Digital air compression station intelligent control method and system based on AI self-learning

Abstract

The invention relates to the technical field of industrial compressed air system intelligent control, in particular to a digital air compression station intelligent control method and system based on AI self-learning, and the system comprises a data collection layer, an AI prediction layer, an optimization control layer and an execution layer. The data acquisition layer realizes accurate acquisition and preprocessing of multi-dimensional data through a three-stage high-precision sensor network and an edge computing node; the AI prediction layer adopts an XGBoost and LSTM mixed model to be combined with an online learning mechanism to realize accurate prediction of the gas consumption demand and the energy efficiency ratio within 1-24 hours; the optimization control layer completes air compressor operation combination optimization and pressure dynamic control based on an improved NSGA-II algorithm and self-adaptive PID adjustment; the execution layer realizes control instruction landing through frequency converter and cluster cooperative scheduling. Meanwhile, an equipment health degree monitoring module, a multi-source data fusion correction module and a waste heat-gas consumption collaborative optimization module are innovatively introduced, and the problems that a traditional air compression station is low in energy efficiency, poor in self-adaption, insufficient in prediction precision and lack of equipment management are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control of industrial compressed air systems, in particular to an AI self-learning-based intelligent control method and system for a digital air compression station. BACKGROUND

[0002] In an industrial production system, an air compression station, as a core power source, provides compressed air for various production equipment, and its energy consumption accounts for 10%-30% of the total industrial electricity consumption, which is a key link for energy saving and emission reduction in the industrial field. With the transformation of manufacturing industry towards intelligence and green, the control mode of traditional air compression stations has been difficult to meet the needs of modern industry for energy efficiency, stability and management accuracy, and its technical defects have gradually become a bottleneck restricting enterprises to reduce costs and increase efficiency.

[0003] Traditional air compression station control mainly relies on manual experience or simple PID adjustment, and has significant deficiencies in supply-demand matching. Because the start-stop and load change of gas-consuming equipment in the production process have randomness and volatility, manual experience cannot capture the dynamic changes of gas demand in real time, resulting in the widespread existence of the "horse pulling a cart" phenomenon - that is, the air compressor is in a long-term full-load operation state, while the actual gas demand is far below the rated capacity of the equipment, causing a large amount of waste of electric energy; at the same time, in the event of a sudden gas demand peak, the supply pressure often drops sharply due to insufficient prediction, affecting the normal operation of production equipment and causing fluctuations in product quality or production interruption. According to industry statistics, the energy waste caused by low supply-demand matching precision of traditional air compression stations accounts for 15%-20% of the total energy consumption, seriously affecting the energy utilization efficiency of enterprises.

[0004] To avoid insufficient pressure at the gas demand peak, traditional air compression stations usually use a conservative pressure setting strategy to maintain the main pipe pressure at a level much higher than the actual demand (generally 0.2-0.3 bar higher than the actual demand). This strategy can ensure the stability of gas supply to a certain extent, but it will significantly increase the operating load of the air compressor. On the one hand, the high pressure setting makes the air compressor need to overcome greater pipe network resistance, resulting in an increase in energy consumption per unit of gas production; on the other hand, long-term high-pressure operation accelerates the wear of internal components of the air compressor, such as piston rings, bearings and other consumable parts, which shortens their service life by more than 30%, increasing equipment maintenance costs and downtime for repair. Data from an air compression station of a chemical company shows that for every 0.1 bar reduction in pressure setting, unit energy consumption can be reduced by 1.2%-1.5%, and the conservative pressure setting under the traditional control mode causes the company to consume more than 100,000 kWh of electricity per year.

[0005] In terms of the coordinated operation of multiple air compressors, traditional air compression stations lack intelligent linkage mechanisms, and the start-stop and load distribution of equipment are mostly based on fixed timing or manual operation. This approach results in frequent loading and unloading of air compressors, and some equipment is in a long-term low-load rate (<50%) operation state. When running at low load, the specific power (kW / m3 / min) significantly increased, and energy efficiency decreased substantially; at the same time, frequent loading and unloading processes generated transient impact current (traditional star-delta starting current was 2-3 times of the rated current), which not only caused impact on the power grid, but also aggravated the fatigue damage of the motor and mechanical parts, further shortening the service life of the equipment. A certain automobile parts factory air compression station once appeared because of 3 air compressors disorderly start-stop, the number of start-stop reached 127 times a day, which led to the burning of the motor of one of the air compressors, the direct economic loss was more than 50,000 yuan, and the indirect loss caused by 8 hours of shutdown was incalculable.

[0006] In recent years, the industry has tried to introduce intelligent technologies such as frequency conversion control and time sequence start-stop to improve the control effect of air compression station. Frequency conversion control adjusts the motor speed to achieve dynamic adjustment of gas production, which to some extent reduces the "loading and unloading" energy consumption, but the single frequency conversion control cannot realize the global optimization of multiple devices, and when multiple frequency converters and power frequency machines are running, there is still a problem of uneven load distribution; the time sequence start-stop technology can control the start and stop of the equipment in a fixed order, but it cannot dynamically adjust according to the real-time gas demand and equipment energy efficiency state, and has poor adaptability. With the development of artificial intelligence technology, AI prediction and optimization algorithms such as XGBoost and genetic algorithm have been gradually introduced into the field of air compression station control, but existing solutions mostly rely on fixed models and lack effective self-learning ability. When production plans are adjusted (such as production load fluctuations caused by changes in order quantity), environmental parameters change (such as a large change in environmental temperature caused by seasonal replacement), or equipment state ages (such as the decrease of air compressor energy efficiency ratio year by year), the prediction accuracy and optimization effect of fixed models will be significantly reduced, making it difficult to adapt to complex and variable working conditions. For example, after the environmental temperature of a certain chemical plant air compression station increased by 10℃ in summer, the gas consumption prediction error increased from 6% to 15%, which led to an increase in the deviation between supply and demand, and the pressure fluctuation range expanded to ±0.2bar, seriously affecting the production stability.

[0007] In addition, the traditional air compression station also has problems such as lack of equipment health management and insufficient comprehensive utilization of energy. Existing systems mostly focus on the operation control of air compressors, lack real-time monitoring of key health indicators such as equipment vibration and bearing temperature, and cannot provide early warning of equipment failure, often resulting in high maintenance costs and long downtime; at the same time, a large amount of waste heat (about 75%-85% of input power) generated during the operation of the air compression station is directly discharged, without being used in coordination with gas demand, plant heating, etc., resulting in energy waste. Therefore, it is an urgent need to develop a digital air compression station intelligent control technology with precise prediction, dynamic optimization, self-learning adaptation, and full life cycle management capabilities to solve the current industry pain points and promote the intelligent upgrading of industrial air compression stations. SUMMARY

[0008] The present application aims to overcome the technical defects of low energy efficiency, poor adaptability, insufficient prediction accuracy, lack of equipment management and insufficient comprehensive utilization of energy in traditional air compression station control, and provides an AI self-learning-based digital air compression station intelligent control method and system, which specifically needs to solve the following problems: the traditional control method has low supply-demand matching accuracy, cannot predict the gas demand in real time, leading to energy waste or insufficient gas supply; the pressure setting is conservative, and the PID parameters are fixed, which cannot be dynamically adjusted according to the pressure fluctuation, resulting in rising energy consumption and pressure oscillation; multiple air compressors lack intelligent collaborative scheduling, frequent loading and unloading, shortened equipment life and low overall energy efficiency; the existing AI control scheme relies on fixed models and has no self-learning ability, making it difficult to adapt to changes in production, environment and equipment state; there is a lack of real-time monitoring of equipment health, weak fault warning capability and high maintenance cost; the waste heat of the air compression station and the gas demand are not optimized in coordination, and the comprehensive utilization rate of energy is low.

