A method and system for self-adaptive regulation of air volume and frequency of a fan

Abstract

The application provides a kind of air volume and frequency adaptive regulation method and system of induced draft fan;The system includes multi-source working condition parameter acquisition module, data preprocessing module, core control module, safety constraint and training convergence module;The method includes the following sub-steps: S1: real-time acquisition of multi-source working condition parameters;S2: the pre-processing of the collected multi-source working condition parameters;S3: realize the dynamic adaptive matching of air volume and frequency of induced draft fan;S4: real-time monitoring of air volume deviation rate and adjusting strategy network strategy parameters, air volume closed loop calibration;S5: realize working condition adaptive switching;S6: safety constraint;S7: system output and execution;The application realizes air volume calibration response and working condition switching regulation by high-frequency acquisition of multi-source working condition parameters, can quickly capture strong dynamic working condition changes such as blast furnace variable load and converter blowing, and ensure that flue gas is promptly extracted, negative pressure is stable, and production safety and environmental protection emission hidden danger caused by regulation lag are completely solved.

Description

Technical Field

[0001] This invention relates to the field of induced draft fan control technology, and in particular to an adaptive control method and system for induced draft fan air volume and frequency. Background Technology

[0002] As an indispensable core auxiliary equipment in dust removal systems of blast furnaces, converters, and iron and steel metallurgy, the induced draft fan's main function is to discharge smelting flue gas and maintain stable negative pressure at the furnace top / flue. Its air volume output is directly related to key aspects such as the stability of metallurgical operations, flue gas recovery and purification effects, production safety, and environmental emissions. With the transformation of steel production towards high efficiency, energy saving, and intelligence, industrial scenarios have placed higher demands on the induced draft fan's operational stability, energy consumption control level, rapid response capability, and adaptive adjustment capability.

[0003] In steel metallurgical production scenarios such as converters, blast furnaces, and ironmaking / steelmaking dust removal systems, operating parameters such as metallurgical intensity, furnace mouth flue gas generation, flue negative pressure, and dust removal system resistance are constantly undergoing drastic dynamic changes (e.g., blast furnace air volume fluctuations, converter blowing cycles, charging and tapping, furnace pressure fluctuations, and changes in dust collector filter bag resistance). This characteristic requires the induced draft fan / dust collector fan to have its air volume output capable of quickly and in real-time adapting to changes in operating conditions, and the operating frequency of the induced draft fan directly determines its air volume output capacity (frequency and air volume are approximately positively correlated). (Relationships); for example, when the blast furnace receives air and the converter enters the tapping stage, the flue gas volume in the flue increases dramatically in an instant. It is necessary to simultaneously increase the frequency of the induced draft fan to increase the air volume, ensure the timely extraction of high-temperature flue gas, maintain the stability of the negative pressure in the flue, and ensure production safety and environmental emissions. When the blast furnace reduces air intake, the converter finishes blowing, or the furnace is shut down to wait for materials, the flue gas volume decreases significantly. It is necessary to promptly reduce the frequency of the induced draft fan to reduce the air volume, avoid long-term high-frequency idling of the fan causing a large waste of electrical energy, and at the same time prevent excessive negative pressure from adversely affecting the furnace process and equipment structure.

[0004] However, existing methods for controlling the air volume and frequency of induced draft fans mainly include manual adjustment, fixed frequency setting, simple PID control, and some adaptive control technologies. These methods are insufficient to meet the safe and efficient operation requirements of blast furnaces and converters under complex operating conditions characterized by strong fluctuations, periodicity, and strong disturbances. The main drawbacks are as follows:

[0005] 1. Problem of lagging dynamic response to operating conditions: Existing control methods (manual, fixed frequency, simple PID) cannot capture strong dynamic changes in operating conditions such as continuous load changes in blast furnace, periodic blowing in converter, sudden changes in flue gas volume, and fluctuations in negative pressure at the furnace top / flue. The adjustment lag time is long, which can easily lead to mismatch between air volume and smelting conditions, loss of negative pressure control, and flue gas overflow, affecting production safety and environmental emissions.

[0006] 2. Low accuracy of air volume and frequency matching: The existing technology has not established an accurate matching model of smelting intensity, flue gas volume, air volume and frequency. Overshoot and oscillation are prone to occur during the adjustment process, and the air volume fluctuates greatly. It is impossible to achieve the optimal matching of air volume and frequency under complex working conditions such as blast furnace load change, converter blowing / stop blowing.

[0007] 3. High energy consumption: Fixed frequency operation and unreasonable traditional adjustment methods cause the induced draft fan to maintain high frequency operation under low load and low flue gas volume conditions, resulting in serious waste of electrical energy. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies by proposing an adaptive control method and system for the air volume and frequency of an induced draft fan.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] An adaptive control system for the air volume and frequency of an induced draft fan, the system comprising a multi-source operating condition parameter acquisition module, a data preprocessing module, a core control module, and a safety constraint and training convergence module;

[0011] The multi-source operating condition parameter acquisition module is used to acquire multi-source operating condition parameters in real time.

[0012] The data preprocessing module is used to preprocess the collected multi-source operating condition parameters;

[0013] The core control module achieves dynamic adaptive matching of induced draft fan air volume and frequency based on the PPO algorithm, including constructing a strategy network and a value network, updating strategy parameters, air volume closed-loop calibration, and adaptive switching of operating conditions.

[0014] The security constraint and training convergence module sets the training convergence criteria for the PPO algorithm model and imposes security constraints through relevant security rules.

[0015] This invention also proposes a method for adaptive control of induced draft fan air volume and frequency, comprising the following sub-steps:

[0016] S1: Real-time acquisition of multi-source operating parameters;

[0017] Multiple types of high-precision industrial sensors are deployed in areas such as the induced draft fan, flue, furnace / furnace top of the blast furnace / converter dust removal system. The multi-source operating condition parameter acquisition module collects multi-source operating condition parameters in real time through high-precision industrial sensors at a preset acquisition frequency (such as 10Hz).

