An industrial energy consumption control method and system

Abstract

The application provides an industrial energy consumption control method and system, wherein the method comprises: collecting energy consumption data through the data acquisition device; sequentially performing high-frequency noise removal and pulse noise removal on the energy consumption data to obtain denoised energy consumption data; extracting time domain features, frequency domain features and industrial mechanism features from the denoised energy consumption data to form feature data; inputting the feature data into a lightweight model deployed in an edge gateway to output future energy consumption prediction data; when the prediction deviation of the future energy consumption prediction data exceeds an energy consumption threshold, automatically generating a device control strategy and controlling the edge executor to execute the device control strategy to dynamically adjust device parameters, so that the removal of data noise, the optimization of model training and deployment, and the real-time and automation of device control can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial energy management, and in particular to an industrial energy consumption control method and system. BACKGROUND

[0002] In industrial production processes, accurate monitoring and effective control of energy consumption data are crucial. With the continuous rise of energy costs and the increasing strictness of environmental protection requirements, how to achieve energy saving and emission reduction has become a key issue in industrial production. Accurate energy consumption monitoring can help enterprises timely identify energy waste links and equipment, while effective energy consumption control strategies can optimize production processes and reduce unnecessary energy consumption, thereby improving the economic benefits and market competitiveness of enterprises.

[0003] In the prior art, the industrial energy consumption monitoring and control system has the following problems:

[0004] Data noise problem: The energy consumption data collected by sensors in industrial environments is easily disturbed by high-frequency noise and impulse noise. High-frequency noise mainly comes from electromagnetic interference, causing data to fluctuate dramatically; impulse noise produces sharp pulses due to sensor faults or external transient disturbances. Traditional denoising methods, such as wavelet transform and median filtering, have limited effectiveness in complex industrial scenarios and cannot meet the needs of high-precision energy consumption monitoring.

[0005] Energy consumption prediction model training and deployment problem: In order to improve the accuracy of energy consumption prediction, existing prediction models often use complex network structures. Such models require a large amount of computing resources and time during training. The usual approach is to first train the model on the cloud and then directly deploy it to the edge device. However, due to the poor adaptability of the model and the limitations of the hardware performance of the edge device, the prediction accuracy and real-time performance are insufficient, and the edge computing resources cannot be fully utilized.

[0006] Timeliness problem of device regulation: Current industrial energy consumption control mainly relies on fixed rules or manual experience, resulting in a lag in device regulation. When energy consumption is abnormal, the system cannot respond quickly, causing energy waste. In high-energy-consuming industries such as steel production, sudden changes in device failure or working conditions can cause energy consumption to rise sharply, and traditional control methods cannot adjust device parameters in a timely manner, affecting production efficiency and cost control. SUMMARY

[0007] The present application provides an industrial energy consumption control method and system to solve the data noise problem, energy consumption prediction model training and deployment problem, and timeliness problem of device regulation in the prior art.

[0008] The application provides an industrial energy consumption control method, an edge controller for an industrial system, the edge controller being connected with a data acquisition device and an actuator respectively, and the method comprises the following steps of:

[0009] acquiring energy consumption data through the data acquisition device;

[0010] sequentially removing high-frequency noise and pulse noise from the energy consumption data to obtain denoised energy consumption data;

[0011] extracting time domain features, frequency domain features and industrial mechanism features from the denoised energy consumption data to form feature data;

[0012] inputting the feature data into a lightweight model deployed on an edge gateway to output future energy consumption prediction data; wherein the lightweight model is a teacher network model trained on a cloud through historical energy consumption data and real labels to generate first energy consumption prediction data, and the first energy consumption prediction data and the real labels are issued to the edge gateway to train a student network model to obtain the lightweight model;

[0013] when a prediction deviation of the future energy consumption prediction data exceeds an energy consumption threshold, automatically generating a device control strategy and controlling the edge actuator to execute the device control strategy to dynamically adjust device parameters.

[0014] The application also provides an industrial energy consumption control system, an edge controller for an industrial system, the edge controller being connected with a data acquisition device and an actuator respectively, and the system comprising the following steps of:

[0015] a data acquisition module for acquiring energy consumption data through the data acquisition device;

[0016] a noise removal module for sequentially removing high-frequency noise and pulse noise from the energy consumption data to obtain denoised energy consumption data;

[0017] a feature extraction module for extracting time domain features, frequency domain features and industrial mechanism features from the denoised energy consumption data to form feature data;

[0018] an energy consumption data prediction module for inputting the feature data into a lightweight model deployed on an edge gateway to output future energy consumption prediction data; wherein the lightweight model is a teacher network model trained on a cloud through historical energy consumption data and real labels to generate first energy consumption prediction data, and the first energy consumption prediction data and the real labels are issued to the edge gateway to train a student network model to obtain the lightweight model;

[0019] The control strategy generation module is configured to automatically generate a device control strategy when the prediction deviation of the future energy consumption prediction data exceeds an energy consumption threshold, and control the edge executor to execute the device control strategy to dynamically adjust the device parameters.

[0020] The present application also provides a computer program product comprising a computer program which, when executed by a processor, implements the industrial energy consumption control method according to any one of the above.

[0021] The industrial energy consumption control method and system provided by the present application can achieve the following technical effects:

[0022] Data noise removal: The energy consumption data is collected by the data acquisition device, and the high-frequency noise and impulse noise are removed to obtain the de-noised energy consumption data, which provides a high-quality data basis for subsequent feature extraction and energy consumption prediction. This data processing procedure can effectively improve the accuracy and reliability of the data, ensuring the precision of the energy consumption prediction.

[0023] Model training and deployment optimization: The first energy consumption prediction data is generated by training the teacher network model on the cloud, and the student network model deployed on the edge gateway is trained in combination with the real labels to obtain a lightweight model. This method utilizes the powerful computing power of the cloud for model training, and uses the knowledge distillation technology to transfer the knowledge of the complex teacher model to the lightweight student model, so that it can run efficiently on the edge device, improving the prediction accuracy and real-time performance of the model.

[0024] Real-time and automatic device control: When the prediction deviation of the future energy consumption prediction value exceeds the energy consumption threshold, a device control strategy is automatically generated, and the edge executor is controlled to execute the device control strategy to dynamically adjust the device parameters. This real-time monitoring and automatic control mechanism can respond to energy consumption changes in a timely manner, avoid energy waste, improve energy utilization efficiency, and reduce production costs. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0026] Figure 1 is a schematic diagram of the distributed hardware architecture for implementing the industrial energy consumption control method provided by the present application.

[0027] Figure 2 is a flowchart of the industrial energy consumption control method provided by the present application.

[0028] Figure 3 is a schematic diagram of the real-time closed-loop control logic provided by the present application.

[0029] Figure 4 is a structural schematic diagram of the industrial energy consumption control system provided by the present application.

[0030] Figure 5 is a structural schematic diagram of the electronic device provided by the present application. DETAILED DESCRIPTION

[0031] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in conjunction with the drawings in the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0032] Firstly, the terms involved in the embodiments of the present application are explained schematically.

[0033] Edge gateway: The edge gateway is a key device deployed in the edge controller, responsible for communication and data transmission between the data acquisition device and the actuator. It usually has data processing capability, can perform preliminary processing and analysis on the acquired energy consumption data, and execute the energy consumption prediction model in the cloud or locally to perform energy consumption prediction.

