Classified reservoir flood control scheduling rule extraction method based on improved XGBoost-LSTM

Abstract

The invention discloses a graded reservoir flood control scheduling rule extraction method based on an improved XGBoost-LSTM. The method comprises the following steps: collecting hydrological data and engineering data; constructing a multi-objective optimization scheduling model, and obtaining a solution set of a multi-objective optimization problem by adopting a multi-objective optimization method of a coupling time correlation weight coefficient, double-population cooperation and an information interaction mechanism; screening the solution set, and based on reservoir optimization outflow in the screened solution set, grading the reservoir optimization outflow by using a K-means clustering algorithm coupled with an adaptive sliding time window; and according to a grading result, extracting a reservoir flood control scheduling rule by using an improved XGBoost-LSTM grading reservoir flood control scheduling rule extraction method to obtain reservoir rule outflow, and verifying the reservoir rule outflow by using an evaluation index. According to the invention, an effective scheduling reference scheme can be quickly provided for reservoir flood control scheduling.

Description

Technical Field

[0001] The invention of this application relates to the technical field of reservoir flood control operation rule extraction, especially a method for extracting hierarchical reservoir flood control operation rules based on improved XGBoost-LSTM. Background Art

[0002] With the continuous expansion of the scale of projects such as reservoirs, the complexity of the real-time flood control joint operation model of the reservoir system is getting higher and higher. The traditional overall joint operation of the basin system faces increasing challenges from modeling to solving. Even in the current situation of rapid progress in computing means, the contradiction between the difficulty of solving the high-dimensional and non-linear mathematical model of real-time flood control operation of reservoir groups and the high-timeliness requirement of flood control decision-making is still very sharp. Especially, the joint operation of the flood control system needs to take into account the safety guarantee of each project and each protected object within the system. Eventually, it often has to face the decision-making problem of selecting an operation plan from many non-inferior alternative plans, constituting a large-scale system and multi-objective non-linear optimization problem. For the modeling research of similar problems, most adopt the method of constructing a multi-objective optimization model and solving it. However, multi-objective optimization has problems such as low computational efficiency, a large number of solutions, and difficulty in differentiating flood coordination relationships, which are limited in practical applications of basin flood control operation.

[0003] In the actual operation of reservoirs, the implementation of reservoir operation usually takes the established reservoir operation rules as the operation criterion. Reservoir operation rules are based on design tasks and operation constraints, and carry out reservoir operation based on a large number of sample series, and then summarize the optimal operation laws of the reservoir. In recent years, computer technologies such as machine learning and data mining have gradually emerged. The XGBoost and LSTM models have achieved many research results in the extraction of non-linear, multi-dimensional, and large-scale system operation rules. However, generally speaking, they are still in the initial stage and still have the following deficiencies: (1) XGBoost cannot directly establish long-term dependencies of time series. The reservoir outflow has connections in both time and space. If XGBoost is applied to the extraction of operation rules for the entire operation period, it may affect the simulation prediction accuracy; (2) The black-box characteristics of artificial intelligence models such as LSTM do not directly consider the complex physical mechanisms involved in the actual operation process, have low interpretability, and have high computational complexity and long training time, reducing the practicality of the operation rules extracted by them in actual operation. If the XGBoost and LSTM models can be improved according to the characteristics of reservoir operation outflow and organically coupled, then an intelligent operation rule extraction model that takes into account both computational efficiency and simulation accuracy can be constructed to quickly provide an effective and physically interpretable operation reference plan for reservoir flood control decision-making. Summary of the Invention

[0004] The invention aims to provide a method for extracting hierarchical reservoir flood control operation rules based on improved XGBoost-LSTM to solve the above problems existing in the prior art.

[0005] Technical solution: Provide a method for extracting hierarchical reservoir flood control scheduling rules based on improved XGBoost-LSTM, including the following steps: Collect hydrological data and engineering data as inputs, solve the reservoir flood control scheduling model using a multi-objective optimization method, and obtain a solution set; Screen the solution set, and based on the optimized reservoir outflow obtained from the screened solution set, use a clustering algorithm to divide it into low-flow processes and high-flow processes; Use the GRU-XGBoost model to extract scheduling rules for the low-flow process to obtain the rule outflow of the low-flow process; use the LSTM model coupled with physical constraints to extract scheduling rules for the high-flow process to obtain the rule outflow of the high-flow process; combine the rule outflows of the high- and low-flow processes into the reservoir rule outflow.

[0006] Preferably, constructing a model and obtaining the solution set of the multi-objective optimization problem includes: Construct a multi-objective optimization scheduling model, set the objective functions as the minimum of the maximum utilization rate of the reservoir's flood control storage capacity, the minimum of the maximum average excess flood volume of the system, and the minimum of the maximum flow at the flood control point, and set the constraint conditions; Based on the objective functions, introduce time-correlated weight coefficients, adjust the priority of the objective functions according to the reservoir flood control scheduling stage, and normalize the objective functions to obtain the objective functions based on the dynamic multi-objective decomposition mechanism; and use the global exploration-local development double-population cooperation and information interaction mechanism to solve, and obtain the solution set of the multi-objective optimization problem.

[0007] Preferably, constructing the objective functions based on the dynamic multi-objective decomposition mechanism includes: Set time-correlated weight coefficients w1, w2, w3, including: The weight coefficient of the utilization rate of the reservoir's flood control storage capacity w1(t) = -sin(πT / t)+1, t∈[0,T]; The weight coefficient of the average excess flood volume of the system w2(t) = sin(πT / t), t∈[0,T]; The weight coefficient of the flow at the flood control point w3(t) = 0.5, t∈[0,T]; where t represents the current time and T represents the total scheduling duration.

