A Substation Ventilation Noise Reduction Design Method Based on Deep Reinforcement Learning

Abstract

This invention discloses a substation ventilation noise reduction design method based on deep reinforcement learning, specifically relating to the field of environmental noise control in power substations. It addresses the technical problem that the increased ventilation resistance after the stacking of silencing components forces the fan to deteriorate, leading to increased noise and making it difficult to simultaneously meet heat dissipation and noise reduction requirements. The method determines the required ventilation volume through thermal balance calculations, calculates the total system resistance and matches the fan operating point, further calculates the resistance distribution entropy and peak surface velocity, and inputs these values ​​into a probabilistic regression model with monotonic constraints. The model outputs the probability value of increased noise to screen candidate combinations. Then, under ventilation and risk constraints, deep reinforcement learning is used to collaboratively optimize the opening location, component combination, and channel parameters, outputting the ventilation and silencing arrangement. Finally, the method solidifies the construction and operation parameters through fluid and acoustic studies, thereby improving the controllability and consistency of the solution and reducing repeated on-site adjustments.

Description

Technical Field

[0001] This invention relates to the field of environmental noise control in power substations, and more specifically, to a substation ventilation noise reduction design method based on deep reinforcement learning. Background Technology

[0002] In urban neighborhoods, indoor or underground substations typically require continuous natural or mechanical ventilation to remove operating heat from the main transformer room to ensure safe equipment operation. Simultaneously, noise generated by the main transformer and its cooling ventilation equipment easily propagates outwards through doors, windows, and air intake / exhaust ducts, becoming a major source of complaints from the surrounding environment. A common practice in existing projects is to install ventilation and sound-absorbing louvers at the air intake / exhaust openings, along with various measures such as ventilation and sound-absorbing doors, sound-absorbing walls, and air intake ducts to suppress noise leakage. For example, patent CN201510320917.4 discloses a substation noise reduction and silencing system that arranges ventilation and sound-absorbing louvers at the air intake / exhaust openings and superimposes sound-absorbing structures to achieve overall noise reduction while satisfying heat dissipation and ventilation requirements.

[0003] However, the aforementioned existing technologies generally emphasize blocking or absorbing sound, often treating silencing components as passive elements that can be directly stacked, easily overlooking the fact that a ventilation system is an integrated whole where airflow, resistance, fan operating conditions, noise, and energy consumption are interdependent: when louvers, baffles, sound-absorbing materials, or longer channels are added to the air inlet and outlet, the resistance of the airflow channel will increase accordingly. In order to maintain the ventilation capacity required for heat dissipation, engineering projects are forced to either increase the fan speed or select a larger power fan; the increase in fan speed and load will bring more obvious aerodynamic and mechanical noise, and the high flow velocity passing over the edges of louvers and baffles may also induce new aerodynamic noise. As a result, a phenomenon that is not intuitive but does exist may occur—it seems that adding more silencing components should make it quieter, but in reality, the increased resistance pushes the fan into a noisier operating state, and may even lead to insignificant improvement or even deterioration of boundary noise; and under the fluctuation of actual operating conditions such as load changes and temperature changes, fixed empirical combination designs are more difficult to balance the continuous optimization of heat dissipation and noise reduction, making it difficult to implement stably and controllably.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a substation ventilation noise reduction design method based on deep reinforcement learning. This method determines the required ventilation volume through thermal balance calculations, calculates the total system resistance and matches the fan operating point, further calculates the resistance distribution entropy and peak surface velocity, and inputs these values ​​into a probabilistic regression model with monotonic constraints. The output noise inverse probability value is used to screen candidate combinations. Then, under ventilation and risk constraints, deep reinforcement learning is used to collaboratively optimize the opening location, component combination, and channel parameters, outputting the ventilation and noise reduction layout. The construction and operation parameters are then solidified through fluid and acoustic studies, thereby improving the controllability and consistency of the solution and reducing repeated on-site adjustments, thus solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] S1: Obtain the heat output and allowable temperature rise limit of the equipment, and calculate the required ventilation volume based on the existing heat balance ventilation volume calculation method;

[0008] S2: Calculate the ventilation resistance of the candidate component combination based on the required ventilation volume to obtain the total system resistance, and determine the fan operating point by combining the fan performance curve. Extract the resistance distribution entropy and peak surface velocity, input the resistance distribution entropy and peak surface velocity into a probabilistic regression model with monotonic constraints, output the noise inverse probability value, and then compare it with the preset risk threshold to determine the availability of the candidate component combination.

[0009] S3: The combination of candidate components deemed usable, along with the required ventilation volume, total system resistance, and fan operating point, is input into a deep reinforcement learning decision-making process, and a ventilation and noise reduction layout scheme is output based on the boundary noise prediction results.

[0010] S4: Perform computational fluid dynamics and acoustic joint verification of the ventilation and noise reduction layout scheme, and generate construction parameters and operation control parameters based on the verification results.

[0011] Furthermore, step S1 includes:

[0012] The total heat generation of the equipment is obtained by adding the heat generation of the main transformer to the rated heating power of the low-voltage switchgear and auxiliary equipment; the upper limit of the allowable temperature rise is determined by looking up the maximum allowable temperature rise value according to the insulation thermal class of the main transformer, and then comparing the maximum allowable temperature rise value with the corresponding indoor allowable temperature rise control value of the industry standard, and taking the smaller value.

[0013] Furthermore, step S1 also includes:

[0014] The required ventilation volume is calculated by dividing the total heat output of the equipment by the product of the specific heat capacity of air at constant pressure, the density of the incoming air, the upper limit of the allowable temperature rise, and the heat exchange efficiency coefficient. The required ventilation volume is set as a hard constraint condition for the calculation of the combined resistance of candidate components and the determination of the fan operating conditions.