[0034] Edge controller: The edge controller is a hardware device installed in the industrial site, directly connected with the data acquisition device and the actuator. It is responsible for coordinating and controlling the operation of the entire industrial energy consumption control system, including data acquisition, noise removal, feature extraction, energy consumption prediction and device regulation, etc.

[0035] Data acquisition device: The data acquisition device is a sensor or instrument used to collect various energy consumption data in industrial production processes, such as current / voltage sensors, smart meters, etc. These devices can collect energy consumption data at a set frequency and transmit them to the edge controller.

[0036] Actuator: The actuator is a device that adjusts the parameters of industrial equipment according to the instructions of the edge controller, such as relays, electric valves, programmable logic controllers (PLC), etc. They can dynamically adjust the operating parameters of the equipment according to the device regulation strategy.

[0037] Wavelet transform: Wavelet transform is a mathematical tool used for signal processing and analysis. It can effectively remove high-frequency noise while preserving the main features of the signal by decomposing the signal into different frequency bands. In the present application, wavelet transform is used for high-frequency noise removal of energy consumption data.

[0038] Adaptive median filtering: Adaptive median filtering is a filtering method that dynamically adjusts the size of the filtering window to remove impulse noise. It can automatically adjust the size of the filtering window according to the noise intensity to achieve the best filtering effect while avoiding excessive smoothing of the signal's main features.

[0039] L1 norm: L1 norm refers to the sum of the absolute values of all elements in a vector. In this invention, L1 norm is used to calculate the sum of the absolute values of the weight matrix in the neural network, so as to evaluate the importance of the weight in the pruning process. The smaller the L1 norm, the smaller the contribution of the weight to the model, which can be considered for pruning first.

[0040] Time domain features: Time domain features refer to statistical features extracted from time series data, such as mean, variance, root mean square, peak value, kurtosis, etc. These features can reflect the statistical characteristics and trend of signals on the time axis.

[0041] Frequency domain features: Frequency domain features refer to features extracted from the frequency spectrum analysis of signals, such as frequency components, power spectral density, dominant frequency, frequency band energy, etc. These features can reflect the energy distribution and frequency characteristics of signals in different frequency bands.

[0042] Industrial mechanism features: Industrial mechanism features are features extracted based on the physical, chemical and engineering principles of industrial production processes, such as equipment operating parameters, production process indicators, energy balance parameters, etc. These features can reflect the internal laws and energy consumption characteristics of industrial production processes.

[0043] Teacher network model: The teacher network model is a complex and high-performance neural network model, usually trained in the cloud. It generates the first energy consumption prediction data by learning historical energy consumption data and real labels, which is used to guide the training of the student network model.

[0044] In order to achieve high-precision prediction, the teacher network model usually adopts a more complex network structure, such as a deep neural network containing multiple hidden layers, a long short-term memory network (LSTM) or a gated recurrent unit (GRU) recurrent neural network structure. These complex structures enable it to capture long-term dependencies and complex nonlinear features in time series data. Taking the LSTM structure as an example, it can effectively remember the key information in historical energy consumption data through the cooperative action of cell state, input gate, forget gate and output gate, and fully utilize these information in the prediction process, thereby improving the accuracy of prediction.

[0045] Student network model: The student network model is a lightweight neural network model deployed on the edge gateway. It is trained by the first energy consumption prediction data generated by the teacher network model and the real label through the knowledge distillation technology to obtain efficient energy consumption prediction ability. In order to adapt to the limited computing resources and storage capacity of edge devices, the student network model usually adopts lightweight network structure such as simplified convolutional neural network, GRU, etc. These structures reduce the parameter quantity and calculation quantity of the model under the premise of ensuring certain prediction performance, so that they can run efficiently on resource-limited devices such as edge gateways.

[0046] GRU layer: GRU layer is the abbreviation of Gated Recurrent Unit layer, which is a neural network layer used to process sequence data. It can effectively capture the time dependence of time series data through the mechanism of update gate and reset gate.

[0047] Fully connected layer: The fully connected layer is a layer in the neural network, in which each neuron is connected to all neurons of the previous layer. The fully connected layer is usually used to combine and classify high-level features to output the final prediction result.

[0048] Secondly, the industrial energy consumption control method of the embodiment of the application runs on a distributed hardware architecture, as shown in Figure 1 The architecture specifically includes the following key components:

[0049] Edge controller: The edge controller is the core control unit of the industrial field, directly connected with various data acquisition devices and actuators. It undertakes the preliminary processing of field data, the issuance of instructions and the coordination between devices. It has the ability of real-time data processing, which can preprocess the collected energy consumption data in a very short time, and start the subsequent energy consumption prediction and device control process according to the method of the application.

[0050] Data acquisition device: Data acquisition devices are widely distributed in various aspects of industrial production to acquire energy consumption data in real time. These devices are diverse, including current / voltage sensors, smart meters, etc. They continuously collect data at a high sampling rate (10Hz-100Hz) set by the user and transmit the data to the edge controller. For example, in a large steel plant, a current sensor installed at the inlet cable of the blast furnace can monitor the power consumption of the blast furnace in real time.

[0051] Executer: The executer is the final execution unit of the edge controller instruction, which adjusts the operating parameters of the industrial equipment accurately according to the received control strategy. For example, adjusting the frequency converter of the motor, controlling the opening of the valve, etc. When the energy consumption prediction shows an abnormality and exceeds the preset threshold, the control instruction sent by the edge controller is converted into specific physical action by the executer, realizing the dynamic adjustment of the equipment. For example, in chemical production, the heating power of the reaction kettle is adjusted in time according to the energy consumption prediction result.

[0052] Edge gateway: The edge gateway is located between the edge controller and the cloud, and plays the role of data transfer and preliminary analysis. On the one hand, it receives the processed data from the edge controller and performs necessary format conversion and aggregation on the data; on the other hand, it deploys a lightweight model to perform preliminary energy consumption prediction analysis on the data, extracts key prediction results, and provides them to the edge controller for further decision-making. At the same time, it is also responsible for forwarding the training results of the teacher network model and other information from the cloud to the edge controller.

[0053] Cloud server: The cloud server undertakes the heavy task of complex model training and data storage, and has a large amount of computing resources and storage capacity. The cloud server collects historical energy consumption data and label information from multiple edge gateways, trains a high-performance teacher network model, and sends the key output (such as the first energy consumption prediction data) and model parameters of the model to the edge gateway to support local lightweight model updating and optimization.

[0054] Each component is connected through a high-speed and reliable industrial network to ensure real-time transmission of data and instructions. For example, in the welding workshop of an automobile manufacturing factory, the energy consumption data of the welding robot is transmitted in real time to the edge controller through the data acquisition device, the edge controller processes the data quickly, and sends the data to the edge gateway, the edge gateway uses the lightweight model to perform preliminary energy consumption evaluation, and feeds back the evaluation results to the edge controller, the edge controller decides whether to adjust the working parameters of the welding robot according to the evaluation results, if necessary, sends the control instruction to the welding robot through the executer. At the same time, the edge gateway uploads the data to the cloud server after aggregation, providing data support for further training of the teacher network model.

[0055] In the foregoing, the terms related to the present invention have been explained in detail, and the hardware architecture for running the method has been analyzed comprehensively. The following content will start from this point and go deep into the specific details of the industrial energy consumption control method of the embodiment of the present invention, by combining Figures 2 to 3 , gradually disassembling each link of the method of the present invention, to clearly present how to realize accurate control and efficient management of industrial energy consumption in actual operation.