[0008] Multiply the time-correlated weight coefficients w1, w2, w3 by the corresponding objective functions, and use the range method to normalize the maximum and minimum values of the objective functions to obtain the objective functions based on the dynamic multi-objective decomposition mechanism.

[0009] Preferably, using the global exploration-local development double-population cooperation and information interaction mechanism for multi-objective optimization solution includes: Construct a global exploration population, determine the number of partitions H and the number of objective functions M, generate reference points N using the Das-Dennis method 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; Construct a local development population, and design a mutation factor U = 0.5×(1 + rand), U ∈ [0.5, 1], and use the DE / current-to-pbest / 1 strategy to generate a mutant vector Vi = Xi + U(Xpbest - Xi) + U(Xr1 - Xr2); where, Xi is the i-th individual in the current population, Xpbest is an individual randomly selected from the top p% of the elite individuals in the population, and Xr1 and Xr2 are two different randomly selected individuals; Select the top a% of the solutions from the global exploration population as candidate solutions A, and select the top a% from the local development population as candidate solutions B to form a candidate pool; Calculate the cosine similarity of the candidate solutions in the candidate pool with the individuals in the target population, and select the individual with the lowest similarity for replacement according to the information interaction mechanism; When the change in the solution sets of the two populations is less than the threshold for K consecutive generations, it is determined that convergence has occurred, and the solution set of the multi-objective optimization problem is obtained.

[0010] Preferably, screen the solution set to obtain the optimized reservoir outflow and cluster it, including: Divide the solution set into 7 categories, including the minimum value point of objective 1, the minimum value point of objective 2, the minimum value point of objective 3, the three-objective equilibrium solution, the solution beneficial to objective 1, the solution beneficial to objective 2, and the solution beneficial to objective 3; Use the hierarchical clustering method based on Euclidean distance to select representative solution sets for the solutions beneficial to each objective, and select a solution with the shortest Euclidean distance from other solutions in each cluster as the representative solution; Calculate the average value of the objective functions of each solution in the equilibrium solution set, and select the solution whose difference between the frequency and 50% is less than 0.01 after sorting the average values as the representative solution set of the equilibrium solution; Integrate the representative solution sets of all categories to obtain the screened solution set; Extract the objective function values, reservoir storage, and optimized reservoir outflow from the screened solution set. Based on the optimized reservoir outflow, use the K-means clustering algorithm with a coupled adaptive sliding time window to divide the optimized reservoir outflow into low-flow processes and high-flow processes as the sample data for the subsequent model.

[0011] Preferably, use the K-means clustering algorithm with a coupled adaptive sliding time window, including: Plot the optimized reservoir outflow in chronological order to obtain the optimized reservoir outflow process curve Q(t), and set the scheduling period of the reservoir as T; Calculate the standard deviation σ of the slope S(t) at each moment of the optimized outflow process curve of the reservoir to reflect the magnitude of local fluctuations; Use the continuous difference method to calculate the local average slope. Set the dynamic sliding time window W according to the comparison result of |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. Calculate the local average slope S within the window; Based on the comparison between the standard deviation σ of the slope within the window and the preset threshold ξ, and the value of the average slope S, divide the optimized outflow of the reservoir into a candidate set of low-flow processes and a candidate set of high-flow processes; Take the means of the candidate set of low-flow processes and the candidate set of high-flow processes as the initial center points, and use the K-means clustering algorithm for clustering to obtain low-flow processes and high-flow processes.

[0012] Preferably, use the GRU-XGBoost model to extract the scheduling rules for low-flow processes, including: Introduce the GRU model, and train the GRU model based on the optimized outflow of the reservoir to obtain the time series characteristics of low-flow processes; Add the time series characteristics of low-flow processes to the model samples, construct the GRU-XGBoost prediction model, optimize the objective function using the second-order Taylor expansion, and control the complexity through the regularization term; Divide the samples into a training set and a validation set. Use the measured inflow of the reservoir, the inflow from the intermediate section, the reservoir storage, the remaining predicted inflow of the reservoir, the value of the objective function, and the time series characteristics of low-flow processes as independent variable factors, and the optimized outflow of the reservoir and the remaining predicted outflow of the reservoir as dependent variable factors to train the GRU-XGBoost model; Input the independent variable factors of the validation set into the GRU-XGBoost model to obtain the predicted regular outflow of the low-flow process, and calculate the fitting error; Based on the fitting error, perform secondary training on the GRU model to obtain the predicted fitting error. Add the predicted regular outflow of the low-flow process and the predicted fitting error to obtain the regular outflow of the low-flow process.

[0013] Preferably, it includes: Based on the independent variable factors of the GRU-XGBoost model, use the gradient boosting algorithm to construct decision trees, generate a set of candidate splitting thresholds using the quantile discretization method, and use multi-threading technology to calculate the splitting gains of each independent variable factor in parallel; Traverse all independent variable factors and their splitting gains of the current node, and use the greedy algorithm to select the point with the largest splitting gain as the splitting condition. 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; Set the learning rate parameter, which is multiplied when updating the predicted value of each newly added tree. By reducing the contribution of a single tree, the model relies more on the weighted results of multiple trees. At the same time, combine the regularization term to prevent overfitting; Integrate the prediction results of each tree to obtain the non - linear mapping relationship between the independent variable factor and the dependent variable factor.

