CN-121745638-B - Lake water quality dynamic optimization regulation and control method

CN121745638BCN 121745638 BCN121745638 BCN 121745638BCN-121745638-B

Abstract

The invention discloses a lake water quality dynamic optimization regulation method, which comprises the steps of obtaining a pre-trained water quality prediction proxy model, constructing a state input sequence containing a historical regulation decision and a current candidate solution, calculating a dynamic factor reflecting regulation loose degree by utilizing a decomposition coordination mechanism based on global total constraint, accordingly determining the current target water flux, constructing a multi-target optimization model, solving by utilizing a multi-target optimization algorithm based on population evolution to obtain the current optimal regulation scheme, executing the scheme, updating the system accumulated flux and the state input sequence, and then rolling to enter the next moment. The invention solves the technical bottlenecks of scarce data, low calculation efficiency and poor dynamic adaptability of the traditional method, and realizes effective dynamic optimization regulation and control of lake water quality.

Inventors

LI XINYU
LIU YANLI
GUAN TIESHENG
LI QIONGFANG
JIN JUNLIANG
DIAO YANFANG
HU CAIHONG
Tang Xiongpeng
WEI JIE

Assignees

水利部交通运输部国家能源局南京水利科学研究院

Dates

Publication Date: 20260508
Application Date: 20260226

Claims (8)

