A multi-objective hierarchical intelligent optimization decision-making method and system for reservoir drought limit water level

Abstract

This invention discloses a multi-objective hierarchical intelligent optimization decision-making method for reservoir drought limit water levels, deployed in a reservoir scheduling automation system. The method includes the following logically strongly coupled and irreversibly sequential steps: constructing a multi-dimensional objective function topology network for drought limit water level scheduling; establishing multi-element marginal triggering rules for drought limit water level scheduling; creating a reservoir scheduling digital twin model considering the activation signals; pre-creating 19 non-redundant weight configuration vector sets in the system rule engine, generated by topological sorting of four types of objective nodes: benefit, risk, cost, and effectiveness; using the drought limit water level sequence of key nodes as decision variables; and implementing a two-level normalization compensation mechanism to eliminate dimensional differences for all non-dominated solution sets. This invention forms a logically strongly coupled six-step irreversible decision pipeline, achieving optimized allocation of computing resources and industrial-grade security against parameter tampering. It is applicable to intelligent drought limit water level scheduling for multi-objective reservoirs involved in flood control, water supply, ecology, power generation, and navigation.

Description

Technical Field

[0001] This invention relates to the field of reservoir scheduling automation technology, and in particular to a multi-objective hierarchical intelligent optimization decision-making method and system for reservoir drought limit water level. Background Technology

[0002] The drought limit water level is a crucial indicator of the shift in reservoir management from normal water supply to emergency water restriction, and its determination requires comprehensive consideration of multiple objectives, including flood control, water supply, ecology, power generation, and navigation. Existing technologies have the following main shortcomings:

[0003] Objective function decentralization: It often adopts independent linear weighted models, lacks a topological expression of the inherent dependencies between the four types of objectives: risk, effectiveness, benefit, and cost, resulting in strong subjectivity in weight allocation.

[0004] Simplified activation conditions: The common practice is to use a single water level threshold for triggering, without logically coupling the three elements of incoming water forecast, downstream water demand and current water level, which makes it difficult to meet the "three-out-of-three" redundant judgment required for project reliability.

[0005] The solution is inefficient: it uses serial traversal for multiple weight configurations, lacks the prediction and pruning mechanism for invalid solution branches, and consumes a lot of computational resources.

[0006] The parameters are not secure: the weight configuration is stored as a software variable, which is susceptible to human tampering or network attacks, and there is no hardware-level circuit breaker protection.

[0007] Evaluation mechanism defects: Single-level normalization is prone to target masking effect, that is, secondary targets interfere with the decision-making of primary targets after normalization, and the schemes have insufficient distinguishability.

[0008] To address the aforementioned problems, this invention proposes a complete technical solution based on topological network compilation, Boolean logic AND gates, digital twin mapping, hierarchical pruning, hardware circuit breaking, and secondary compensation normalization. Summary of the Invention

[0009] The purpose of this invention is to provide a multi-objective hierarchical intelligent optimization decision-making method and system for reservoir drought limit water level, thereby solving the aforementioned problems existing in the prior art.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] A multi-objective hierarchical intelligent optimization decision-making method for reservoir drought limit water level, deployed in a reservoir scheduling automation system, includes the following logically strongly coupled and sequentially irreversible steps:

[0012] S10: Construct a multi-dimensional objective function topology network for drought-limited water level scheduling. The network nodes should include at least four types of functional nodes: risk avoidance objective, benefit objective, project effectiveness objective, and management cost objective. Each node is connected by directed edges to form an objective dependency graph. The dependency graph is then compiled into a set of scheduling period evaluation functions that can be computed in parallel.

[0013] S20: Establish a multi-element marginal triggering rule for drought-limited water level scheduling. The rule will use the current water level status, future water inflow forecast, and downstream water demand forecast to perform Boolean logic three-to-three AND gate circuit operation based on engineering reliability requirements, and output a unique activation signal for automatic triggering control command.

[0014] S30: Create a digital twin model for reservoir scheduling that takes into account activation signals. The model maps the decision variable domain to the water level-discharge phase space trajectory during the scheduling period by solving the water balance state equation and the set of engineering constraint inequalities simultaneously.

[0015] S40: In the system rule engine, 19 non-redundant weight configuration vector sets are pre-formed, which are generated by topological sorting of four types of target nodes: benefit, risk, cost, and effectiveness. The vector sets form a progressive calling sequence based on the solution space hierarchical pruning mechanism. The calling sequence achieves optimized allocation of computing resources through the prediction and pruning of invalid solution branches. The weight configuration vectors are locked in a read-only state through the hardware circuit breaker mechanism during the system operation cycle.

[0016] S50: The key node drought limit water level sequence is used as the decision variable. The weight configuration vectors in the vector set are called. The multi-objective optimization algorithm with adaptive parameter dynamic calibration mechanism is used to solve the corresponding non-dominated solution set in parallel. The algorithm population size and number of iterations are adjusted in real time according to the objective function dimension. The lower limit of the population size is 200 and the lower limit of the number of iterations is 500.

[0017] S60: A two-level normalization compensation mechanism to eliminate dimensional differences is implemented for all non-dominated solution sets. The compensation mechanism eliminates the target occlusion effect caused by single-level normalization and enhances the distinguishability of the schemes through the serial processing of the first-level sub-target normalization and the second-level comprehensive benefit normalization. Based on the priority of the calling sequence of the hierarchical pruning mechanism, the comprehensive effect index is automatically calculated, and the drought limit water level scheduling scheme corresponding to the optimal solution of the index and its confidence assessment report are output.

[0018] In some specific embodiments, the four types of functional nodes in the S10 multidimensional objective function topology network are specifically:

[0019] Risk avoidance node: The flood control objective function minF1 is used to calculate the maximum flood frequency that the remaining flood control reservoir capacity can cope with at the drought limit water level during the flood season; its calculation formula is:

[0020] In the formula: The maximum flood frequency (in billions of m³) that the drought limit water level can cope with during the flood season. 3 ); The discharge volume (100 million m³) is required to meet flood control requirements. 3 ); The curve showing the relationship between flood volume and flood frequency; This is the reservoir water level-capacity curve;

[0021] Beneficial benefit nodes include water supply objective function maxF2, ecological objective function maxF3, power generation objective function maxF4, and shipping objective function maxF5;

[0022] The formula for calculating the target flow rate guarantee rate of water supply is: ;

[0023] In the formula: The water supply guarantee rate is represented by n, which represents the total number of time periods. The reservoir discharge flow rate (m³) during time period t 3 / s); The reservoir discharge flow (m³) corresponding to meeting water supply demand during time period t. 3 / s);

[0024] The formula for calculating the flow guarantee rate for ecological targets is: ;

[0025] In the formula: To ensure ecological flow guarantee rate; Let m be the reservoir discharge flow rate corresponding to meeting ecological needs during time period t. 3 / s;

[0026] The formula for calculating the total power generation target is: ;

[0027] ;

[0028] In the formula: Electricity generation (100 million kWh); The power generation of the reservoir during time period t is 100 million kWh. Acceleration due to gravity (m / s²) 2 ); The power generation flow rate (m³) of the power plant during time period t 3 / s); The generator efficiency of the power station; The density of water (kg / m³) 3 ); The difference in hydropower head (m) during time period t at the power station. The power generation duration (h) of the power station during time period t;

[0029] The formula for calculating the navigation guarantee rate for shipping targets is: ;

[0030] In the formula: For navigation guarantee rate; The water level (m) of the reservoir during time period t; Water level (m) required for navigation;

[0031] Project effectiveness milestone: The water storage objective function maxF6 is used to calculate the average maximum water storage level over a multi-year water storage period;

[0032] ;

[0033] In the formula: The average highest water level is (m); N is the total number of scheduling cycles (years). The water level (m) at the start of the reservoir's storage period in the i-th year; This represents the water level at the end of the reservoir's storage period in year i.

[0034] Management cost node: The objective function minF7 for the number of activations is used to calculate the proportion of days the drought limit water level is activated; its calculation formula is:

[0035] ;

[0036] The dependency graph is compiled using the Kahn algorithm to generate a directed acyclic graph. Topological sorting generates 19 non-redundant weight configuration vector sets. These vector sets are formed by stacking the dual-mode benefit nodes and the independent configurations of the project effectiveness nodes to form a four-level progressive structure: Level 1: 1 pure benefit mode; Level 2: 2 benefit / total benefit and risk combinations; Level 3: 4 benefit / total benefit and cost combinations; Level 4: 12 benefit / total benefit fully coupled with risk, cost, and effectiveness. Among them, the pure risk, pure cost, and risk + cost configurations lack the input of the first-level basic nodes and do not have independent optimization value. They are only included in the vector set as a reference group for quantitative evaluation of the optimization gain after introducing the benefit / effectiveness target and do not participate in the main process solution of the progressive call sequence.

[0037] In some specific embodiments, the three-factor threshold parameters of the Boolean logic 3-to-3 AND gate circuit in step S20 are:

[0038] The future number of days X is set to 10 days;

[0039] Designed for low water flow The flow rate is set in three phases: 13,800 m³ / s during the main flood season, 11,000 m³ / s during the water storage period, and 5,600 m³ / s during the dry season.

[0040] The flood control requirement is a flow rate Q_f of 28000 m³ / s;

[0041] The warning water level Z_y is set at 35m;

[0042] When all three elements meet the threshold conditions, the AND gate outputs a high-level activation signal; otherwise, it outputs a low-level signal to maintain normal operation.

