Electromechanical equipment energy-saving management and control system and method based on intelligent monitoring

Abstract

The invention discloses an electromechanical equipment energy-saving management and control system and method based on intelligent monitoring, and relates to the field of intelligent energy-saving management and control, and the method comprises the steps: firstly carrying out the filtering and excitation validity judgment of original sensor data, and screening out data with a parameter identification value; on the basis, a regression feature vector representing the current physical state of the equipment is established through physical feature extraction and linear conversion; then, algorithm gains are dynamically adjusted through variable forgetting factors based on posterior errors, time-varying tracking is carried out on model parameters, a projection correction mechanism under physical constraints is introduced, and the updated parameters are forced to be located in a reasonable physical feasible region all the time. And finally, performing global optimization under the constraint of working conditions based on the self-adaptive energy efficiency model corrected in real time, thereby ensuring that the control model can accurately follow the aging process of the equipment, and realizing accurate energy-saving control of the whole life cycle.

Description

Technical Field

[0001] This application relates to the field of intelligent energy-saving management and control, and more specifically, to an energy-saving management and control system and method for electromechanical equipment based on intelligent monitoring. Background Technology

[0002] Existing energy-saving control technologies for electromechanical equipment mainly rely on static physical models or black-box models trained offline based on historical data. These technologies generally suffer from poor adaptability to changes in the physical characteristics of equipment throughout its entire lifecycle and a lack of physical constraints. Current technologies typically rely on static assumptions, assuming that key physical parameters such as heat transfer coefficients, compressor efficiency, and friction coefficients are constant values ​​or functions that only change with load. However, in actual long-term operating environments, electromechanical equipment is inevitably affected by physical factors such as heat exchanger fouling, lubricant deterioration, and mechanical component wear, leading to irreversible, nonlinear, time-varying degradation of its physical characteristics over time (i.e., soft faults). This degradation causes a significant drift between the actual physical parameters of the equipment after a period of operation and the initially calibrated model parameters, resulting in a serious model mismatch problem. Since pure physical models struggle to accurately capture this complex microscopic dynamic degradation process, and purely data-driven models lacking physical mechanism constraints are prone to outputting control commands that violate physical principles (such as negative pressure or ultra-high efficiency) under extreme conditions or in regions with sparse samples, this can lead to control safety hazards. Furthermore, existing control models are mostly "open-loop" applications, meaning that once the model is trained or calibrated, it is used for a long time without an online, automated "residual observation-parameter correction" closed-loop mechanism to combat equipment aging in real time. This results in energy-saving control strategies often having to calculate the optimal operating point based on outdated model parameters, causing the calculated control parameters to deviate from the actual optimal state. This not only fails to achieve the expected energy-saving effect but also causes continuous and hidden energy waste.

[0003] Therefore, there is an urgent need for an energy-saving management and control scheme for electromechanical equipment that can integrate physical mechanism constraints and can perceive and adaptively correct equipment aging characteristics in real time based on monitoring data. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this application provides an energy-saving control method for electromechanical equipment based on intelligent monitoring, comprising: S1: performing low-pass filtering and excitation validity prediction on the raw sensor data stream of the collected electromechanical equipment to obtain an excitation validity flag and a filtered dataset; S2: extracting physical features and linearizing the filtered dataset to obtain a regression feature vector and an observation scalar representing the physical state at the current moment; S3: when the excitation validity flag indicates validity, determining a variable forgetting factor based on the posterior error of the previous moment, and determining a gain vector that can suppress noise interference by combining the regression feature vector and the covariance matrix of the previous moment; S4: updating and projecting the model parameters of the previous moment under physical constraints based on the gain vector, the variable forgetting factor, and the observation scalar to obtain the corrected model parameters and the covariance matrix at the current moment; S5: constructing an adaptive energy efficiency prediction model based on the corrected model parameters, and performing global optimization on the adaptive energy efficiency prediction model under the operating condition constraints corresponding to the filtered dataset to obtain the optimal setpoint; S6: issuing the optimal setpoint to the electromechanical equipment for control execution.

[0005] This application also provides an energy-saving control system for electromechanical equipment based on intelligent monitoring, comprising: an electromechanical equipment data processing module, used to perform low-pass filtering and excitation validity prediction on the raw sensor data stream of the collected electromechanical equipment to obtain an excitation validity flag and a filtered dataset; a feature extraction and linearization transformation module, used to extract physical features and perform linearization transformation on the filtered dataset to obtain a regression feature vector and an observation scalar representing the physical state at the current moment; and a gain vector generation module, used to determine a variable forgetting factor based on the posterior error of the previous moment when the excitation validity flag indicates validity, and combine the regression features... The gain vector and the covariance matrix of the previous time step determine the gain vector that can suppress noise interference; the parameter update and projection correction module is used to update and project the model parameters of the previous time step under physical constraints based on the gain vector, the variable forgetting factor and the observation scalar to obtain the corrected model parameters and the covariance matrix of the current time step; the optimal setpoint determination module is used to construct an adaptive energy efficiency prediction model based on the corrected model parameters, and to perform global optimization of the adaptive energy efficiency prediction model under the operating condition constraints corresponding to the filtered dataset to obtain the optimal setpoint; the control execution module is used to send the optimal setpoint to the electromechanical equipment for control execution.

[0006] Compared with existing technologies, this application provides an energy-saving control system and method for electromechanical equipment based on intelligent monitoring. Addressing the problems of physical characteristic drift, static model mismatch, and unreliable output from unconstrained black-box models caused by equipment aging, this application first filters and determines the excitation validity of raw sensor data, selecting data with parameter identification value. Based on this, a regression feature vector representing the current physical state of the equipment is established through physical feature extraction and linearization transformation. Then, the algorithm gain is dynamically adjusted using a variable forgetting factor based on posterior error to track the model parameters over time, and a projection correction mechanism under physical constraints is introduced to force the updated parameters to always remain within a reasonable physical feasible domain. Finally, an adaptive energy efficiency model based on real-time correction performs global optimization under operating condition constraints. By constructing a closed loop of perception-identification-constraint-optimization, this ensures that the control model can accurately follow the aging process of the equipment, thereby achieving precise energy-saving control throughout its entire lifecycle. Attached Figure Description

[0007] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings.

[0008] Figure 1 This is a flowchart of an energy-saving control method for electromechanical equipment based on intelligent monitoring, according to an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of data flow in an energy-saving management and control method for electromechanical equipment based on intelligent monitoring, according to an embodiment of this application.

[0010] Figure 3 This is a flowchart of step S4 in the energy-saving control method for electromechanical equipment based on intelligent monitoring according to an embodiment of this application.

[0011] Figure 4 This is a block diagram of an energy-saving control system for electromechanical equipment based on intelligent monitoring, according to an embodiment of this application. Detailed Implementation

[0012] The embodiments of this application will now be described in more detail with reference to the accompanying drawings. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0013] In view of the shortcomings in the above-mentioned technical fields, this application proposes an energy-saving control method for electromechanical equipment based on intelligent monitoring. Figure 1 This is a flowchart of an energy-saving control method for electromechanical equipment based on intelligent monitoring, according to an embodiment of this application. Figure 2 This is a schematic diagram of data flow in an energy-saving control method for electromechanical equipment based on intelligent monitoring, according to an embodiment of this application. Figure 1 and Figure 2As shown, the energy-saving control method for electromechanical equipment based on intelligent monitoring according to an embodiment of this application includes: S1: performing low-pass filtering and excitation validity prediction on the raw sensor data stream of the collected electromechanical equipment to obtain an excitation validity flag and a filtered dataset; S2: performing physical feature extraction and linearization transformation on the filtered dataset to obtain a regression feature vector and an observation scalar used to characterize the physical state at the current moment; S3: when the excitation validity flag indicates validity, determining a variable forgetting factor based on the posterior error of the previous moment, and determining a gain vector that can suppress noise interference by combining the regression feature vector and the covariance matrix of the previous moment; S4: updating and projecting the model parameters of the previous moment under physical constraints based on the gain vector, the variable forgetting factor, and the observation scalar to obtain the corrected model parameters and the covariance matrix at the current moment; S5: constructing an adaptive energy efficiency prediction model based on the corrected model parameters, and performing global optimization on the adaptive energy efficiency prediction model under the operating condition constraints corresponding to the filtered dataset to obtain the optimal setpoint; S6: issuing the optimal setpoint to the electromechanical equipment for control execution.

