Cross-basin scheduling method based on knowledge driving and time sequence progressive constraint processing

Abstract

The application discloses a cross-basin scheduling method based on knowledge driving and time sequence progressive constraint processing, and comprises the following steps: acquiring hydraulic connection topology and hydrological supply-demand boundary data, and quantitatively generating a time-varying constraint state sequence representing the optimization difficulty of each scheduling period; according to the constraint processing strategy, an optimal target scheduling scheme meeting preset engineering limitations is searched in a time sequence progressive evolution solving process; at a stage conversion node of the time sequence progressive evolution, historical characteristic mapping is performed to construct an initial population of the next stage, and the final converged target scheduling scheme is converted into actual regulation instructions of the reservoir group. By introducing continuous difficulty quantitative indexes and a structured knowledge migration mechanism, the application effectively solves the problems of constraint processing rigidity and blind search in the prior art, and improves the safety, convergence speed and target benefit of the cross-basin reservoir scheduling scheme.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of inter-basin reservoir scheduling, and particularly relates to an inter-basin scheduling method based on knowledge driving and time sequence progressive constraint processing. BACKGROUND

[0002] With the advancement of various water diversion and supply projects in China, the demand for inter-basin water resource management is increasing. Under the complex demands of multiple projects, long time periods and multiple targets, how to improve the response speed and effect through scientific scheduling has great scientific significance and practical application value.

[0003] The existing inter-basin reservoir scheduling methods mainly include traditional optimization methods based on dynamic programming, machine learning methods based on reinforcement learning, and heuristic algorithm methods based on particle swarm optimization. These methods have their own advantages in dealing with scheduling problems of different scales. For example, traditional methods are suitable for small and medium-sized problems, heuristic algorithms have advantages in dealing with nonlinear and multi-objective problems, and machine learning methods focus on simulating the interaction of agents in a dynamic environment.

[0004] However, the existing technology still faces severe challenges in dealing with large-scale complex constraints, mainly in the quantitative lack of optimization difficulty, the rigidity of constraint processing mechanism and the superficialization of historical experience utilization. Specifically, the existing scheme often lacks continuous quantitative evaluation of the optimization difficulty of the scheduling period (such as supply and demand tension or risk level), which leads to the inability of the algorithm to identify key bottleneck periods. In addition, the use of unified or simple binary constraint processing mechanism makes it difficult to achieve fine balance between different risk levels (such as flood control safety and ecological water supply), which easily leads the algorithm to fall into a dead loop or converge to an infeasible solution under strong constraints. Moreover, the use of historical data is mostly limited to simple point-to-point prediction or random initialization, lacking structured extraction and migration of historical good scheduling patterns, resulting in large search blindness and low computational efficiency in long-term optimization. SUMMARY

[0005] The application aims to provide an inter-basin scheduling method based on knowledge driving and time sequence progressive constraint processing to solve the above problems existing in the prior art.

[0006] The technical scheme of the application is an inter-basin scheduling method based on knowledge driving and time sequence progressive constraint processing, which comprises:

[0007] Based on the obtained water power connection topology and hydrological supply and demand boundary data of the inter-basin reservoir group, a time-varying constraint state sequence representing the optimization difficulty of each scheduling period is quantitatively generated;

[0008] The time-varying constraint state sequence is used to dynamically regulate the constraint processing strategy, and in the time sequence progressive evolution solving process, the target scheduling scheme that meets the preset project limit is optimized stage by stage;

[0009] At the stage transition node of the time sequence progressive evolution, the historical feature mapping is performed to construct an initial population of the next stage, and the finally converged target scheduling scheme is converted into actual regulation instructions of the reservoir group.

[0010] Beneficial effects, the present application effectively solves the problems of constraint processing rigidity and search blindness in the prior art by introducing continuous difficulty quantization indicators and a structured knowledge transfer mechanism, and improves the safety, convergence speed and target benefit of the cross-basin reservoir scheduling scheme. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 A step flowchart of a cross-basin scheduling method based on knowledge driving and time sequence progressive constraint processing provided for an embodiment of the present application.

[0012] Figure 2 A step flowchart of quantitatively generating a time-varying constraint state sequence provided for an embodiment of the present application.

[0013] Figure 3 A step flowchart of executing weighted constraint control based on risk stratification provided for an embodiment of the present application.

[0014] Figure 4 A step flowchart of executing distributed learning and differential prediction provided for an embodiment of the present application. DETAILED DESCRIPTION

[0015] In order to enable personnel in the art to better understand the present application scheme, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0016] It should be noted that the terms include and have as well as any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to the clearly listed steps or units, but can include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0017] As shown in Figure 1 A cross-basin scheduling method based on knowledge driving and time sequence progressive constraint processing includes the following steps:

[0018] Based on the obtained hydraulic connection topology and hydrological supply and demand boundary data of the cross-basin reservoir group, a time-varying constraint state sequence representing the optimization difficulty of each scheduling period is quantitatively generated.

[0019] In other words, the hydraulic connection topology and hydrological supply-demand boundary data of the inter-basin reservoir group are acquired, preprocessed, and then the preprocessed hydraulic connection topology and hydrological supply-demand boundary data are obtained; and based on the preprocessed hydraulic connection topology and hydrological supply-demand boundary data, a time-varying constraint state sequence representing the optimization difficulty of each scheduling period is quantitatively generated.

[0020] In this embodiment, the inter-basin reservoir group is a hydraulic system including multiple reservoirs, water diversion projects, and water receiving areas. The hydraulic connection topology specifically refers to physical structure information such as the upstream and downstream relationship between reservoirs, the connection relationship between the starting point and the ending point of the water diversion line, etc. For example, in a certain actual project, the hydraulic system can include the Three Gorges Reservoir, the Danjiangkou Reservoir, and the Jiangjiang Hanbu Project connecting the two, as well as the South-to-North Water Diversion Midline Project for water supply from Danjiangkou to the north. The hydrological supply-demand boundary data refers to the basic input sequence driving the system operation, mainly including the inflow of each reservoir, the water demand of each water receiving area, the design water delivery capacity of each water diversion project, and the reservoir capacity constraint boundary of each reservoir. These data are usually long time series, such as ten-day or monthly historical data covering the past 50 years.

[0021] Further, in order to ensure the consistency and accuracy of subsequent calculations, the original data collected is standardized and preprocessed before quantitative generation is performed. Specifically, for possible data missing or outliers, the system uses linear interpolation to complete, ensuring the continuity of the time series. At the same time, different types of data are uniformly converted into flow volume form, and the physical unit is unified to cubic meters per second or the total water volume of the corresponding period, and the time scale of all data is unified to ten-day or monthly sequences.

[0022] On this basis, the time-varying constraint state sequence is quantitatively generated to evaluate the system operation pressure. The current inflow is analyzed for each period to determine whether it can meet all water demand and water diversion requirements, and the optimization difficulty of the period is judged in combination with the engineering constraints. The generated time-varying constraint state sequence is a sequence of indicators corresponding to the scheduling period. The indicators can be discrete classification labels, such as easy optimization or difficult optimization, or continuous numerical scores for accurately quantifying the urgency of the current period. This provides a core basis for subsequent dynamic strategy adjustment.

[0023] The time-varying constraint state sequence is used to dynamically regulate the constraint processing strategy, and in the time sequence progressive evolution solving process, the target scheduling scheme that meets the preset engineering limit is optimized stage by stage.

[0024] In this embodiment, the time-sequential progressive evolution represents a phased optimization solution framework. Unlike traditional one-time optimization of the entire period, the entire scheduling period is divided into several stages, or the determination of decision variables is made one period at a time as time goes on. It can effectively reduce the dimension of large-scale problems and avoid the curse of dimensionality. The constraint handling strategy refers to how to deal with solutions that violate physical limits in the optimization algorithm, including penalty function method, feasibility rules, etc., while dynamic regulation means that the constraint handling strategy is not fixed, but is adjusted in real time according to the time-varying constraint state sequence.

[0025] Specifically, when the time-varying constraint state sequence indicates that the current period optimization difficulty is low, the system can adopt a more relaxed constraint handling strategy, such as allowing a certain degree of constraint violation to exchange for a wider search space, or simplifying the calculation process to improve speed. Conversely, when it indicates that the optimization difficulty is high, the system will automatically switch to a strict constraint handling strategy, forcing the solution to meet all key constraints, or increasing computing resources to ensure that a feasible solution is found. This makes the algorithm efficient and robust in different situations. Exemplarily, the target scheduling scheme refers to the final output of the optimal decision variable sequence, mainly including the discharge flow of each reservoir in each period, the water transfer amount of each water transfer project, and the operation water level of each reservoir, etc. Meeting engineering constraints means that the target scheduling scheme must strictly comply with the physical constraints such as the upper and lower limits of the water level of the reservoir, the ecological red line of the discharge flow, and the upper limit of the water transfer capacity of the water transfer project. The optimization process is optimized by means of evolutionary algorithms such as differential evolution algorithm or genetic algorithm, which iteratively reproduce the population and constantly approach the global optimal solution.

[0026] At the stage transition node of the time-sequential progressive evolution, the historical feature mapping is performed to construct the initial population of the next stage, and the final converged target scheduling scheme is converted into the actual regulation and control instructions of the reservoir group.

