A water resource allocation analysis and decision-making method based on multi-objective collaborative optimization

Abstract

The application discloses a water resource allocation analysis and decision-making method based on multi-target cooperative optimization, belongs to the technical field of multi-target optimization based on machine learning, and comprises the following steps: collecting multi-source heterogeneous water resource data for pretreatment; based on a federal pulse learning framework, a hybrid mechanism network model is constructed, pulse threshold coding is performed on multi-source heterogeneous data tensors, and sparse excitation-lateral inhibition double-loop feature extraction and pulse consistency weighted federal aggregation are performed; a multi-factor optimization mechanism is introduced, a knowledge transfer enhanced multi-target snake vulture optimization algorithm is improved, and daily water demand prediction values are subjected to multi-source multi-user cooperative allocation optimization; a multi-scenario allocation scheme set is objectively weighted and closeness degree sorting is performed; and the allocation schemes at the forefront in each scenario are decomposed and coded and are sent to an executing mechanism for water quantity control. The application aims to realize intelligent generation of multi-target optimization allocation schemes and provide efficient optimization schemes for water resource allocation.

Description

Technical Field

[0001] This invention belongs to the field of multi-objective optimization technology based on machine learning, and particularly relates to a water resource allocation analysis and decision-making method based on multi-objective collaborative optimization. Background Technology

[0002] Currently, my country's water resource allocation industry has shifted from the traditional engineering-based water supply allocation model to a comprehensive allocation and governance model that emphasizes "safety assurance, efficient utilization, ecological priority, and coordinated balance." Application scenarios cover all areas, including basin-level water resource coordination and allocation, cross-regional water transfer project allocation, urban water supply and management, irrigation district agricultural water-saving allocation, and industrial water recycling optimization. Driven by national policies for smart water management, intelligent allocation decision-making technologies are gradually being implemented, and the industry as a whole is rapidly developing towards digital modeling, dynamic sensing, intelligent decision-making, and precise allocation.

[0003] Currently, existing analysis and decision-making technologies still have unresolved technical bottlenecks: 1. The use of continuous value models results in dense computational redundancy, leading to insufficient multi-scale feature capture and centralized training causing data privacy barriers; 2. Existing optimization algorithms have an imbalance between global exploration and local development when solving high-dimensional constrained water resource allocation, resulting in the discarding of infeasible solution information and the inability to reuse cross-task experience. Summary of the Invention

[0004] This invention discloses a water resource allocation analysis and decision-making method based on multi-objective collaborative optimization. By deeply modifying the core parameters of the optimization mechanism, introducing new mechanisms, and applying hybrid operators, it intelligently generates multi-objective optimization allocation schemes to solve the aforementioned technical bottleneck problems. Therefore, the present invention provides the following technical solution: This invention discloses a water resource allocation analysis and decision-making method based on multi-objective collaborative optimization, specifically including: S1, collecting multi-source heterogeneous water resource data and preprocessing it to obtain a standardized multi-source heterogeneous data tensor; S2, based on a federated impulse learning framework, using a dual-loop sparse impulse attention and adaptive LIF spiking neural network model to perform impulse threshold encoding on the multi-source heterogeneous data tensor, as well as sparse excitation-lateral inhibition dual-loop feature extraction and impulse consistency weighted federated aggregation, outputting the daily water demand prediction value for each water-using unit; S3, using a hemispherical prey escape energy factor to drive... A dynamic dual-mode displacement search method is used, which integrates the constraint-improved incentive factor CII and the knowledge transfer-enhanced multi-objective guilder optimization algorithm with cross-task knowledge capsule dual-layer archive maintenance. This method optimizes the daily water demand forecast by multi-source and multi-user collaborative configuration, resulting in a Pareto configuration scheme set. S4: The Pareto configuration scheme set for multiple scenarios is objectively weighted and ranked by proximity to obtain the top 3 configuration schemes in each scenario. S5: The top 3 configuration schemes in each scenario are decomposed and encoded at the instruction granularity and issued to the execution mechanism for water control.

[0005] Furthermore, its S1 includes: S101. Collect and combine data from sensor equipment, hydrology, meteorology and water use statistics obtained from monitoring stations as raw heterogeneous data; S102. Perform missing value imputation, standardization, and spatiotemporal alignment on the original heterogeneous data to obtain a standardized multi-source heterogeneous data tensor.

[0006] Furthermore, its S2 includes: S201. Using an adaptive pulse threshold coding method based on local variance, the multi-source data tensor is dynamically binarized to generate a sparse binary pulse sequence. S202. Using the grouped projection and sparse masking attention method, only the attention weights of bidirectional pulse time step pairs are calculated for the sparse binary pulse sequence to obtain the attention features of the feedforward excitation loop. S203. Calculate the L2 norm of the attention feature of the feedforward excitation circuit as the activation intensity. The formula for calculating the activation intensity is as follows: In the formula, is Let L2 be the norm of the attention output matrix of the feedforward excitation loop; Let be the attention output value of the c-th channel at time step t; T is the time step number; t is the time index; the competition intensity score of each channel is obtained through softmax normalization, and the output feature of the lateral suppression loop is calculated by setting the suppression intensity parameter, the formula of which is: SAttn LI =SAttn FF ⊙(1-τ·C), where SAttnLI The lateral inhibition loop feature is represented by ⊙; element-wise multiplication is represented by ⊙; C is the competition intensity score vector; SAttn FF τ represents the output feature matrix of the feedforward excitation loop; τ is the inhibition strength control parameter. A fusion gating network is constructed, and the fusion weights are dynamically determined based on the features of the feedforward excitation loop and the lateral inhibition loop. After concatenating the feedforward and inhibition features, gating coefficients are generated through a linear layer and a Sigmoid activation. The gating coefficients are used as weights to weight and fuse the features of the feedforward excitation loop and the lateral inhibition loop to obtain the dual-loop attention features. The dual-loop attention features are added to the original input residuals, and layer normalization is performed to stabilize the training, resulting in the fused attention features.

