The first i individual; X pbest For the population p A randomly selected individual from the % elite individuals; X r1 、 X r2 are two different individuals randomly selected from the entire population ( r 1≠ r 2≠ i ); U is the variation factor;

[0105] In the crossover phase, the variable components that meet the constraints in the parent generation are preferentially selected to reduce the probability of generating infeasible solutions;

[0106] Step S23c: Select the top 20% of the non-dominated sorted solutions from the global population as candidate solutions A, and select the top 20% of the fitness from the local population as candidate solutions B to form a candidate pool;

[0107] Step S23d: Calculate the target space cosine similarity between candidate solutions A and B in the candidate pool and the target population individuals. Based on the information exchange mechanism, select the individual with the lowest similarity in the target population for each candidate solution to replace it, ensuring the introduction of diversity. If the minimum similarity between the candidate solution and the target population exceeds 0.9, the migration is abandoned.

[0108] Step S23e: When the solution sets of the two populations no longer change significantly within multiple generations, convergence is achieved and the solution set of the multi-objective optimization problem is obtained. Otherwise, go to step S23a for iterative calculation.

[0109] In this embodiment, the Das-Dennis method is used to generate initial reference points, and reference points are added according to the sparse solution set density area. The solution set shape is adaptive, and the DE / current-to-pbest / 1 strategy (differential evolution algorithm) is adopted. The top 20% elite individuals are used to guide the mutation direction, which is conducive to accelerating the convergence speed. According to the information interaction mechanism, the target space cosine similarity between the candidate solution and the target population individuals is calculated, and the individual with the lowest similarity is selected for replacement, effectively filling the blank area of the solution set.

[0110] According to one aspect of the present application, step S3 is further:

[0111] Step S31, divide the solution set into 7 categories: Category 1 is the minimum point of target 1; Category 2 is the minimum point of target 2; Category 3 is the minimum point of target 3; Category 4 is the solution that is balanced among targets 1, 2 and 3; Category 5 is the solution that is beneficial to target 1; Category 6 is the solution that is beneficial to target 2; Category 7 is the solution that is beneficial to target 3.

[0112] Step S32: Use the hierarchical clustering method based on Euclidean distance to select representative solution sets for categories 5, 6, and 7. For each category, set C A set of initial solutions, calculate the Euclidean distance between each solution in the category, and generate the Euclidean distance matrix composed of all solutions in the category D ,in D i，j For solution x i and y i The Euclidean distance of the objective function, set the number of representative solutions of this category C’ , generate a decision tree, and cluster the solutions in this category into C’ In each solution cluster, a solution with the shortest Euclidean distance to other solutions is selected as the representative solution set;

[0113] Step S33: average the objective function of each solution in the solution set of category 4, sort the average values from large to small, calculate the frequency using the expected value formula, and the solutions whose frequency differs from 50% by less than 0.01 form the representative solution set of category 4;

[0114] Step S34: Integrate the representative solution sets of category 1, category 2, category 3, and category 4, category 5, category 6, and category 7 to obtain a filtered solution set;

[0115] Step S35: extract the objective function value, the reservoir storage at each time step, and the optimized outflow of the reservoir from the filtered solution set; based on the optimized outflow of the reservoir in the filtered solution set, use the K-means clustering algorithm coupled with an adaptive sliding time window to divide the optimized outflow of the reservoir into a low-flow process and a high-flow process; and based on the low-flow process and the high-flow process, organize the above-extracted variables and hydrological data into the form of input data for the XGBoost and LSTM models, respectively, as samples for the XGBoost and LSTM models.

[0116] In this embodiment, a comprehensive screening method is proposed by combining the hierarchical clustering method and the physical meaning of the solution frontier to screen out a limited number of representative solutions from the solution frontier as research objects, thereby enhancing the pertinence and focus of learning, and solving the problem that there are a large number of solution solutions with small differences and high similarities between them. Directly using all solutions as input data for machine learning will result in excessively high computational costs and low efficiency.

