A method and apparatus for controlling a power system

Abstract

A method and apparatus for controlling equipment in a power system relates to data monitoring technology in nuclear power system control. The aim is to address the problem that existing data repair technologies are insufficient when industrial sensor data is missing, thus affecting equipment control and increasing safety hazards. The method includes: real-time acquisition of time-series data from industrial sensors and statistical analysis of the characteristics of each data column; matching a repair model based on the statistical characteristics and physical properties of each data column; generating a computational task queue based on the missing data rate; using the matched repair model to predict and fill in the missing locations of corresponding columns, generating a full-scale repair data matrix; constructing a mask matrix to record the coordinates of data where the filling operation occurs; fusing the full-scale repair data matrix with the mask matrix; and generating control commands based on the fused matrix and inputting them into the equipment to control its operation.

Description

Technical Field

[0001] This invention relates to power system control technology, specifically to data monitoring technology for nuclear power systems in nuclear power system control. Background Technology

[0002] During the operation of industrial equipment, it is necessary to collect real-time physical quantities such as temperature, pressure, flow rate, liquid level, and rotational speed through industrial sensors. These data are real-time representations of the equipment's operating status. For example, pressure reflects the fluid state in pipelines or containers, temperature reflects the heat exchange process, flow rate reflects the material conveying status, and liquid level reflects the material inventory in storage tanks or reactors.

[0003] The control system determines the operating status of the equipment based on sensor data and then performs further control. However, uncontrollable factors such as electromagnetic interference, signal transmission packet loss, sensor aging or short-term failures, or equipment start-up and shutdown can lead to missing sensor data. This makes the controller's judgment of the equipment's operating status inaccurate and generates incorrect control commands, affecting the safe operation of the equipment. Therefore, it is necessary to estimate and compensate for missing measurements based on the dynamic changes of sensor timing signals, and input the compensated state quantities into the industrial control system to adjust actuators such as valve opening, pump frequency, heating power, cooling water flow, and fan speed.

[0004] Existing processing methods fall short in addressing this challenge. While traditional statistical imputation methods (such as mean imputation and linear interpolation) are computationally simple, they often obscure the fluctuation characteristics of data when dealing with industrial parameters with complex dynamic trends, creating "pseudo-smooth" curves that violate physical laws. For example, simple interpolation cannot reproduce the instantaneous pressure changes caused by valve opening and closing in fluid pipelines, and this loss of physical information can easily mislead subsequent fault diagnosis models.

[0005] With the introduction of composite deep learning, remediation methods based on single deep learning models (such as RNNs) have become increasingly popular. However, feedback from practical applications shows that these models suffer from a significant adaptability paradox: complex neural networks perform excellently when processing massive amounts of nonlinear data, but often overfit when faced with simple linear parameters or small sample data, and consume enormous computational resources; conversely, simple filtering algorithms cannot cope with complex conditions involving long-term time-series dependencies. Furthermore, most existing remediation tools are black-box in nature, with the remediated data being mixed with the original data, lacking an intuitive audit trail mechanism. This makes it impossible for engineers to distinguish between actual measurements and algorithmic estimates during review, which is unacceptable in the rigorous field of industrial safety.

[0006] In summary, in the process of industrial equipment control, the data repair effect of existing processing methods for missing industrial sensor data is not ideal, which not only affects equipment control but also increases safety hazards. Summary of the Invention

[0007] The purpose of this invention is to solve the above-mentioned problems in the prior art and provide a method and apparatus for controlling equipment in a power system, which does not affect the control of the equipment even if sensor data is missing, thereby reducing safety risks.

[0008] A device control method for a power system according to the present invention includes: real-time acquisition of time-series data from industrial sensors, and statistical analysis of the statistical characteristics of each data column, the statistical characteristics including the missing rate and the maximum consecutive missing length; matching a repair model for each data column based on the statistical characteristics and physical properties of each data column, the repair model including a Kalman filter model suitable for linear data columns containing Gaussian noise, a long short-term memory network model suitable for nonlinear data columns with long time-series dependencies, and an unscented Kalman filter model suitable for strongly nonlinear data columns; sorting all columns to be repaired according to the missing rate to generate a computational task queue, and predicting and filling the missing positions of the corresponding columns using the matched repair model according to the computational task queue to generate a full repair data matrix; constructing a Boolean mask matrix with the same structure as the original data column matrix to record the data coordinates of the filling operation; fusing the full repair data matrix with the mask matrix to obtain a fusion matrix; generating a state vector based on the fusion matrix, generating control commands based on the state vector, and inputting the control commands to the device to control the device operation.

[0009] Optionally, the method further includes: using the mask matrix to drive a rendering engine to apply visual saliency markers to the data units of the repaired region in the fusion matrix.

[0010] Optionally, the process noise covariance of the Kalman filter model for:

[0011]

[0012] For the first The fundamental process noise covariance of physical quantities of each sensor; This is the missing rate weighting coefficient; This represents the missing rate of the data column; The weighting coefficient is the length of consecutive missing characters. Indicates as of the date The current consecutive missing length at each moment; Indicates the preset reference missing length.

[0013] Optionally, the final repair value of the Kalman filter model for:

[0014]

[0015] in, and The first The minimum and maximum allowable values ​​of the physical quantities corresponding to each industrial sensor. For the first The sensor at the first The repair value at each moment.