[0009] The technical solution adopted by the present application to solve its technical problems is: an AI self-learning-based digital air compression station intelligent control method, comprising the following steps: S1, data acquisition and preprocessing: collecting the pressure, flow, temperature, power data and environmental temperature and humidity, atmospheric pressure data of the air compression station main pipe, branch pipe and end gas point through a three-level high-precision sensing network, the pressure sensor has an accuracy of 0.1%FS and a sampling frequency of ≥10Hz, the flow measurement error is <1.5%, and the power acquisition adopts a Hall effect current sensor; the edge computing node performs data cleaning using the 3σ principle, extracts time domain features, frequency domain features and EER energy efficiency ratio working condition features; S2, AI accurate prediction: constructing an XGBoost and LSTM hybrid prediction model, the XGBoost inputs historical gas consumption, production plan OEE and a cubic term expansion of environmental temperature trend term, the LSTM network contains 64 neuron input layer, 32 neuron hidden layer and Dropout layer, and predicts the period term using Huber loss (δ=1.0) and Nadam optimizer (lr=0.001), and weighted fusion obtains 1-24 hour gas demand curve and confidence interval; the model is evaluated at 0 o'clock every day, when MAE>8%, the model is fine-tuned by fixing the weights of the first three layers of XGBoost to realize incremental training using near 30-day data and K-means clustering key samples; S3, optimization control decision: using an improved NSGA-II algorithm, the chromosome coding contains air compressor state (0 stop, 1 variable frequency 50-100%, 2 power frequency) and main pipe pressure setting, and the optimal running combination is solved by taking the minimum total energy consumption, minimum gas supply deviation and load rate close to 75% as objective functions; combined with adaptive PID, Kp is adjusted to 0.8x and 1.5x of the baseline value, and Ti is adjusted to 1.2x and 0.7x of the baseline value (the baseline value is determined by the Ziegler-Nichols method) according to the pressure fluctuation ±0.1bar and ±0.3bar interval respectively. S4, execution and feedback: the frequency converter dynamically adjusts the rotating speed to 30-50 Hz, the air compressor cluster is scheduled according to the rules of 40-80% basic load of the frequency converter, high energy efficiency, priority of less running equipment, and real-time operation data is collected and fed back to the AI prediction layer to update the model.

[0010] Specifically, in S1, the time domain features are 5-minute sliding average flow and pressure fluctuation variance, the frequency domain features are extracted by FFT to extract 0.1-1Hz frequency band energy, and the EER energy efficiency ratio = actual gas production (m 3 / min) / input power (kW).

[0011] Specifically, in S2, the mixed prediction model output also includes energy efficiency prediction results, and the air compressor with EER <75% is marked as an inefficient device and fed back to the optimization control layer for priority adjustment.

[0012] Specifically, in S3, the constraint conditions of the improved NSGA-II algorithm include total gas production ≥ predicted demand and single machine load rate ∈ [50%, 100%].

[0013] Specifically, it also includes a device health monitoring step: collecting air compressor vibration and bearing temperature data, constructing a health evaluation model based on CNN-LSTM, and calculating a health index in real time. When the index <60, a maintenance warning is triggered.

[0014] Specifically, in S1, the data preprocessing also uses multi-source data fusion correction: through Kalman filtering to fuse pressure and flow data, correct measurement deviation caused by sensor drift, and ensure data accuracy.

[0015] Specifically, it also includes a waste heat-gas collaborative optimization step: collecting hot water temperature and flow data of the air compression station waste heat recovery system, establishing a waste heat output and gas demand correlation model, and scheduling air compressors with high waste heat output when optimizing operation combination to improve overall energy utilization.

[0016] The AI self-learning-based digital air compression station intelligent control system comprises a data acquisition layer, a three-level high-precision sensing network (piezoelectric pressure sensor, thermal + vortex flow sensor, Hall effect current sensor, temperature and humidity / air pressure sensor) and an edge computing node (deploying data cleaning and feature extraction algorithms); an AI prediction layer, deploying an XGBoost module, an LSTM module, a weighted fusion layer and an online learning module, to realize air demand and energy efficiency ratio prediction; an optimization control layer, comprising an improved NSGA-II algorithm optimization engine and a self-adaptive PID adjustment module, to output optimal operation combination and pressure set value; an execution layer, comprising a frequency converter (controlling motor speed 30-50Hz, starting current ≤1.2 times rated current) and an air compressor cluster scheduling unit, to execute control instructions; and further comprising a device health monitoring module (vibration / bearing temperature sensor, CNN-LSTM evaluation model), a multi-source data fusion module (Kalman filtering algorithm) and a waste heat coordination module (waste heat data acquisition, correlation model).

[0017] Specifically, the edge computing node adopts an NVIDIA Jetson AGX Orin, deploys a lightweight TensorRT model (FP16 precision) and realizes real-time control with a cycle ≤100ms; communication adopts Modbus RTU (sensor-edge node), MQTT+JSON (edge node-cloud platform) and RESTAPI (cloud platform-user terminal), and data transmission adopts AES-256 encryption.

[0018] Specifically, the system has fault tolerance capability: when a failure is predicted, switching to a "flow-pressure" lookup table method, and when communication is interrupted, the edge node runs the latest optimization strategy autonomously; PC / APP remote monitoring is supported, and the fault prediction accuracy is >90%.

[0019] The present application has the following advantages: Energy efficiency is significantly improved: through a hybrid prediction model to accurately predict air demand, an improved NSGA-II algorithm to optimize operation combination, and a self-adaptive PID to dynamically adjust pressure, the comprehensive energy consumption is reduced by 5%-10%; at the same time, waste heat-air coordination optimization improves the comprehensive energy utilization rate by 15%-20%, solving the problems of high energy consumption and energy waste in traditional air compression stations.

[0020] Strong adaptive ability: online learning mechanism ensures daily evaluation and incremental update of the model, which can adapt to production plan adjustment (such as OEE change), environmental parameter change (such as temperature fluctuation) and equipment state aging in real time, with a prediction error stable control at <5%, avoiding the problem of precision decline of fixed models when the working condition changes.