[0117] According to one aspect of the present application, step S35 is further as follows:

[0118] Step S35a: Based on the optimized outflow of the reservoir in the filtered solution set, the optimized outflow time series of the reservoir is plotted into a reservoir optimized outflow process curve Q(t), and the scheduling period of the reservoir is set to T;

[0119] Step S35b, calculating the standard deviation σ of the slope S(t) of the reservoir optimized outflow process curve at each moment, which is used to reflect the magnitude of local fluctuations. When σ is large, it means that the data fluctuations in this period are severe, otherwise, the fluctuations are relatively stable;

[0120] Step S35c, use the continuous difference method to calculate the local average slope. First, set a dynamic sliding window according to the seasonal changes in reservoir operation. If |S(t)|<σ, set the sliding time window W∈[t-α*T, t+α*T]. If (t-α*T)<0 or (t+α*T)>T, then W∈[0, t+α*T]) or W∈[t-α*T, T]. If |S(t)|>σ, set the sliding time window W∈[t-β*T, t+β*T], where β<α<1. At this time, the local average slope S=(Q(t-1 / 2*W)- Q(t+1 / 2*W)) / W. If S is significantly positive or negative, it means that the reservoir optimization outflow process within the sliding time window has an obvious continuous upward or downward trend.

[0121] Step S35d: The optimized outflow process of the reservoir within the sliding time window is divided into a low-flow process candidate set and a high-flow process candidate set. If the slope standard deviation σ within the window exceeds the preset threshold ξ, it means that the data fluctuates violently during the period. At the same time, if the average slope S is close to 0 (i.e., there is no obvious overall upward or downward trend), it can be classified as a low-flow process candidate set. If the average slope S is significantly positive or negative, it means that the data as a whole shows a continuous upward or downward trend. At this time, even if there is a certain fluctuation, as long as σ does not exceed the preset threshold ξ, it is classified as a high-flow process candidate set.

[0122] Step S35e: Use the mean of the low-flow process candidate set and the high-flow process candidate set obtained in step S35d as the initial center point, and use the K-means clustering algorithm to perform clustering to obtain low-flow processes and high-flow processes.

[0123] In this embodiment, the K-means clustering algorithm is improved based on the characteristics that the low-flow process of reservoir optimization outflow has large fluctuations and the high-flow process has a clear trend of change. This solves the problem that the traditional K-means clustering results do not conform to the seasonal and cyclical fluctuations of reservoir scheduling, and the outflow in the flood season is attributed to the high-flow process, and the outflow in the dry season is attributed to the low-flow process.

[0124] According to one aspect of the present application, step S4 is further:

[0125] Step S41: Introduce a gated recurrent unit (GRU) model, train the GRU model based on reservoir outflow optimization, obtain time series features of the low-flow process, add the time series features of the low-flow process to the model sample to obtain a new sample, build a GRU-XGBoost model based on the new sample, and use the GRU-XGBoost model to extract reservoir flood control scheduling rules for the low-flow process to obtain regular outflow of the low-flow process;

[0126] Step S42: Based on the model sample and the water level-discharge capacity curve, an LSTM model coupled with physical constraints is constructed, and the LSTM model coupled with physical constraints is used to extract reservoir flood control scheduling rules for high flow processes to obtain regular outflow for high flow processes;

[0127] Step S43: Arrange the regular outflow during low flow and the regular outflow during high flow in a time series to obtain the regular outflow of the reservoir, define the water balance error, peak error, excess discharge capacity constraint percentage, and target value error evaluation indicators, and verify the output of the model based on the optimized outflow of the reservoir and the regular outflow of the reservoir using the root mean square error, deterministic coefficient, and the above-mentioned customized evaluation indicators.