[0027] Specifically, the phase transition node refers to the time when the optimization process transitions from the current optimization phase to the next optimization phase. For example, when the optimization of the first t time periods is completed, the optimization of the t+1 time period is ready to start, which triggers the phase transition. The historical feature mapping uses existing historical optimization experience or knowledge base to guide the initialization of the new phase; this is not a simple random initialization, but through the analysis of the system state characteristics of the current phase, similar scenarios are found in the historical data, and the successful experience in history is transferred to the current problem. Building the initial population of the next phase is to convert the high-quality solutions or knowledge mapped from the history into the initial individuals of the new round of evolutionary algorithm, which can provide a high-quality starting point for the algorithm, avoid starting from zero, and thus shorten the convergence time and improve the solution quality. The final converged target scheduling scheme will be converted into specific control instructions, such as sending specific opening degree instructions to the reservoir gate control system or sending start-stop and flow instructions to the pump station, to realize the actual control of the physical reservoir group.

[0028] As shown in Figure 2 In one possible implementation, the time-varying constraint state sequence is specifically a scheduling difficulty score sequence; the time-varying constraint state sequence is quantified to generate, including:

[0029] Based on the hydrological supply and demand boundary data, the hydrological supply and demand difficulty component representing the degree of approximation of the supply and demand gap to the reservoir capacity boundary is calculated.

[0030] In this embodiment, the hydrological supply and demand difficulty component can quantify the objective optimization pressure brought by natural hydrological conditions and rigid supply and demand relationship. Exemplarily, the hydrological supply and demand difficulty component is composed of three parts: supply and demand contradiction, dead water level risk, and full reservoir risk. Specifically, for each reservoir i and each time period t, the basic supply and demand gap term is calculated. If the sum of the water demand and the designed water transfer capacity in the current time period exceeds the inflow, a positive difficulty value is generated, and the larger the gap, the higher the difficulty value. Reflects the direct pressure brought by water resource shortage. At the same time, the approximation degree of the reservoir capacity boundary is calculated. On the one hand, the dead water level approximation risk is concerned, that is, when the reservoir capacity at the end of the last time period approaches the minimum reservoir capacity corresponding to the dead water level, it means that the available regulation reservoir capacity is very small, and the water supply guarantee rate is threatened, so a corresponding difficulty penalty is generated. On the other hand, the full reservoir risk is concerned, that is, when the reservoir capacity approaches the maximum reservoir capacity corresponding to the normal high water level or the flood control limit water level, it means that the flood control situation is severe, and the discharge must be increased, which may conflict with the water transfer target, and also generates a difficulty penalty.

[0031] Optionally, the hydrological supply and demand difficulty component D hydro can be represented as the sum of three weighted terms: D hydro_i_t =α1 * max(0,D demand_i_t -Q in_i_t +Q transmax_i_t) + a2*max(0, V min_i -V i_tprev ) + a3*max(0, V i_tprev -V max_i ); wherein a1, a2, a3 are preset positive weight parameters, respectively corresponding to the attention degree of the supply-demand gap, the dead water level risk and the full reservoir risk; D demand_i_t is the water demand, Q in_i_t is the inflow, Q transmax_i_t is the designed water transfer capacity, V min_i and V max_i are the minimum and maximum reservoir capacities respectively, V i_tprev is the reservoir capacity at the end of the previous period. Preferably, the designed water transfer capacity and the upper and lower limits of the reservoir capacity can take different values at different periods, so as to reflect the design conditions in different operation stages.

[0032] According to the preset constraint risk level definition, the distribution weights of the high-risk constraints and the low-risk constraints in the current period are counted, and a constraint risk difficulty component is generated.

[0033] Specifically, in order to realize fine risk management, the system pre-establishes a constraint risk level definition table. The definition table divides all physical constraints into different risk levels. For example, the upper limit of water level constraint related to dam safety and dike flood control safety is defined as a high-risk constraint. The constraints related to ecological base flow and general water supply targets are defined as medium or low-risk constraints, and a small amount of adjustment for a short time is allowed in extreme cases. Exemplarily, the constraint risk difficulty component D risk_i_t can be realized by the following linear weighting formula: D risk_i_t = g1*R high_i_t + g2*R low_i_t ; wherein g1, g2 are risk level weight coefficients, and g1 is much larger than g2, so as to reflect the absolute attention to high-risk factors; R high_i_t , R low_i_t respectively represent the number or the corresponding characteristic value of the high-risk constraints and the low-risk constraints effective in the period. Through the constraint risk difficulty component, the system can identify the key period.

[0034] The hydrological supply-demand difficulty component and the constraint risk difficulty component are multi-dimensionally coupled to generate a scheduling difficulty score sequence.

[0035] In the embodiment, multi-dimensional coupling refers to the fusion of the two difficulty components with different sources and different physical meanings into a unified comprehensive index, i.e. the scheduling difficulty score sequence D tThe fusion process preferably adopts a weighted summation manner, but more complex nonlinear mapping is also supported. Through coupling, the system not only considers the objective hydrological physical conditions, but also considers the human-set risk preference, and combines the feedback information of the actual operation of the algorithm, so that the scheduling difficulty is quantified in all directions. Exemplarily, the calculation formula of the scheduling difficulty score sequence D t may be: D t = λ1 * D hydro_t + λ2 * D risk_t ; wherein λ1, λ2 are normalized fusion weights; D hydro_t and D risk_t are the weighted aggregation values of the hydrological supply-demand difficulty component and the constraint risk difficulty component corresponding to each reservoir respectively. The generated scheduling difficulty score sequence D t is a one-dimensional numerical sequence that changes over time, and the higher the value, the greater the comprehensive optimization difficulty of the period, and a more robust or more aggressive strategy needs to be taken.

[0036] As shown in Figure 3 , in a further implementation manner, a time-varying constraint state sequence is used to dynamically regulate the constraint processing strategy, including performing weighted constraint control based on risk stratification, specifically:

[0037] The pre-stored physical constraint conditions of the whole period are decomposed into a high-risk constraint set and a low-risk constraint set.

[0038] Specifically, based on the determined risk level definition, the system clearly cuts and groups all constraint conditions originally mixed together at the mathematical set level. The high-risk constraint set K high contains all constraint items marked as high risk, such as the constraint that the reservoir capacity cannot exceed the flood control limit water level V limit . The low-risk constraint set K low contains the remaining low-risk constraint items, such as the constraint that the discharge flow at a certain time needs to meet the minimum ecological flow Q eco . This provides a structural basis for subsequent stratified processing.

[0039] The scheduling difficulty score sequence is used as a non-uniform amplification coefficient to calculate the high-risk weighted violation degree of each individual under the high-risk constraint set and the low-risk weighted violation degree under the low-risk constraint set, and a stratified weighted violation degree vector is constructed.

[0040] In this embodiment, the non-uniform amplification coefficient refers to using the scheduling difficulty score D t to dynamically adjust the punishment degree of constraint violation. For any optimization individual j, when calculating its constraint violation, the difficulty score is no longer simply summed as a weight. Specifically, for each constraint item k in the high-risk constraint set, its violation degree violj_k_t will be assigned a larger base weight a high and multiplied by a difficulty amplification factor (1 + p * D t ), meaning that the cost of violating the high-risk constraint will be significantly amplified in difficult periods. Exemplarily, the calculation formula of the high-risk weighted violation degree CV high_j is: CV high_j =∑ t ∑ k∈Khigh ( a high *(1 + p * D t )*viol j_k_t ); similarly, the calculation formula of the low-risk weighted violation degree CV low_j is: CV low_j =∑ t ∑ k∈Klow ( a low *(1 + p * D t )*viol j_k_t ); where p is the difficulty sensitivity coefficient; a high is the base weight of the high-risk constraint; a low is the base weight of the low-risk constraint; viol j_k_t is the violation degree of individual j to constraint item k at period t. The finally constructed hierarchical weighted violation degree vector is a vector [CV high_j , CV low_j ] containing two elements, which carries more information about the risk structure than the single violation degree value.

[0041] Based on the hierarchical weighted violation degree vector, hierarchical dominance logic is executed in the environment selection process to preferentially eliminate individuals whose high-risk weighted violation degree is worse than that of competitors, and when the high-risk weighted violation degrees are the same, the dominance relationship is determined according to the low-risk weighted violation degree, so as to screen the preferred population for the next generation evolution.

[0042] Specifically, environment selection is a key link in evolutionary algorithms that determines which individuals can survive to the next generation. The hierarchical dominance logic changes the traditional Pareto dominance or simple constraint dominance rule. When comparing the pros and cons of two individuals j1 and j2, compare their high-risk weighted violation degrees. If CV high_j1 is less than CV high_j2 , it is directly determined that individual j1 is better than j2, regardless of how j2 performs on the low-risk constraint or the objective function. It embodies the principle of safety first, that is, first to ensure that no catastrophic events such as dam collapse occur. Only when the high-risk violation degrees of the two individuals are equal (usually both are 0, that is, both satisfy the high-risk constraint), the system will further compare their low-risk weighted violation degrees. If CV low_j1 is less than CV low_j2Then, j1 is determined to be better than j2. Only when both of them perform consistently on the two-layer risk constraints, their objective function values, such as water shortage rate, are compared. Through the layer-by-layer screening mechanism, the population converges to the direction of meeting the safety bottom line first, then pursues more refined compliance, and finally pursues the maximization of benefit.