[0007] Furthermore, its S2 also includes: S204. Using the Poisson pulse coding method, the fused attention features are transformed into pulse sequences and input into a four-layer adaptive time constant LIF spiking neural network for time series modeling. By replacing gradient learning and membrane potential linear decoding, the daily water demand prediction value of each water-using unit is output. At the same time, the parameters of the dual-loop sparse pulse attention and adaptive LIF spiking neural network model, as well as the pulse firing mode matrix of each layer, are recorded for federated aggregation. S205. Using a pulse matching aggregation method based on a federated pulse learning framework, pulse consistency scores are calculated for the pulse firing pattern matrices recorded in each layer. Based on the pulse consistency scores, the parameters of the dual-loop sparse pulse attention and adaptive LIF spiking neural network model are weighted and aggregated to obtain a global spiking neural network water demand prediction model, which is used to perform inference and update the daily water demand prediction values ​​of each water-using unit in real time.

[0008] Furthermore, its S3 includes: S301. Obtain the daily water demand forecast value updated in real time for each water-using unit, construct a target-constrained optimization model with the objectives of water supply guarantee rate, water shortage, economic benefits and ecological satisfaction, and use the historical knowledge capsules activated by similarity filtering in the cross-task knowledge base to extract elite solutions as seeds and inject them into the initial population to obtain a hot-start initial population and empty elite files that integrate historical experience and random diversity. S302. Construct a hemispherical prey escape energy factor for the current iteration. The formula for calculating the energy factor is as follows: E in the formula 能量 (t) represents the escape energy factor value of the hemispherical prey at generation t; E0 is the initial escape energy; t is the current iteration generation; T is the total number of iterations; γ is the decay steepness control parameter; η is the iteration progress ratio; f is the oscillation amplitude control parameter; 震荡 This is the oscillation frequency parameter; For Sigmoid decay term; It is a sinusoidal oscillation term; S303. Based on the hemispherical prey escape energy factor, adaptively switch between energy-sensing spiral search or constraint-sensing gradient descent mode. Guided by the non-dominated solutions and population mean in the current elite files, perform dual-mode displacement update on the individuals in the population to generate offspring candidate populations.

[0009] Furthermore, its S3 also includes: S304. Using the Boundary Information Potential Energy (BIP) index and the Constraint Improvement Incentive Factor (CII), the parent population and offspring candidate population are merged with the current elite archive. The resulting set is then evaluated hierarchically and the archive is maintained to obtain a new generation of population and an updated elite archive. After the iteration ends, the Pareto configuration scheme set is constructed from the non-dominated feasible solutions in the final elite archive. S305. Using the crowding distance filtering and context feature encoding method, elite solutions are extracted from the Pareto configuration scheme set and encapsulated into knowledge capsules. Redundancy management is performed through cosine similarity filtering and replacement to obtain the updated cross-task knowledge base.

[0010] Furthermore, its S4 includes: S401. Obtain the daily water demand forecast for each water-using unit, construct four configuration scenarios (abundant, normal, low, and extreme drought) using the hydrological frequency classification method, and correct the constraint parameters. For each scenario, call the knowledge transfer-enhanced multi-objective guilder optimization algorithm to solve the problem and obtain the objective function matrix corresponding to the Pareto non-dominated configuration scheme set. S402. Use the entropy weight method to perform homogenization, normalization and information entropy calculation on the objective function matrix corresponding to the Pareto non-dominated configuration scheme set to obtain the objective weight vector of each objective in the comprehensive evaluation. S403. Based on the objective weight vector, the normalized Pareto solution set is weighted. By calculating the weighted Euclidean distance between each solution and the positive and negative ideal solutions, the relative proximity is obtained and sorted in descending order. Finally, the top 3 configuration solutions with the best overall performance in each scenario are obtained.

[0011] Furthermore, its S5 includes: S501. For each scenario, calculate the top 3 configuration schemes based on overall performance to obtain the engineering optimization control instruction sequence corresponding to each scheme. S502. Distribute the engineering optimization control command sequence corresponding to each scheme to the lower execution mechanism for execution and monitor and feedback status data in real time.

[0012] Compared with the prior art, the present invention achieves at least one of the following beneficial effects: 1. By introducing a federated impulse learning framework, continuous values ​​are transformed into sparse impulse sequences using impulse thresholding. A dual-loop sparse impulse attention mechanism is designed, and a lateral inhibition loop simulates competition in the biological cortex to remove redundant feature channels. At the same time, an impulse consistency weighted aggregation strategy is used to replace federated averaging to avoid loss of impulse discrete information. This addresses the problems of dense computational redundancy in existing continuous value models, which leads to insufficient multi-scale feature capture, and centralized training which causes data privacy barriers.

[0013] 2. By constructing a hemispherical prey escape energy factor, the search behavior is dynamically regulated through a three-stage nonlinear decay. In the high-energy stage, logarithmic spiral exploration is performed to cover the space. In the low-energy stage, constraint-aware gradient descent is introduced, injecting the target improvement direction and boundary attraction into the position update simultaneously. A two-layer archive maintenance strategy is designed, using the boundary information potential energy index containing constraint improvement incentive factors within the task to guide the near-feasible solution toward the feasible region in a directional manner. At the cross-task layer, the optimized elite solution is encapsulated into a knowledge capsule. Through similarity filtering, the hot-start reuse of historical configuration experience is realized, which solves the problem that the existing optimization algorithm loses the balance between global exploration and local development when solving high-dimensional constrained water resource allocation, resulting in the discarding of infeasible solution information and the inability to reuse cross-task experience.