[0014] In step S1, the raw sensor data stream of the collected electromechanical equipment is subjected to low-pass filtering and excitation validity prediction to obtain the excitation validity flag and the filtered dataset. It should be understood that in the actual operating environment of industrial sites, the operating data of electromechanical equipment collected by sensors is often accompanied by complex operating conditions. Power grid fluctuations, electromagnetic interference, and the sensor's own measurement noise will cause a large amount of high-frequency random noise to be mixed in the raw data. If this noise directly enters the subsequent model parameter identification stage, the algorithm will mistakenly identify the noise as changes in the physical characteristics of the equipment, thereby causing model parameter jitter or even divergence. On the other hand, electromechanical equipment is in a stable state for most of its operating time. At this time, the data contains a large amount of redundant information and lacks dynamic features sufficient to stimulate model parameter updates. If such stable data is continuously used for parameter iteration, it will lead to singular covariance matrix in the adaptive algorithm or the estimator becoming dormant due to lack of excitation, thus weakening the model's ability to capture sudden changes in operating conditions. Therefore, in order to ensure that the energy efficiency prediction model constructed subsequently can be processed based on real, effective and physically meaningful data, this application conducts rigorous signal cleaning and excitation effectiveness screening to remove noise interference and identify dynamic data segments with parameter identification value.

[0015] In one exemplary embodiment of this application, step S1 includes: S11, performing multi-channel signal synchronization and digital low-pass filtering on the original sensor data stream to obtain a filtered dataset; S12, performing instantaneous rate of change estimation based on Euclidean distance on the filtered dataset from the previous time step and the filtered dataset from the current time step to obtain a data rate of change index reflecting the degree of system disturbance; S13, performing a logical comparison and decision between the data rate of change index and a preset dead zone threshold to obtain an excitation valid flag.

[0016] The operation method is detailed below: Before elaborating on the specific implementation of step S1 and its sub-steps, it is first necessary to clarify the physical meaning of each component in the original sensor data stream and its acquisition method. Specifically, the original sensor data stream here includes three-phase voltage, three-phase current, active power, power factor, working fluid inlet temperature, working fluid outlet temperature, working fluid flow rate, system pressure, operating frequency, valve opening degree, start / stop status, ambient dry-bulb temperature, and relative humidity. This data stream is a digital mirror that can comprehensively map the operating status of electromechanical equipment. Among them, three-phase voltage and three-phase current are used to monitor the power quality on the power supply side and the load current balance of the motor. Active power directly reflects the actual power consumption of the equipment at present, while the power factor reflects the efficiency level of power utilization. The above electrical parameters are collected through multi-functional intelligent power meters installed in the equipment control cabinet, or directly read from the high-precision register values ​​inside the frequency converter via the RS485 communication interface. In terms of thermodynamic performance, the inlet and outlet temperatures of the working fluid characterize the enthalpy change of the fluid before and after passing through the equipment. Combined with the working fluid flow rate collected by electromagnetic or ultrasonic flow meters, the real-time cooling or heating capacity of the equipment can be calculated using thermodynamic formulas. System pressure is obtained from pressure transmitters installed at key nodes to monitor pipeline resistance or compressor discharge pressure, preventing surge or overpressure operation. As inputs for control, operating frequency directly determines the rotational speed of rotating components, valve opening reflects the adjustment state of throttling elements to fluid resistance, and start / stop status is a logical Boolean value indicating whether the equipment is in an operating cycle. These three parameters are generally read directly from the memory address of a field-programmable logic controller (PLC) or distributed control system (DCS). Furthermore, considering the significant impact of the external environment on the efficiency of air-cooled equipment or cooling towers, ambient dry-bulb temperature and relative humidity need to be collected by temperature and humidity transmitters deployed around the equipment.

[0017] After obtaining the above data, step S11 is implemented first. In the monitoring network of electromechanical equipment, different physical quantities are often acquired through different sensors and acquisition modules. For example, electrical parameters may come from smart meters, while temperature and pressure parameters may come from PLCs or analog acquisition cards. This results in slight differences in the timestamps of various data arriving at the central processing unit. First, a unified sampling clock reference is established, for example, setting the sampling period to 1 second. For each sampling moment, the nearest original data frame before and after the current moment for each channel is retrieved. Linear interpolation or zero-order hold is used to align the data of all channels to the same time segment, forming a synchronized original data vector. Subsequently, in order to filter out the high-frequency white noise superimposed on the real physical signal, a first-order exponentially weighted moving average filter is applied to each synchronized physical quantity channel. The core logic of this digital low-pass filter is to use the current measurement value and the filtered estimate value of the previous moment for weighted fusion. Its mathematical expression is that the current filter value is equal to the filter coefficient multiplied by the current original measurement value plus (1 minus the filter coefficient) multiplied by the filter value of the previous moment. Here, the filter coefficient is a preset constant between 0 and 1, used to balance the smoothness of the signal and the response hysteresis. For example, for physical quantities with high thermal inertia, such as the working fluid outlet temperature, the filter coefficient can be set relatively small, such as 0.05, to greatly smooth temperature fluctuations. For electrical quantities such as operating frequency or active power, the filter coefficient can be set slightly larger, such as 0.2, to preserve their rapid changing trends. After this processing step, the output filtered dataset is a high signal-to-noise ratio vector that has removed glitches and is time-synchronized. This vector contains the current best estimates of the aforementioned thirteen physical quantities, providing a clean data foundation for subsequent feature extraction. For example, if the instantaneous three-phase voltage value acquired by the original sensor at a certain moment is 382.5 volts, accompanied by a noise fluctuation of 0.8 volts, and the filtered output at the previous moment is 380.0 volts, if a filter coefficient of 0.1 is used, the filtered output at the current moment will be calculated as 380.25 volts, thus suppressing the masking of the true trend of the data by instantaneous noise.

[0018] Next, step S12 is implemented. A buffer is allocated in memory to store the filtered dataset output from the previous sampling time. Once the filtered dataset for the current time is generated, a vector difference operation is first performed, subtracting the data vector from the previous time from the current data vector to obtain an increment vector reflecting the magnitude of change of each physical quantity within the sampling period. Since the dataset contains physical quantities of different dimensions and orders of magnitude—for example, changes in active power may be at the kilowatt level, while changes in power factor are only at the 0.01 level—directly calculating the Euclidean distance would cause large-value physical quantities to dominate the rate of change index, masking the changes in small-value physical quantities. Therefore, this step introduces a diagonal weighting matrix. The diagonal elements of this matrix correspond to the normalized weights of each physical quantity. These weights are set based on the statistical characteristics of historical data; specifically, each weight value is the square of the reciprocal of the standard deviation of that physical quantity in historical operating data, or the square of the reciprocal of the span of its normal operating range. By introducing this weighted matrix, all changes in physical quantities are mapped to the same dimensionless metric space. Subsequently, the rate of change of data is calculated. The process is as follows: , This is the increment vector. Geometrically, this calculation is equivalent to calculating the displacement distance of the current state point relative to the previous state point in a weighted multidimensional space. This index... It is a non-negative scalar, and its magnitude directly reflects the degree of disturbance or excitation intensity of the equipment's operating state. For example, when the equipment is in a stable operating period with constant speed and constant load, the changes in various physical quantities are minimal, and the calculated values ​​are... The value will approach zero; however, when the equipment receives a frequency converter command to adjust the speed, or when the external load changes abruptly (such as valve opening adjustment), the relevant physical quantities will undergo significant jumps, leading to... The value increased rapidly.