[0015] Furthermore, step S2 includes:

[0016] The total system resistance is calculated by summing the friction resistance of each straight channel and the local resistance of each louver, baffle, elbow and sound absorber along the airflow path; the operating point of the fan is determined by the intersection of the fan performance curve and the system characteristic curve formed by the relationship between the total system resistance and the flow rate.

[0017] Furthermore, step S2 also includes:

[0018] The pressure drop distribution entropy is obtained by calculating the pressure drop ratio sequence based on the ratio of the pressure drop of each channel segment or component to the total system resistance, and the pressure drop ratio sequence is obtained by Shannon information entropy calculation; the peak face velocity is calculated based on the flow distribution results of the ventilation network, and the maximum value is taken as the peak face velocity.

[0019] Furthermore, step S2 also includes:

[0020] The inverse risk coefficient is obtained by inputting the drag distribution entropy and the peak surface velocity into a probabilistic regression model with monotonic constraints. The model imposes a risk non-decreasing constraint on the peak surface velocity and a risk non-increasing constraint on the drag distribution entropy. The inverse risk coefficient is compared with a preset risk threshold to determine the availability of candidate component combinations.

[0021] Furthermore, step S3 includes:

[0022] The deep reinforcement learning environment adopts the Markov decision process framework. The state space includes the required ventilation volume and total system resistance, the fan operating point, the current component combination code, and operating parameters. The action space corresponds to the available candidate component combination index. After the action is executed, the total system resistance and the fan operating point are deterministically updated.

[0023] Furthermore, step S3 also includes:

[0024] The boundary noise prediction value is obtained by accumulating the insertion loss of the silencing components and superimposing the logarithmic sum of the noise source strength of the fan and the main transformer. The reward function is mainly based on the negative boundary noise prediction value and secondarily penalizes the excessive resistance and insufficient ventilation margin. The PPO algorithm drives the agent interaction optimization and outputs the ventilation and silencing arrangement scheme with the optimal component combination.

[0025] Furthermore, step S4 includes:

[0026] The computational fluid dynamics model uses the finite volume method and hybrid mesh generation. Boundary conditions include velocity inlet, pressure outlet, and wind-driven reference frame. The flow field simulation uses the Realizable k-ε turbulence model and SIMPLE algorithm. The actual ventilation volume is checked to be no less than the required ventilation volume, and the simulation system's total resistance and the fan's operating point are matched.

[0027] Furthermore, step S4 also includes:

[0028] The acoustic simulation uses a hybrid CFD / CAE method and FW-H acoustic analog integral to calculate boundary noise. The A-weighted sound pressure level is checked to ensure it meets the predicted value and specifications. Based on the check results, the precise dimensional construction parameters of the noise reduction components and the control parameters of the variable frequency operation curve of the fan are generated.

[0029] The technical effects and advantages of the substation ventilation noise reduction design method based on deep reinforcement learning in this invention are as follows:

[0030] This invention uses thermal balance calculations to determine the required ventilation volume as a unified constraint and resistance calculations and fan performance matching to determine the operating condition point as an engineering feasibility benchmark. Based on this, it introduces resistance distribution entropy and peak surface velocity to pre-judge the risk of noise amplification in candidate component combinations. This transforms the implicit problems of resistance increase caused by the superposition of silencing components, forcing the fan into a high-noise operating state, and local high-speed-induced spontaneous dynamic noise from empirical judgments into calculable and screenable intermediate constraints. This provides a stable boundary for subsequent deep reinforcement learning to optimize ventilation and silencing arrangements within the feasible domain. Finally, through fluid mechanics and acoustics, it solidifies into a constructable and operable parameter system, ensuring that ventilation capacity, fan operating status, and boundary noise control converge consistently within the same technical chain. This reduces the probability of noise amplification or repeated rework under different operating conditions, improving the controllability, stability, and consistency of ventilation and noise reduction design in engineering implementation. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating a substation ventilation noise reduction design method based on deep reinforcement learning according to the present invention. Detailed Implementation

[0032] 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 embodiments of the present invention, and not all embodiments. 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.

[0033] Example 1: Figure 1 This invention presents a substation ventilation noise reduction design method based on deep reinforcement learning, comprising:

[0034] S1: Obtain the heat output and allowable temperature rise limit of the equipment, and calculate the required ventilation volume based on the existing heat balance ventilation volume calculation method, which serves as the ventilation constraint for subsequent design solutions.

[0035] S2: Calculate the ventilation resistance of the candidate component combination based on the required ventilation volume to obtain the total system resistance, and determine the fan operating point by combining the fan performance curve. Extract the resistance distribution entropy and peak surface velocity, input the resistance distribution entropy and peak surface velocity into a probabilistic regression model with monotonic constraints, output the noise inverse probability value, and then compare it with the preset risk threshold to determine the availability of the candidate component combination.

[0036] S3: The candidate component combinations that are determined to be available are combined with the required ventilation volume, total system resistance and fan operating conditions. The input is a deep reinforcement learning decision process, and the ventilation and noise reduction layout scheme is output based on the boundary noise prediction results.

[0037] S4: Perform computational fluid dynamics and acoustic joint verification of the ventilation and noise reduction layout scheme, and generate construction parameters and operation control parameters based on the verification results.

[0038] This invention proposes a ventilation and noise reduction design method for substation equipment. Addressing the systemic contradiction in existing engineering projects where noise reduction components increase ventilation resistance, forcing fans into higher noise operating states, the method establishes a complete design chain from determining ventilation requirements and screening candidate solutions to collaborative optimization and engineering verification. First, the method determines the ventilation requirements to ensure heat dissipation safety based on traditional heat balance principles, treating this as a rigid constraint throughout the entire process. Then, under this constraint, resistance and fan operating conditions are matched for candidate ventilation and noise reduction component combinations. Simultaneously, risk characteristics reflecting resistance structures and extreme local flow states are extracted. A comprehensive analysis yields a risk assessment result, and threshold judgment is performed to preemptively eliminate high-risk combinations that could lead to increased noise. Based on this, the available combinations are input into a deep reinforcement learning decision optimization process. Under the premise of satisfying ventilation constraints and operational controllability, a ventilation and noise reduction layout scheme is output. The scheme is then solidified into constructable and operable parameters through joint fluid and acoustic verification. Through this overall concept, this invention achieves collaborative constraints and consistent implementation of ventilation capacity, fan operating status, and boundary noise control, improving the controllability and engineering applicability of substation ventilation and noise reduction design.