[0056] Figure 2is one of the flowcharts of the industrial energy consumption control method provided by the present application, as shown in Figure 2 The method comprises the following steps:

[0057] Step 201: collecting energy consumption data by the data acquisition device.

[0058] In step 201, various data acquisition devices are needed to collect energy consumption data in the industrial production process. This process is the starting point of the entire energy consumption control method, and its accuracy and completeness are directly related to the effect of subsequent processing. These data acquisition devices mainly include current / voltage sensors, smart meters, etc. They are precisely installed at key positions in the factory, such as the incoming cable of the main equipment, and the power input end of each distribution cabinet and main equipment. These sensors will collect data at a high sampling rate of 10Hz to 100Hz, so as to capture the subtle changes in energy consumption data and provide the possibility of timely discovering abnormal energy consumption. In order to ensure the consistency and accuracy of the data, all devices use a unified time synchronization mechanism to avoid data confusion caused by time differences and ensure that data collected from different devices can be analyzed and processed in the same time coordinate system.

[0059] Before the data is transmitted to the edge controller, the data acquisition device will also perform preliminary preprocessing on the raw data. This step is crucial because it can effectively improve data quality. The preprocessing process includes range conversion and filtering operations to remove obvious noise and interference. For example, the signal collected by the current sensor may be affected by electromagnetic interference and produce burrs. Simple filtering processing can preliminarily smooth these burrs and improve the reliability of the data. The preprocessed data is then transmitted in real time to the edge controller through an industrial network (such as Ethernet, fieldbus, etc.). In the data transmission process, in order to ensure the integrity and accuracy of the data, appropriate data checking and error correction mechanisms are adopted. For example, CRC checking and other methods are used to check the transmitted data. Once data errors are found, retransmission is performed immediately to ensure that the data received by the edge controller is consistent with the data sent by the data acquisition device. This process lays a solid data foundation for subsequent noise removal, feature extraction, and energy consumption prediction steps, ensuring the efficient operation of the entire energy consumption control system.

[0060] Step 202: sequentially removing high-frequency noise and pulse noise from the energy consumption data to obtain denoised energy consumption data.

[0061] First, wavelet transform is used to remove high-frequency noise from the energy consumption data. Wavelet transform can effectively distinguish the frequency bands of signals and noise by decomposing the signal into different frequency bands. Specifically, the energy consumption data can be decomposed into low-frequency approximation coefficients and multiple high-frequency detail coefficients using high-pass and low-pass filters. Then, adaptive threshold processing is applied to the high-frequency detail coefficients to remove high-frequency noise. The threshold processing method dynamically adjusts the threshold according to the noise intensity to ensure that the main features of the signal are preserved while removing noise. After threshold processing, the high-frequency detail coefficients are combined with the low-frequency approximation coefficients to perform inverse wavelet transform, resulting in intermediate energy consumption data with high-frequency noise removed.

[0062] Next, adaptive median filtering is used to remove impulse noise from the intermediate energy consumption data. Adaptive median filtering can automatically adjust the size of the filtering window according to the noise intensity. The noise intensity index of the intermediate energy consumption data is determined by a pre-set noise intensity evaluation model. According to the noise intensity index, the filtering window size is dynamically adjusted. For example, if the noise intensity index is less than 30%, the filtering window size is 3x3; if the noise intensity index is between 30% and 50%, the filtering window size is 5x5; if the noise intensity index is greater than 50%, the filtering window size is 7x7. Through this adaptive approach, impulse noise can be effectively removed while avoiding excessive smoothing of the main features of the signal.

[0063] Through the above two steps of processing, the final denoised energy consumption data is obtained. This process not only improves the signal-to-noise ratio of the data, but also provides a high-quality data foundation for subsequent feature extraction and energy consumption prediction, ensuring the accuracy and reliability of the entire energy consumption control system.

[0064] Step 203, extracting time domain features, frequency domain features and industrial mechanism features from the denoised energy consumption data to form feature data.

[0065] In step 203, feature extraction is performed on the denoised energy consumption data to form feature data for energy consumption prediction. This process aims to extract representative and discriminative features from the processed energy consumption data to provide high-quality input for the subsequent energy consumption prediction model. Feature extraction mainly includes the following aspects:

[0066] 1) Time domain feature extraction. Time domain features reflect the statistical characteristics and trends of energy consumption data on the time axis. The following time domain features are extracted: Mean: reflects the average level of energy consumption data; Variance: measures the fluctuation of energy consumption data; Root Mean Square (RMS): reflects the effective value of energy consumption data; Peak: records the maximum and minimum values of energy consumption data, used to identify extreme energy consumption situations; Kurtosis: measures the degree of peak of energy consumption data distribution.

[0067] 2) Frequency domain feature extraction. Frequency domain features reflect the energy distribution and frequency characteristics of energy consumption data in different frequency bands. This embodiment uses Fast Fourier Transform (FFT) to convert time domain signals to frequency domain signals and extracts the following frequency domain features: dominant frequency: the frequency component with the highest energy in the energy consumption data, determined by finding the frequency corresponding to the maximum value in the frequency spectrum; power spectral density (PSD): measures the energy distribution of different frequency components; band energy: calculates the total energy in a specific frequency band to analyze the energy consumption contribution of different frequency bands.

[0068] 3) Industrial mechanism feature extraction. Industrial mechanism features are based on the physical, chemical and engineering principles of industrial production processes and can reflect the internal causes of energy consumption. The following industrial mechanism features are extracted: equipment operating parameters such as motor speed, pump flow, furnace temperature, etc., which are directly related to the energy consumption of equipment; production process indicators such as molten iron temperature in steel production, conversion rate of chemical reactions, clinker output in cement production, etc., which reflect the energy consumption efficiency of production processes; energy balance parameters such as the ratio of input power to output power of equipment, thermal efficiency of the system, etc., used to evaluate the energy utilization efficiency.

[0069] The extracted time domain features, frequency domain features and industrial mechanism features are integrated to form a feature dataset. Each sample in the feature dataset contains multiple feature dimensions to describe the multi-aspect characteristics of energy consumption data. These feature data will be used as input to train and run the lightweight model to achieve accurate prediction and control of industrial energy consumption. By considering time domain, frequency domain and industrial mechanism features comprehensively, the characteristics of energy consumption data can be fully captured, improving the accuracy and reliability of energy consumption prediction.

[0070] Step 204, input the feature data into the lightweight model deployed by the edge gateway, and output future energy consumption prediction data.

[0071] Wherein, the lightweight model is a teacher network model trained by historical energy consumption data and real labels in the cloud, generating first energy consumption prediction data, and the first energy consumption prediction data and real labels are issued to the student network model for training to obtain.

[0072] In step 204, the feature data extracted in step 203 is input into the lightweight model deployed by the edge gateway. This model is specially designed to efficiently run on resource-constrained edge devices while providing accurate energy consumption prediction. The construction and training process of the lightweight model is as follows:

[0073] 1) Preprocessing and input of feature data. Before inputting the feature data into the lightweight model, the data is first standardized. This step allows the model to train and predict more stably by scaling the data to a specific range (usually 0 to 1 or -1 to 1). For example, using the Z-score standardization method, the formula is:

[0074] x normalized = (x—μ) / σ

[0075] where x is the feature value, μ is the mean of the feature, and σ is the standard deviation of the feature. Standardized feature data can eliminate the dimensional differences between different features, improve the convergence speed of the model and the prediction accuracy.

