Intelligent air compressor energy-saving control method and system

Abstract

The application discloses an intelligent air compressor energy-saving control method and system, and belongs to the technical field of energy-saving control. The method comprises the following steps: air compressor data acquisition, air compressor data preprocessing, air compressor energy consumption prediction, air compressor energy-saving strategy optimization and air compressor energy-saving control. The scheme is based on the calculation of adaptive fusion weights based on aggregated features, the generation of local working condition representation vectors, the introduction of a physical prior mechanism, the obtaining of local enhanced feature vectors, the obtaining of joint feature vectors based on dynamic gating coefficients, the generation of global fusion feature vectors based on time series attention weights, and the obtaining of the energy consumption prediction value of the air compressor based on a calibration coefficient vector. An energy consumption-stability compound reward item, a device health-compliance reward item and a gas use-regulation cost reward item are constructed, a multi-objective weighted reward function is obtained based on a load adaptive weight, an optimal energy-saving strategy of the air compressor is formed, and safe, efficient and accurate intelligent energy-saving control is realized.

Description

Technical Field

[0001] This invention belongs to the field of energy-saving control technology, specifically referring to an intelligent energy-saving control method and system for air compressors. Background Technology

[0002] Energy-saving control methods for air compressors utilize big data analysis and intelligent control algorithms to dynamically adjust operating strategies in real time, helping to achieve efficient and low-consumption operation of air compressors and improve energy utilization efficiency. However, existing energy-saving control methods for air compressors suffer from several problems. They fail to dynamically capture the complex correlations between multiple characteristics under air compressor operating conditions and ignore the inherent physical laws of air compressor operation, resulting in insufficient physical rationality of prediction results and an inability to accurately support energy-saving control decisions. Furthermore, existing energy-saving control methods face conflicts in multi-objective optimization, including energy consumption, equipment health, and air demand, making it difficult to balance multiple objectives and adapt to dynamic operating conditions. This leads to low efficiency in generating energy-saving strategies and difficulty in obtaining optimal energy-saving operating parameters, ultimately hindering the realization of intelligent energy-saving control for air compressors. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent energy-saving control method and system for air compressors. Addressing the problems of existing air compressor energy-saving control methods failing to dynamically capture the complex relationships between multiple features under air compressor operating conditions and neglecting the inherent physical laws of air compressor operation, resulting in insufficient physical rationality of prediction results and inability to accurately support air compressor energy-saving control decisions, this solution constructs an edge adjacency matrix and an in-edge adjacency matrix based on dynamic correlation strength. It calculates adaptive fusion weights based on aggregated features to obtain bidirectional fusion features. A physical prior mechanism is introduced to weight and fuse local operating condition representation vectors and physical prior feature vectors to obtain local enhanced feature vectors. A joint feature vector is obtained based on dynamic gating coefficients. A global fusion feature vector is generated based on temporal attention weights. Finally, a calibration coefficient is used to... This method obtains predicted energy consumption values ​​for air compressors, improving the accuracy of these predictions and enabling safe, efficient, and precise intelligent energy-saving control. Addressing the conflict between multiple objectives in existing air compressor energy-saving control methods—energy consumption, equipment health, and gas demand—this method fails to balance multiple objectives and adapt to dynamic operating conditions, resulting in low efficiency in generating energy-saving strategies and difficulty in obtaining optimal energy-saving operating parameters, thus hindering intelligent energy-saving control of air compressors. This solution constructs a composite reward item for energy consumption and stability, a reward item for equipment health and compliance, and a reward item for gas consumption and adjustment costs. Based on load adaptive weighting, a multi-objective weighted reward function is obtained. A reinforcement learning agent is trained using a near-end strategy optimization algorithm to form the optimal energy-saving strategy for the air compressor. This provides a precise basis for intelligent air compressor energy-saving control, significantly reducing air compressor energy consumption and improving energy utilization efficiency.

[0004] The technical solution adopted by this invention is as follows: This invention provides an intelligent energy-saving control method for air compressors, which includes the following steps:

[0005] Step S1: Air compressor data acquisition;

[0006] Step S2: Air compressor data preprocessing;

[0007] Step S3: Predict air compressor energy consumption;

[0008] Step S4: Optimize the energy-saving strategy for the air compressor;

[0009] Step S5: Air compressor energy-saving control.

[0010] Further, in step S1, the air compressor data acquisition involves collecting historical air compressor time-series data; the historical air compressor time-series data includes timestamps, air compressor operating data, air compressor air consumption data, environmental operating condition data, and energy consumption values, with the energy consumption values ​​used as data tags.

[0011] Furthermore, in step S2, the air compressor data preprocessing involves cleaning, encoding, and normalizing the historical air compressor time-series data. After aligning the preprocessed data with a uniform time step, it is divided into a training dataset and a test dataset.

[0012] Further, in step S3, the air compressor energy consumption prediction is based on a deep neural network to construct an air compressor energy consumption prediction model to predict the energy consumption value of the air compressor; specifically, it includes the following steps:

[0013] Step S31: Construction of dynamic feature association graph; For the feature sequence of each time step in the training dataset, construct a dynamic feature association graph. Each node in the graph corresponds to a feature dimension. Based on the association weight and combined with the load rate of the real-time working conditions, calculate the dynamic association strength between nodes and construct the edge adjacency matrix and the in edge adjacency matrix respectively.