[0018] The high-precision industrial sensors include frequency sensors, pressure sensors, air volume sensors, etc.

[0019] The multi-source operating parameters include fan operating parameters, process operating parameters, environmental compensation parameters, and auxiliary parameters;

[0020] The fan operating parameters include the fan's current operating frequency, fan outlet pressure, fan actual air volume, fan motor power, fan current, and fan vibration value.

[0021] The process operating parameters include furnace / furnace top negative pressure, blast furnace smelting intensity / converter blowing signal, flue gas resistance / pressure difference, flue gas temperature, etc.

[0022] The environmental compensation parameters include ambient temperature, etc.

[0023] The auxiliary parameters include the opening status of the flue valve and auxiliary parameters that characterize the resistance to flue gas flow.

[0024] S2: Preprocess the collected multi-source operating condition parameters;

[0025] The data preprocessing module preprocesses the collected multi-source operating parameters, including standardization, noise filtering, and outlier removal.

[0026] Specifically, the standardization process involves the data preprocessing module collecting historical data of various multi-source operating parameters in advance and calculating the historical maximum and minimum values ​​of each multi-source operating parameter.

[0027] Record the historical maximum value as The historical minimum value is The parameters of each multi-source operating condition are standardized using the following normalization formula:

[0028] ;

[0029] Where x represents the collected multi-source operating condition parameters, These are the normalized multi-source operating condition parameters;

[0030] Noise filtering: A sliding window mean filtering algorithm is used to set an independent sliding window for each multi-source operating condition parameter, and the average value of the normalized multi-source operating condition parameter within the sliding window is calculated. This average value is the multi-source operating condition parameter value after filtering to eliminate noise.

[0031] Outlier removal: The data preprocessing module presets safety thresholds for each multi-source operating condition parameter, iterates through each multi-source operating condition parameter, identifies multi-source operating condition parameters that exceed the corresponding safety threshold as outliers and removes them, and supplements missing data through linear interpolation to obtain the preprocessed multi-source operating condition parameters.

[0032] S3: Achieve dynamic adaptive matching between induced draft fan air volume and frequency;

[0033] The core control module constructs a policy network and a value network through the PPO algorithm, and calculates the frequency adjustment of the wind turbine inverter based on the preprocessed multi-source operating parameters.

[0034] Includes the following sub-steps:

[0035] S31: Constructing strategy networks and value networks;

[0036] The core control module constructs a policy network and a value network using the PPO algorithm. Both the policy network and the value network adopt a 3-layer fully connected neural network structure.

[0037] The three-layer fully connected neural network structure includes an input layer, a hidden layer, and an output layer; the input layer contains 6 neurons; the hidden layer includes two layers, denoted as hidden layer 1 and hidden layer 2, each with 32 neurons, using the ReLU activation function; the output layer contains 1 neuron, which outputs the frequency adjustment amount of the wind turbine inverter;

[0038] The formula for the policy network is as follows:

[0039] ;

[0040] in, , , , For strategy parameters; This is the frequency adjustment amount; , , The weight matrix is ​​randomly initialized, where It is 6 rows and 32 columns. 32 rows and 32 columns It consists of 32 rows and 1 column; For activation function, For activation functions; The initial bias vector, 32 rows and 1 column 32 rows and 1 column It is arranged in 1 row and 1 column (consistent with the number of neurons in the corresponding layer);

[0041] The value network adopts the exact same network structure as the policy network;

[0042] S32: Optimize the policy parameters of the policy network and output the optimal frequency adjustment amount;

[0043] Includes the following sub-steps:

[0044] S321: Initialize the core parameters of the PPO algorithm;

[0045] Load the policy network and value network, and initialize the core parameters of the PPO algorithm;

[0046] The core parameters of the PPO algorithm include the discount factor γ and the learning rate. , trimming factor ε, experience collection round length, strategy update frequency, and preset range of frequency adjustment amount;

[0047] S322: Generates initial frequency adjustment amount;

[0048] Set the duration of each collection cycle and the total number of collection cycles included in each experience collection round; obtain the corresponding multi-source working condition data in units of collection cycle duration; each collection cycle corresponds to a collection cycle number, let the collection cycle number be t, and the value range of t is [1, total number of collection cycles];

[0049] The current operating frequency of the fan, the outlet pressure of the fan, the actual air volume of the fan, the ambient temperature, the dynamic target air volume of the fan, and the air volume deviation corresponding to each collection cycle are respectively formed into independent input vectors S, which are input into the strategy network. The strategy network calculates and outputs the preliminary frequency adjustment amount through the formula, and the preliminary frequency adjustment amount is limited to a preset range.

[0050] S323: Calculate cumulative discount rewards;

[0051] The instantaneous reward for each collection period is calculated using the following multi-objective weighted reward function;

[0052] ;

[0053] in, =1- / ; =1- / a; =1- / ;

[0054] in, The instant reward for the t-th gathering cycle in a single experience gathering round. , , These are the weighting coefficients for each reward item. + + =1; For air volume accuracy reward, For frequency smoothing rewards, As an energy consumption reward;

[0055] Record the actual air volume of the fan in the current data collection period as: The preset dynamic target air volume for the fan is Air volume deviation value The dynamic target air volume is Updated in real time based on smelting conditions; ΔQ=| - |;

[0056] , where a is the initial frequency adjustment amount for the t-th collection cycle in a single experience collection round, and a is the set frequency adjustment threshold. For the fan motor power in the t-th collection cycle of a single experience collection round, This is the preset rated power of the fan motor;

[0057] Calculate the cumulative discount reward using the following formula:

[0058] ;

[0059] Where T is the total number of collection cycles within a single experience collection round, and k is the summation index variable. The cumulative discount reward value for the t-th collection cycle in a single experience collection round;