[0076] Feature data is sampled with a fixed time window to form the model input sequence. For example, the past 10 time steps of feature data are selected as an input sample to capture the time series characteristics of energy consumption changes. The setting of this time window can ensure that the model can fully utilize historical feature information when predicting, improving the accuracy of the prediction.

[0077] 2) Structure and training of lightweight model. The lightweight model mainly consists of an input layer, a GRU layer, a fully connected layer, and an output layer. The input layer receives preprocessed feature data; the GRU layer contains multiple neurons that process sequence data and capture temporal dependencies; the fully connected layer combines and classifies high-level features; and the output layer finally outputs future energy consumption prediction data. For example, a typical lightweight model structure includes an input layer, a GRU layer containing 64 neurons, a fully connected layer containing 32 neurons, and an output layer.

[0078] Cloud teacher network model training: In the cloud, a large amount of historical energy consumption data and real labels are used to train the teacher network model. The teacher network model structure is complex, such as the LSTM + XGBoost integrated model, which has strong feature learning and prediction ability. During training, hyperparameter optimization methods such as grid search are used to adjust the model's parameters to minimize prediction error. For example, after multiple iterations of optimization on the training set, the root mean square error (RMSE) on the validation set can be reduced to less than 0.05, indicating that the model has high prediction accuracy.

[0079] Knowledge distillation trains student network model: the first energy consumption prediction data generated by the teacher network model and the real label are issued to the edge gateway. The student network model is trained on the edge gateway based on these data. During the training process, a hybrid loss function is used, which includes not only the KL divergence between the first energy consumption prediction data and the second energy consumption prediction data output by the student network model, but also the mean square error between the second energy consumption prediction data and the real label. For example, the calculation formula of the hybrid loss function is:

[0080] L = a x L KL + (1 - a) x L MSE

[0081] Where a is the weight coefficient, usually set to 0.7, L KL is the KL divergence, L MSE is the mean square error. In this way, the student network model not only learns the information of the real label, but also inherits the generalization ability of the teacher network model, so as to realize efficient and accurate energy consumption prediction on edge devices.

[0082] 3) Optimization and verification of the model. During the training process of the lightweight model, a series of optimization techniques are used to further improve the model performance. For example, L1 regularization method is used to delete unimportant weights, reduce the parameter quantity and calculation amount of the model, and avoid overfitting. In addition, the floating point parameters are converted to INT8 format for quantization optimization, reducing the storage and calculation requirements of the model. After these optimization measures, the size of the lightweight model can be reduced from 4.2MB to 1.1MB, and the inference speed is improved by 3 times, meeting the resource constraints and real-time requirements of edge devices.

[0083] In the model verification stage, an independent test set is used to evaluate the lightweight model. The test set includes energy consumption data under various working conditions to ensure the generalization ability of the model. In the steel production scene, the prediction accuracy of the lightweight model on the test set can reach root mean square error (RMSE) 0.06 and mean absolute error (MAE) 0.03, indicating that the model can accurately predict the energy consumption trend. At the same time, by comparing the change trend of the actual energy consumption value and the predicted energy consumption value, the recognition ability of the model to energy consumption mutation is verified. For example, in a certain test case, when the device fails and the energy consumption rises sharply, the lightweight model can detect this change in time and trigger the warning mechanism, providing timely decision support for device regulation.

[0084] Through the above steps, the extracted feature data is input into the lightweight model to realize accurate prediction of future energy consumption. This process fully utilizes the powerful computing capacity of the cloud and the real-time processing capacity of the edge device, and transfers the knowledge of the complex teacher network model to the lightweight student network model through the knowledge distillation technology, so that the student network model can efficiently run on the edge device, improve the prediction accuracy and real-time performance of the model, and meet the actual needs of industrial energy consumption control.

[0085] Step 205, when the prediction deviation of the future energy consumption prediction data exceeds the energy consumption threshold, automatically generating a device control strategy and controlling the edge executor to execute the device control strategy to dynamically adjust the device parameters.

[0086] In step 205, based on the deviation between the future energy consumption prediction data and the actual energy consumption value, dynamic regulation of the industrial device is realized to ensure that the energy consumption is maintained within the expected range.

[0087] Firstly, the system continuously compares the future energy consumption prediction data with the actual monitored real energy consumption value, and calculates the prediction deviation. Once the prediction deviation exceeds the set energy consumption threshold, the system triggers the early warning mechanism in time. The energy consumption threshold is adaptively adjusted according to the characteristics of the target control device, covering different categories such as large inertia devices, fast response devices and auxiliary devices. For example, for large inertia devices such as heating furnaces, the threshold is appropriately relaxed to 20% to reduce the regulation frequency; and for fast response devices such as fans and pumps, the threshold is tightened to 10% to improve the regulation accuracy.

[0088] Subsequently, the system automatically generates a device control strategy. The strategy covers key information such as prediction deviation type, target device, specific parameters and adjustment range. Taking the heating furnace in steel production as an example, when the predicted energy consumption is high, the strategy may be to reduce the power of the burner, and the adjustment range is set to-10% to-5%; if the energy consumption is predicted to be low, the motor operating frequency is increased accordingly, and the adjustment range is controlled to+5% to+10%.

[0089] After the control strategy is generated, the system is implemented through the edge executor. The edge controller sends accurate instructions to the executor, and the executor dynamically adjusts the device operating parameters in real time according to the instructions. For example, adjust the motor frequency converter to change the motor speed, or adjust the valve opening to control the flow. In chemical production, according to the energy consumption prediction and deviation, the system adjusts the heating power of the reaction kettle in time to ensure that the reaction process is efficient and energy-saving.

[0090] The technical effects realized by the industrial energy consumption control method provided by the embodiment of the present application include:

[0091] Data noise removal: The energy consumption data is collected by the data acquisition device and high-frequency noise and impulse noise are removed to obtain denoised energy consumption data, providing a high-quality data basis for subsequent feature extraction and energy consumption prediction. This data processing procedure can effectively improve the accuracy and reliability of the data, ensuring the precision of the energy consumption prediction.

[0092] Model training and deployment optimization: The first energy consumption prediction data is generated by training the teacher network model on the cloud, and the student network model deployed on the edge gateway is trained with real labels to obtain a lightweight model. This approach takes advantage of the powerful computing power of the cloud for model training and uses knowledge distillation technology to transfer the knowledge of the complex teacher model to the lightweight student model, enabling it to run efficiently on edge devices and improving the prediction accuracy and real-time performance of the model.

[0093] Real-time and automated device regulation: When the prediction deviation of future energy consumption prediction values exceeds the energy consumption threshold, a device regulation strategy is automatically generated, and the edge actuator is controlled to execute the device regulation strategy for dynamic adjustment of device parameters. This real-time monitoring and automatic regulation mechanism can respond to energy consumption changes in a timely manner, avoid energy waste, improve energy utilization efficiency, and reduce production costs.

[0094] In this embodiment, after receiving the energy consumption data, different industrial devices are first classified and identified, such as motors, furnaces, pumps, etc. The energy consumption data of different devices has different noise characteristics, for example, the energy consumption data of a motor may contain periodic high-frequency noise, while the energy consumption data of a furnace may be subject to significant low-frequency interference. According to the device type, parameters such as the decomposition level of wavelet transform, wavelet basis function, and initial window size of adaptive median filtering are dynamically adjusted. For example, for motor devices, a higher decomposition level (such as 5 layers) and a suitable wavelet basis function (such as sym5) are used to more finely remove high-frequency noise; for furnace devices, the decomposition level is appropriately reduced (such as 3 layers), and a wavelet basis function that can better cope with low-frequency interference (such as bior3.5) is selected, while the initial window size of adaptive median filtering is adjusted to better adapt to the energy consumption data characteristics of different devices.

