A sound source array arrangement optimization method based on a particle swarm algorithm

Abstract

The present application belongs to the technical field of acoustics optimization, and particularly relates to a sound source array arrangement optimization method based on a particle swarm algorithm, comprising: obtaining cooling tower parameters for joint simulation, extracting droplet concentration distribution data and normalizing the droplet concentration distribution data into mist droplet weights, and constructing a sound field prediction model and an optimization space; decoding the particles to be evaluated into the sound field prediction model to obtain a sound pressure level distribution, combining the mist droplet weights and an energy consumption penalty term to calculate fitness; dynamically adjusting an inertia weight based on a discrete degree ratio and a stagnation degree ratio, and performing directional reseeding to a high-concentration low-sound-pressure blind area when premature convergence is encountered; and decoding the optimal particle after convergence for review calculation and converting the optimal particle into a parameter file. The present application eliminates the sound pressure blind area in the high-concentration area, and realizes the unification of algorithm convergence quality and engineering water collection targets in the sound source array arrangement optimization.

Description

Technical Field

[0001] This invention belongs to the field of acoustic optimization technology, specifically relating to a method for optimizing the arrangement of sound source arrays based on particle swarm optimization algorithm. Background Technology

[0002] In water-intensive industries such as thermal power generation and steel metallurgy, natural draft cooling towers generate a large number of fine droplets with a diameter of 5 to 50 micrometers during operation. These droplets escape from the tower with the rising airflow, causing considerable water loss. Acoustic agglomeration water collection technology utilizes a high-intensity sound field to induce differentiated vibration, collision, aggregation, and gravity settling of these fine droplets. This is currently the mainstream approach to overcome the limitations of traditional mechanical water separators. Since the effective radius of a single column-type air-acoustic transducer is limited, engineering practice necessitates the deployment of an array of dozens to hundreds of sound sources within the cooling tower. The spatial location, orientation, and number of these sound sources directly determine the coverage quality of the sound field within the tower, and consequently, the overall water collection efficiency.

[0003] For the design of sound source array layouts, existing technologies have developed a design approach combining acoustic simulation and intelligent optimization: First, acoustic software such as COMSOL is used to perform batch sound field simulations on different array schemes to build training samples; then, a backpropagation neural network is used to fit the mapping relationship between the layout parameters and the sound field distribution to form a rapid prediction model; finally, swarm intelligence algorithms such as particle swarm optimization are used to perform global optimization on the prediction model, outputting the array scheme that maximizes the sound pressure level inside the tower. Compared with purely manual experience-based layout, this approach shows significant improvements in both design efficiency and coverage.

[0004] However, the internal flow field of a cooling tower exhibits a significant non-uniformity, characterized by high speed at the center and low speed around the periphery. Similarly, the droplet concentration is spatially non-uniform, with high-concentration droplet areas often concentrated in specific rings and recirculation zones above the packing layer. Against this backdrop, the standard particle swarm optimization algorithm reveals deep-seated flaws when performing array optimization: the decision variables for array arrangement simultaneously include discrete quantities like the number of sound sources and continuous quantities like position coordinates and installation angles, creating a high-dimensional hybrid optimization space. The standard particle swarm algorithm relies on the unidirectional attraction of globally optimal particles, making it highly susceptible to rapid absorption by a locally dominant scheme in the early iteration stages, leading to a sharp loss of population diversity and premature convergence. More critically, the locally optimal array obtained through premature convergence suffers from a mismatch between the sound field energy distribution and the spatial distribution of droplet concentration—sound pressure energy is excessively concentrated in the high-speed zone at the center of the airflow, while the edge high-concentration rings, where droplet concentration is highest and the sound field is most needed, form a low-sound-pressure blind zone with insufficient effective sound pressure. This blind zone firmly locks in water collection efficiency, and further increasing the number of sound sources or sound power only increases energy consumption without eliminating the blind zone. How to prevent premature convergence in the particle swarm optimization process and match the convergence result with the spatial distribution of droplet concentration has become a core problem that urgently needs to be solved in the optimization of sound source array layout. Summary of the Invention

[0005] This invention provides a sound source array arrangement optimization method based on particle swarm optimization algorithm to solve the technical problem that premature convergence of standard particle swarm optimization algorithm in high-dimensional hybrid optimization space causes mismatch between the sound field energy distribution and droplet concentration spatial distribution of the convergent array.