[0016] Optionally, the input feature vector of the Long Short-Term Memory network model for:

[0017]

[0018] in, This represents the historical valid values ​​or repaired values ​​prior to the missing points. This indicates the equipment control quantity at the previous moment. This indicates the length of consecutive missing data up to the previous time step. For missing indicator variables, A value of 1 indicates that the previous time step was a valid measured value. A value of 0 indicates that the value was missing in the previous time step; For the first The attribute encoding of physical quantities corresponding to each sensor.

[0019] Optionally, the long short-term memory network model incorporates a physical constraint loss function during training and model updates. :

[0020]

[0021] in, For the prediction error of effective observation points, Indicates the first The sensor at the first The measured values ​​at each sampling time; To control response constraints; Indicates the effect of the change in control quantity on the first The expected impact of each physical quantity; For boundary penalty terms; and These are the weighting coefficients; This represents the total number of sampling points in the parameter column; This is the boundary penalty function.

[0022] Optionally, the nonlinear state-space model of the unscented Kalman filter model is expressed as:

[0023]

[0024]

[0025] Subscript express Time, subscript Indicates the first One sensor, Indicates the first The sensor at the first The state vector of the physical quantity at each moment. This represents a nonlinear state transition function, used to describe how the device state changes with time and control input; It is a nonlinear observation function used to describe the relationship between sensor measurements and internal states; for Time-based control input; This is process noise; To observe noise; Indicates the first One sensor in The actual sampled value at that moment.

[0026] Optionally, the method further includes: generating the first consecutive missing length based on the mask matrix and the current consecutive missing length. A sensor or state variable in Reliability of Time Control The control commands are adjusted based on the control confidence level to obtain the final control commands. :

[0027]

[0028] in, This indicates the control commands generated based on the repair data. This indicates the security control strategy.

[0029] This invention also provides a device for controlling a power system, comprising: a data feature extraction unit, adapted to acquire time-series data from industrial sensors in real time and statistically analyze the statistical features of each data column, the statistical features including the missing rate and the maximum consecutive missing length; a repair model matching unit, adapted to match a repair model for each data column based on the statistical features and physical properties of each data column, the repair model including a Kalman filter model suitable for linear data columns containing Gaussian noise, a long short-term memory network model suitable for nonlinear and long-term time-dependent data columns, and an unscented Kalman filter model suitable for strongly nonlinear data columns; and a queue construction and repair unit, adapted to... The missing data rate is used to sort all columns to be repaired, generating a computational task queue. Based on the computational task queue, a matching repair model is used to predict and fill in the missing positions of the corresponding columns, generating a full repair data matrix. A mask matrix construction unit is adapted to construct a Boolean mask matrix with the same structure as the original data column matrix, used to record the data coordinates where the filling operation occurs. A data fusion unit is adapted to fuse the full repair data matrix with the mask matrix to obtain a fusion matrix. A control unit is adapted to generate a state vector based on the fusion matrix, generate control commands based on the state vector, and input the control commands to the device to control the device operation.

[0030] The present invention also provides an electronic device, comprising: a memory; a processor; and a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the device control method of the power system described above.

[0031] The present invention has the following effects:

[0032] First, it achieves an optimal balance between repair accuracy and computational efficiency. Through an adaptive routing mechanism, it avoids using complex LSTM networks to process simple linear data, and also avoids using simple interpolation methods to handle complex nonlinear trends, ensuring that each type of data is matched with the most cost-effective algorithm.

[0033] Second, it possesses strong engineering applicability and security. Compared to pure script tools, the visual audit view provided by this invention allows engineers to identify man-made data at a glance, greatly reducing the decision-making risks caused by data tampering or misoperation, and providing a security barrier for the secondary use of industrial data.

[0034] Third, it significantly lowers the barrier to data governance. Through integrated GUI interaction and fully automated batch processing logic, users do not need a deep programming or algorithm background; they can complete professional-grade data repair work with just a simple file loading operation, significantly improving the efficiency of data governance in industrial settings.

[0035] This invention uses multi-model adaptive missing data compensation to restore the industrial state variables at the missing time to estimated values ​​that conform to the dynamic laws and physical constraints of the equipment, and further participates in closed-loop control calculations. This enables the control system to maintain safe, continuous and stable control output even when sensors are missing for a short time or continuously. It is suitable for predictive maintenance, process optimization and energy efficiency management of industrial equipment. Attached Figure Description

[0036] Figure 1 This is a schematic flowchart of a device control method for a power system according to an embodiment of this application;

[0037] Figure 2 This is a schematic diagram of the data repair system architecture according to an embodiment of this application;

[0038] Figure 3 This is a schematic diagram illustrating the principle of missing feature fingerprint extraction and data health profiling according to an embodiment of this application;

[0039] Figure 4 This is a schematic diagram of linear missing value repair logic based on Kalman filtering according to an embodiment of this application;

[0040] Figure 5 This is a schematic diagram of an LSTM-based nonlinear temporal recursive repair structure according to an embodiment of this application;

[0041] Figure 6 This is a visual diagram illustrating the mask auditing mechanism and repair results according to an embodiment of this application;

[0042] Figure 7 This is a schematic diagram of the structure of a power system equipment control device according to an embodiment of this application;