[0095] Specifically, in the denoising process of industrial energy consumption data, wavelet transform is a powerful tool that can effectively remove high-frequency noise and preserve the key features of the signal. Specifically, the energy consumption data is decomposed through a high-pass filter and a low-pass filter to obtain low-frequency approximation coefficients and multiple high-frequency detail coefficients. These high-frequency detail coefficients include first, second, and third detail coefficients, which correspond to high-frequency noise in different frequency bands, and the frequency bands decrease in turn. The low-frequency approximation coefficient includes the fourth detail coefficient, which mainly preserves the main trend of the signal.

[0096] The energy consumption data is decomposed using a multi-resolution analysis (MRA) framework. MRA represents a signal as a combination of approximation and detail parts at different resolutions by decomposing the signal layer by layer. In this embodiment, a suitable wavelet function (such as db4 wavelet) and decomposition level (such as 3-level decomposition) are selected to ensure that the noise components in different frequency bands can be effectively separated while the main features of the signal are preserved. For example, when decomposing the motor energy consumption data, 3-level decomposition can separate the signal into 3 high-frequency detail parts and 1 low-frequency approximation part, each detail part corresponding to noise in a different frequency range, thereby providing a basis for subsequent targeted denoising processing.

[0097] Next, the high-frequency detail coefficients obtained by decomposition are subjected to adaptive threshold processing. The adaptive threshold here is dynamically adjusted according to the noise intensity to ensure that the high-frequency noise is removed while the main features of the signal are preserved to the greatest extent. In this way, high-frequency noise can be effectively suppressed while avoiding excessive smoothing that would distort the signal.

[0098] For each high-frequency detail coefficient (first, second, and third detail coefficient) obtained by decomposition, its noise intensity is calculated. The threshold is dynamically adjusted according to the noise intensity, and the threshold calculation formula is:

[0099] λ = σ

[0100] where σ is the noise standard deviation and N is the signal length. The noise standard deviation can be estimated by calculating the median absolute deviation (MAD) of the high-frequency detail coefficients.

[0101] The high-frequency detail coefficients are processed using soft threshold or hard threshold methods. The soft threshold method reduces noise by shrinking the coefficients towards zero, with the formula:

[0102] d new = sign(d) · max(|d| - λ, 0)

[0103] The hard threshold method directly sets the coefficients with absolute values less than the threshold to zero, with the formula:

[0104] d new =

[0105] where d is the original coefficient and d new is the processed coefficient.

[0106] In practical applications, appropriate threshold processing methods can be selected according to specific noise characteristics and signal features. For example, when processing energy consumption data containing sharp impulse noise, the soft threshold method may be more effective because it can smoothly reduce the impact of noise; while when processing data with lower noise intensity and clearer signal features, the hard threshold method may more directly remove noise components.

[0107] Finally, the high-frequency detail coefficients and low-frequency approximation coefficients after adaptive threshold processing are subjected to inverse wavelet transform. The inverse wavelet transform recombines the decomposition coefficients through a synthesis filter bank to reconstruct the energy consumption data with high-frequency noise components removed. This process ensures that the signal retains the main trend and key features of the original signal while removing high-frequency noise, providing a higher quality data basis for subsequent steps such as impulse noise removal and feature extraction.

[0108] The signal-to-noise ratio (SNR) and root mean square error (RMSE) of the reconstructed signal and the original signal are calculated to evaluate the denoising effect. If the denoised signal fails to meet the expected SNR or RMSE indicators, the parameters such as wavelet function, decomposition level, threshold calculation method, etc. can be adjusted according to the actual situation to optimize the denoising process until satisfactory denoising effect is obtained. For example, in the energy consumption monitoring system of a certain steel plant, by continuously adjusting the parameters of wavelet transform, the SNR of the denoised energy consumption data is finally improved by 20 dB, and the RMSE is reduced by 30%, effectively improving the accuracy of energy consumption monitoring.

[0109] Further, after the removal of high-frequency noise, there may still be impulse noise remaining in the intermediate energy consumption data. In order to further improve the data quality, adaptive median filtering is used to remove impulse noise from the intermediate energy consumption data.

[0110] Firstly, the noise intensity index of the intermediate energy consumption data is determined by a preset noise intensity evaluation model. This evaluation model quantifies the intensity of noise based on statistical characteristics of the data, such as variance, peak value, and other factors. According to the noise intensity index obtained by evaluation, the size of the filtering window and the data range of the intermediate energy consumption data are dynamically adjusted. Specifically, if the noise intensity index is less than 30%, indicating a low noise level, the filtering window size is set to 3x3, and the intermediate energy consumption data containing the first, second, third, and fourth detail coefficients after wavelet inverse transform is used for filtering processing. Such a setting can remove impulse noise while preserving more signal details. When the noise intensity index is between 30% and 50%, the noise level is moderate, the filtering window size is adjusted to 5x5, and the intermediate energy consumption data only includes the second, third, and fourth detail coefficients after wavelet inverse transform. This adjustment can remove more noise while avoiding excessive smoothing that leads to loss of signal characteristics. When the noise intensity index exceeds 50%, it means that the noise level is high, and stronger filtering processing is needed. Therefore, the filtering window size is increased to 7x7, and the intermediate energy consumption data only retains the third and fourth detail coefficients after wavelet inverse transform. In this way, strong impulse noise can be effectively removed while minimizing the impact on the main characteristics of the signal. After determining the filtering window size and data range, the adjusted intermediate energy consumption data is processed to remove impulse noise, and the final denoised energy consumption data is obtained. This process can flexibly adjust the filtering strategy according to different noise intensities, ensuring effective removal of impulse noise under various noise conditions and providing high-quality data support for subsequent energy consumption analysis and prediction.

[0111] In one specific embodiment of the present application, after determining the filtering window size and data range, the adjusted intermediate energy consumption data is processed to remove impulse noise. Specifically, the intermediate energy consumption data is divided into multiple intervals with the same size as the filtering window. For each data point in the interval, the data points are first sorted in ascending order, and then the median is calculated. If the deviation of the center data point value from the median exceeds the preset two standard deviation range, the center data point is determined as an impulse noise point, and it is replaced by the median. This process can effectively remove impulse noise and preserve the main characteristics of the signal. If the data point does not exceed the deviation range, the original value is retained. Then, the filtering window slides sequentially until the entire data sequence is processed, and the final denoised energy consumption data is output.

[0112] Further, in the industrial energy consumption control method of the present application, the training and optimization of the lightweight model are key links to achieve efficient energy consumption prediction. The training process, pruning strategy, and fine-tuning method of the model are described below.

[0113] Bi-LSTM as the teacher network, containing 2 layers of hidden layers (128 neurons per layer), the input layer receives a 100-dimensional feature vector (including time domain, frequency domain and process characteristics), and the output layer predicts the energy consumption value in the future 1 hour. The model structure is as follows: input layer (100) → bidirectional LSTM layer (128) → bidirectional LSTM layer (128) → fully connected layer (1) → output layer.

[0114] The student network uses a single-layer GRU (64 neurons), which reduces the number of parameters by 60% compared to the teacher network. The structure is as follows: input layer (100) → GRU layer (64) → fully connected layer (32) → fully connected layer (1) → output layer.