[0060] S324: Calculate the advantage function and the policy optimization objective function;

[0061] Let the input vector corresponding to the t-th collection cycle in a single experience collection round be denoted as... ,Will The input is fed into the constructed value network, and the output of the value network is the value network's response to the input vector. The estimated value of is denoted as . , The network parameters are preset for the value network;

[0062] Advantage function ;

[0063] Optimize the objective function using a maximization strategy To achieve the goal, the strategy parameters are updated and iterated:

[0064] Strategy optimization objective function ;

[0065] in, For strategy parameters, The clipping constraint function is... The interval range is [ , ];

[0066] r represents the probability ratio between the old and new strategies. t (θ)= ;in, For the old strategy probability, The probability of the new strategy; The frequency adjustment amount for the t-th collection cycle in a single experience collection round. ; Adjust the frequency amount The input is fed into the Sigmoid activation function, and the output is the old policy probability. ;Optimize the objective function using a maximization strategy The specific value is the target, and the output is the optimized policy parameters, which yields the optimized policy network. The input vector is then used as the target. The input is fed into the optimized policy network, which outputs a frequency adjustment value. This frequency adjustment value is then input into the Sigmoid activation function to output the new policy probability. ; The empirical sampling distribution;

[0067] S325: Perform iterative updates to the strategy parameters;

[0068] Based on the advantage function calculated in step S324 Strategy optimization objective function The policy parameters of the policy network are updated using the stochastic gradient ascent method.

[0069] During each update, the probability ratio between the old and new strategies is limited by the clipping constraint function. Within the preset range;

[0070] Set the policy update to be performed every n experience collection rounds, where n is a positive integer; repeat steps S322-S325 until the training convergence criterion is met.

[0071] Specifically, the training convergence criterion is as follows:

[0072] The training convergence criterion is determined to be met when all three of the following conditions are met simultaneously.

[0073] (1) The average reward value fluctuation over n consecutive experience collection rounds is less than or equal to a set percentage (e.g., 5%).

[0074] The formula for calculating the average reward value fluctuation is as follows:

[0075] ;

[0076] in, The average reward value over n consecutive experience collection rounds. The cumulative discount reward value for the i-th experience collection round;

[0077] (2) The air volume control accuracy is within the set range (e.g., [-2%, 2%]) for n consecutive experience collection rounds.

[0078] The air volume control accuracy is defined as follows: The formula for calculating airflow control accuracy is as follows:

[0079] *100%;

[0080] (3) The accuracy rate of working condition identification is greater than or equal to the set percentage (e.g., 99.5%).

[0081] The policy network extracts features from the preprocessed multi-source working condition parameters and identifies the working conditions corresponding to the experience collection rounds; the actual results of each working condition are known.

[0082] The working conditions identified by the policy network are compared with the actual results of the working conditions. If they match, the identification is considered correct; otherwise, it is considered incorrect.

[0083] The accuracy of condition identification is calculated using the following formula:

[0084] ;

[0085] in, To improve the accuracy of working condition identification, The number of correctly identified working condition samples, The total number of operating condition samples identified;

[0086] Once all the above conditions are met, the training is considered to have converged, offline pre-training is stopped, the policy parameters of the current policy network are solidified, and the solidified policy network parameters are saved to the policy library.

[0087] Simultaneously, for both low-load and high-load operating conditions, specialized training is conducted using historical operating data (multi-source operating parameters, energy consumption data, etc.) corresponding to the operating conditions to form the optimal control strategy for the corresponding operating conditions, which is stored in the strategy library; the optimal control strategy includes the optimal operating frequency of the wind turbine inverter, the target air volume, etc.

[0088] S4: Real-time monitoring of airflow deviation rate and adjustment of strategy parameters of strategy network to perform airflow closed-loop calibration;

[0089] Record the air volume deviation rate as , = / *100%, ΔQ=| - |;

[0090] The The actual air volume of the fan after preprocessing in step S2, For the dynamic target air volume of the fan, This refers to the air volume deviation value.

[0091] Preset airflow deviation rate The threshold (e.g., 3%) is set. When the threshold is exceeded, the airflow accuracy reward becomes a negative reward, triggering the calibration mechanism; otherwise, the airflow accuracy reward becomes a positive reward, maintaining the current strategy parameters.

[0092] After the calibration mechanism is triggered, the policy parameters of the policy network are adjusted by optimizing the objective function through the dominance function and the policy, outputting a new frequency adjustment amount Δf, and calculating the airflow deviation rate. Repeat the above steps until the air volume deviation rate is reached. Within the preset correction range;

[0093] S5: Enables adaptive switching of operating conditions;

[0094] Once the strategy network identifies the corresponding operating condition, it switches to the optimal control strategy for this operating condition in the strategy library; and based on the optimal control strategy, it re-acquires the corresponding optimal operating frequency and target air volume to complete the operating condition switching adjustment.

[0095] S6: Apply safety constraints;

[0096] The safety constraints include frequency safety constraints, pressure safety constraints, and overload safety constraints;

[0097] Frequency safety constraints: The safe frequency range of the wind turbine inverter is preset (e.g., 20Hz-50Hz). The Δf output by the strategy network must ensure that the final frequency does not exceed this range; otherwise, the frequency limit will be triggered and the frequency will be forcibly adjusted to the boundary value.

[0098] Pressure safety constraint: Real-time monitoring of the fan outlet pressure; if it exceeds a preset threshold (e.g., ±500Pa), an emergency shutdown command is immediately triggered, and an audible and visual alarm is issued simultaneously.

[0099] Overload safety constraint: Monitor the fan motor current. If it exceeds 110% of the rated current, automatically reduce the frequency (1Hz each time, 0.2s interval) until the current returns to the rated range.