[0014] 3. Four scenarios—abundance, normal, low, and extreme drought—are constructed, and optimization algorithms are called to generate Pareto solution sets for each scenario. The entropy weight method is used to objectively assign weights to the solution sets based on their dispersion on each objective. TOPSIS is used to calculate the weighted Euclidean distance and relative proximity between each solution and the positive and negative ideal solutions. The top 3 preferred configuration solutions for each scenario are output in descending order of proximity. This solves the problem that existing technologies rely on human experience in solution selection, which makes it difficult to adapt to the dynamic changes of multiple scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating a water resource allocation analysis and decision-making method based on multi-objective collaborative optimization according to the present invention. Figure 2 This is a schematic diagram of the implementation process of S2 of the present invention; Figure 3 This is a schematic diagram of the implementation process of S3 of the present invention. Detailed Implementation

[0017] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0018] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0019] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0020] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0021] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0022] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0023] The provided examples are as follows: Please see the appendix Figure 1A water resource allocation analysis and decision-making method based on multi-objective collaborative optimization is proposed, specifically including: S1, collecting multi-source heterogeneous water resource data and preprocessing it to obtain a standardized multi-source heterogeneous data tensor; S2, based on a federated impulse learning framework, using a dual-loop sparse impulse attention and adaptive LIF spiking neural network model, performing impulse threshold encoding on the multi-source heterogeneous data tensor, as well as sparse excitation-lateral inhibition dual-loop feature extraction and impulse consistency weighted federated aggregation, outputting the daily water demand prediction value for each water-using unit; S3, using a hemispherical prey escape energy factor to drive... The dual-mode displacement search method, combined with the constraint-improved incentive factor CII and the knowledge transfer-enhanced multi-objective guilder optimization algorithm for cross-task knowledge capsule dual-layer archive maintenance, is used to optimize the daily water demand forecast through multi-source and multi-user collaborative configuration, resulting in a Pareto configuration scheme set. S4: The Pareto configuration scheme set for multiple scenarios is objectively weighted and ranked by proximity to obtain the top 3 configuration schemes in each scenario. S5: The top 3 configuration schemes in each scenario are decomposed and encoded at the instruction granularity and issued to the execution mechanism for water control.

[0024] The core technical points of this invention include: 1- Constructing a dual-loop sparse pulse attention architecture under a federated pulse learning framework, using pulse threshold encoding to drive feedforward excitation and lateral inhibition loops, replacing continuous dense attention; 2- Constructing a pulse matching aggregation strategy, using pulse firing pattern consistency scores to replace traditional federated averaging gradient weighting, overcoming the loss of discrete information in federated aggregation of spiking neural networks; 3- Constructing a dual-mode displacement search mechanism driven by a hemispherical prey escape energy factor, integrating logarithmic spiral exploration and constraint-aware gradient descent; 4- Constructing a two-layer knowledge transfer archive maintenance strategy that integrates constraint improvement incentive factors, using the directionality of boundary information potential energy indicators to guide the evolution of near-feasible solutions, and achieving structured reuse of historical configuration experience through cross-task knowledge capsule encapsulation and similarity filtering.

[0025] In specific implementation, S1 includes: S101. Collect and combine data from sensor equipment, hydrology, meteorology and water use statistics obtained from monitoring stations as raw heterogeneous data; S102. Perform missing value imputation, standardization, and spatiotemporal alignment on the original heterogeneous data to obtain a standardized multi-source heterogeneous data tensor.

[0026] Its missing value imputation, normalization processing, and spatiotemporal alignment include: Step 1) Use linear interpolation to fill in missing values. The formula is: In the formula x t t is the imputation estimate for the missing time t; t is the time index; t1 and t2 are the most recent valid observation times before and after the missing segment, respectively; and t1 and t2 are the actual observation values ​​at times t1 and t2, respectively; (t-t1) is the offset of the missing time from the starting point; The rate of change per unit time; Step 2) Perform Min-Max normalization on each feature sequence, mapping it to the [0,1] interval. The formula is: In the formula x 标准 The standardized value; x min x is the minimum value of the feature in the historical record; max The maximum value of the feature in the historical record; (xx) min (x) represents the offset of the current value relative to the minimum value; max -x min The numerical range of features; Step 3) Based on Steps 1) to 2), the standardized data are aligned according to the "day" time granularity and spatial partition (sub-basin / administrative partition) to form a synchronous time series with multi-source features. For scenarios involving cross-administrative regions (or sensitive water source data), the above preprocessing process of each participating region runs on the local edge server to form a regional local data tensor as a standardized multi-source heterogeneous data tensor.

[0027] For specific implementation details, please refer to the appendix. Figure 2 Its S2 includes: S201. Using an adaptive pulse threshold coding method based on local variance, the multi-source data tensor is dynamically binarized to generate a sparse binary pulse sequence. S202. Using the grouped projection and sparse masking attention method, only the attention weights of bidirectional pulse time step pairs are calculated for the sparse binary pulse sequence to obtain the attention features of the feedforward excitation loop. S203. Calculate the L2 norm of the attention feature of the feedforward excitation circuit as the activation intensity. The formula for calculating the activation intensity is as follows: In the formula, is Let L2 be the norm of the attention output matrix of the feedforward excitation loop; Let be the attention output value of the c-th channel at time step t; T is the time step number; t is the time index; the competition intensity score of each channel is obtained through softmax normalization, and the output feature of the lateral suppression loop is calculated by setting the suppression intensity parameter, the formula of which is: SAttn LI =SAttn FF ⊙(1-τ·C), where SAttn LI The lateral inhibition loop feature is represented by ⊙; element-wise multiplication is represented by ⊙; C is the competition intensity score vector; SAttn FFτ is the output feature matrix of the feedforward excitation loop; τ is the inhibition strength control parameter (typically 0.3); a fusion gating network is constructed, and the fusion weights are dynamically determined based on the features of the feedforward excitation loop and the lateral inhibition loop. After concatenating the feedforward and inhibition features, gating coefficients are generated through a linear layer and Sigmoid activation. The gating coefficients are used as weights to weight and fuse the features of the feedforward excitation loop and the lateral inhibition loop to obtain the dual-loop attention features; the dual-loop attention features are added to the original input residuals, and layer normalization is performed to stabilize the training to obtain the fused attention features.