[0128] According to one aspect of the present application, step S41 is further as follows:

[0129] Step S41a: introducing a gated recurrent unit (GRU) model, training the GRU model based on reservoir outflow optimization, and obtaining the time series characteristics of the low flow process;

[0130] Step S41b: Add the low-flow process time series features to the model sample to obtain a new sample, build a GRU-XGBoost prediction model based on the new sample, define the objective function, use the second-order Taylor expansion to optimize the objective function, and control the complexity through the regularization term;

[0131] Step S41c, dividing the new samples into a training set and a validation set, using the measured inflow of the reservoir, the interval water flow from the outflow section of each reservoir to the flood control section, the reservoir storage, the remaining predicted inflow of the reservoir, the objective function value, and the time series characteristics of the low-flow process as the independent variable factors of the GRU-XGBoost model, and the optimized outflow of the reservoir and the remaining predicted outflow of the reservoir as the dependent variable factors of the GRU-XGBoost model. The model is trained by a parallel tree generation algorithm, a greedy algorithm is used for feature splitting, and the shrinkage learning rate is introduced to control overfitting, and a nonlinear mapping relationship between the independent variable factors and the dependent variable factors is established;

[0132] Step S41d: Input the independent variable factors of the validation period into the model to obtain the prediction results of the corresponding dependent variable factors, extract the predicted low-flow process regular outflow from the prediction results, and calculate the fitting error based on the predicted low-flow process regular outflow and the low-flow process in the reservoir optimized outflow;

[0133] Step S41e: Based on the fitting error, the GRU model is trained twice to obtain the predicted fitting error, and the low-flow process rule outflow is obtained based on the predicted low-flow process rule outflow and the predicted fitting error, wherein the low-flow process rule outflow is the sum of the predicted low-flow process rule outflow and the predicted fitting error.

[0134] In this embodiment, a two-stage modeling approach is used to improve the XGBoost model by utilizing the time series features extracted by the GRU model and the error correction mechanism, solving the problem that XGBoost cannot directly establish long-term dependencies in time series, enabling the model to obtain more accurate prediction results.

[0135] According to one aspect of the present application, step S41c is further as follows:

[0136] Step S41c1: Based on the independent variable factors of the GRU-XGBoost model, a gradient boosting algorithm is used to construct a decision tree, a quantile discretization method is used to generate a set of candidate split thresholds, and based on the candidate split thresholds, the split gains of the respective variable factors are calculated in parallel using multi-threading technology;

[0137] Gain =0.5[ G L 2 / ( H L +λ)+ G R 2 / ( H R +λ)-( G L + G R ) 2 / ( H L + H R +λ)]-γ;

[0138] Where: G L 、 G R is the sum of the first-order gradients (first-order derivatives of the loss function) of the left and right child node samples; H L 、 HR is the sum of the second-order gradients (second-order derivatives of the loss function) of the left and right child node samples; λ is the regularization coefficient; γ is the split gain threshold parameter;

[0139] Step S41c2: traverse all independent variable factors and their splitting gains of the current node, use a greedy algorithm to iteratively select the optimal splitting point, set the maximum splitting gain as the splitting condition, and repeat the above process for the left and right child nodes until the maximum tree depth is reached or the gain is lower than the set threshold;

[0140] Step S41c3: Set a learning rate parameter. When the prediction value of each newly added tree is updated, multiply it by the learning rate parameter. By reducing the contribution of a single tree, the model is made more dependent on the weighted results of multiple trees. At the same time, a regularization term is used to prevent overfitting.

[0141] Step S41c4: Integrate the prediction results of each tree to obtain a nonlinear mapping relationship between the independent variable factors and the dependent variable factors.

[0142] In this embodiment, quantile discretization is used to reduce the search space for split thresholds. Multithreaded parallel computation of the split gains for each variable factor significantly accelerates model training. Based on an improved gain formula (including the regularization term λ and the threshold γ), only the local optimal solution is selected at each split, avoiding the complexity of the global search. The γ in the formula, serving as the split gain threshold, actively controls the growth of tree branches, preventing the generation of redundant subtrees and reducing memory consumption.

[0143] According to one aspect of the present application, step S42 is further as follows:

[0144] Step S42a: Based on the model sample and the water level-discharge capacity curve, a loss function is constructed, and an LSTM model using coupled physical constraints is built, wherein the LSTM model includes a forget gate, an input gate, a long-term state unit of the current input, a long-term state unit at the current moment, and an output gate. The loss function includes a root mean square error penalty term, a water balance constraint penalty term, a high flow weighted penalty term, a discharge capacity constraint penalty term, and a non-negative constraint penalty term.