[0014] Preferably, the specific steps for extracting the scheduling rules for the high - flow process include: Based on the model samples and the water - level - discharge capacity curve, construct a loss function that includes the root - mean - square error penalty term, 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, and build an LSTM model coupled with physical constraints; Divide the model samples into a training period and a validation period. Use the measured reservoir inflow, inflow from the intermediate area, reservoir storage, remaining predicted reservoir inflow, and objective function value as independent variable factors, and use the optimized reservoir outflow and remaining predicted reservoir outflow as dependent variable factors to train the LSTM model coupled with physical constraints; Input the independent variable factors in the validation period into the LSTM model coupled with physical constraints to obtain the regular outflow of the high - flow process.

[0015] Preferably, building an LSTM model coupled with physical constraints includes: Based on the reservoir storage, water - level - storage curve, and water - level - discharge curve data in the model samples, obtain the reservoir discharge capacity; Set parameters y1, y2, y3, y4, 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 reservoir outflow, and parameter y5 is the predicted remaining reservoir outflow value; Based on the measured reservoir inflow, parameter y1, remaining predicted reservoir inflow, parameter y5, end - of - period reservoir storage, and reservoir storage, calculate the water - balance constraint penalty term according to the principle of water balance; Based on the optimized reservoir outflow, parameter y3, and parameter y4, calculate the weight coefficient according to the change in the size of the reservoir outflow to obtain the high - flow weighted penalty term; Compare the magnitudes of parameter y1 and parameter y2. If parameter y1 is greater than parameter y2, then calculate the discharge - capacity constraint penalty term, otherwise this penalty term is 0; Compare the magnitude of parameter y1 and 0, and calculate the non - negative constraint penalty term; Add the LSTM model loss function and the above - mentioned constraint penalty terms to form the loss function of the LSTM model coupled with physical constraints.

[0016] Preferably, calculate the water - balance constraint penalty term according to the following formula: loss1i = [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]; Among them, loss1 i is the penalty term for the water balance constraint of reservoir i, M is the number of samples extracted for each training in the training set, Q i ,t is the measured inflow of the reservoir, Δt is the unit time interval, WU i,t is the remaining predicted inflow of the reservoir, V i,T+ 1 is the end - of - period storage of the reservoir, V i,t is the storage of the reservoir; Calculate the high - flow weighted penalty term according to the following formula: loss2 i = [1 / M × ∑(QO i ,t / y4 × (QO i ,t - y1 t )) 2 1 / 2 , t ∈ [1, M]; Among them, loss2 i is the high - flow weighted penalty term of reservoir i, QO i,t is the optimized outflow of the reservoir.

[0017] Calculate the penalty term for the discharge capacity constraint according to the following formula: loss3 i = 1 / M × ∑max{(y1 t - y2 t ), 0} 2 , t ∈ [1, M]; Among them, loss3 i is the penalty term for the discharge capacity constraint of reservoir i; Calculate the penalty term for the non - negative constraint according to the following formula: loss4 i = 1 / M × ∑min{y1 t , 0} × O, t ∈ [1, M]; Among them, loss4 i is the penalty term for the non - negative constraint of reservoir i, O is a large constant; ​​Add the original loss function of the LSTM model 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.

[0018] Beneficial effects: By coupling the K - means clustering algorithm with an adaptive sliding time window, the present invention scientifically and reasonably divides the reservoir outflow during the reservoir operation period into low - flow processes and high - flow processes, and proposes a method for extracting hierarchical reservoir flood control operation rules based on improved XGBoost - LSTM. The XGBoost and LSTM models are organically coupled, and the improved models are used to extract reservoir flood control operation rules for low - flow processes and high - flow processes respectively. Using this method can quickly obtain high - precision reservoir regulated outflow, and can quickly provide an effective scheduling reference plan for reservoir flood control operation. Description of the Drawings

[0019] Figure 1 is a flowchart of the method of the present invention.

[0020] Figure 2 is a schematic diagram of adding physical constraints to the loss function of the LSTM model.

[0021] Figure 3 is a comparison chart of the reservoir regulated outflow obtained by using this method for extracting operation rules and the reservoir regulated outflow obtained by a single model for a certain reservoir in the study area.

[0022] Figure 4 is a flowchart of the present invention using the K - means clustering algorithm with an adaptive sliding time window. Detailed Embodiments

[0023] The present invention will be described in detail below with reference to the drawings and specific embodiments.

[0024] As Figures 1 to 4 shown, the following technical solutions are proposed.

[0025] According to one aspect of the present application, a method for extracting hierarchical reservoir flood control operation rules based on improved XGBoost - LSTM includes the following steps: Step S1: Collect hydrological data and engineering data. The above - mentioned hydrological data includes the measured reservoir inflow at each time step, the inflow between the reservoir outflow section and the flood control section, the remaining predicted reservoir inflow, the remaining predicted reservoir outflow, and the reservoir end - of - period storage. The engineering data includes the water level - storage capacity curve and the water level - discharge capacity curve of the reservoir. In this embodiment, the above-mentioned data can be obtained from public data, aiming to deeply understand the basic situation and hydrological characteristics of the research area, and provide basic information and data support for subsequent model design and construction.