1. A lake water quality dynamic optimization regulation method is characterized by comprising the following steps: acquiring a pre-trained water quality prediction agent model; constructing a state input sequence at the current moment, wherein the state input sequence comprises executed regulation decision information at the historical moment and candidate solution information to be evaluated at the current moment; Calculating a dynamic factor reflecting the current time regulation and control loose degree by utilizing a global constraint decomposition mechanism based on global total constraint in the residual optimization period, and determining the current day target water flux according to the dynamic factor; constructing a multi-objective optimization model on the same day, wherein an objective function of the multi-objective optimization model comprises a water quality exceeding punishment item calculated based on a prediction result output by a water quality prediction agent model aiming at a state input sequence, and a water quantity deviation punishment item calculated based on a target water flux on the same day; Solving a multi-objective optimization model on the same day by using a multi-objective optimization algorithm based on population evolution to obtain an optimal regulation scheme on the same day; outputting the optimal regulation scheme on the same day, updating the accumulated flux and the state input sequence of the system according to the optimal regulation scheme, and rolling to enter the next moment until the regulation of the whole optimization period is completed; calculating a dynamic factor reflecting the regulation and control loose degree at the current moment by using a global constraint decomposition mechanism, wherein the dynamic factor comprises the following components: Acquiring a preset global water quantity control target of a regulating unit, and acquiring accumulated executed water flux at the moment before the current moment; Calculating to obtain residual water budget in a residual optimization period based on the global water control target and accumulated executed water flux; Acquiring an original reference residual water quantity required by an original plan in a pre-stored residual optimization period under a non-optimization scene; calculating the ratio of the residual water budget to the original reference residual water, and determining the ratio as a dynamic factor; calculating the ratio of the remaining water budget to the original reference remaining water is based on the following formula: α i (t) =(T i -C i (t-1) )/S i (t) ; Wherein alpha i (t) is the dynamic factor of the ith regulating and controlling unit at the T moment, T i is the global water quantity control target of the regulating and controlling unit, C i (t-1) is the accumulated executed water flux of the regulating and controlling unit at the T-1 moment, and S i (t) is the original reference residual water quantity required by the regulating and controlling unit from the T moment to the end of the optimization period under the non-optimization scene.
2. The method of claim 1, wherein constructing the state input sequence for the current time comprises: Setting a time review window with a length of L, and constructing a time sequence containing L historical moment state vectors; For the control variable part in the time sequence, the following rules are assigned: For the historical moment before the optimization starting moment, adopting a pre-acquired original observation value; for the historical time which is positioned after the optimization starting time and before the current time, adopting an optimized value in the current day optimal regulation scheme which is actually executed at the corresponding time; and for the current moment, adopting a value corresponding to the candidate solution to be evaluated, which is generated by a multi-objective optimization algorithm.
3. The method of claim 1, wherein calculating a dynamic factor reflecting a current time regulation and control loose degree using a global constraint decomposition mechanism, further comprises numerical correction of the dynamic factor: Judging whether the current moment is the end moment of the optimization period, if so, calculating a dynamic factor based on the original reference water quantity at the current moment only without involving the subsequent period, and if not, judging whether the original reference residual water quantity is smaller than a preset minimum threshold value: If the dynamic factor is smaller than the minimum threshold value, the dynamic factor is directly set to a preset default value, and zero removal errors are avoided; if the dynamic factor is greater than or equal to the minimum threshold value, calculating according to the formula of claim 1 to obtain a dynamic factor calculated value, and judging whether the dynamic factor calculated value exceeds a preset reasonable interval, if so, cutting off the dynamic factor to the boundary value of the reasonable interval, otherwise, directly adopting the dynamic factor calculated value as a final dynamic factor.
4. The method of claim 1, wherein determining the current day target water flux using a dynamic factor comprises: acquiring the current day original water flux of a pre-stored non-optimal scene; and constructing a smooth mapping function, substituting the dynamic factor into the smooth mapping function to obtain a flow regulation coefficient, and determining the product of the current day original water flux and the flow regulation coefficient as the current day target water flux.
5. The method of claim 1, wherein the objective function of the multi-objective optimization model of the day comprises: The water quality exceeding punishment item is obtained by calculating the ratio of the difference value of the predicted water quality index output by the water quality prediction agent model and the preset water quality target value to the preset water quality target value, and the punishment value is only generated when the predicted water quality index is inferior to the preset water quality target value; A water quantity deviation punishment term is calculated based on the relative deviation between the optimized water flux corresponding to the candidate solution to be evaluated and the current day target water flux generated by the multi-target optimization algorithm; wherein, the water quality exceeding punishment item and the water quantity deviation punishment item are calculated in a relative error mode so as to eliminate dimension and order-of-magnitude differences among different physical quantities.
6. The method of claim 1, wherein solving the current day multi-objective optimization model using a multi-objective optimization algorithm based on population evolution, specifically comprising iterative optimization using elite selection strategies that balance water quality guidance with population diversity, the elite selection strategies comprising: non-dominant sorting is carried out on the combined population formed by the parent population and the offspring population, so as to obtain a plurality of pareto front layering; Calculating the water quality standard reaching evaluation score and the crowding degree distance of each individual in the combined population; acquiring a current optimization progress, and dynamically adjusting a water quality guiding weight and a diversity maintaining weight according to the current optimization progress; calculating a comprehensive score of each individual based on the adjusted water quality guiding weight, the diversity maintaining weight, the water quality standard reaching evaluation score and the crowding degree distance; And sequencing individuals according to the comprehensive scores in the same pareto front layering layer, and selecting elite individuals entering the next generation population based on the sequencing result.
7. The method of claim 6, wherein calculating a composite score for each individual comprises: For each individual, calculating a water quality standard reaching assessment score based on the weighted deviation of the predicted value and the target value of each corresponding water quality index; respectively carrying out normalization processing on the water quality standard reaching evaluation score and the crowding degree distance to obtain a normalized water quality score and a normalized crowding degree distance; Weighting the normalized water quality score by using a water quality guiding weight, and weighting the normalized crowding degree distance by using a diversity maintaining weight; the difference between the weighted normalized water quality score and the weighted normalized crowdedness distance is calculated and determined as a composite score, wherein a lower composite score indicates a higher individual preference level.
8. The method of claim 6, wherein dynamically adjusting the water quality guidance weight and the diversity maintaining weight based on the current optimization schedule comprises: determining the ratio of the current iteration algebra to the preset maximum iteration algebra as the current optimization progress; presetting a first progress threshold and a second progress threshold, wherein the second progress threshold is larger than the first progress threshold; When the current optimization progress is smaller than a first progress threshold, setting the diversity maintaining weight to be larger than the water quality guiding weight so as to focus on the diversity exploration of the population; when the current optimization progress is greater than or equal to a second progress threshold, setting the water quality guiding weight to be greater than the diversity maintaining weight so as to focus on the water quality to reach the standard and converge; when the current optimization progress is between the first progress threshold and the second progress threshold, the water quality guiding weight and the diversity maintaining weight are set to be balance values.