[0043] Figure 8 This is a schematic diagram of the structure of a computing device according to an embodiment of this application. Detailed Implementation

[0044] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0045] As used herein, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context indicates otherwise. It should be further understood that the terms “comprising,” “including,” indicate the presence of the stated feature, operation, element, component, item, kind, and / or group, but do not preclude the presence, occurrence, or addition of one or more other features, operations, elements, components, items, kinds, and / or groups. The terms “or” and “and / or” as used herein are interpreted as inclusive, or mean any one or any combination thereof. Thus, “A, B, or C” or “A, B, and / or C” means “any one of: A; B; C; A and B; A and C; B and C; A, B, and C.” Exceptions to this definition occur only when combinations of elements, functions, or operations are inherently mutually exclusive in some manner.

[0046] Figure 1 This is a schematic flowchart of a device control method for a power system according to an embodiment of this application. Figure 1 As shown, an embodiment of the present application provides a device control method for a power system, which includes the following steps S1 to S6.

[0047] Step S1: Data Feature Extraction

[0048] In step S1, time-series data from industrial sensors are collected in real time, and statistical characteristics of each data column are calculated. The statistical characteristics include the missing rate and the maximum consecutive missing length.

[0049] Typically, the time-series data collected from industrial sensors is stored in a table, with data from the same sensor stored in the same column according to time sequence. The process involves iterating through the numerical data columns and calculating the missing rate for each column. and maximum consecutive missing length .in, =Number of missing points / Total points = The maximum number of consecutive missing points.

[0050] by Figure 3 The original data column shown is a column containing 417 consecutive sampling points (not all sampling points are shown in the figure), and 50 consecutive sampling points are missing. Therefore, the missing rate of this parameter column (i.e., the data column) is 12%, the maximum consecutive missing length is 50, and the status is consecutive block missing.

[0051] Step S2: Repair model matching

[0052] In step S2, a model is matched for each data column based on its data characteristics and physical properties. The model includes a Kalman filter model suitable for linear data columns containing Gaussian noise, a long short-term memory network model suitable for nonlinear data columns with long time-series dependence, and an unscented Kalman filter model suitable for strongly nonlinear data columns.

[0053] This step matches a model to each data column based on a pre-established model matching mechanism. The mechanism is as follows: for linear data columns containing Gaussian noise, a Kalman filter (KF) model is selected; for data columns with nonlinear and long-term time-series dependencies, a Long Short-Term Memory (LSTM) network model is selected; and for strongly nonlinear system data columns, an unscented Kalman filter (UKF) model is selected. Each data column has physical properties, such as temperature, pressure, flow rate, liquid level, rotational speed, and current. Based on the data characteristics obtained in step S1 and the physical properties of the data columns, a matching model is selected.

[0054] Step S3: Queue Construction and Repair

[0055] In step S3, all columns to be repaired are sorted according to the missing rate to generate a computation task queue. Based on the computation task queue, the missing positions of the corresponding columns are predicted and filled using a matching model combined with historical data to generate a full-scale repair data matrix. .

[0056] Step S4: Construct the mask matrix

[0057] Construct a Boolean mask matrix with the same structure as the original data column matrix. Used to record the data coordinates where the filling operation occurs. The mask matrix is ​​preferably a binary mask matrix. Indicates the first Line number Listed as data to be repaired. Indicates the first Line number Listed as original measurement values.

[0058] Steps S3 and S4 can be performed simultaneously.

[0059] Step S5, Data Fusion

[0060] The full-scale repair data matrix is ​​fused with the mask matrix to obtain a fused matrix. Based on this, the mask matrix can be used to drive the rendering engine to apply visual saliency markers to the data units in the repaired region, for example, according to... coordinate pair matrix The corresponding data units are marked with a preset background color to generate a traceable audit report file.

[0061] by Figure 6 For example, an original matrix is ​​generated based on the original parameter columns. Data in the second row and second column, the second row and third column, and the third row and second column of this matrix are missing. A mask matrix with the same structure as this matrix is ​​constructed. In this mask matrix, the second row and second column, the second row and third column, and the third row and second column are all 1, and the other positions are all 0. The predicted data for these three positions are 56.7, 48.1, and 52.3, respectively, obtained through model repair. The original matrix is ​​then filled with these three predicted data in sequence according to the positions where the element value in the mask matrix is ​​1, and the filled data is highlighted to obtain the audit report file.

[0062] Step S6: Generate control commands based on the fusion matrix to control the operation of the device.

[0063] In step S6, based on the timing characteristics of the data in the fusion matrix, the data in the fusion matrix is ​​restored into a state vector and input to the control system. The control system generates control commands based on the state vector and transmits the control commands to the equipment so as to control the operating state of the equipment.

[0064] In this embodiment of the application, industrial sensor data can be represented as:

[0065]

[0066] in, Indicates the first The sensor at the first The measured values ​​at each sampling time; Indicates the sensor number, such as temperature sensor. Pressure sensor Flow sensor Liquid level sensor Motor current sensor wait; Indicates the sampling time; Indicates the total number of sampling points; This indicates the number of sensors. In this application embodiment, unless otherwise specified, the subscript... All indicate the first One sensor.