[0019] Finally, proceed to step S13. Before this step, a dead zone threshold needs to be determined in advance. This dead zone threshold is obtained by monitoring the device for a period of time while it is in an absolutely stable operating state, and recording the calculated dead zone threshold during this stable period. The sequence is analyzed, and its mean and peak values ​​are calculated. The dead zone threshold is set slightly higher than that under steady-state conditions. A value representing the peak value, for example, 1.2 times the peak value. This threshold represents the inherent background noise of the system or the upper limit of small fluctuations within the allowable range. In specific implementation, a digital comparator is configured to use the rate of change of the current data obtained in step S12. Input the non-inverting input of the comparator and input the preset dead-time threshold to the inverting input. If If the input signal exceeds the dead-time threshold, the comparator outputs a high level or a logic true value, determining that the current input signal contains not only noise but also valid dynamic excitation reflecting the physical characteristics of the device. Therefore, the excitation validity flag is asserted as valid (True). This means that the current data point carries new physical information that can be used to correct model parameters. Conversely, if... If the value is less than or equal to the dead zone threshold, the comparator outputs a low level or a logic false value, indicating that the system is in a steady state or quiescent region. Data fluctuations are mainly caused by measurement noise and lack discriminative value; therefore, the excitation validity flag is set to invalid (False). To prevent the algorithm from idling on data with no information, leading to parameter drift or overfitting noise, subsequent steps will keep the model parameters unchanged, i.e., freeze the learning process. Specifically, consider an industrial chiller unit. At time t, the original operating frequency is 46.01Hz, and the active power is 131.0kW. In step S11, after synchronization and EWMA processing with a filtering coefficient of 0.1, the filtered operating frequency is 45.01Hz, and the active power is 120.2kW. The buffered data from the previous time t-1 shows a filtered operating frequency of 44.90Hz and an active power of 119.0kW. In step S12, the increments are calculated: frequency increment is 0.11Hz, and power increment is 1.2kW. In the pre-defined weighting matrix, the weight corresponding to the frequency is 100, which is 1 / 0.1. 2 The weight corresponding to power is 0.01, which is 1 / 10. 2 Assuming the changes in other physical quantities are zero, the quadratic form is calculated as: (0.11)^2 × 100 + (1.2)^2 × 0.01 = 1.2244. Finally, the square root is taken to obtain the rate of change index. =1.107. In step S13, the preset dead-zone threshold is 0.5. The logic comparator determines that 1.107 > 0.5, and the condition is met. Therefore, the processing unit outputs the excitation validity flag as True. Conversely, if the power changes by only 0.1kW, the calculated index may only be 0.05, which is lower than the threshold of 0.5. In this case, the flag is False, and the system will ignore this small fluctuation, maintain the original model parameters, and avoid incorrect learning of noise.

[0020] In step S2, physical features are extracted and linearized from the filtered dataset to obtain the regression feature vector and observation scalar representing the physical state at the current moment. Correspondingly, the energy consumption characteristics of electromechanical equipment are essentially determined by physical laws such as fluid mechanics, thermodynamics, and electromagnetism. These physical laws typically exhibit highly nonlinear coupling relationships between the original variables; for example, the shaft power of a water pump is proportional to the cube of its rotational speed, and the energy consumption of a chiller unit depends on the product of flow rate and temperature difference. However, the original sensor data stream only provides discrete observations of each physical quantity. Directly inputting this unprocessed nonlinear data into parameter identification algorithms based on linear assumptions (such as recursive least squares or Kalman filtering) will lead to model non-convergence or significant prediction bias. Therefore, to enable subsequent adaptive algorithms to accurately track the time-varying characteristics of equipment physical parameters under low computational load, this application constructs a bridge connecting the physical mechanism and the linear regression algorithm before parameter updates. This step aims to decouple the complex nonlinear physical process and map it into a linear mathematical structure in the parameter space by introducing prior knowledge from the physics field.

[0021] In one exemplary embodiment of this application, step S2 includes: S21, parsing and extracting the independent state variables and dependent variables required to construct the energy consumption equation from the filtered dataset to obtain a structured set of extracted variables; S22, performing a mechanism-based nonlinear feature transformation on the independent state variables in the set of extracted variables to obtain a list of transformed features; S23, mapping and assembling the list of transformed features into a linearized regression feature vector, and mapping the dependent variables in the set of extracted variables into observed scalars.

[0022] The operation method is detailed as follows: First, implement step S21. First, call the device physical model mapping table pre-stored in non-volatile memory. This mapping table is a configuration file established during the device commissioning phase based on device type (e.g., centrifugal chiller, screw air compressor, etc.) and sensor wiring definitions, creating a unique correspondence between sensor channel IDs and physical model variable names. In this embodiment, based on a preset thermodynamic mechanism model, the device's energy consumption characteristics are defined as primarily dependent on the heat load state and ambient thermal lift. Therefore, according to the mapping table, from the fully filtered dataset containing thirteen dimensions, five core variables required to construct this specific energy consumption equation are selectively selected and extracted into the extracted variable group. For example, the mapping table specifies that the active power corresponding to channel CH01 is denoted as... The inlet temperature of the working fluid corresponding to channel CH05 is denoted as The outlet temperature of the working fluid corresponding to channel CH06 is denoted as The working fluid flow rate corresponding to channel CH07 is denoted as The ambient dry-bulb temperature corresponding to channel CH12 is denoted as Based on this mapping relationship, the values ​​of the aforementioned variables are accurately extracted from the filtered dataset. During this process, to prevent ill-conditioned matrix operations or excessively large condition numbers due to significant differences in the magnitudes of different physical quantities, the extracted variables are standardized. For example, power is converted to kilowatts (kW), flow rate to cubic meters per hour, and temperature is kept at degrees Celsius. The resulting structured set of extracted variables is then a cleaned, aligned set of physical quantities with unified units. Continuing with the example in step S1, if the standardized variable value extracted at the current time t is: active power... =120.2kW, working fluid flow rate =120.0m 3 / h, working fluid inlet temperature =12.0℃, working fluid outlet temperature =7.0℃, ambient dry bulb temperature =30.0℃.

[0023] Next, step S22 is implemented. A nonlinear combination calculation is performed on the extracted variable set based on the built-in thermodynamic mechanism formula. This thermodynamic mechanism formula is a mathematical model describing the energy conversion process of the equipment. Taking a chiller unit as an example, its basic energy consumption model is described as follows: total energy consumption consists of the cooling load energy consumption generated by work, the compression energy consumption to overcome the pressure difference (pressure ratio) between condensing and evaporating pressures, and fixed losses such as mechanical friction. Specifically, this formula incorporates power... Represented as several physical characteristic terms With the parameters to be identified A linear combination, i.e. In this embodiment, three key physical characteristic terms, or basis functions, are constructed: the first characteristic term... Characterizing the current cooling load, according to the calorimetric formula Calculate the product of flow rate and inlet / outlet temperature difference: Substitute the numerical values ​​into the calculation: =120.0×(12.0-7.0)=600.0. This term has a clear physical meaning, representing the intensity of the current heat load borne by the equipment. The second characteristic term... The effect of ambient temperature on condensation pressure is characterized by the thermal lift of the system being approximately equal to the difference between the ambient temperature and the outlet water temperature. Substitute the numerical values ​​into the calculation: =30.0 - 7.0 = 23.0. This item reflects the ease or difficulty of heat exchange under current environmental conditions. The third characteristic item. Set to a constant of 1, this term captures fixed losses independent of the load, such as control circuit power consumption and mechanical no-load friction, i.e., the intercept term. After the above calculations, a transformed feature term list containing [600.0, 23.0, 1.0] is generated. This process transforms the originally nonlinear flow rate and temperature difference relationship into a relationship with respect to parameters. In contrast, the linear feature space allows physical laws to be explicitly encoded into the data features.

[0024] Finally, in step S23, the feature list obtained in step S22 is stacked sequentially to assemble a regression feature vector in column vector form. Represented in mathematical notation as: The vector Geometrically, this represents the coordinate position of the current operating condition in the physical characteristic space. Simultaneously, the dependent variable, active power, extracted and standardized in step S21... Directly mapped to the observed scalar ,Right now: =120.2. Thus, the operating state of the physical equipment is abstracted into a set of standard linear regression input pairs. This set of data clearly shows that at the current moment, when the equipment is under physical conditions with a heat load characteristic of 600.0 and a thermal lift characteristic of 23.0, its actual observed energy consumption is 120.2kW.

[0025] In step S3, when the excitation validity flag indicates validity, the variable forgetting factor is determined based on the posterior error of the previous time step, and the gain vector that can suppress noise interference is determined by combining the regression feature vector and the covariance matrix of the previous time step. It is understandable that in the full life-cycle management of electromechanical equipment, physical parameters characterizing the energy efficiency of the equipment (such as the heat transfer coefficient of heat exchangers, the friction resistance coefficient of mechanical components, etc.) have dual time-scale evolution characteristics: on the one hand, they exhibit slow, irreversible aging drift accumulated over operating time; on the other hand, they exhibit rapid, reversible random disturbances caused by transient load fluctuations or sensor measurement noise. Traditional parameter identification algorithms, if using fixed forgetting factors or gain settings, often face an irreconcilable contradiction: a smaller forgetting factor, while improving the algorithm's tracking speed of time-varying parameter characteristics, greatly amplifies the impact of measurement noise, causing severe oscillations in model parameters during steady-state operation; while a larger forgetting factor, although effectively smoothing noise, causes a significant lag in the model's response to sudden operating conditions or actual aging trends, resulting in model prediction bias. To address this challenge, this application proposes an adaptive gain mechanism that dynamically adjusts the algorithm's memory length and parameter update weights based on the statistical characteristics of the current data's incentive effectiveness and historical prediction errors.