[0028] The above-mentioned dynamic binarization of multi-source data tensors includes: Step 1) Calculate the local variance of the multi-source heterogeneous data tensor at each time step. The formula for calculating the local variance is: In the formula V 方差,t x represents the local variance at time step t (reflecting the overall fluctuation of multi-source water resource data at time t); t,i d represents the standardized value of the i-th feature at time step t; d represents the feature dimension of the standardized multi-source data tensor for the current region. For each feature, the squared deviation from its time-step mean is used (to measure the instantaneous fluctuation of a single feature); based on local variance, an adaptive threshold is generated through a learnable linear transformation and the Sigmoid function, with the following formula: In the formula θ is the adaptive threshold; θ0 is the base threshold (default is 0.3); σ is the Sigmoid function; W s and b s These are learnable parameters; Step 2) Based on Step 1), use step functions to generate pulses and construct a pulse sequence matrix S, whose set of step functions is: In the formula s 脉冲,t,i Let θ be a single element in the t-th row and i-th column of the pulse sequence matrix S (multiple pulses constitute the pulse sequence matrix S); Θ is the step function.

[0029] The above calculation of attention weights for bidirectional pulse time step pairs includes: Step 1) Divide the pulse sequence matrix S uniformly into G disjoint subspaces along the feature dimension, and perform linear projection of the query, key, and value respectively. The formula set is as follows: In the formula Q g K g and V g For querying the matrix, key matrix, and value matrix; S g This is the g-th pulse sequence; , and These are the projection weight matrices for the query, key, and value of the g-th group (which determine the information to focus on at the current time step, the retrieveable information carried at each time step, and the time-series dependency weights, respectively); g is the grouped index. Step 2) Based on Step 1), construct the sparse pulse mask matrix M by defining the sparse mask. g The sparse pulse mask is defined by the following formula: In the formula M g [i,j] is the sparse impulse mask matrix M g The element in the i-th row and j-th column (as a single element of the matrix); d G c represents the feature channel dimension for each group; c is the channel index. S is an existential quantifier. g [i,c] represents the pulse value of the c-th channel at time step i in the g-th group; st is the conditional identifier; [1,d] G [] represents the range of values ​​for the channel; Step 3) Based on Step 2), additive masking is performed using sparse impulse masks, assigning extremely negative values ​​to non-connected locations so that their weights tend to zero after softmax. The sparse attention calculation formula is as follows: In the formula S 注意力,g This is the final attention score matrix after applying a sparse mask; λ is the transpose of the key matrix; λ is the masking constant (as a penalty term, with a value of 10). -9 ); This is the scaled similarity score matrix; λ·(M g -1) is an additive penalty term; (M) g -1) is the mask transformation matrix; d k Hidden dimensions for each attention head; Q g Let S be the query matrix for the g-th group; 注意力,g The softmax formula is as follows: In the formula The attention output of the feedforward excitation loop is obtained by concatenating the attention outputs of all G groups along the feature dimension to restore the original dimension d, thus obtaining the attention feature SAttn of the feedforward excitation loop. FF .

[0030] In specific implementation, S2 also includes: S204. Using the Poisson pulse coding method, the fused attention features are transformed into pulse sequences and input into a four-layer adaptive time constant LIF spiking neural network for time series modeling. By replacing gradient learning and membrane potential linear decoding, the daily water demand prediction value of each water-using unit is output. At the same time, the parameters of the dual-loop sparse pulse attention and adaptive LIF spiking neural network model, as well as the pulse firing mode matrix of each layer, are recorded for federated aggregation. S205. Using a pulse matching aggregation method based on a federated pulse learning framework, pulse consistency scores are calculated for the pulse firing pattern matrices recorded in each layer. Based on the pulse consistency scores, the parameters of the dual-loop sparse pulse attention and adaptive LIF spiking neural network model are weighted and aggregated to obtain a global spiking neural network water demand prediction model, which is used to perform inference and update the daily water demand prediction values ​​of each water-using unit in real time.

[0031] The method for transforming fused attention features into pulse sequences is as follows: treat the continuous fused attention features as emissivity, use the Bernoulli function to sample each time step t and feature dimension i to obtain the actual pulse output, and construct a pulse sequence from all the pulse outputs.

[0032] Its four-layer adaptive time constant LIF spiking neural network, and alternative gradient learning and membrane potential linear decoding include: Step 1) Construct a four-layer LIF spiking neural network. The network parameters include: 1-Spiking hidden layer 1 (512 adaptive ATC-LIF neurons, fully connected input); 2-Spiking hidden layer 2 (256 ATC-LIF neurons, fully connected); 3-Spiking output layer (M water-unit neurons, accumulated membrane potential readout without spiking reset). A discrete-time update equation is used for each LIF neuron layer, with the following formula: In the formula Let l be the membrane potential of the j-th neuron in layer l at simulation step n (used to determine the firing pulse); l is the network layer index; n is the simulation time step. This represents the membrane potential of the previous simulation step; This is the sum of the synaptic currents input to the neuron at step n; This is the leakage term (used to simulate the passive decay of the membrane potential of biological neurons over time); the adaptive time constant is calculated using the following formula: In the formula τ is the membrane time constant of the j-th neuron in the l-th layer at the current membrane potential; 基础 The basic time constant (typically 2.0); τ 自适应 The adaptive amplitude is set to 5.0 (typical value 5.0); α is the temperature coefficient (set to 10.0); V is the current membrane potential (i.e., the membrane potential at the previous simulation step); V 阈值 The pulse firing threshold (typically 1.0); For the exponential term (used to convert the membrane potential shift into a smooth transition in the range [0,1]); when When the pulse is applied, the membrane potential is reset. Step 2) Based on Step 1), to address the non-differentiability issue of pulse delivery, the gradient is replaced by arctan during training. The replacement formula is: In the formula ω is the partial derivative of the pulse output o with respect to the membrane potential V (i.e., the gradient required for back propagation); ≈ is the approximate sign; β is the width parameter (typically 2.0). β is the scaling factor (used to adjust the maximum magnitude of the substitute gradient); β·(VV) 阈值 ) represents the scaled offset; V represents the current membrane potential (calculated using the discrete-time update equation described above). V is the derivative of the arctan function (giving the network the strongest learning signal near the critical state); 阈值 The threshold is set to an adaptive threshold (typically 1.0). For the output layer of the network, no pulse reset is generated; instead, the membrane potential over the entire simulation period is accumulated and passed through a learnable linear decoder to obtain the predicted water demand. The water demand calculation formula is as follows: In the formula Let N be the predicted water demand of the m-th water-using unit at time step t (i.e., the daily predicted water demand of each water-using unit); t is the time index; m is the water-using unit index; N s n is the total number of simulation time steps; n is the simulation time step index. The membrane potential of the m-th neuron in the output layer at the n-th simulation step; Let be the decoding weights of the m-th output neuron; This is the decoding bias for the m-th output neuron.