[0043] In another possible implementation, the quantification of the time-varying constraint state sequence further comprises: monitoring the proportion of feasible solutions and the average violation degree of the population in the time sequence progressive evolution process, calculating the algorithm convergence difficulty component; and coupling the hydrological supply and demand difficulty component, the constraint risk difficulty component, and the algorithm convergence difficulty component in multiple dimensions to generate a scheduling difficulty score sequence.

[0044] In the embodiment, the algorithm convergence difficulty component D alg is used to reflect the actual encounter of the optimization algorithm in the current search space. It is a dynamic feedback index derived from the iterative process of the algorithm. Specifically, every certain iteration interval, the system will count the proportion of feasible solution individuals in the current population that meet all constraints, denoted as p feas . If the proportion is low, it means that the feasible region is narrow and the search difficulty is great. At the same time, the system also calculates the average constraint violation degree of all infeasible solutions in the population, denoted as mean CV . The greater the violation degree, the farther the distance from the feasible region, and the more difficult the convergence. Exemplarily, the algorithm convergence difficulty component can be constructed as a combination of the negative correlation function of the proportion of feasible solutions and the positive correlation function of the average violation degree: D alg_t_g =η1*(1 -p feas_t_g ) +η2*mean CV_t_g ; wherein η1 and η2 are weight parameters, g represents the current evolution generation number; D alg_t_g is the convergence difficulty degree of the optimization algorithm at evolution generation number g in time period t. In order to obtain more stable evaluation, the D alg_t_g of the recent several generations is usually processed by sliding average, thereby obtaining the smoothed time period algorithm convergence difficulty component.

[0045] In another possible embodiment, the time-varying constraint state sequence is specifically a sliding window constraint tightness sequence; and the quantification of the time-varying constraint state sequence comprises:

[0046] Based on the hydrological supply and demand boundary data, the basic tightness is calculated, which represents the proportion of the single reservoir supply and demand gap in the water transfer capacity.

[0047] In the embodiment, the basic tightness CTI baseThe supply-demand balance state of a single reservoir in a specific period can be quantified. The basic tension is obtained by calculating the difference between the total demand and the total supply, and normalizing it relative to the capacity of the water transfer project. Specifically, for reservoir i in period t, the total demand includes the local basin discharge demand D i_t and the designed water transfer capacity TC i_t of the external water transfer project, and the total supply is the natural inflow I i_t of the reservoir in this period. The calculation formula of the basic tension is: CTI base_i_t = (D i_t + TC i_t - I i_t ) / TC i_t . When CTI base is greater than 0, it means that the natural inflow is not enough to support both discharge and full-load water transfer at the same time, and the reservoir storage or water transfer must be reduced, and the larger the value, the more acute the contradiction; when CTI base is less than or equal to 0, it means that the inflow is abundant, and the system is in a relatively relaxed state. This provides a basis for subsequent complex calculations.

[0048] According to the real-time reservoir capacity state of the reservoir in the storage period or the discharge period, a reservoir capacity state correction coefficient reflecting the reservoir capacity adjustment margin is constructed, and the basic tension is weighted and corrected by using the reservoir capacity state correction coefficient to obtain the single reservoir correction tension.

[0049] In this embodiment, considering that the same water gap means different scheduling difficulties at different reservoir capacity levels, a reservoir capacity state correction coefficient ω cap is introduced. The system identifies whether the current period t is in the storage period or the discharge period of the reservoir. In the storage period, the main task of the reservoir is to raise the water level for later use, and at this time, if the reservoir capacity is low, the storage task is urgent, and any additional water transfer will exacerbate the tense situation. Therefore, the lower the reservoir capacity, the larger the correction coefficient should be. Specifically, if in the storage period, the calculation formula of the correction coefficient is: ω cap_i_t =1+γ*(V max_i_t -V i_t ) / (V max_i_t -V min_i_t ); wherein V i_t is the current reservoir capacity, V max_i_t and V min_i_t are the upper and lower limits of the reservoir capacity in this period, and γ is a regulation coefficient. Conversely, in the discharge period, if the reservoir capacity is high, the flood control pressure is large, and the discharge demand competes with the water transfer capacity for the water transfer channel, so the higher the reservoir capacity, the larger the correction coefficient. At this time, the formula is adjusted to: ω cap_i_t =1+γ*(V i_t -V min_i_t ) / (V max_i_t -Vmin_i_t ); the base tension is multiplied by the coefficient, i.e. CTI i_t = CTI base_i_t *ω cap_i_t , so as to obtain the single reservoir modified tension containing the reservoir storage state information.

[0050] According to the hydraulic connection topology, the system weight of each reservoir is determined, the spatial aggregation of the single reservoir modified tension of each reservoir is performed, and a time decay factor is introduced to perform forward-looking aggregation on the spatial aggregation value in a preset window in the future, so as to generate a sliding window constraint tension sequence.

[0051] Specifically, in order to reflect the overall tension degree of the entire inter-basin system, the tension degrees of each single reservoir need to be spatially aggregated. Different reservoirs have different positions in the system, and the pivotal reservoirs as the water sources of multiple water transfer projects are obviously more critical. Therefore, the system weight λ i is calculated based on the number n i_out of water transfer projects of which the reservoir is the water source: λ i = n i_out / Σ j n j_out ; wherein n i_out is the number of water transfer projects of which the reservoir i is the water source, and n j_out is the number of water transfer projects of which the reservoir j is the water source; the modified tension of each reservoir is weighted and summed by using the system weight, so as to obtain the instantaneous tension at the system level. Further, considering the time lag and continuity of reservoir scheduling, the current decision needs to consider the future pressure. Therefore, the system constructs a sliding window constraint tension SW-CTI. The index not only contains the tension of the current period t, but also aggregates the tension in the future window W t . The window length W t can be dynamically determined according to the total storage capacity of the system: the less the storage, the shorter the forward-looking window, and the more attention to the present survival. Exemplarily, W t = min(W max , ceil((V sys_t -V sys_min ) / Q avg )); wherein W max is the upper limit of the window length, V sys_t is the total storage of the system at the period t, V sys_min is the minimum safe storage of the system, Q avg is the average water supply of the system, and ceil represents the upward rounding operation; in the window, a decay factor β (0<β<1) is introduced to weightedly sum the future tension: SW-CTI t =Σ k=0 Wt-1 (β k *CTIsys_t+k ); where β k is the decay weight of future tension; CTI sys_t+k is the instantaneous tension at time period t+k; the final generated SW-CTI sequence is a smooth and forward-looking continuous indicator, capable of early sensing future drought or flood risks.

[0052] In further embodiments, the time-varying constraint state sequence is utilized to dynamically regulate the constraint handling strategy, including performing tension-based algorithm parameter adaptation, specifically:

[0053] Establishing a nonlinear mapping relationship between the sliding window constraint tension sequence and the algorithm control parameters to generate an adaptive evolution parameter set.

[0054] Specifically, the system utilizes the calculated SW-CTI values to adjust the key control parameters of the evolutionary algorithm in real time, achieving adaptive algorithm behavior. This mapping relationship is usually nonlinear to enhance the algorithm's response capability in extreme cases. The generated adaptive evolution parameter set contains a complete set of parameter configurations for constraint handling, search range, and population diversity maintenance.

[0055] The adaptive evolution parameter set includes: a dynamic relaxation threshold that decays exponentially with tension index, used to control the tolerance of infeasible solutions; a dynamic optimization range that inversely scales with tension, used to determine the length of the time period for current decision variables participating in evolution; and a dynamic elite retention ratio positively associated with tension, used to determine the number of preferred individuals retained during population update.

[0056] In this embodiment, the dynamic adjustment logic of the three core parameters is defined in detail. The dynamic relaxation threshold ε t determines the tolerance of the algorithm to constraint violations. When SW-CTI is high (tension), ε t should be quickly reduced to force the solution to strictly satisfy the constraints; when SW-CTI is low (relaxed), ε t can be larger, allowing exploration. Its mapping relationship is exponential decay: ε t =ε max *exp(-κ*SW-CTI t ); where ε max is the maximum allowed relaxation, and κ is the sensitivity coefficient. In the dynamic optimization range Range t , when the constraint is relaxed, the algorithm can optimize longer future time periods at the same time to pursue global optimality; when the constraint is tight, the range should be contracted to concentrate algorithmic power to solve current feasibility problems. Its mapping relationship is inversely scaled in steps or linearly. In the dynamic elite retention ratio Elite Ratio_tWhen the environment is harsh (stressful), the feasible solution is scarce, and the proportion of elite preservation needs to be increased to prevent the loss of high-quality genes; when the environment is superior, the proportion of preservation can be reduced to increase diversity. The mapping relationship is a positive linear correlation.

[0057] Adaptive evolution parameter sets are injected into the iterative process of time sequence progressive evolution in real time, dynamically adjusting the search behavior of the algorithm.

[0058] Specifically, the calculated parameter set is directly applied to the evolution operation of the current generation. For example, in the environmental selection stage, the current ε t is directly used; in the mutation operation, only the genes within the current Range t are disturbed; and when generating the next generation population, the corresponding number of parent elites is strictly intercepted according to Elite Ratio_t . This makes the algorithm seem to have the ability to perceive the environment and automatically adjust the parameter set according to SW-CTI, thereby efficiently navigating in complex solution space.