[0100] S7: System output and execution;

[0101] After multiple rounds of experience collection and iterative updates, the optimal frequency adjustment is finally determined to be the frequency adjustment that satisfies the training convergence criterion.

[0102] The core control module converts the optimal frequency adjustment into a control signal that the frequency converter can recognize and outputs it to the induced draft fan frequency converter to adjust the fan speed in real time.

[0103] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0104] This invention achieves air volume calibration response and operating condition switching adjustment by high-frequency acquisition of multi-source operating condition parameters and combining the real-time strategy update mechanism of the PPO algorithm. It can quickly capture strong dynamic operating condition changes such as blast furnace load change and converter blowing, ensure timely exhaust of flue gas and stable negative pressure, and completely solve the production safety and environmental emission hazards caused by adjustment lag.

[0105] This method constructs a policy network and a value network of a three-layer fully connected neural network to establish a precise matching model. It combines a multi-objective weighted reward function to guide policy optimization, thereby stabilizing the air volume control accuracy, effectively avoiding overshoot and oscillation, ensuring the stability of furnace top pressure and flue negative pressure, and improving the flue gas recovery and dust removal effect.

[0106] Through dynamic adaptive adjustment, the fan frequency follows the changes in smelting load and flue gas volume in real time: the frequency is automatically reduced under low load conditions and the frequency is precisely increased under high load conditions. At the same time, the intelligent agent is guided to learn energy-saving strategies through energy consumption reward items to avoid high-frequency idling, significantly reduce the energy consumption of fan operation, and improve the economic efficiency of equipment operation. Attached Figure Description

[0107] Figure 1 This is a flowchart illustrating the steps of an adaptive control method for the air volume and frequency of an induced draft fan according to the present invention. Detailed Implementation

[0108] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0109] An adaptive control system for the air volume and frequency of an induced draft fan, the system comprising a multi-source operating condition parameter acquisition module, a data preprocessing module, a core control module, and a safety constraint and training convergence module;

[0110] The multi-source operating condition parameter acquisition module is used to acquire multi-source operating condition parameters in real time.

[0111] The data preprocessing module is used to preprocess the collected multi-source operating condition parameters;

[0112] The core control module achieves dynamic adaptive matching of induced draft fan air volume and frequency based on the PPO algorithm, including constructing a strategy network and a value network, updating strategy parameters, air volume closed-loop calibration, and adaptive switching of operating conditions.

[0113] The security constraint and training convergence module sets the training convergence criteria for the PPO algorithm model and imposes security constraints through relevant security rules.

[0114] like Figure 1 As shown, the present invention also proposes an adaptive control method for the air volume and frequency of an induced draft fan, comprising the following sub-steps:

[0115] S1: Real-time acquisition of multi-source operating parameters;

[0116] Multiple types of high-precision industrial sensors are deployed in areas such as the induced draft fan, flue, furnace / furnace top of the blast furnace / converter dust removal system. The multi-source operating condition parameter acquisition module collects multi-source operating condition parameters in real time through high-precision industrial sensors at a preset acquisition frequency (such as 10Hz).

[0117] The high-precision industrial sensors include frequency sensors, pressure sensors, air volume sensors, etc.

[0118] The multi-source operating parameters include fan operating parameters, process operating parameters, environmental compensation parameters, and auxiliary parameters;

[0119] The fan operating parameters include the fan's current operating frequency, fan outlet pressure, fan actual air volume, fan motor power, fan current, and fan vibration value.

[0120] The process operating parameters include furnace / furnace top negative pressure, blast furnace smelting intensity / converter blowing signal, flue gas resistance / pressure difference, flue gas temperature, etc.

[0121] The environmental compensation parameters include ambient temperature, etc.

[0122] The auxiliary parameters include the opening status of the flue valve and auxiliary parameters that characterize the resistance to flue gas flow.

[0123] S2: Preprocess the collected multi-source operating condition parameters;

[0124] The data preprocessing module preprocesses the collected multi-source operating parameters, including standardization, noise filtering, and outlier removal.

[0125] Specifically, the standardization process involves the data preprocessing module collecting historical data of various multi-source operating parameters in advance and calculating the historical maximum and minimum values ​​of each multi-source operating parameter.

[0126] Record the historical maximum value as The historical minimum value is The parameters of each multi-source operating condition are standardized using the following normalization formula:

[0127] ;

[0128] Where x represents the collected multi-source operating condition parameters, These are the normalized multi-source operating condition parameters;

[0129] Noise filtering: A sliding window mean filtering algorithm is used to set an independent sliding window for each multi-source operating condition parameter, and the average value of the normalized multi-source operating condition parameter within the sliding window is calculated. This average value is the multi-source operating condition parameter value after filtering to eliminate noise.

[0130] Outlier removal: The data preprocessing module presets safety thresholds for each multi-source operating condition parameter, iterates through each multi-source operating condition parameter, identifies multi-source operating condition parameters that exceed the corresponding safety threshold as outliers and removes them, and supplements missing data through linear interpolation to obtain the preprocessed multi-source operating condition parameters.

[0131] S3: Achieve dynamic adaptive matching between induced draft fan air volume and frequency;

[0132] The core control module constructs a policy network and a value network through the PPO algorithm, and calculates the frequency adjustment of the wind turbine inverter based on the preprocessed multi-source operating parameters.

[0133] Includes the following sub-steps:

[0134] S31: Constructing strategy networks and value networks;

[0135] The core control module constructs a policy network and a value network using the PPO algorithm. Both the policy network and the value network adopt a 3-layer fully connected neural network structure.