[0006] This invention provides a sound source array layout optimization method based on particle swarm optimization (PSO), comprising: S1, obtaining the structural geometric parameters and operating boundary conditions of the cooling tower, performing joint simulation to extract droplet concentration distribution data, normalizing the droplet concentration as the droplet weight, establishing a sound field prediction model with sound source parameters as input and sound pressure level distribution as output, and constructing an optimization space for PSO based on the sound field prediction model; S2, decoding the particles to be evaluated in the optimization space into array schemes and inputting them into the sound field prediction model to obtain the sound pressure level distribution, determining the effective coverage grid based on the sound pressure level distribution and aggregation threshold, and based on the accumulated droplet weight value corresponding to the effective coverage grid, the accumulated droplet weight value of the entire tower, and the objective value based on the air supply boundary. The energy consumption penalty term is calculated in reverse to determine the fitness of the particles to be evaluated; S3, during the particle swarm iteration process, the inertia weight is dynamically adjusted based on the dispersion ratio and stagnation ratio of the population fitness to update the particle velocity. When the population is determined to be in a premature convergence state, the sound field evaluation result corresponding to the current global optimal array is extracted to mark the high-concentration low sound pressure blind zone, and the sound source coordinates of some particles in the population are forcibly migrated to the neighborhood of the high-concentration low sound pressure blind zone to perform directional reseeding; S4, when the population meets the convergence criterion, the current global optimal particle is extracted and decoded into the final sound source array scheme. The final sound source array scheme is back-substituted into the full-size acoustic finite element model for verification calculation. After the verification is passed, it is converted into the parameter file required for engineering construction.

[0007] Its effect is as follows: by constructing droplet weights based on the spatial distribution of droplet concentration obtained from the flow field simulation in the cooling tower, the evaluation benchmark of the optimization space directly corresponds to the actual distribution of droplets; by constructing a fitness function based on the energy consumption penalty term objectively derived from the droplet weight coverage ratio and the air supply boundary, the optimization objective changes from maximizing the sound pressure level to matching the sound field energy distribution with the droplet distribution, thus mathematically blocking the ineffective optimization path of relying on piling up sound sources to obtain indicators.

[0008] Furthermore, the structural geometric parameters and operating boundary conditions of the cooling tower are obtained through joint simulation to extract droplet concentration distribution data. The droplet concentration is normalized and used as the droplet weight, including: A three-dimensional computational fluid dynamics model is constructed by obtaining the internal outline of the cooling tower, the elevation of the packing layer, the location of the water collector, the layout of the flue gas duct, and the operating parameters. The velocity field, temperature field, and humidity field of the rising airflow inside the tower were jointly simulated, and the discrete phase model was used to track the trajectory of the droplets and extract the spatial distribution data of droplet concentration and particle size in the three-dimensional space inside the tower. The space within the cooling tower where sound sources can be placed is discretized into a three-dimensional grid at a predetermined resolution, and the droplet mass concentration of all three-dimensional grids is normalized and used as the droplet weight of each three-dimensional grid.

[0009] Furthermore, a sound field prediction model is established with sound source parameters as input and sound pressure level distribution as output. Based on this model, an optimization space for the particle swarm optimization algorithm is constructed, including: Using the number of sound sources, the three-dimensional spatial coordinates of each sound source, and the installation angle as input variables, and the sound pressure level distribution of all grid units in the tower as output variables, the input and output sample pairs are obtained by batch sound field simulation of array schemes with the range of sound source number and location parameter values ​​in the coverage optimization space. A neural network is used to fit the input-output mapping relationship to complete the construction of the sound field prediction model; Encode a single array scheme as a particle, and limit the range of sound source coordinates to the grid set corresponding to the installable structural surface to construct the optimization space of the particle swarm algorithm.

[0010] Its effects are as follows: it uses batch sound field simulation to generate samples, establishes a mapping from sound source arrangement parameters to sound pressure level distribution through neural networks to obtain a sound field prediction model, encodes the array scheme as particles and delineates the optimization space in combination with actual installation constraints, replaces time-consuming finite element simulation with neural networks, greatly reduces the computational overhead of particle swarm iterative evaluation, and embeds engineering installation constraints into the optimization space to ensure that the optimized scheme has physical feasibility.

[0011] Furthermore, based on the cumulative droplet weights corresponding to the effective coverage grid, the cumulative droplet weights of the entire tower, and the energy consumption penalty term objectively deduced from the air supply boundary, the fitness of the particle to be evaluated is calculated, including: using the ratio of the cumulative droplet weights corresponding to the effective coverage grid to the cumulative droplet weights of the entire tower as the droplet weight coverage ratio; objectively deducing the energy consumption penalty coefficient based on the number of sound sources operating at full power that the air compressor system can support under rated operating conditions with its maximum stable air supply capacity, and the maximum gain that the droplet weight coverage ratio can obtain when the number of sound sources increases to the number of sound sources operating at full power; calculating the energy consumption penalty term based on the actual number of sound sources used by the particle to be evaluated, the upper limit of the number of sound sources constrained by installation space and air supply capacity, and the energy consumption penalty coefficient; and using the difference between the droplet weight coverage ratio and the energy consumption penalty term as the fitness of the particle to be evaluated.

[0012] Its effect is that by constructing a fitness function that simultaneously incorporates the matching degree of sound field energy with the spatial distribution of droplets and the constraints of the air supply boundary, the particle swarm always focuses on eliminating the insufficient sound pressure in the high-concentration area as the core of optimization during the iteration process, rather than simply pursuing the numerical improvement of the average sound pressure level across the entire field. This fundamentally corrects the deviation between the optimization direction and the water collection target.