[0026] In one exemplary embodiment of this application, step S3 includes: S31, determining a variable forgetting factor based on the posterior error of the previous time step when the excitation valid flag is true; S32, determining the gain numerator vector based on the covariance matrix and regression eigenvector of the previous time step; S33, determining the gain denominator scalar based on the quadratic form of the regression eigenvector with respect to the covariance matrix of the previous time step and in combination with the variable forgetting factor; S34, performing Kalman gain vector synthesis on the gain numerator vector and the gain denominator scalar to obtain a gain vector used to determine the weight of the current observation.

[0027] The operation method is detailed below: Before proceeding, it is necessary to clarify three key internal state variables of the algorithm: the so-called posterior error of the previous time step. This refers to the residual obtained by re-predicting the input using the updated parameters at the previous sampling time t-1, after the model parameters have been updated and corrected. It reflects the final fit of the model to the data at that time after incorporating information from the previous time step and is a key indicator of the model's current accuracy. This value is stored in the algorithm's history buffer. (Previous time step covariance matrix) This is a symmetric positive definite matrix with dimensions equal to the number of parameters to be identified, which is 3×3 in this embodiment. In Kalman filtering or recursive least squares algorithms, this matrix geometrically represents the confidence ellipsoid of the parameter estimation error. The larger the value of its diagonal elements, the less confident the algorithm is in the current estimate of the corresponding parameter, meaning it believes the parameter may have a large deviation and therefore tends to assign a larger correction weight in subsequent updates; conversely, if the value is very small, it indicates that the algorithm believes the parameter has converged and tends to maintain the status quo. Regression eigenvector This is the standardized vector output in step S2, which includes the product of flow rate and temperature difference, thermal lift, and intercept term. .

[0028] First, implement step S31. First, read the excitation validity flag output from step S1. If this flag is False, it indicates that the current data mainly consists of noise or the system is in a static state, and does not contain sufficient valid information to drive parameter updates. In this case, to prevent the algorithm from being misled by noise, the variable forgetting factor is forcibly adjusted. Set to 1.0. Mathematically, this means the algorithm's memory window is stretched to infinity, maximizing the weight given to retaining historical information, thus keeping the covariance matrix unchanged and effectively freezing the parameter update process. If the flag is True, indicating that the current data contains effective dynamic stimuli, the forgetting rate is dynamically adjusted based on the accuracy of historical predictions. The specific calculation formula is as follows:

[0029] in, This is the preset lower limit of the forgetting factor, which is determined based on the expected frequency of equipment operating condition switching. The more frequent and drastic the changes in operating conditions, the lower this value should be set to ensure tracking speed, and vice versa, to ensure steady-state accuracy, such as 0.95. The preset sensitivity coefficient is set based on the decay preference of historical error memory. The closer the value is to 1, the more tolerant it is to historical errors. The forgetting factor is only significantly reduced when the error continues or increases significantly, such as 0.99. This is a preset noise variance benchmark, obtained by statistically analyzing the variance of the model prediction error over a period of time when the device is running in steady state. It represents the inherent, unavoidable random noise level of the system, such as 10.0. The physical meaning of this formula is: when the posterior error of the previous time step... When the square of the subtrahend is very small, meaning the model prediction is very accurate, the subtrahend term approaches 0. As the value approaches 1.0, the algorithm tends to maintain a steady state; however, when the error increases significantly (i.e., the model mismatch is severe), the subtrahend term increases. Rapidly decreases and approaches the lower limit Smaller This means the algorithm will quickly forget distant historical data and instead place greater trust in the latest data, thus accelerating the expansion of the covariance matrix and giving the algorithm stronger tracking capabilities. For example, using specific numerical values: [Preset...] =0.95, =0.99, =10.0. If the model prediction was extremely accurate in the previous time step, the posterior error would be 10.0. =0.1, then (1-0.99)×0.01 / 10=0.00001, =1 - 0.00001 = 0.99999 ≈ 1.0, the algorithm remains in steady state. If a sudden change occurred in the previous time step, causing the posterior error to spike to... =5.0, then (1-0.99)×25 / 10=0.025, =1-0.025=0.975. At this point, the forgetting factor decreases, indicating that the algorithm will place greater emphasis on new data in subsequent steps.

[0030] Next, step S32 is executed. The calculation is performed only when the excitation validity flag is True; otherwise, the vector is set to zero. The covariance matrix from the previous time step is retrieved from memory. The regression feature vector at the current time Perform matrix and vector multiplication operations: This step utilizes the parameter uncertainty information contained in the covariance matrix to perform a weighted projection on the current physical eigenvectors. If If the variance of a certain parameter is large (high uncertainty), the feature of that dimension will be amplified in the product result; conversely, it will be reduced. Continuing the example from step S2, the feature vector... At this point, the algorithm is in its initial running stage, and the covariance matrix... It is a diagonal matrix, with each diagonal element being 0.01 (representing a certain degree of uncertainty), i.e. The calculation process is as follows:

[0031] The obtained gain numerator vector The directional information of the eigenvectors is preserved, but their magnitudes are scaled by the covariance matrix.

[0032] Then, step S33 is performed to calculate the gain denominator scalar used for normalization. This scalar is a numerical value, and its calculation formula combines the quadratic form of the variable forgetting factor and the eigenvector:

[0033] The second half of the formula In fact, it is the result calculated in step S32. With feature vectors The dot product (or the squared magnitude of the eigenvectors in the covariance metric space). If the flag is False, the denominator is set to 1.0 to prevent subsequent division by zero errors. If True, the full calculation is performed. Continuing with the numerical example above, given... , For example, the variable forgetting factor calculated in step S31 =0.975. First, calculate the quadratic form term: =3605.3 Then add the forgetting factor: =0.975+3605.3=3606.275, this denominator value reflects the energy intensity of the input signal.

[0034] Finally, step S34 is performed, which involves dividing the vector by the scalar, i.e., using the gain numerator vector obtained in step S32. Divide by the scalar of the gain denominator obtained in step S33 The formula is The gain vector This is the final output of this step. It's a column vector with the same dimensions as the number of model parameters; in this example, it's 3-dimensional. It precisely quantifies the direction and step size along which the model parameters should respond to the prediction error at the current moment. Substitute the values ​​into the numerical calculation: Analysis of the calculation results reveals that the first element (corresponding to the heat load coefficient) has the largest gain value (0.001663), indicating that the current observation data contains the most information about the heat load. Therefore, the algorithm will primarily utilize the current error to correct the heat load-related parameters. The gain corresponding to the intercept term is extremely small, indicating that the current data contributes little to the correction of the fixed loss parameter. This result is consistent with physical facts, because in the input features... (600.0) is much larger than the other items.

[0035] In a preferred exemplary embodiment of this application, step S3 includes: S3-1, when the excitation validity flag is true, calculating the stochastic gradient adaptive forgetting factor of the gradient flow based on the posterior error of the previous time step to obtain the variable forgetting factor; S3-2, determining the gain numerator vector based on the covariance matrix and regression eigenvector of the previous time step; S3-3, determining the gain denominator scalar based on the quadratic form of the regression eigenvector with respect to the covariance matrix of the previous time step and in combination with the variable forgetting factor; S3-4, performing Kalman gain vector synthesis on the gain numerator vector and the gain denominator scalar to obtain the gain vector used to determine the weight of the current observation. It is worth noting that the processing methods for S3-2, S3-3, and S3-4 are the same as before, and are omitted here; the focus is on the implementation details of S3-1.

[0036] It is understandable that in the complex electromagnetic and mechanical environment of industrial sites, the data collected by sensors will inevitably be affected by non-Gaussian impulse noise. This large-amplitude instantaneous noise is captured as a huge posterior error in conventional algorithms. If the forgetting factor is adjusted simply based on an empirical heuristic formula based on the square of the posterior error, the algorithm will mistakenly identify such noise as a drastic change in the physical parameters of the equipment, thus drastically reducing the forgetting factor in an attempt to quickly track this pseudo-change. This misjudgment causes the algorithm to discard valuable historical steady-state information and instead fit meaningless random noise, leading to instantaneous expansion of the parameter covariance matrix and violent oscillations in the parameter estimates, ultimately resulting in unstable control output. Therefore, in order to endow the algorithm with the ability to intelligently distinguish between instantaneous noise and sudden changes in true parameters, this application introduces a gradient-flow stochastic gradient adaptive forgetting factor calculation method, upgrading the forgetting factor into a dynamic parameter that can be trained via gradient descent. By tracking the flow of error sensitivity to the forgetting factor, robust parameter identification is achieved.