[0039] The heat generated during the operation of the main transformer room in indoor or underground substations needs to be continuously ventilated to ensure equipment safety. At the same time, noise control requirements strictly limit external propagation. The existing design of stacked sound-absorbing components is prone to increasing fan noise due to increased resistance. In order to avoid the mutual constraint between ventilation and noise reduction, it is necessary to first accurately determine the ventilation volume required for heat dissipation as a hard constraint, so as to provide a unified benchmark for subsequent component combination selection and optimization decisions.

[0040] S101: Obtain the heat generated by the equipment.

[0041] The total heat generation of the main transformer room is the core basis for determining ventilation requirements. All heat sources must be fully covered and actual operating load changes must be considered in order to avoid the ventilation volume calculation deviating from the actual heat dissipation requirements.

[0042] First, collect the rated capacity, load loss, no-load loss, and current or predicted load rate of the main transformer. The calculation process for the heat generation of the main transformer is as follows:

[0043] The no-load loss is kept constant and used directly. For the load loss, the load rate is first squared, then multiplied by the load loss value. The result of the squared calculation is then added to the no-load loss to obtain the heat output of the main transformer. Simultaneously, the rated heat output of the low-voltage switchgear and auxiliary equipment is collected. The total equipment heat output is calculated by directly adding the heat output of the main transformer to the rated heat output of the low-voltage switchgear and auxiliary equipment.

[0044] For example, the on-site engineer reads the no-load loss from the nameplate of the main transformer as a fixed value, obtains the load loss value from the equipment manual, squares the load rate provided by the dispatcher on the day, multiplies it by the load loss, and adds the no-load loss to obtain the heat generation of the main transformer. Then, the heat dissipation data of the low-voltage switchgear is summarized, and finally, the total equipment heat generation record table is formed.

[0045] S102: Obtain the maximum allowable temperature rise.

[0046] After determining the total heat output of the equipment, it is necessary to establish the maximum permissible rise in indoor air temperature relative to the intake air temperature. This serves as the temperature boundary for the safe operation of the equipment, ensuring that the ventilation design neither exceeds the temperature limit nor leaves a margin.

[0047] The process for obtaining the upper limit of allowable temperature rise is as follows: First, based on the insulation thermal class of the main transformer (oil-immersed or dry-type), consult the technical data provided by the equipment manufacturer to obtain the maximum allowable oil temperature or winding temperature rise value; at the same time, refer to the design value of the ambient air temperature in the area where the substation is located; then calculate the difference between the maximum allowable temperature rise and the design value of the ambient air temperature, and compare it with the national or industry standards for transformer operating temperature, taking the smaller value as the final value.

[0048] For example, the engineer opens the main transformer technical manual, finds the winding temperature rise limit corresponding to the insulation class, calculates the difference based on local summer extreme temperature records, consults power industry standards for confirmation, and finally fills in the allowable upper limit value of temperature rise in the design form.

[0049] S103: Calculate the required ventilation volume.

[0050] Once the total heat generation of the equipment and the upper limit of the allowable temperature rise are determined, the air volume flow rate required to maintain the indoor temperature below the upper limit can be accurately calculated by using the principle of convective heat transfer balance of air carrying heat, thus transforming the heat dissipation requirement into a unified ventilation constraint.

[0051] The required ventilation volume is calculated using a heat balance equation. First, the product of the specific heat capacity of air at constant pressure, the density of the incoming air, and the upper limit of the allowable temperature rise is calculated. Then, the heat transfer efficiency coefficient is introduced and multiplied by this product to form the denominator. The numerator is the total heat generation of the equipment. Finally, the total heat generation of the equipment is divided by this denominator to obtain the required ventilation volume. Among these, the specific heat capacity of air at constant pressure is taken as a fixed value under standard atmospheric conditions, and the density of the incoming air is corrected according to altitude and temperature.

[0052] For example, the engineer fills in the total heat generation of the equipment calculated above in the calculation table as the numerator; looks up the specific heat capacity of air at constant pressure and the density of incoming air in the table, multiplies the three (specific heat capacity of air at constant pressure, density of incoming air, and upper limit of allowable temperature rise) in sequence, and then multiplies them by the heat exchange efficiency coefficient to obtain the denominator; after dividing the numerator by the denominator, the required ventilation volume value is recorded in the table and marked as a hard constraint condition for the subsequent combination of all components and matching of the fan.

[0053] Step S1 completes the calculation of total equipment heat generation, determines the upper limit of allowable temperature rise, and calculates the required ventilation volume based on the principle of heat balance, forming a precise quantitative constraint on heat dissipation requirements, ensuring that the entire ventilation and noise reduction design process always prioritizes meeting the safe operating temperature requirements of the main transformer room equipment.

[0054] Step S1 has determined the required ventilation volume as a hard constraint for heat dissipation. However, the stacking of noise-absorbing components can easily lead to increased fan noise due to uneven resistance distribution or localized high flow velocities. In order to eliminate risky combinations in advance, it is necessary to quantify the resistance structure and flow field characteristics of candidate component combinations under the required ventilation volume constraint, form comparable criteria, and ensure that subsequent deep reinforcement learning is performed within a feasible domain without the risk of increased noise.