[0014] Step S32: Bidirectional feature fusion; Based on the outgoing edge adjacency matrix and the incoming edge adjacency matrix, calculate the corresponding degree matrix and perform symmetric normalization, extract the outgoing edge aggregation feature and the incoming edge aggregation feature, calculate the adaptive fusion weight based on the L1 norm mean of the aggregation feature, and perform weighted fusion to obtain the bidirectional fusion feature. Then, combine the outgoing edge adjacency matrix and the incoming edge adjacency matrix to extract the interaction feature, perform global average pooling on the interaction feature, and generate a local working condition representation vector.

[0015] Step S33: Physical Prior Enhancement; Calculate the load rate, specific power and environmental correction coefficient of the air compressor operation, and normalize them respectively. Introduce the physical prior mechanism, encode it through a lightweight fully connected layer to obtain the physical prior feature vector, calculate the dynamic fusion weight, and fuse the local operating condition representation vector and the physical prior feature vector with weight to obtain the local enhanced feature vector after physical prior enhancement.

[0016] Step S34: Global temporal feature extraction; Input the local enhanced feature vectors of all time steps into a bidirectional gated recurrent unit to extract temporal dependent features, and obtain the global temporal feature vector through a linear projection layer;

[0017] Step S35: Global fusion feature generation; concatenate the local enhanced feature vector with the global temporal feature vector, generate dynamic gating coefficients through a gating mechanism, and weightedly fuse to obtain the joint feature vector of each time step. Normalize the temporal attention score between the joint feature vector and the global temporal feature vector to obtain the temporal attention weight of each time step. Weightedly sum the joint feature vectors of all time steps to generate the global fusion feature vector.

[0018] Step S36: Feature calibration; Based on the average load rate, specific power and environmental correction coefficient of the entire time series, generate a calibration coefficient vector, perform dimension-by-dimensional calibration on the global fused feature vector, and output the predicted energy consumption value of the air compressor through the regression layer.

[0019] Further, in step S4, the air compressor energy-saving strategy optimization is based on reinforcement learning to construct an air compressor energy-saving strategy optimization model, optimize the adjustable operating parameters of the air compressor, and generate the optimal energy-saving strategy for the air compressor; specifically, it includes the following steps:

[0020] Step S41: State and action space setting; define the air compressor operating data, air compressor air consumption data, and environmental condition data together as the state space, forming a state vector; define the set of values ​​of the air compressor's adjustable operating parameters within its feasible range as the action space, forming an action vector.

[0021] Step S42: Construct a multi-objective weighted reward function; integrate energy consumption saving rate and energy consumption stability to construct an energy consumption-stability composite reward item; integrate operating parameter compliance and equipment health status to construct an equipment health-compliance reward item; integrate gas demand satisfaction and adjustment cost to construct a gas consumption-adjustment cost reward item; based on the current load rate, recent average load rate, and load rate change, generate a load adaptive weight vector, and perform weighted fusion of the three reward items to construct a multi-objective weighted reward function;

[0022] Step S43: Optimal energy-saving strategy generation; The reinforcement learning agent is trained based on the near-end policy optimization algorithm. The policy network takes the state vector as input and outputs the probability distribution of each adjustable operating parameter in the action vector. The policy network parameters are iteratively optimized by maximizing the expected cumulative discount reward to obtain the trained air compressor energy-saving strategy optimization model. The parameters at this time are used as the optimal policy network parameters. The current state vector is input into the policy network trained based on the optimal policy network parameters, and the optimal energy-saving operating parameters of the air compressor corresponding to the current state are output to form the optimal energy-saving strategy of the air compressor.

[0023] Further, in step S5, the air compressor energy-saving control involves collecting real-time air compressor time-series data, which includes timestamps, air compressor operating data, air compressor air consumption data, and environmental condition data. After data cleaning, encoding, and normalization, the real-time air compressor time-series data is input into the trained air compressor energy consumption prediction model and air compressor energy-saving strategy optimization model to obtain the optimal energy-saving strategy for the air compressor and send it to the air compressor controller to achieve intelligent energy-saving control of the air compressor.

[0024] The present invention provides an intelligent energy-saving control system for air compressors, comprising an air compressor data acquisition module, an air compressor data preprocessing module, an air compressor energy consumption prediction module, an air compressor energy-saving strategy optimization module, and an air compressor energy-saving control module;

[0025] The air compressor data acquisition module collects historical air compressor time-series data and sends the data to the air compressor data preprocessing module;

[0026] The air compressor data preprocessing module performs data cleaning, data encoding and data normalization to construct training datasets and test datasets, and sends the data to the air compressor energy consumption prediction module.

[0027] The air compressor energy consumption prediction module constructs an edge adjacency matrix and an in-edge adjacency matrix based on dynamic correlation strength, calculates adaptive fusion weights based on aggregated features to obtain bidirectional fusion features, generates a local operating condition representation vector, introduces a physical prior mechanism, and weights and fuses the local operating condition representation vector and the physical prior feature vector to obtain a local enhanced feature vector, obtains a joint feature vector based on dynamic gating coefficients, generates a global fusion feature vector based on temporal attention weights, obtains the predicted energy consumption value of the air compressor based on the calibration coefficient vector, and sends the data to the air compressor energy-saving strategy optimization module.