[0059] In a preferred implementation, for each individual j, when evaluating its fitness, the original constraint violation degree viol j_k_t is calculated for each time period t and each constraint type k. When the constraint is not violated, it is recorded as 0. According to the preset risk classification, the constraint types are divided into a high-risk constraint set K high and a low-risk constraint set K low (may contain medium and low risk). The scheduling difficulty score D t is used to define the weight coefficient w k (t) for each time period t and each constraint type k: w k (t) =α high * (1 +ρ* D t ), if k∈K high ; w k (t) =α low * (1+ρ* D t ), if k∈K low ; where α high >α low >0 is the basic weight of different risk levels; ρ≥0 is the difficulty sensitivity coefficient. The high-risk constraint violation degree summary value CV high_j and the low-risk constraint violation degree summary value CV low_j are calculated for individual j: CV high_j =Σ t Σ k∈Khigh w k (t)*viol j_k_t ; CV low_j =Σ t Σ k∈Kloww k (t)*viol j_k_t . When the environment is selected, the following comparison rules are defined for two candidate individuals j1, j2: if CV high_j1 < CV high_j2 , then j1 is better than j2; if CV high_j1 = CV high_j2 and CV low_j1 < CV low_j2 , then j1 is better than j2; if the above two are equal, comparison is made according to multi-objective non-dominated sorting and congestion degree criteria. In the presence of ε-CDP, high time-varying relaxation parameter ε high_g and low time-varying relaxation parameter ε low_g may be further introduced: only when CV high_j > ε high_g or CV low_j > ε low_g , is considered as a violation of constraints; ε high_g , ε low_g monotonically decrease with the generation g. The decreasing speed of ε high_g may be set faster than that of ε low_g so as to force the satisfaction of high-risk constraints earlier.

[0060] According to one aspect of the present application, historical feature mapping is performed to construct the initial population of the next stage, including structured retrieval based on the scenario-pattern library, specifically:

[0061] A historical scenario-pattern library containing the mapping relationship between historical scenario feature vectors and low-dimensional scheduling pattern parameters is constructed; multi-dimensional statistical analysis is performed on the hydrological supply-demand boundary data of the current stage to extract the current scenario feature vector containing flow statistical features and constraint difficulty features; the weighted spatial distance between the current scenario feature vector and historical records is calculated in the historical scenario-pattern library to retrieve the target scheduling pattern parameters that best match the current working condition.

[0062] In the present embodiment, the historical scenario-pattern library is a knowledge database that is pre-constructed and continuously updated, and the core unit stored therein is the scenario-pattern pair. In order to achieve efficient storage and retrieval, the system needs to define the structure of the scenario feature vector. The scenario feature vector can accurately depict the hydrological and engineering characteristics of the scheduling period or stage with a small number of numerical features. Specifically, for any time window, the scenario feature vector F scenario may be composed of the following components: the average inflow Q mean , inflow variance Q var , inflow skewness coefficient Q skew , total water demand D total , water demand peak value and its relative position D peak , and the proportion of high-difficulty periods within the window Dhigh_ratio . In which, the determination of high difficulty period can reuse the difficulty score or tension index. The corresponding low-dimensional scheduling mode parameter is the compressed expression of the high-quality historical solution under this scenario. Directly storing high-dimensional time series (such as daily water transfer amount) will occupy a large amount of space and is difficult to generalize, so the system uses dimension reduction technology for encoding. One preferred implementation is to use principal component analysis method to project the long sequence onto the first K principal components, and retain the coefficients as the mode parameter P mode ; another implementation is to divide the time series into several equal length segments, calculate the mean of each segment, and construct a low-dimensional mean vector. In the retrieval stage, the system performs the same statistical analysis on the hydrological supply and demand data of the current optimization stage to generate the current scenario feature vector F new . The similarity between the current scenario and the historical scenario F history in the library is calculated using the weighted spatial distance formula. Exemplarily, the distance calculation formula can be expressed as: dist = sqrt (Σ i (w i * (F new_i - F history_i ) 2 )); Where w i is the weight of the i-th feature dimension, used to balance the influence of different dimension features, for example, the weight of the inflow mean may be set higher than the skewness coefficient. The system traverses the mode library, selects one or more historical records with the smallest distance dist, extracts the corresponding mode parameter P mode as the target scheduling mode parameter.

[0063] In some optional embodiments, in order to improve retrieval efficiency, KD tree or local sensitive hashing technology can be used to index the mode library, so as to quickly locate the nearest neighbor in a large amount of historical data. In addition, time similarity constraints can also be introduced, such as preferentially retrieving records of the same period or adjacent years, to take advantage of the seasonal regularity of hydrology.

[0064] In further embodiments, the scenario-mode library based structured retrieval also includes mode decoding and population generation, specifically:

[0065] Using the pre-configured inverse transformation matrix or basis function, the target scheduling mode parameter is decoded and restored to the full-period main trend scheduling trajectory; taking the main trend scheduling trajectory as the distribution center, introducing random noise to control the variance intensity of the time-varying constraint state sequence, generating an initial population containing a predetermined number of individuals in the next stage, and injecting the initial population in the next stage into the next stage calculation of the time sequence progressive evolution.

[0066] In the present embodiment, the decoding restoration is to reconvert the low-dimensional abstract parameters into high-dimensional scheduling process curves with physical meaning. If principal component analysis encoding is adopted, the decoding process is to multiply the mode parameters P mode by the transpose of the principal component feature vector matrix and add the mean vector if necessary, so as to reconstruct the approximate water release sequence or reservoir capacity sequence; if piecewise mean encoding is adopted, the piecewise mean is expanded to each specific time period by interpolation or step function. The restored sequence is referred to as the principal trend scheduling trajectory Q trend , which represents the scheduling baseline that has been proven to be effective under similar historical scenarios. Preferably, in order to adapt to the current specific minor differences, the system generates an initial population with diversity by adding random noise around the principal trend scheduling trajectory. The strength of the noise is not fixed but is regulated by the time-varying constraint state sequence. Specifically, in periods of high constraint tightness or difficult scheduling, the solution feasible region is narrow, and the system should apply a smaller noise variance to make the generated individuals closely around the historical experience, ensuring feasibility; while in periods of loose constraints, the system applies a larger noise variance to encourage the algorithm to explore a wider solution space.

[0067] Exemplarily, the sampling formula can be expressed as: Q new_i_t = Q trend_i_t + ζ * σ t * randn; wherein Q new_i_t is the decision variable value of the i-th newly generated individual at the t time period, Q trend_i_t is the principal trend value, randn is a standard normal distribution random number, ζ is the overall disturbance coefficient; σ t is the dynamic standard deviation of the time period, which is negatively correlated with the scheduling difficulty score D t or the sliding window tightness SW-CTI of the time period or is obtained through a specific function mapping. The generated initial population, which inherits the historical wisdom and maintains the necessary evolutionary vitality, is directly injected into the next stage of evolutionary calculation as the parent population.

[0068] In another possible implementation, the time-varying constraint state sequence is specifically a binary optimization state sequence; quantifying the generated time-varying constraint state sequence includes:

[0069] For each scheduling time period, the inflow of each reservoir and the sum of the downstream demand and the designed water release capacity are obtained and compared to determine the single-reservoir supply-demand balance state; if the inflow of all reservoirs in the scheduling time period is greater than the sum of the downstream demand and the designed water release capacity, the scheduling time period is marked as an easy-to-optimize period, otherwise it is marked as a difficult-to-optimize period, and a binary optimization state sequence composed of easy-to-optimize period marks and difficult-to-optimize period marks is generated.

[0070] In this embodiment, the generation logic of binary optimization state sequence is intuitive and computationally efficient. For any reservoir in the system, aggregate all its rigid outflow demands in this period, i.e., the minimum downstream discharge demand D req and the design maximum water transfer capacity TC design of all water transfer projects that take this reservoir as the water source. Compare this total demand with the current natural inflow I in . The specific judgment logic is as follows: if for each reservoir i in the system, I in_i is greater than or equal to D req_i plus TC design_i , it means that the water resources are naturally abundant in this period, and all targets can be met without using reservoir storage, and there may even be waste water, so there is no complex supply and demand game, and it is defined as an easy-to-optimize period, i.e., a Rich period. On the contrary, as long as the inflow of any reservoir is less than its total demand, it means that balance must be maintained by consuming reservoir capacity or reducing water transfer, involving complex constraint trade-offs, so it is defined as a difficult-to-optimize period, i.e., a Poor period. The system arranges this series of Boolean judgment results in chronological order to generate a binary sequence composed of Rich and Poor labels, which directly guides the subsequent optimization branch selection.

[0071] In further embodiments, the time-varying constraint state sequence is used to dynamically regulate the constraint handling strategy, including performing phased constraint control, specifically:

[0072] In the process of time sequence gradual evolution, read the label of the current optimization period in the binary optimization state sequence; if the label is a difficult-to-optimize period, perform the full constraint checking strategy and select individuals based on the constraint dominance principle; if the label is an easy-to-optimize period, perform the relaxation-tightening strategy, ignore the physical constraints of the current new period in the early evolution, and introduce a dynamic convergence relaxation threshold after the phase transition condition is met, and select feasible solutions based on the ε constraint dominance principle.