[0145] Step S42b, dividing the model sample into a training period and a validation period, using the measured inflow of the reservoir, the water flow from the outflow section of each reservoir to the flood control section, the reservoir storage, the remaining predicted inflow of the reservoir, and the objective function value as the independent variable factors of the LSTM model using coupled physical constraints, and the optimized outflow of the reservoir and the remaining predicted outflow of the reservoir as the dependent variable factors of the LSTM model using coupled physical constraints, inputting the independent variable factors and the dependent variable factors of the training period into the LSTM model with coupled physical constraints for training, and obtaining a mapping relationship between the independent variable factors and the dependent variable factors;

[0146] Step S42c: Input the independent variable factors of the verification period into the model to calculate the prediction results of the corresponding dependent variable factors, that is, obtain the regular outflow of the high flow process.

[0147] According to one aspect of the present application, step S42a is further:

[0148] Step S42a1: Reading the measured reservoir inflow, the remaining predicted reservoir inflow, the remaining predicted reservoir outflow, the optimized reservoir outflow, the reservoir storage, the end-of-period storage, the reservoir's water level-capacity curve, and the water level-discharge curve data in the model sample, and obtaining the reservoir discharge capacity based on the reservoir storage, water level-capacity curve, and water level-discharge curve data;

[0149] Step S42a2: Setting parameters y 1. Parameters y 2. Parameters y 3. Parameters y 4. Parameters y 5, where the parameters y 1 is the parameter to be assigned to the value of the reservoir regular outflow predicted in step S42b, parameter y 2 is the parameter to be assigned to the reservoir discharge capacity. y 3 is the parameter y The maximum value of 1, parameter y 4 is the maximum value of the optimized outflow of the reservoir, parameter y 5 is the parameter to be assigned to the value of the predicted value of the remaining predicted outflow of the reservoir obtained by step S42b;

[0150] Step S42a3: Based on the measured inflow and parameters of the reservoir y 1. Reservoir remaining forecast inflow and parameters y 5. The end-of-period storage capacity and the storage capacity of the reservoir are used to calculate the water balance constraint penalty item according to the water balance principle;

[0151] loss 1 i =[1 / M ×∑(( Q i,t - y 1t )×Δ t +( WU i,t - y 5t )-( V i,T+1 - V i,t )) 2 ] 1 / 2 t ∈[1, M ];

[0152] Where: loss 1 i For reservoirs i Water balance constraint penalty item; M The number of samples drawn for each training set; Q i,t is the measured inflow to the reservoir; Δ t is the unit time interval; WU i,t Forecast inflow remaining for the reservoir; V i,T+1 is the storage capacity of the reservoir at the end of the period; V i,t It is the storage capacity of the reservoir;

[0153] Step S42a4: Optimize outflow and parameters based on reservoir y 3 and parameters y 4. Calculate the weight coefficient according to the change of reservoir outflow size, and multiply the weight coefficient by the reservoir optimized outflow and parameter y The difference of 1 results in a high-traffic weighted penalty term;

[0154] loss 2 i =[1 / M ×∑( QO i,t / y 4×( QO i,t - y 1t )) 2 ] 1 / 2 t ∈[1, M ];

[0155] Where: loss 2 i For reservoirs i A high-traffic weighted penalty term; QO i,t Optimizing outflow from reservoirs;

[0156] Step S42a5: Compare parameters y 1 with parameters y 2, if the parameter y 1 is greater than the parameter y 2, then calculate the root mean square error of the two and obtain the discharge capacity constraint penalty term. If the parameter y 1 is less than the parameter y 2, then the discharge capacity constraint penalty term is 0;

[0157] loss 3i =1 / M ×∑max{( y 1t - y 2t ),0} 2 t ∈[1, M ];

[0158] Where: loss 3 i For reservoirs i The discharge capacity constraint penalty item;

[0159] Step S42a6: Compare parameters y The size of 1 and 0, the absolute value of the smaller value between the two and the larger value O Multiply them together to get the non-negative constraint penalty term;

[0160] loss 4 i =1 / M ×∑min{ y 1t ,0}× O t ∈[1, M ];

[0161] Where: loss 4 i For reservoirs i The non-negative constraint penalty term;

[0162] Step S42a7: Extract the LSTM model loss function, add the water balance constraint penalty term, the high flow weighted penalty term, the discharge capacity constraint penalty term, and the non-negative constraint penalty term to the LSTM model loss function to form a new loss function, and establish an LSTM model coupled with physical constraints.