[0028] The air compressor energy-saving strategy optimization module constructs an energy consumption-stability composite reward item, an equipment health-compliance reward item, and an air consumption-adjustment cost reward item. Based on the load adaptive weight, it obtains a multi-objective weighted reward function, trains a reinforcement learning agent based on the near-end strategy optimization algorithm, forms the optimal energy-saving strategy for the air compressor, and sends the data to the air compressor energy-saving control module.

[0029] The air compressor energy-saving control module collects real-time air compressor timing data, obtains the corresponding optimal energy-saving strategy for the air compressor, and performs intelligent energy-saving control of the air compressor.

[0030] The beneficial effects achieved by the present invention using the above solution are as follows:

[0031] (1) In view of the problems that existing air compressor energy-saving control methods cannot dynamically capture the complex correlation between multiple features under the air compressor operating conditions and ignore the inherent physical laws of air compressor operation, resulting in insufficient physical rationality of prediction results and inability to accurately support air compressor energy-saving control decisions, this scheme constructs an edge adjacency matrix and an in-edge adjacency matrix based on dynamic correlation strength, so that feature correlation is closely combined with the real-time operating conditions of air compressor; based on aggregated feature calculation, adaptive fusion weight is calculated to obtain bidirectional fusion features and generate local operating condition representation vectors to improve the representation ability of local features; a physical prior mechanism is introduced to weight and fuse the local operating condition representation vector and the physical prior feature vector to obtain a local enhanced feature vector, so that the feature representation fits the actual physical laws of air compressor energy consumption and improves the physical credibility of local features; a joint feature vector is obtained based on dynamic gating coefficients, and a global fusion feature vector is generated based on temporal attention weights to improve the representation ability of global features; the predicted energy consumption value of air compressor is obtained based on calibration coefficient vector to improve the accuracy of energy consumption prediction value, thereby realizing safe, efficient and accurate intelligent energy-saving control.

[0032] (2) In view of the existing air compressor energy-saving control methods, there is a conflict between multiple objectives of energy consumption, equipment health and gas demand. It is impossible to take into account multiple objectives and adapt to dynamic working conditions, resulting in low efficiency of energy-saving strategy generation and difficulty in obtaining the optimal energy-saving operating parameters, thus failing to realize intelligent energy-saving control of air compressors. This solution constructs an energy consumption-stability composite reward item, an equipment health-compliance reward item and a gas consumption-adjustment cost reward item. Based on the load adaptive weight, a multi-objective weighted reward function is obtained, which is adapted to the energy-saving core, equipment safety and gas supply guarantee requirements, respectively, to realize multi-objective collaborative consideration and adapt to changes in working conditions. Based on the near-end strategy optimization algorithm, a reinforcement learning agent is trained to form the optimal energy-saving strategy of air compressor, which provides accurate basis for intelligent air compressor energy-saving control, significantly reduces the energy consumption of air compressors and improves energy utilization efficiency. Attached Figure Description

[0033] Figure 1A flowchart illustrating an intelligent air compressor energy-saving control method provided by the present invention;

[0034] Figure 2 A schematic diagram of an intelligent air compressor energy-saving control system provided by the present invention;

[0035] Figure 3 This is a flowchart illustrating step S3;

[0036] Figure 4 This is a flowchart illustrating step S4.

[0037] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0038] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0039] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0040] Example 1, see Figure 1 The present invention provides an intelligent energy-saving control method for air compressors, which includes the following steps:

[0041] Step S1: Air compressor data acquisition; collect historical air compressor time-series data;

[0042] Step S2: Air compressor data preprocessing; perform data cleaning, data encoding and data normalization to construct training datasets and test datasets;

[0043] Step S3: Air compressor energy consumption prediction; Based on dynamic correlation strength, construct edge adjacency matrix and ingress edge adjacency matrix, calculate adaptive fusion weight based on aggregated features to obtain bidirectional fusion features, generate local operating condition representation vector, introduce physical prior mechanism, weighted fusion of local operating condition representation vector and physical prior feature vector to obtain local enhanced feature vector, obtain joint feature vector based on dynamic gating coefficient, generate global fusion feature vector based on temporal attention weight, and obtain the predicted energy consumption value of air compressor based on calibration coefficient vector;

[0044] Step S4: Optimize the energy-saving strategy of the air compressor; construct a composite reward item for energy consumption and stability, a reward item for equipment health and compliance, and a reward item for gas consumption and adjustment costs. Obtain a multi-objective weighted reward function based on load adaptive weights. Train a reinforcement learning agent based on the near-end strategy optimization algorithm to form the optimal energy-saving strategy for the air compressor.

[0045] Step S5: Air compressor energy-saving control; collect real-time air compressor timing data, obtain the corresponding optimal energy-saving strategy for the air compressor, and perform intelligent energy-saving control of the air compressor.

[0046] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the air compressor data acquisition is to collect historical air compressor time series data.

[0047] The historical air compressor time series data includes timestamps, air compressor operation data, air compressor air consumption data, environmental operating condition data, and energy consumption values, with energy consumption values ​​used as data tags;

[0048] The air compressor operating data includes exhaust pressure, exhaust temperature, intake pressure, intake temperature, motor speed, stator temperature, motor current, motor voltage, motor power, rated motor power, oil-gas separator pressure difference, cooling medium flow rate, cooling medium inlet and outlet temperatures, running time, intake valve opening, number of start-stop cycles, and inverter output frequency.

[0049] The air compressor's air consumption data includes air consumption and air pressure.

[0050] The environmental operating condition data includes ambient temperature, ambient humidity, and atmospheric pressure.