[0073] In this embodiment, for the period marked as Poor, the algorithm adopts a robust conservative strategy. Because the feasible region is small at this time, any solution that violates the constraints can lead to an infeasible actual schedule. Therefore, when the environment is selected, the standard constraint dominance principle (CDP) is directly adopted. That is, the feasible solution is preferred; if none is feasible, the solution with smaller violation degree is selected; if all are feasible, the solution with better objective function is selected. This makes the algorithm strictly adhere to the feasibility bottom line in difficult times. For the period marked as Rich, the algorithm adopts a relaxation-tightening strategy from loose to strict. In the early stage of optimization in the Rich phase, that is, stage a, the algorithm temporarily ignores the constraint conditions of the newly added period, allowing the population to expand rapidly in the infeasible region to find the potential area of the objective function. When the population evolution stagnates or reaches the preset generation, the stage conversion condition is triggered, and stage b is entered. At this time, a dynamic convergence relaxation threshold ε is introduced. The initial value ε0 of ε can be set as the maximum constraint violation degree of the population when entering stage b, and then decays with the increase of the iteration number.

[0074] Exemplarily, the decay formula of ε can be specifically represented as: ε k = ε0*(1 - (k - gs) / (G cur -gs)) cp ; where k is the current cumulative iteration number, gs is the starting iteration number when entering stage b, G cur is the total maximum iteration number allowed in the current period, and cp is a parameter for controlling the decay rate. When the environment is selected, as long as the constraint violation degree of the individual is less than the current relaxation threshold ε k , it is considered as a feasible solution to participate in the dominance ordering. The algorithm allows to fully utilize the loose characteristics in the easy-to-optimize period and jump out of the local optimum.

[0075] As Figure 4 shown, according to another aspect of the present application, the historical feature mapping is performed to construct the initial population of the next stage, including performing solution distribution learning and differential prediction, specifically:

[0076] The mean and variance of the decision variables in the final population of the last stage are counted, a Gaussian distribution model of each dimension is constructed, and a distribution learning sub-population is generated by sampling; the Euclidean distance of the individuals in the current population and the individuals in the last stage population in the objective space is calculated to determine the nearest neighbor correspondence relationship, the difference evolution direction vector of the decision variables is calculated based on the determined nearest neighbor correspondence relationship, and the decision variable sequence is linearly extrapolated and predicted to generate a differential prediction sub-population; the distribution learning sub-population and the differential prediction sub-population are merged to serve as the initial population of the next stage, which is injected into subsequent calculations.

[0077] In this embodiment, to achieve a smooth transition from time period t to time period t+1, two complementary mechanisms are employed to generate the new population. The first is solution distribution learning, assuming that the optimal solution distributions of adjacent time periods have statistical similarity. The system calculates the mean μ and variance σ of the final population in the previous stage across each decision variable dimension. 2 Construct a multidimensional Gaussian distribution N(μ, σ) 2 N / 2 new individuals are randomly sampled from this distribution to form a distribution-learning subpopulation. This ensures that the new population inherits the overall statistical characteristics of the parent generation. The next step is the solution prediction strategy, which utilizes point-to-point evolutionary inertia for extrapolation. For each individual *ind* in the current population, the system searches for the individual *ind* in the previous stage population with the smallest Euclidean distance in the objective function space. prev As corresponding entities, the Euclidean distance is calculated based on the standardized target value. After finding the correspondence, the system calculates the evolutionary direction of the decision variables. For the water diversion variable, the evolutionary direction D1 is directly the sequence of individual ind minus ind. prev The sequence; for storage capacity variables, first calculate their respective time-period storage capacity difference sequences Diff, with the evolutionary direction D2 being the time-period storage capacity difference sequence Diff of the current population individual ind. ind Subtract the ind of the individual corresponding to the previous stage prev Diff of the time-by-time reservoir capacity difference series prev Based on the evolutionary direction, the variable sequence of the new individual is generated through linear extrapolation: Var new = Var ind + r*D; where r is a random number between 0 and 1; Var ind Let be the sequence of decision variables for the current population individual 'ind'; D is the differential evolution direction vector. For the storage capacity sequence, the difference sequence needs to be predicted first, and then the absolute storage capacity value is restored by accumulation. The N / 2 differential prediction subpopulations generated in this way can capture the dynamic trend of the solution's evolution over time. After merging the two, they constitute the initial population for the next stage, which has both statistical stability and dynamic trend.

[0078] In an exemplary embodiment, assuming that during a certain scheduling period t, for the Three Gorges Reservoir (as the key reservoir i), the boundary data collected by the system is as follows: the inflow rate I during this period. in The flow rate is 15,000 cubic meters per second; the minimum downstream discharge demand includes ecological and shipping needs. req The flow rate is 5000 cubic meters per second; the designed water diversion capacity TC of the Yangtze River-to-Han River Water Diversion Project during this period is 1000 cubic meters per second. The current real-time reservoir capacity V... curr The capacity is 20 billion cubic meters, while the minimum allowable reservoir capacity V during this period is... min With a capacity of 15 billion cubic meters and a maximum reservoir capacity of V maxis 393 million cubic meters. According to the basic binary classification logic, the total demand D is calculated = D req + TC = 5000 + 1000 = 6000 cubic meters per second. Comparing the inflow 15000 with the total demand 6000, since 15000 is much greater than 6000, condition I in is met. Assuming that other reservoirs in the system also meet similar conditions, this period is marked as a Rich (easy to optimize) period, and the ε-constraint dominance principle (ε-CDP) strategy will be enabled subsequently. According to the continuous tension index logic, the basic tension CTI base is calculated, CTI base = (5000 + 1000 - 15000) / 1000 = -9.0, indicating that the surplus water is extremely large, 9 times the water transfer capacity. Considering the reservoir capacity correction, assuming that it is currently in the storage period, the correction coefficient ω cap is calculated using the storage period formula. Assuming that the adjustment coefficient γ is 0.5. Then ω cap = 1 + 0.5 * (393 - 200) / (393 - 150) = 1 + 0.5 * 193 / 243 ≈ 1.397. The final single reservoir corrected tension CTI = -9.0 * 1.397 ≈ -12.57. It indicates that the current period is extremely relaxed, and can be greatly explored or aggressively scheduled. Assuming that the average inflow Q mean in the quarter window in which this period is located is 12000 cubic meters per second, and the inflow variance Q var is 2000. The system constructs the feature vector F new = [12000, 2000]. In the historical library, it is found that the scenario F hist = [11500, 2100] is closest to it. Therefore, the system extracts the scheduling mode parameters under this scenario in 1998, and decodes them as the initial solution main trend for the current period, avoiding cold start.

[0079] In one possible embodiment, it further includes constructing a cross-basin scheduling optimization model, specifically:

[0080] Based on the actual engineering needs of the cross-basin reservoir group, a multi-objective optimization mathematical model is constructed, which includes minimizing the water supply deficiency rate and maximizing the downstream water demand guarantee rate.

[0081] Specifically, in order to quantify the pros and cons of the scheduling scheme, the system defines two core optimization objectives. The first objective is to minimize the water demand deficiency rate downstream of the Danjiangkou Reservoir, which includes ecological flow guarantee. The mathematical expression is constructed as follows: calculate the sum of the outflow Q out_dan_t of the Danjiangkou Reservoir at time period t and the water transfer amount Q trans_yb_t of the Yijiang River to Hanjiang River project at time period t, and compare this sum with the downstream water demand Ddown_t A comparison is made. If the sum is less than the water demand, there is a water shortage. Exemplarily, the water shortage rate calculation formula is: F1 = (1 / T) *∑ t=1 T ((D down_t - (Q out_dan_t + Q trans_yb_t )) / D down_t ) 2 ; wherein T is the total number of scheduling periods, and the summation is only performed for the water shortage period, i.e., the square difference is calculated when the water supply is less than the water demand, to punish serious water shortage events. The second objective is to minimize the water supply target loss rate of the South-to-North Water Diversion Middle Route. The mathematical expression is constructed as follows: the difference between the actual water diversion amount Q trans_mid_t of the South-to-North Water Diversion Project at period t and the water diversion target D mid_t of the middle route project. Exemplarily, the water supply target loss rate calculation formula is: F2 = (1 / T) *∑ t=1 T ((D mid_t - Q trans_mid_t ) / D mid_t ) 2 ; similarly, the formula smooths the water supply process through the square term, making the water supply more uniform and avoiding large and small fluctuations.

[0082] All targets are standardized to eliminate dimensional differences, and multi-objective evaluation is performed by weighting or Pareto dominance.

[0083] In this embodiment, since the physical dimensions or numerical ranges of the above two objective functions may differ, in order to facilitate subsequent unified optimization, especially to adapt to the CDP strategy, the system performs standardization processing on all target values. The standardization formula is: f prime_i = (f i - f min ) / (f max - f min ); wherein f i is the original calculated target value, f prime_i is the standardized target value, f max and f min are the maximum and minimum values of the target in the current population or historical record. After processing, all target values are mapped to the interval of 0 to 1.

[0084] A chromosome coding strategy based on real matrix is adopted, and the water diversion amount of the water diversion project and the reservoir capacity are taken as independent decision variables, and other state variables are derived through the water balance equation.