[0136] The three-layer fully connected neural network structure includes an input layer, a hidden layer, and an output layer; the input layer contains 6 neurons; the hidden layer includes two layers, denoted as hidden layer 1 and hidden layer 2, each with 32 neurons, using the ReLU activation function; the output layer contains 1 neuron, which outputs the frequency adjustment amount of the wind turbine inverter;

[0137] The formula for the policy network is as follows:

[0138] ;

[0139] in, This refers to the frequency adjustment amount of the wind turbine inverter. , , The weight matrix is ​​randomly initialized, where It is 6 rows and 32 columns. 32 rows and 32 columns It consists of 32 rows and 1 column; For activation function, For activation functions; The initial bias vector, 32 rows and 1 column 32 rows and 1 column It is arranged in 1 row and 1 column (consistent with the number of neurons in the corresponding layer);

[0140] Specifically, S is a 6×1 dimensional input vector; It is a weight matrix with 6 rows and 32 columns; *S is a 6×32 weight matrix multiplied by a 6×1 dimension input vector, outputting a 32-dimensional vector (32×1), which corresponds to the original outputs of the 32 neurons in hidden layer 1; the ReLU activation function is applied to this 32-dimensional vector. This yields a 32-dimensional (32×1) feature vector, which is the final output of hidden layer 1;

[0141] Hidden layer 2 corresponds to 32 neurons, then the weight matrix... It has 32 rows and 32 columns. Multiplying by the final output of hidden layer 1, a 32 (32×1) dimensional feature vector is output, corresponding to the original outputs of the 32 neurons in hidden layer 2; the ReLU activation function is then applied to this 32-dimensional vector. This yields a 32-dimensional (32×1) feature vector, which is the final output of hidden layer 2;

[0142] If the output layer corresponds to one neuron, then the weight matrix... The final output of hidden layer 2 is transposed into a 1×32 row vector and a weight matrix, with a 32x1 matrix. Multiplying the data, the output is 1D data. This data is then activated by the Sigmoid function to output the frequency adjustment of the wind turbine inverter. ;

[0143] The value network adopts the exact same network structure as the policy network;

[0144] S32: Optimize the policy parameters of the policy network and output the optimal frequency adjustment amount;

[0145] Includes the following sub-steps:

[0146] S321: Initialize the core parameters of the PPO algorithm;

[0147] Load the policy network and value network, and initialize the core parameters of the PPO algorithm;

[0148] The core parameters of the PPO algorithm include the discount factor γ and the learning rate. , trimming factor ε, experience collection round length, strategy update frequency, and preset range of frequency adjustment amount;

[0149] In this embodiment, the initial discount factor γ = 0.95, the learning rate is 0.001, the pruning factor is 0.2, the experience collection round length is 100 steps (each step corresponds to one collection cycle), and the policy update frequency is once every 10 rounds. ∈[-5Hz, +5Hz];

[0150] S322: Generates initial frequency adjustment amount;

[0151] Set the duration of each collection cycle and the total number of collection cycles included in each experience collection round; obtain the corresponding multi-source working condition data in units of collection cycle duration; each collection cycle corresponds to a collection cycle number, let the collection cycle number be t, and the value range of t is [1, total number of collection cycles];

[0152] The current operating frequency of the fan, the outlet pressure of the fan, the actual air volume of the fan, the ambient temperature, the dynamic target air volume of the fan, and the air volume deviation corresponding to each collection cycle are respectively formed into independent input vectors S, which are input into the strategy network. The strategy network calculates and outputs the preliminary frequency adjustment amount through the formula. The preliminary frequency adjustment amount is limited to a preset range (such as [-5Hz, +5Hz]).

[0153] S323: Calculate cumulative discount rewards;

[0154] The instantaneous reward for each collection period is calculated using the following multi-objective weighted reward function;

[0155] ;

[0156] in, =1- / ; =1- / a; =1- / ;

[0157] in, The instant reward for the t-th gathering cycle in a single experience gathering round. , , These are the weighting coefficients for each reward item. + + =1; For air volume accuracy reward, For frequency smoothing rewards, As an energy consumption reward;

[0158] Record the actual air volume of the fan in the current data collection period as: The preset dynamic target air volume for the fan is Air volume deviation value The dynamic target air volume is Updated in real time based on smelting conditions; ΔQ=| - |;

[0159] , where a is the initial frequency adjustment amount for the t-th collection cycle in a single experience collection round, and a is the set frequency adjustment threshold. For the fan motor power in the t-th collection cycle of a single experience collection round, This is the preset rated power of the fan motor;

[0160] Calculate the cumulative discount reward using the following formula:

[0161] ;

[0162] Where T is the total number of collection cycles within a single experience collection round, and k is the summation index variable. The cumulative discount reward value for the t-th collection cycle in a single experience collection round;

[0163] S324: Calculate the advantage function and the policy optimization objective function;

[0164] Let the input vector corresponding to the t-th collection cycle in a single experience collection round be denoted as... ,Will The input is fed into the constructed value network, and the output of the value network is the value network's response to the input vector. The estimated value of is denoted as . , The network parameters are preset for the value network;

[0165] Advantage function ;

[0166] Optimize the objective function using a maximization strategy To achieve the goal, the strategy parameters are updated and iterated:

[0167] Strategy optimization objective function ;

[0168] in, For strategy parameters, The clipping constraint function is... The interval range is [ , ];

[0169] r represents the probability ratio between the old and new strategies. t (θ)= ;in, For the old strategy probability, The probability of the new strategy; The frequency adjustment amount for the t-th collection cycle in a single experience collection round. ; Adjust the frequency amount The input is fed into the Sigmoid activation function, and the output is the old policy probability. ;Optimize the objective function using a maximization strategy The specific value is the target, and the output is the optimized policy parameters, which yields the optimized policy network. The input vector is then used as the target. The input is fed into the optimized policy network, which outputs a frequency adjustment value. This frequency adjustment value is then input into the Sigmoid activation function to output the new policy probability. ; The empirical sampling distribution;

[0170] S325: Perform iterative updates to the strategy parameters;

[0171] Based on the advantage function calculated in step S324 Strategy optimization objective function The policy parameters of the policy network are updated using the stochastic gradient ascent method.