[0021] Device life extension: inverter control reduces start-up impact (start-up current ≤ 1.2 times rated current), cluster scheduling avoids frequent loading and unloading (device start-stop times / day from 127 to 41, reduced by 67.7%), health monitoring early warning of failure, maintenance cost reduced by more than 30%, service life extended by 2-3 years.

[0022] High running stability: three-level sensor network and multi-source data fusion ensure data accuracy, adaptive PID control reduces pressure fluctuation from ±0.2bar to ±0.1bar (reduced by 50%), fault tolerance mechanism avoids downtime caused by prediction failure or communication interruption, and ensures continuous and stable production.

[0023] High level of intelligent management: supports PC / APP remote monitoring, fault prediction accuracy > 90%, realizes "unattended, less maintenance" of air compression station, reduces labor cost; at the same time, full-dimensional data acquisition and analysis provide data support for air compression station operation and maintenance optimization, and promote the transformation of industrial air compression station to digital and intelligent. BRIEF DESCRIPTION OF DRAWINGS

[0024] The present application will be further described below in conjunction with the drawings and examples.

[0025] Figure 1 The flowchart of the AI self-learning based digital air compression station intelligent control method provided by the present application; Figure 2 The architecture diagram of the AI self-learning based digital air compression station intelligent control system provided by the present application. DETAILED DESCRIPTION

[0026] In order to make the technical means, creative features, purposes and effects realized by the present application easy to understand, the present application will be further described below in conjunction with specific embodiments.

[0027] As shown in Figure 1 The AI self-learning based digital air compression station intelligent control method described in the present application: The method includes four core steps of data collection and preprocessing, AI accurate prediction, optimization control decision, execution and feedback, and integrates device health monitoring, multi-source data fusion correction and waste heat-gas utilization collaborative optimization modules, the specific process is as follows: (1) Data collection and preprocessing The comprehensive and accurate collection of multi-dimensional data at the air compression station is achieved through a three-level high-precision sensing network. Piezoelectric pressure sensors with a precision of 0.1% FS are installed at the main pipe, branch pipe, and end gas points of the air compression station. The sampling frequency is set to ≥10 Hz to ensure real-time capture of pressure dynamic changes at different positions. The flow measurement uses a combination of thermal mass flow meters and vortex flow meters to control the measurement error to <1.5% and accurately obtain the compressed air flow data at each link. For power monitoring, a Hall effect current sensor is installed for each air compressor to collect real-time power parameters such as input power and power factor. At the same time, temperature and humidity sensors and barometers are integrated to collect environmental temperature and humidity and atmospheric pressure data for subsequent working condition correction in energy efficiency ratio calculation.

[0028] The edge computing node uses the NVIDIA Jetson AGX Orin hardware platform to deploy data cleaning and feature extraction algorithms. In the data cleaning stage, the 3σ principle (i.e., data deviating from the rolling mean by 3 times the rolling standard deviation is considered abnormal) is used to detect and remove abnormal data. Specifically, the outlier_detection function is written in Python, the sliding window is set to 30 data points, and after calculating the rolling mean and rolling standard deviation, the abnormal values are identified and replaced with the interpolation of adjacent normal data. In the feature extraction stage, three types of key features are extracted: time domain features (5-minute sliding average flow, pressure fluctuation variance, reflecting the overall trend and stability of the data), frequency domain features (extracting 0.1-1 Hz frequency band energy through FFT transformation, which is the characteristic frequency of the air equipment start-stop, and can indirectly judge the production equipment running state), and working condition features (energy efficiency ratio EER = actual gas production (m 3 / min) / input power (kW), directly reflecting the real-time energy efficiency level of the air compressor).

[0029] In addition, a multi-source data fusion correction module is introduced, and Kalman filtering algorithm is used to fuse and process pressure and flow data. Due to the drift error of different sensors (such as zero drift of pressure sensors after long-term operation), relying solely on data from a single sensor can easily lead to measurement deviation. Kalman filtering establishes a state equation and an observation equation, combines the correlation of pressure and flow data, and real-time corrects the measurement error to ensure that the data accuracy meets the subsequent prediction and control requirements.

[0030] (2) AI precise prediction An XGBoost and LSTM hybrid prediction model is constructed to achieve precise prediction of 1-24 hour gas demand and energy efficiency ratio. This hybrid model fully combines the strong ability of XGBoost to handle nonlinear features and the advantage of LSTM to capture time series periodic features, and the specific architecture is as follows: The XGBoost module is responsible for trend item prediction, and the input features include: historical air consumption (selecting air consumption data with a lag of 1 h, 4 h, and 24 h to cover short-term, medium-term, and long-term historical rules), production plan OEE (device comprehensive efficiency, reflecting production load level, and the higher the OEE, the greater the air demand), and environmental temperature (expanding to the 3rd order through a polynomial to fit the nonlinear relationship between environmental temperature and air consumption, such as the increased air consumption of air equipment for heat dissipation in summer, which indirectly leads to increased compressed air consumption). XGBoost learns the mapping relationship between the above features and air demand trends by building a multi-decision tree ensemble model.

[0031] The LSTM module is responsible for periodic item prediction, and the network structure is built through the Keras framework of Python: the input layer contains 64 LSTM neurons, the input shape is (24, 10) (i.e. 24 time steps, each time step contains 10 features, including historical periodic air consumption and device start-stop frequency), and return_sequences=True is set to output sequence data; the hidden layer contains 32 LSTM neurons, and a Dropout(0.2) layer is added in the middle to prevent overfitting; the output layer is a Dense neuron with a relu activation function to output the periodic item prediction value. The Huber loss function (δ=1.0) is used for model training, which is robust to outliers and avoids interference from extreme air consumption data during model training; the optimizer is Nadam (learning rate lr=0.001), which combines the advantages of Nesterov momentum and Adam adaptive learning rate to speed up model convergence.

[0032] The hybrid model outputs the final prediction result through a weighted fusion layer, and the weights are dynamically adjusted according to the prediction period: in short-term prediction (1-6 hours), the LSTM periodic item weight is set to 0.6 (because short-term air demand is greatly affected by device start-stop cycles), and the XGBoost trend item weight is set to 0.4; in long-term prediction (7-24 hours), the XGBoost trend item weight is increased to 0.6 (because long-term air demand depends more on production plans and historical trends), and the LSTM periodic item weight is reduced to 0.4. At the same time, the model output also includes the EER prediction result, and the air compressor with EER < 75% is marked as an inefficient device and fed back to the subsequent optimization control layer for priority adjustment.

[0033] To ensure the model adapts to changes in operating conditions, an online learning mechanism is designed: at 0 o'clock every day, the model evaluation is automatically triggered, and the mean absolute error (MAE) of the predicted value and the actual gas consumption is calculated; if MAE > 8%, the incremental training process is started. During incremental training, the normal operation data of the last 30 days are retained, and key operating condition samples (such as extreme gas consumption peaks, equipment switching operating conditions, etc. rare samples) are selected through the K-means clustering algorithm to avoid data distribution drift; the transfer learning strategy is adopted, the weights of the first 3 layers of decision trees in XGBoost are fixed (the learned general features are retained), and only the top layer of decision tree weights is fine-tuned to reduce the amount of training data and computational overhead, ensuring the efficiency of model updating.