[0043] In some specific embodiments, the water balance equation in step S30 is specifically as follows: Where f2() is the water level-storage capacity curve, f2'() is its inverse function, θ is the flow-water conversion coefficient, and Δt is the time period length. This refers to the amount of loss due to evaporation and leakage. The water level at the end of time period t in the reservoir. The initial water level of the reservoir during time period t is in meters. Let m be the inflow rate of the reservoir during time period t. 3 / s; The outbound flow rate m during time period t 3 / s;

[0044] The system of engineering constraint inequalities specifically includes five types of constraints:

[0045] Discharge capacity constraints: In the formula: , Let m represent the minimum and maximum discharge rates of the reservoir in the t-th time period, respectively. 3 / s, determined by downstream flood control, maximum discharge capacity, and reservoir water level fluctuations;

[0046] Power generation flow constraints: In the formula: For power generation flow rate m 3 / s; The maximum power generation flow rate m at time t 3 / s, determined by the unit's flow capacity corresponding to the water level at time t;

[0047] Unit output constraints: In the formula: This represents the reservoir's output in kW during time period t. , The minimum and maximum output limits for time period t are kW, and the minimum and maximum output limits are determined according to the dispatching procedures.

[0048] Reservoir water level constraints: In the formula: and The minimum and maximum water levels m in time period t are respectively determined based on the control water level at that time.

[0049] Daily water level fluctuation constraints: In the formula: The upper limit of the reservoir water level fluctuation during time period t, in meters;

[0050] In step S30, the flow phase space trajectory mapping maps the reservoir state variables to a three-dimensional phase space. The feasibility of the scheduling scheme is predicted by analyzing the trajectory curvature and flow field stability. If the feasibility is lower than the threshold, the pruning mechanism in step S40 is triggered to remove the branch of the scheme in advance.

[0051] In some specific embodiments, the topological sorting generation process in step S40 includes:

[0052] Step a: Extract the loop structure from the dependency graph in step S10;

[0053] Step b: Introduce virtual zero nodes to relax constraints in the subgraph containing loops;

[0054] Step c: Perform topological sorting on the relaxed directed acyclic graph using the Kahn algorithm;

[0055] Step d: Map the sorted node sequence to 19 weight configuration vectors;

[0056] The 19 weight configuration vector sets are specifically structured in a four-level progressive manner:

[0057] Level 1: 1 type, weight vector W1=[0, 0, 1, 1, 1, 1, 0];

[0058] Level 2: 2 types, with weight vectors W2-W3 based on Level 1, W1 is scaled or shifted to introduce risk weight 1;

[0059] Level 3: 4 types, with weight vectors W4-W7 further introducing a management cost weight of 1 based on Level 1 and Level 2;

[0060] Level 4: 12 types, weight vector W8-W 19 Based on the first three levels, an engineering effectiveness weight of 1 is introduced, and the scaling or translation weight vector W1 is subtracted. Then, similar terms are merged, resulting in four types, with the weight vector W8-W. 19 Based on the first three levels, an engineering effectiveness weight of 1 is introduced, comprising 8 types;

[0061] All vectors are normalized to form a non-redundant weight configuration vector set;

[0062] The execution logic of the S40 level pruning mechanism is as follows:

[0063] After solving for the non-dominated solution set corresponding to the first-level weight vector W1, the distribution density of the solution set in the two-dimensional projection space of flood control and water storage is analyzed.

[0064] If the projection density is lower than the preset threshold, all vectors W4 and W5 in the second level that do not contain flood control or water storage weights are determined to be invalid branches and are directly removed without further solution.

[0065] The pruning operation allows subsequent calculations to skip at least 5 weight vectors, thus optimizing the allocation of computational resources.

[0066] Step S40, the hardware fuse mechanism, is implemented in the following way:

[0067] During the system power-on initialization phase, 19 weight configuration vectors are loaded from flash memory into the FPGA's block random access memory;

[0068] After loading is complete, the write enable pin of the configuration register is fixed, physically cutting off the external modification path;

[0069] During operation, the weight vector is used as a read-only constant in the parallel solution of step S50. Any illegal modification instruction will trigger a hardware watchdog reset.

[0070] In some specific embodiments, the specific rules of the adaptive parameter dynamic calibration mechanism in step S50 are as follows:

[0071] When the sum of the dimensions of the four objective functions is ≤10, the population size is set to 200 and the number of iterations is set to 500.

[0072] For every 5 increases in the target dimension, the population size increases by 50, and the number of iterations increases by 100.

[0073] When the number of remaining weight vectors after the current stage pruning is ≤5, the lower limit of the number of iterations is reduced to 300 to accelerate convergence.

[0074] In some specific embodiments, the first-level processing of the secondary normalization compensation mechanism in step S60 is as follows:

[0075] The four sub-objectives of water supply, ecology, power generation, and shipping are normalized to the [0,1] interval using the range transformation method.

[0076] The inverse range transformation method is used for the sub-objectives of flood control and activation cost;

[0077] The project's effectiveness target is achieved by normalizing the average maximum water level over multiple years to [0,1] using the range transformation method.

[0078] In some specific embodiments, the second-level normalization process in step S60 is as follows:

[0079] The normalized values ​​of water supply, ecology, power generation, and shipping output from the first level are weighted at 0.25:0.25:0.25:0.25 to obtain the comprehensive benefit index B;

[0080] The comprehensive effect index C is obtained by weighting B again with the normalized flood control, water storage, and activation costs by 0.25:0.25:0.25:0.25;

[0081] The two-level weight allocation structure creates a compensatory effect. When the single-level normalization causes a conflict between the dimensions of the benefit target and the cost target, the second level recalibrates the scale to eliminate the conflict.

[0082] In some specific embodiments, the confidence assessment report of step S60 includes three main visualization components:

[0083] A radar chart comparing the various options shows the achievement status of the 19 options on the seven objectives.

[0084] Pareto front projection plot shows the distribution of the non-dominated solution set in the benefit-risk two-dimensional space;

[0085] The confidence interval diagram of the drought limit water level process is given, along with the water level process line and 90% confidence band of the optimal solution.

[0086] A smart optimization decision-making system for reservoir drought limit water level, based on the same concept, is deployed in the industrial control network of the reservoir dispatch center, including:

[0087] The objective function topology network construction module is used to execute step S10 and output the objective dependency graph.

[0088] The edge-triggered rule engine module is used to execute step S20 and has built-in Boolean logic AND gate circuits;

[0089] The digital twin model calculation module is used to execute step S30 and integrates a water balance state equation solver.

[0090] The weight configuration vector management module is used to execute step S40, which implements hardware circuit breaking through FPGA;

[0091] The parallel optimization solution module is used to execute step S50, and adopts adaptive parameter dynamic calibration.

[0092] The solution evaluation report generation module is used to execute step S60 and output the optimized solution with confidence level;

[0093] The industrial data interface unit is configured to communicate bidirectionally with the reservoir SCADA system. When the result of the three-element Boolean logic operation in step S20 is true, it automatically sends a flow control command to SCADA and receives execution feedback to form a closed-loop control.

[0094] The weight configuration vector management module uses Xilinx Zynq series FPGA. The 19 weight vectors are fixed in the form of IP cores. After the system is powered on, they are loaded into the DDR memory of the optimization solution module within 100ms through the AXI DMA interface. After loading is completed, the fuse bit is set, and any subsequent rewrite operation will trigger a hardware abnormal interrupt.

[0095] The parallel optimization solution module is deployed on a GPU-accelerated computing server, supporting the concurrent operation of 19 NSGA-II algorithm instances corresponding to weight vectors. The instances share digital twin model data through the PCIe bus, and the concurrent solution efficiency is 10-15 times higher than that of serial solution.

[0096] The beneficial effects of this invention are:

[0097] Fundamental improvement in computational efficiency: By using a hierarchical pruning mechanism to pre-select solutions in the solution space and combining it with a GPU parallel acceleration architecture, invalid computation paths are significantly reduced, enabling the multi-objective optimization solution time to meet the rigid requirements of real-time decision-making for reservoir scheduling. This completely solves the industry pain points of high computational redundancy and slow response of traditional traversal methods.

[0098] Significantly improved decision-making accuracy: The two-level normalization compensation mechanism creatively eliminates the target obscuring effect inherent in single-level normalization by separating the normalization benchmarks of benefits and costs, ensuring that the optimal solution can fully reflect the comprehensive trade-off between flood control, water storage, benefits, and costs, and that the solution selection results have high reliability and repeatability.

[0099] Industrial-grade security protection: The FPGA hardware circuit breaker mechanism elevates the weight configuration from a software variable to a circuit-level constant, and implements parameter anti-tampering through physical layer write protection. The system has the ability to resist malicious attacks and human error, and its security level meets the protection standards of reservoir scheduling automation systems.

[0100] Full-objective collaborative optimization: For the first time, the topology network compilation technology constructs four types of target nodes into a directed dependency graph. Through relational graph compilation and topology sorting, it realizes the inherent logical coupling of risk-effectiveness-benefit-cost objectives, avoiding scheduling strategy conflicts caused by the isolated design of objective functions in traditional methods.

[0101] Enhanced engineering reliability: The Boolean logic 3-to-3 AND gate triggering mechanism upgrades the activation condition from a single threshold judgment to multi-factor redundant voting, achieving a response speed of milliseconds, which significantly improves the reliability and anti-interference capability of drought-limited water level scheduling.