[0037] Based on this, in a preferred exemplary embodiment of this application, step S3-1, when the excitation validity flag is true, calculates the stochastic gradient adaptive forgetting factor of the gradient flow based on the posterior error of the previous time step to obtain the variable forgetting factor, including: The gradient of the cost function with respect to the variable forgetting factor is calculated based on the posterior error from the previous time step and the sensitivity vector from the time step before that, and the variable forgetting factor is updated along the negative gradient direction. This step aims to use the idea of ​​stochastic gradient descent (SGD) to find the optimal value of the forgetting factor that minimizes the current prediction error. To achieve this calculation, a higher-order auxiliary variable parameter sensitivity vector is introduced. This mathematically quantifies the cumulative impact of minute changes in the forgetting factor on model parameters after being propagated through a complex RLS iterative formula. Based on this, an instantaneous cost function is defined, which is the square of the posterior error at the previous time step, and the partial derivative (gradient) of this cost function with respect to the variable forgetting factor is calculated using the chain rule. The physical meaning of this gradient is to indicate whether the forgetting factor should be increased or decreased to reduce prediction error. In specific calculations, the processing unit uses the posterior error from the previous time step. The regression feature vector of the previous time step And the sensitivity vector stored in the buffer at the time two steps prior, i.e., time t-2. Based on the provided technical corpus, the updated formula is as follows:

[0038] In the formula, and These are the forgetting factors for the current and previous time steps, respectively; The preset gradient descent learning rate is set to a minimum value, such as 10. -4 It is used to control the smoothness of the adjustment and prevent the forgetting factor from jumping due to excessive single error; For truncation functions, the calculation result is forced to be limited to a certain range. The closed interval is used, such as [0.95, 1.0], to ensure the stability of the algorithm. The terms in the formula... This approximates the gradient of the cost function. The effect of this step is that if the large error is caused by noise, the noise and the feature vector... Uncorrelated and random in direction, the expectation of this gradient term tends to 0 in the long run, making Maintaining a stable high-level state suppresses the influence of noise; if the cause of a large error is a sudden change in parameters, the gradient term will continue to point in the same direction, driving... The rapid descent allows the algorithm to quickly adapt to new operating conditions.

[0039] The sensitivity vector is recursively evolved based on the gain vector and covariance matrix from the previous time step to obtain the sensitivity vector from the previous time step. The purpose of this step is to provide a basis for the gradient calculation at the next time step, i.e., to update the higher-order auxiliary variables. . Mathematically, it is defined as the partial derivative of the model parameter estimate with respect to the forgetting factor. This quantifies the cumulative impact of minute changes in the forgetting factor on model parameters after being propagated through a complex RLS iterative formula. Based on the gain vector from the previous time step... Covariance matrix and regression feature vectors The state is updated according to the following state transition equation:

[0040] In the formula, This is the updated sensitivity vector from the previous time step; It is the identity matrix; the first term This reflects the attenuation characteristics of historical sensitivity information; the second item The source term represents how the current error and data characteristics stimulate new sensitivities. This step ensures that the algorithm can remember historical adjustment paths, making the adjustment of the forgetting factor no longer short-sighted but based on long-term parameter evolution trends. A concrete numerical example illustrates the above process: if an electromechanical device is operating at time t, the forgetting factor at the previous time t-1... =0.995. Preset learning rate. =0.0001, lower limit of forgetting factor =0.95. At the previous moment, the sensor detected a large power fluctuation, leading to a posterior error. =-5.0kW (predicted value is too high). The corresponding regression feature vector is (Flow-temperature product, thermal lift, intercept). The sensitivity vector stored in the buffer from the time step two moments ago is... First, gradient flow calculation is performed. Then, the scalar product is calculated. =600×0.01+23×0.01+1×0.001=6.231. Next, calculate the gradient update term: =-0.0001×(-5.0)×6.231=0.0031155. Update the forgetting factor: =0.995 + 0.0031155 = 0.9981155. After truncation (not exceeding 1.0), the final result is... ≈0.998. In this example, although a large error of -5.0kW occurred, the gradient calculation results suggested that the forgetting factor should be increased (or kept high). The algorithm judged that the error might originate from noise or that the model was not misfitting, and therefore did not blindly reduce the forgetting factor, thus avoiding overfitting. Subsequently, the known gain vector from the previous time step was used... Covariance Matrix renew The value is stored in a buffer for calculation at time t+1. Based on the variable forgetting factor obtained by this preferred implementation method, since it eliminates the interference of instantaneous noise during the calculation process and retains the sensitivity to the sudden change of the real parameters, the gain vector calculated in subsequent steps S3-2-S3-4 is more accurate and reliable, avoiding the spurious expansion of the covariance matrix caused by noise misjudgment.

[0041] In step S4, the model parameters of the previous time step are updated and projected under physical constraints based on the gain vector, the variable forgetting factor, and the observation scalar to obtain the corrected model parameters and the covariance matrix at the current time step. It is understandable that in traditional electromechanical equipment control model applications, the problem of deviation between mathematical optimality and physical reality is commonly faced. Since parameter identification algorithms such as recursive least squares or Kalman filtering are essentially data fitting processes based on statistical principles, their core objective is to minimize the sum of squared prediction errors. Under ideal conditions with sufficient data, mathematical fitting results can usually reflect physical reality. However, in the complex environment of industrial sites, when encountering sudden large-amplitude measurement noise from sensors, singular values ​​generated by data transmission packet loss, or sparse data samples due to extreme operating conditions of the equipment, purely mathematical algorithms may calculate model parameters that violate physical common sense in order to forcibly fit these abnormal data points. For example, they may derive negative heat transfer coefficients, compressor efficiencies exceeding 100%, or negative mechanical friction losses. If mathematically correct but physically absurd model parameters are directly used for subsequent control decisions, they can easily lead to dangerous controller outputs, failing to save energy and potentially damaging equipment. Therefore, after performing the mathematical calculations for parameter updates using gain vectors, a safety valve based on physical mechanisms needs to be introduced. This involves using a projection correction mechanism under physical constraints to forcibly limit the updated parameters to a reasonable physically feasible domain, ensuring that the model evolves within the framework of physical laws. This results in corrected model parameters that possess both data adaptability and conform to objective physical laws.

[0042] Figure 3 This is a flowchart of step S4 in the energy-saving control method for electromechanical equipment based on intelligent monitoring according to an embodiment of this application. Figure 3 As shown, in an exemplary embodiment of this application, step S4 includes: S41, using the model parameters of the previous time step to pre-estimate the regression feature vector to obtain the calculation result, and subtracting the calculation result from the observed scalar to obtain the prior error reflecting the bias of the old model; S42, based on the gain vector and the prior error, correcting the model parameters of the previous time step to obtain the preliminary updated parameters, and using the variable forgetting factor to update the covariance matrix of the previous time step to obtain the covariance matrix of the current time step; S43, performing physical feasible region projection correction on each element of the preliminary updated parameters to obtain the corrected model parameters.

[0043] The operation method is detailed below: At this time, the processing unit has obtained the observation scalar at the current time t. That is, the measured active power is 120.2kW, and the regression eigenvector is... ,Right now The gain vector calculated in step S3 ,Right now and the variable forgetting factor That is, 0.975. In addition, the algorithm's memory buffer also stores the model parameters from the previous time step. covariance matrix at the previous time step The previous time-step model parameters here are a column vector with the same dimensions as the feature vector, containing coefficients describing the physical characteristics of the device at the end of the previous sampling period. When the device first runs, these parameters are initialized to the device's factory nameplate data or offline calibration values; in subsequent runs, they are the corrected results output from step S4 of the previous time-step. As in this embodiment, the previous time-step model parameters... The first element, 0.1900, represents the energy consumption coefficient corresponding to the product of temperature difference per unit flow rate (related to the reciprocal of cooling efficiency), the second element, 0.2000, represents the energy consumption coefficient corresponding to the unit thermal lift, and the third element, 2.0000, represents the fixed mechanical loss intercept.