[0033] For specific implementation details, please refer to the appendix. Figure 3 Its S3 includes: S301. Obtain the daily water demand forecast value updated in real time for each water-using unit, construct a target-constrained optimization model with the objectives of water supply guarantee rate, water shortage, economic benefits and ecological satisfaction, and use the historical knowledge capsules activated by similarity filtering in the cross-task knowledge base to extract elite solutions as seeds and inject them into the initial population to obtain a hot-start initial population and empty elite files that integrate historical experience and random diversity. S302. Construct a hemispherical prey escape energy factor for the current iteration. The formula for calculating the energy factor is as follows: E in the formula 能量 (t) represents the escape energy factor value of the hemispherical prey at generation t; E0 is the initial escape energy; t is the current iteration generation; T is the total number of iterations; γ is the decay steepness control parameter; η is the iteration progress ratio; f is the oscillation amplitude control parameter; 震荡 This is the oscillation frequency parameter; For Sigmoid decay term; It is a sinusoidal oscillation term; S303. Based on the hemispherical prey escape energy factor, adaptively switch between energy-sensing spiral search or constraint-sensing gradient descent mode. Guided by the non-dominated solutions and population mean in the current elite files, perform dual-mode displacement update on the individuals in the population to generate offspring candidate populations.

[0034] The calculation steps for the above-mentioned targets and constraints regarding water supply guarantee rate, water shortage, economic benefits, and ecological satisfaction are as follows: Step 1) The formula for calculating the water supply guarantee rate is: In the formula, f1(x) is the objective function value of water supply guarantee rate; x is the decision variable vector; m is the water-using unit index; M is the total number of water-using units; n 水源 For water source indexing; For the nth 水源 The actual amount of water allocated from one water source to the m-th user; The predicted water demand of the m-th water-using unit at time step t (obtained from S2); The water supply satisfaction rate for user m; Step 2), the formula for calculating water shortage is: In the formula, f2(x) is the objective function value of the total water shortage; This represents the water shortage for user m; the meanings of the other symbols are the same as in step 1). Step 3) The formula for calculating economic benefits is: In the formula, f3(x) is the economic benefit objective function value (used to measure the total net economic benefit generated by the water resource allocation scheme); N is the total number of water sources; p m The water consumption benefit per unit for user m; For water source n 水源 The unit cost of supplying water to user m; Net unit benefit; the meanings of the remaining symbols are the same as in step 2); Step 4) The formula for calculating ecological satisfaction is: In the formula, f4(x) is the objective function value for ecological flow satisfaction; N 生态 A specific collection of water sources that can provide water for the ecosystem (such as groundwater that is not used for river ecosystems). For water source n 水源 The amount of water allocated to ecological water use units; Q 生态 The ecological water demand threshold (determined based on hydrological methods); To meet the ecological water supply requirements.

[0035] The above method utilizes historical knowledge capsules activated through similarity filtering in a cross-task knowledge base to extract elite solutions as seeds for the initial population. This involves constructing a feature vector for the current scenario, retrieving historical knowledge capsules from the cross-task knowledge base, and for each capsule, calculating the cosine similarity between its scenario features and the current features. The cosine similarity calculation formula is as follows: , where sim( , The cosine similarity between the current scenario and the historical scenario (used as a basis for determining whether a knowledge capsule can be activated). The feature vector of the current configuration scenario; These are the context feature vectors stored in the historical knowledge capsule; It is the product of the lengths of two vectors; The dot product of two feature vectors (serving as the original measure of similarity); only if sim( , When )≥δ (δ=0.7), the capsule is activated, and the elite solution set archived in the activated capsule is taken as the knowledge seed (the knowledge seed accounts for no more than 30% to maintain diversity), and together with the randomly generated solution, they form the initial population.

[0036] Using adaptive switching energy-aware spiral search as mode one and constraint-aware gradient descent as mode two, mode one mimics the circling and diving of a guilder bird in a high-energy state, using a logarithmic spiral to cover a wide space. Its candidate solution generation formula is as follows: In the formula The vector of new candidate solutions (representing a new water allocation scheme) generated by spiral search for individual i in pattern one; x i Let E be the current solution vector for individual i (i.e., the starting position of the search); 能量 (t) represents the energy factor value of the hemispherical prey escape at generation t (constructed via S302); x 非支配 This is a non-dominated solution randomly selected from the current elite archives; α is the exponential growth factor (used to simulate the dynamic change of the guilder's hovering radius); cos(2πψ) is the periodic oscillation factor (giving the search direction a periodic rotational characteristic); α·(x r1 -x r2 ) represents a diverse perturbation component (random diffusion superimposed along the main spiral direction, causing the search to be incompletely dominated by the guided solution); x r1 and x r2 These are the first and second individuals randomly selected from the current population (i.e., the solution vectors); (x r1 -x r2) represents the random difference perturbation term; b·ψ represents the spiral trajectory control parameter; for mode two, mimicking the precise capture of the guilder bird when energy is insufficient, the target improvement and constraint repair information are simultaneously injected into the updated candidate solution generation formula set as follows: In the formula The vector of new candidate solutions generated by constrained perceptual gradient descent for individual i in mode 2; △ 下降 The gradient descent displacement vector oriented towards the target; △ 引力 λ1 is the boundary gravity displacement vector; λ1 is the step size decay coefficient (set to 0.5 to control the overall magnitude of the gradient descent step size and prevent over-correction); f(x) i ) represents the four-dimensional objective function vector for individual i (consisting of the aforementioned objectives of water supply guarantee rate, water shortage, economic benefits, and ecological satisfaction); sign is the sign function; x 违反 The population with the lowest constraint violation and the distance x i The most recent individual; f 基准 is the arithmetic mean of the target vectors of all individuals in the current population (serving as a benchmark for the overall performance of the population); ⊙ is the element-wise multiplication operator; x 精英 -x i To guide the solution of the gap between individuals and elites across various decision-making dimensions; CVD(x i ) for individual x i The constraint violation degree (used to measure the overall degree of individual constraint violation, which is composed of the cumulative violation of four types of constraints: water supply capacity exceeding the limit, water conveyance capacity exceeding the limit, reservoir capacity exceeding the limit, and insufficient ecological base flow. Each violation is truncated as a non-negative difference between the actual value and the allowable upper limit). It is a minimal constant (to prevent division by zero); β 边界 The boundary driving coefficient is set to 0.3 to control the overall strength of the boundary gravity. For new candidate solution vectors generated after updating using Mode 1 or Mode 2, if any component exceeds the upper or lower bound of the variable (the lower bound is the known constraint constant matrix determined by the physical characteristics of the water conveyance project), the reflection boundary processing is used, and the generated new candidate solution vector is stored in the offspring candidate population.