[0034] (3) Optimal control decision An optimization control strategy combining improved NSGA-II algorithm and adaptive PID regulation is adopted to realize the optimization of air compressor operation combination and dynamic control of mother pipe pressure, as follows: The improved NSGA-II algorithm is used to solve the multi-objective optimal operation combination to achieve the goals of minimum total energy consumption, minimum gas supply deviation, and balanced device load rate (tending to 75% optimal load rate). The core design of the algorithm includes: Chromosome coding: integer coding is used, and the chromosome structure is [air compressor 1 state, air compressor 2 state,..., air compressor N state, total pipe pressure setting]. The air compressor state coding rule is 0 = stop, 1 = variable frequency operation (load rate 50-100%), and 2 = fixed frequency operation (rated load rate); the total pipe pressure setting value is coded with 0.01 bar precision to ensure the pressure control accuracy.

[0035] Objective function: a three-objective minimization function is constructed: where, is the input power of the i-th air compressor, n is the total number of air compressors, and the minimum achieves the reduction of total energy consumption of the air compression station; where, is the actual gas supply at time j, is the predicted demand at time j, and m is the total number of time series, and the minimum achieves accurate matching of gas supply and demand; where, is the load rate of the k-th running air compressor, and p is the number of running air compressors, and the minimum achieves the optimal interval of the load rate of each air compressor tending to 75%, avoiding inefficient operation.

[0036] Constraint conditions: 1. Total gas production constraint (sufficient gas supply): ,in, Let be the air output of the i-th air compressor. To predict total demand throughout the entire time period.

[0037] 2. Single-machine load rate constraint (energy efficiency guarantee): ( Ensure that the load rate of each operating air compressor is in the high-efficiency range to avoid energy efficiency degradation due to excessively low load rate.

[0038] Algorithm optimization process: Through fast non-dominated sorting, crowding calculation and elite retention strategy, a Pareto optimal solution set is generated, and then the final operation combination scheme is selected according to the actual needs of the air compressor station (such as energy consumption priority or stability priority).

[0039] Adaptive PID control is used to dynamically adjust the head pipe pressure setpoint, solving the problem of poor adjustment performance of traditional fixed PID parameters when pressure fluctuates. First, the PID baseline parameters (Kp0, Ti0, Td0) are determined using the Ziegler-Nichols method: Under pure proportional control, Kp is gradually increased until the system exhibits constant-amplitude oscillations, and the oscillation period T0 and the critical proportional coefficient Kp are recorded. cr Calculate the baseline parameter Kp0 = 0.6 × Kp cr Ti0 = 0.5 × T0, Td0 = 0.125 × T0; then the parameters are dynamically adjusted according to the real-time pressure fluctuation range, and the specific rules are as follows: When the pressure fluctuation is ≤ ±0.1 bar (system is stable), Kp = 0.8 × Kp0, Ti = 1.2 × Ti0, reduce the proportional action and increase the integral action to avoid pressure oscillation caused by over-adjustment; When the pressure fluctuation is greater than ±0.3 bar (large system deviation), Kp = 1.5 × Kp0 and Ti = 0.7 × Ti0, the proportional action is enhanced, the integral time is shortened, and the deviation correction speed is accelerated.

[0040] In addition, an integrated waste heat-gas utilization synergistic optimization module is used: it collects hot water temperature (usually 80-100℃) and hot water flow data of the air compressor station's waste heat recovery system, and establishes a correlation model between waste heat output and air compressor operating status (e.g., the waste heat output of a fixed frequency compressor is 20%-30% higher than that of a variable frequency compressor); a waste heat utilization benefit term (e.g., waste heat output × plant heating demand coefficient) is added to the objective function of the improved NSGA-II algorithm; when optimizing the operation combination, air compressors with high waste heat output are prioritized for scheduling, and the waste heat is used for plant heating or production process heating to improve the overall energy utilization rate.

[0041] (4) Execution and feedback The execution layer realizes the landing execution of the optimization control instruction through the frequency converter and the air compressor cluster scheduling unit, and feeds back the running data to the AI prediction layer in real time to form a closed-loop control.

[0042] Frequency converter control: configure a high-performance frequency converter for each variable frequency air compressor, dynamically adjust the motor speed to the range of 30-50Hz according to the load rate instruction output by the optimization control layer (this speed interval corresponds to 50-100% load rate of the air compressor, and the energy efficiency is higher). The frequency converter uses vector control technology to ensure that the speed regulation accuracy is ≤1Hz, so that the gas production accurately matches the gas demand; at the same time, the frequency converter limits the motor starting current to within 1.2 times of the rated current (the traditional star-triangle starting current is 2-3 times of the rated current), reduces the starting impact, and prolongs the service life of the motor and mechanical parts.

[0043] Air compressor cluster cooperative scheduling: according to the optimization operation combination scheme, follow the "three priority" scheduling rules: Variable frequency machines prefer to bear basic load (40-80% load rate) and use the adjustment flexibility of variable frequency machines to respond to gas demand fluctuations; High EER efficiency machines are preferentially started, and the current most energy-efficient machines are selected for operation through real-time EER monitoring; Devices with less cumulative running time are preferentially started to balance the running time of each device and avoid excessive wear of some devices.

[0044] Running data feedback: the execution layer collects the actual load rate, input power, exhaust temperature, vibration value and other running data of the air compressor in real time, transmits them to the edge computing node through the communication protocol, and feeds them back to the AI prediction layer after preprocessing, which is used for model evaluation and incremental training to ensure continuous optimization of the control strategy.

[0045] (5) Equipment health monitoring In order to realize the whole life cycle management of the equipment, the equipment health monitoring module is introduced: vibration sensors (sampling frequency ≥1kHz) and temperature sensors (measurement accuracy ±0.5℃) are arranged at key parts such as motors, bearings and cylinders of each air compressor to collect vibration acceleration and bearing temperature data in real time; a health evaluation model based on CNN-LSTM is constructed, the CNN module extracts the frequency domain features of vibration data (such as frequency spectrum peak value and harmonic component), the LSTM module learns the time evolution law of temperature and vibration data, and outputs a health index of 0-100 (100 for the best state); when the health index is <60, the system automatically triggers a maintenance warning and pushes the warning information to the PC / APP terminal to prompt the maintenance personnel to repair the equipment and prevent the fault from expanding.