[0102] Technical applicability and economic efficiency: This invention is applicable to various types of comprehensive reservoirs with functions such as flood control, water supply, ecology, power generation and navigation. The hardware modules are highly reusable. Under the premise of moderately increasing costs, the system deployment generates scheduling benefits far exceeding those of traditional methods, and has excellent engineering application value and promotion prospects. Attached Figure Description

[0103] Figure 1 This is a schematic diagram of the scheduling process of the present invention;

[0104] Figure 2 This is a flowchart of the multi-objective hierarchical intelligent optimization decision-making method of the present invention;

[0105] Figure 3 This is a schematic diagram of the drought limit water level calculation results under the 19 target schemes of the present invention;

[0106] Figure 4 This is a schematic diagram illustrating the optimal result of optimizing the implementation of the present invention with the objectives of flood control, water storage, and activation. Detailed Implementation

[0107] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0108] Reference Figure 1 , Figure 2 , Figure 3 and Figure 4 The multi-objective hierarchical intelligent optimization decision-making method for reservoir drought limit water level, as shown, is deployed in a reservoir scheduling automation system and includes the following logically strongly coupled and sequentially irreversible steps:

[0109] S10: Construct a multi-dimensional objective function topology network for drought-limited water level scheduling. The network nodes should include at least four types of functional nodes: risk avoidance objective, benefit objective, project effectiveness objective, and management cost objective. Each node is connected by directed edges to form an objective dependency graph, and the dependency graph is compiled into a set of scheduling period evaluation functions that can be computed in parallel.

[0110] This step specifically involves: Step S10, constructing a multi-dimensional objective function topology network for drought-limited water level scheduling. This network consists of four types of functional nodes: risk avoidance objective, benefit objective, project effectiveness objective, and management cost objective. Each type of node carries a specific scheduling objective dimension, and they are connected by directed edges to form an objective dependency graph, clarifying the logical relationships between the objectives.

[0111] The risk avoidance target node quantifies the reservoir's ability to cope with floods during the flood season, and its function output is the maximum flood frequency that can be safely handled during the scheduling period. The benefit target node aggregates four sub-targets: water supply, ecology, power generation, and navigation. Each sub-target is quantified through indicators such as flow guarantee rate or power generation, ultimately forming a comprehensive benefit value. The project effectiveness target node calculates the average highest water level during the impoundment period, reflecting the reservoir's water storage capacity. The management cost target node tracks the number of times the drought limit water level is activated, quantifying the economic cost of scheduling management.

[0112] The nodes are connected by directed edges, forming a target dependency graph. For example, the risk avoidance node points to the project effectiveness node, indicating that flood control limits restrict the upper limit of water storage; the project effectiveness node points in reverse to the benefit sub-node, indicating that the level of water storage directly affects the realization of benefits such as water supply and power generation.

[0113] The dependency graph is compiled into a set of scheduling evaluation functions that can be computed in parallel using topological sorting. The compilation process employs the Kahn algorithm, arranging the nodes in the graph according to their dependency order to generate the computation sequence for each node within the scheduling period. The evaluation function of each node is dynamically calculated based on the output of its dependent predecessor nodes, ultimately forming a set of functions that can be executed in parallel. This set of functions is invoked daily within the scheduling period, achieving rapid solution through parallel computing resources such as GPUs, thus meeting the real-time requirements of reservoir scheduling.

[0114] S20: Establish a multi-factor marginal triggering rule for drought-limited water level scheduling. The rule uses three factors—current water level status, future water inflow forecast, and downstream water demand forecast—to perform Boolean logic 3-to-3 AND gate operation based on engineering reliability requirements, and outputs a unique activation signal for automatically triggering the control command.

[0115] Specifically, this step involves: S20 establishing multi-factor marginal triggering rules for drought-limited water level scheduling, upgrading the traditional single threshold judgment to a three-choice, three-redundant voting mechanism based on engineering reliability requirements, and implementing automatic triggering through hardware-level Boolean logic and gate circuits.

[0116] The three-element marginal triggering rule comprises three independent judgment elements, all of which must be simultaneously satisfied to activate the scheduling. Element 1 monitors the current water level status. It is deemed satisfied when the real-time water level of the reservoir is continuously lower than the dynamic drought limit water level threshold. The dynamic threshold changes with the scheduling period and is determined by key control points such as the flood limit water level transition node and the start node of the water storage period. Element 2 assesses future inflow forecasts. It is deemed satisfied when the forecasted inflow for the next 10 days is simultaneously lower than the design low water flow and less than the flood control requirement flow. The design low water flow is divided into three levels: 13,800 cubic meters per second during the main flood season, 11,000 cubic meters per second during the water storage period, and 5,600 cubic meters per second during the dry season. The flood control requirement flow is constant at 28,000 cubic meters per second. Element 3 judges downstream water demand forecasts. It is deemed satisfied when the forecasted water level at key downstream sections is lower than the 35-meter warning water level for the next 10 days.

[0117] The three-factor determination result is integrated into a Boolean logic AND gate circuit, implemented using a three-input AND gate chip or an internal lookup table in an FPGA. Each input pin receives the comparator output corresponding to the factor; a high level (3.3V) is input when the threshold is met, and a low level (0V) is input otherwise. The AND gate outputs a high-level activation signal only when all three inputs are high; if any input is low, the output remains low. This three-to-three voting structure meets the redundancy design requirements of hydraulic engineering, avoiding false triggering due to single sensor failure or prediction deviation, with a response delay of less than 10 milliseconds.

[0118] The activation signal is a square wave with a pulse width of no less than 100 milliseconds, which is transmitted to the DMA controller in subsequent steps via an isolated optocoupler to trigger model calculation. If any element condition is no longer met during the calculation process, the AND gate output immediately goes low, and the receiver detects the falling edge, stops the calculation, and returns to standby mode. This physical layer-based solidified logic transforms business rules into circuit logic, solving the security risks of traditional software judgment being susceptible to human intervention, and realizing a complete hardware-automated trigger chain from data acquisition and threshold comparison to logical voting.

[0119] S30: Create a digital twin model of reservoir scheduling that takes into account activation signals. The model maps the decision variable domain to the water level-flow phase space trajectory during the scheduling period by solving the water balance state equation and the set of engineering constraint inequalities simultaneously.

[0120] This step specifically involves: Step S30 creating a digital twin model of reservoir scheduling that considers the activation signal. The model consists of a water balance state equation and a set of engineering constraint inequalities, and calculations are initiated only when the output of the Boolean logic AND gate in Step S20 is high. The activation signal directly triggers the clock enable pin of the model solver via the DMA controller, ensuring strict synchronization between the calculation and real-time water and rainfall conditions. If the signal is removed during the calculation, the model immediately saves its current state to SRAM and enters sleep mode, resuming from the breakpoint upon the next activation, forming a pulse-like working mode of edge triggering-calculation-interruption-recovery.

[0121] The water balance equation is

[0122] ,

[0123] in () represents the water level-reservoir capacity characteristic curve, and bidirectional cross-referencing is achieved using piecewise cubic spline interpolation.

[0124] () is its inverse function, solved using Newton's iteration method with an iteration tolerance of 0.001 meters; θ is taken as 0.0864, ​​and Δt is 86400 seconds. This equation maps the reservoir capacity change (inflow-outflow-loss) over a period of time to the water level change, and is the core of the model's state evolution.

[0125] The system of engineering constraint inequalities includes five terms: discharge capacity constraint. The lower limit is set at 5700 cubic meters per second, and the upper limit is determined by the discharge capacity curve corresponding to the maximum opening of the gate. Power generation flow constraints Ensure that the current does not exceed the unit's overcurrent capacity; Unit output constraints It is necessary to avoid the vibration zone of the water turbine; Reservoir water level constraints The lower limit is the dead water level of 250.00 meters, and the upper limit is the flood control limit or normal storage level; daily water level fluctuation constraints. The maximum height is 2 meters per day to prevent the risk of reservoir bank landslides. The model employs a simultaneous solution strategy, substituting the state equations into a system of inequalities to form a system of differential-algebraic equations. The decision variables are the key node drought limit water level sequences (encoded as 15-dimensional real vector chromosomes). During the solution process, all constraints are applied simultaneously in every time period; any outflow exceeding a constraint is considered invalid. All solutions are rejected by the feasible region, ensuring the physical validity of the solution. The optimization algorithm generates one set of solutions each time. The sequence immediately invokes the state equation for recursive calculation. The sequence is substituted into a system of inequalities for verification, thus realizing the coupled simulation of state variables and control variables in the time domain. Phase space trajectory mapping will daily water level ,flow Storage capacity V t Mapped to a three-dimensional state space, a continuous trajectory is formed. The geometric features of the trajectory have clear physical meaning: excessive curvature indicates frequent opening and closing of the gate with excessive amplitude constraints; flow field divergence indicates that the state equation iteration does not converge; oscillation indicates that the scheme is unsustainable. When the curvature or stability index is lower than the preset feasibility threshold, the sequence is determined to be an invalid solution, and the subsequent pruning mechanism is directly triggered to remove it in advance, without the need to complete full-time calculations, thus achieving pre-optimization of computational resources. Model initialization is triggered by the rising edge of the activation signal. The current water level is read from the EEPROM as the initial value, and the inflow forecast for the next 10 days is loaded, which is determined according to the scheduling period. , , By setting threshold parameters, the algorithm population is reset. Those skilled in the art can implement the complete solution process using Pyomo or MATLAB without introducing undisclosed algorithms. The innovation lies in introducing phase space geometric analysis into feasibility assessment through simultaneous solution, elevating numerical verification to trajectory morphology prediction, and providing feedforward criteria for pruning. This represents a cross-disciplinary combination of scheduling optimization theory and geometric topology.