[0067] Define missing indicator variables:

[0068]

[0069] No. Missing rate of sensor data series Defined as:

[0070]

[0071] Missing rate It can reflect the data integrity of the sensor within a sampling period.

[0072] No. Maximum consecutive missing length of sensor data columns Defined as:

[0073]

[0074] Maximum consecutive missing length Used to determine whether the missing data in a data column is a short-term discrete missing data or a long-term continuous block missing data.

[0075] Based on the physical properties of the data column Missing rate Maximum consecutive missing length and data change characteristics Select repair model:

[0076]

[0077] in, Indicates the first Repair model corresponding to each sensor data column; It represents physical properties, such as temperature, pressure, flow rate, liquid level, rotational speed, and current; Used to describe the dynamic characteristics of data, such as stationarity, abrupt changes, periodicity, and degree of nonlinearity; This represents the adaptive matching function of the model.

[0078] This application provides several repair models in its embodiments.

[0079] 1. An improved Kalman filter model.

[0080] The Kalman filter model is suitable for repairing linear or weakly nonlinear data. For industrial parameters such as pressure, liquid level, slowly changing temperature, and motor current, which have relatively stable changing trends and whose noise approximately follows a Gaussian distribution, the improved Kalman filter model is preferred for missing value repair. The repair logic of this model is as follows: Figure 4 As shown.

[0081] Data repair using an improved Kalman filter model requires establishing state variables:

[0082]

[0083] in, Indicates the first The sensor at the first The state vector of a physical quantity at each moment, for example Indicates the first Temperature state vector at each moment Indicates the first The pressure state vector at each moment; Indicates the first The first industrial physical quantity, i.e., the first... Data from the first sensor, in the... The state value at each time point can be obtained by sensor measurement or by estimation by the repair model; This indicates the rate at which the physical quantity changes with time. In the embodiments of this application, unless otherwise specified, the subscript... All indicate the first That moment.

[0084] The linear state-space model of this repair model is as follows:

[0085]

[0086]

[0087] in, For the first The sensor at the first The actual collected value at each moment; For the first The state transition matrix of a physical quantity is used to describe the natural evolution of the physical quantity over time. The physical quantity refers to the first The physical quantity corresponding to each sensor is as follows: for example, if the first sensor is a temperature sensor, then the first physical quantity is temperature; if the second sensor is a pressure sensor, then the second physical quantity is pressure. For the first The control input matrix for each physical quantity; The control quantity output by the control system at the previous moment, such as valve opening, pump speed, inverter frequency, heating power or cooling power; For the first The observation matrix of each physical quantity; This is process noise; Indicates the first At the [time]th moment The observation noise of each physical quantity.

[0088] When the sampling interval is When the physical quantity approximately satisfies the assumption of uniform change, we can take:

[0089]

[0090] in, This indicates the time interval between two adjacent sampling points.

[0091] Compared to conventional Kalman filtering, the advantage of this repair model is that the process noise covariance does not use a fixed value, but rather adaptively adjusts based on the missing rate of the sensor data in that column and the current consecutive missing length.

[0092]

[0093] in, Indicates the first The sensor at the first Adaptive process noise covariance at each time step; For the first The fundamental process noise covariance of physical quantities of each sensor; This is the missing rate weighting coefficient; The weighting coefficient is the length of consecutive missing characters. Indicates as of the date The current consecutive missing length at each moment; Indicates the preset reference missing length.

[0094] The time update process is as follows:

[0095]

[0096]

[0097] in, This represents the estimate of the current prior state obtained based on the state predicted at the previous moment; This represents the prior estimation error covariance.

[0098] When the There are valid observations at any given time, i.e. At that time, perform a measurement update:

[0099]

[0100]

[0101]

[0102] in, Kalman gain; To observe the noise covariance; It is an identity matrix.

[0103] When the Sensor data is missing at any given moment, i.e. At that time, instead of performing a measurement update, the following command is executed:

[0104]

[0105]

[0106]

[0107] in, That is, the first The sensor at the first The repair value at each moment.

[0108] To prevent the repair values ​​from exceeding the reasonable operating range of industrial equipment, this invention further introduces physical constraints:

[0109]

[0110] in, This is the final repair value after physical constraints have been applied; and The first The minimum and maximum allowable values ​​for a physical quantity, such as upper and lower limits for pressure, temperature, liquid level, or flow rate.

[0111] This improvement enables Kalman filtering not only to repair missing data by utilizing temporal continuity, but also to adjust the estimation results by combining the severity of the missing data with the physical boundaries of the equipment, thus avoiding the generation of repair values ​​that do not conform to the operating rules of the equipment in long-term missing scenarios.

[0112] 2. Improved LSTM model

[0113] LSTM models are suitable for repairing nonlinear long-series parameters. For industrial parameters with nonlinear, hysteretic, or periodic variation characteristics, such as feedwater flow rate, reactor temperature, pipeline flow rate, load current, and valve-controlled fluid pressure, a Long Short-Term Memory (LSTM) network is used to repair missing values. The repair logic of this model is as follows: Figure 5 As shown in the figure, ; Indicates the first The recursive repair state at any given moment; Indicates the hidden state of LSTM. This represents the recursive input vector at the next moment after the fusion and repair results.