[0013] Furthermore, the inertia weights are dynamically adjusted based on the population fitness dispersion ratio and stagnation ratio to update particle velocities, including: The current dispersion of the population is characterized by the standard deviation of the fitness values ​​of all particles, and the dispersion ratio is calculated by combining the standard deviation of the fitness values ​​of the initial population. Record the number of stagnant generations in which the global optimal fitness has not been improved, and calculate the stagnation ratio in combination with the preset observation window length; The inertial weight of the current generation is calculated based on the inertial weight baseline value, the dispersion ratio, and the stagnation ratio, and the particle velocity is updated by combining the individual historical best position and the global best position.

[0014] Furthermore, when the population is determined to be in a premature convergence state, the sound field evaluation results corresponding to the current global optimal array are extracted to mark the high-concentration low sound pressure blind zone, and the sound source coordinates of some particles in the population are forcibly migrated to the neighborhood of the high-concentration low sound pressure blind zone to perform directional reseeding. This includes: when the stagnation ratio reaches the upper limit of the observation window and the dispersion ratio is lower than the historical lowest point, the population is determined to be in a premature convergence state; in all grids, grids with droplet weights higher than the average droplet weight of the entire tower and sound pressure levels not reaching the aggregation threshold are screened out, and the spatial region formed by them is marked as the high-concentration low sound pressure blind zone; some particles with lower fitness rankings are selected, and their sound source coordinate components are forcibly migrated to the installable structural surface grid adjacent to the high-concentration low sound pressure blind zone, and the installation angle component is reset towards the center of the high-concentration low sound pressure blind zone to perform directional reseeding.

[0015] Its effect is that, through dynamic adjustment of inertial weight driven by the ratio of stagnation to dispersion based on the population's own operational data, and the blind zone-guided directional reseeding method, the particle swarm can be directionally guided to the neighborhood of the high-concentration, low-sound-pressure blind zone to continue exploring when it falls into premature convergence, thus fundamentally eliminating the blind zone locking problem caused by premature convergence of the standard particle swarm in the high-dimensional mixed space.

[0016] Furthermore, the final sound source array scheme is substituted back into the full-size acoustic finite element model for verification calculations, including: Verify the sound field distribution of the final sound source array scheme under the conditions of actual tower boundary, wall reflection and flue gas medium attenuation; Check whether the sound pressure level in the high-concentration, low-sound-pressure blind zone exceeds the aggregation threshold; If the deviation between the verification result and the evaluation result of the sound field prediction model exceeds the allowable range, the training samples of the sound field prediction model will be updated with the verified sound field data, and the optimization will be re-executed with the current optimal array as the initial population center until the verification is passed.

[0017] Furthermore, after the review is approved, the data is converted into parameter files required for the project construction, including: The optimal array scheme is converted into a parameter file containing the positioning coordinates of the mounting bracket, the load-bearing structure number where the fixing point is located, the direction and diameter allocation of the air source branch pipe, and the operating frequency and start / stop grouping table of each sound source. Sound sources located in the vicinity of high-concentration, low-sound-pressure blind zones are marked as key protected sound sources, while sound sources located in low-concentration zones are marked as intermittently operating sound sources.

[0018] Its effect is that by integrating optimal particle decoding, full-size acoustic verification and engineering parameter file generation into a coherent post-processing flow, it realizes a complete mapping of the optimization results from the algorithm space to the engineering space. The final output installation parameter file can be directly used to guide on-site construction and operation scheduling, eliminating the translation barrier between algorithm output and engineering implementation.

[0019] Furthermore, the effective coverage grid is determined by combining the sound pressure level distribution and the aggregation threshold, including: On a scaled-down experimental platform, the sound pressure level was gradually increased and the aggregation and sedimentation rate was observed using droplets of the dominant particle size measured in the field. The sound pressure level at the point when the aggregation and sedimentation rate begins to show a significant inflection point is taken as the aggregation threshold. Grids with sound pressure levels not lower than the aggregation threshold are marked as effective coverage grids.

[0020] Furthermore, encoding a single array scheme as a particle includes: The first part of the particle position vector is set as the sound source quantity parameter, and the second part is set as the spatial coordinate component and installation angle component of each candidate sound source in turn. The range of installation angle is limited to the interval from vertically upward to horizontally pointing towards the tower axis.

[0021] The beneficial effects are as follows: In the final output array scheme, the sound field energy is concentrated in the area with the highest droplet concentration. There are no longer blind spots with insufficient effective sound pressure in the high-concentration area. Without increasing the total energy consumption of the sound source, the effective area of ​​sound wave aggregation completely covers the droplet-rich space within the tower. This eliminates the constraint of layout mismatch on water collection efficiency and avoids wasting sound energy in ineffective areas, achieving a balance between algorithm convergence quality and engineering water collection objectives in the sound source array layout optimization. Attached Figure Description

[0022] Figure 1This is a flowchart of a sound source array layout optimization method based on particle swarm optimization.