[0055] The resistance distribution entropy, in a dimensionless form, characterizes the concentration of total system resistance among various components and channel sections. It can distinguish between resistance spikes caused by a few components and the essential difference between uniformly distributed resistance and similar total resistance, thus making explicit the risk of deterioration caused by local high resistance, which is often overlooked in existing designs. The peak surface velocity, in meters per second, characterizes the most unfavorable local surface wind speed extreme value at the critical section. It can directly capture local high-speed and strong turbulent states that cannot be reflected by the average wind speed, providing necessary preliminary criteria for identifying spontaneous dynamic noise and additional pressure loss that may occur in louvers, deflectors, and narrow sections. The two complement each other—one structural and one extreme—allowing for the quantification of implicit risks beyond total resistance and operating points into identifiable intermediate quantities without increasing the amount of initial data acquisition. This enables the early screening of high-risk component combinations and the convergence constraint of the feasible region, reducing ineffective exploration in subsequent optimization and verification, and improving the controllability and consistency of the initial design and engineering implementation.

[0056] S201: The total system resistance is calculated from the ventilation resistance calculation.

[0057] Once the required ventilation volume is determined, the resistance introduced by the candidate component combination must be precisely quantified under a fixed flow constraint in order to reveal the extent to which the silencing components affect the overall pressure drop.

[0058] The ventilation resistance calculation process employs a separate solution algorithm for friction loss and local resistance. First, a ventilation topology is established based on the channel connectivity of candidate component combinations, distinguishing between series channel segments and parallel branches. When the channel is a purely series system, the flow rate for each segment is taken as the required ventilation volume. When parallel branches exist, the branch flow distribution is first determined based on the resistance characteristics of each branch, and then the average flow velocity of each segment is determined accordingly. Subsequently, the friction loss is calculated for each straight channel segment using the Darcy-Weisbach formula, and the local pressure drop is calculated for each local component such as louvers, baffles, elbows, and sound absorbers based on the local resistance coefficient. These components are then accumulated along the actual flow path to obtain the total system resistance.

[0059] For example, engineers draw the channel layout of candidate component combinations in design software, label the straight sections, louvers, and baffles separately, input the length, hydraulic diameter, and roughness of the straight sections, look up the friction coefficient from the table and perform a product operation, input the geometric angle of the louvers, look up the local resistance coefficient from the table and perform a product operation, and finally add up all the pressure drop values ​​and record them in the table to form the total system resistance value.

[0060] S202: Determine the operating point of the fan by combining the fan performance curve.

[0061] After quantifying the total system resistance, it is necessary to determine the actual load state of the fan in order to judge whether the resistance change has driven the fan into the high-noise operating range.

[0062] The process of determining the operating point of a fan includes: obtaining the performance curve of a pre-selected fan; establishing a system characteristic curve based on the resistance calculation results; and determining the intersection of the system characteristic curve and the fan performance curve near the required ventilation volume as the operating point to obtain the corresponding operating air volume, total pressure, speed, and shaft power. When variable frequency control is used, the fan speed is adjusted so that the intersection point satisfies the requirement that the operating air volume is not lower than the required ventilation volume and the total pressure can overcome the total system resistance, thereby determining the achievable operating point.

[0063] S203: Calculate the resistance distribution entropy.

[0064] The total system resistance reflects the overall load, but resistance concentrated in a few components can easily lead to local noise sources, so it is necessary to introduce a distribution uniformity criterion for quantification.

[0065] The calculation of resistance distribution entropy employs the Shannon information entropy algorithm. First, the pressure drop contribution list is reused to calculate the proportion of each pressure drop: each pressure drop value is divided by the total system resistance, resulting in a sequence of proportions with a sum of 1. Then, each proportion is multiplied by its base-2 logarithm, and the negative value is taken, followed by summation to obtain the dimensionless resistance distribution entropy. Items with a proportion of zero are defined as having a contribution of 0. The resistance distribution entropy ranges from 0 to the logarithm of the number of pressure drop contribution items (base-2); a larger value indicates a more uniform resistance distribution.

[0066] A pressure drop contribution list is a structured record table that lists the pressure drop values ​​corresponding to each pressure drop source according to the actual airflow path and local components during the calculation of ventilation resistance for candidate component combinations. In the context of this application, this "list" is not a single parameter, but a detailed collection of multiple pressure drop records. This includes pressure drop along the flow path from straight channels, as well as local pressure drop from components such as louvers, baffles, elbows, and sound absorbers. Each record reflects the specific contribution of different parts to the total system resistance. The total system resistance is obtained by summing all pressure drop values ​​along the actual flow path. Furthermore, the pressure drop values ​​in this list are used to calculate the pressure drop percentage sequence, thereby obtaining the resistance distribution entropy. The acquisition process can be formally described as follows: First, under the constraint of the required ventilation volume, establish a ventilation topology based on the channel connectivity of the candidate component combinations, distinguish between series channel segments and parallel branches, and determine the flow rate and average cross-sectional velocity of each segment; then, calculate the pressure drop along the flow path for each straight channel segment using the Darcy-Weisbach formula, and calculate the local pressure drop for each local component such as louvers, baffles, elbows, and sound absorbers based on the local resistance coefficient; finally, record the pressure drop values ​​corresponding to each segment and component in a table according to the actual flow path to form a pressure drop contribution list. This list is essentially a detailed pressure drop table generated from the preceding resistance calculation; subsequent steps simply reuse this detailed table, rather than regenerating another set of original data.

[0067] For example, the engineer extracts all pressure drop values ​​from the aforementioned table, divides each value by the total system resistance to obtain the percentage, lists it in a new table, calculates the base-2 logarithm of each percentage, multiplies it by the negative of the percentage itself, and then sums them up. Finally, the resistance distribution entropy value is recorded at the bottom of the table.

[0068] S204: Calculate the peak surface velocity.

[0069] Beyond resistance distribution, uneven flow distribution in local cross-sections can create high-velocity zones, requiring network analysis to identify extreme flow conditions.