[0115] For the cloud teacher network model, the teacher network model is trained on the cloud server using historical energy consumption data and real labels. The teacher network uses an integrated model architecture of LSTM and XGBoost to fully leverage the advantages of both: LSTM can capture long-term dependencies in time series data, while XGBoost is good at processing structured data nonlinear features.

[0116] The training data set contains hourly energy consumption data for the past year, divided into training set, validation set and test set in the ratio of 7:2:1. The input features include time domain features (mean, variance, etc.), frequency domain features (dominant frequency, power spectral density, etc.) and industrial mechanism features (device operating parameters, production process indicators, etc.), and the output is the energy consumption value in the future 1 hour.

[0117] During optimization, grid search is used to optimize hyperparameters such as the number of hidden layer neurons in LSTM and the tree depth of XGBoost, and mean squared error (MSE) is used as the loss function. After 100 rounds of iterative training, the MSE on the validation set is reduced to 0.045, indicating that the teacher network has high prediction accuracy.

[0118] For the student network model: The first energy consumption prediction data generated by the teacher network and the real label are issued to the edge gateway for training the student network model. The student network uses a lightweight GRU architecture, which includes an input layer, a GRU layer (64 neurons), a fully connected layer (32 neurons) and an output layer.

[0119] When training the student network, a hybrid loss function is used:

[0120] L = α × L KL + (1 - α) × L MSE

[0121] Where, L KLis the KL divergence between the teacher network output and the student network output, measuring the similarity of the two probability distributions; L MSE is the mean square error between the student network output and the true label; a is set to 0.7 to balance the influence of the two.

[0122] When the MSE of the student network on the validation set is less than 0.05, the training is considered complete, and a lightweight model is obtained. The inference delay of this model on edge devices is controlled within 300ms, meeting the real-time requirements.

[0123] After training, the L1 norm of the connection weights of each neuron in the GRU layer and the fully connected layer is calculated. The L1 norm reflects the sum of the absolute values of the weights, and the smaller the value, the lower the importance of the connection.

[0124] According to the contribution difference of neurons in the training process, a personalized pruning ratio is developed. For example, for the GRU layer, the top 80% of the L1 norm weights are retained; for the fully connected layer, the top 90% of the L1 norm weights are retained. This differentiated pruning strategy takes into account the characteristics of different layers: the GRU layer is responsible for capturing time series features and has a greater impact on model accuracy, so more weights are retained; the fully connected layer mainly performs feature combination and can be pruned more to reduce the parameter amount.

[0125] After pruning, the parameter amount of the GRU layer is reduced by about 20%, and the parameter amount of the fully connected layer is reduced by about 10%. The total parameter amount of the model is reduced from 4.2MB to 1.8MB, with a compression rate of more than 50%.

[0126] After pruning, the range of neurons to be frozen is dynamically determined based on the activation frequency or gradient change of neurons during training. For example, for neurons with low activation frequency and small gradient change, they are preferentially included in the frozen range.

[0127] where the activation frequency refers to the proportion of the number of times a neuron is activated to the total number of training samples during model training. The activation of a neuron is usually determined by its input signal and weight, and when the weighted sum of the input signal and weight exceeds a certain threshold, the neuron is activated.

[0128] Neurons with high activation frequency play an important role in processing most input data and may be key feature extractors for the model to make predictions. These neurons have a greater influence on the output of the model and are usually retained. For example, a neuron is activated 800 times out of 1000 training samples, with an activation frequency of 80%. This indicates that the neuron is involved in information processing in most cases and contributes significantly to the performance of the model, so it should be preferentially retained during fine-tuning.

[0129] Gradient change refers to the absolute value change of the gradient of the loss function with respect to the neuron weights during the model training process. The gradient reflects the direction and rate of change of the loss function under the current weights, and a large gradient change means that the weight has a large fluctuation in its impact on the loss function during training.

[0130] Neurons with large gradient changes have unstable effects on the loss function during training, which may have a significant impact on performance improvement in some cases, but may introduce noise or instability in other cases. By analyzing the gradient change, we can filter out neurons that are stable and have a significant impact on model performance. For example, during training, the gradient of a certain neuron weight fluctuates greatly at different iteration steps, and its absolute value is sometimes large and sometimes small. This indicates that the neuron's role in the training process is unstable, and its contribution to the model may need to be further evaluated. If a neuron has a small gradient change and a small absolute gradient, it means that its impact on the loss function is relatively stable and small during training, and it can be considered for adjustment or freezing during fine-tuning.

[0131] After freezing the top 50% of neurons in the GRU layer, the remaining part of the model is fine-tuned using the local data of the edge device (the energy consumption data of the last 7 days). During fine-tuning, only the connection weights of the unfrozen neurons are trained to further optimize the model performance.

[0132] After fine-tuning, the MSE of the model on the test set is reduced to 0.038, which improves the accuracy compared to before pruning, and the inference speed is increased by about 20%, significantly improving the running efficiency of the edge device.

[0133] Then the fine-tuned lightweight model is deployed to the edge gateway. The edge gateway is responsible for receiving preprocessed feature data and running the model in real time for energy consumption prediction. During deployment, ensure that the model is compatible with the hardware architecture of the edge device, and perform format conversion and optimization on the model to adapt to the running environment of the device.

[0134] After deploying the lightweight model, input the feature data into the lightweight model to output future energy consumption prediction data. To ensure that the feature data accurately reflects the energy consumption characteristics of the current industrial production process, the system first updates and filters the feature data based on real-time running data. This means that the system will continuously monitor and integrate new running data to adjust the selection range and weight of the feature data, ensuring the timeliness and relevance of the feature data. For example, when production tasks change or new equipment is put into operation, the system can identify and incorporate new feature data related to them in a timely manner, while reducing or eliminating the weight of feature data with low relevance to old production tasks or equipment.

[0135] Next, the updated feature data is input into the lightweight model deployed by the edge gateway. This lightweight model has adaptive capabilities, allowing it to automatically adjust its internal parameters based on changes in the statistical characteristics of the input feature data. Specifically, the model can automatically adjust the neuron connection weights of the GRU layer and the parameters of the fully connected layer. This adaptive adjustment mechanism enables the model to flexibly respond to energy consumption prediction needs under different working conditions, without the need for frequent human intervention or retraining, ensuring the accuracy and reliability of the prediction results.

[0136] The feature data is sequentially processed through the GRU layer, fully connected layer, and output layer in the lightweight model. In this process, the GRU layer is responsible for capturing the temporal dependence in time series data, while the fully connected layer further combines and processes the feature data. The final output layer generates future energy consumption prediction data, providing key basis for energy consumption management and device control in industrial production processes.

[0137] In step 205, the system automatically adjusts the operating parameters of industrial equipment by real-time monitoring and analyzing energy consumption data to achieve optimal control of energy consumption. Specifically, the system first collects the actual energy consumption value of the industrial system, which is a key basis for evaluating energy consumption prediction accuracy and developing control strategies. Within each set time period, the system calculates the prediction deviation between future energy consumption prediction data and actual energy consumption value, which directly reflects the gap between prediction and actual situation.

[0138] The mean absolute percentage error (MAPE) is used to measure the prediction deviation:

[0139]

[0140] where y i is the actual energy consumption value, and y is the future energy consumption prediction data. When MAPE>15%, the device control strategy is triggered.