[0026] Step S2: Construct a multi-objective optimal operation model. Read the measured reservoir inflow, the inflow between each reservoir outflow section and the flood control section, the water level-storage curve, and the water level-discharge capacity curve into the model. Set the objective function and constraints, and use a multi-objective optimization method that couples the time-correlated weight coefficient, double-population cooperation, and information interaction mechanism to solve the problem, obtaining the solution set of the multi-objective optimization problem. In this embodiment, the multi-objective optimization method that couples the time-correlated weight coefficient, double-population cooperation, and information interaction mechanism can significantly improve the dynamic adaptability, solution efficiency, and decision-making reliability of the model.

[0027] Step S3: Screen the solution set, and extract the objective function values, the reservoir storage at each time step, and the data of the optimized reservoir outflow in the screened solution set. Based on the optimized reservoir outflow in the screened solution set, use the K-means clustering algorithm that couples an adaptive sliding time window to divide the optimized reservoir outflow into a low-flow process and a high-flow process. Based on the low-flow process and the high-flow process, organize the above-extracted data into the form of the input data of the subsequent GRU-XGBoost and LSTM model that couples physical constraints respectively, as model samples. In this embodiment, using the K-means clustering algorithm that couples an adaptive sliding time window to divide the optimized reservoir outflow into a low-flow process and a high-flow process solves the problems of "misclassification" and "omission of classification" when using the traditional K-means clustering algorithm for classifying the optimized reservoir outflow.

[0028] Step S4: Introduce a gated recurrent unit (GRU) model. Based on the optimized reservoir outflow, train the GRU model to obtain the time-series characteristics of the low-flow process. Based on the model samples, the time-series characteristics of the low-flow process, and the water level-discharge capacity curve, build the GRU-XGBoost and LSTM model that couples physical constraints. Use the GRU-XGBoost model to extract the reservoir flood control operation rules for the low-flow process, obtaining the rule outflow of the low-flow process. Use the LSTM model that couples physical constraints to extract the reservoir flood control operation rules for the high-flow process, obtaining the rule outflow of the high-flow process. Arrange the rule outflow of the low-flow process and the rule outflow of the high-flow process in continuous time series to obtain the reservoir rule outflow, and use the optimized reservoir outflow to verify the output of the model.

[0029] In this embodiment, by adopting the method of hierarchical extraction of operation rules, the two models are organically coupled, which fully improves the efficiency and accuracy of the model.

[0030] According to one aspect of the present application, step S2 is further as follows: Step S21: Construct a multi-objective optimal scheduling model, read the measured reservoir inflow, the inflow between each reservoir outflow section and the flood control section, the water level-storage capacity curve, and the water level-discharge capacity curve into the model, and set the objective functions as follows: the minimum of the maximum flood control storage utilization rate of the reservoir, the minimum of the maximum average excess flood volume of the system, and the minimum of the maximum flow at the flood control point. The constraint conditions include water balance constraint, water level constraint, initial condition constraint of the scheduling period, boundary condition constraint of the scheduling period, outflow discharge constraint, discharge variation constraint, and non-negative variable constraint; Step S22: Based on the objective functions, introduce a time-related weight coefficient, adjust the priority of the objective functions according to the reservoir flood control scheduling stage, and normalize the objective functions to obtain the objective functions based on the dynamic multi-objective decomposition mechanism; Step S23: According to the objective functions based on the dynamic multi-objective decomposition mechanism, use the global exploration-local development double-population cooperation and information interaction mechanism to perform multi-objective optimization solution to obtain the solution set of the multi-objective optimization problem.

[0031] According to one aspect of the present application, step S22 is further as follows: Step S22a: Set the time-related weight coefficient w 1 , w 2 , w 3 , where w 1 is the weight coefficient of the flood control storage utilization rate of the reservoir, w 2 is the weight coefficient of the average excess flood volume of the system, w 3 is the weight coefficient of the flow at the flood control point. The formulas are as follows: w 1 ( t ) = -sin(π T / t ) + 1 t ∈[0, T ; w 2 ( t ) = sin(π T / t ) t ∈[0, T ; w 3 ( t ) = 0.5 t ∈[0, T ; Step S22b: Multiply the time - related weight coefficient w 1 , w 2 , w 3 with the corresponding objective function, and normalize the objective function using the range method to obtain the objective function based on the dynamic multi - objective decomposition mechanism; F i = ( F i - F i，min ) / ( F i，max - F i，min ); In this embodiment, according to the stage characteristics of reservoir flood control scheduling, a time - related weight coefficient is introduced. In the initial stage of scheduling, it is necessary to give priority to pre - discharging, focusing on minimizing the utilization rate of flood control storage capacity. In the middle stage of scheduling, it is necessary to reduce the peak flood and reduce the downstream flood risk, focusing on minimizing the average excess flood volume of the system. In the later stage of scheduling, it is necessary to discharge the reservoir water as much as possible, focusing on minimizing the utilization rate of flood control storage capacity to prepare for the next flood, better adapting to the needs of different scheduling stages. And using the range method to normalize the objective function can effectively eliminate the dimensional differences between different objectives, making the multi - objective optimization more effective.