[0126] S40: In the system rule engine, 19 non-redundant weight configuration vector sets are pre-formed, which are generated by topological sorting of four types of target nodes: benefit, risk, cost, and effectiveness. The vector sets form a progressive calling sequence based on the solution space hierarchical pruning mechanism. The calling sequence optimizes the allocation of computing resources by predicting and pruning invalid solution branches. The weight configuration vectors are locked in a read-only state through a hardware circuit breaker mechanism during the system operation cycle.

[0127] Specifically, this step involves S40 pre-creating a weight configuration vector set in the system rule engine. This vector set is automatically generated by a topology sorting algorithm from four categories of target nodes: benefit, risk, cost, and effectiveness. The benefit category includes four sub-objectives: water supply, ecology, power generation, and shipping. The risk category includes flood control objectives, the cost category includes activation frequency objectives, and the effectiveness category includes water storage objectives. The topology sorting uses the Kahn algorithm to perform hierarchical sorting on the directed acyclic graph composed of the four types of nodes, generating 19 non-redundant weight configuration vectors. Each vector is a seven-dimensional array, corresponding to the weight allocation of the seven objectives.

[0128] The vector set forms a progressive four-level call sequence: the first level is a pure benefit layer, the second level is a benefit and risk superposition layer, the third level is a benefit and cost superposition layer, and the fourth level is a fully coupled layer of benefit, risk, cost, and effectiveness. The call sequence is executed based on a solution space hierarchical pruning mechanism: after solving the first level, the distribution density of the non-dominated solution set in the flood control-water storage two-dimensional projection space is analyzed. If the density is lower than a preset threshold, vectors in subsequent levels that do not contain flood control or water storage are determined to be invalid branches and are directly pruned. The pruning operation disables the corresponding solution task through a flag, skips the GPU computation queue, and saves GPU memory and computing core resources.

[0129] The weight configuration vectors are locked in a read-only state during the system runtime via a hardware fuse mechanism. This is implemented as follows: upon system power-up, 19 vectors are loaded from QSPI flash memory to the FPGA's on-chip RAM via DMA. After loading, a fuse code is written to the SLCR configuration register, physically disconnecting the write enable pin. After the fuse is broken, the vectors participate in calculations as ROM; any modification instruction triggers a hardware exception interrupt. This mechanism ensures that the scheduling scheme cannot be maliciously tampered with or accidentally modified, achieving a security level that meets industrial control standards.

[0130] S50: Using the key node drought limit water level sequence as the decision variable, calling the weight configuration vectors in the vector set, and using a multi-objective optimization algorithm with an adaptive parameter dynamic calibration mechanism to solve in parallel to obtain the corresponding non-dominated solution set, wherein the algorithm population size and iteration number are adjusted in real time according to the objective function dimension, with a lower limit of 200 for the population size and a lower limit of 500 for the iteration number.

[0131] Specifically, step S50 uses the key node drought limit water level sequence as the decision variable. This sequence consists of time nodes with control significance during the scheduling period and their corresponding water level values, encoded as a real-number vector chromosome. The gene values ​​are restricted by the constraints in S30. This step calls 19 pre-built weight configuration vectors from S40. Each vector is stored in the FPGA block RAM as a read-only constant and is read sequentially via the AXI-Lite bus. Each time a vector is read, an independent optimization algorithm instance is started, and the 19 instances form a parallel solution queue.

[0132] This invention employs the NSGA-II multi-objective optimization algorithm as the solution framework, comprising three core operators: fast non-dominated sorting, crowding distance calculation, and elite retention strategy. The algorithm configuration is determined through an adaptive parameter dynamic calibration mechanism, with a lower limit of 200 for the population size and 500 for the number of iterations. The lower limits are maintained when the objective function dimension is ≤10; for every 5-fold increase in dimension, the population size increases by 50 and the number of iterations increases by 100, ensuring sufficient search diversity in the high-dimensional objective space. If the number of remaining vectors after previous pruning is ≤5, the lower limit for the number of iterations is reduced to 300 to accelerate convergence. The crossover operator uses simulated binary crossover, with an initial distribution exponent of 20 decreasing to 5 with each iteration; the mutation operator uses polynomial mutation, with an initial distribution exponent of 20 increasing with each iteration; the crossover probability is 0.9, and the mutation probability is 0.1.

[0133] Parallel solving is deployed on a GPU-accelerated server, employing CUDA streaming parallelism. Each weight vector corresponds to an independent CUDA stream, and streams share S30 digital twin model data to reduce memory copying. The GPU kernel function evaluates individual fitness in parallel using a single-instruction multi-threaded mode, with a thread block size of 128 and a grid size dynamically adjusted based on the population. After solving, each non-dominated solution set Fᵢ contains 20 to 70 non-dominated solutions, representing approximately 10%-20% of the population. Non-dominated solutions consist of chromosome sequences and seven-dimensional objective function values, stored in an SSD array for use in the S60 steps. The adaptive calibration mechanism and GPU parallel architecture reduce the total solution time for the 19-way solution to within 15 minutes, an order of magnitude improvement in efficiency compared to CPU serial processing.

[0134] S60: A two-level normalization compensation mechanism to eliminate dimensional differences is implemented for all non-dominated solution sets. The compensation mechanism eliminates the target occlusion effect caused by single-level normalization and enhances the distinguishability of the schemes through the serial processing of the first-level sub-target normalization and the second-level comprehensive benefit normalization. Based on the priority of the calling sequence of the hierarchical pruning mechanism, the comprehensive effect index is automatically calculated, and the drought limit water level scheduling scheme corresponding to the optimal solution of the index and its confidence assessment report are output.

[0135] This step specifically involves: Step S60 receiving all non-dominated solution sets output by S50 and executing a two-stage normalization compensation mechanism. The first stage eliminates dimensional differences in the seven objective function values: Benefit-related sub-objectives (water supply guarantee rate, ecological guarantee rate, navigation guarantee rate) are linearly mapped to the 0-1 interval with 100% as the baseline; power generation is normalized with the historical maximum value as the baseline. Cost and risk-related sub-objectives are processed in reverse: flood control risk is reverse-normalized with the historical maximum frequency as the baseline; activation cost is reverse-mapped with the total number of scheduling days as the baseline; and water storage effectiveness is linearly mapped with the interval from dead water level to normal water level as the baseline. After the first stage is completed, each non-dominated solution is transformed into a seven-dimensional standard vector.

[0136] The second level constructs a comprehensive benefit index B by taking the arithmetic mean of the four normalized values ​​of water supply, ecology, power generation, and navigation output from the first level with equal weights. The value range of B is 0-1. Subsequently, the B value and the normalized values ​​of flood control, water storage, and activation cost are recalibrated using the range transformation method to eliminate scale differences. The final comprehensive effect index C is calculated using equal weights: C = 0.25 × B + 0.25 × Flood Control + 0.25 × Water Storage + 0.25 × Activation Cost, where flood control and activation cost are values ​​after reverse scoring. The C value calculation traverses all non-dominated solution sets. Within each solution set, solutions are sorted in descending order of C value. The solution with the largest global C value is the recommended scheme, and its corresponding key node drought limit water level sequence constitutes an executable dispatch instruction.

[0137] S60 simultaneously generates a confidence assessment report, including radar charts, Pareto front projections, and confidence intervals for each scheme. The report generates 100 simulated trajectories using Monte Carlo perturbations and calculates a 90% confidence band; a narrower bandwidth indicates stronger robustness. This two-stage serial normalization structure, through separate benefit aggregation and comprehensive trade-off steps, creatively eliminates the masking effect of secondary objectives on primary benefits in single-stage normalization, significantly improving scheme differentiation. All calculations are automatically performed by the rules engine, outputting a PDF report for archiving.

[0138] In some specific embodiments,

[0139] The four types of functional nodes in the S10 multidimensional objective function topology network are as follows:

[0140] Risk avoidance node: The flood control objective function minF1 is used to calculate the maximum flood frequency that the remaining flood control reservoir capacity corresponding to the drought limit water level during the flood season can cope with;

[0141] Benefit nodes include water supply objective function maxF2, ecological objective function maxF3, power generation objective function maxF4, and navigation objective function maxF5. F2 and F3 are quantified as flow guarantee rate, and F4 and F5 are quantified as flow guarantee rate, power generation and navigation guarantee rate, respectively.

[0142] Project effectiveness milestone: The water storage objective function maxF6 is used to calculate the average maximum water storage level over a multi-year water storage period;

[0143] Management cost node: The objective function minF7 for the number of activations is used to calculate the proportion of days the drought limit water level is activated;

[0144] The dependency graph is compiled using the Kahn algorithm to generate a directed acyclic graph (DAG). Topological sorting generates 19 non-redundant weight configuration vector sets. These vector sets are formed by stacking the dual-mode benefit nodes and the independent configurations of the project effectiveness nodes to create a four-level progressive structure: Level 1: one pure benefit mode; Level 2: two combinations of benefit / total benefit and risk; Level 3: four combinations of benefit / total benefit and cost; Level 4: twelve fully coupled benefit / total benefit with risk, cost, and effectiveness. The pure risk, pure cost, and risk + cost configurations lack input from the first-level basic nodes and do not have independent optimization value. They are only included in the vector set as a benchmark group for quantitatively evaluating the optimization gain after introducing the benefit / effectiveness target and do not participate in the main process solution of the progressive call sequence. In this embodiment, the main steps include determining the scheduling target and objective function, determining the drought limit water level scheduling method, constructing the drought limit water level scheduling model, constructing multi-level solution targets, solving the drought limit water level algorithm, and optimizing the results.