[0163] In this embodiment, in order to improve the applicability and accuracy of the LSTM model in the extraction of scheduling rules for high-flow processes, and to address the problem of insufficient physical interpretability of traditional data-driven models, water balance constraints, high-flow weighting, discharge capacity constraints, and non-negative constraint penalty terms are added to the loss function, effectively limiting the solution space of the LSTM model when solving practical problems, and ensuring that the output of the model is closer to the actual scheduling rules.

[0164] In summary, by applying the hierarchical reservoir flood control scheduling rule extraction method based on improved XGBoost-LSTM proposed in this application, it is possible to obtain a multi-objective solution set with stage characteristics that conforms to reservoir flood control scheduling, scientifically and reasonably divide the optimized reservoir outflow into low-flow process and high-flow process, improve the XGBoost and LSTM models, and organically couple the improved models to obtain reservoir regular outflow, achieve a balance between computing resources and accuracy, and quickly provide an efficient and feasible reference solution for reservoir flood control scheduling.

[0165] According to one aspect of the present application, a specific output of the present invention is regular outflow of a reservoir.

[0166] The technical concept of this invention is as follows: To address the multi-objective flood control scheduling problem of a reservoir cluster, a multi-objective model was constructed to solve this problem. However, the computational time is long, making it difficult to apply to actual flood control scheduling decision-making scenarios. This problem directly affects the timeliness of flood control decision-making. To this end, we first attempted to introduce an LSTM to fit the outflow process of the reservoir cluster in an existing multi-objective optimized flood control scheduling scheme, thereby extracting the flood control scheduling rules for the reservoir cluster. The extracted flood control scheduling rules for the reservoir cluster were successfully extracted, with a certainty coefficient of 0.8.

[0167] However, research has revealed that LSTM still has the following issues: outflow fitting does not meet actual physical constraints and lacks physical interpretability; the input data volume is large, and the computation is complex. To address these issues, the following solutions are proposed: physical mechanism constraints are introduced into the LSTM model loss function, including water balance constraints, high-flow process constraints, discharge capacity constraints, and non-negativity constraints; a K-means clustering algorithm is used to divide reservoir outflow into low-flow and high-flow processes; the high-flow process is fitted using the improved LSTM model described above, while the low-flow process is fitted using the XGBoost model, using a coupled model to improve computational efficiency.

[0168] However, in actual use, if the traditional K-means clustering algorithm is used to pre-process the outflow of a reservoir group, the following problems will arise: some periods are flood seasons, and some periods are dry seasons. The scheduling of reservoirs has different rules in different seasons or periodic changes. The traditional K-means clustering results do not conform to the seasonal and periodic fluctuations of reservoir scheduling. The outflow in the flood season will be attributed to the high-flow process, and the outflow in the dry season will be attributed to the low-flow process.

[0169] Therefore, the following improvements are made to the K-means clustering algorithm: the standard deviation of the slope of the reservoir optimized outflow process curve at each moment is calculated; the local average slope is calculated to determine whether the reservoir optimized outflow process within the dynamic sliding window set according to the seasonal changes in reservoir scheduling has an obvious continuous upward or downward trend; according to the calculated slope standard deviation and average slope within the window, the reservoir optimized outflow process curve is divided into a low-flow process candidate set and a high-flow process candidate set, and the mean of the candidate set is used as the initial center point. The K-means clustering algorithm is used for clustering to obtain the low-flow process and the high-flow process.

[0170] At the same time, in the actual process, the XGBoost model is used to fit the low-flow process, and the following problems arise: the low-flow process of the reservoir has continuity and time characteristics, but XGBoost as a decision tree integration method does not have time dependence, and the predicted value shows large fluctuations and is not smooth enough.