[0046] As shown in Figure 2 The AI self-learning-based digital air compression station intelligent control system of the present application: The system corresponds to the above control method, comprising a data acquisition layer, an AI prediction layer, an optimization control layer, an execution layer and an auxiliary function module, and each layer cooperates to realize intelligent control of the air compression station. The specific structure is as follows: (1) Data acquisition layer Containing a three-level high-precision sensor network and an edge computing node. The sensor network includes: Pressure sensor sub-network: piezoelectric pressure sensors are arranged at the main pipe, branch pipe and end gas point. The model is selected as B D SENSORS DMP 331 (accuracy 0.1% FS, range 0-1.6 MPa), the sampling frequency is set to 10 Hz, and the data is transmitted through 4-20 mA analog signal; Flow sensor sub-network: ABB FMT 3000 thermal mass flowmeter (measurement range 0-100 m 3 / min, error <1.0%) is used in the main pipe, and Yokogawa AXF vortex flowmeter (measurement range 0-50 m 3 / min, error <1.5%) is used in the branch pipe to realize layered monitoring of flow; Electricity sensor sub-network: LEM LA 55-P Hall effect current sensor (measurement range 0-500 A, accuracy ±0.5%) is installed at the input end of each air compressor, and voltage sensor is used to collect input power and power factor; Environment and equipment state sensor sub-network: Sensirion SHT31 temperature and humidity sensor (accuracy ±2% RH, ±0.3°C) and BMP280 atmospheric pressure sensor (accuracy ±1 hPa) are arranged in the air compression station room, and PCB352C33 vibration sensor (sensitivity 100 mV / g) and PT100 platinum resistance temperature sensor (measurement range -200-850°C, accuracy ±0.1°C) are arranged at the key parts of the air compressor.

[0047] The edge computing node uses NVIDIA Jetson AGX Orin development board, which carries Linux operating system, deploys TensorRT lightweight inference engine (FP16 precision), ensures that the real-time control period is less than or equal to 100 ms; runs the data cleaning, feature extraction, and multi-source data fusion algorithm written in Python, and simultaneously realizes communication interaction with the sensor and the AI prediction layer.

[0048] (2) AI prediction layer Deployed in edge computing nodes or cloud platforms (selected according to data volume and real-time requirements), including XGBoost module, LSTM module, weighted fusion layer and online learning module. The XGBoost module is implemented using the XGBoost open source library, with a tree depth of 8, a learning rate of 0.05, and estimators = 200; the LSTM module is built through the Keras framework, and the early stopping strategy (patience = 5) is used during training to avoid overfitting; the weighted fusion layer dynamically adjusts the weight according to the prediction period, and outputs the gas demand curve, confidence interval (95% confidence) and energy efficiency ratio prediction results; the online learning module includes model evaluation, sample screening and incremental training submodules, and automatically executes the model update process every day.

[0049] (3) Optimization control layer Including improved NSGA-II algorithm optimization engine and adaptive PID adjustment module. The improved NSGA-II algorithm is written in MATLAB, with a population size of 50, an iteration number of 100, a crossover probability of 0.8, and a mutation probability of 0.1; the adaptive PID adjustment module is written in C language to control the program, integrated into the real-time operating system of the edge computing node, collects pressure data every 10ms, dynamically adjusts PID parameters and outputs control instructions.

[0050] In addition, the optimization control layer also integrates a waste heat coordination module, including a waste heat data acquisition interface (communicates with the waste heat recovery system PLC) and a correlation model calculation unit, which incorporates waste heat utilization benefits into the optimization target.

[0051] (4) Execution layer Including frequency converter and air compressor cluster scheduling unit. The frequency converter uses HuiChuan MD500 series (adapted to 132kW motor, speed range 0-50Hz, control accuracy ±0.01Hz), supports Modbus RTU communication protocol, receives the speed instruction of the optimization control layer; the air compressor cluster scheduling unit is realized by PLC, according to the optimization operation combination scheme, outputs the relay control signal, realizes the start-stop and operation mode switching (variable frequency / working frequency) of the air compressor.

[0052] (5) Auxiliary function module Device health monitoring module: including vibration / temperature data acquisition submodule, CNN-LSTM health evaluation model (deployed in edge computing nodes), early warning push submodule (pushed to PC / APP terminal through MQTT protocol); Communication module: adopts multi-layer communication protocol, sensor and edge node communicate through Modbus RTU (RS485 bus), edge node and cloud platform communicate through MQTT+JSON protocol (4G / ethernet), cloud platform and user terminal communicate through REST API, data transmission adopts AES-256 encryption to ensure data security; Fault tolerance module: has the ability of failure prediction tolerance (switches to "flow-pressure" lookup table method, that is, according to real-time flow to query the preset pressure setting value table) and communication interruption tolerance (edge node runs the last optimization strategy independently until communication is restored); Remote monitoring module: supports PC end (Web interface) and APP end (Android / iOS) remote monitoring, real-time display of air compression station running parameters (pressure, flow, energy consumption, equipment health degree), historical data curve, fault warning information, remote start-stop control and parameter setting are supported.

[0053] In order to verify the actual application effect of the AI self-learning based digital air compression station intelligent control method and system, an implementation experiment was carried out in the air compression station of an automobile parts manufacturing enterprise, and the specific implementation process and results are as follows: (1) Experimental environment and equipment configuration The air compression station of the enterprise provides compressed air for the stamping, welding and assembly production lines in the factory area. The original control system adopts traditional manual + simple PID control, and there are problems such as high energy consumption, large pressure fluctuation and frequent equipment start-stop. Before the experiment, the air compression station was intelligently transformed, and the configuration is as follows: Air compressor equipment: 4 sets of 132kW screw air compressors, of which 2 sets are variable frequency air compressors (model Kaimeng JN 110-8-II-BP) and 2 sets are fixed frequency air compressors (model Kaimeng JN110-8-II), the rated discharge pressure of the air compressor is 0.8MPa, and the rated discharge capacity is 22m 3 / min (single); Pipe network system: total pipe diameter DN200, branch pipe diameter DN100-DN150, total pipe volume 85m 3 ; Data acquisition layer configuration: Pressure sensor: BDSensors DMP 331 piezoelectric pressure sensor is arranged at the total pipe (1), branch pipe (4, corresponding to 4 workshop main pipes), and end gas point (8, 2 key gas points are selected in each workshop), the accuracy is 0.1%FS, and the sampling frequency is 10Hz; Flow sensor: ABB FMT 3000 thermal mass flowmeter is arranged at the total pipe (measurement range 0-100m 3 / min); 4 branch pipes are respectively arranged with horizontal river AFX vortex flow meters (measurement range 0-50 m 3 / min); Power sensor: LEM LA 55-P Hall effect current sensor is additionally installed at the incoming line end of each air compressor, which is matched with a Siemens voltage sensor to collect input power and power factor; Environmental and equipment state sensor: Sensirion SHT31 temperature and humidity sensor and BMP280 atmospheric pressure sensor are arranged in the machine room; PCB 352C33 vibration sensor is arranged at the motor end and bearing end of each air compressor, and PT100 temperature sensor is arranged at the cylinder end; Edge computing node: NVIDIA Jetson AGX Orin is adopted, which is equipped with TensorRT 8.5 to deploy data preprocessing, AI prediction and optimization control algorithm, and the real-time control cycle is set to 80 ms; Execution layer configuration: original HuiChuan MD500 frequency converter (adapted to 132 kW motor) of the variable frequency air compressor and a relay control module (communicating with the PLC) additionally installed on the power frequency air compressor; Communication and monitoring: Modbus RTU (sensor-edge node), MQTT+JSON (edge node-Alibaba cloud platform) and RESTAPI (cloud platform-user terminal) communication are adopted, data transmission is encrypted by AES-256; Web monitoring interface is deployed on the PC side, Android / iOS application is developed on the APP side, and remote monitoring and early warning are supported.