[0114] First, establish the input vector:

[0115]

[0116] in, Indicates the first The sensor at the first The LSTM input feature vector at time step; Indicates the length of the sliding window; This indicates the historical valid value or repaired value prior to the missing point; This indicates the equipment control parameters at the previous moment, such as valve opening, pump speed, heating power, and cooling power. Indicates whether the observation from the previous time step is missing; Indicates the length of consecutive missing data up to the previous time step; It is a static characteristic or physical property quantity, representing the first... Each sensor corresponds to a certain attribute code of a physical quantity.

[0117] The specific structure of this LSTM model is as follows.

[0118] The Gate of Oblivion is:

[0119]

[0120] The input gate is:

[0121]

[0122] The candidate cell status is as follows:

[0123]

[0124] Cell status updated to:

[0125]

[0126] The output gate is:

[0127]

[0128] The hidden state is:

[0129]

[0130] The predicted repair value is:

[0131]

[0132] in, This is the forget gate, used to determine how much historical operational information is retained; This is the input gate, used to determine the impact of the current input on the state update; Candidate cell state; This represents the cellular state, used to store long-term temporal dependencies; For output gate; It is in a hidden state; This represents the Sigmoid activation function; Represents the hyperbolic tangent function; Represents element-wise product; , , , , These represent the weight matrices for the forget gate, input gate, candidate cell state, output gate, and predicted repair, respectively. , , , , These represent the forget gate, input gate, candidate cell state, output gate, and bias term for predicted repair, respectively.

[0133] The improvement of this model lies in the fact that the input of LSTM not only includes historical sensor data, but also incorporates control variables, missing indicator variables, consecutive missing lengths, and physical properties. The changes in real measurement data are smooth, while the changes caused by missing data are large fluctuations, which enables the model to distinguish between real measurement changes and information breaks caused by missing data.

[0134] When the When data is missing at a given moment, the system uses a recursive repair method:

[0135]

[0136] When multiple consecutive time steps are missing, the repaired value from the previous time step will be used as the input for the prediction of the next time step:

[0137]

[0138] To improve the consistency between the repair value and industrial physical laws, embodiments of this application introduce a physical constraint loss function during training and / or model updating. Typically, a physical constraint loss function is introduced during the training or model update process. This is sufficient, but if the sensor ages due to prolonged use, a physical constraint loss function needs to be introduced during both training and model updates. Physical constraint loss function The expression is as follows:

[0139]

[0140] Among them, the first item The second term represents the prediction error of the effective observation points. This is a control response constraint term used to constrain the physical relationship between the sensor's changing trend and the control input. Indicates the effect of the change in control quantity on the first The expected impact of a physical quantity, such as an increase in valve opening typically leading to an increase in flow rate, and an increase in heating power typically leading to a rise in temperature; the third item For boundary penalty terms, and These are the weighting coefficients.

[0141] The boundary penalty function can be defined as:

[0142]

[0143] in, and These are the minimum and maximum allowable values ​​for the industrial physical quantity, respectively.

[0144] The improvements made to conventional LSTM in this application embodiment enable LSTM to not only fit time series mathematically, but also to repair missing values ​​by combining device control inputs and physical boundaries, thereby improving the availability of the repaired values ​​for subsequent device control.

[0145] 3. Improved Unscented Kalman Filter Model UKF

[0146] This model is suitable for repairing strongly nonlinear data. For strongly nonlinear systems, such as reactor temperature changes, complex pipeline pressure fluctuations, multi-actuator coupled control systems, and industrial processes with significant hysteresis and nonlinear response, the embodiments of this application use an improved unscented Kalman filter model for missing value repair.

[0147] The nonlinear state-space model is represented as:

[0148]

[0149]

[0150] in, Indicates the first The sensor at the first The state vector of the physical quantity at each moment. For the first The nonlinear state transition function corresponding to each physical quantity is used to describe the law of change of equipment state with time and control input; It is a nonlinear observation function used to describe the relationship between sensor measurements and internal states; This is the control input from the previous moment; This is process noise; To observe noise.

[0151] Let the state dimension be The state at the previous time step was estimated as follows: covariance is Then construct the Sigma point:

[0152]

[0153]

[0154] The improvement of this model lies in the fact that the repair process does not simply rely on historical values ​​for black-box prediction. Instead, it incorporates industrial control inputs and operating conditions into the model input, enabling the repaired values ​​to reflect the actual operating status of the equipment. For example, when the pump speed increases, the repair result of the flow sensor should reflect the trend of flow rate increasing with speed; when the valve opening decreases, the repair result of pipeline pressure or flow rate should reflect the corresponding physical change direction.

[0155] Furthermore, embodiments of this application introduce a mask-based loss function into the model to avoid mixing up the original valid measurements with the missing filler values ​​during model parameter calibration. When When filling in values ​​for the algorithm, the mask variable The value is 1, when When the original measurement value is used, the mask variable The value is 0.

[0156] During model training or online calibration, the original valid measurements should be used first to calculate the error:

[0157]

[0158] in, This indicates the total number of sampling points for a data column.

[0159] In addition, to ensure that the repair results comply with the physical constraints of industrial equipment, this application embodiment also adds physical constraint items:

[0160]

[0161] in, and These represent the upper and lower limits of the physical quantity of the sensor, such as the upper limit of pipeline pressure, the upper limit of safe temperature, and the minimum allowable flow rate. Indicates the rate of change of the repair value; This represents the reference rate of change calculated based on the equipment's empirical model or adjacent sensors. , , These are the weighting coefficients.