[0145] Scheduling target and objective function

[0146] The reservoir scheduling objectives are determined based on the reservoir scheduling tasks. Risk-related objectives are mainly flood control, while benefit-related objectives include ecology, water supply, navigation, and power generation. At the same time, the reservoir's water storage objectives and drought limit water level effectiveness objectives must also be considered.

[0147] Drought-limited water level regulation requires comprehensive consideration of the benefits and risks of reservoirs in flood control, water supply, power generation, navigation, and ecology. Furthermore, as an emergency regulation mechanism, its activation carries certain costs, and the frequency of activation is also a consideration. Reservoirs themselves have water storage targets and tasks, and water storage should also be considered as an objective. In practical applications, corresponding objective functions should be established based on specific reservoir regulation objectives.

[0148] (1) Flood control objectives

[0149] The drought limit water level during the flood season is the upper limit of temporary water storage during the flood season. When the drought limit water level is higher than the flood limit water level, the flood frequency that can be coped with is calculated based on the remaining reservoir capacity corresponding to the drought limit water level during the flood season as the objective function for flood control risk.

[0150] ;

[0151] In the formula: This represents the maximum flood frequency that the drought limit water level can handle during the flood season; To verify the reservoir capacity corresponding to the flood level or design flood level (in billions of cubic meters) 3; The discharge volume of 100 million cubic meters is required to meet flood control requirements. 3 ; The curve showing the relationship between flood volume and flood frequency; This is the reservoir water level-capacity curve.

[0152] (2) Water supply target

[0153] The water supply target considers the water demand of the reservoir's water supply targets and the corresponding reservoir discharge flow, with the flow guarantee rate as the objective function:

[0154] ;

[0155] In the formula: The water supply guarantee rate is represented by n, which represents the total number of time periods. The reservoir discharge flow rate (m³) during time period t 3 / s; The reservoir discharge flow rate (m) corresponding to meeting water supply demand during time period t. 3 / s.

[0156] (3) Ecological goals

[0157] The ecological objectives mainly examine the ecological assessment flow at key downstream ecological sections or the basic ecological discharge from reservoirs, with the ecological flow guarantee rate as the objective function.

[0158] ;

[0159] In the formula: To ensure ecological flow guarantee rate; Let m be the reservoir discharge flow rate corresponding to meeting ecological needs during time period t. 3 / s.

[0160] (4) Power generation target

[0161] The power generation target uses the amount of electricity generated as the objective function for power generation benefits.

[0162] ;

[0163] ;

[0164] In the formula: The power generation is 100 million kWh; The reservoir's power generation during period t is 100 million kWh. The acceleration due to gravity is m / s² 2 ; The power generation flow rate m during time period t of the power station 3 / s; The generator efficiency of the power station; The density of water is kg / m³ 3 ; The difference in hydropower head (m) during time period t at the power station; Let h be the power generation duration of the power station during time period t.

[0165] (5) Shipping objectives

[0166] The shipping objective function uses the upstream and downstream navigation guarantee rate, i.e., the time required to meet shipping demand levels within the shipping timeframe, as the shipping objective function.

[0167] ;

[0168] In the formula: For navigation guarantee rate; The initial water level of the reservoir during time period t is in meters. The required water level is m for navigation.

[0169] (6) Project effectiveness target (highest water level)

[0170] The water storage target uses the highest water level reached during the storage period as the objective function for water storage.

[0171] ;

[0172] In the formula: The average highest water level is represented by m; N is the total number of scheduling cycles, in years. Let be the water level at the start of the reservoir's storage period in year i, in meters. This represents the water level at the end of the i-th year of the reservoir's storage period.

[0173] (7) Management cost target (number of starts)

[0174] The proportion of days the drought-limited water level was activated was used as the utility objective function.

[0175] ;

[0176] In some specific embodiments, the three-factor threshold parameters of the Boolean logic 3-to-3 AND gate circuit in step S20 are:

[0177] The future number of days X is set to 10 days;

[0178] Designed for low water flow The flow rate is set in three phases: 13,800 m³ / s during the main flood season, 11,000 m³ / s during the water storage period, and 5,600 m³ / s during the dry season.

[0179] The flood control requirement is a flow rate Q_f of 28000 m³ / s;

[0180] The warning water level Z_y is set at 35m;

[0181] When all three elements meet the threshold conditions, the AND gate outputs a high-level activation signal; otherwise, it outputs a low-level signal to maintain normal operation.

[0182] In this embodiment, the drought-limited water level scheduling method

[0183] The drought limit water level serves as a primary condition for initiating reservoir drought relief operations and a reference indicator for guiding these operations. When the reservoir's expected inflow remains consistently low, reaching the designed low-water level, and the reservoir level approaches or reaches the drought limit water level, downstream water demand will be affected. In such cases, drought limit water level operation should be initiated. The initiation conditions for drought limit water level operation must comprehensively consider upstream inflow, reservoir level, and downstream drought conditions. If these conditions are not met, only one or two conditions may be used as the activation target. When the drought limit water level operation activation conditions are met, the reservoir's basic water demand is used as the water supply target for water supply operations.

[0184] Conditions for activating drought-limited water level regulation: The current reservoir water level is lower than the drought-limited water level Z. hx，t The forecast indicates that the water inflow will be lower than the designed low flow rate Q within the next X days. h And the flow rate is less than the flood control requirement Q. f Downstream key section water level Z sh，t The water level will be below the warning level Z in the next X days. y .

[0185] Drought-limited water level control method: After the activation conditions are met, the outflow from the reservoir will be regulated to raise the water level to the drought-limited water level as much as possible. If the outflow is less than the current water intake demand, the outflow will be adjusted according to the current water intake demand.

[0186] Construction of a reservoir scheduling model considering drought-limited water levels

[0187] Based on the reservoir scheduling objectives and procedures, the reservoir scheduling process needs to meet multiple constraints, and a reservoir scheduling model is established accordingly.

[0188] The constraints are as follows:

[0189] In some specific embodiments, the water balance equation in step S30 is specifically as follows: Where f2() is the water level-storage capacity curve, f2'() is its inverse function, θ is the flow-water conversion coefficient, and Δt is the time period length. This refers to the amount of loss due to evaporation and leakage. The water level at the end of time period t in the reservoir. The water level of the reservoir during time period t is in meters. Let m be the inflow rate of the reservoir during time period t. 3 / s; The outbound flow rate m during time period t 3 / s;

[0190] The system of engineering constraint inequalities specifically includes five types of constraints:

[0191] Discharge capacity constraints: In the formula: , Let m represent the minimum and maximum discharge rates of the reservoir in the t-th time period, respectively. 3 / s, determined by downstream flood control, maximum discharge capacity, and reservoir water level fluctuations;

[0192] Power generation flow constraints: In the formula: For power generation flow rate m 3 / s; The maximum power generation flow rate m at time t 3 / s, determined by the unit's flow capacity corresponding to the water level at time t;

[0193] Unit output constraints: In the formula: This represents the reservoir's output in kW during time period t. , The minimum and maximum output limits for time period t are in kW, and the minimum and maximum output limits are determined according to the dispatching procedures.

[0194] Reservoir water level constraints: In the formula: and The minimum and maximum water levels m in time period t are respectively determined based on the control water level at that time.

[0195] Daily water level fluctuation constraints: In the formula: The upper limit of the reservoir water level fluctuation during time period t, in meters;

[0196] In step S30, the flow phase space trajectory mapping maps the reservoir state variables (water level, reservoir capacity, outflow) to a three-dimensional phase space. The feasibility of the scheduling scheme is predicted by the trajectory curvature and flow field stability analysis. If the feasibility is lower than the threshold, the pruning mechanism in step S40 is triggered to remove the branch of the scheme in advance.

[0197] In some specific embodiments, the topological sorting generation process in step S40 includes:

[0198] Step a: Extract the loop structure from the dependency graph in step S10;

[0199] Step b: Introduce virtual zero nodes to relax constraints in the subgraph containing loops;

[0200] Step c: Perform topological sorting on the relaxed directed acyclic graph using the Kahn algorithm;

[0201] Step d: Map the sorted node sequence to 19 weight configuration vectors;

[0202] The 19 weight configuration vector sets are specifically structured in a four-level progressive manner:

[0203] Level 1: 1 type, weight vector W1=[0, 0, 1, 1, 1, 1, 0];

[0204] Level 2: 2 types, with weight vectors W2-W3 based on Level 1, W1 is scaled or shifted to introduce risk weight 1;

[0205] Level 3: 4 types, with weight vectors W4-W7 further introducing a management cost weight of 1 based on Level 1 and Level 2;

[0206] Level 4: 12 types, weight vector W8-W 19 Based on the first three levels, an engineering effectiveness weight of 1 is introduced, and the scaling or translation weight vector W1 is subtracted. Then, similar terms are merged, resulting in four types, with the weight vector W8-W. 19 Based on the first three levels, an engineering effectiveness weight of 1 is introduced, comprising 8 types;

[0207] All vectors are normalized to form a non-redundant weight configuration vector set;

[0208] The execution logic of the S40 level pruning mechanism is as follows:

[0209] After solving for the non-dominated solution set corresponding to the first-level weight vector W1, the distribution density of the solution set in the two-dimensional projection space of flood control and water storage is analyzed.