[0032] According to one aspect of the present application, step S23 is further as follows: Step S23a: Construct a global exploration population. According to the objective function based on the dynamic multi - objective decomposition mechanism, determine the number of partitions H and the number of objective functions M , and use the Das - Dennis method to generate reference points N , N = ( H+M - 1) + ( M - 1), generate all non - negative integer combinations that satisfy x 1 + x 2 +…+ x M = H for each objective axis, normalize the non - negative integer combinations, map them to the objective space, and then use the perpendicular distance to associate the objective function based on the dynamic multi - objective decomposition mechanism to the nearest reference point. Then, according to the current solution set distribution density, increase the reference points in the sparse areas and reduce the reference points in the dense areas, and adjust the individual fitness evaluation to preferentially select the solutions that satisfy all constraints; Step S23b: Construct a local development population and design a mutation factor U = 0.5×(1 + rand), U ∈[0.5, 1], and adopt the DE / current-to-pbest / 1 strategy to balance the convergence speed and diversity; V i = X i + U ( X pbest - X i ) + UX r1 - X r2 ); In the formula: V i is the generated mutation vector; X i is the i th individual in the current population; X pbest is an individual randomly selected from the top p % of the elite individuals in the population; X r1 , X r2 are two different individuals randomly selected from the entire population ( r 1 ≠ r 2 ≠ i ); U is the mutation factor; And in the crossover stage, preferentially select the variable components that satisfy the constraints in the parent generation to reduce the probability of generating infeasible solutions; Step S23c: Select the top 20% of the non-dominated sorted solutions from the global population as candidate solution A, and select the top 20% of the fitness from the local population as candidate solution B to form a candidate pool; Step S23d: Calculate the objective space cosine similarity between candidate solution A and candidate solution B in the candidate pool and the individuals in the target population. According to the information interaction mechanism, select the individual with the lowest similarity in the target population for each candidate solution to replace, ensuring the introduction of diversity. If the minimum similarity between the candidate solution and the target population exceeds 0.9, then abandon this migration; Step S23e: When the solution sets of the two populations no longer change significantly for multiple consecutive generations, then it converges to obtain the solution set of the multi-objective optimization problem. Otherwise, go back to step S23a for iterative calculation.

[0033] In this embodiment, an initial reference point is generated based on the Das-Dennis method, reference points are added according to the sparse region of the solution set density to adapt to the shape of the solution set, and the DE / current-to-pbest / 1 strategy (differential evolution algorithm) is adopted. The first 20% of the elite individuals are used to guide the mutation direction, which is beneficial to accelerating the convergence speed. According to the information interaction mechanism, the cosine similarity of the target space between the candidate solution and the individuals of the target population is calculated, and the individual with the lowest similarity is selected for replacement, effectively supplementing the blank area of the solution set.

[0034] According to one aspect of the present application, step S3 is further as follows: Step S31: The solution set is divided into 7 categories: Category 1 is the minimum point of objective 1; Category 2 is the minimum point of objective 2; Category 3 is the minimum point of objective 3; Category 4 is the solution where objectives 1, 2, and 3 are balanced; Category 5 is the solution beneficial to objective 1; Category 6 is the solution beneficial to objective 2; Category 7 is the solution beneficial to objective 3.

[0035] Step S32: The hierarchical clustering method based on Euclidean distance is used to select representative solution sets for Categories 5, 6, and 7. For each category, a set of C initial solutions is set, the Euclidean distance between each solution in this category is calculated, and a Euclidean distance matrix formed by all solutions in this category is generated D , where D i，j is the Euclidean distance of the objective function between solution x i and y i . The number of representative solutions for this category is set C’ , a decision tree is generated, and the solutions in this category are clustered into C’ ones. In each solution cluster, a solution with the shortest Euclidean distance to other solutions is selected as the representative solution set; Step S33: The objective functions of each solution in the solution set of Category 4 are averaged, the averages are sorted from largest to smallest, and the frequency is calculated using the expected value formula. The solutions with the difference between the frequency and 50% less than 0.01 form the representative solution set of Category 4; Step S34: The representative solution sets of Categories 1, 2, 3, and Categories 4, 5, 6, 7 are integrated to obtain the filtered solution set; Step S35: Extract the objective function values, reservoir storage at each time step, and optimized reservoir outflow from the filtered solution set. Based on the optimized reservoir outflow in the filtered solution set, use the K-means clustering algorithm with a coupled adaptive sliding time window to classify the optimized reservoir outflow into low-flow processes and high-flow processes. Then, based on the low-flow processes and high-flow processes, 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.

[0036] In this embodiment, combining the hierarchical clustering method and the physical meaning of the solution set frontier, a comprehensive screening method is proposed to screen out a limited number of representative solutions from the solution set frontier as the research object, enhancing the pertinence and focus of learning, and solving the problem that the number of solution set schemes is large, the differences between schemes are small, and the similarity is high. Directly using all schemes as input data for machine learning will result in too high computational costs and low efficiency.

[0037] According to one aspect of the present application, step S35 is further as follows: Step S35a: Based on the optimized reservoir outflow in the filtered solution set, plot the optimized reservoir outflow in series as the optimized reservoir outflow process curve Q(t), and set the scheduling period of the reservoir as T. Step S35b: Calculate the standard deviation σ of the slope S(t) at each moment of the optimized reservoir outflow process curve to reflect the magnitude of local fluctuations. When σ is large, it indicates that the data fluctuates violently during this period, and vice versa, the fluctuations are relatively stable. Step S35c: Use the method of continuous difference to calculate the local average slope. First, set a dynamic sliding window according to the seasonal variation of 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 indicates that the optimized reservoir outflow process within this sliding time window has an obvious continuous upward or downward trend.

[0038] Step S35d: Divide the optimized outflow process of the reservoir within the sliding time window into a low-flow process candidate set and a high-flow process candidate set. If the standard deviation σ of the slope within the window exceeds the preset threshold ξ, it indicates that the data fluctuates violently during this 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 into the low-flow process candidate set; if the average slope S is significantly positive or negative, it indicates that the data shows a continuous upward or downward trend as a whole. At this time, even if there are certain fluctuations, as long as σ does not exceed the preset threshold ξ, it is classified into the high-flow process candidate set. Step S35e: Use the means of the low-flow process candidate set and the high-flow process candidate set obtained in step S35d as the initial center points, and perform clustering using the K-means clustering algorithm to obtain the low-flow process and the high-flow process.