[0037] In practical implementation, S3 also includes: S304. Using the Boundary Information Potential Energy (BIP) index and the Constraint Improvement Incentive Factor (CII), the parent population and offspring candidate population are merged with the current elite archive. The resulting set is then evaluated hierarchically and the archive is maintained to obtain a new generation of population and an updated elite archive. After the iteration ends, the Pareto configuration scheme set is constructed from the non-dominated feasible solutions in the final elite archive. S305. Using the crowding distance filtering and context feature encoding method, elite solutions are extracted from the Pareto configuration scheme set and encapsulated into knowledge capsules. Redundancy management is performed through cosine similarity filtering and replacement to obtain the updated cross-task knowledge base.

[0038] Its tiered evaluation and record maintenance include: Step 1) Based on the degree of constraint violation, divide the merged set of the parent population, the offspring candidate population and the current elite file into three layers: feasible domain layer, boundary information enhancement layer and severely infeasible layer. Step 2) Based on Step 1), perform fast non-dominated sorting on the feasible domain layer, and add files in order from low to high front layer level until the capacity is reached. If the capacity is exceeded when adding any front layer, calculate the congestion distance for that layer and discard it in descending order of congestion distance. Step 3) Based on Step 1), for the boundary information enhancement layer, since there are still empty slots in the file, the boundary information potential (BIP) of this layer is calculated using the following formula: In the formula, BIP(x) represents the final boundary information potential energy (used to comprehensively measure the information value of near-feasible solutions near the constraint boundary); x represents the candidate solution to be evaluated; CVD(x) represents the constraint violation degree of solution x; Dist 距离 (x) represents the crowding distance of the solution x in the target space; CVD(x) 父代 ) represents the degree of constraint violation of the parent individual that produced x; The constraint improvement rate index (measures the progress of the offspring in constraint repair relative to the parent, i.e., serves as the constraint improvement incentive factor CII). For a minimum value (taking the value 10) -8 To prevent division by zero); Step 4) Based on Step 1), select NP individuals (population size, i.e. the total number of individuals maintained in each generation of evolution) from the merged set using tournament selection (based on Pareto dominance and constraint violation) to form a new generation of population.

[0039] In specific implementation, S4 includes: S401. Obtain the daily water demand forecast for each water-using unit, construct four configuration scenarios (abundant, normal, low, and extreme drought) using the hydrological frequency classification method, and correct the constraint parameters. For each scenario, call the knowledge transfer-enhanced multi-objective guilder optimization algorithm to solve the problem and obtain the objective function matrix corresponding to the Pareto non-dominated configuration scheme set. S402. Use the entropy weight method to perform homogenization, normalization and information entropy calculation on the objective function matrix corresponding to the Pareto non-dominated configuration scheme set to obtain the objective weight vector of each objective in the comprehensive evaluation. S403. According to the objective weight vector, weight the normalized Pareto solution set, calculate the weighted Euclidean distances between each solution and the positive and negative ideal solutions, obtain the relative closeness degree and sort it in descending order, and finally obtain the top 3 configuration solutions with the best comprehensive performance in each scenario.

[0040] The construction of four configuration scenarios of abundant, normal, dry, and extremely dry conditions includes: 1. For the abundant water scenario with a water inflow frequency where the hydrological frequency P (i.e., annual runoff) ≤ 25%, the available water volume of each water source takes the relatively abundant value at the corresponding frequency; 2. For the normal water scenario with a water inflow frequency where 25% < P ≤ 75%, the available water volume takes the value of a normal water year; 3. For the dry water scenario with a water inflow frequency where 50% < P ≤ 75%, the available water volume takes the value of a dry water year; 4. For the extremely dry scenario with a water inflow frequency where P > 75%, the available water volume takes the value of an extremely dry year, and the ecological base flow constraint can be appropriately relaxed but default penalties are included.

[0041] It respectively calls the knowledge transfer enhanced multi-objective secretary bird optimization algorithm for solution, including: Step 1): For each scenario, use this scenario as the query key, calculate the cosine similarity with each historical knowledge capsule in the cross-task knowledge base, activate the capsules with a similarity ≥ 0.7, inject the elite solutions among them as knowledge seeds into the initial population to form a hot-start population that combines historical experience and random diversity, and simultaneously initialize an empty external elite archive; Step 2): After entering the iterative loop based on Step 1), execute the optimization process of S3 in sequence for each generation, update the elite archive and select a new generation of population; Step 3): Based on Step 2), when the iteration reaches the maximum number of generations T (set to 200), form the Pareto configuration solution set for this scenario with all non-dominated feasible solutions in the final elite archive, and each solution is attached with its four-dimensional objective function vector (the four configuration scenarios are executed independently). Finally, four groups of Pareto solution sets are obtained, including the water supply guarantee rate (the larger the better), the water shortage volume (the smaller the better), the economic benefit (the larger the better), and the ecological satisfaction degree (the larger the better).