[0051] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S2, the air compressor data preprocessing is to perform data cleaning, data encoding and data normalization on the historical air compressor time series data. After aligning the preprocessed data according to a uniform time step, it is divided into training dataset and test dataset in a 7:3 ratio.

[0052] The data cleaning process involves identifying and processing invalid data, filling missing values ​​with the mean, and detecting and processing outliers using the 3σ principle.

[0053] The data encoding uses One-Hot encoding to convert categorical data into numerical data.

[0054] The data normalization method uses the max-min scaling method to unify numerical data to the range [0, 1].

[0055] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S3, the air compressor energy consumption prediction is based on a deep neural network to construct an air compressor energy consumption prediction model to predict the energy consumption value of the air compressor; specifically, it includes the following steps:

[0056] Step S31: Construction of Dynamic Feature Association Graph; The association relationships of various feature dimensions of the air compressor change dynamically with the load rate of the operating condition. Static association analysis cannot match the real-time operating condition requirements of energy-saving control. For the feature sequence of each time step in the training dataset, a dynamic feature association graph is constructed. Each node in the graph corresponds to a feature dimension. Based on the association weight and combined with the load rate of the real-time operating condition, the dynamic association strength between nodes is calculated, and the edge adjacency matrix and the incoming edge adjacency matrix are constructed respectively. This allows the feature association relationship to be deeply bound to the real-time operating condition of the air compressor, improving the targeting of information extraction of energy-saving related features. The formulas used are as follows:

[0057] ;

[0058] ;

[0059] ;

[0060] In the formula, and These represent the dynamic association strengths between node i and node j at time step t, and between node j and node i, respectively. Here, λ is the eigenvalue corresponding to node i at time step t, and λ is the correlation decay coefficient. , These are learnable self-loop weight coefficients. , It is the learnable association weight between node i and node j. W max It is the largest learnable effective association weight. , It is the operating condition correction factor for time step t. , , and These are the load rate and motor power at time step t, and RLF and These are the rated load rate and rated motor power, respectively, RLF=1. and These are the outgoing edge adjacency value of node i to node j and the incoming edge adjacency value of node j to node i, respectively, at time step t. and These represent the dynamic association strengths of node i to node k and node k to node i at time step t, respectively. t is the time step index, i, j, and k are node indices, and N is the total number of nodes. It is a smoothing term. ;

[0061] Step S32: Bidirectional Feature Fusion; unidirectional extraction of air compressor features results in information omissions and a lack of targeted representation of local operating conditions, affecting the accuracy of energy consumption prediction and thus restricting energy-saving control decisions. Based on the outgoing and incoming adjacency matrices, the corresponding degree matrices are calculated and symmetrically normalized. A two-layer graph convolutional network aggregates the outgoing and incoming edge association information of each node, extracting outgoing and incoming edge aggregation features. Adaptive fusion weights are calculated based on the L1 norm mean of the aggregated features, and weighted fusion yields bidirectional fused features. Interaction features are then extracted by combining the outgoing and incoming adjacency matrices, and global average pooling is performed on these interaction features to generate a local operating condition representation vector. This accurately captures the real-time local operating condition features required for energy-saving control, providing a precise local feature foundation for subsequent physical prior enhancement. The formulas used are as follows:

[0062] ;

[0063] ;

[0064] ;

[0065] In the formula, It is the adaptive fusion weight of the outgoing edge aggregation features at time step t. and These are the outgoing edge aggregation features and incoming edge aggregation features obtained by GCN aggregation at time step t, respectively. and They are and The mean of the L1 norm of node features, and These are the bidirectional fusion features and interaction features at time step t, respectively. and These are the outgoing edge adjacency matrix and the incoming edge adjacency matrix at time step t, respectively. It is the ReLU activation function. and These are the learnable outbound edge weight matrix and the bias vector, respectively. and These are the learnable incoming edge weight matrix and the bias vector, respectively.

[0066] Step S33: Physical Prior Enhancement; Pure data-driven feature extraction ignores the physical mechanism of air compressor energy saving, leading to a disconnect between feature representation and actual air compressor energy consumption patterns, resulting in insufficient physical rationality of prediction results; The load rate, specific power, and environmental correction coefficient of the air compressor are calculated and normalized respectively. A physical prior mechanism is introduced, encoded through a lightweight fully connected layer to obtain a physical prior feature vector. Dynamic fusion weights are calculated, and the local operating condition representation vector and the physical prior feature vector are weighted and fused to obtain a locally enhanced feature vector after physical prior enhancement; This ensures that the feature representation conforms to the actual physical laws of air compressor energy consumption, improving the physical credibility and accuracy of energy consumption prediction in energy-saving control scenarios; the formula used is as follows:

[0067] ;

[0068] ;

[0069] ;

[0070] In the formula, , and They are , and The normalization result, and These are the specific power at time step t and the environmental correction factor, respectively. Q std This is the displacement under standard operating conditions. , and These are the ambient temperature and atmospheric pressure at time step t, respectively, and c. T c P c0 and c0 are the learnable temperature weights, pressure weights, and basic correction bias, respectively. , , , W is the encoded physical prior feature vector at time step t. phy and b phy These are the learnable encoding weight matrix and the bias vector, respectively. It is the dynamic fusion weight at time step t. and These are the learnable, dynamically fused weight matrix and the bias vector, respectively. It's a vector concatenation operation. It is the Sigmoid activation function, W fusion and b fusion These are the learnable feature fusion weight matrix and the bias vector, respectively. It is the local operating condition representation vector at time step t. It is the locally enhanced feature vector at time step t;