[0171] Therefore, the following improvements were made: the gated recurrent unit (GRU) was introduced, and the GRU was first trained to extract the temporal features of the low-flow process. This feature and other factors were used as the input of XGBoost, and the XGBoost model was used to fit the low-flow process of reservoir outflow; the fitting error of XGBoost was calculated, and the GRU was then trained a second time to learn the temporal variation trend of the error, and the fitting value of XGBoost was corrected according to the error with temporal characteristics.

[0172] In this embodiment, after all the above steps, a regular outflow of the reservoir can be obtained, which quickly provides an effective scheduling reference plan for reservoir flood control scheduling, and has high technical innovation and practicality. The above describes the preferred embodiments of the present invention in detail, but the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solution of the present invention, and these equivalent transformations all fall within the scope of protection of the present invention.

Claims

1. A hierarchical reservoir flood control scheduling rule extraction method based on improved XGBoost-LSTM, characterized in that: The following steps are involved: Collect hydrological data and engineering data as input, use multi-objective optimization method to solve the reservoir flood control operation model and obtain the solution set; The solution set is screened, and based on the reservoir optimized outflow obtained from the screened solution set, a clustering algorithm is used to divide it into low-flow process and high-flow process; The GRU-XGBoost model is used to extract scheduling rules for low-flow processes and obtain the regular outflow of low-flow processes; An LSTM model coupled with physical constraints is used to extract scheduling rules for high-flow processes and obtain the regular outflow of high-flow processes; Combine the regular outflows of high and low flow processes into the regular outflow of the reservoir; Extract scheduling rules for high-traffic processes, including: Based on the model samples and the water level-discharge capacity curve, a loss function was constructed that included a root mean square error penalty, a water balance constraint penalty, a high flow weighted penalty, a discharge capacity constraint penalty, and a non-negative constraint penalty. An LSTM model coupled with physical constraints was then built and trained. The independent variable factors of the validation period are input into the LSTM model coupled with physical constraints to obtain the regular outflow of high-flow processes; Among them, the water balance constraint penalty term loss1 i = [1 / M×∑((Q i,t -y1 t )×Δt+(WU i,t -y5 t )-(V i,T +1-V i,t )) 2 ] 1 / 2 , t∈[1,M]; where loss1 i is the water balance constraint penalty term of reservoir i, M is the number of samples extracted from the training set for each training, Q i,t is the measured inflow to the reservoir, Δt is the unit time interval, WU i,t is the remaining forecast inflow of the reservoir, V i,T+1 is the storage capacity of the reservoir at the end of the period, V i,t It is the storage capacity of the reservoir; High traffic weighted penalty loss2 i = [1 / M×∑(QO i ,t / y4×(QO i ,t-y1 t )) 2 ] 1 / 2 , t∈[1,M]; where loss2 i is the weighted penalty term for high flow of reservoir i, QO i,t Optimizing outflow from reservoirs; y1 is the predicted regular outflow value of the reservoir, y4 is the maximum value of the optimized outflow of the reservoir, and parameter y5 is the predicted residual forecast outflow value of the reservoir; Discharge capacity constraint penalty loss3 i = 1 / M×∑max{(y1 t -y2 t ),0} 2 , t∈[1,M]; where loss3 i is the discharge capacity constraint penalty term of reservoir i; Non-negative constraint penalty term loss4 i = 1 / M×∑min{y1 t ,0}×O,t∈[1,M]; where loss4 i is the non-negative constraint penalty term of reservoir i, O is a large constant; parameter y2 is the reservoir discharge capacity.

2. The method according to claim 1, characterized in that Build models and obtain solutions to multi-objective optimization problems, including: Construct a multi-objective optimization scheduling model, set the objective function as minimizing the maximum flood control storage capacity utilization rate of the reservoir, minimizing the maximum average excess flood volume of the system, minimizing the maximum flow at the flood control point, and set constraints; Based on the objective function, the time-related weight coefficient is introduced, the priority of the objective function is adjusted according to the reservoir flood control scheduling stage, and the objective function is normalized. The objective function based on the dynamic multi-objective decomposition mechanism is obtained and solved to obtain the solution set of the multi-objective optimization problem.