[0054] The experimental period is divided into two stages: the first stage (1 month) adopts the traditional control mode, and the running data is recorded as a comparison benchmark; the second stage (3 months) enables the intelligent control system of the application, continuously collects running data and analyzes the effect. During the experiment, the production plan of the factory is basically stable, the daily production time is 16 hours, the average gas consumption is about 52 m 3 / min, and the environmental temperature changes in the range of 15-35℃ (covering the season replacement from late spring to early summer).

[0055] (II) Data acquisition and preprocessing implementation In the second stage of the experiment, the data acquisition layer realizes multi-dimensional data acquisition and preprocessing according to the preset scheme: Data acquisition: the pressure sensor collects the pressure data of the main pipe, branch pipe and end gas point in real time, such as the end gas point pressure of the stamping workshop, which fluctuates greatly during the production peak period (9:00-11:00), and the original data shows that the pressure value changes between 0.72-0.85 MPa; the flow sensor records the daily average flow of the main pipe 52.3 m 3 / min, of which the daily average flow of the stamping workshop is 18.5 m 3 / min (35.4%), welding workshop 15.2 m 3 / min (29.1%), assembly workshop 14.8 m 3 / min (28.3%), other auxiliary equipment 3.8 m 3 / min (7.2%); the power sensor collects the daily total power consumption of 4 air compressors 16820 kWh, including the daily power consumption of the frequency conversion machine 6230 kWh per unit and the daily power consumption of the power frequency machine 8560 kWh per unit; the environmental sensor records the daily average environmental temperature 25.6℃, the relative humidity 62%, and the atmospheric pressure 101.2kPa; the equipment state sensor collects the air compressor motor vibration acceleration between 0.2-0.8g and the bearing temperature between 65-85℃.

[0056] Data cleaning: the edge computing node runs the outlier_detection function (window 30) to detect outliers for the collected pressure and flow data. For example, at a certain time, the flow sensor of the welding workshop branch pipe appears an abnormal value of 120 m 3 / min (far beyond the normal range) due to electromagnetic interference. The system identifies this abnormal value by the 3σ principle and replaces it with the average value of the previous and next 10 normal data (48.5 m 3 / min), ensuring data continuity; after data cleaning, the abnormal rate of pressure data is reduced from 3.2% to 0.5%, and the abnormal rate of flow data is reduced from 2.8% to 0.4%.

[0057] Feature extraction: time domain features (5-minute sliding average flow, pressure fluctuation variance), frequency domain features (FFT extracts 0.1-1Hz frequency band energy), and working condition features (EER) are extracted. Taking one frequency conversion air compressor as an example, the 5-minute sliding average flow fluctuates between 45-55 m 3 / min, the pressure fluctuation variance is 0.002MPa² (0.008MPa² in traditional control), the 0.1-1Hz frequency band energy appears a peak value at the start and stop of production equipment (such as 10:00 puncher die change), and the average EER is 0.23 m 3 / (min·kW) (0.20 m 3 / (min·kW) in traditional control).

[0058] Multi-source data fusion correction: Kalman filter is used to fuse pressure and flow data. For example, due to zero drift, the total pipe pressure sensor measures 0.03MPa lower than the actual value. Through Kalman filter combined with flow data (pressure should remain constant when flow is stable), the pressure measurement value is corrected to the actual value in real time. The deviation of the corrected pressure data from the standard pressure gauge measurement value is less than 0.005MPa, ensuring that the data accuracy meets the subsequent prediction requirements.

[0059] (III) AI precise prediction implementation Deploy XGBoost and LSTM hybrid prediction model on edge computing node, and carry out 1-24 hour gas demand and energy efficiency ratio prediction experiment: Model training data preparation: Collect 1 month of historical data (720 hours in total) during the first phase of the experiment (traditional control), including hourly gas consumption, OEE, ambient temperature, EER, etc. Divide them into training set (504 hours) and validation set (216 hours) in the ratio of 7:3; At the same time, collect production plan data (such as order quantity, production shift), and convert them into OEE indicators (OEE=equipment running time x performance efficiency x pass rate / planned production time).

[0060] XGBoost module training: input features are lag 1h, 4h, 24h gas consumption, OEE, and cubic expansion of ambient temperature (T, T 2 , T 3 ). Optimize hyperparameters through grid search: tree depth 8, learning rate 0.05, estimators=200, subsample=0.8, colsample_bytree=0.8. After training, the XGBoost module has a gas consumption trend item prediction MAE of 4.2% on the validation set.

[0061] LSTM module training: input data is 24 time steps x 10 features (including hourly gas consumption, device start-stop times, vibration mean, etc.), and output is 1-hour periodic item prediction value. The network structure is built according to the preset, and batch_size=32, epochs=100, early stopping strategy patience=5 are used during training. After training, the LSTM module has a periodic item prediction MAE of 3.8% on the validation set.

[0062] Hybrid model fusion: short-term prediction (1-6 hours) LSTM weight 0.6, XGBoost weight 0.4, long-term prediction (7-24 hours) XGBoost weight 0.6, LSTM weight 0.4. After fusion, the hybrid model has a 1-24 hour gas consumption prediction MAE of 3.5% on the validation set, which is significantly lower than the prediction error of XGBoost alone (MAE 4.2%) and LSTM alone (MAE 4.0%).

[0063] Energy efficiency ratio prediction: based on the gas consumption prediction value output by the hybrid model and the historical EER data, an EER prediction sub-model is established to predict the real-time EER of each air compressor. During the experiment, the EER prediction MAE is 2.8%, and when the predicted EER of a certain frequency machine is <0.20 m 3 / (min·kW) (i.e. <75% average EER), the system marks it as an inefficient device and feeds back to the optimization control layer.

[0064] Online learning mechanism verification: The model MAE was automatically evaluated at 0 o'clock every day. The model MAE was stable between 3.5%-4.0% in the two weeks before the experiment, and did not need to be retrained. On the 15th day, the model MAE increased to 8.2% due to production plan adjustment (OEE decreased from 85% to 70%), triggering incremental training. During incremental training, the last 30 days of data (including the 5 days of production plan adjustment) and 20 key working condition samples (such as OEE sudden drop, gas peak) were selected for K-means clustering. The first 3 layers of XGBoost were fixed, and the top layer was fine-tuned. After that, the model MAE decreased to 3.8%, restoring the accurate prediction ability.