[0162] The final loss function is:

[0163]

[0164] in, To account for observation error loss, For equipment dynamic losses, For physical boundary loss, and These are the weighting coefficients for the corresponding loss terms.

[0165] The observation error loss is:

[0166]

[0167] in, Indicates the first One sensor in Valid measurement value at time, Indicates the first One sensor in The model prediction at time 10:00. This indicates that the data is valid. When the data is valid, When data is missing, Therefore, this loss is used to train the model using only valid sensor measurements.

[0168] The dynamic loss of the equipment is as follows:

[0169]

[0170] in, This indicates the first determination based on equipment status, control input, and external disturbances. A function of the rate of change of a physical quantity. This loss is used to constrain the predicted trends of temperature, pressure, flow rate, or liquid level changes by the model, ensuring they align with the actual dynamics of industrial equipment. Indicates the first Each state variable corresponds to the lag time of the control input, measured in the number of sampling periods, and is used to characterize the response delay that exists when the control input acts on the state variable. Indicates the first External disturbance vectors or auxiliary input variables at each sampling time, including environmental disturbances, load changes, boundary condition changes, or other variables that are not modeled as control inputs but affect the dynamic response of the system. Indicates the first The scale normalization factor for each state quantity The standard deviation, range, or preset reference scale of the state quantity in the training samples can be taken to eliminate the influence of differences in the dimensions and numerical ranges of different state quantities on the loss function. This loss is used to constrain the temperature, pressure, flow rate, or liquid level change trends predicted by the model, so that they are consistent with the actual dynamic laws of industrial equipment. Normalized sample size representing the dynamic law loss term:

[0171]

[0172] in, Indicates the first The state variable is in the first... The validity weights or mask coefficients at each sampling time. When the data is valid at that moment and participates in the dynamic constraint calculation, When data is missing, abnormal, or not included in the calculation at that moment, .

[0173] The physical boundary loss is:

[0174]

[0175] in, and They represent the first The lower and upper limits of the physical quantities corresponding to each sensor; This represents the total number of sensors. When the predicted value is within the allowable range, the physical boundary loss is zero; when the predicted value exceeds the device's safe range, the loss function penalizes the prediction result.

[0176] The loss function described above not only learns the time series variation characteristics based on existing sensor data, but is also subject to the joint constraints of equipment dynamics and physical boundaries, thereby reducing the probability of generating abnormal repair values ​​that exceed the sensor range, equipment safety range, or normal variation patterns of industrial processes.

[0177] The increase in equipment dynamic law loss and physical boundary loss makes UKF repair results not only have the ability to fit time series, but also constrained by sensor range, equipment safety boundary and industrial process change law, thereby avoiding the generation of abnormal repair values ​​that exceed the physical range of the equipment.

[0178] The following details the process of generating, predicting, and updating the Sigma point in the UKF model described above.

[0179] Let the state equation of the nonlinear industrial system be:

[0180]

[0181] in, Nonlinear functions that represent the changes in equipment status with time and control input, such as the heat balance function where temperature is affected by both heating power and cooling water flow, and the fluid dynamic function where pressure is affected by pump speed and valve opening. This represents the sensor observation function, used to describe the nonlinear mapping between the internal state of the device and the sensor measurement values; This is process noise; To observe noise.

[0182] for dimensional state vector The mean was estimated based on the previous time step. Covariance generate Sigma points:

[0183]

[0184] in, The scaling parameter is used. Substituting the Sigma point into the nonlinear state equation of the industrial equipment, the predicted Sigma point is obtained:

[0185]

[0186] The predicted state mean is:

[0187]

[0188] The predicted covariance is:

[0189]

[0190] in, and These are the Sigma point mean weights and covariance weights, respectively. The basic process noise covariance represents the physical quantity corresponding to the sensor.

[0191] Substituting the predicted Sigma point into the observation function, we obtain the predicted observation value:

[0192]

[0193] When sensor data is valid, updates are made based on the observed values; when sensor data is missing, predicted observed values ​​are used as missing data correction values.

[0194]

[0195] in, Indicates the first One sensor in Missing repair value at any given moment; This represents the predicted value of the sensor observation vector; Represents the extraction of the first predicted observation vector. Each component.

[0196] The improvement of this model lies in the nonlinear function of UKF. and These parameters are not only provided by mathematical models but also correspond to industrial physical quantities. For example, in a pumping system, pressure can be related to pump speed, valve opening, and pipeline resistance; in a heat exchange system, temperature can be related to heating power, cooling water flow rate, and medium flow rate; and in a storage tank system, liquid level can be related to feed flow rate and discharge flow rate. By incorporating these physical variables into the nonlinear state equations, the missing value repair results can reflect the equipment's operating mechanism, rather than simply relying on manually filled rules.

[0197] Based on the above models, this application embodiment also defines the following model adaptive matching rules: Model matching is based on sensor physical properties, missing features, and equipment operating conditions for adaptive routing. Specifically, let:

[0198]

[0199] in,( ) indicates the first The repair model selected for each sensor; Represents the model routing function. This indicates the degree of nonlinearity of the physical quantity corresponding to the sensor; This indicates the correlation between the physical quantity and other sensor physical quantities; This represents the response time constant of the physical quantity. For example, temperature has a large thermal inertia time constant, while pressure and flow rate may have faster response characteristics.