Description

Lake water quality dynamic optimization regulation and control method Technical Field The invention relates to the technical field of water environment pollution control and water resource management, in particular to a lake water quality dynamic optimization regulation method. Background The lake water environment management is a complex system engineering which relates to the coupling of multiple processes such as hydrodynamic force, water quality and engineering regulation, and the optimal regulation scheme of the water quality is scientifically determined, so that the lake water environment management is a key for realizing the standard treatment of lakes and the effective utilization of water resources. Particularly in the drainage and drainage lakes in agricultural irrigation areas, the lake entering pollution load and the water resource allocation show obvious nonlinear response characteristics, and the accurate analysis of the water environment capacity under the complex working conditions is realized by establishing a high-fidelity and low-time-consuming response regulation system, so that the method has important technical value for improving the stability of the lake ecological system and the water quality standard reaching precision. Currently, in the prior art, a mechanism model is mostly adopted to perform manual scene simulation, or historical monitoring data is utilized to train a proxy model to perform rolling optimization. The scene simulation method relies on expert experience to set a limited scheme, so that it is difficult to ensure that a globally optimal solution is found, and it is difficult to reveal an inherent trade-off relationship among multiple targets. When the mechanism model is directly coupled with the optimization algorithm, the method is limited by the problem that the single operation of the model consumes longer time, and the timeliness requirement of real-time analysis is difficult to meet. The agent model trained based on pure historical data has serious defects in prediction reliability when facing the historical non-existing situations of low pollution load, high water supplementing flow and the like, and the existing regulation and control framework mostly adopts static decomposition rules when processing global total constraint, so that the real-time dynamic change of the system is difficult to deal with. The prior art is difficult to balance between computational timeliness, predictive generalization capability and dynamic constraint accuracy. The method has the advantages that the agent model has weak out-of-distribution prediction capability on out-of-distribution scenes, the rolling optimization process lacks explicit description on the cumulative effect of decisions, and the global total constraint is difficult to consider global convergence and local flexibility in real-time scheduling. Because the agent model is difficult to effectively reproduce the physical characteristics of the mechanism model and the constraint decomposition in the rolling time domain lacks feedback correction, the regulation and control scheme generated by the optimization algorithm is often invalid in a complex dynamic environment. How to construct a regulating and controlling method which has high generalization capability, can capture decision feedback and can adaptively coordinate global constraint is a technical problem to be solved in the current lake treatment field. Disclosure of Invention The invention aims to provide a lake water quality dynamic optimization regulation method, which aims to solve at least one of the problems in the prior art. The technical scheme is that the lake water quality dynamic optimization regulation method comprises the following steps: acquiring a pre-trained water quality prediction agent model; constructing a state input sequence at the current moment, wherein the state input sequence comprises executed regulation decision information at the historical moment and candidate solution information to be evaluated at the current moment; Calculating a dynamic factor reflecting the current time regulation and control loose degree by utilizing a global constraint decomposition mechanism based on global total constraint in the residual optimization period, and determining the current day target water flux according to the dynamic factor; constructing a multi-objective optimization model on the same day, wherein an objective function of the multi-objective optimization model comprises a water quality exceeding punishment item calculated based on a prediction result output by a water quality prediction agent model aiming at a state input sequence, and a water quantity deviation punishment item calculated based on a target water flux on the same day; Solving a multi-objective optimization model on the same day by using a multi-objective optimization algorithm based on population evolution to obtain an optimal regulation scheme on the same day; outputting the