[0039] In this embodiment, according to the characteristics of large fluctuations in the low-flow process of the reservoir's optimized outflow and clear change trends in the high-flow process, the K-means clustering algorithm is improved, solving the problem that the traditional K-means clustering results do not conform to the seasonal and periodic fluctuations of reservoir operation, and will classify the outflows in the wet season into the high-flow process and the outflows in the dry season into the low-flow process.

[0040] According to one aspect of the present application, step S4 is further as follows: Step S41: Introduce a gated recurrent unit (GRU) model. Based on the optimized outflow of the reservoir, train the GRU model to obtain the time series characteristics of the low-flow process. Add the time series characteristics of the low-flow process to the model samples to obtain new samples. Based on the new samples, build a GRU-XGBoost model, and use the GRU-XGBoost model to extract the reservoir flood control scheduling rules for the low-flow process to obtain the rule-based outflow of the low-flow process. Step S42: Based on the model samples and the water level-discharge capacity curve, build an LSTM model coupled with physical constraints, and use the LSTM model coupled with physical constraints to extract the reservoir flood control scheduling rules for the high-flow process to obtain the rule-based outflow of the high-flow process. Step S43: Arrange the rule-based outflows of the low-flow process and the high-flow process in chronological order to obtain the rule-based outflow of the reservoir. Define evaluation indicators such as water balance error, peak error, percentage of over-discharge capacity constraint, and target value error. Based on the optimized outflow of the reservoir and the rule-based outflow of the reservoir, use the root mean square error, coefficient of determination, and the above-defined evaluation indicators to verify the output of the model.

[0041] According to one aspect of the present application, step S41 is further as follows: Step S41a: Introduce a gated recurrent unit (GRU) model. Based on the optimized outflow of the reservoir, train the GRU model to obtain the time series characteristics of the low-flow process. Step S41b: Incorporate the low-flow process time series characteristics into the model samples to obtain new samples. Based on the new samples, construct a GRU-XGBoost prediction model, define the objective function, optimize the objective function using second-order Taylor expansion, and control the complexity through regularization terms; Step S41c: Divide the new samples into a training set and a validation set. The measured reservoir inflow, the inflow between each reservoir outflow section and the flood control section, the reservoir storage, the remaining predicted reservoir inflow, the objective function value, and the low-flow process time series characteristics are used as the independent variable factors of the GRU-XGBoost model, and the optimized reservoir outflow and the remaining predicted reservoir outflow are used as the dependent variable factors of the GRU-XGBoost model. Train the model using the parallel tree generation algorithm, perform feature splitting using the greedy algorithm, introduce the Shrinkage learning rate to control overfitting, and establish a non-linear mapping relationship between the independent variable factors and the dependent variable factors; Step S41d: Input the independent variable factors during the validation period into the model to obtain the prediction results of the corresponding dependent variable factors. Extract the regular outflow of the predicted low-flow process from the prediction results, and calculate the fitting error based on the regular outflow of the predicted low-flow process and the low-flow process in the optimized reservoir outflow; Step S41e: Based on the fitting error, perform secondary training on the GRU model to obtain the predicted fitting error. Based on the regular outflow of the predicted low-flow process and the predicted fitting error, obtain the regular outflow of the low-flow process, where the regular outflow of the low-flow process is the sum of the regular outflow of the predicted low-flow process and the predicted fitting error.

[0042] In this embodiment, a two-stage modeling method is used. By utilizing the time series characteristics extracted by the GRU model and the error correction mechanism, the XGBoost model is improved, solving the problem that the XGBoost model cannot directly establish long-term dependencies in time series, enabling the model to obtain more accurate prediction results.

[0043] According to one aspect of the present application, Step S41c is further as follows: Step S41c1: Based on the independent variable factors of the GRU-XGBoost model, construct decision trees using the gradient boosting algorithm, generate a candidate split threshold set using the quantile discretization method, and based on the candidate split threshold set, use multi-threading technology to parallelly calculate the split gain of each independent variable factor; 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; In the formula: G L and G R are the sum of the first-order gradients (the first derivative of the loss function) of the left and right child node samples; H L and H R are the sum of the second-order gradients (the second derivative of the loss function) of the left and right child node samples; λ is the regularization coefficient; γ is the split gain threshold parameter; Step S41c2: Traverse all independent variable factors and their split gains of the current node, use the greedy algorithm to iteratively select the optimal split point, set the maximum 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 to stop the iteration; Step S41c3: Set the learning rate parameter. When updating the predicted value of each newly added tree, multiply it by the learning rate parameter, and by reducing the contribution of a single tree, make the model more dependent on the weighted results of multiple trees. At the same time, combine the regularization term to prevent overfitting; Step S41c4: Integrate the predicted results of each tree to obtain the non-linear mapping relationship between the independent variable factor and the dependent variable factor.

[0044] In this embodiment, by reducing the split threshold search space through quantile discretization and combining multi-threaded parallel computing of the split gains of each independent variable factor, the model training is significantly accelerated. Based on the 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 global search. γ in the formula is used as the split gain threshold, which can actively control the branch growth of the tree, prevent the generation of redundant subtrees, and reduce memory consumption.