[0200] When sensor data is linearly stationary, noise is approximately Gaussian, and the response process has obvious inertial characteristics, the KF model is selected; when sensor data has obvious periodicity, long time-series dependence, or complex nonlinear fluctuations, the LSTM model is selected; when there is a strong nonlinear coupling relationship between sensor data and control input, the UKF model is selected.

[0201] Once the data is repaired, it will be used as feedback input to the controller to participate in closed-loop control of the equipment's operating status, as well as display, storage, and statistical analysis.

[0202] Let the repaired sensor state vector be:

[0203]

[0204] in, The temperature after repair. For the pressure after repair, For the restored traffic, The liquid level after repair. The corrected rotational speed. The control system compares the corrected state vector with the target setpoint:

[0205]

[0206] in, Define a value vector for the target, such as target temperature, target pressure, target flow rate, or target liquid level; To control errors, the controller generates control commands based on the errors:

[0207]

[0208] in, This represents the control law, which can be PID control, model predictive control, or a logic control algorithm in an industrial control system. Indicates the equipment's safety status or operating mode, and control commands. This may include valve opening adjustment, variable frequency pump frequency adjustment, heating power adjustment, cooling water valve opening adjustment, and feeding speed adjustment.

[0209] For example, in a pipeline delivery system, when flow sensor data is temporarily missing, LSTM or UKF can be used to repair and obtain the current flow estimate. and with target traffic Compare:

[0210]

[0211] The controller adjusts the frequency of the variable frequency pump or the valve opening based on this error:

[0212]

[0213] in, Indicates the first The flow control command for each control cycle can represent the frequency command of the variable frequency pump or the valve opening command. Indicates the target traffic; Indicates the current flow control deviation; , and These represent the proportional coefficient, integral coefficient, and derivative coefficient of flow control, respectively. This indicates the sampling period of the control system. It is determined by the control quantity. The system can actually change the pump speed or valve opening, thereby changing the pipeline flow and adjusting the physical operating state of industrial equipment.

[0214] For example, in heat exchange or heating systems, when temperature sensor data is missing, the system uses KF or UKF to obtain a temperature estimate. and with target temperature Compare:

[0215]

[0216] in, For the first Temperature control deviation per control cycle (i.e., the current control cycle), The target temperature; The estimated temperature after repair is obtained using KF or UKF.

[0217] controller data Adjust the heating power or cooling water valve opening:

[0218]

[0219] in, Indicates the first The temperature control command for each control cycle can represent the heater output power, steam valve opening, or cooling water regulating valve opening. , and These represent the proportional coefficient, integral coefficient, and derivative coefficient of temperature control, respectively.

[0220] This control action can directly affect the temperature change of the equipment, and belongs to the control of industrial physical processes.

[0221] Meanwhile, to avoid false triggering of control when the confidence of the repair value is insufficient, the embodiments of this application can generate control confidence based on the mask matrix and the length of consecutive missing values. :

[0222]

[0223] in, Indicates the first The sensor or state variable in the first Reliability of time-based control Indicates the current consecutive missing length. This is the decay coefficient. The longer the continuous loss time, the more... The smaller the value, the higher the uncertainty of the repair value. The controller can then... Adjust the control weights to generate the following control instructions:

[0224]

[0225] in, This indicates the control commands generated based on the repair data. This represents a safety control strategy, such as maintaining the current valve opening, reducing pump frequency, switching to a conservative operating mode, or triggering manual confirmation. Each safety strategy for each physical quantity corresponds to a numerical value, which can be substituted into a formula to calculate the final control command. By introducing control reliability, the embodiments of this application can not only achieve missing data repair but also improve the operational safety of industrial control systems under sensor malfunctions.

[0226] Through the above-mentioned repair models, model matching methods, and control command generation methods, the equipment control method of the power system in this application embodiment can produce the following technical effects:

[0227] First, missing data is repaired by using the state equations of industrial equipment, control inputs, and sensor observation relationships, so that the repaired values ​​conform to industrial physical laws such as temperature inertia, pressure continuity, dynamic changes in flow rate, and conservation of liquid level, avoiding non-physical abrupt changes or excessive smoothing caused by simple interpolation.

[0228] Second, through the adaptive routing mechanism of the three repair models KF, LSTM and UKF, different types of industrial sensor data are matched with different repair models, thereby improving repair accuracy and computational efficiency.

[0229] Third, the repair position is recorded by a mask matrix, and the mask information is further used for control confidence calculation, so that the control system can distinguish between the original measurement value and the algorithm repair value, reducing the risk of miscontrol caused by missing sensor data.

[0230] Fourth, the repaired data is input into PLC, DCS, SCADA or edge controller to adjust actuators such as valves, pumps, heaters, cooling systems, and feeding mechanisms, so that the data repair results directly affect the operating status of industrial equipment, forming a closed-loop control process of "sensor data acquisition - missing detection - physical constraint repair - control command generation - actuator adjustment".