[0085] In this embodiment, in order to adapt to the characteristics of multi-project and long time period of the cross-basin system, the system designs a unique matrix encoding mode. The individual is represented as a matrix, the number of rows of the matrix corresponds to the category of the decision variable, and the number of columns corresponds to the number of scheduling periods T. Exemplarily, the matrix structure is as follows: the first row represents the water transfer sequence of the Jiangjiang Hanjiang project in each period; the second row represents the water transfer sequence of the Middle Route of the South-to-North Water Diversion Project in each period; the third row represents the reservoir capacity sequence of the Danjiangkou Reservoir in each period; and the fourth row represents the reservoir capacity sequence of the Three Gorges Reservoir in each period. When performing cross variation operation, in order to facilitate calculation, the system expands the matrix into a one-dimensional long vector by row. When decoding, the system reads the reservoir capacity sequence and the water transfer sequence, combines the inflow data, and uses the reservoir water balance equation, i.e. the end-of-period storage = initial-of-period storage + period inflow - period outflow - evaporation loss, to reversely calculate the discharge of each reservoir. The core state (reservoir capacity) and control variable (water transfer) of the system are controlled, and the hydraulic coupling relationship of the reservoir group can be effectively handled.

[0086] In some optional embodiments, in terms of the specific implementation framework of time sequence progressive evolution, in addition to the advancing mode of each time period (i.e. step length of 1), the advancing mode of each window can also be adopted. For example, each time an optimization containing a time window of 3 time periods is rolled forward, only the decision variable of the first time period in the window is fixed, and the variables of the subsequent two time periods are reserved as the initial values of the next optimization. This rolling horizon control (RHC) strategy can have better robustness when dealing with strong time lag systems. Further, in the historical feature mapping link, in addition to using weighted Euclidean distance for retrieval, cosine similarity, Mahalanobis distance or sequence distance measurement based on dynamic time warping (DTW) can also be used. For the encoding of pattern parameters, in addition to principal component analysis and piecewise mean, self-encoder neural network can also be used for nonlinear dimension reduction to extract deeper potential features.

[0087] In an optional implementation, in the constraint processing strategy, in addition to the ε-CDP method, the non-corresponding penalty function method or the gradient-based repair operator in multi-objective optimization can also be combined. In particular, for linear constraints, the simplex method or the projection operator can be directly used to project the infeasible solution to the boundary of the feasible region, thereby accelerating the convergence. These variants may differ in specific algorithm operators, but as long as they are based on the time-varying constraint state sequence proposed in the present application to dynamically trigger or adjust these operators, they should fall within the protection scope of the present application.

[0088] In a possible embodiment, the cross-basin scheduling method based on knowledge driving and time sequence progressive constraint processing can also be used for:

[0089] Step one, collect data of the inter-basin reservoir system, including water network topology, reservoir inflow, reservoir storage constraints, minimum release, water transfer capacity, water transfer target, etc., and conduct data quality inspection and preprocessing.

[0090] Step two, based on reservoir inflow data, release needs and designed water transfer capacity, classify all scheduling periods according to the period division standard to obtain the period classification result time flags ; consider all variables, optimize without considering constraints, and when the phase conversion condition is reached, proceed to the next step.

[0091] Specifically, the period division standard is as follows: for an inter-basin multi-reservoir system (R1, R2, …, RN) containing N reservoirs, reservoir i has corresponding reservoir inflow inflow N at time t i t , release demand rf i t and designed water transfer capacity (the sum of the designed flow of the water transfer project with the reservoir as the water source) trc i t If inflow i t > rf i t + trc i t , it indicates that the water resources of reservoir i are abundant at time t, and there is no obvious conflict between the release demand and the water transfer demand, and reservoir i is easy to optimize in the period. If all reservoirs of the reservoir system are easy to optimize in the period, the period is marked as Rich, otherwise marked as Poor. For convenience of subsequent optimization, all targets are standardized, and the formula is as follows: f norm = (f-f min ) / (f max -f min ); where f is the original target value; f norm is the standardized target value; f max , f min are the maximum and minimum values of the target. The flow of phase conversion judgment is as follows: for the current population pop t , calculate the phase conversion judgment index value:

[0092] Obj=∑ i=1 N ∑ j=1 M ∣f norm i j (x)∣;

[0093] value=10 M-2Obj / M x N;

[0094] where M is the target number, N is the population size, f norm i j is the normalized objective value of the ith individual on the jth objective; Obj is the absolute sum of the normalized objective values of the current population on all individuals and all objectives.

[0095] For the current population pop g and the previous population pop g-1 , the stage conversion judgment index value g of the current population pop g is calculated, and the stage conversion judgment index value g-1 of the previous population pop g-1 is calculated. If |value g - value g-1 | < λ, it is considered that the current population meets the stage conversion condition, where λ is a threshold parameter.

[0096] Step three, using a time sequence progressive constraint processing mechanism, the population is optimized by time period.

[0097] Specifically, the time sequence progressive constraint processing mechanism is that if the current time period serial number is cur, when the population is iterated, only the variables of [1, cur] are subjected to cross mutation operation, and the remaining variables remain unchanged, and when the constraint is considered, only the constraint conditions in the [1, cur] time period are considered. For the time period cur marked as poor, environmental selection is performed based on the CDP strategy. The CDP strategy is that if both individuals are infeasible solutions, the constraint violation degrees of the two individuals are compared. If CV1< CV2, individual one dominates individual two. If individual one is a feasible solution and individual two is an infeasible solution, individual one dominates individual two. If individual one and individual two are both feasible solutions, the dominance relationship between individual one and individual two is determined according to the objective value. The non-dominated solution of the two individuals is selected. For the time period marked as rich, it is traversed backward until the next time period marked as poor is found rich end , and stage a is entered. First, only the constraints of the [0, cur-1] time are considered, and environmental selection is performed based on the CDP strategy. When the stage conversion condition is reached, stage b is entered, and the constraints of [0, rich end ] are considered, and environmental selection is performed based on the adaptive ε-CDP strategy. The ε-CDP strategy is that if the constraint violation degree CV is less than ε, the solution is regarded as a feasible solution, and the rest is the same as the CDP strategy. The calculation formula of ε is:

[0098] ε0 = max(CV tb i ), i ∈ (1, N); ε = ε0 · (1-(gb / G cur )) 2 G cur =(cur / T)·(βG max -g s );

[0099] Where ε0 is the maximum constraint violation degree of an individual in the population at the moment when phase b is just included, and CV tb i When transitioning to stage b (t) b The constraint violation degree of individual i within the population (constraints in the time interval [1, cur] are considered only), where N is the number of individuals in the population; g s For the number of iterations to enter stage b, g b G represents the number of iterations consumed after entering phase b. cur The maximum allowed number of iterations for the current time period, where cur is the current time period number, T is the total number of time periods, β is the parameter, and G... max This represents the maximum number of iterations allowed in the entire optimization process.

[0100] Step 4: Utilize historical solution learning strategies to generate the initial population for the next time period during stage transition, including solution distribution learning and solution prediction strategies.

[0101] Specifically, the historical solution learning strategy includes a solution prediction strategy and a solution distribution learning strategy. The solution prediction strategy is as follows: when the optimization time interval changes from cur to cur+1, pop the current population. t For each individual `ind` within the target space, select the individual with the smallest Euclidean distance from the previous generation population as the corresponding individual `ind`. cor When making predictions, only variables within the time interval [1, cur] are considered. For the water transfer variable, based on the time interval [1, cur], the water transfer sequence of ind (tran1, tran2, ..., tran) is... cur ), with ind cor Water diversion sequence (tran) cor1 tran cor2 , ...,tran corcur ), calculate the potential variation direction D1 of the water transfer volume variable: D1=(tran1,tran2,…,tran2) cur )- (tran cor1 tran cor2 , ...,tran corcur For the reservoir capacity sequence, the difference sequence Diff of the reservoir capacity is calculated for each time period: Diff = (diff1, diff1, ..., diff) T ); diff t =c t-c t-1 ;where c t Let be the reservoir capacity at time t. Based on the difference sequences Diff1 and Diff2 of ind, the potential variation direction D2 = Diff1 - Diff2 is calculated. The predicted water transfer sequence is then calculated: Pre1 = r·D1 + (tran1, tran2, ..., tran cur ); where r is a random number in (0, 1). Calculate the predicted difference sequence: Pre2 = r·D2 + Diff1. Based on the initial reservoir capacity, restore the difference sequence to the reservoir water level sequence. For the water diversion sequence and reservoir capacity sequence, predict the variables within the [1, cur] time period using the above method. For the variables within the cur+1 time period, randomly initialize them within the value range. For the variables within the [cur+2, T] time period, keep the ind value unchanged, and generate the initial individuals for the new time period. The solution distribution learning strategy is: based on the previous generation population pop g-1 The mean μ and variance σ of each dimension variable are calculated, and a Gaussian distribution N(μ, σ) is constructed. New individuals are generated by sampling the Gaussian distribution of the corresponding variables in the [1, cur] time period. Two methods, the prediction strategy and the distribution learning strategy, are used to generate N / 2 individuals respectively.

[0102] Step 5: Repeat steps 3 and 4 until all time periods have been optimized or the maximum number of iterations has been reached, and output the final optimization result.