[0045] According to one aspect of the present application, step S42 is further as follows: Step S42a: Based on the model samples and the water level-discharge capacity curve, construct a loss function, and build an LSTM model with coupled physical constraints. The LSTM model includes a forget gate, an input gate, a long-term state unit of the current input, a long-term state unit of 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.

[0046] Step S42b: Divide the model samples into a training period and a validation period. The measured reservoir inflow, the inflow between each reservoir outflow section and the flood control section, the reservoir storage, the remaining predicted reservoir inflow, and the objective function value are used as independent variable factors of the LSTM model with coupled physical constraints. The optimized reservoir outflow and the remaining predicted reservoir outflow are used as dependent variable factors of the LSTM model with coupled physical constraints. Input the independent variable factors and dependent variable factors in the training period into the LSTM model with coupled physical constraints for training to obtain the mapping relationship between the independent variable factors and the dependent variable factors. Step S42c: Input the independent variable factors in the validation period into the model to calculate the predicted results of the corresponding dependent variable factors, that is, obtain the regular outflow of the high-flow process.

[0047] According to one aspect of the present application, step S42a is further as follows: Step S42a1: Read 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 reservoir storage, the water level-storage curve and the water level-discharge curve data of the reservoir in the model samples. Based on the above reservoir storage, water level-storage curve and water level-discharge curve data, obtain the reservoir discharge capacity. Step S42a2: Set parameters y 1 , parameter y 2 , parameter y 3 , parameter y 4 , parameter y 5 , where parameter y 1 is the parameter to be assigned to the value of the regular reservoir outflow predicted in step S42b, parameter y 2 is the parameter to be assigned to the reservoir discharge capacity, parameter y 3 is the maximum value of parameter y 1 , parameter y 4 is the maximum value of the optimized reservoir outflow, parameter y 5 is the parameter to be assigned to the predicted value of the remaining predicted reservoir outflow predicted in step S42b; Step S42a3: Based on the measured reservoir inflow, parameter y 1 , the remaining predicted reservoir inflow, parameter y 5 , the end-of-period reservoir storage and the reservoir storage, calculate the water balance constraint penalty term according to the water balance principle. 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 ; In the formula: loss 1 i is the penalty term for the water balance constraint of the reservoir i ; M is the number of samples extracted for each training in the training set; Q i,t is the measured inflow of the reservoir; Δ t is the unit time interval; WU i,t is the remaining predicted inflow of the reservoir; V i,T+1 is the end - of - period storage of the reservoir; V i,t is the storage of the reservoir; Step S42a4, based on the optimized outflow of the reservoir, parameter y 3 and parameter y 4 , calculate the weight coefficient that changes according to the size of the reservoir outflow. Multiply this weight coefficient by the difference between the optimized outflow of the reservoir and parameter y 1 to obtain the high - flow weighted penalty term; loss 2 i =[1 / M ×∑( QO i,t / y 4 ×( QO i,t - y 1t )) 2 1 / 2 t ∈[1, M ; In the formula: loss 2​​i is the reservoir i 's high-flow weighted penalty term; QO i,t is the optimized outflow of the reservoir; Step S42a5, compare parameter y 1 with parameter y 2 If parameter y 1 is greater than parameter y 2 , then calculate the root mean square error of the two to obtain the discharge capacity constraint penalty term. If parameter y 1 is less than parameter y 2 , then the discharge capacity constraint penalty term is 0; loss 3 i = 1 / M × ∑ max{( y 1t - y 2t ), 0} 2 t ∈ [1, M ; In the formula: loss 3 i is the discharge capacity constraint penalty term of the reservoir i ; Step S42a6, compare parameter y 1 with 0, multiply the absolute value of the smaller value between the two by a large value O to obtain the non-negativity constraint penalty term; loss 4 i = 1 / M × ∑ min{ y 1t , 0} × O t ∈ [1, M ; In the formula: loss 4 i is the non-negativity constraint penalty term of the reservoir i ; Step S42a7, extract the loss function of the LSTM model, add the water balance constraint penalty term, high-flow weighted penalty term, discharge capacity constraint penalty term and non-negativity constraint penalty term to the loss function of the LSTM model to form a new loss function, and establish an LSTM model with coupled physical constraints.

[0048] In this embodiment, to improve the applicability and accuracy of the LSTM model in the extraction application of high-flow process scheduling rules, aiming at the problem of insufficient physical interpretability of traditional data-driven models, a water balance constraint, high-flow weighting, discharge capacity constraint, and non-negative constraint penalty term are added to the loss function, effectively restricting the solution set space of the LSTM model when solving practical problems and ensuring that the output of the model is closer to the actual scheduling rules.

[0049] In summary, by applying the hierarchical reservoir flood control scheduling rule extraction method based on the improved XGBoost-LSTM proposed in this application, a multi-objective solution set that conforms to the stage characteristics of reservoir flood control scheduling can be obtained, scientifically and reasonably dividing the reservoir optimized outflow into low-flow processes and high-flow processes, improving the XGBoost and LSTM models, and organically coupling the improved models to obtain the reservoir regular outflow, achieving the balance between computing resources and accuracy, and quickly providing an efficient and feasible reference scheme for reservoir flood control scheduling.

[0050] According to one aspect of the present application, the specific output of the present invention is the reservoir regular outflow.