[0023] Figure 2 This is a comparison of the global optimal fitness convergence curves of a particle swarm optimization algorithm based on a resampling mechanism in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] An embodiment of the sound source array arrangement optimization method based on particle swarm optimization algorithm provided by this invention: like Figure 1 As shown, the sound source array layout optimization method based on particle swarm optimization includes the following steps: S1, obtain cooling tower parameters for joint simulation and construct optimization space.

[0026] In one optional embodiment, the structural geometric parameters and operating boundary conditions of the cooling tower are acquired, including the internal contour of the tower body, the elevation of the packing layer, the location of the water collector, the layout of the flue gas duct, and operating parameters such as circulating water flow rate, inlet and outlet water temperature, and ambient temperature and humidity. Based on the acquired parameters, a full-size three-dimensional computational fluid dynamics model of the cooling tower is constructed to jointly simulate the velocity field, temperature field, and humidity field of the rising airflow inside the tower; simultaneously, a discrete phase model is used to track the trajectory of droplets and extract the spatial distribution data of droplet concentration and particle size in the three-dimensional space inside the tower. It should be noted that there are many software platforms in the art that can achieve joint simulation of the flow field and droplet trajectory inside the tower. Those skilled in the art can choose different simulation tools according to actual engineering conditions, and COMSOL is not the only option.

[0027] Furthermore, the space within the cooling tower where sound sources can be placed is discretized into a three-dimensional grid at a predetermined resolution. Each three-dimensional grid records its center coordinates, droplet mass concentration at that location, and whether it is located on a surface where the sound source can be installed. The droplet mass concentration of all three-dimensional grids is normalized and used as the droplet weight for each grid; that is, the higher the concentration of the three-dimensional grid, the greater its weight. These weights are derived entirely from flow field simulation data and do not contain any subjective empirical assignments. Since the spatial distribution of droplet concentration within the cooling tower is not uniform, high-concentration droplet areas are often concentrated in specific annular zones above the packing layer and in the airflow recirculation zone. If the evaluation weights of each three-dimensional grid are the same, the improvement of sound field coverage in high-concentration areas and the improvement of coverage in low-concentration areas will be treated equally during the optimization process, resulting in the algorithm lacking sufficient resolution for the high-concentration areas that truly determine water collection efficiency. The introduction of droplet weights ensures that the sound pressure coverage contribution from high-concentration three-dimensional grids in the subsequent fitness evaluation is far greater than that from low-concentration three-dimensional grids, providing a physical basis consistent with the water collection mechanism for the optimization direction.

[0028] Next, a sound field prediction model is pre-established, taking sound source parameters as input and sound pressure level distribution as output. Using the number of sound sources, their three-dimensional spatial coordinates, and installation angle as input variables, and the sound pressure level distribution of all three-dimensional grids within the tower as output variables, COMSOL is used to perform batch sound field simulations on array schemes covering the range of sound source number and location parameter values ​​in the optimization space. This yields input-output sample pairs with a number objectively determined based on the optimization space dimension and neural network structure; for example, the number of sample pairs can be 500. A neural network is used to fit the input-output mapping relationship, with the training set mean square error not exceeding a preset multiple of the validation set mean square error as the training convergence criterion; for example, this preset multiple can be 1.05. After completing the sound field prediction model, this model replaces full-scale finite element simulation in each fitness evaluation of subsequent particle swarm iterations, compressing the computation time of a single evaluation from hours to milliseconds, making iterative optimization of large-scale populations possible within an engineering-acceptable time window.

[0029] Then, based on the sound field prediction model, an optimization space for the particle swarm optimization algorithm is constructed: a single array scheme is encoded as a particle, with the first part of the particle position vector set as the number of sound sources, and the second part set as the spatial coordinate components and installation angle components of each candidate sound source in sequence; the range of sound source coordinate values ​​is limited to the grid set corresponding to the installable structural surface, and the range of installation angle values ​​is limited to the interval from vertically upward to horizontally pointing towards the tower axis. Thus, the abstract array arrangement problem is transformed into a particle optimization problem in a spatially constrained high-dimensional hybrid space. The boundary of the optimization space is objectively determined by the physical installation conditions of the cooling tower, without relying on manually preset empirical ranges.

[0030] By integrating flow field simulation, droplet distribution extraction, and the construction of a fast neural network prediction model into a unified preprocessing workflow, an optimization space that simultaneously carries acoustic response mapping and droplet concentration spatial information is established. This enables each fitness evaluation of the subsequent particle swarm optimization algorithm to obtain dual information on sound field distribution and droplet distribution in millisecond-level response time, laying a data foundation for achieving accurate matching between sound field energy distribution and droplet distribution.

[0031] S2, calculate the fitness of the particle to be evaluated.