[0071] Step S34: Global Time Feature Extraction; Air compressor energy consumption has strong time-series dependence. Local enhancement features only depict the operating conditions of a single time step and cannot capture the time-series dependence pattern of the entire time series. The local enhancement feature vectors of all time steps are input into a bidirectional gated recurrent unit to extract time-series dependence features. The global time feature vector is obtained through a linear projection layer. The bidirectional GRU simultaneously captures the historical and future time-series dependence of air compressor operation, adapting to the time-series variation characteristics of energy consumption in energy-saving control. The formulas used are as follows:

[0072] ;

[0073] In the formula, It is the last time step T max The hidden state vector of the feedforward GRU. B is the hidden state vector of the backward GRU at time step 1. global It is a global time feature vector. and These are the learnable global temporal feature projection weight matrix and the bias vector, respectively.

[0074] Step S35: Global Fusion Feature Generation; The contribution of features at each time step to global energy consumption prediction is not differentiated, resulting in inefficient feature information fusion and weakening of time step features for key energy-saving operating conditions; The local enhanced feature vector is concatenated with the global time feature vector, and dynamic gating coefficients are generated through a gating mechanism. Weighted fusion is then performed to obtain the joint feature vector for each time step. The temporal attention score between the joint feature vector and the global time feature vector is normalized to obtain the temporal attention weight for each time step. The joint feature vectors of all time steps are then weighted and summed to generate the global fusion feature vector; This adapts to the feature information requirements of different energy-saving operating conditions of the air compressor, strengthens the information weight of core operating conditions in energy-saving control, and improves the representation ability of global features and the accuracy of energy consumption prediction; The formula used is as follows:

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] In the formula, , , and These are the dynamic gating coefficients at time step t, the joint feature vector, the temporal attention score, and the temporal attention weights, W. gate and b gate These are the learnable gated weight matrix and bias vector, W. att It is a learnable temporal attention weight matrix. , It is a time step index. It is a time step Temporal attention score, F final It is a globally fused feature vector. It is a transpose operation;

[0080] Step S36: Feature Calibration; The global fusion feature suffers from an imbalance in dimensional feature weights and lacks calibration based on the average physical prior parameters of the entire air compressor sequence, failing to accurately support energy-saving control decisions. Based on the average load rate, specific power, and environmental correction coefficients of the entire time series, a calibration coefficient vector is generated. The global fusion feature vector is then calibrated dimensionally. Using the test dataset, the Adam optimizer is used to minimize the mean squared error loss function, iteratively updating all learnable parameters in the model to obtain the trained air compressor energy consumption prediction model. The predicted energy consumption value of the air compressor is output through a regression layer. This integrates feature calibration with the overall operating patterns of the air compressor, improving the accuracy of feature representation. The formulas used are as follows:

[0081] ;

[0082] ;

[0083] In the formula, δ is the calibration coefficient vector. , and These are the average values ​​of load factor, specific power, and environmental correction factor within the time series, respectively, W. cal and b cal These are the learnable mapping weight matrix and bias vector, U. cal It is a learnable physical prior-aided calibration weight matrix, W reg and b reg These are the learnable regression head weight matrix and the bias vector, respectively. This is the predicted energy consumption value of the air compressor.

[0084] By performing the above operations, this solution addresses the problems of existing air compressor energy-saving control methods, which fail to dynamically capture the complex relationships between multiple features under air compressor operating conditions and ignore the inherent physical laws of air compressor operation, resulting in insufficient physical rationality of prediction results and inability to accurately support air compressor energy-saving control decisions. Based on dynamic correlation strength, this solution constructs edge adjacency matrices and ingress adjacency matrices, closely integrating feature correlation with the real-time operating conditions of the air compressor. Based on aggregated feature calculation, adaptive fusion weights are calculated to obtain bidirectional fusion features, generating local operating condition representation vectors and improving the representational ability of local features. A physical prior mechanism is introduced to weightedly fuse the local operating condition representation vectors and physical prior feature vectors to obtain locally enhanced feature vectors, ensuring that feature representations align with the actual physical laws of air compressor energy consumption and improving the physical reliability of local features. Based on dynamic gating coefficients, a joint feature vector is obtained, and based on temporal attention weights, a global fusion feature vector is generated, improving the representational ability of global features. Based on calibration coefficient vectors, the predicted energy consumption value of the air compressor is obtained, improving the accuracy of the predicted energy consumption value, thereby achieving safe, efficient, and accurate intelligent energy-saving control.

[0085] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S4, the air compressor energy-saving strategy optimization is based on reinforcement learning to construct an air compressor energy-saving strategy optimization model, optimize the adjustable operating parameters of the air compressor, and generate the optimal energy-saving strategy for the air compressor; specifically, it includes the following steps:

[0086] Step S41: State-action space setting; define the air compressor operating data, air compressor air consumption data, and environmental condition data together as the state space, forming a state vector. The set of adjustable operating parameters of the air compressor within its feasible range is defined as the action space, which is then used to construct the action vector. Among them, g t and m t These are the state vector and action vector corresponding to the t-th time step, respectively, Q. t C t and E t These are the air compressor operation data vector, air compressor air consumption data vector, and environmental condition data vector at time step t, respectively. t q t v t n t and r tThese are the exhaust pressure, cooling medium flow rate, intake valve opening, number of start / stop cycles, and inverter output frequency at the t-th time step. The feasible ranges of the air compressor's adjustable operating parameters before normalization are: exhaust pressure 0.6~1.2Mpa, cooling medium flow rate 5~50m³ / h, intake valve opening 0~100%, number of start / stop cycles 0~5 times / hour, and inverter output frequency 20~50Hz.