[0210] If the projection density is lower than the preset threshold, all vectors W4 and W5 in the second level that do not contain flood control or water storage weights are determined to be invalid branches and are directly removed without further solution.

[0211] The pruning operation allows subsequent calculations to skip at least 5 weight vectors, thus optimizing the allocation of computational resources.

[0212] Step S40, the hardware fuse mechanism, is implemented in the following way:

[0213] During the system power-on initialization phase, 19 weight configuration vectors are loaded from flash memory into the FPGA's block random access memory;

[0214] After loading is complete, the write enable pin of the configuration register is fixed, physically cutting off the external modification path;

[0215] During operation, the weight vector is used as a read-only constant in the parallel solution of step S50. Any illegal modification instruction will trigger a hardware watchdog reset.

[0216] In some specific embodiments, the specific rules of the adaptive parameter dynamic calibration mechanism in step S50 are as follows:

[0217] When the sum of the dimensions of the four objective functions is ≤10, the population size is set to 200 and the number of iterations is set to 500.

[0218] For every 5 increases in the target dimension, the population size increases by 50, and the number of iterations increases by 100.

[0219] When the number of remaining weight vectors after the current stage pruning is ≤5, the lower limit of the number of iterations is reduced to 300 to accelerate convergence.

[0220] In some specific embodiments, the first-level processing of the secondary normalization compensation mechanism in step S60 is as follows:

[0221] The four sub-objectives of water supply, ecology, power generation, and shipping are normalized to the [0,1] interval using the range transformation method.

[0222] The inverse range transformation method is used for the sub-objectives of flood control and activation cost (the smaller the value, the higher the score);

[0223] The project's effectiveness target is achieved by normalizing the average maximum water level over multiple years to [0,1] using the range transformation method.

[0224] In some specific embodiments, the second-level normalization process in step S60 is as follows:

[0225] The normalized values ​​of water supply, ecology, power generation, and shipping output from the first level are weighted at 0.25:0.25:0.25:0.25 to obtain the comprehensive benefit index B;

[0226] The comprehensive effect index C is obtained by weighting B again with the normalized flood control, water storage, and activation costs by 0.25:0.25:0.25:0.25;

[0227] The two-level weight allocation structure creates a compensatory effect. When the single-level normalization causes a conflict between the dimensions of the benefit target and the cost target, the second level recalibrates the scale to eliminate the conflict.

[0228] In some specific embodiments, the confidence assessment report of step S60 includes three main visualization components:

[0229] A radar chart comparing the various options shows the achievement status of the 19 options on the seven objectives.

[0230] Pareto front projection plot shows the distribution of the non-dominated solution set in the benefit-risk two-dimensional space;

[0231] The confidence interval diagram of the drought limit water level process is given, along with the water level process line and 90% confidence band of the optimal solution.

[0232] In this embodiment, a multi-level target scheme is constructed.

[0233] Taking into account the benefits and risks of reservoir drought-limited water level regulation, while also considering the disruption to existing drought relief regulations and maximizing the original functions of the reservoir, this paper proposes a principle of balancing benefits, risks, costs, and effectiveness. It employs a tiered, multi-faceted approach to design a multi-level objective optimization process and eliminates ineffective objectives. In practical application, if a particular objective is especially important, only optimization objectives containing that objective can be selected.

[0234] (1) First level, considering the combination of objectives with comprehensive benefits:

[0235]

[0236] (2) The second level is a combination of objectives that takes into account both comprehensive benefits and risks.

[0237]

[0238] (3) The third level is a combination of objectives that takes into account comprehensive benefits, risks and management costs.

[0239]

[0240] (4) Level 4: A combination of objectives that take into account benefits, risks, management costs, and project effectiveness.

[0241]

[0242] (5) Meaningless goals

[0243]

[0244] Solving and optimizing the drought limit water level for a multi-level objective combination

[0245] (1) Decision variables: The drought limit water level at key nodes is used as the decision variable, and the drought limit water level is calculated by interpolation.

[0246] (2) Solution process: The drought limit water level is input into the scheduling model as a decision variable, and the outflow, power generation and end-of-period water level are output. The objective function is calculated based on the end-of-period water level. Intelligent optimization algorithms such as genetic algorithm and particle swarm optimization algorithm are used to construct a multi-level solution objective with a multi-level objective scheme as the optimization objective. The non-dominated solution set of each objective is obtained by setting a reasonable population number and number of iterations.

[0247] (3) Comparison indexes and methods: The objective function values ​​in the non-dominated solution set are normalized, and the optimal solution under the current objective combination is determined by the weight combination in the scheme setting in "multi-level objective scheme construction".

[0248] The drought limit water level obtained from solving the multi-level objectives is used as the solution set. Benefits, risks, costs, and project effectiveness indicators are evaluated sequentially, namely power generation, navigation, water supply, ecology, flood control, activation of the drought limit water level, and the highest water level during the storage period. These indicators are then normalized. Power generation, navigation, water supply, and ecology are summed with equal weights of 1:1:1:1 to form the benefit indicator. After calculating the benefit indicator, it is normalized again and summed with equal weights of 1:1:1:1 for benefits, risks, costs, and project effectiveness. The comprehensive effect indicator is calculated, and the scheme with the largest result is the recommended scheme.

[0249] If there are special needs, such as when a certain goal is particularly important, the weight can be calculated by referring to subjective weight calculation methods such as expert scoring.

[0250] Example: Taking a reservoir as an example, the functions of the reservoir are flood control, power generation, navigation, ecology and water supply.

[0251] Scheduling target and objective function

[0252] Based on the reservoir's functions, and considering the effectiveness of the drought limit water level and the project's achievement of objectives, the reservoir's functional objectives are flood control, power generation, navigation, ecology, and water supply. The management cost objective is the number of times the drought limit water level is activated, and the project's achievement of objectives is the highest water level during the multi-year average storage period.

[0253] (1) Flood control objective: In order to ensure downstream flood control needs, the downstream discharge flow shall not exceed 32,700 m³ within 7 days. 3 Under the scenario of / s, the frequency of floods that the reservoir's drought limit water level can handle is shown below. The flood control objective function is calculated based on the drought limit water level during the flood season.

[0254]

[0255] (2) Water supply target

[0256] To ensure water supply for downstream cities and irrigation (April-September), the reservoir outflow is required to be at least 7000 m³ / h. 3 / s and 8500m 3 / s, calculate the water supply guarantee rate based on the daily outflow from the reservoir output by the model.

[0257] (3) Ecological objectives: It is known that the target for ecological discharge at key downstream sections is 5700m. 3 / s, calculate the ecological flow guarantee rate based on the daily outflow from the reservoir output by the model.

[0258] (4) Power generation target: The total power generation is calculated by summing the daily power generation output from the model.

[0259] (5) Shipping objectives: The upstream shipping requirement is that the water level is not lower than 260m during the non-flood season. The navigation guarantee rate is calculated based on the daily water level output by the model.

[0260] (6) Project effectiveness target: Calculate the highest water level during the multi-year average water storage period based on the water level output by the model.

[0261] (7) Management cost target: The number of times the model is activated is directly output by the model.

[0262] Dispatch method: Drought limit water level dispatch activation conditions: The current reservoir water level is lower than the drought limit water level Zhx,t; the forecast for the next X(10) days is that the inflow will be lower than the design low flow rate Qh (13800m3 / s during the flood season, 11000m3 / s during the storage period). 3 / s, dry season 5600m3 / s) and less than the flood control requirement flow Qf (28000m3 / s); downstream key section water level Zsh, t will be lower than the warning water level Zy (35m) in the next X (10) days. Drought limit water level scheduling method: after the activation conditions are met, the reservoir discharge flow is regulated to raise the water level to the drought limit water level as much as possible. If the outflow is less than the current water demand, the discharge will be based on the current water demand, that is: 8500m from April to September. 3 / s, 7000m at other times 3 / s.

[0263] Model construction: based on scheduling procedures and Figure 1 The scheduling process is used to construct a scheduling model.

[0264] Multi-level solution objective construction: Based on the multi-level objective combination in "Multi-level objective scheme construction", the following objectives are constructed.

[0265]

[0266] The above objective was solved, and the results are as follows: Figure 3 As shown in the table below, the scheduling effects of each scheme are calculated using the "scheduling objective and objective function" evaluation index.

[0267]

[0268] The scheduling effects of each scheme are calculated based on the comprehensive effect evaluation index of "multi-level target combination drought limit water level solution and optimization". The table below shows the scheduling effects of each scheme.

[0269]

[0270] Reference Figure 4 Option 17 yields the optimal result, representing the optimization with the objectives of flood control, water storage, and activation. The specific results are as follows:

[0271]

[0272] A smart optimization decision-making system for reservoir drought limit water level, based on the same concept, is deployed in the industrial control network of the reservoir dispatch center, including:

[0273] The objective function topology network construction module is used to execute step S10 and output the objective dependency graph.

[0274] The edge-triggered rule engine module is used to execute step S20 and has built-in Boolean logic AND gate circuits;

[0275] The digital twin model calculation module is used to execute step S30 and integrates a water balance state equation solver.

[0276] The weight configuration vector management module is used to execute step S40, which implements hardware circuit breaking through FPGA;

[0277] The parallel optimization solution module is used to execute step S50, and adopts adaptive parameter dynamic calibration.