3. The method according to claim 2, characterized in that Construct an objective function based on a dynamic multi-objective decomposition mechanism, including: Set the time-related weight coefficients w1, w2, and w3, including: The weight coefficient of reservoir flood control capacity utilization rate is w1(t) = -sin(πT / t)+1, t∈[0,T]; System average excess flood weight coefficient w2(t) = sin(πT / t), t∈[0,T]; The flood control point flow weight coefficient w3(t) = 0.5, t∈[0,T]; where t represents the current time and T represents the total scheduling duration; The time-related weight coefficients w1, w2, and w3 are multiplied by the corresponding objective function, and the maximum and minimum values of the objective function are normalized using the range method to obtain the objective function based on the dynamic multi-objective decomposition mechanism.

4. The method according to claim 2, characterized in that Solve the objective function based on the dynamic multi-objective decomposition mechanism, including: Construct a global exploration population and a local development population; select candidate solutions from the global exploration population and the local development population to form a candidate pool; Calculate the target space cosine similarity between the candidate solutions in the candidate pool and the target population individuals in sequence, and select the individual with the lowest similarity for replacement based on the information interaction mechanism; When the solution sets of the two populations change less than a threshold for consecutive K generations, convergence is determined and the solution set of the multi-objective optimization problem is obtained.

5. The method according to claim 1, characterized in that The solution set is screened to obtain the optimized outflow of the reservoir and cluster it, including: The solution set is divided into 7 categories, including the minimum points of objectives 1, 2, and 3, the three-objective equilibrium solution, and the solution that is beneficial to objectives 1, 2, and 3; A hierarchical clustering method based on Euclidean distance is used to select a representative solution set for the solutions that are beneficial to each goal, and a solution with the shortest Euclidean distance to other solutions is selected as the representative solution in each cluster; Calculate the average value of the objective function of each solution in the equilibrium solution set, sort the average values and select the equilibrium solution as the representative solution set; integrate the representative solution sets of all categories to obtain the filtered solution set; The objective function value, reservoir storage and optimized outflow of the reservoir are extracted from the filtered solution set. Based on the optimized outflow of the reservoir, the K-means clustering algorithm coupled with an adaptive sliding time window is used to divide the optimized outflow of the reservoir into low-flow process and high-flow process, which are used as sample data for subsequent models.

6. The method according to claim 5, characterized in that Use the K-means clustering algorithm coupled with an adaptive sliding time window, including: The optimized outflow sequence of the reservoir is plotted as the optimized outflow process curve Q(t), and the scheduling period of the reservoir is set as T; Calculate the standard deviation σ of the slope S(t) of the reservoir's optimized outflow process curve at each moment to reflect the magnitude of local fluctuations; The local average slope is calculated using the continuous difference method. A dynamic sliding time window W is set based on the comparison between |S(t)| and σ. When |S(t)|<σ, W∈[t-α·T, t+α·T]; when |S(t)|>σ, W∈[t-β·T, t+β·T], where β<α<1. The local average slope S within the window is calculated. Based on the comparison between the slope standard deviation σ within the window and the preset threshold ξ, as well as the value of the average slope S, the reservoir optimized outflow is divided into a low-flow process candidate set and a high-flow process candidate set; The mean of the low-flow process candidate set and the high-flow process candidate set are used as the initial center point, and the K-means clustering algorithm is used to cluster the low-flow process and the high-flow process are obtained.

7. The method according to claim 1, characterized in that The GRU-XGBoost model is used to extract scheduling rules for low-traffic processes, including: The GRU model is introduced and trained based on the optimized outflow of the reservoir to obtain the time series characteristics of the low flow process; The time series characteristics of low-flow processes were added to the model samples, and the GRU-XGBoost prediction model was constructed and preliminarily trained. The second-order Taylor expansion was used to optimize the objective function, and the complexity was controlled by the regularization term. The independent variable factors of the validation set are input into the GRU-XGBoost model to obtain the regular outflow for predicting low flow process and calculate the fitting error; The GRU model is trained twice based on the fitting error to obtain the predicted fitting error. The predicted regular outflow during low flow process is added to the predicted fitting error to obtain the regular outflow during low flow process.

Images