[0087] Step S42: Construction of a multi-objective weighted reward function; Air compressor energy-saving control involves multiple conflicting objectives: energy consumption, equipment health, and gas demand. A single reward function cannot simultaneously address multiple objectives, and fixed reward weights cannot adapt to dynamic operating conditions. This step integrates energy consumption efficiency and energy consumption stability to construct a composite reward item; it also integrates operational parameter compliance and equipment health status to construct an equipment health-compliance reward item; and it integrates gas demand satisfaction and adjustment costs to construct a gas consumption-adjustment cost reward item. Based on the current load rate, recent average load rate, and load rate change, a load adaptive weight vector is generated. The three reward items are then weighted and integrated to construct the multi-objective weighted reward function. The energy consumption-stability item balances energy saving rate and energy consumption stability to avoid operational fluctuations caused by energy saving; the equipment health-compliance item ensures equipment safety and parameter compliance during energy-saving operations; the gas consumption-adjustment cost item balances gas demand and parameter adjustment costs; and the load adaptive weight allows the priority of multiple objectives to dynamically change with the air compressor load rate, perfectly adapting to the dynamic operating conditions of energy-saving control and achieving optimal synergy among multiple objectives. The formulas used are as follows:

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] In the formula, , and These are the energy consumption-stability composite reward item, the equipment health-compliance reward item, and the gas consumption-regulation cost reward item at the t-th time step, respectively. and These are the hyperbolic tangent activation function and the softmax activation function, respectively. It is the historical average energy consumption value under the same operating conditions. This is the energy consumption value of the air compressor at time step t. and These are the average and standard deviation of energy consumption over the past 5 time steps, respectively. and These are the values ​​of the w-th adjustable parameter at the t-th and t-1-th time steps, respectively. , and These are the baseline value, maximum allowable value, and minimum allowable value of the w-th adjustable parameter, respectively. , , , It is an indicator function. and These are the stator temperature and exhaust temperature at the t-th time step, respectively. It is the temperature safety threshold. , It is the gas pressure at time step t. and These are the minimum pressure and the target gas pressure, respectively. , ω t It is the load adaptive weight vector at time step t. , and They are , and The weights, W ω and b ω These are the learnable weight matrix and the bias vector, respectively. It is the average load factor over the past 5 time steps. D is the change in load factor at time step t. t It is the multi-objective weighted reward value at the t-th time step;

[0094] Step S43: Optimal energy-saving strategy generation; A reinforcement learning agent is trained based on a proximal policy optimization algorithm. The policy network takes the state vector as input and outputs the probability distribution of each adjustable operating parameter in the action vector. The value network evaluates the long-term value of the state. The policy network parameters are iteratively optimized by maximizing the expected cumulative discount reward. When the cumulative discount reward changes less than a threshold of 10⁻ in five consecutive iterations on the test dataset... 4 Training is complete when the preset maximum number of iterations (1000) is reached, resulting in a trained air compressor energy-saving strategy optimization model. The parameters at this point are used as the optimal strategy network parameters. The current state vector is input into the strategy network trained based on these optimal parameters, and the optimal energy-saving operating parameters for the air compressor corresponding to the current state are output, forming the optimal energy-saving strategy for the air compressor. The formula used is as follows:

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] in, It is the expected cumulative discount reward function value. These are the learnable parameters of the policy network, and G is the index of the future time step. It is a discount factor. , It is the multi-objective weighted reward value at the (t+G)th time step. It is the complete trajectory generated by the interaction between the reinforcement learning agent and the environment. and These are the policy network parameters before and after the update. and These are the strategies and Sampling trajectory Find the expected value. and These are the strategies used. and In g t Choose m t The probability value, Z t It is the dominant function value. , Yes The estimate, It is the value network for g t Output state value estimate, It is a clipping function. It is the cutting factor. , These are the optimal policy network parameters. These are the optimal energy-saving operating parameters for the air compressor at time step t. , and They are respectively composed of parameters , and Defined policy network.

[0100] By performing the above operations, this solution addresses the problem in existing air compressor energy-saving control methods where there are conflicts between multiple objectives—energy consumption, equipment health, and gas demand—making it impossible to balance multiple objectives and adapt to dynamic operating conditions. This results in low efficiency in generating energy-saving strategies and difficulty in obtaining optimal energy-saving operating parameters, ultimately hindering intelligent energy-saving control of air compressors. This solution constructs a composite reward item for energy consumption and stability, a reward item for equipment health and compliance, and a reward item for gas consumption and adjustment costs. Based on load adaptive weighting, a multi-objective weighted reward function is obtained, adapting to the core energy-saving requirements, equipment safety, and gas supply guarantee needs, respectively. This achieves multi-objective collaborative consideration and adapts to changes in operating conditions. Furthermore, a reinforcement learning agent is trained based on a near-end strategy optimization algorithm to form the optimal energy-saving strategy for the air compressor, providing accurate data for intelligent air compressor energy-saving control, significantly reducing air compressor energy consumption, and improving energy utilization efficiency.