[0103] In a detailed experimental case, two water diversion projects are considered: the Yangtze River Diversion Project and the South-to-North Water Diversion Project (Middle Route), comprising two reservoirs: the Three Gorges Reservoir and the Danjiangkou Reservoir. Data includes the ten-day inflow rates of both reservoirs, their minimum downstream discharge requirements, their ten-day storage capacity ranges, the downstream water demand of the Danjiangkou Reservoir (including ecological flow), and the maximum ten-day water diversion flow rates of both projects. Based on this data, a cross-basin reservoir scheduling model is constructed. Considering actual engineering needs, the optimization objectives are to minimize the downstream water demand deficit rate of the Danjiangkou Reservoir (including ecological flow) and the minimum water supply target deficit rate of the Middle Route Project, with the downstream water demand of the Three Gorges Reservoir as a constraint. Constraints are set, including reservoir capacity constraints, downstream discharge flow constraints, water conveyance capacity constraints, and water balance constraints. A cross-basin reservoir system topology is established. The optimization objectives are to minimize the downstream water demand deficit rate of the Danjiangkou Reservoir (including ecological flow) and the minimum water supply target deficit rate of the Middle Route Project. To ensure a more uniform water supply, the following formula is used as the objective function of the model in subsequent solution steps. Specifically, the mathematical expression for minimizing the water demand deficit rate (including ecological flow) downstream of Danjiangkou is as follows:

[0104] F need =1 / T∑ t=1 T ((rf tdan +tran t yjbh -need t ) / need t ) 2 ; Obj1 = Min(F need ) ;

[0105] wherein rft t dan is the discharge of the Danjiangkou Reservoir at time t, tran t yjbh is the water transfer amount of the Danjiang-Kong-to-Han River Project at time t, need t is the water demand of the Danjiangkou downstream at time t, F need is the mean square value of the Danjiangkou downstream water supply target shortage rate, and Obj1 is the standardized target value of the Danjiangkou downstream.

[0106] The mathematical expression of the minimum water supply target shortage rate of the middle line is as follows:

[0107] F target = 1 / T∑ t=1 T ((tran t nsbd -target t nsbd ) / target t nsbd ) 2 ; Obj2 = Min(F target ) ;

[0108] wherein tran t nsbd is the water transfer amount of the South-to-North Water Transfer Project at time t, target t nsbd is the water transfer target of the South-to-North Water Transfer Middle Line Project at time t, F target is the mean square value of the South-to-North Water Transfer Middle Line water supply target shortage rate, and Obj2 is the standardized target value of the South-to-North Water Transfer Project.

[0109] The model is solved by using the cross-basin scheduling optimization method based on knowledge driving and time sequence progressive constraint processing, and the individual is represented as: ind = [tran1 yjbh , tran2 yjbh , …, tran T yjbh ; tran1 nsbd , tran2 nsbd , …, tran T nsbd ; V1 dan , V2 dan , …, VT dan ; V1 san , V2 san ,..., V T san ]. The matrix from top to bottom represents the water diversion amount of the Yangtze River water supplement to the Hanjiang River project, the water diversion amount of the South-to-North Water Diversion Project, the reservoir capacity of the Danjiangkou Reservoir, and the reservoir capacity of the Three Gorges Reservoir. When performing cross variation, the matrix is expanded by row for convenience of operation. The discharge of the reservoir and other variables can be calculated by using the water balance equation through the variables in the matrix. According to inflow i t and rf i t + trc i t , it is judged whether the reservoir i is easy to optimize at the t period. It is judged that the Danjiangkou Reservoir is easy to optimize at the t period if the inflow inflow dan t of the Danjiangkou Reservoir at the t period is greater than the discharge need need dan t and the design water diversion flow trc nsbd t of the South-to-North Water Diversion Project. If the inflow inflow san t of the Three Gorges Reservoir at the t period is greater than the discharge need need san t and the design water diversion flow trc nsbd t of the Yangtze River water supplement to the Hanjiang River project, the Three Gorges Reservoir is easy to optimize at the t period. If the Danjiangkou Reservoir and the Three Gorges Reservoir are both easy to optimize, the period is marked as rich, otherwise, if any one of the reservoirs is not easy to optimize, the period is marked as poor. Finally, the period marking array time flags is obtained.

[0110] The cross variation is performed using the difference operator to generate offspring, and the fusion population P merge is formed with the parent population. The N individuals are selected according to the crowding distance and non-dominated relationship to form the next generation population Pop g+1 without considering the constraint violation of the individuals. The period index cur is initialized as 1. The value value g of the population Pop g is calculated, and if the inequality |value g -value g-1 | < λ is satisfied, it is considered that the stage conversion condition is met, and the stage conversion is performed. The mark of the current period, i.e., time flags[Cur], if marked as rich, the rich phase optimization algorithm is used, if marked as poor, the poor phase optimization algorithm is used. In the optimization process, the differential evolution operator is used to generate offspring, but only the variables in the period of [1, cur] are changed, and the variables in the remaining periods remain consistent with the parent. Among them, the poor phase optimization algorithm is: after the offspring is generated, evaluation is carried out, and the constraint violation matrix of the individual after evaluation is as follows: cv matrix = [cv1 1 , cv2 1 , …, cv T 1 ; cv1 2 , cv2 2 , …, cv T 2 ; …; cv1 con , cv2 con , …, cv T con ];

[0111] Wherein cv t i represents the violation degree of the individual in the t period of the i constraint condition, and con represents the number of constraint conditions. For the convenience of subsequent judgment, when calculating the constraint violation degree, it is uniformly converted into flow volume. When the environment is selected, the constraint violation degree of the individual is recorded as CV cur , that is, only the constraints in the period of [1, cur] are considered, and the calculation formula is as follows: CV cur =∑ i=1 con ∑ t= 1 cur cv t i . After the individual is evaluated, the offspring and the parent are combined, and based on the CDP strategy, N individuals are selected from them as the next generation population. When comparing the constraint violation degree, CV cur is used for comparison. After each iteration is completed, it is judged whether the current population meets the phase conversion condition, if it meets, the next step is entered, if it does not meet, the iteration is continued.

[0112] The rich phase optimization algorithm is: if the current period is marked as rich, record the current period index cur, traverse time flags backward until the next period is marked as poor, record the index of this period rich end . The rich phase optimization algorithm includes two sub-stages. Stage a: after the offspring is generated and evaluated, only the constraints in the period of [1, cur-1] are considered, and CV cur-1As the constraint violation degree of individuals, the environment is selected based on the CDP strategy. After obtaining the next generation population, it is determined whether the current population meets the stage conversion condition. If it does not meet the condition, iteration is continued. If it meets the condition, the iteration number g after entering stage b is recorded s , the CV of all individuals is calculated cur-1 , let ε0 be equal to the maximum CV cur-1 , enter stage b. Stage b: after the offspring generation evaluation, record the iteration number g after entering stage b b , calculate ε. Considering the constraints in the period [1, rich end ], calculate the CV richend As the constraint violation degree of individuals, the environment is selected based on the CDP strategy. After obtaining the next generation population, it is determined whether the current population meets the stage conversion condition. If it does not meet the condition, iteration is continued. If it meets the condition, the iteration number g after entering stage b is recorded

[0113] After the population meets the stage conversion condition, solution distribution learning and solution prediction strategy are performed. Solution distribution learning is: according to the final population before stage conversion, the mean μ and variance σ of variables in (num R +num tr )T dimensions are calculated in turn, where num R is the number of reservoirs in the system, and num tr is the number of water transfer projects in the system; (num R +num tr )T Gaussian distributions N i，j (μ, σ) are constructed, and a distribution matrix is constructed: [N 1，1 (μ, σ), N 1，2 (μ, σ), …, N 1，T (μ, σ); N 2，1 (μ, σ), N 2，2 (μ, σ), …, N 2，T (μ, σ); …; N numR+numtr，1 (μ, σ), N numR+numtr，2 (μ, σ), …, N numR+numtr，T (μ, σ)];

[0114] The Gaussian distribution corresponding to the period [1, cur] is sampled to generate new individuals.

[0115] Let the ind1 variable matrix be: [tran1 yjbh , tran2 yjbh , …, tran T yjbh ; tran1 nsbd , tran2nsbd ,..., tran T nsbd ; V1 dan , V2 dan ,..., V T dan ; V1 san , V2 san ,..., V T san ].

[0116] Let the variable obtained from the Gaussian distribution N i，j (μ, σ) be pre1 i，j , and generate offspring. The solution prediction strategy randomly selects N / 2 individuals from the population as parents, and generates N / 2 offspring individuals in the above manner.

[0117] The solution prediction strategy is as follows: according to the final population pop t before the stage transformation, and the final population pop t-1 of the last stage, the target value matrix is obtained, and the Euclidean distance between the individuals of the two populations is calculated, as follows: d = sqrt(∑ i=1 M (Obj i t - Obj i t -1 ) 2 ); where Obj i t is the i-th target value of the individual in pop t . For the individual in pop t , the individual in pop t-1 with the smallest Euclidean distance is regarded as the corresponding individual ind cur , and when predicting, only the variables in the [1, cur] period are considered. For the water transfer amount variable, according to the [1, cur] period, the water transfer amount sequence of ind and the water transfer amount sequence of indcor, the mutation direction D1 is calculated, for the reservoir capacity sequence, the difference sequence Diff and the predicted water transfer amount sequence and the difference sequence are calculated, and according to the initial reservoir capacity of reservoir scheduling, the difference sequence is restored to the reservoir water level sequence by accumulating the difference sequence by time period. For the variables in the cur+1 period, they are randomly initialized within the value range, and the values of the variables in [cur+2, T] remain unchanged, and the initial individual in the new period is generated. The solution prediction strategy randomly selects N / 2 individuals from the population as parents, and generates N / 2 offspring individuals in the above manner. A total of N individuals are generated by the two methods, which are used as the initial population after the stage transformation. Repeat until the evaluation budget is exhausted.