[0044] First, implement step S41. Call the model parameters from the previous time step. The regression feature vector at the current time Perform a preliminary calculation. This is essentially a simulation of what the current energy consumption should be if the physical characteristics of the equipment remain unchanged. The calculation formula is the dot product of the eigenvector and the parameter vector: Substitute specific values ​​into the calculation: =120.6, the calculation results show that the energy consumption predicted based on the old parameters is 120.6 kW. Subsequently, the actual observed scalar values ​​were read. =120.2kW, and the difference between the two is used to obtain the prior error. : =120.2-120.6=-0.4, the error value is negative (-0.4kW), which indicates that the actual observed value is lower than the model prediction value. This means that the actual operating efficiency of the equipment is higher than expected by the old model, or that the parameters of the old model are too large (possibly because the equipment has just undergone cleaning and maintenance, or the efficiency has increased due to the change of operating conditions). Therefore, the parameters need to be corrected downward.

[0045] Next, step S42 is performed. First, the parameters are updated using the gain vector calculated in step S3. Prior error Mapping to the parameter space yields preliminary update parameters. The calculation formula is: Substitute the numerical values ​​into the calculation:

[0046] Since the prior error is negative, all parameters receive a slight downward adjustment. Furthermore, because the first element in the gain vector is the largest, the correction magnitude for the first parameter (heat load coefficient) is also the largest, which aligns with the data distribution. The result obtained at this point... It is the result of purely mathematical operations. At the same time, the covariance matrix at the current moment needs to be updated. This is to prepare for the calculation at the next time step. Its update formula is based on the derivation of the Recursive Least Squares (RLS) algorithm:

[0047] in the formula This is the identity matrix. The physical meaning of this step is that the terms within the parentheses represent the reduction in uncertainty (covariance) of the model parameters due to the acquisition of new observational information; and dividing by the variable forgetting factor... In this example, a value of 0.975 serves as the function of an expanded covariance matrix. Because... Dividing by 1 will increase the value of the matrix elements, which means that the algorithm artificially adds a little uncertainty to prevent the covariance matrix from approaching zero over time, i.e., the algorithm becomes overconfident and stops learning, thus maintaining the algorithm's sensitivity to future parameter changes.

[0048] Finally, step S43 is performed to update the initial parameters. Performing physical feasible region projection correction is to prevent the algorithm from diverging under extreme noise. First, a predefined set of physical parameter feasible regions is loaded. This set defines the physical lower bound for each parameter i. and the Upper Realm These boundary values ​​are determined based on thermodynamic laws and equipment design specifications. For example, for the first parameter... (Heat load energy consumption coefficient), which physically cannot be less than 0 (otherwise negative energy consumption will occur), nor can it be too large (limited by the Carnot cycle efficiency limit), is set. =0.15, =0.25. For the second parameter... (Thermal lift coefficient), set =0.10, =0.30. For the third parameter... (Fixed losses), setting =0.5, =5.0. Iterate through each element in the initial update parameter vector, applying either the hard diagonal projection algorithm or the logical truncation method. Check the first item: =0.1893348. Determine if 0.15 ≤ 0.1893348 ≤ 0.25; the result is within the range, so retain the original value. Check the second item: =0.1999744. Determine if 0.10 ≤ 0.1999744 ≤ 0.30; the result is within the range, so retain the original value. Check the third item: =1.9999988. Determine 0.5 ≤ 1.9999988 ≤ 5.0; the result is within the range, so retain the original value. In an abnormal situation, due to instantaneous strong noise interference, the preliminary parameters calculated in step S42 may be affected. It became 0.12 (below the lower bound of 0.15). At this point, the projection algorithm will force... Change to lower bound value =0.15. This processing is equivalent to vertically projecting the estimated points that fall into the non-physical region back onto the boundary of the physical feasible region in the parameter space. After this round of verification, the final corrected model parameters are output. In this example, it is Approximately This set of parameters not only incorporates the latest observational data to the greatest extent mathematically, but is also strictly constrained within a reasonable range physically.

[0049] In step S5, an adaptive energy efficiency prediction model is constructed based on the corrected model parameters, and the optimal setpoint is obtained by global optimization of the adaptive energy efficiency prediction model under the operating condition constraints corresponding to the filtered dataset. Accordingly, after obtaining the corrected model parameters that can accurately reflect the current aging state and physical characteristics of the equipment, the core task of energy-saving management shifts from model identification to the optimization control stage. Traditional control strategies are often based on PID regulation or fixed rule tables. Although these methods can maintain the stability of process variables, they cannot actively find the operating point with the lowest energy consumption under multiple constraints of load demand and equipment safety. For example, under the premise of meeting the same cooling capacity demand, it can be achieved by increasing the flow rate and reducing the temperature difference, or by reducing the flow rate and increasing the temperature difference, but the total energy consumption of the water pump and compressor corresponding to these two combinations is completely different. Moreover, due to the drift of equipment physical parameters over time (such as a decrease in heat exchange efficiency), the optimal operating point originally calibrated at the factory may have become an inefficient zone. Based on this, this application constructs an adaptive energy efficiency prediction model that incorporates the latest physical parameters and performs real-time global optimization within a strict safety and process constraint space, thereby calculating the true energy-saving optimal solution at the current moment.

[0050] In one exemplary embodiment of this application, step S5 includes: S51, constructing an adaptive objective function that can sense the current aging level of the device based on the corrected model parameters; S52, constructing a feasible region of the operational variables based on the device safety specifications and the load requirements in the filtered dataset to obtain a feasible region constraint set containing inequalities and equality constraints; S53, performing a minimum search on the adaptive objective function within the feasible region constraint set to obtain the optimal setpoint.

[0051] The operation method is detailed below: At the current time t, after the correction in step S4, the obtained model parameter vector is: These correspond to the heat load coefficient, heat lift coefficient, and fixed loss intercept, respectively. At this point, it is necessary to find a set of optimal controllable variable values ​​that minimizes the predicted total energy consumption of the equipment under the current environmental and load conditions. First, implement step S51. First, separate the roles of the variables in the physical model. Define those quantities that cannot be changed within the current control cycle and can only be passively accepted as uncontrollable environmental state variables, denoted as... This mainly includes the ambient dry-bulb temperature obtained directly from the filtered dataset. Examples include 30.0℃ and load requirements that the process must meet (such as flow rate or cooling capacity requirements). Quantities that can be adjusted by the controller are defined as controllable operating variables, denoted as... In this embodiment, it mainly refers to the working fluid outlet temperature setpoint. ,Right now The target value. Although the traffic It is also a variable, but determined by the terminal load or coupled with the temperature difference. In this simplified model, the main optimization variable is the outlet water temperature setpoint. Next, the function is restructured. Review the energy consumption equation structure in step S2: In this step, the corrected parameters will be... Substitute into the equation, and Treated as the independent variable to be solved For example, the current working fluid inlet temperature. =12.0℃, flow rate =120.0m 3 / h (These are the current states) Ambient temperature =30.0℃. Therefore, the energy consumption prediction function is constructed as about ,Right now Functions: , simplify this formula: It is important to note that the above linear equations only describe a specific local linearization point. In more complex practical nonlinear models, the compressor power and... Often exhibiting a non-linear relationship (e.g.) The lower the value, the lower the COP (Coefficient of Performance), and the higher the power output. In this explanation, to maintain consistency with the linear feature transformation in step S2, the linear form is used. With more refined physical modeling (introducing quadratic features), the actual constructed objective function would be: The function form reflects: The lower the outlet water temperature, the greater the cooling capacity, but the more work the compressor does (the first item, pressure ratio energy consumption, increases sharply). The higher the pressure, the smaller the temperature difference. Although the pressure ratio decreases, it may not be able to meet the cooling capacity requirements (an implicit constraint). This function Because the latest version is used Therefore, it has already sensed the current aging level of the equipment (for example, if...). An increase in the value indicates a decrease in heat exchange efficiency, meaning that more energy is required to achieve the same temperature difference.

[0052] Next, proceed to step S52. This requires combining the equipment safety specifications with the load requirements of the filtered data set. The equipment safety specifications are stored in the controller's configuration database and contain the hard limits specified at the time of manufacture. For example: 1. Anti-freeze protection: working fluid outlet temperature The temperature must not fall below 4.0℃, otherwise the evaporator will freeze and crack. ≥4.0. 2. Compressor safe operating range: The outlet temperature should not be too high, otherwise effective cooling will be impossible. The upper limit is set at 15.0℃. ≤15.0. The load requirement stems from current process requirements. For example, based on the filtered dataset collected in step S1, the current return water temperature... =12.0℃, and the terminal process requires that the water supply temperature must be able to remove at least 500kW of heat. According to the formula If the flow rate is fixed at 120, and ≈1.163kW / (m3·K), then It must be greater than 500 / (1.163×120)≈3.58℃. This means ≥3.58, that is Combining the above inequality constraints, we obtain the operational variables. Feasible domain: In addition, there may be equality constraints, such as in some precision temperature control scenarios where the rate of temperature change is controlled by a flow valve, which is simplified here to the aforementioned interval constraints. This set of feasible region constraints This ensures that, regardless of the optimization results, the equipment will not operate in hazardous areas or in a state that fails to meet process requirements.