[0070] The peak face velocity calculation process employs a ventilation network flow distribution algorithm. First, the candidate component combination is equivalent to a network diagram composed of branches and nodes. The resistance characteristics of each branch are established based on the aforementioned pressure drop calculation to establish a quadratic resistance relationship. The equivalent resistance coefficient of each branch is determined. Under the constraint of the required total ventilation volume, the Hardy Cross iterative method or matrix solution method is used to balance the flow conservation of nodes and the pressure drop closure. Initially, the flow of each branch is assumed, and iterative adjustments are made until the inflow and outflow of all nodes are equal and the sum of the loop pressure drops is 0, thus obtaining the actual flow of each branch. Subsequently, the face velocity is calculated at a key section. The flow through this section is divided by the effective flow area, and finally, the maximum value of all face velocities is taken as the peak face velocity, in meters per second.

[0071] S205: Calculate the risk coefficient of reverse increase and determine availability.

[0072] The drag distribution entropy and peak surface velocity characterize the structural and flow field features, respectively, and must be comprehensively transformed into a single risk index to achieve automated availability determination of candidate combinations.

[0073] The calculation and determination process of the reverse increase risk coefficient adopts a probabilistic regression model with monotonic constraints. First, an input vector containing the drag distribution entropy and peak surface velocity is constructed and fed into a pre-trained model. A gradient boosting tree or neural network with monotonically increasing or decreasing constraints is used to force the partial derivative of the peak surface velocity to be non-negative and the partial derivative of the drag distribution entropy to be non-positive. The model is trained with historical or simulated samples. The samples contain the input vector and a binary label indicating whether there is a reverse increase in wind turbine noise or a perceptible increase in local aerodynamic noise. After the model outputs the probability, it is calibrated by the Sigmoid function (Platt scaling) to obtain a dimensionless reverse increase risk coefficient, ranging from 0 to 1. Then, it is compared with a preset risk threshold, which is determined by engineering experience or offline calibration. If the reverse increase risk coefficient is less than or equal to the preset risk threshold, the candidate component combination is determined to be usable; otherwise, it is discarded. The usable combination, along with the total system drag and operating condition point, is used for subsequent steps.

[0074] For example, engineers input the drag distribution entropy and peak surface velocity into the interface of a pre-trained model software. After clicking "calculate," the resulting inverse risk coefficient is displayed in the results column. This coefficient is then compared with a pre-set risk threshold value. If the coefficient is low, it is marked as usable in the candidate combination list, and the relevant parameters are saved.

[0075] In one embodiment, the probabilistic regression model with monotonic constraints is constructed using a gradient boosting decision tree framework to ensure that the model output monotonically increases with respect to peak surface velocity and monotonically decreases with respect to drag distribution entropy, while providing a calibrated probability output as a noise-inverse probability value.

[0076] The data collection process first generates a large number of candidate component combination samples through computational fluid dynamics simulation software, or extracts measured operating condition data from historical substation engineering projects. Each sample contains an input feature vector (resistance distribution entropy and peak surface velocity) and a binary label (a label value of 1 indicates that the fan noise has increased or the local aerodynamic noise has been perceptibly increased, and a label value of 0 indicates that it has not occurred). The total number of samples is usually between several hundred and several thousand to cover the range of variations in different loads, temperatures and component combinations.

[0077] The model construction process uses the LightGBM library to implement a gradient boosting decision tree. A binary classification target is selected as the base model, and log loss is used as the loss function. The number of trees is set to be between 100 and 500, the maximum depth is limited to 4 to 8 layers, the number of leaf nodes in each tree is limited to 16 to 64, and the learning rate is set to be between 0.01 and 0.1. To impose monotonic constraints, the model parameters explicitly specify monotonone_constraints as [-1,1], where the first element corresponds to the input feature of resistance distribution entropy, which is forced to decrease monotonically, and the second element corresponds to the input feature of peak surface velocity, which is forced to increase monotonically, to ensure that the physical meaning of the model is consistent.

[0078] The training process is divided into two stages: initial training and probability calibration. The initial training uses five-fold cross-validation to divide the dataset. The training fold is fitted with the above parameters, and the validation fold is evaluated for log loss and AUC. After training, the Platt scaling method is used to fit the sigmoid function on the independent output set to map the original model scores to calibration probabilities. Alternatively, the isotonic regression method is used to directly perform nonparametric calibration on the probability output to ensure that the probability value of noise inversion is reliably distributed in the range of 0 to 1.

[0079] The optimization process involves adjusting hyperparameters, including the number of trees, learning rate, maximum depth, and number of leaf nodes, through grid search or Bayesian optimization. The optimization objective is to maximize the average AUC of cross-validation, while monitoring compliance with monotonic constraints and the risk of overfitting. Finally, a set of parameters is selected to solidify the model. Engineers run the training script on a computer, observe the loss decrease curve and constraint satisfaction indicators in the iteration log, and save the model file and calibration parameters after training. In subsequent design processes, the noise inversion probability value prediction used for candidate component combinations is directly loaded.

[0080] Step S2 completes the calculation of the total system resistance of the candidate component combination and the determination of the fan operating condition point under the constraint of the required ventilation volume. The uniformity of resistance distribution and local extreme flow state are quantified by resistance distribution entropy and peak surface velocity. The probability value of noise reversal is output by the probabilistic regression model with monotonic constraints to realize availability determination and form a list of available combinations without the risk of noise reversal.

[0081] Step S2 has identified available candidate component combinations without the risk of increased noise and provided the corresponding required ventilation volume, total system resistance, and fan operating conditions. However, the number of available combinations is large, and direct enumeration makes it difficult to minimize boundary noise under multiple operating conditions. To efficiently search for the optimal synergy between ventilation and noise reduction, a deep reinforcement learning decision-making process must be introduced. Using available combinations as actions, optimization is driven by boundary noise prediction to output a ventilation and noise reduction layout scheme.

[0082] S301: Constructing a deep reinforcement learning environment.

[0083] The available candidate component combinations need to be optimized for noise reduction under fixed ventilation constraints. First, the Markov decision process environment is defined to ensure that the agent learns a physically consistent decision strategy.