[0032] In an optional embodiment, for each particle to be evaluated in the optimization space, the particle is decoded into a specific array scheme, input into the constructed sound field prediction model, and the sound pressure level distribution formed by the array at each three-dimensional grid within the tower is quickly obtained. The effective coverage grid is determined by combining the sound pressure level distribution with the aggregation threshold. It is then determined whether each three-dimensional grid reaches the effective aggregation sound pressure, and the three-dimensional grids with sound pressure levels not lower than the aggregation threshold are marked as effective coverage grids. This aggregation threshold is not set manually, but is determined by droplet aggregation mechanism experiments: on a scaled-down experimental platform, using field-measured dominant droplets as the object, the sound pressure level is gradually increased while observing the aggregation and settling rate. The sound pressure level corresponding to the point where the aggregation and settling rate begins to show a significant inflection point is taken as the aggregation threshold, strongly binding the threshold to the physical response characteristics of the droplets and eliminating the human bias introduced by empirically set thresholds.

[0033] Furthermore, the fitness function is defined as:

[0034] In the formula, To effectively cover the accumulated value of the fog droplet weights corresponding to the grid, the value is obtained by summing the fog droplet weights of all effectively covered grids element by element. The total droplet weight is the sum of all droplet weights in the tower, obtained by summing the droplet weights of all three-dimensional grids element by element; the ratio of the sum of droplet weights corresponding to the effectively covered grid to the total droplet weight is taken as the droplet weight coverage ratio. The actual number of sound sources used for the particles to be evaluated; The upper limit for the number of sound sources is constrained by installation space and gas supply capacity; This represents the energy consumption penalty coefficient. The first term of the fitness function is the droplet weight coverage ratio. It only generates a benefit when the sound field energy falls exactly on the three-dimensional grid with high droplet concentration. The sound pressure energy falling on the low concentration area does not generate a benefit. Therefore, the fitness level is directly equivalent to the degree of matching between the sound field energy distribution and the droplet distribution, providing an evaluation benchmark consistent with the water collection mechanism for subsequent optimization.

[0035] Next, the energy consumption penalty coefficient This is based on objective reverse engineering, not on subjective experience-based settings. It determines the number of sound sources operating at full power that the air compressor system can support under rated operating conditions with its maximum stable air supply capacity. The value is obtained by dividing the rated air supply of the air compressor nameplate by the full-power air consumption of a single sound source and rounding down. The minimum number of sound sources in the optimization space is objectively determined by the lower limit of the number of sound sources in the optimization space. Indicates the number of sound sources from Increase to full power operation quantity At that time, the maximum gain that can be obtained by the droplet weight coverage ratio is determined by adjusting the number of sound sources in the training samples. All simulation records within the interval are obtained by taking the difference between the maximum and minimum coverage ratios. From this, the energy consumption penalty coefficient can be objectively calculated. The relation is:

[0036] When the number of sound sources increases from Increase to At that time, based on the actual number of sound sources used for the particles to be evaluated, the upper limit of the number of sound sources constrained by installation space and air supply capacity, and the energy consumption penalty coefficient, the energy consumption penalty term is calculated, and the increment of the penalty term is... It is exactly equal to the maximum coverage gain that can be obtained by simply stacking sound sources within that range. This mathematically eliminates the option of blindly increasing the number of sound sources, forcing the algorithm to improve fitness by optimizing spatial layout rather than simply adding more hardware. The difference between the fog droplet weight coverage ratio and the energy consumption penalty term is used as the fitness of the particle being evaluated.

[0037] In this way, a fitness function was constructed that simultaneously incorporates the matching degree between sound field energy and droplet spatial distribution as well as the air supply boundary constraints into the evaluation. This ensures that the particle swarm always focuses on eliminating insufficient sound pressure in high-concentration areas as the core of optimization during the iteration process, rather than simply pursuing the numerical improvement of the average sound pressure level across the entire field. This corrects the deviation between the optimization direction and the water collection target in the standard scheme from the evaluation level.

[0038] S3, adjust the inertia weights and perform targeted reseeding.

[0039] In an optional embodiment, during the particle swarm iteration process, the velocity update formula of the standard particle swarm algorithm is:

[0040] In the formula, and The first The particle in the first The speed and position of the generation For inertial weights, , As a learning factor, , It is a random number. This represents the optimal position in the individual history of this particle. This represents the globally optimal position for the population. Inertia weight. Typically, a fixed value is taken or the value decreases linearly with the number of iterations; the particle's search behavior is entirely determined by... and Traction occurs when the global optimal position corresponds to a premature local solution; the entire population will rapidly converge on it and lose the ability to explore other regions.