[0101] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the air compressor energy-saving control involves collecting real-time air compressor time-series data. The real-time air compressor time-series data includes timestamps, air compressor operating data, air compressor air consumption data, and environmental operating condition data. After data cleaning, data encoding, and data normalization, the real-time air compressor time-series data is input into the trained air compressor energy consumption prediction model and air compressor energy-saving strategy optimization model to obtain the optimal energy-saving strategy for the air compressor and send it to the air compressor controller to realize intelligent energy-saving control of the air compressor.

[0102] Example 7, see Figure 2 Based on the above embodiments, this embodiment provides an intelligent air compressor energy-saving control system, including an air compressor data acquisition module, an air compressor data preprocessing module, an air compressor energy consumption prediction module, an air compressor energy-saving strategy optimization module, and an air compressor energy-saving control module.

[0103] The air compressor data acquisition module collects historical air compressor time-series data and sends the data to the air compressor data preprocessing module;

[0104] The air compressor data preprocessing module performs data cleaning, data encoding and data normalization to construct training datasets and test datasets, and sends the data to the air compressor energy consumption prediction module.

[0105] The air compressor energy consumption prediction module constructs an edge adjacency matrix and an in-edge adjacency matrix based on dynamic correlation strength, calculates adaptive fusion weights based on aggregated features to obtain bidirectional fusion features, generates a local operating condition representation vector, introduces a physical prior mechanism, and weights and fuses the local operating condition representation vector and the physical prior feature vector to obtain a local enhanced feature vector, obtains a joint feature vector based on dynamic gating coefficients, generates a global fusion feature vector based on temporal attention weights, obtains the predicted energy consumption value of the air compressor based on the calibration coefficient vector, and sends the data to the air compressor energy-saving strategy optimization module.

[0106] The air compressor energy-saving strategy optimization module constructs an energy consumption-stability composite reward item, an equipment health-compliance reward item, and an air consumption-adjustment cost reward item. Based on the load adaptive weight, it obtains a multi-objective weighted reward function, trains a reinforcement learning agent based on the near-end strategy optimization algorithm, forms the optimal energy-saving strategy for the air compressor, and sends the data to the air compressor energy-saving control module.

[0107] The air compressor energy-saving control module collects real-time air compressor timing data, obtains the corresponding optimal energy-saving strategy for the air compressor, and performs intelligent energy-saving control of the air compressor.

[0108] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0109] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0110] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent energy-saving control method for air compressors, characterized in that: The method includes the following steps: Step S1: Air compressor data acquisition; collect historical air compressor time-series data; Step S2: Air compressor data preprocessing; perform data cleaning, data encoding and data normalization to construct training datasets and test datasets; Step S3: Air compressor energy consumption prediction; Based on dynamic correlation strength, construct edge adjacency matrix and ingress edge adjacency matrix, calculate adaptive fusion weight based on aggregated features to obtain bidirectional fusion features, generate local operating condition representation vector, introduce physical prior mechanism, weighted fusion of local operating condition representation vector and physical prior feature vector to obtain local enhanced feature vector, obtain joint feature vector based on dynamic gating coefficient, generate global fusion feature vector based on temporal attention weight, and obtain the predicted energy consumption value of air compressor based on calibration coefficient vector; Step S4: Optimize the energy-saving strategy of the air compressor; construct a composite reward item for energy consumption and stability, a reward item for equipment health and compliance, and a reward item for gas consumption and adjustment costs. Obtain a multi-objective weighted reward function based on load adaptive weights. Train a reinforcement learning agent based on the near-end strategy optimization algorithm to form the optimal energy-saving strategy for the air compressor. Step S5: Air compressor energy-saving control; collect real-time air compressor timing data, obtain the corresponding optimal energy-saving strategy for the air compressor, and perform intelligent energy-saving control of the air compressor.

2. The intelligent air compressor energy-saving control method according to claim 1, characterized in that: In step S3, the air compressor energy consumption prediction is based on a deep neural network to construct an air compressor energy consumption prediction model and predict the energy consumption value of the air compressor; specifically, it includes the following steps: Step S31: Dynamic feature association graph construction; For the feature sequence of each time step in the training dataset, construct a dynamic feature association graph. Based on the association weight and combined with the load rate of the real-time working conditions, calculate the dynamic association strength between nodes, and construct the edge adjacency matrix and the in edge adjacency matrix respectively. Step S32: Bidirectional feature fusion; Based on the outgoing edge adjacency matrix and the incoming edge adjacency matrix, calculate the corresponding degree matrix and perform symmetric normalization, extract outgoing edge aggregation features and incoming edge aggregation features, calculate adaptive fusion weights based on the L1 norm mean of the aggregation features, and perform weighted fusion to obtain bidirectional fusion features. Then, combine the outgoing edge adjacency matrix and the incoming edge adjacency matrix to extract interaction features, perform global average pooling on the interaction features, and generate local working condition representation vectors. Step S33: Physical prior reinforcement; Step S34: Global temporal feature extraction; Input the local enhanced feature vectors of all time steps into a bidirectional gated recurrent unit to extract temporal dependent features, and obtain the global temporal feature vector through a linear projection layer; Step S35: Global fusion feature generation; Step S36: Feature calibration; Based on the average load rate, specific power and environmental correction coefficient of the entire time series, generate a calibration coefficient vector, perform dimension-by-dimensional calibration on the global fused feature vector, and output the predicted energy consumption value of the air compressor through the regression layer.