[0141] The system sets adaptive adjustment energy consumption thresholds for different types of target control devices, including at least large inertia devices, fast response devices, and auxiliary devices. Such classification and individualized threshold setting enable the system to more accurately respond to the energy consumption variation characteristics of different devices, improving the targeting and effectiveness of control. For example, for large inertia devices such as furnaces, the energy consumption threshold can be appropriately relaxed to 20% to reduce the frequency of control; while for fast response devices such as fans and pumps, the energy consumption threshold is tightened to 10% to improve control accuracy.

[0142] Further, the system compares the prediction deviation of each time period with the corresponding energy consumption threshold. If the prediction deviation of N consecutive time periods exceeds the set energy consumption threshold, it indicates that the current energy consumption condition has exceeded the acceptable range, and the system thus triggers the generation of the corresponding device control strategy. These strategies explicitly indicate the type of prediction deviation, the target device that needs to be controlled, the specific device parameters, and the acceptable adjustment range, thereby guiding the executor to accurately adjust the parameters of the device and achieving effective control of energy consumption. For example, see Table 1. When the predicted energy consumption is high, the strategy may include reducing the burner power of the heating furnace, with an adjustment range set at -10% to -5%; if the predicted energy consumption is low, the corresponding operating frequency of the motor is increased, with an adjustment range controlled at +5% to +10%.

[0143] Table 1

[0144]

[0145] In addition, after generating the control strategy, the system executes these strategies through the edge executor to dynamically adjust the device parameters. The executor modifies the device parameters according to the specific instructions in the control strategy. For example, adjusting the frequency converter of the motor to change the motor speed, or adjusting the valve opening to control the flow. In chemical production, according to the energy consumption prediction and deviation, the system can timely adjust the heating power of the reaction kettle to ensure that the reaction process is both efficient and energy-saving.

[0146] After executing the control strategy, the system continuously monitors the changes in energy consumption of the device and compares the actual energy consumption value with the predicted value to evaluate the effectiveness of the control strategy. According to the evaluation results, the system automatically adjusts the model parameters and the control strategy to further optimize the energy consumption control effect, thereby forming a real-time closed-loop control of "perception - analysis - control", as shown in Figure 3 The closed-loop feedback mechanism ensures that the system can continuously adapt to the dynamic changes of industrial production and continuously improve the energy consumption management level.

[0147] Through actual verification, after the system is applied in high-energy-consuming industries such as steel and chemical industry, the energy utilization rate is significantly improved by 10%-15%, the annual electricity saving can reach more than 8.5 million yuan, and the equipment operation and maintenance cost is reduced by 20%, which has proved the significant benefits in actual production.

[0148] The industrial energy consumption control system provided by the embodiments of the present application is described below. The industrial energy consumption control system described below can be correspondingly referred to the industrial energy consumption control method described above.

[0149] The embodiments of the present application provide an industrial energy consumption control system, as shown in Figure 4 , comprising:

[0150] The data acquisition module 401 is configured to acquire energy consumption data through the data acquisition device.

[0151] The noise removal module 402 is configured to sequentially perform high-frequency noise removal and pulse noise removal on the energy consumption data to obtain denoised energy consumption data.

[0152] The feature extraction module 403 is configured to extract time domain features, frequency domain features and industrial mechanism features from the denoised energy consumption data to form feature data.

[0153] The energy consumption data prediction module 404 is configured to input the feature data into a lightweight model deployed on an edge gateway to output future energy consumption prediction data. The lightweight model is a teacher network model trained on a cloud through historical energy consumption data and real labels to generate first energy consumption prediction data, and the first energy consumption prediction data and the real labels are issued to the edge gateway to train a student network model to obtain.

[0154] The control strategy generation module 405 is configured to automatically generate a device control strategy when a prediction deviation of the future energy consumption prediction data exceeds an energy consumption threshold, and control the edge actuator to execute the device control strategy to dynamically adjust device parameters.

[0155] Figure 5 An example of an entity structure diagram of an electronic device is shown in Figure 5 As shown, the electronic device can include a processor 510, a communications interface 520, a memory 530 and a communications bus 540, wherein the processor 510, the communications interface 520 and the memory 530 communicate with each other through the communications bus 540. The processor 510 can invoke logical instructions in the memory 530 to execute an industrial energy consumption control method, which includes: acquiring energy consumption data through the data acquisition device; sequentially performing high-frequency noise removal and pulse noise removal on the energy consumption data to obtain denoised energy consumption data; extracting time domain features, frequency domain features and industrial mechanism features from the denoised energy consumption data to form feature data; inputting the feature data into a lightweight model deployed on an edge gateway to output future energy consumption prediction data; when a prediction deviation of the future energy consumption prediction data exceeds an energy consumption threshold, automatically generating a device control strategy, and controlling the edge actuator to execute the device control strategy to dynamically adjust device parameters.

[0156] Further, the logic instructions in the memory 530 described above can be implemented in the form of software functional units and sold or used as independent products, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that make contributions to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0157] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the industrial energy consumption control method provided by the above-mentioned methods. The method comprises: collecting energy consumption data by the data acquisition device; sequentially performing high-frequency noise removal and pulse noise removal on the energy consumption data to obtain denoised energy consumption data; extracting time domain features, frequency domain features and industrial mechanism features from the denoised energy consumption data to form feature data; inputting the feature data into a lightweight model deployed by an edge gateway to output future energy consumption prediction data; when the prediction deviation of the future energy consumption prediction data exceeds an energy consumption threshold, automatically generating a device control strategy, and controlling the edge executor to execute the device control strategy to dynamically adjust the device parameters.

[0158] In another aspect, the present application also provides a computer program product, which comprises a computer program, the computer program can be stored on a non-transitory computer readable storage medium, and the computer program can be executed by a processor to enable a computer to execute the industrial energy consumption control method provided by the above-mentioned methods. The method comprises: collecting energy consumption data by the data acquisition device; sequentially performing high-frequency noise removal and pulse noise removal on the energy consumption data to obtain denoised energy consumption data; extracting time domain features, frequency domain features and industrial mechanism features from the denoised energy consumption data to form feature data; inputting the feature data into a lightweight model deployed by an edge gateway to output future energy consumption prediction data; when the prediction deviation of the future energy consumption prediction data exceeds an energy consumption threshold, automatically generating a device control strategy, and controlling the edge executor to execute the device control strategy to dynamically adjust the device parameters.

[0159] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purposes of the embodiments according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0160] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An industrial energy consumption control method, characterized by, An edge controller for an industrial system, the edge controller being connected with a data acquisition device and an actuator respectively, the method comprising: acquiring energy consumption data through the data acquisition device; removing high-frequency noise and impulse noise from the energy consumption data in sequence to obtain denoised energy consumption data; extracting time-domain features, frequency-domain features and industrial mechanism features from the denoised energy consumption data to form feature data; inputting the feature data into a lightweight model deployed on an edge gateway to output future energy consumption prediction data; wherein the lightweight model is generated by training a teacher network model on a cloud through historical energy consumption data and real labels, and the first energy consumption prediction data and the real labels are issued to the edge gateway to train a student network model to obtain the lightweight model; when the prediction deviation of the future energy consumption prediction data exceeds an energy consumption threshold, automatically generating a device control strategy and controlling the edge actuator to execute the device control strategy to dynamically adjust the device parameters; the lightweight model comprises an input layer, a GRU layer, a fully connected layer and an output layer, the GRU layer comprises a plurality of neurons, and the fully connected layer comprises a plurality of input neurons and output neurons; after the training of the lightweight model is completed, the method further comprises: pruning: calculating the L1 norm of the weight matrix corresponding to the connection between each neuron of the GRU layer and the L1 norm of the weight matrix corresponding to the connection between the input neurons and the output neurons of the fully connected layer; formulating a personalized pruning ratio according to the contribution difference of different neurons or connections in the GRU layer and the fully connected layer during the training process; pruning the GRU layer and the fully connected layer according to the personalized pruning ratio, and retaining the corresponding proportion of weights with smaller L1 norm; wherein the weights represent the importance of the connection between neurons, and the smaller the absolute value of the L1 norm, the lower the importance of the connection; fine-tuning: dynamically determining the range of neurons to be frozen based on the activation frequency or gradient change of the neurons during the training process; freezing the corresponding neurons in the GRU layer based on the frozen range, and continuing to train the remaining part of the initial lightweight model through the local data of the edge device to obtain the final lightweight model.