[0041] Furthermore, to address this deficiency, this invention makes two improvements to the velocity update method. The first is to dynamically adjust the inertia weight based on the ratio of dispersion to stagnation of population fitness to update the particle velocity. After each iteration, the standard deviation of all particle fitness values ​​is used to update the velocity. Characterizes the current dispersion of the population, combined with the fitness standard deviation of the initial population. Calculate the dispersion ratio Simultaneously record the global optimal fitness. Stagnant Algebras with No Continuous Promotion Combined with the preset observation window length Calculate the degree of stagnation The inertia weight for the current generation is calculated based on the baseline inertia weight, the dispersion ratio, and the stagnation ratio, and updated using the following formula:

[0042] In the formula, For example, the existing inertial weight benchmark values ​​for standard particle swarm optimization. The value can be 0.729. This is calculated using the above formula each time. Then, if the calculation result is lower than the preset lower limit value Then Forced to For example, The value of can be 0.1. In this relation, The larger the value, the deeper the population has stagnated, corresponding to the fact that the global optimal position has not changed substantially for a long time. At this time, the inertia weight is reduced to inhibit the inertia of particles along the current direction and encourage particles to respond more to the attraction between the individual's historical optimal position and the global optimal position. A larger value indicates a higher degree of dispersion in the population fitness, meaning the particles are still exploring within a larger range. In this case, the inertia weight is increased to maintain the exploration capability. Both adjustment values ​​are derived from the population's own operational data, without introducing external empirical constants, ensuring that the change in the inertia weight always matches the actual search state of the population.

[0043] When the degree of stagnation is greater than (The stagnant algebra has reached the upper limit of the observation window) and the degree of dispersion is higher than Below the current population iteration history When the minimum value is reached (i.e., the current dispersion is lower than the historical low), the population is judged to have entered a premature convergence state, triggering the second improvement, namely the blind zone-guided directional reseeding method. After triggering the premature convergence judgment, the current population structure is retained, the sound field evaluation results corresponding to the current global optimal array are extracted, and the grids with droplet weights higher than the average droplet weight of the entire tower and sound pressure levels not reaching the aggregation threshold are screened out in all three-dimensional grids. The spatial region formed by these grids is marked as a high-concentration low sound pressure blind zone. Then, some particles with lower fitness rankings are selected, and their sound source coordinate components are forcibly migrated to the installable structural surface grid adjacent to the high-concentration low sound pressure blind zone. The installation angle component is reset towards the center of the high-concentration low sound pressure blind zone to perform directional reseeding. The remaining components retain their original values, and the iteration continues according to the velocity update formula.

[0044] The essence of premature convergence is that the population has never effectively sampled the area where the blind zone is located. Although random re-initialization restores diversity, the probability of the blind zone being sampled again remains very low. However, directing some particles to the neighborhood of the blind zone is equivalent to directly injecting the mismatch information between the sound field energy distribution and the droplet distribution into the population. This allows the particle swarm to conduct targeted exploration of the energy gap area while retaining existing search results, fundamentally ensuring that the optimization process always revolves around the ultimate goal of eliminating the sound pressure blind zone in the high-concentration area. It should be noted that the specific proportion of the later particles can be flexibly adjusted by those skilled in the art according to the population size and premature convergence trigger frequency. This invention does not impose a mandatory limit on this proportion.

[0045] Thus, through the dynamic adjustment of inertial weights driven by the stagnation ratio and dispersion ratio, and the synergistic effect of the blind zone-guided directional reseeding method, the particle swarm can be directionally guided to the neighborhood of the high-concentration, low-sound-pressure blind zone to continue exploring when it falls into premature convergence. This fundamentally eliminates the blind zone locking problem caused by premature convergence of the standard particle swarm in high-dimensional mixed space, ensuring that the optimization trajectory always remains consistent with the engineering goal of eliminating insufficient sound pressure in the high-concentration region.

[0046] S4 decodes the optimal particle kernel and converts the parameter file.

[0047] In an optional embodiment, when the population satisfies the convergence criterion, i.e., the global optimal fitness is within a preset observation window... No further improvements are needed. When the temperature falls back to the normal range above the historical low, the current global best particle is extracted and decoded into the final sound source array scheme. This scheme includes the total number of sound sources, the three-dimensional installation coordinates of each sound source within the cooling tower, the installation orientation of each sound source, and the layer affiliation of each sound source.

[0048] Furthermore, the final sound source array scheme is substituted back into a full-scale acoustic finite element model for verification calculations. This verifies the sound field distribution of the final sound source array scheme under the conditions of real tower boundaries, wall reflections, and flue gas medium attenuation, with a focus on checking whether the sound pressure level in the high-concentration, low-sound-pressure dead zone exceeds the aggregation threshold. Since the sound field prediction model is obtained by neural network fitting, it has certain interpolation errors in the sparse sample region near the optimization space boundary. If the deviation between the verification result and the evaluation result of the sound field prediction model exceeds the allowable range, the training samples of the sound field prediction model are updated with the verified sound field data, and the optimization is re-executed with the current optimal array as the initial population center until the verification passes. This iterative verification process ensures that the sound field distribution of the final output array scheme under the real physical boundary is consistent with the algorithm evaluation result, eliminating the influence of neural network prediction errors on the convergence quality.