[0172] During each update, the probability ratio between the old and new strategies is limited by the clipping constraint function. Within the preset range;

[0173] Set the policy update to be performed every n experience collection rounds, where n is a positive integer; repeat steps S322-S325 until the training convergence criterion is met.

[0174] Specifically, the update amount of the strategy parameters is denoted as... , For the stochastic gradient ascent method, optimize the objective function of the policy. The gradient vector obtained by calculating the gradient. = * ; through learning rate However, it is not known that scaling the gradient vector is necessary to ensure that the parameter adjustment range is appropriate each time, avoiding policy oscillation due to excessive step size, and preventing inefficient training due to excessively small step size.

[0175] Specifically, the training convergence criterion is as follows:

[0176] The training convergence criterion is determined to be met when all three of the following conditions are met simultaneously.

[0177] (4) The average reward value fluctuation over n consecutive experience collection rounds is less than or equal to a set percentage (e.g., 5%).

[0178] The formula for calculating the average reward value fluctuation is as follows:

[0179] ;

[0180] in, The average reward value over n consecutive experience collection rounds. The cumulative discount reward value for the i-th experience collection round;

[0181] (5) The air volume control accuracy is within the set range (e.g., [-2%, 2%]) for n consecutive experience collection rounds.

[0182] The air volume control accuracy is defined as follows: The formula for calculating airflow control accuracy is as follows:

[0183] *100%;

[0184] (6) The accuracy rate of working condition identification is greater than or equal to the set percentage (e.g., 99.5%).

[0185] The policy network extracts features from the preprocessed multi-source working condition parameters and identifies the working conditions corresponding to the experience collection rounds; the actual results of each working condition are known.

[0186] Specifically, the strategy network extracts features through preprocessed multi-source operating parameters, including basic operating features such as fan outlet pressure and fan motor current, environmental features such as flue flow rate, and equipment status features such as motor temperature rise.

[0187] The extracted features are matched with feature templates in the strategy library; the feature templates in the strategy library include high load condition feature templates, low load condition feature templates, etc.

[0188] The weighted cosine similarity algorithm was used to obtain the weighted cosine similarity between the extracted features and the feature templates for high-load and low-load conditions, respectively.

[0189] A similarity interval between high-load and low-load operating conditions is pre-defined. If the calculated weighted cosine similarity falls within any of the above intervals, the condition is determined to be the operating condition corresponding to that interval. If the calculated weighted cosine similarity does not fall within any of the above intervals, the condition is determined to be a transitional operating condition.

[0190] The working conditions identified by the policy network are compared with the actual results of the working conditions. If they match, the identification is considered correct; otherwise, it is considered incorrect.

[0191] The accuracy of condition identification is calculated using the following formula:

[0192] ;

[0193] in, To improve the accuracy of working condition identification, The number of correctly identified working condition samples, The total number of operating condition samples identified;

[0194] Once all the above conditions are met, the training is considered to have converged, offline pre-training is stopped, the policy parameters of the current policy network are solidified, and the solidified policy network parameters are saved to the policy library.

[0195] Simultaneously, for both low-load and high-load operating conditions, specialized training is conducted using historical operating data (multi-source operating parameters, energy consumption data, etc.) corresponding to the operating conditions to form the optimal control strategy for the corresponding operating conditions, which is stored in the strategy library; the optimal control strategy includes the optimal operating frequency of the wind turbine inverter, the target air volume, etc.

[0196] S4: Real-time monitoring of airflow deviation rate and adjustment of strategy parameters of strategy network to perform airflow closed-loop calibration;

[0197] Record the air volume deviation rate as , = / *100%, ΔQ=| - |;

[0198] The The actual air volume of the fan after preprocessing in step S2, For the dynamic target air volume of the fan, This refers to the air volume deviation value.

[0199] Preset airflow deviation rate The threshold (e.g., 3%) is set. When the threshold is exceeded, the airflow accuracy reward becomes a negative reward, triggering the calibration mechanism; otherwise, the airflow accuracy reward becomes a positive reward, maintaining the current strategy parameters.

[0200] After the calibration mechanism is triggered, the policy parameters of the policy network are adjusted by optimizing the objective function through the dominance function and the policy, outputting a new frequency adjustment amount Δf, and calculating the airflow deviation rate. Repeat the above steps until the air volume deviation rate is reached. Within the preset correction range.

[0201] S5: Enables adaptive switching of operating conditions;

[0202] Once the strategy network identifies the corresponding operating condition, it switches to the optimal control strategy for this operating condition in the strategy library; and based on the optimal control strategy, it re-acquires the corresponding optimal operating frequency and target air volume to complete the operating condition switching adjustment.

[0203] Specifically, low-load operating conditions: corresponding to conditions such as converter shutdown, blast furnace stable tapping, and waiting for material, the amount of flue gas is small, and the frequency of the blowers needs to be reduced to achieve energy-saving operation;

[0204] High-load conditions: When the converter is tapping steel, measuring temperature, or blast furnace is blowing air, the flue gas volume increases dramatically. It is necessary to increase the frequency of the blowers, maintain stable negative pressure, and ensure timely exhaust of the flue gas.

[0205] S6: Apply safety constraints;

[0206] The safety constraints include frequency safety constraints, pressure safety constraints, and overload safety constraints;

[0207] Frequency safety constraints: The safe frequency range of the wind turbine inverter is preset (e.g., 20Hz-50Hz). The Δf output by the strategy network must ensure that the final frequency does not exceed this range; otherwise, the frequency limit will be triggered and the frequency will be forcibly adjusted to the boundary value.

[0208] Pressure safety constraint: Real-time monitoring of the fan outlet pressure; if it exceeds a preset threshold (e.g., ±500Pa), an emergency shutdown command is immediately triggered, and an audible and visual alarm is issued simultaneously.