2. The industrial energy consumption control method of claim 1, wherein training the student network model by issuing the first energy consumption prediction data and the real labels to the edge gateway, and training the student network model based on a hybrid loss function, if the mean square error between the prediction result of the trained student network model on the validation set and the real label is less than a threshold, the lightweight model is obtained; the hybrid loss function comprises: the KL divergence between the first energy consumption prediction data and the second energy consumption prediction data output by the student network model, and the mean square error between the second energy consumption prediction data and the real label. removing high-frequency noise and impulse noise from the energy consumption data in sequence to obtain denoised energy consumption data, specifically comprising:

3. The industrial energy consumption control method of claim 1, wherein, ​ The energy consumption data is subjected to wavelet transform to remove high-frequency noise, to obtain intermediate energy consumption data; The intermediate energy consumption data is subjected to adaptive median filtering to remove impulse noise, to obtain denoised energy consumption data.

4. The industrial energy consumption control method of claim 3, wherein, The energy consumption data is subjected to wavelet transform to remove high-frequency noise, to obtain intermediate energy consumption data, specifically including: The energy consumption data is subjected to high-pass filter and low-pass filter, to be decomposed into low-frequency approximation coefficient and multiple high-frequency detail coefficients; the multiple high-frequency detail coefficients include first detail coefficient, second detail coefficient and third detail coefficient in turn with decreasing frequency band, and the low-frequency approximation coefficient includes fourth detail coefficient; The first detail coefficient, the second detail coefficient and the third detail coefficient are subjected to adaptive threshold processing respectively; wherein the threshold is dynamically adjusted according to noise intensity; The first detail coefficient, the second detail coefficient, the third detail coefficient and the fourth detail coefficient subjected to adaptive threshold processing are subjected to wavelet inverse transform, to eliminate high-frequency noise component, to obtain denoised intermediate energy consumption data.

5. The industrial energy consumption control method of claim 4, wherein, The intermediate energy consumption data is subjected to adaptive median filtering to remove impulse noise, to obtain denoised energy consumption data, specifically including: The noise intensity index of the intermediate energy consumption data is determined through a preset noise intensity evaluation model; The data range of the intermediate energy consumption data is adjusted according to the noise intensity index and the size of the filtering window; wherein, if the noise intensity index is less than 30%, the size of the filtering window is 3*3, and the intermediate energy consumption data includes the first detail coefficient, the second detail coefficient, the third detail coefficient and the fourth detail coefficient subjected to wavelet inverse transform; if the noise intensity index is 30%~50%, the size of the filtering window is 5*5, and the intermediate energy consumption data includes the second detail coefficient, the third detail coefficient and the fourth detail coefficient subjected to wavelet inverse transform; if the noise intensity index is greater than 50%, the size of the filtering window is 7*7, and the intermediate energy consumption data includes the third detail coefficient and the fourth detail coefficient subjected to wavelet inverse transform; The intermediate energy consumption data subjected to data range adjustment is subjected to impulse noise elimination according to the size of the filtering window, to obtain denoised energy consumption data.

6. The industrial energy consumption control method of claim 1, wherein, The feature data is input into a lightweight model deployed by an edge gateway, to output future energy consumption prediction data, specifically including: The feature data is dynamically updated and screened according to real-time running data, to adjust the selection range and weight of the feature data; The feature data is input into a lightweight model deployed by an edge gateway, to enable the lightweight model to automatically adjust the internal parameters of the model according to the statistical characteristic changes of the feature data, wherein the internal parameters of the model include the neuron connection weight of the GRU layer and the parameters of the full connection layer; The feature data is sequentially subjected to processing by the GRU layer, the full connection layer and the output layer, to output the future energy consumption prediction data.

7. The industrial energy consumption control method of claim 1, wherein, When the prediction deviation of the future energy consumption prediction data exceeds an energy consumption threshold, a device regulation strategy is automatically generated, specifically including: Collecting real energy consumption values of the industrial system, and calculating prediction deviations of the future energy consumption prediction data and the real energy consumption values in each time period; Comparing the prediction deviations of each time period with the energy consumption threshold respectively, wherein the energy consumption threshold is adaptively adjusted according to the type of target control equipment, and the type of target control equipment at least includes large inertia equipment, fast response equipment and auxiliary equipment; If the prediction deviations of continuous N time periods exceed the energy consumption threshold, a device control strategy is generated; wherein the device control strategy includes prediction deviation type, target control equipment, device parameter and adjustment range.

8. An industrial energy consumption control system, characterized by, An edge controller for an industrial system, the edge controller is connected with a data acquisition device and an actuator respectively, and the system comprises: A data acquisition module for acquiring energy consumption data through the data acquisition device; A noise removal module for sequentially removing high-frequency noise and impulse noise from the energy consumption data to obtain denoised energy consumption data; A feature extraction module for extracting time domain features, frequency domain features and industrial mechanism features from the denoised energy consumption data to form feature data; An energy consumption data prediction module for inputting the feature data into a lightweight model deployed in an edge gateway to output future energy consumption prediction data; wherein the lightweight model is a teacher network model trained in the cloud through historical energy consumption data and real labels to generate first energy consumption prediction data, and the first energy consumption prediction data and the real labels are issued to the edge gateway to train a student network model to obtain; A control strategy generation module for automatically generating a device control strategy when the prediction deviation of the future energy consumption prediction data exceeds an energy consumption threshold, and controlling an edge actuator to execute the device control strategy to dynamically adjust the device parameter; The lightweight model comprises an input layer, a GRU layer, a fully connected layer and an output layer, the GRU layer comprises a plurality of neurons, and the fully connected layer comprises a plurality of input neurons and output neurons; After the training of the lightweight model is completed, it further comprises: Pruning: calculating the L1 norm of the weight matrix corresponding to the connection between each neuron of the GRU layer and the L1 norm of the weight matrix corresponding to the connection between the input neurons and the output neurons of the fully connected layer; According to the contribution difference of different neurons or connections in the GRU layer and the fully connected layer in the training process, a personalized pruning ratio is formulated; According to the personalized pruning ratio, the GRU layer and the fully connected layer are pruned respectively, and the corresponding proportion of weights with smaller L1 norm is retained; wherein the weight represents the importance of the connection between neurons, and the smaller the absolute value of the L1 norm, the lower the importance of the connection; Fine-tuning: dynamically determining the range of neurons to be frozen based on the activation frequency or gradient change of the neurons in the training process; Freezing the corresponding neurons in the GRU layer based on the frozen range, and continuing to train the remaining part of the initial lightweight model through the local data of the edge device to obtain the final lightweight model.

9. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, implements the industrial energy consumption control method according to any one of claims 1 to 7.

Images