[0065] (Four) Optimization of control decision implementation Improved NSGA-II algorithm and adaptive PID adjustment module were deployed on the edge computing node to implement air compressor operation combination optimization and pressure control: Improved NSGA-II algorithm optimization experiment: Chromosome coding: 4 air compressor states (0 = stop, 1 = frequency conversion, 2 = power frequency) + total pipe pressure setting (0.7-0.8 MPa, accuracy 0.01 MPa), chromosome length is 5, such as [1,0,2,0,0.75] represents 1# frequency conversion, 2# stop, 3# power frequency operation, 4# stop, total pipe pressure 0.75 MPa.

[0066] Objective function calculation: taking total energy consumption , gas supply deviation , load rate balance as the target, the predicted gas consumption at a certain time during the experiment was 55m 3 / min, after 100 iterations of the algorithm, the Pareto optimal solution set was generated, and the optimal solution with minimum total power 245kW, <5% (gas supply 55.2m 3 / min), <0.1 (load rate 72%-78%) was selected: [1,1,2,0,0.74], i.e. 1# and 2# frequency conversion, 3# power frequency operation, 4# stop, total pipe pressure 0.74 MPa.

[0067] Optimization effect: compared with traditional control (randomly starting and stopping 2 frequency conversion and 1 power frequency, total power 278kW, gas supply 58.3m 3 / min, load rate 60%-85%), after optimization, total energy consumption decreased by 11.9%, gas supply deviation decreased by 5.7%, and load rate balance improved by 17.6%.

[0068] Adaptive PID adjustment experiment: Benchmark parameter determination: through Ziegler-Nichols method, under pure proportional control, Kp gradually increases to the critical proportional coefficient Kp = 2.5, the oscillation period T0 = 10 s, the benchmark parameters Kp0 = 1.5, Ti0 = 5 s, Td0 = 1.25 s are recorded. cr =2.5, the oscillation period T0 = 10 s, the benchmark parameters Kp0 = 1.5, Ti0 = 5 s, Td0 = 1.25 s are recorded.

[0069] Dynamic parameter adjustment: when the total pipe pressure fluctuation is ± 0.08 bar (stable working condition), Kp = 0.8 x 1.5 = 1.2, Ti = 1.2 x 5 = 6 s, after adjustment, the pressure fluctuation quickly converges to ± 0.05 bar; when the production load suddenly increases, resulting in a pressure fluctuation of -0.35 bar (large deviation working condition), Kp = 1.5 x 1.5 = 2.25, Ti = 0.7 x 5 = 3.5 s, the pressure is restored to the set value 0.74 MPa within 8 s, and the fluctuation is controlled within ± 0.1 bar.

[0070] Pressure control effect: during the experiment, the total pipe pressure fluctuation range is reduced from ± 0.2 bar of traditional control to ± 0.1 bar, a reduction of 50%, meeting the requirements of production equipment for pressure stability (allowable fluctuation ± 0.15 bar).

[0071] Waste heat-gas collaborative optimization experiment: Waste heat data collection: the waste heat recovery system uses a shell and tube heat exchanger, collects hot water temperature (85-95℃), hot water flow rate (5-8 m 3 / h) data, and establishes a waste heat output model: waste heat output = 0.7 x input power (kW) x running time (h) when the frequency converter is running, waste heat output = 0.6 x input power (kW) x running time (h).

[0072] Collaborative optimization: add the waste heat utilization benefit item (waste heat output x 0.8, 0.8 is the heating demand coefficient of the factory area) to the objective function of the improved NSGA-II algorithm, when the factory area needs heating, the frequency converter with high waste heat output is preferentially dispatched. During the experiment, the optimization running combination on a heating day is [0, 0, 2, 2, 0.76] (2 frequency converters running), the waste heat output reaches 320 kWh, meeting 30% of the heating demand of the factory area, compared with the non-collaborative optimization (waste heat output 240 kWh), the energy comprehensive utilization rate increases by 33.3%.

[0073] (Five) Implementation and feedback and equipment health monitoring Implementation and feedback: Frequency converter control: according to the load rate command output by the optimization control layer, the load rate of 1# frequency converter air compressor is 75%, the frequency converter adjusts the motor speed to 42 Hz (corresponding to gas production 16.5 m 3 / min); 2# VFD air compressor load rate 72%, speed 40Hz (gas production 15.8m 3 / min); 3# power frequency air compressor rated speed 50Hz (gas production 22m 3 / min), total gas production 54.3m 3 / min, deviating from the predicted demand of 55m 3 / min by 1.3%, accurately matching the gas demand. At the same time, the inverter starting current is controlled at 180A (rated current 220A), which is only 40% of the traditional star-delta starting current (450A), reducing the impact on the power grid.

[0074] Cluster scheduling: According to the rules of "frequency priority, high energy efficiency priority, and less running time priority", the average daily running time of the 4 air compressors during the experiment was 1# 10.2h, 2# 9.8h, 3# 8.5h, and 4# 8.1h, with a difference of less than 2.1h, avoiding the imbalance phenomenon of traditional control, where one air compressor runs for 14h and another for only 5h per day; the number of equipment start-stop times decreased from 127 times / day in traditional control to 41 times / day, a reduction of 67.7%, significantly reducing mechanical wear and tear.

[0075] Data feedback: The execution layer collects air compressor operation data (load rate, power, temperature, vibration) in real time, transmits it to the edge computing node every 10s, and feeds it back to the AI prediction layer after preprocessing, which is used for daily model evaluation. For example, at a certain time, the actual gas consumption was 54.8m 3 / min, the predicted value was 55m 3 / min, with a deviation of 0.4%, and the model MAE maintained between 3.5%-4.0%.

[0076] Equipment health monitoring: Data collection: Vibration sensors collect air compressor motor vibration acceleration in real time (sampling frequency 1kHz), and temperature sensors collect bearing temperature, storing data every 5 minutes; on the 45th day of the experiment, the bearing temperature of 3# power frequency machine increased from normal 75℃ to 88℃, and the vibration acceleration increased from 0.3g to 0.7g.

[0077] Health assessment: The CNN-LSTM model extracts the frequency domain features of vibration data (such as the appearance of a new harmonic peak at 200Hz) and temperature change trend, and calculates the health index from 85 to 58 (<60 warning threshold).

[0078] Early warning and maintenance: the system pushes early warning information through the APP: "3# power frequency machine bearing temperature abnormally high, vibration increases, health degree 58, suggest stop maintenance"; after receiving the early warning, the operation and maintenance personnel stop the machine and find that the bearing lubricating grease is aging, after replacing the lubricating grease, the bearing temperature returns to 72℃, the vibration acceleration decreases to 0.28g, and the health degree index rises to 82, avoiding the equipment failure caused by bearing burning (under traditional control, the bearing damage needs to be found, and the maintenance cost is more than 20,000 yuan, and the downtime is 8 hours).