[0278] The solution evaluation report generation module is used to execute step S60 and output the optimized solution with confidence level;

[0279] The industrial data interface unit is configured to communicate bidirectionally with the reservoir SCADA system. When the result of the three-element Boolean logic operation in step S20 is true, it automatically sends a flow control command to SCADA and receives execution feedback to form a closed-loop control.

[0280] The weight configuration vector management module uses Xilinx Zynq series FPGA. The 19 weight vectors are fixed in the form of IP cores. After the system is powered on, they are loaded into the DDR memory of the optimization solution module within 100ms through the AXI DMA interface. After loading is completed, the fuse bit is set, and any subsequent rewrite operation will trigger a hardware abnormal interrupt.

[0281] The parallel optimization solution module is deployed on a GPU-accelerated computing server, supporting the concurrent operation of 19 NSGA-II algorithm instances corresponding to weight vectors. The instances share digital twin model data through the PCIe bus, and the concurrent solution efficiency is 10-15 times higher than that of serial solution.

[0282] This invention proposes an intelligent optimization decision-making system for reservoir drought limit water levels, deployed within the industrial control network of the reservoir dispatch center, aiming to achieve efficient and intelligent dispatching decisions for drought limit water levels. The system consists of multiple functional modules, which work collaboratively to complete the entire optimization process from target construction to scheme evaluation.

[0283] Objective Function Topology Network Construction Module: This module executes step S10 to construct a multi-dimensional objective function topology network. Network nodes include four types of functional nodes: risk avoidance objectives, benefit objectives, engineering effectiveness objectives, and management cost objectives. Nodes are connected by directed edges to form an objective dependency graph, clarifying the logical relationships between objectives. The objective dependency graph output by this module serves as the input basis for subsequent steps, providing a structural framework for compiling the scheduling period evaluation function group. Marginal Triggering Rule Engine Module: Executes step S20 and incorporates a Boolean logic 3-to-3 AND gate circuit based on engineering reliability requirements. This module receives three factors as input: current water level status, future inflow forecast, and downstream water demand forecast. The AND gate circuit outputs a high-level activation signal, triggering subsequent modules to start calculations, only when all three factors simultaneously meet preset threshold conditions. This module ensures the timeliness and reliability of scheduling decisions, avoiding false triggering by a single condition. Digital Twin Model Calculation Module: Executes step S30 and integrates a water balance state equation solver. This module maps the decision variable domain to the water level-flow phase space trajectory during the scheduling period by simultaneously solving the water balance equation and the set of engineering constraint inequalities. The model employs a simultaneous solution strategy to ensure coupled simulation of state and control variables in the time domain, providing an accurate water level-flow process for subsequent optimization. The weight configuration vector management module executes step S40, using a Xilinx Zynq series FPGA to implement hardware circuit breaking. Nineteen weight configuration vectors are embedded in the FPGA as IP cores and are rapidly loaded into the optimization module's DDR memory within 100ms via the AXI DMA interface after system power-on. After loading, the circuit breaker bit is set; any attempt to rewrite the weight vectors will trigger a hardware exception interrupt, ensuring the safety and stability of the weight configuration during system operation. The parallel optimization module executes step S50 and is deployed on a GPU-accelerated computing server. This module supports the concurrent execution of NSGA-II algorithm instances corresponding to the 19 weight vectors, with each instance sharing digital twin model data via the PCIe bus. Concurrent solution efficiency is 10-15 times higher than traditional serial methods, significantly shortening computation time and meeting the real-time requirements of reservoir scheduling. The scheme evaluation report generation module executes step S60 and outputs optimized schemes with confidence levels. This module performs a two-level normalization compensation mechanism on all non-dominated solution sets, eliminating the target masking effect caused by single-level normalization and enhancing scheme distinguishability. Finally, it outputs the drought limit water level scheduling scheme corresponding to the optimal solution and its confidence evaluation report, providing a scientific basis for scheduling decisions. The industrial data interface unit is configured for bidirectional communication with the reservoir SCADA system. When the three-factor Boolean logic operation result output by module S20 is true, this unit automatically sends flow control commands to the SCADA system and receives execution feedback, forming a closed-loop control. This unit ensures that system decisions can be applied to actual scheduling operations in real time and accurately, achieving intelligent scheduling control.Through the collaborative work of its various modules, this system optimizes the entire process from target setting to scheme evaluation, significantly improving the efficiency and reliability of reservoir drought limit water level scheduling and providing an intelligent solution for reservoir scheduling and management.

[0284] By adopting the above-disclosed technical solution of this invention, the following beneficial effects are obtained:

[0285] This invention achieves efficient, accurate, and safe reservoir drought-limit water level scheduling by constructing an intelligent optimization decision-making system. The system employs hierarchical pruning and GPU parallel acceleration, significantly improving computational efficiency and reducing solution time to 1 / 10 to 1 / 15 of traditional methods, meeting real-time scheduling requirements. A two-level normalization compensation mechanism eliminates the target occlusion effect, enhances scheme differentiation, and ensures accurate and reliable decision-making. An FPGA hardware fuse mechanism solidifies weight configurations, providing industrial-grade security and eliminating the risk of parameter tampering. Multi-dimensional objective function topology networks and topology sorting compilation technology achieve full-objective collaborative optimization, strengthening the overall coordination of the scheduling scheme. Boolean logic 3-to-3 AND gate circuits improve engineering reliability, ensuring timely and accurate execution of scheduling instructions. This system is applicable to various types of comprehensive utilization reservoirs, offering high deployment economics and significant benefits, and possesses broad engineering application value and promotion prospects. The above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the invention, and these improvements and modifications should also be considered within the scope of protection of the invention.

Claims

1. A multi-objective hierarchical intelligent optimization decision-making method for reservoir drought limit water level, deployed in a reservoir scheduling automation system, characterized in that, This includes the following logically tightly coupled and sequentially irreversible steps: S10: Construct a multi-dimensional objective function topology network for drought-limited water level scheduling. The network nodes include at least four types of functional nodes: risk avoidance objective, benefit objective, project effectiveness objective, and management cost objective. Each node is connected by directed edges to form an objective dependency graph. The dependency graph is then compiled into a scheduling period evaluation function set that can be computed in parallel. S20: Establish a multi-element marginal triggering rule for drought-limited water level scheduling. The rule uses the current water level status, future water inflow forecast, and downstream water demand forecast as three elements to perform Boolean logic three-to-three AND gate circuit operation based on engineering reliability requirements, and outputs a unique activation signal for automatic triggering control command. S30: Create a digital twin model of reservoir scheduling that takes into account the activation signal. The model maps the decision variable domain to the water level-flow phase space trajectory during the scheduling period by solving the water balance state equation and the set of engineering constraint inequalities simultaneously. S40: In the system rule engine, 19 non-redundant weight configuration vector sets are pre-formed, which are generated by topological sorting of four types of target nodes: benefit, risk, cost, and effectiveness. The vector sets form a progressive calling sequence based on the solution space hierarchical pruning mechanism. The calling sequence achieves optimized allocation of computing resources through invalid solution branch prediction and pruning. The weight configuration vectors are locked in a read-only state through a hardware circuit breaker mechanism during the system operation cycle. S50: Using the key node drought limit water level sequence as the decision variable, call the weight configuration vectors in the vector set, and use a multi-objective optimization algorithm with an adaptive parameter dynamic calibration mechanism to solve in parallel to obtain the corresponding non-dominated solution set. The algorithm population size and number of iterations are adjusted in real time according to the objective function dimension, with a lower limit of 200 for the population size and a lower limit of 500 for the number of iterations. S60: Perform a two-level normalization compensation mechanism to eliminate dimensional differences on all non-dominated solution sets. The compensation mechanism eliminates the target occlusion effect caused by single-level normalization and enhances the distinguishability of the schemes through the serial processing of the first-level sub-target normalization and the second-level comprehensive benefit normalization. Based on the priority of the calling sequence of the hierarchical pruning mechanism, the comprehensive effect index is automatically calculated, and the drought limit water level scheduling scheme corresponding to the optimal solution of the index and its confidence assessment report are output.