3. The intelligent air compressor energy-saving control method according to claim 2, characterized in that: In step S33, the physical prior enhancement involves calculating the load rate, specific power, and environmental correction coefficient of the air compressor, normalizing them, introducing a physical prior mechanism, encoding them through a lightweight fully connected layer to obtain a physical prior feature vector, calculating dynamic fusion weights, and weightedly fusing the local operating condition representation vector and the physical prior feature vector to obtain a locally enhanced feature vector after physical prior enhancement.

4. The intelligent air compressor energy-saving control method according to claim 2, characterized in that: In step S35, the global fusion feature generation involves concatenating the local enhanced feature vector with the global temporal feature vector, generating dynamic gating coefficients through a gating mechanism, weighted fusion to obtain the joint feature vector for each time step, normalizing the temporal attention score between the joint feature vector and the global temporal feature vector to obtain the temporal attention weight for each time step, and weighted summing of the joint feature vectors for all time steps to generate the global fusion feature vector.

5. The intelligent air compressor energy-saving control method according to claim 1, characterized in that: In step S4, the air compressor energy-saving strategy optimization is based on reinforcement learning to construct an air compressor energy-saving strategy optimization model, optimize the adjustable operating parameters of the air compressor, and generate the optimal energy-saving strategy for the air compressor. Specifically, the following steps are included: Step S41: State and action space setting; define the air compressor operating data, air compressor air consumption data, and environmental condition data together as the state space, forming a state vector; define the set of values ​​of the air compressor's adjustable operating parameters within its feasible range as the action space, forming an action vector. Step S42: Constructing the multi-objective weighted reward function; Step S43: Optimal energy-saving strategy generation; Based on the near-end policy optimization algorithm, a reinforcement learning agent is trained. The policy network parameters are iteratively optimized by maximizing the expected cumulative discount reward to obtain the trained air compressor energy-saving strategy optimization model. The parameters at this time are used as the optimal policy network parameters. The current state vector is input into the policy network trained based on the optimal policy network parameters, and the optimal energy-saving operation parameters of the air compressor corresponding to the current state are output to form the optimal energy-saving strategy of the air compressor.

6. The intelligent air compressor energy-saving control method according to claim 5, characterized in that: In step S42, the construction of the multi-objective weighted reward function is to integrate energy consumption saving rate and energy consumption stability to construct an energy consumption-stability composite reward item; to integrate operating parameter compliance and equipment health status to construct an equipment health-compliance reward item; and to integrate gas demand satisfaction and adjustment cost to construct a gas consumption-adjustment cost reward item. Based on the current load rate, recent average load rate, and load rate change, a load adaptive weight vector is generated. The three reward items are then weighted and fused to construct a multi-objective weighted reward function.

7. The intelligent air compressor energy-saving control method according to claim 1, characterized in that: In step S1, the air compressor data acquisition involves collecting historical air compressor time-series data. The historical air compressor time-series data includes timestamps, air compressor operating data, air compressor air consumption data, environmental operating condition data, and energy consumption values, with the energy consumption values ​​used as data tags.

8. The intelligent air compressor energy-saving control method according to claim 1, characterized in that: In step S5, the air compressor energy-saving control involves collecting real-time air compressor time-series data, which includes timestamps, air compressor operating data, air compressor air consumption data, and environmental condition data. After data cleaning, encoding, and normalization, the real-time air compressor time-series data is input into the trained air compressor energy consumption prediction model and air compressor energy-saving strategy optimization model to obtain the optimal energy-saving strategy for the air compressor and send it to the air compressor controller to achieve intelligent energy-saving control of the air compressor.

9. An intelligent air compressor energy-saving control system, used to implement the intelligent air compressor energy-saving control method as described in any one of claims 1-8, characterized in that: It includes an air compressor data acquisition module, an air compressor data preprocessing module, an air compressor energy consumption prediction module, an air compressor energy-saving strategy optimization module, and an air compressor energy-saving control module; The air compressor data acquisition module collects historical air compressor time-series data and sends the data to the air compressor data preprocessing module; The air compressor data preprocessing module performs data cleaning, data encoding and data normalization to construct training datasets and test datasets, and sends the data to the air compressor energy consumption prediction module. The air compressor energy consumption prediction module constructs an edge adjacency matrix and an in-edge adjacency matrix based on dynamic correlation strength, calculates adaptive fusion weights based on aggregated features to obtain bidirectional fusion features, generates a local operating condition representation vector, introduces a physical prior mechanism, and weights and fuses the local operating condition representation vector and the physical prior feature vector to obtain a local enhanced feature vector, obtains a joint feature vector based on dynamic gating coefficients, generates a global fusion feature vector based on temporal attention weights, obtains the predicted energy consumption value of the air compressor based on the calibration coefficient vector, and sends the data to the air compressor energy-saving strategy optimization module. The air compressor energy-saving strategy optimization module constructs an energy consumption-stability composite reward item, an equipment health-compliance reward item, and an air consumption-adjustment cost reward item. Based on the load adaptive weight, it obtains a multi-objective weighted reward function, trains a reinforcement learning agent based on the near-end strategy optimization algorithm, forms the optimal energy-saving strategy for the air compressor, and sends the data to the air compressor energy-saving control module. The air compressor energy-saving control module collects real-time air compressor timing data, obtains the corresponding optimal energy-saving strategy for the air compressor, and performs intelligent energy-saving control of the air compressor.

Images