[0209] Overload safety constraint: Monitor the fan motor current. If it exceeds 110% of the rated current, automatically reduce the frequency (1Hz each time, 0.2s interval) until the current returns to the rated range.

[0210] S7: System output and execution;

[0211] After multiple rounds of experience collection and iterative updates, the optimal frequency adjustment is finally determined to be the frequency adjustment that satisfies the training convergence criterion.

[0212] The core control module converts the optimal frequency adjustment into a control signal that the frequency converter can recognize and outputs it to the induced draft fan frequency converter to adjust the fan speed in real time.

[0213] Simultaneously, the adjustment data (frequency, air volume, energy consumption, deviation rate) is uploaded to the monitoring terminal for operation and maintenance personnel to view, and is also stored as historical data for subsequent model iterations.

[0214] This method, through the PPO algorithm's automatic condition identification capability (identification accuracy ≥ 99.5%), can accurately distinguish typical operating conditions such as high / low load, automatically call the corresponding optimal control strategy in the strategy library, and achieve adaptive matching throughout the entire smelting cycle without manual intervention, adapting to complex scenarios such as converter periodic blowing and blast furnace continuous variable load.

[0215] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A method for adaptive control of air volume and frequency of an induced draft fan, characterized in that: Including the following step: S1: Real-time acquisition of multi-source operating parameters; The multi-source operating condition parameter acquisition module acquires multi-source operating condition parameters in real time at a preset acquisition frequency through pre-deployed high-precision industrial sensors. S2: Preprocess the collected multi-source operating condition parameters; The data preprocessing module preprocesses the collected multi-source operating parameters, including standardization, noise filtering, and outlier removal. S3: Achieve dynamic adaptive matching between induced draft fan air volume and frequency; S31: Constructing strategy networks and value networks; S32: Optimize the policy parameters of the policy network and output the optimal frequency adjustment amount; S4: Real-time monitoring of airflow deviation rate and adjustment of strategy parameters of strategy network to perform airflow closed-loop calibration; S5: Enables adaptive switching of operating conditions; After identifying the operating condition through the strategy network, the system switches to the optimal control strategy corresponding to this operating condition in the strategy library; and based on the optimal control strategy, it re-acquires the corresponding optimal operating frequency and target air volume to complete the operating condition switching and adjustment. S6: Apply safety constraints; The safety constraints include frequency safety constraints, pressure safety constraints, and overload safety constraints; S7: System output and execution; After multiple rounds of experience collection and iterative updates, the optimal frequency adjustment is finally determined to be the frequency adjustment that satisfies the training convergence criterion. The core control module converts the optimal frequency adjustment into a control signal that the frequency converter can recognize and outputs it to the induced draft fan frequency converter to adjust the fan speed in real time.

2. The adaptive control method for induced draft fan air volume and frequency as described in claim 1, characterized in that: In step S1, the multi-source operating parameters include fan operating parameters, process operating parameters, environmental compensation parameters, and auxiliary parameters; The fan operating parameters include the fan's current operating frequency, fan outlet pressure, fan actual air volume, and fan motor power; The process operating parameters include furnace / furnace top negative pressure, blast furnace smelting intensity / converter blowing signal, flue gas resistance / pressure difference, and flue gas temperature; The environmental compensation parameters include ambient temperature; The auxiliary parameters include the opening status of the flue valve and auxiliary indicators of flue gas flow resistance.

3. The adaptive control method for induced draft fan air volume and frequency as described in claim 1, characterized in that: The specific details of step S31 are as follows: The core control module constructs a policy network and a value network using the PPO algorithm. Both the policy network and the value network adopt a 3-layer fully connected neural network structure. The three-layer fully connected neural network structure includes an input layer, a hidden layer, and an output layer; the input layer contains 6 neurons; the hidden layer includes two layers, denoted as hidden layer 1 and hidden layer 2, each with 32 neurons, using the ReLU activation function; the output layer contains 1 neuron, which outputs the frequency adjustment amount of the wind turbine inverter; The formula for the policy network is as follows: ； in, , , , For strategy parameters; This is the frequency adjustment amount; , , The weight matrix is ​​randomly initialized in the policy parameters, where It is 6 rows and 32 columns. 32 rows and 32 columns It consists of 32 rows and 1 column; For activation function, For activation functions; This is the initial bias vector in the policy parameters. 32 rows and 1 column 32 rows and 1 column It consists of 1 row and 1 column; The value network adopts the exact same network structure as the policy network.

4. The adaptive control method for induced draft fan air volume and frequency as described in claim 1, characterized in that: Step S32 includes the following sub-steps: S321: Initialize the core parameters of the PPO algorithm; Load the policy network and value network, and initialize the core parameters of the PPO algorithm; The core parameters of the PPO algorithm include discount factor γ, pruning coefficient ε, experience collection round length, strategy update frequency, and preset range of frequency adjustment amount. S322: Generates initial frequency adjustment amount; Set the duration of each collection cycle and the total number of collection cycles included in each experience collection round; obtain the corresponding multi-source working condition data in units of collection cycle duration; each collection cycle corresponds to a collection cycle number, let the collection cycle number be t, and the value range of t is [1, total number of collection cycles]; The current operating frequency of the fan, the outlet pressure of the fan, the actual air volume of the fan, the ambient temperature, the dynamic target air volume of the fan, and the air volume deviation corresponding to each collection cycle are respectively formed into independent input vectors S, which are input into the strategy network. The strategy network calculates and outputs the preliminary frequency adjustment amount through the formula, and the preliminary frequency adjustment amount is limited to a preset range. S323: Calculate cumulative discount rewards; S324: Calculate the advantage function and the policy optimization objective function; S325: Perform iterative updates to the strategy parameters.