2. The method according to claim 1, characterized in that, The four types of functional nodes in the S10 multidimensional objective function topology network are as follows: Risk avoidance node: The flood control objective function minF1 is used to calculate the maximum flood frequency that the remaining flood control reservoir capacity can cope with at the drought limit water level during the flood season; its calculation formula is: In the formula: This represents the maximum flood frequency that the drought limit water level can handle during the flood season; To verify the reservoir capacity (in billions of cubic meters) corresponding to the flood level or design flood level. 3 ) To meet flood control requirements, the discharge volume (100 million m³) 3 ); The curve showing the relationship between flood volume and flood frequency; This is the reservoir water level-capacity curve; Beneficial benefit nodes include water supply objective function maxF2, ecological objective function maxF3, power generation objective function maxF4, and shipping objective function maxF5; The formula for the target flow guarantee rate of water supply is: ; In the formula: The water supply guarantee rate is represented by n, which represents the total number of time periods. The reservoir discharge flow rate (m³) during time period t 3 / s); The reservoir discharge flow (m³) corresponding to meeting water supply demand during time period t. 3 / s); The formula for calculating the flow guarantee rate for ecological targets is: ; In the formula: To ensure ecological flow guarantee rate; Let m be the reservoir discharge flow rate that meets ecological needs during time period t. 3 / s; The formula for calculating the total power generation target is as follows: ； ； In the formula: Electricity generation (100 million kWh); The power generation of the reservoir during time period t is 100 million kWh. Acceleration due to gravity (m / s²) 2 ); The power generation flow rate (m³) of the power plant during time period t 3 / s); The generator efficiency of the power station; The density of water (kg / m³) 3 ); The difference in hydropower head (m) during time period t at the power station. The power generation duration (h) of the power station during time period t; The formula for calculating the navigation guarantee rate for shipping targets is: ; In the formula: For navigation guarantee rate; The water level (m) of the reservoir during time period t; Water level (m) required for navigation; Project effectiveness milestone: The water storage objective function maxF6 is used to calculate the average maximum water storage level over a multi-year water storage period; ； In the formula: The average highest water level is (m); N is the total number of scheduling cycles (years). The water level (m) at the start of the reservoir's storage period in the i-th year; This represents the water level at the end of the reservoir's storage period in year i. Management cost node: The objective function minF7 for the number of activations is used to calculate the proportion of days the drought limit water level is activated; its calculation formula is: ; The dependency graph is compiled using the Kahn algorithm to generate a directed acyclic graph. Topological sorting generates 19 non-redundant weight configuration vector sets. These vector sets are formed by stacking the dual-mode benefit nodes and the independent configurations of the project effectiveness nodes to form a four-level progressive structure: Level 1: 1 pure benefit mode; Level 2: 2 benefit / total benefit and risk combinations; Level 3: 4 benefit / total benefit and cost combinations; Level 4: 12 benefit / total benefit fully coupled with risk, cost, and effectiveness. Among them, the pure risk, pure cost, and risk + cost configurations lack the input of the first-level basic nodes and do not have independent optimization value. They are only included in the vector set as a reference group for quantitative evaluation of the optimization gain after introducing the benefit / effectiveness target and do not participate in the main process solution of the progressive call sequence.

3. The method according to claim 1, characterized in that, The three threshold parameters for the Boolean logic 3-to-3 AND gate circuit in step S20 are: The future number of days X is set to 10 days; Design low water flow Q h The flow rate is set in three phases: 13,800 m³ / s during the main flood season, 11,000 m³ / s during the water storage period, and 5,600 m³ / s during the dry season. The flood control requirement is a flow rate Q_f of 28000 m³ / s; The warning water level Z_y is set at 35m; When all three elements meet the threshold conditions, the AND gate outputs a high-level activation signal; otherwise, it outputs a low-level signal to maintain normal operation.

4. The method according to claim 1, characterized in that, The water balance equation in step S30 is as follows: Where f2() is the water level-storage capacity curve, f2'() is its inverse function, θ is the flow-water conversion coefficient, and Δt is the time period length. This refers to the amount of loss due to evaporation and leakage. The water level at the end of time period t in the reservoir. The initial water level of the reservoir at time t (m); The inflow rate of the reservoir during time period t (m³) 3 / s); The outbound flow rate (m³) during time period t 3 / s); The set of engineering constraint inequalities specifically includes five types of constraints: Discharge capacity constraints: In the formula: , Represent the minimum and maximum discharge rates (m³) of the reservoir in the t-th time period, respectively. 3 The value is determined based on downstream flood control, maximum discharge capacity, and reservoir water level fluctuations. Power generation flow constraints: In the formula: Power generation flow (m 3 / s); The maximum power generation flow rate at time t (m³) 3 / s), determined by the unit's flow capacity corresponding to the water level at time t; Unit output constraints: In the formula: This represents the power output (kW) of the reservoir during time period t. , The minimum and maximum output limits (kW) for time period t are determined according to the dispatching procedures. Reservoir water level constraints: In the formula: and These are the minimum and maximum water levels (m) for time period t, determined based on the control water level at that time. Daily water level fluctuation constraints: In the formula: The upper limit of the reservoir water level fluctuation during time period t (m); In step S30, the flow phase space trajectory mapping maps the reservoir state variables to a three-dimensional phase space. The feasibility of the scheduling scheme is predicted by analyzing the trajectory curvature and flow field stability. If the feasibility is lower than the threshold, the pruning mechanism in step S40 is triggered to remove the branch of the scheme in advance.

5. The method according to claim 1, characterized in that, The topological sorting generation process in step S40 includes: Step a: Extract the loop structure from the dependency graph in step S10; Step b: Introduce virtual zero nodes to relax constraints in the subgraph containing loops; Step c: Perform topological sorting on the relaxed directed acyclic graph using the Kahn algorithm; Step d: Map the sorted node sequence to 19 weight configuration vectors; The 19 weight configuration vector sets are specifically a four-level progressive structure: Level 1: 1 type, weight vector W1=[0, 0, 1, 1, 1, 1, 0]; Level 2: 2 types, with the weight vector W2-W3 based on Level 1, W1 is scaled or shifted to introduce risk weight 1; Level 3: 4 types, with weight vectors W4-W7 further introducing a management cost weight of 1 based on Level 1 and Level 2; Level 4: 12 types, weight vector W8-W 19 Based on the first three levels, an engineering effectiveness weight of 1 is introduced, and the scaling or translation weight vector W1 is subtracted. Then, like terms are merged, resulting in four weight vectors W8-W. 19 Based on the first three levels, an engineering effectiveness weight of 1 is introduced, comprising 8 types; All vectors, after being normalized, constitute the non-redundant weight configuration vector set. The execution logic of the hierarchical pruning mechanism in step S40 is as follows: After solving for the non-dominated solution set corresponding to the first-level weight vector W1, the distribution density of the solution set in the two-dimensional projection space of flood control and water storage is analyzed. If the projection density is lower than the preset threshold, all vectors W4 and W5 in the second level that do not contain flood control or water storage weights are determined to be invalid branches and are directly removed without further solution. The pruning operation causes subsequent calculations to skip at least 5 weight vectors, thereby optimizing the allocation of computing resources. The hardware circuit breaker mechanism in step S40 is implemented in the following way: During the system power-on initialization phase, 19 weight configuration vectors are loaded from flash memory into the FPGA's block random access memory; After loading is complete, the write enable pin of the configuration register is fixed, physically cutting off the external modification path; During operation, the weight vector is used as a read-only constant in the parallel solution of S50. Any illegal modification instruction will trigger a hardware watchdog reset.

6. The method according to claim 1, characterized in that, The specific rules of the adaptive parameter dynamic calibration mechanism in step S50 are as follows: When the sum of the dimensions of the four objective functions is ≤10, the population size is set to 200 and the number of iterations is set to 500. For every 5 increases in the target dimension, the population size increases by 50, and the number of iterations increases by 100. When the number of remaining weight vectors after the current stage pruning is ≤5, the lower limit of the number of iterations is reduced to 300 to accelerate convergence.

7. The method according to claim 1, characterized in that, The first-level processing of the secondary normalization compensation mechanism in step S60 is as follows: The four sub-objectives of water supply, ecology, power generation, and shipping are normalized to the [0,1] interval using the range transformation method. The inverse range transformation method is used for the sub-objectives of flood control and activation cost; The project's effectiveness target is achieved by normalizing the average maximum water level over multiple years to [0,1] using the range transformation method.

8. The method according to claim 7, characterized in that, The second-level normalization process in step S60 is as follows: The normalized values ​​of water supply, ecology, power generation, and shipping output from the first level are weighted at 0.25:0.25:0.25:0.25 to obtain the comprehensive benefit index B; The comprehensive effect index C is obtained by weighting B again with the normalized flood control, water storage, and activation costs by 0.25:0.25:0.25:0.25; The two-level weight allocation structure creates a compensation effect. When the single-level normalization causes a conflict between the dimensions of the benefit target and the cost target, the second level recalibrates the scale to eliminate the conflict.

9. The method according to claim 1, characterized in that, The confidence assessment report for step S60 includes three main visualization components: A radar chart comparing the various options shows the achievement status of the 19 options on the seven objectives. Pareto front projection plot shows the distribution of the non-dominated solution set in the benefit-risk two-dimensional space; The confidence interval diagram of the drought limit water level process is given, along with the water level process line and 90% confidence band of the optimal solution.

10. A smart optimization decision-making system for reservoir drought limit water level, characterized in that, The system is deployed in the industrial control network of the reservoir dispatch center, and includes: The objective function topology network construction module is used to execute step S10 and output the objective dependency graph. The edge-triggered rule engine module is used to execute step S20 and has built-in Boolean logic AND gate circuits; The digital twin model calculation module is used to execute step S30 and integrates a water balance state equation solver. The weight configuration vector management module is used to execute step S40, which implements hardware circuit breaking through FPGA; The parallel optimization solution module is used to execute step S50, and adopts adaptive parameter dynamic calibration. The solution evaluation report generation module is used to execute step S60 and output the optimized solution with confidence level; The industrial data interface unit is configured to communicate bidirectionally with the reservoir SCADA system. When the result of the three-element Boolean logic operation in step S20 is true, it automatically sends a flow control command to SCADA and receives execution feedback to form a closed-loop control. The weight configuration vector management module uses a Xilinx Zynq series FPGA. The 19 weight vectors are fixed in the form of IP cores. After the system is powered on, they are loaded into the DDR memory of the optimization solution module within 100ms through the AXI DMA interface. After loading is completed, the fuse bit is set, and any subsequent rewrite operation will trigger a hardware abnormal interrupt. The parallel optimization solution module is deployed on a GPU-accelerated computing server, supporting the concurrent operation of 19 NSGA-II algorithm instances corresponding to weight vectors. Each instance shares digital twin model data through the PCIe bus, and the concurrent solution efficiency is 10-15 times higher than that of serial solution.

Images