[0049] Next, after verification, the optimal array scheme is converted into parameter files required for engineering construction. These files include the positioning coordinates of the mounting brackets, the load-bearing structure numbers where the fixing points are located, the routing and diameter allocation of the gas source branch pipes, and the operating frequency and start / stop grouping table for each sound source. Sound sources located in the vicinity of the high-concentration, low-sound-pressure blind zone are marked as key protection sources, ensuring priority gas supply pressure during operation scheduling. Sound sources located in the low-concentration zone are marked as intermittently operating sources, allowing for priority shutdown under low-load conditions to save gas consumption. Through this grouping and labeling, the differentiated evaluation of high-concentration and low-concentration zones during the optimization phase is mapped to differentiated scheduling methods during the operation phase, extending the engineering value of the optimization results to the daily operation and maintenance level.

[0050] In this way, by integrating optimal particle decoding, full-size acoustic verification, and engineering parameter file generation into a coherent post-processing workflow, a complete mapping of the optimization results from the algorithm space to the engineering space is achieved. The final output parameter file can be directly used to guide on-site construction and operation scheduling, eliminating the translation barrier between algorithm output and engineering implementation.

[0051] like Figure 2 As shown, it can be observed that the existing methods tend to flatten out after completing the main convergence in the early stage of iteration, and the subsequent gain is extremely limited; while the method of the present invention shows a clear step-like leap in the convergence curve after each resampling trigger, and the final global optimal fitness is significantly higher than that of the existing methods. This indicates that the resampling mechanism of the present invention can effectively break the convergence stagnation caused by the decline of population diversity and continuously promote the improvement of global optimization ability.

[0052] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing the arrangement of sound source arrays based on particle swarm optimization, characterized in that, include: S1. Obtain the structural geometric parameters and operating boundary conditions of the cooling tower for joint simulation to extract droplet concentration distribution data. Normalize the droplet concentration and use it as the droplet weight. Establish a sound field prediction model with sound source parameters as input and sound pressure level distribution as output. Construct the optimization space for the particle swarm optimization algorithm based on the sound field prediction model. S2. For the particles to be evaluated in the optimization space, decode them into array schemes and input them into the sound field prediction model to obtain the sound pressure level distribution. Combine the sound pressure level distribution with the aggregation threshold to determine the effective coverage grid. Based on the accumulated droplet weights corresponding to the effective coverage grid, the accumulated droplet weights of the entire tower, and the energy consumption penalty term objectively deduced from the air supply boundary, calculate the particle weights to be evaluated. S3, during the particle swarm iteration process, the inertia weight is dynamically adjusted based on the dispersion ratio and stagnation ratio of the population fitness to update the particle velocity. When the population is determined to be in a premature convergence state, the sound field evaluation result corresponding to the current global optimal array is extracted to mark the high-concentration low sound pressure blind zone, and the sound source coordinates of some particles in the population are forcibly migrated to the neighborhood of the high-concentration low sound pressure blind zone to perform directional reseeding; S4, when the population meets the convergence criterion, the current global optimal particle is extracted and decoded into the final sound source array scheme. The final sound source array scheme is then substituted back into the full-size acoustic finite element model for verification calculation. After the verification is passed, it is converted into the parameter file required for engineering construction.

2. The sound source array arrangement optimization method based on particle swarm optimization algorithm according to claim 1, characterized in that, The structural geometric parameters and operating boundary conditions of the cooling tower are obtained through joint simulation to extract droplet concentration distribution data. The droplet concentration is normalized and used as the droplet weight, including: A three-dimensional computational fluid dynamics model is constructed by obtaining the internal outline of the cooling tower, the elevation of the packing layer, the location of the water collector, the layout of the flue gas duct, and the operating parameters. The velocity field, temperature field, and humidity field of the rising airflow inside the tower were jointly simulated, and the discrete phase model was used to track the trajectory of the droplets and extract the spatial distribution data of droplet concentration and particle size in the three-dimensional space inside the tower. The space within the cooling tower where sound sources can be placed is discretized into a three-dimensional grid at a predetermined resolution, and the droplet mass concentration of all three-dimensional grids is normalized and used as the droplet weight of each three-dimensional grid.

3. The sound source array layout optimization method based on particle swarm optimization algorithm according to claim 1, characterized in that, A sound field prediction model is established, taking sound source parameters as input and sound pressure level distribution as output. Based on this model, an optimization space for the particle swarm optimization algorithm is constructed, including: Using the number of sound sources, the three-dimensional spatial coordinates of each sound source, and the installation angle as input variables, and the sound pressure level distribution of all grid units in the tower as output variables, the input and output sample pairs are obtained by batch sound field simulation of array schemes with the range of sound source number and location parameter values ​​in the coverage optimization space. A neural network is used to fit the input-output mapping relationship to complete the construction of the sound field prediction model; Encode a single array scheme as a particle, and limit the range of sound source coordinates to the grid set corresponding to the installable structural surface to construct the optimization space of the particle swarm algorithm.