[0079] (Six) Experimental results and analysis After the second stage (3 months) of the experiment, the key indicators of the traditional control and the intelligent control of the application were compared, and the results are shown in the following table:

[0080] From the above table, it can be seen that the intelligent control system of the application is significantly better than the traditional control in various key indicators: In terms of energy consumption, the unit energy consumption decreases from 0.104kWh / m 3 to 0.095kWh / m 3 , with an increase of 9.5%, according to the daily gas consumption of 832m 3 (16 hours x 52m 3 / min), and the electricity price of 0.6 yuan / kWh, the monthly electricity bill can be saved by 832x0.6x(0.104-0.095)x30≈1354 yuan, and the annual electricity bill is more than 16,000 yuan; In terms of running stability, the pressure fluctuation is reduced by 50%, avoiding the production equipment failure caused by unstable pressure (such as the unqualified rate of welding points in the welding workshop caused by pressure fluctuation, from 1.2% to 0.3%); In terms of equipment management, the start-stop frequency is reduced by 67.7%, the maintenance cost is reduced by 34.7%, the fault prediction accuracy is 92.3%, the equipment life is effectively prolonged, and the downtime loss is reduced; In terms of energy comprehensive utilization, through waste heat optimization, the energy comprehensive utilization rate is increased by 21.8%, further reducing the overall energy consumption of the enterprise.

[0081] At the same time, the system shows good self-adaptability during the experiment: when the environmental temperature rises from 15℃ to 35℃ (seasonal change), and the OEE decreases from 85% to 70% (production plan adjustment), the gas consumption prediction MAE is always stable at about 3.8%, and there is no significant increase; in the communication interruption test (manually disconnecting the edge node from the cloud platform for 2 hours), the edge node runs the latest optimization strategy independently, the pressure fluctuation is controlled within ±0.12bar, and the production is not affected, verifying the fault tolerance ability of the system.

[0082] In conclusion, the AI self-learning-based digital air compression station intelligent control method and system can realize the goals of energy consumption reduction, stability improvement, equipment protection and intelligent management in actual industrial applications, and has good practicability and popularization value.

[0083] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. An AI self-learning-based digital air compression station intelligent control method, characterized in that, Comprising the following steps: S1, data acquisition and preprocessing: Collecting pressure, flow, temperature, electric quantity data and environmental temperature and humidity, atmospheric pressure data of air compression station main pipe, branch pipe and terminal gas point through high-precision sensor network arranged in three stages, the precision of pressure sensor is 0.1%FS, the sampling frequency is ≥10Hz, the flow measurement error is <1.5%, and the electric quantity is collected by Hall effect current sensor; the edge computing node uses 3σ principle for data cleaning, extracting time domain features, frequency domain features and EER energy efficiency ratio working condition features; S2, AI accurate prediction: Constructing XGBoost and LSTM mixed prediction model, XGBoost inputting historical gas consumption, production plan OEE and three times expansion of environmental temperature trend term, LSTM network containing 64 neuron input layer, 32 neuron hidden layer and Dropout layer, predicting period term with Huber loss and Nadam optimizer, and obtaining 1-24 hour gas demand curve and confidence interval by weighted fusion; evaluating the model at 0 o'clock every day, and when MAE>8%, using near 30-day data+K-means clustering key samples to realize incremental training by fixing XGBoost top 3 layer weight fine-tuning; S3, optimization control decision: Using improved NSGA-II algorithm, chromosome coding containing air compressor state and main pipe pressure setting, solving optimal operation combination with minimum total energy consumption, minimum gas supply deviation and load rate close to 75% as objective function; combining adaptive PID, adjusting Kp to 0.8x and 1.5x of baseline value and Ti to 1.2x and 0.7x of baseline value according to pressure fluctuation ±0.1bar and ±0.3bar interval respectively, and determining baseline value by Ziegler-Nichols method; S4, execution and feedback: Adjusting the speed of frequency converter to 30-50Hz, scheduling air compressor cluster according to the rule that frequency converter bears 40-80% basic load, high energy efficiency power frequency machine is preferred, and equipment is preferred when running less, and feeding real-time operation data to AI prediction layer to update model.

2. The AI self-learning-based intelligent control method for a digitalized air compression station according to claim 1, characterized in that: The time domain features in S1 are 5-minute moving average flow rate and pressure fluctuation variance, and the frequency domain features are the energy in the 0.1-1 Hz band extracted by FFT. The EER energy efficiency ratio = actual gas production (m 3 / min) / input power (kW). 3.The AI self-learning-based intelligent control method of a digitalized air compression station according to claim 1, characterized in that: The mixed prediction model output in S2 also includes energy efficiency ratio prediction result, and air compressors with EER<75% are marked as low-efficiency equipment and fed back to optimization control layer for priority adjustment.

4. The AI self-learning-based digital air compression station intelligent control method according to claim 1, characterized in that: The constraint conditions of improved NSGA-II algorithm in S3 include total gas production ≥ predicted demand and single machine load rate ∈[50%, 100%].

5. The AI self-learning-based digital air compression station intelligent control method according to claim 1, characterized in that: It also includes equipment health monitoring step: collecting air compressor vibration and bearing temperature data, constructing CNN-LSTM-based health evaluation model, and calculating health index in real time, and triggering maintenance warning when the index is less than 60.

6. The AI self-learning-based digital air compression station intelligent control method according to claim 1, characterized in that: The data preprocessing in S1 also uses multi-source data fusion correction: fusing pressure and flow data by Kalman filter to correct measurement deviation caused by sensor drift and ensure data accuracy.

7. The AI self-learning-based intelligent control method for a digitalized air compression station according to claim 1, characterized in that: It also includes waste heat-gas collaborative optimization step: collecting hot water temperature and flow data of air compression station waste heat recovery system, establishing waste heat output and gas demand correlation model, and preferentially scheduling air compressors with high waste heat output when optimizing operation combination to improve overall energy utilization rate. 8.The AI self-learning-based digital air compression station intelligent control system of the control method according to any one of claims 1-7, characterized in that, Comprising: Data acquisition layer, containing three high-precision sensing network and edge computing nodes; AI prediction layer, deploying XGBoost module, LSTM module, weighted fusion layer and online learning module, realizing gas demand and energy efficiency ratio prediction; optimization control layer, containing improved NSGA-II algorithm optimization engine and adaptive PID adjustment module, outputting optimal operation combination and pressure set value; execution layer, including frequency converter and air compressor cluster scheduling unit, executing control instructions; also including equipment health monitoring module, multi-source data fusion module and waste heat coordination module. 9.The AI self-learning-based digital air compression station intelligent control system according to claim 8, characterized in that: Edge computing nodes use NVIDIA Jetson AGX Orin, deploy lightweight TensorRT model, real-time control cycle ≤100ms; communication uses Modbus RTU, MQTT+JSON, REST API, data transmission uses AES-256 encryption.

10. The AI self-learning-based digital air compression station intelligent control system according to claim 9, characterized in that: The system has fault tolerance capability: when the prediction fails, switch to "flow-pressure" table lookup method, when the communication is interrupted, the edge node runs the latest optimization strategy independently; Support PC / APP remote monitoring.

Images