[0084] The deep reinforcement learning environment construction process adopts the standard Markov decision process framework. State space. Defined as a multidimensional vector, including the required ventilation volume Current total system resistance Fan operating point (operating air volume) Total pressure Rotation speed ), current component combination code (one-hot or embedded vector representation with available combination index), and current operating condition parameters (load rate). Ambient temperature Action space Defined as a discrete set of actions, directly corresponding to the available candidate component combination indexes output in step S2 (limited in number, typically tens to hundreds). After an action is executed, the component combinations, total system resistance, and fan operating point in the state are updated immediately. The transfer function deterministically maps actions to the new state (reusing the S2 resistance calculation and fan curve intersection logic). This environment construction ensures that the state fully represents the current performance of the ventilation system, facilitating agent learning of the impact of component selection on noise.

[0085] S302: Designing the reward function and predicting boundary noise.

[0086] After defining the state and action, a reward function must be designed to guide the agent to converge toward low boundary noise, while strictly penalizing any behavior that violates ventilation constraints.

[0087] Boundary noise prediction employs a simplified acoustic propagation model. First, the insertion loss of each noise-absorbing component is calculated for the component combination under the current state. (Determined by looking up tables or empirical formulas based on component type, frequency band, and face wind speed, such as louver insertion loss) , Surface velocity refers to the average velocity (unit: m / s) of air flowing through key cross-sections such as louvers, used to calculate the dependence of louver insertion loss on airflow velocity; baffles , Indicates the number of bends in the baffle, referring to the number of times the airflow bends in the baffle structure, used to calculate the dependence of the baffle insertion loss on the geometry; sound absorption channel. in (where is the sound absorption coefficient); then the total insertion loss is calculated. (Contribution of superimposed door and wall structure). Indicates door insertion loss, which refers to the noise reduction effect of a ventilated, soundproof door (unit: dB). Indicates wall insertion loss, referring to the noise attenuation of the sound-absorbing wall (unit: dB); fan noise source intensity. Based on operating condition points, look up the fan sample spectrum or formula. , The reference sound power level of the wind turbine at the reference speed and reference power is obtained directly from the noise sample data or standard test report provided by the wind turbine manufacturer. This reference speed value is consistent with the reference point in the fan performance curve; The reference shaft power value at the corresponding base speed is also derived from wind turbine sample data; the main transformer's noise source strength For a fixed spectrum; final boundary noise prediction value (A-weighted sound pressure level, unit dB(A)) The calculation formula is as follows:

[0088] ;

[0089] in This refers to the additional propagation loss of sound waves as they propagate along the air intake and exhaust ducts due to distance attenuation, wall reflection, and absorption by sound-absorbing materials. It is determined by looking up tables or empirical formulas using the duct length, cross-sectional area, wall absorption coefficient, and frequency band. For example, in a straight duct... Estimation is performed (L is the channel length, D is the equivalent diameter, α is the sound absorption coefficient, and f is the frequency). This is the environmental background noise correction value, used to deduct the interference of non-target noise in the environment measured or calculated on-site. It is directly obtained from the actual background noise test data of the substation site or the typical value of the corresponding acoustic environment functional zone in the national standard (GB3096-2008). For example, the nighttime background noise correction value for Class II urban areas is 30 dB(A).

[0090] reward function Designed for

[0091] ;

[0092] in This is the penalty coefficient (ranging from 10 to 100). This is the maximum permissible total pressure for the fan. The bonus is primarily based on negative boundary noise, while a secondary penalty is applied for excessive drag or insufficient ventilation margin, ensuring that optimization always adheres to thermal constraints.

[0093] S303: Execute deep reinforcement learning decision-making processes.

[0094] Once the environment and rewards are determined, agent training or inference is performed to search for the optimal combination of components.

[0095] The deep reinforcement learning decision-making process employs the Proximal Policy Optimization (PPO) algorithm (Actor-Critic framework). First, the Actor network (multilayer perceptron, input state dimension, output action probability distribution) and the Critic network (output state value function) are initialized. For new projects, a pre-trained model (trained offline based on historical substation samples) is used, or a random policy is started. Within multiple episodes (each episode is a complete component selection sequence, typically 1 to 10 steps long to support combination adjustments), the agent interacts with the environment to collect trajectories, calculates the advantage function, and then performs PPO pruning objective optimization.

[0096] ;

[0097] in The probability ratio is used to characterize the ratio of the probability of the new and old strategies on the same action. For advantage estimation, it is used to measure the gain of the action relative to the average level. The pruning scope is used to limit the policy update magnitude; simultaneously, value loss and entropy regularization are optimized to improve exploration. After training convergence (or direct greedy selection in inference mode), the ventilation and noise reduction layout scheme corresponding to the action sequence with the highest cumulative reward is output, including the optimal component combination, layout position, corresponding total system resistance, and fan operating point. This process efficiently traverses the available combination space, minimizing boundary noise.

[0098] For example, an engineer loads a PPO agent into a deep learning framework, inputs the required ventilation volume and a list of available combinations for the current project, runs hundreds of episodes, observes the convergence of the reward curve, and selects the component combination corresponding to the episode with the highest reward as the final layout scheme.

[0099] Step S3, based on the available candidate component combinations, constructs a state space and action space containing the required ventilation volume and fan operating conditions. Designs a reward function that is based on the negative boundary noise prediction value and penalizes constraint violations. The decision-making process is executed through the ProximalPolicyOptimization algorithm, and the ventilation and noise reduction layout scheme with the minimum boundary noise is output.

[0100] Step S3 has output the ventilation and noise reduction scheme with the minimum boundary noise. However, the simplified model may ignore three-dimensional turbulence and complex sound propagation details. To ensure that the scheme meets heat dissipation constraints and achieves reliable noise reduction, a joint computational fluid dynamics and acoustic verification must be performed, and construction parameters and operation control parameters should be generated based on the verification results.

[0101] S401: Construct a computational fluid dynamics model.