[0042] The above-mentioned co-directionalization, normalization, and information entropy calculation processing include: Step 1): For the water shortage volume among the four objectives, which is a cost-type index (the smaller the better), and the rest are benefit-type indexes. Before calculating the entropy weight, the reciprocal transformation method needs to be used to convert the cost-type index into a benefit-type index. The reciprocal transformation formula is: , where is the positive value of the j-th solution on objective 2 (i.e., the total water shortage volume); f j2 is the original value of the j-th solution on objective 2 (i.e., the total water shortage volume); is a minimum constant (to prevent division by zero); This is a reciprocal transformation operation (used to reverse the direction; the smaller the water shortage, the larger the reciprocal); other objectives remain unchanged, and at this time all objectives satisfy "the larger the better"; Step 2) Based on Step 1), the linear scaling transformation method (i.e., Sum normalization) is used to eliminate the influence of dimensions and transform each target value into a probability distribution form. For the m-th target, its information entropy value is used to reflect the degree of dispersion of the target's distribution in the Pareto solution set. The formula for calculating its information entropy is: In the formula e m Let J be the information entropy of the m-th objective (a measure of the information content of the objective in the Pareto solution set); J is the total number of solutions in the Pareto solution set (used to determine the maximum possible value of the entropy). The original entropy after taking the negative (i.e., converted into a positive information metric); p is the natural logarithm of the ratio (i.e., the core term of self-information in information theory); jm Let be the normalized scaling value of the j-th scheme on the m-th objective; Step 3) Calculate the objective weights using the following formula: In the formula w m d represents the objective weight of the m-th objective. m e represents the information redundancy (i.e., the difference coefficient) of the m-th target; m The information entropy of the m-th target (used as input for calculating redundancy); This is the sum of the redundancy of the four target information.

[0043] The set of weighted Euclidean distance formulas for calculating the distance between each scheme and the positive and negative ideal solutions are as follows: In the formula and v represents the weighted Euclidean distance from the j-th scheme to the positive and negative ideal solutions, respectively (measuring the difference between scheme j and the optimal and worst solutions); jm This is the weighted standardized value of the j-th scheme on the m-th objective; and These are the positive and negative ideal solution values ​​for the m-th objective, respectively. and Let be the squares of the deviations of scheme j from the positive and negative ideal solutions on target m, respectively; the formula for calculating its relative closeness is: In the formula, C j Let be the overall relative similarity of the j-th scheme.

[0044] In specific implementation, its S5 includes: S501. For each scenario, calculate the top 3 configuration schemes based on overall performance to obtain the engineering optimization control instruction sequence corresponding to each scheme. S502. Distribute the engineering optimization control command sequence corresponding to each scheme to the lower execution mechanism for execution and monitor and feedback status data in real time.

[0045] The above solution calculation for the top 3 configuration schemes in each scenario includes: Step 1) For each water source n in the configuration scheme, summarize its total allocation to all M water-using units, and perform parameter calculations according to the water source project type, including dam-type water sources, pump station-type water sources, and water diversion channel / pipeline valve-type water sources; Step 2) Based on Step 1) for the known current water level of the reservoir for dam-type water sources, the current water storage is obtained by interpolation using a pre-set water level-storage capacity relationship table. The target water storage after configuration is calculated using the following formula: V 目标 =V 实际 -Q 计划 +Q 预报 In the formula, V 目标 V represents the target water storage capacity of reservoir n at the end of the configuration period (i.e., the water storage state the reservoir should reach after configuration); 实际 Q represents the actual water storage capacity of reservoir n at the beginning of the current configuration period; 计划 Q represents the total planned discharge from reservoir n during the specified time period; 预报 To configure the forecast of inflow water volume within the specified time period; Step 3) Based on Step 1), for pump station-type water sources, the formula for calculating the total water supply is as follows: In the formula The total water supply flow required for pump station n (i.e., the amount of water that the pump station needs to supply to all users, extracted from the configuration plan); K n R represents the number of pumps that need to be turned on at pump station n; k is the pump number index; n The uniform operating speed of each pump in pumping station n (based on the characteristic curve provided by the pump manufacturer); q k (R n Let R be the speed of the k-th water pump. n Flow characteristic function; Step 4) Based on Step 1), for valves in water diversion channels / pipelines, the valve opening is obtained by reverse lookup using the pre-calibrated opening-flow curve (the factory characteristic curve provided by the valve manufacturer). Step 5) Based on steps 1) to 4), organize the control parameters of all engineering nodes into a standard control instruction structure. Each instruction includes: 1-unique equipment identifier; 2-parameter type; 3-calculated specific value; 4-instruction execution period.

[0046] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

Claims

1. A water resource allocation analysis and decision-making method based on multi-objective collaborative optimization, characterized in that, Specifically, it includes: S1. Collect multi-source heterogeneous water resource data and preprocess it to obtain a standardized multi-source heterogeneous data tensor. S2. Based on the federated impulse learning framework, a dual-loop sparse impulse attention and adaptive LIF spiking neural network model are used to perform impulse threshold encoding on multi-source heterogeneous data tensors, as well as sparse excitation-lateral inhibition dual-loop feature extraction and impulse consistency weighted federated aggregation, to output the daily water demand prediction value of each water-using unit. S3. Using a dual-mode displacement search method driven by a hemispherical prey escape energy factor and incorporating the constraint-improved incentive factor CII, as well as a knowledge transfer-enhanced multi-objective guilder optimization algorithm with cross-task knowledge capsule dual-layer archive maintenance, the daily water demand forecast is optimized by multi-source and multi-user collaborative configuration to obtain the Pareto configuration scheme set. S4. Objectively weight and rank the Pareto configuration schemes for multiple scenarios based on their relevance to obtain the top 3 configuration schemes in terms of overall performance for each scenario. S5. For the top 3 configuration schemes in each scenario, decompose and encode them at the instruction granularity and send them to the execution mechanism for water volume control. S3 includes: S301. Obtain the daily water demand forecast value updated in real time for each water-using unit, construct a target-constrained optimization model with the objectives of water supply guarantee rate, water shortage, economic benefits and ecological satisfaction, and use the historical knowledge capsules activated by similarity filtering in the cross-task knowledge base to extract elite solutions as seeds and inject them into the initial population to obtain a hot-start initial population and empty elite files that integrate historical experience and random diversity. S302. Construct a hemispherical prey escape energy factor for the current iteration. The formula for calculating the energy factor is as follows: E in the formula 能量 (t) represents the escape energy factor value of the hemispherical prey at generation t; E0 is the initial escape energy; t is the current iteration generation; T represents the total number of iterations; γ is the decay steepness control parameter. η is the iteration progress ratio; f is the oscillation amplitude control parameter; 震荡 This is the oscillation frequency parameter; For Sigmoid decay term; It is a sinusoidal oscillation term; S303. Based on the hemispherical prey escape energy factor, adaptively switch between energy-sensing spiral search or constraint-sensing gradient descent mode. Guided by the non-dominated solutions and population mean in the current elite files, perform dual-mode displacement update on the individuals in the population to generate offspring candidate populations.