[0053] Finally, implement step S53. The adaptive objective function constructed in step S51 is then... The set of feasible region constraints defined in step S52 As input, it invokes a nonlinear programming solver such as Sequential Quadratic Programming (SQP) or the interior-point method. The mathematical essence of this solution process can be described as solving a constrained extremum problem: The solver first calculates the gradient of the objective function at the current search point. The search is performed iteratively along the direction of the fastest energy decrease (i.e., the negative gradient direction). In this embodiment, the simplified objective function derived in step S51 is used. (or may include higher-order terms reflecting pressure ratio characteristics), which are related to the outlet water temperature. The first derivative is negative, which physically means that as the set temperature increases... As the pressure ratio of the compressor increases, the overall energy consumption decreases, leading to a downward trend in total energy consumption. However, this decrease is not unlimited; the solver's search path quickly reaches the boundary of the feasible region calculated in step S52. Specifically, the solver searches within the interval [4.0, 8.42]. When the algorithm attempts to further improve... When reducing energy consumption, the upper limit of the inequality constraint determined by the process load requirements will be triggered. ≤8.42. At this point, the constraint becomes valid or active. The solver verifies the Kuhn-Tucker (KKT) conditions on the boundary using the Lagrange multiplier method, determining that, under the premise of satisfying all safety and load constraints, the point of minimum energy consumption lies precisely on the upper boundary of the feasible region. Therefore, the algorithm converges and outputs the optimal setpoint. =8.42℃. This calculation result has significant physical and engineering implications: it shows that under the current operating conditions of a return water temperature of 12℃ and an ambient temperature of 30℃, in order to meet the cooling demand of 500kW at the terminal, the equipment does not need to rigidly maintain 7.0℃ as in traditional control methods, but can safely raise the outlet water temperature to 8.42℃. Through this adjustment, the pressure ratio that the compressor needs to overcome is reduced, thereby minimizing energy consumption without sacrificing the process compliance rate. Ultimately, this 8.42℃, obtained through physical model correction and full-constraint optimization calculation, was established as the optimal setpoint.

[0054] In step S6, the optimal setpoint is sent to the electromechanical equipment for control execution. That is, in the previous steps, the global optimization algorithm based on physical constraints has calculated the theoretically optimal operating setpoint under the current operating conditions, namely the optimal setpoint of the working fluid outlet temperature, 8.42℃. However, this optimal solution currently only exists in the memory of the computing unit and is an abstract mathematical floating-point value. Electromechanical equipment in the industrial field, such as the PLC controller of a chiller unit or the DDC module of a variable frequency pump, cannot directly understand the mathematical results output by the advanced algorithm. They rely on specific industrial communication protocols and strictly defined electrical signal standards to drive the physical actuators. Furthermore, large electromechanical equipment has significant thermal and mechanical inertia; from receiving the command to the actual change in physical state and reaching a new steady state, a non-negligible physical response process is required. If the next round of data acquisition is performed immediately after the command is issued, non-stationary data in a transient state will be collected, which will seriously interfere with subsequent model parameter identification and may even lead to control oscillations. Therefore, in order to translate the energy-saving decisions on the computing side into actual actions on the physical side and to ensure the correct timing logic of the closed-loop control, this application ultimately implements a standardized instruction issuance and state synchronization process.

[0055] In one exemplary embodiment of this application, step S6 includes: mapping the optimal setpoint to an industrial bus register address and sending it to the device controller, and triggering a control execution completion signal after waiting for a preset response stabilization time.

[0056] The operation method is detailed below: First, the protocol mapping process is executed. This process establishes a translation bridge between the algorithm domain and the control domain. A preset industrial bus point table is then invoked, which defines the unique identifier of each physical control object in the communication network. In this embodiment, the optimal setpoint is considered. =8.42℃. The point table indicates that the corresponding device controller (such as a Siemens S7-1500 PLC) supports the Modbus TCP communication protocol. The logical name of the target controlled object is the chiller outlet water temperature setpoint, and its register address in the Modbus protocol is 40105 (holding register area). The point table also specifies that the data format of this register is a 16-bit unsigned integer (UINT16), and the scaling factor is 0.1. This means that the controller internally stores the temperature value in integer form, with a value of 1 representing 0.1℃. Based on this rule, a numerical conversion calculation is performed: the floating-point number 8.42 is divided by the scaling factor 0.1 to obtain 84.2, and then rounded to 84 according to the rounding principle. If the controlled object is the operating frequency of the variable frequency water pump (such as the optimal solution obtained in the previous optimization step being frequency), and the field actuator receives a 4-20mA analog control signal, then a digital-to-analog conversion mapping is required. For example, given that the inverter's frequency range of 0-50Hz corresponds to the DAC's (digital-to-analog converter) digital input range of 0-4095 (12-bit precision), if the optimal frequency is 35Hz, then the mapped value is (35 / 50)×4095=2866. In this case, the final generated data packet to be sent contains: target device IP address, function code 06 (write to a single register), register address 40105, and write data value 84.

[0057] Subsequently, the command issuance phase begins, establishing a connection with the field edge computing gateway or directly with the equipment controller via the industrial Ethernet interface. The encapsulated data packet is sent to the equipment controller via the TCP / IP protocol stack. Upon receiving the message, the equipment controller parses the write command and updates the value 84 to its internal memory address 40105. The controller's logic program periodically scans this memory address. When it detects that the setpoint has changed from 70 (7.0℃) to 84 (8.42℃), it immediately triggers the internal PID control loop. For chiller units, the internal controller compares the current actual outlet water temperature with the new setpoint. If the actual temperature is significantly lower than the new setpoint, it determines that the load demand has decreased. Subsequently, the controller outputs physical commands to drive the compressor's guide vane mechanism to reduce its opening or reduce the speed of the variable frequency compressor to reduce the refrigerant flow, causing the outlet water temperature to gradually rise and approach 8.42℃.

[0058] Finally, the execution of the action and the stabilization wait are performed. After the command is successfully issued and a write success confirmation frame is received from the controller, the process does not immediately end; instead, a software countdown timer is started. The duration of this timer is set to a preset response stabilization time. This time parameter is determined based on the thermodynamic time constant of the equipment. During the system commissioning phase, technicians will apply a step response test to the equipment, for example, by abruptly changing the set temperature by 1°C and recording the time required for the outlet water temperature to stabilize within the target value ±0.1°C error band. For chillers with high inertia, this process may take 300 to 600 seconds. If testing confirms 300 seconds (5 minutes), during the timer's operation, upper-level data acquisition and parameter identification tasks are suspended, or the data acquired during this period is marked as transitional data and not used for model updates to prevent the algorithm from learning unstable physical characteristics. When the 300-second countdown ends, the physical equipment has completed the adjustment process, and the actual outlet water temperature stabilizes around 8.42°C, and the system enters a new steady state. At this point, a control execution completion signal is triggered. This signal serves as an interrupt or status flag, notifying the main control program that the current control cycle has been closed, and that step S1 can be safely activated to begin a new round of sampling, identification, and optimization, thereby achieving cyclical adaptive energy-saving control.

[0059] In summary, the energy-saving control method for electromechanical equipment based on intelligent monitoring, as described in the embodiments of this application, addresses the problems of physical characteristic drift, static model mismatch, and unreliable output of unconstrained black-box models caused by equipment aging in existing technologies. First, the original sensor data is filtered and the excitation validity is determined to select data with parameter identification value. Based on this, a regression feature vector representing the current physical state of the equipment is established through physical feature extraction and linearization transformation. Then, the algorithm gain is dynamically adjusted using a variable forgetting factor based on posterior error to track the model parameters over time, and a projection correction mechanism under physical constraints is introduced to force the updated parameters to always remain within a reasonable physical feasible domain. Finally, an adaptive energy efficiency model based on real-time correction performs global optimization under operating condition constraints. By constructing a closed loop of perception-identification-constraint-optimization, this ensures that the control model can accurately follow the aging process of the equipment, thereby achieving precise energy-saving control throughout its entire lifecycle.