[0051] The technical idea of the present invention is roughly as follows: To solve the problem of multi-objective flood control scheduling of reservoir groups, a multi-objective model is constructed to solve the multi-objective flood control scheduling problem. However, the calculation time is long and it is difficult to be applied to the actual flood control scheduling decision-making scenario. And this problem directly affects the high timeliness of flood control decision-making. Therefore, first, an attempt is made to introduce LSTM to fit the outflow process of the reservoir group of the existing multi-objective optimized flood control scheduling scheme, extract the flood control scheduling rules of the reservoir group, and successfully extract the flood control scheduling rules of the reservoir group, with the determination coefficient reaching 0.8.

[0052] However, it is found in the research that LSTM also has the following problems: The fitted outflow does not meet the actual physical constraint conditions and has insufficient physical interpretability; the input data volume is large and the calculation is complex. To solve these problems, the following solutions are provided: Physical mechanism constraints are introduced into the loss function of the LSTM model, including water balance constraint, high-flow process constraint, discharge capacity constraint, and non-negative constraint; the K-means clustering algorithm is used to divide the reservoir outflow into low-flow processes and high-flow processes; the improved LSTM model is used to fit the high-flow process, and the XGBoost model is introduced to fit the low-flow process, and a coupled model is used to improve the calculation efficiency.

[0053] However, in the actual use process, if the traditional K-means clustering algorithm is used to preprocess the outflow of the reservoir group, the following problems occur: In some periods, it is the flood season, while in some periods, it is the dry season. The reservoir scheduling has different rules under different seasons or periodic changes. The traditional K-means clustering results do not conform to the seasonal and periodic fluctuations of the reservoir scheduling, and will classify the outflows in the flood season as high-flow processes and the outflows in the dry season as low-flow processes.

[0054] Therefore, the following improvements are made to the K-means clustering algorithm: Calculate the standard deviation of the slopes at each moment of the reservoir optimized outflow process curve; Calculate the local average slope to judge whether there is an obvious continuous upward or downward trend in the reservoir optimized outflow process within the dynamically sliding window set according to the seasonal changes of the reservoir scheduling; According to the calculated standard deviation of the slopes and the average slope within the window, divide the reservoir optimized outflow process curve into a low-flow process candidate set and a high-flow process candidate set, take the mean value of the candidate set as the initial center point, and use the K-means clustering algorithm for clustering to obtain the low-flow process and the high-flow process.

[0055] At the same time, in the actual process, when using the XGBoost model to fit the low-flow process, the following problems occur: The low-flow process of the reservoir has continuity and time characteristics, but XGBoost, as a decision tree ensemble method, does not have time dependence, and the predicted values show large fluctuations and are not smooth enough.

[0056] Therefore, the following improvements are made: Introduce the gated recurrent unit GRU. First, train GRU to extract the time series characteristics of the low-flow process, use this feature and other factors as the input of XGBoost, and use the XGBoost model to fit the low-flow process of the reservoir outflow; Calculate the fitting error of XGBoost, and then retrain GRU to learn the time series change trend of the error, and correct the fitting value of XGBoost according to the error with time series characteristics.

[0057] In this embodiment, the regular outflow of the reservoir can be obtained through all the above steps, which can quickly provide an effective scheduling reference plan for the reservoir flood control scheduling, and has high technical innovation and practicality. The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept scope of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A method for extracting flood control dispatching rules for hierarchical reservoirs 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 a low-flow process and a 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; 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.

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 to minimize the maximum flood control storage capacity utilization rate of the reservoir, minimize the maximum average excess flood volume of the system, minimize 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, and the solution set of the multi-objective optimization problem is obtained.

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, 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; The target space cosine similarity between the candidate solutions in the candidate pool and the target population individuals is calculated sequentially, and the individual with the lowest similarity is selected for replacement according to 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 reservoir optimized outflow and clustering, including: The solution set is divided into 7 categories, including the minimum points of objectives 1, 2, and 3, the equilibrium solutions of the three objectives, and the solutions that are beneficial to objectives 1, 2, and 3; A hierarchical clustering method based on Euclidean distance is used to select representative solution sets for 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 screened solution set; The objective function value, reservoir storage and optimized outflow of the reservoir are extracted from the screened 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 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 into the optimized outflow process curve Q(t), and the dispatch 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. The dynamic sliding time window W is set according to the comparison result of |S(t)| and σ. When |S(t)|<σ, W∈[t-α·T, t+α·T], and when |S(t)|>σ, W∈[t-β·T, t+β·T], where β<α<1, and the local average slope S in the window is calculated. Based on the comparison between the slope standard deviation σ in the window and the preset threshold ξ, as well as the value of the average slope S, the reservoir optimal 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 taken 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-flow processes, including: The GRU model is introduced and trained based on the reservoir outflow optimization to obtain the time series characteristics of the low flow process; The time series characteristics of low-flow processes are added to the model samples, and the GRU-XGBoost prediction model is constructed and preliminarily trained. The second-order Taylor expansion is used to optimize the objective function, and the complexity is 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 processes 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.

8. The method according to claim 1, characterized in that The specific steps for extracting dispatch rules for high-flow processes include: Based on the model samples and the water level-discharge capacity curve, a loss function including root mean square error penalty, water balance constraint penalty, high flow weighted penalty, discharge capacity constraint penalty and non-negative constraint penalty is constructed, and an LSTM model coupled with physical constraints is 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.

9. The method according to claim 8, characterized in that Water balance constraint penalty 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 into 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 in reservoir i, QO i,t Optimizing outflows 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 remaining forecast outflow value of the reservoir.

10. The method according to claim 9, characterized in that 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 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 discharge capacity of the reservoir.

Images