2. The water resource allocation analysis and decision-making method based on multi-objective collaborative optimization according to claim 1, characterized in that, S1 includes: S101. Collect and combine data from sensor equipment, hydrology, meteorology and water use statistics obtained from monitoring stations as raw heterogeneous data; S102. Perform missing value imputation, standardization, and spatiotemporal alignment on the original heterogeneous data to obtain a standardized multi-source heterogeneous data tensor.

3. The water resource allocation analysis and decision-making method based on multi-objective collaborative optimization according to claim 2, characterized in that, S2 includes: S201. Using an adaptive pulse threshold coding method based on local variance, the multi-source data tensor is dynamically binarized to generate a sparse binary pulse sequence. S202. Using the grouped projection and sparse masking attention method, only the attention weights of bidirectional pulse time step pairs are calculated for the sparse binary pulse sequence to obtain the attention features of the feedforward excitation loop. S203. Calculate the L2 norm of the attentional characteristics of the feedforward excitation circuit as the activation intensity. The formula for calculating the activation intensity is as follows: In the formula Let L2 be the norm of the attention output matrix of the feedforward excitation loop; For the c-th channel at time t 索引 Attention output value at each time step; T 时间 t represents the number of time steps. 索引 For time indexing; through softmax normalization, the competition intensity score of each channel is obtained, and the output characteristics of the lateral suppression loop are calculated by setting the suppression intensity parameter, the formula of which is: SAttn LI =SAttn FF ⊙(1-τ·C), where SAttn LI The lateral inhibition loop feature is represented by ⊙; element-wise multiplication is represented by ⊙; C is the competition intensity score vector; SAttn FF τ represents the output feature matrix of the feedforward excitation loop; τ is the inhibition strength control parameter. A fusion gating network is constructed, and the fusion weights are dynamically determined based on the features of the feedforward excitation loop and the lateral inhibition loop. After concatenating the feedforward and inhibition features, gating coefficients are generated through a linear layer and a Sigmoid activation. The gating coefficients are used as weights to weight and fuse the features of the feedforward excitation loop and the lateral inhibition loop to obtain the dual-loop attention features. The dual-loop attention features are added to the original input residuals, and layer normalization is performed to stabilize the training, resulting in the fused attention features.

4. The water resource allocation analysis and decision-making method based on multi-objective collaborative optimization according to claim 3, characterized in that, S2 further includes: S204. Using the Poisson pulse coding method, the fused attention features are transformed into pulse sequences and input into a four-layer adaptive time constant LIF spiking neural network for time series modeling. By replacing gradient learning and membrane potential linear decoding, the daily water demand prediction value of each water-using unit is output. At the same time, the parameters of the dual-loop sparse pulse attention and adaptive LIF spiking neural network model, as well as the pulse firing mode matrix of each layer, are recorded for federated aggregation. S205. Using a pulse matching aggregation method based on a federated pulse learning framework, pulse consistency scores are calculated for the pulse firing pattern matrices recorded in each layer. Based on the pulse consistency scores, the parameters of the dual-loop sparse pulse attention and adaptive LIF spiking neural network model are weighted and aggregated to obtain a global spiking neural network water demand prediction model, which is used to perform inference and update the daily water demand prediction values ​​of each water-using unit in real time.

5. The water resource allocation analysis and decision-making method based on multi-objective collaborative optimization according to claim 1, characterized in that, S3 further includes: S304. Using the Boundary Information Potential Energy (BIP) index and the Constraint Improvement Incentive Factor (CII), the parent population and offspring candidate population are merged with the current elite archive. The resulting set is then evaluated hierarchically and the archive is maintained to obtain a new generation of population and an updated elite archive. After the iteration ends, the Pareto configuration scheme set is constructed from the non-dominated feasible solutions in the final elite archive. S305. Using the crowding distance filtering and context feature encoding method, elite solutions are extracted from the Pareto configuration scheme set and encapsulated into knowledge capsules. Redundancy management is performed through cosine similarity filtering and replacement to obtain the updated cross-task knowledge base.

6. The water resource allocation analysis and decision-making method based on multi-objective collaborative optimization according to claim 5, characterized in that, S4 includes: S401. Obtain the daily water demand forecast for each water-using unit, construct four configuration scenarios (abundant, normal, low, and extreme drought) using the hydrological frequency classification method, and correct the constraint parameters. For each scenario, call the knowledge transfer-enhanced multi-objective guilder optimization algorithm to solve the problem and obtain the objective function matrix corresponding to the Pareto non-dominated configuration scheme set. S402. Use the entropy weight method to perform homogenization, normalization and information entropy calculation on the objective function matrix corresponding to the Pareto non-dominated configuration scheme set to obtain the objective weight vector of each objective in the comprehensive evaluation. S403. Based on the objective weight vector, the normalized Pareto solution set is weighted. By calculating the weighted Euclidean distance between each solution and the positive and negative ideal solutions, the relative proximity is obtained and sorted in descending order. Finally, the top 3 configuration solutions with the best overall performance in each scenario are obtained.

7. A water resource allocation analysis and decision-making method based on multi-objective collaborative optimization according to claim 6, characterized in that, S5 includes: S501. Perform scheme calculations on the top 3 configuration schemes with the best overall performance in each scenario to obtain the engineering optimization control instruction sequence corresponding to each scheme. S502. Distribute the engineering optimization control command sequence corresponding to each scheme to the execution mechanism for execution and monitor and feedback status data in real time.

Images