4. The sound source array arrangement optimization method based on particle swarm optimization algorithm according to claim 1, characterized in that, The fitness of the particle to be evaluated is calculated based on the cumulative droplet weights corresponding to the effective coverage grid, the cumulative droplet weights of the entire tower, and the energy consumption penalty term objectively deduced from the air supply boundary. This includes: using the ratio of the cumulative droplet weights corresponding to the effective coverage grid to the cumulative droplet weights of the entire tower as the droplet weight coverage ratio; objectively deducing the energy consumption penalty coefficient based on the number of sound sources operating at full power that the air compressor system can support under rated operating conditions with maximum stable air supply capacity, and the maximum gain that the droplet weight coverage ratio can obtain when the number of sound sources increases to the number of sound sources operating at full power; calculating the energy consumption penalty term based on the actual number of sound sources used by the particle to be evaluated, the upper limit of the number of sound sources constrained by installation space and air supply capacity, and the energy consumption penalty coefficient; and using the difference between the droplet weight coverage ratio and the energy consumption penalty term as the fitness of the particle to be evaluated.

5. The sound source array arrangement optimization method based on particle swarm optimization algorithm according to claim 1, characterized in that, Dynamically adjusting inertia weights based on the population fitness dispersion ratio and stagnation ratio to update particle velocities includes: The current dispersion of the population is characterized by the standard deviation of the fitness values ​​of all particles, and the dispersion ratio is calculated by combining the standard deviation of the fitness values ​​of the initial population. Record the number of stagnant generations in which the global optimal fitness has not been improved, and calculate the stagnation ratio in combination with the preset observation window length; The inertial weight of the current generation is calculated based on the inertial weight baseline value, the dispersion ratio, and the stagnation ratio, and the particle velocity is updated by combining the individual historical best position and the global best position.

6. The sound source array layout optimization method based on particle swarm optimization algorithm according to claim 1, characterized in that, When the population is determined to be in a premature convergence state, the sound field evaluation results corresponding to the current global optimal array are extracted to mark the high-concentration low sound pressure blind zone. The sound source coordinates of some particles in the population are forcibly migrated to the neighborhood of the high-concentration low sound pressure blind zone to perform directional reseeding. This includes: when the stagnation ratio reaches the upper limit of the observation window and the dispersion ratio is lower than the historical low point, the population is determined to be in a premature convergence state; in all grids, grids with droplet weights higher than the average droplet weight of the entire tower and sound pressure levels not reaching the aggregation threshold are screened out, and the spatial region formed by them is marked as the high-concentration low sound pressure blind zone; some particles with lower fitness rankings are selected, and their sound source coordinate components are forcibly migrated to the installable structural surface grid adjacent to the high-concentration low sound pressure blind zone, and the installation angle component is reset towards the center of the high-concentration low sound pressure blind zone to perform directional reseeding.

7. The sound source array arrangement optimization method based on particle swarm optimization algorithm according to claim 1, characterized in that, The final sound source array scheme was then substituted back into a full-size acoustic finite element model for verification calculations, including: Verify the sound field distribution of the final sound source array scheme under the conditions of actual tower boundary, wall reflection and flue gas medium attenuation; Check whether the sound pressure level in the high-concentration, low-sound-pressure blind zone exceeds the aggregation threshold; If the deviation between the verification result and the evaluation result of the sound field prediction model exceeds the allowable range, the training samples of the sound field prediction model will be updated with the verified sound field data, and the optimization will be re-executed with the current optimal array as the initial population center until the verification is passed.

8. The method for optimizing the arrangement of sound source arrays based on particle swarm optimization as described in claim 1, characterized in that, After review and approval, the parameters are converted into the necessary files for project construction, including: The optimal array scheme is converted into a parameter file containing the positioning coordinates of the mounting bracket, the load-bearing structure number where the fixing point is located, the direction and diameter allocation of the air source branch pipe, and the operating frequency and start / stop grouping table of each sound source. Sound sources located in the vicinity of high-concentration, low-sound-pressure blind zones are marked as key protected sound sources, while sound sources located in low-concentration zones are marked as intermittently operating sound sources.

9. The method for optimizing the arrangement of sound source arrays based on particle swarm optimization as described in claim 1, characterized in that, The effective coverage grid is determined by combining sound pressure level distribution and aggregation threshold, including: On a scaled-down experimental platform, the sound pressure level was gradually increased and the aggregation and sedimentation rate was observed using droplets of the dominant particle size measured in the field. The sound pressure level at the point when the aggregation and sedimentation rate begins to show a significant inflection point is taken as the aggregation threshold. Grids with sound pressure levels not lower than the aggregation threshold are marked as effective coverage grids.

10. The sound source array arrangement optimization method based on particle swarm optimization algorithm according to claim 3, characterized in that, Encoding a single array scheme as a particle includes: The first part of the particle position vector is set as the sound source quantity parameter, and the second part is set as the spatial coordinate component and installation angle component of each candidate sound source in turn. The range of installation angle is limited to the interval from vertically upward to horizontally pointing towards the tower axis.

Images