[0102] Once the ventilation and noise reduction layout is determined, a high-fidelity three-dimensional computational domain must first be established to reproduce the actual geometry and flow environment, ensuring the reliability of subsequent simulation results.

[0103] The computational fluid dynamics model was constructed using a finite volume method discretization framework. First, based on the component combinations and layout of the ventilation and noise reduction scheme, the complete three-dimensional geometry of the main transformer room, air intake and exhaust ducts, louvers, baffles, sound absorbers, fans, and doors and windows was reconstructed in CAD software. The computational domain was extended to the outdoor boundary (at least 10 times the feature length from the opening to avoid the influence of outlet backflow). A hybrid meshing strategy was adopted (hexahedral structured mesh in the core flow region, tetrahedral unstructured mesh with prism layers near the walls and complex components), with the total number of meshes controlled to be in the millions to tens of millions. Boundary conditions were set, including: the inlet was a velocity inlet (initial estimate based on the required ventilation volume). The model employs a pressure outlet, a non-slip wall with an applied wall function, and a dynamic reference frame or sliding mesh to simulate impeller rotation in the fan region. This model construction ensures that the geometry and boundaries faithfully adhere to the layout scheme, providing a foundation for accurate flow field solutions.

[0104] S402: Perform flow field simulation and verify ventilation performance.

[0105] After the geometric model is established, the three-dimensional incompressible turbulent field needs to be solved to verify whether the actual ventilation volume, total system resistance and peak surface velocity meet the constraints.

[0106] The flow field simulation process employs a Reynolds-averaged Navier-Stokes (RANS) equation closure algorithm, using a Realizable k-ε turbulence model (transport equations are turbulent kinetic energy). With dissipation rate The governing equations are:

[0107] Continuity equation

[0108] ;

[0109] Momentum equation

[0110] ;

[0111] in For turbulent viscosity, These are the model coefficients.

[0112] The solution employs the SIMPLE algorithm with pressure-velocity coupling and second-order upwind discretization. The residuals of the convection term converge to 10^{-5}, and those of the turbulence term converge to 10^{-4}.

[0113] The verification process is as follows: the actual ventilation volume is obtained by integrating the outlet surface flow rate. With the required ventilation volume Comparison (requirements) The total resistance of the simulated system is obtained by calculating the total pressure difference between the inlet and outlet. Verification was performed by matching the operating conditions of the wind turbine; the maximum face wind speed at key sections was extracted. If the ventilation performance does not meet the requirements, mark the deviation for subsequent adjustments.

[0114] For example, engineers import the layout plan CAD file, divide the grid, set the fan speed boundary, run the solver to observe the convergence of the residual curve, export the flow integral report and velocity cloud map, and confirm that the actual ventilation volume reaches the required ventilation volume level.

[0115] S403: Perform acoustic simulation and check boundary noise.

[0116] After the flow field converges, noise generation and propagation need to be calculated based on transient or average flow field data to verify the effectiveness of boundary noise control.

[0117] The acoustic simulation process employs a hybrid CFD / CAE approach. First, the steady-state flow field is derived as the average background. If broadband noise needs to be captured, the simulation is switched to unsteady DES or LES to acquire transient pressure fluctuations. Noise source extraction uses a broadband noise source model (Proudman and Lilley acoustic power formulas) or directly uses the FW-H acoustic analogy integral.

[0118] Far-field sound pressure

[0119] ;

[0120] The FW-H acoustic analogy integral formula is used to calculate far-field sound pressure disturbances. ,in For the location of the observation point, For observation time; For the velocity of sound in the ambient medium; integral surface A control surface (such as a solid wall or virtual surface) is used to surround the noise source. The component of the Lighthill stress tensor in the radial direction from the source point to the observation point represents the fluid momentum and pressure disturbance. for The material derivative (the rate of change of the source point's motion over time). The distance from the source point to the observation point; This represents the radial component of the Mach number vector at the source point; The dot product of the surface normal vector and the Mach number vector; all source terms are evaluated under retarded time to account for the acoustic propagation delay; the first term in the formula corresponds to the unipolar and dipole contributions of the thickness and the loaded noise, and the second term is the motion effect correction.

[0121] The boundary noise verification process is as follows: Calculate the A-weighted sound pressure level at a designated outdoor receiving point (1 meter above the opening, meeting environmental noise standards). , and the predicted value Compare with the standard limits; simultaneously check the characteristic frequency of the fan and the peak value of aerodynamic noise in the spectrum. If the boundary noise exceeds the standard or the deviation from the prediction exceeds the preset tolerance, record the deviation.

[0122] S404: Generate construction parameters and operation control parameters based on the verification results.

[0123] After the flow field and acoustic calibration are completed, the verified scheme needs to be converted into specific parameters that can be implemented. If there are any deviations, they should be fine-tuned and recalibrated.

[0124] The parameter generation process employs an iterative adjustment logic. If all verification indicators are met (actual ventilation volume is not lower than the required ventilation volume, boundary noise is not higher than the predicted value and specifications), then the parameters are directly extracted: construction parameters include the precise dimensions of each sound-absorbing component (louver angle, baffle spacing, sound-absorbing material thickness), channel geometric coordinates, and fan installation location; operation control parameters include the fan variable frequency operation curve (speed or frequency corresponding to different load rates). Ambient temperature The system includes temperature control logic (increasing the rotation speed when the indoor temperature rises close to the upper limit of the allowable temperature rise). If there is a deviation, the component parameters are adjusted based on sensitivity analysis (such as reducing the local resistance coefficient or optimizing the layout), and S402 and S403 are re-executed until they pass. Finally, a complete parameter table and verification report are output.

[0125] For example, after engineers check the sound pressure level report at the acoustic receiving point and confirm that it is lower than the standard, they measure the louver blade spacing and sound-absorbing layer thickness from the model, draw up detailed construction drawings, and compile a control table for the fan frequency as temperature changes.