[0060] Figure 4 This is a block diagram of an energy-saving control system for electromechanical equipment based on intelligent monitoring, according to an embodiment of this application. Figure 4As shown, the energy-saving control system 100 for electromechanical equipment based on intelligent monitoring according to an embodiment of this application includes: an electromechanical equipment data processing module 110, used to perform low-pass filtering and excitation validity prediction on the raw sensor data stream of the collected electromechanical equipment to obtain an excitation validity flag and a filtered dataset; a feature extraction and linearization transformation module 120, used to perform physical feature extraction and linearization transformation on the filtered dataset to obtain a regression feature vector and an observation scalar used to characterize the physical state at the current moment; and a gain vector generation module 130, used to determine a variable forgetting factor based on the posterior error of the previous moment when the excitation validity flag indicates validity, and combine it with regression... The eigenvector and the covariance matrix of the previous time step determine the gain vector that can suppress noise interference; the parameter update and projection correction module 140 is used to update and project the model parameters of the previous time step under physical constraints based on the gain vector, the variable forgetting factor and the observation scalar to obtain the corrected model parameters and the covariance matrix of the current time step; the optimal setpoint determination module 150 is used to construct an adaptive energy efficiency prediction model based on the corrected model parameters, and to perform global optimization on the adaptive energy efficiency prediction model under the operating condition constraints corresponding to the filtered dataset to obtain the optimal setpoint; the control execution module 160 is used to send the optimal setpoint to the electromechanical equipment for control execution.

[0061] Here, those skilled in the art will understand that the specific operations of each step in the above-described intelligent monitoring-based energy-saving control system for electromechanical equipment have been referenced above. Figures 1 to 3 The energy-saving management method for electromechanical equipment based on intelligent monitoring has been described in detail, and therefore, its repeated description will be omitted.

Claims

1. A method for energy-saving management and control of electromechanical equipment based on intelligent monitoring, characterized in that, include: S1: Perform low-pass filtering and excitation validity prediction on the raw sensor data stream of the collected electromechanical equipment to obtain the excitation validity flag and the filtered dataset; S2: Perform physical feature extraction and linearization transformation on the filtered dataset to obtain the regression feature vector and observation scalar used to characterize the physical state at the current moment; S3: When the excitation valid flag indicates that the excitation is valid, the variable forgetting factor is determined based on the posterior error of the previous time step, and the gain vector that can suppress noise interference is determined by combining the regression feature vector and the covariance matrix of the previous time step. S4: Based on the gain vector, variable forgetting factor and observation scalar, the model parameters of the previous time step are updated and projected under physical constraints to obtain the corrected model parameters and the covariance matrix of the current time step. S5: Construct an adaptive energy efficiency prediction model based on the corrected model parameters, and perform global optimization of the adaptive energy efficiency prediction model under the working condition constraints corresponding to the filtered dataset to obtain the optimal set point. S6: Send the optimal setpoint to the electromechanical equipment for control execution.

2. The energy-saving control method for electromechanical equipment based on intelligent monitoring according to claim 1, characterized in that, The raw sensor data stream includes three-phase voltage, three-phase current, active power, power factor, working fluid inlet temperature, working fluid outlet temperature, working fluid flow rate, system pressure, operating frequency, valve opening, start / stop status, ambient dry-bulb temperature, and relative humidity. Step S1 includes: Multi-channel signal synchronization and digital low-pass filtering are performed on the raw sensor data stream to obtain the filtered dataset. The instantaneous rate of change of the filtered dataset from the previous time step and the filtered dataset from the current time step is estimated based on Euclidean distance to obtain a data rate of change index that reflects the degree of system disturbance. Logical comparison and judgment are performed between the data change rate index and the preset dead zone threshold to obtain the effective excitation flag.

3. The energy-saving control method for electromechanical equipment based on intelligent monitoring according to claim 1, characterized in that, Step S2 includes: The independent state variables and dependent variables required to construct the energy consumption equation are parsed and extracted from the filtered dataset to obtain a structured set of extracted variables. A mechanism-based nonlinear feature transformation is performed on the independent state variables in the extracted variable group to obtain a list of transformed feature terms; The transformed feature list is mapped and assembled into a linearized regression feature vector, and the dependent variable in the extracted variable set is mapped into an observed scalar.

4. The energy-saving control method for electromechanical equipment based on intelligent monitoring according to claim 1, characterized in that, Step S3 includes: When the excitation validity flag is true, the variable forgetting factor is determined based on the posterior error of the previous time step. Based on the covariance matrix and regression eigenvector of the previous time step, determine the gain numerator vector; Based on the quadratic form of the regression eigenvector with respect to the covariance matrix of the previous time step and combined with the variable forgetting factor, the scalar of the gain denominator is determined; Kalman gain vector synthesis is performed on the gain numerator vector and the gain denominator scalar to obtain the gain vector used to determine the weight of the current observation.

5. The energy-saving control method for electromechanical equipment based on intelligent monitoring according to claim 1, characterized in that, Step S3 includes: When the excitation valid flag is true, the stochastic gradient adaptive forgetting factor of the gradient flow is calculated based on the posterior error of the previous time step to obtain the variable forgetting factor. Based on the covariance matrix and regression eigenvector of the previous time step, determine the gain numerator vector; Based on the quadratic form of the regression eigenvector with respect to the covariance matrix of the previous time step and combined with the variable forgetting factor, the scalar of the gain denominator is determined; Kalman gain vector synthesis is performed on the gain numerator vector and the gain denominator scalar to obtain the gain vector used to determine the weight of the current observation.

6. The energy-saving control method for electromechanical equipment based on intelligent monitoring according to claim 5, characterized in that, When the activation validity flag is true, the stochastic gradient adaptive forgetting factor of the gradient flow is calculated based on the posterior error of the previous time step to obtain the variable forgetting factor, including: The gradient of the cost function with respect to the variable forgetting factor is calculated based on the posterior error of the previous time step and the sensitivity vector of the time step before that, and the variable forgetting factor is updated along the negative gradient direction. The sensitivity vector is recursively evolved based on the gain vector and covariance matrix of the previous time step to obtain the sensitivity vector of the previous time step.

7. The energy-saving control method for electromechanical equipment based on intelligent monitoring according to claim 1, characterized in that, Step S4 includes: The regression feature vector is estimated using the model parameters from the previous time step to obtain the calculation results, and the difference between the calculation results and the observed scalar is calculated to obtain the prior error reflecting the bias of the old model. Based on the gain vector and prior error, the model parameters of the previous time step are corrected to obtain the initial updated parameters, and the covariance matrix of the previous time step is updated using the variable forgetting factor to obtain the covariance matrix of the current time step. Physically feasible region projection correction is performed on each element of the initially updated parameters to obtain the corrected model parameters.

8. The energy-saving control method for electromechanical equipment based on intelligent monitoring according to claim 1, characterized in that, Step S5 includes: Based on the corrected model parameters, an adaptive objective function capable of sensing the current aging level of the device is constructed. Based on the equipment safety specifications and the load requirements in the filtered dataset, the feasible region of the operational variables is constructed to obtain a feasible region constraint set containing inequalities and equality constraints. The optimal setpoint is obtained by performing a minimum search on the adaptive objective function within the feasible region constraint set.

9. The energy-saving control method for electromechanical equipment based on intelligent monitoring according to claim 1, characterized in that, Step S6 includes: mapping the optimal setpoint to an industrial bus register address and sending it to the device controller, and triggering a control execution completion signal after waiting for a preset response stabilization time.

10. An energy-saving control system for electromechanical equipment based on intelligent monitoring, characterized in that, include: The electromechanical equipment data processing module is used to perform low-pass filtering and excitation validity prediction on the raw sensor data stream of the collected electromechanical equipment to obtain the excitation validity flag and the filtered dataset. The feature extraction and linearization transformation module is used to extract physical features and perform linearization transformation on the filtered dataset to obtain regression feature vectors and observation scalars that characterize the physical state at the current moment. The gain vector generation module is used to determine the variable forgetting factor based on the posterior error of the previous time step when the excitation valid flag indicates that the excitation is valid, and to determine the gain vector that can suppress noise interference by combining the regression feature vector and the covariance matrix of the previous time step. The parameter update and projection correction module is used to update and project the model parameters of the previous time step under physical constraints based on the gain vector, the variable forgetting factor and the observation scalar to obtain the corrected model parameters and the covariance matrix of the current time step. The optimal setpoint determination module is used to construct an adaptive energy efficiency prediction model based on the corrected model parameters, and to perform global optimization of the adaptive energy efficiency prediction model under the working condition constraints corresponding to the filtered dataset to obtain the optimal setpoint. The control execution module is used to send the optimal setpoint to the electromechanical equipment for control execution.

Images