5. The adaptive control method for induced draft fan air volume and frequency as described in claim 4, characterized in that: The specific details of step S323 are as follows: The instantaneous reward for each collection period is calculated using the following multi-objective weighted reward function; ； in, =1- / ; =1- / a; =1- / ; in, The instant reward for the t-th gathering cycle in a single experience gathering round. , , These are the weighting coefficients for each reward item. + + =1; For air volume accuracy reward, For frequency smoothing rewards, As an energy consumption reward; Record the actual air volume of the fan in the current data collection period as: The preset dynamic target air volume for the fan is Air volume deviation value The dynamic target air volume is Updated in real time based on smelting conditions; ΔQ=| - |; , where a is the initial frequency adjustment amount for the t-th collection cycle in a single experience collection round, and a is the set frequency adjustment threshold. For the fan motor power in the t-th collection cycle of a single experience collection round, This is the preset rated power of the fan motor; Calculate the cumulative discount reward using the following formula: ； Where T is the total number of collection cycles within a single experience collection round, and k is the summation index variable. The cumulative discount reward value for the t-th collection cycle in a single experience collection round.

6. The adaptive control method for induced draft fan air volume and frequency as described in claim 4, characterized in that: The specific details of step S324 are as follows: Let the input vector corresponding to the t-th collection cycle in a single experience collection round be denoted as... ,Will The input is fed into the constructed value network, and the output of the value network is the value network's response to the input vector. The estimated value of is denoted as . Advantage function ; Optimize the objective function using a maximization strategy To achieve the goal, the policy parameters are updated and iterated: Strategy optimization objective function ; in, For strategy parameters, The clipping constraint function is... The interval range is [ , ]; r represents the probability ratio between the old and new strategies. t (θ)= ;in, For the old strategy probability, The probability of the new strategy; The frequency adjustment amount for the t-th collection cycle in a single experience collection round. ; Adjust the frequency amount The input is fed into the Sigmoid activation function, and the output is the old policy probability. ;Optimize the objective function using a maximization strategy The specific value is the target, and the output is the optimized policy parameters, which yields the optimized policy network. The input vector is then used as the target. The input is fed into the optimized policy network, which outputs a frequency adjustment value. This frequency adjustment value is then input into the Sigmoid activation function to output the new policy probability. ; This is an empirical sampling distribution.

7. The adaptive control method for induced draft fan air volume and frequency as described in claim 4, characterized in that: The specific details of step S325 are as follows: Based on the advantage function calculated in step S324 Strategy optimization objective function The policy parameters of the policy network are updated using the stochastic gradient ascent method. During each update, the probability ratio between the old and new strategies is limited by the clipping constraint function. Within the preset range; Set the policy update to be performed every n experience collection rounds, where n is a positive integer; repeat steps S322-S325 until the training convergence criterion is met. The training convergence criteria are as follows: The training convergence criterion is determined to be met when all three of the following conditions are met simultaneously. (1) The average reward value fluctuation over n consecutive experience collection rounds is less than or equal to a set percentage (e.g., 5%). The formula for calculating the average reward value fluctuation is as follows: ； in, The average reward value over n consecutive experience collection rounds. The cumulative discount reward value for the i-th experience collection round; (2) The air volume control accuracy is within the set range (e.g., [-2%, 2%]) for n consecutive experience collection rounds. The air volume control accuracy is set as follows: The formula for calculating airflow control accuracy is as follows: *100%； (3) The accuracy rate of working condition identification is greater than or equal to the set percentage (e.g., 99.5%). The policy network extracts features from the preprocessed multi-source working condition parameters and identifies the working conditions corresponding to the experience collection rounds; the actual results of each working condition are known. The working conditions identified by the policy network are compared with the actual results of the working conditions. If they match, the identification is considered correct; otherwise, it is considered incorrect. The accuracy of condition identification is calculated using the following formula: ； in, To improve the accuracy of working condition identification, The number of correctly identified working condition samples, The total number of operating condition samples identified; Once all the above conditions are met, the training is considered to have converged, offline pre-training is stopped, the policy parameters of the current policy network are solidified, and the solidified policy network parameters are saved to the policy library. Simultaneously, for both low-load and high-load operating conditions, specific training is conducted using historical operating data corresponding to the operating conditions to form the optimal control strategy for the corresponding operating conditions, which is stored in the strategy library; the optimal control strategy includes the optimal operating frequency of the fan inverter and the target air volume.

8. The adaptive control method for induced draft fan air volume and frequency as described in claim 1, characterized in that: The specific details of step S4 are as follows: Record the air volume deviation rate as , = / *100%, ΔQ=| - |; The The actual air volume of the fan after preprocessing in step S2, For the dynamic target air volume of the fan, This refers to the air volume deviation value. Preset airflow deviation rate The threshold is set, and when the threshold is exceeded, the airflow accuracy reward is negative, triggering the calibration mechanism; otherwise, the airflow accuracy reward is positive, maintaining the current strategy parameters. After the calibration mechanism is triggered, the policy parameters of the policy network are adjusted by optimizing the objective function through the dominance function and the policy, outputting a new frequency adjustment amount Δf, and calculating the airflow deviation rate. Repeat the above steps until the air volume deviation rate is reached. Within the preset correction range.

9. An adaptive control system for the air volume and frequency of an induced draft fan for implementing the method according to any one of claims 1-8, characterized in that: The system includes a multi-source operating condition parameter acquisition module, a data preprocessing module, a core control module, and a safety constraint and training convergence module. The multi-source operating condition parameter acquisition module is used to acquire multi-source operating condition parameters in real time. The data preprocessing module is used to preprocess the collected multi-source operating condition parameters; The core control module achieves dynamic adaptive matching of induced draft fan air volume and frequency based on the PPO algorithm, including constructing a strategy network and a value network, updating strategy parameters, air volume closed-loop calibration, and adaptive switching of operating conditions. The security constraint and training convergence module sets the training convergence criteria for the PPO algorithm model and imposes security constraints through relevant security rules.

Images