[0118] The algorithm parameters are set as follows in the test solution of the above-mentioned data set: the initial population size is set to 100, the population size N is set to 100, the maximum evaluation times and iteration times are set to 400000 and 4000. The crossover probability of the differential evolution operator is set to 0.7, and the mutation probability is set to 0.5. The parameter λ is set to 0.001, and the parameter β is set to 0.8. The missing rates of the middle line water supply and the downstream water demand in different years can be obtained.

[0119] Overall, the cross-basin scheduling method based on knowledge driving and time sequence progressive constraint processing acquires the hydraulic topology of the reservoir group and the hydrological supply and demand data, quantitatively generates a time-varying constraint state sequence representing the optimization difficulty of each scheduling period, such as scheduling difficulty score or sliding window constraint tightness; utilizes the sequence to dynamically regulate the constraint processing strategy, controls the risk by layers or adjusts the parameters adaptively, and optimizes stage by stage in the time sequence progressive evolution process; and at the stage conversion, executes historical feature mapping based on the scenario-pattern library to construct the initial population of the next stage.

[0120] The application adopts multi-dimensional scheduling difficulty field construction or continuous constraint tightness index (CTI), realizes continuous and accurate quantification of the optimization difficulty of the period by fusing hydrological supply and demand, risk constraints and algorithm convergence state, or calculating the supply and demand tightness based on the sliding window, so that the system can identify and focus on the key bottleneck period. The risk layered constraint processing based on the difficulty field or the parameter adaptive mechanism based on the tightness realizes the flexible processing of secondary constraints under the premise of guaranteeing the safety bottom line (such as flood control), effectively avoids the stagnation of the algorithm under strong constraints. The structured historical migration strategy based on the scenario-pattern library realizes the structured reuse of historical high-quality scheduling rules by extracting scenario feature vectors and retrieving historical mode parameters, cooperating with main trend decoding and controlled disturbance, overcomes the blindness of random initialization, and improves the cold start efficiency.

[0121] The above describes the preferred embodiments of the application in detail, but the application is not limited to the specific details in the above embodiments, and various equivalent transformations can be made to the technical solutions of the application within the technical concept range of the application, and these equivalent transformations all belong to the protection range of the application.

Claims

1. A cross-basin scheduling method based on knowledge driving and time sequence progressive constraint processing, characterized in that, The application comprises the following: Based on the obtained water connection topology of the inter-basin reservoir group and the hydrological supply-demand boundary data, a time-varying constraint state sequence representing the optimization difficulty of each scheduling period is quantitatively generated; The time-varying constraint state sequence is used to dynamically regulate the constraint processing strategy, and in the process of time sequence progressive evolution, the target scheduling scheme meeting the preset engineering limit is optimized stage by stage; At the stage conversion node of the time sequence progressive evolution, historical feature mapping is performed to construct the initial population of the next stage, and the final converged target scheduling scheme is converted into actual regulation instructions of the reservoir group; The time-varying constraint state sequence is specifically a scheduling difficulty score sequence; the time-varying constraint state sequence is quantitatively generated, including: Based on the hydrological supply-demand boundary data, the hydrological supply-demand difficulty component representing the degree of approximation of the supply-demand gap and the reservoir capacity boundary is calculated; According to the predefined constraint risk level definition, the distribution weights of high-risk constraints and low-risk constraints in the current period are counted to generate the constraint risk difficulty component; The hydrological supply-demand difficulty component and the constraint risk difficulty component are coupled in multiple dimensions to generate the scheduling difficulty score sequence; The time-varying constraint state sequence is used to dynamically regulate the constraint processing strategy, including performing weighted constraint control based on risk stratification, specifically: The pre-stored physical constraint conditions of the whole period are divided into a high-risk constraint set and a low-risk constraint set; The scheduling difficulty score sequence is used as a non-uniform amplification coefficient to calculate the high-risk weighted violation degree of individuals under the high-risk constraint set and the low-risk weighted violation degree of individuals under the low-risk constraint set, and a stratified weighted violation degree vector is constructed; Based on the stratified weighted violation degree vector, stratified dominance logic is executed in the environmental selection process to eliminate individuals with worse high-risk weighted violation degrees than competitors, and when the high-risk weighted violation degrees are the same, the low-risk weighted violation degree is used to determine the dominance relationship, so as to screen the optimal population for the next generation evolution; The time-varying constraint state sequence is specifically a sliding window constraint tension sequence; the time-varying constraint state sequence is quantitatively generated, including: Based on the hydrological supply-demand boundary data, the basic tension representing the proportion of the single-reservoir supply-demand gap in the water transfer capacity is calculated; The reservoir capacity state correction coefficient reflecting the reservoir capacity adjustment margin is constructed according to the real-time reservoir capacity state of the reservoir in the storage period or the discharge period, and the basic tension is weighted and corrected according to the correction coefficient to obtain the single-reservoir corrected tension; According to the water connection topology, the system weight of each reservoir is determined, the single-reservoir corrected tension of each reservoir is spatially aggregated, and a time decay factor is introduced to perform forward aggregation on the spatial aggregation value in the future preset window to generate the sliding window constraint tension sequence.

2. The method of claim 1, wherein, The time-varying constraint state sequence is used to dynamically regulate the constraint processing strategy, including performing algorithm parameter self-adaptation based on the tension, specifically: A nonlinear mapping relationship between the sliding window constraint tension sequence and the algorithm control parameters is established to generate an adaptive evolution parameter set, including a dynamic relaxation threshold that decays exponentially with the tension index, used to control the tolerance of infeasible solutions, a dynamic optimization range that inversely scales with the tension, used to determine the length of the time period of the decision variables currently participating in evolution, and a dynamic elite retention ratio that positively correlates with the tension, used to determine the number of preferred individuals to be retained when the population is updated; The adaptive evolution parameter set is injected into the iterative process of the time-series progressive evolution in real time to dynamically adjust the search behavior of the algorithm.

3. The method of claim 1, wherein, Historical feature mapping is performed to construct the initial population of the next stage, including structured retrieval based on a scenario-pattern library, specifically: A historical scenario-pattern library containing the mapping relationship between historical scenario feature vectors and low-dimensional scheduling pattern parameters is constructed. Multidimensional statistical analysis is performed on the current stage's hydrological supply and demand boundary data to extract a current scenario feature vector containing flow statistical characteristics and constraint difficulty characteristics. The weighted spatial distance between the current scenario feature vector and historical records is calculated in the historical scenario-pattern library to retrieve the target scheduling pattern parameters that best match the current working condition.

4. The method of claim 3, wherein, Structured retrieval based on the scenario-pattern library also includes pattern decoding and population generation, specifically: The target scheduling pattern parameters are decoded and restored to the full-period main trend scheduling trajectory using a preconfigured inverse transformation matrix or basis function. The main trend scheduling trajectory is used as the distribution center, and random noise with a variance intensity regulated by the time-varying constraint state sequence is introduced for sampling to generate an initial population of the next stage containing a predetermined number of individuals, which is injected into the next stage of the time-series progressive evolution calculation.

5. The method of claim 1, wherein, The time-varying constraint state sequence is a binary optimization state sequence; quantifying the time-varying constraint state sequence includes: For each scheduling period, the inflow and discharge demand of each reservoir are obtained and compared with the sum of the design water transfer capacity to determine the single-reservoir supply and demand balance state. If the inflow of all reservoirs in the scheduling period is greater than the sum of the discharge demand and the design water transfer capacity, the scheduling period is marked as an easy-to-optimize period, otherwise it is marked as a difficult-to-optimize period, generating a binary optimization state sequence composed of easy-to-optimize period markers and difficult-to-optimize period markers.

6. The method of claim 5, wherein, The time-varying constraint state sequence is used to dynamically regulate the constraint handling strategy, including performing phased constraint control, specifically: During the time-series progressive evolution process, the current optimization period's marker in the binary optimization state sequence is read. If the marker is a difficult-to-optimize period, the full constraint checking strategy is executed, and the population individuals are selected based on the constraint dominance principle. If the marker is an easy-to-optimize period, the relaxation-tightening strategy is executed, which ignores the physical constraints of the current newly added period in the early evolution stage, and introduces a dynamic relaxation threshold after the stage transition condition is met, and selects feasible solutions based on the ε constraint dominance principle.

7. The method of claim 1, wherein, Historical feature mapping is performed to construct the initial population of the next stage, including performing solution distribution learning and difference prediction, specifically: The mean and variance of the decision variables in the final population of the previous stage are statistically analyzed to construct a Gaussian distribution model for each dimension, and a distribution learning sub-population is generated by sampling. Calculate the Euclidean distance between the current population individuals and the last stage population individuals in the target space, determine the nearest neighbor correspondence, calculate the differential evolution direction vector of the decision variable according to the nearest neighbor correspondence, and perform linear extrapolation prediction on the decision variable sequence to generate a differential prediction sub-population; Merge the distribution learning sub-population and the differential prediction sub-population as the initial population of the next stage.

Images