[0126] Step S4 is based on the ventilation and noise reduction layout scheme output in step S3. The actual flow field is solved and the ventilation performance is checked by a high-fidelity computational fluid dynamics model. Then, based on the flow field results, acoustic simulation is performed to check the boundary noise. Finally, based on the check results, accurate construction parameters and operation control parameters are generated to ensure that the entire design achieves a reliable and traceable noise reduction effect while meeting the heat dissipation constraints.

[0127] Specifically, the above are merely preferred embodiments of this application and are not intended to limit this application.

[0128] Other preset parameters whose value range and acquisition logic are not explicitly specified in this invention can be pre-calibrated through offline simulation testing, or set to fixed values ​​according to on-site operating procedures.

[0129] In the description of this specification, references to terms such as "an embodiment," "an example," and "a specific example" indicate that a particular feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0130] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A substation ventilation and noise reduction design method based on deep reinforcement learning, characterized in that, Including the following steps: S1: Obtain the heat output and allowable temperature rise limit of the equipment, and calculate the required ventilation volume based on the existing heat balance ventilation volume calculation method; S2: Based on the required ventilation volume, the ventilation resistance of the candidate component combination is calculated to obtain the total system resistance. The fan operating point is determined by combining the fan performance curve. The resistance distribution entropy and peak face velocity are extracted. The resistance distribution entropy is obtained by calculating the pressure drop ratio sequence according to the ratio of the pressure drop of each channel section or component to the total system resistance. The pressure drop ratio sequence is then calculated using Shannon information entropy. The peak face velocity is calculated according to the ventilation network flow distribution results for each key section, and the maximum value is taken. The resistance distribution entropy and peak face velocity are input into a probabilistic regression model with monotonic constraints. The output noise inverse probability value is then compared with a preset risk threshold to determine the availability of the candidate component combination. S3: The combination of candidate components deemed usable, along with the required ventilation volume, total system resistance, and fan operating point, is input into a deep reinforcement learning decision-making process, and a ventilation and noise reduction layout scheme is output based on the boundary noise prediction results. S4: Perform computational fluid dynamics and acoustic joint verification of the ventilation and noise reduction layout scheme, and generate construction parameters and operation control parameters based on the verification results.

2. The substation ventilation and noise reduction design method based on deep reinforcement learning according to claim 1, characterized in that, Step S1 includes: The total heat generation of the equipment is obtained by adding the heat generation of the main transformer to the rated heating power of the low-voltage switchgear and auxiliary equipment; the upper limit of the allowable temperature rise is determined by looking up the maximum allowable temperature rise value according to the insulation thermal class of the main transformer, and then comparing the maximum allowable temperature rise value with the corresponding indoor allowable temperature rise control value of the industry standard, and taking the smaller value.

3. The substation ventilation and noise reduction design method based on deep reinforcement learning according to claim 2, characterized in that, Step S1 also includes: The required ventilation volume is calculated by dividing the total heat output of the equipment by the product of the specific heat capacity of air at constant pressure, the density of the incoming air, the upper limit of the allowable temperature rise, and the heat exchange efficiency coefficient. The required ventilation volume is set as a hard constraint condition for the calculation of the combined resistance of candidate components and the determination of the fan operating conditions.

4. The substation ventilation and noise reduction design method based on deep reinforcement learning according to claim 3, characterized in that, Step S2 includes: The total system resistance is calculated by summing the friction resistance of each straight channel and the local resistance of each louver, baffle, elbow and sound absorber along the airflow path; the operating point of the fan is determined by the intersection of the fan performance curve and the system characteristic curve formed by the relationship between the total system resistance and the flow rate.

5. The substation ventilation and noise reduction design method based on deep reinforcement learning according to claim 4, characterized in that, Step S2 also includes: The noise inversion probability value is obtained by inputting the drag distribution entropy and the peak surface velocity into a monotonic constraint probabilistic regression model. The model imposes a risk non-decreasing constraint on the peak surface velocity and a risk non-increasing constraint on the drag distribution entropy. The noise inversion probability value is compared with a preset risk threshold to determine the availability of candidate component combinations.

6. The substation ventilation noise reduction design method based on deep reinforcement learning according to claim 5, characterized in that, Step S3 includes: The deep reinforcement learning environment adopts the Markov decision process framework. The state space includes the required ventilation volume and total system resistance, the fan operating point, the current component combination code, and operating parameters. The action space corresponds to the available candidate component combination index. After the action is executed, the total system resistance and the fan operating point are deterministically updated.

7. The substation ventilation noise reduction design method based on deep reinforcement learning according to claim 6, characterized in that, Step S3 also includes: The boundary noise prediction value is obtained by accumulating the insertion loss of the silencing components and superimposing the logarithmic sum of the noise source strength of the fan and the main transformer. The reward function is mainly based on the negative boundary noise prediction value and secondarily penalizes the excessive resistance and insufficient ventilation margin. The PPO algorithm drives the agent interaction optimization and outputs the ventilation and silencing arrangement scheme with the optimal component combination.

8. The substation ventilation noise reduction design method based on deep reinforcement learning according to claim 7, characterized in that, Step S4 includes: The computational fluid dynamics model uses the finite volume method and hybrid mesh generation. Boundary conditions include velocity inlet, pressure outlet, and wind-driven reference frame. The flow field simulation uses the Realizable k-ε turbulence model and SIMPLE algorithm. The actual ventilation volume is checked to ensure that it is not lower than the required ventilation volume, and the simulation system's total resistance and the fan's operating point are matched.

9. The substation ventilation noise reduction design method based on deep reinforcement learning according to claim 8, characterized in that, Step S4 also includes: The acoustic simulation uses a hybrid CFD / CAE method and FW-H acoustic analog integral to calculate boundary noise. The A-weighted sound pressure level is checked to ensure it meets the predicted value and specifications. Based on the check results, the precise dimensional construction parameters of the noise reduction components and the control parameters of the variable frequency operation curve of the fan are generated.

Images