[0231] In summary, the embodiments of this application utilize sensor measurement, state estimation, industrial process modeling, and closed-loop control technologies to solve the technical problems of unstable control, false alarms, misadjustments, and increased equipment operation risks caused by missing sensor data during the operation of industrial equipment. They also achieve the technical effects of improving control continuity, enhancing equipment operation safety, and improving the reliability of industrial data.

[0232] The following section uses the restoration of historical sensor data in a simulated nuclear power circulating water system as an example to illustrate the detailed process of the power system equipment control method according to an embodiment of this application. The system exports operating data in Excel format from an industrial database at fixed time intervals. The data columns include sampling time, water supply temperature T, etc. s Return water temperature T rMain pipe pressure P, circulating water flow rate Q, water tank level H, valve opening u, pump frequency f, and outdoor temperature T a Due to factors such as packet loss in field communication, sensor maintenance, equipment start-up and shutdown, and electromagnetic interference, some data columns contain discrete or continuous missing data, which needs to be repaired while preserving the traceability of the original data.

[0233] The specific execution process is as follows:

[0234] Step 1, Data Feature Extraction.

[0235] After reading the Excel file, the system first identifies the time index column and then iterates through the columns of numerical sensors. For each column, the system calculates the missing rate R. miss and maximum consecutive missing length L gap For example, in an Excel spreadsheet, the water supply temperature T... s There are 8 missing data points, with a missing rate of 0.56% and a maximum consecutive missing length of 2; the main pipe pressure P has 60 missing data points, with a missing rate of 4.17% and a maximum consecutive missing length of 50; the circulating water flow rate Q has 35 missing data points, with a missing rate of 2.43% and a maximum consecutive missing length of 6; the water tank level H has 15 missing data points, with a missing rate of 1.04% and a maximum consecutive missing length of 4; the valve opening u and pump frequency f data are complete.

[0236] Based on the above statistical results, a data health profile is formed: water supply temperature T s The following columns show mild discrete missing values: main pipe pressure P (continuous block missing), circulating water flow rate Q (moderate missing with dynamic fluctuations), and tank level H (mild continuous missing). If the system detects that a column has no valid observations at all, it marks that column as non-automatically repairable and prompts for manual confirmation or the introduction of other data sources.

[0237] Step 2, Repair model matching

[0238] The system combines missing features and physical attributes of data columns to select the corresponding repair model from the algorithm library.

[0239] For water supply temperature T sFor the water tank level H, due to its relatively slow change and strong continuity and inertia characteristics during short-term missing periods, the system selects a Kalman filter model for repair. For the main pipe pressure P, although its overall change is relatively stable during normal operation, there are missing data points for 50 consecutive sampling points; therefore, a Kalman filter model is also selected to maintain the pressure change trend through state prediction. For the circulating water flow rate Q, due to its influence from pump frequency f and valve opening u, it exhibits significant nonlinearity and time-series correlation; therefore, the system selects an LSTM model for repair. If parameters with strong nonlinear relationships exist in other scenarios, such as branch pressure difference ΔP where the pressure difference has a significant relationship with the square of the flow rate, the UKF model can be selected to estimate the state at the missing time points through unscented transformation.

[0240] Step 3: Generate a binary mask matrix M with the same row and column dimensions as the original data matrix based on the Excel spreadsheet. For the actual measurement values ​​already present in the original file, the corresponding positions in matrix M are marked as 0.

[0241] Step 4, Queue Construction and Repair

[0242] Sort the columns to be repaired according to their missing rate from highest to lowest, and construct a task queue. The task queue order is: main pressure P, circulating water flow rate Q, water tank level H, and supply water temperature T. s Repairs are performed column by column in the order of the queue.

[0243] For the main pressure P, a linear state-space model is established, where the state variables can be set as the pressure value and its rate of change. All sampling points of the main pressure P parameter series are traversed sequentially according to sampling time. When a sampling point is reached and the pressure observation value exists, time and measurement updates are performed. When a sampling point is reached but the pressure value is missing, the measurement update is skipped, and the prior estimate obtained from the Kalman filter is used as the corrected value for that moment, continuing the recursion to the next moment. This maintains the continuity of the pressure trend during periods of 50 consecutive missing sampling points, avoiding abrupt broken lines caused by simple linear interpolation.

[0244] For the circulating water flow rate Q, a sliding window-based LSTM prediction model is constructed. The model input includes valid historical sequences of circulating water flow rate Q, pump frequency f, valve opening u, and main pipe pressure P from the 30 consecutive sampling points preceding the missing point. The output is the predicted circulating water flow rate for the current sampling point. When a sampling point's Q value is missing, the LSTM model generates a predicted value for that moment and fills the missing position with this value. Subsequently, this predicted value serves as one of the inputs for recursive predictions at subsequent moments, used to repair consecutive or adjacent missing points. This preserves the nonlinear dynamic characteristics of flow rate variations with pump frequency and valve opening.

[0245] For the water tank level H and the water supply temperature T sA Kalman filter model is used for smoothing and repair. Since these two types of parameters do not change drastically in a short period of time, the Kalman filter can estimate the missing points by utilizing the continuity of the states at different times, thus avoiding data distortion caused by mean filling.

[0246] For strongly nonlinear parameters requiring UKF processing, a nonlinear state equation is first constructed based on the system's physical relationships. For example, in branch pressure difference repair, the nonlinear equation can be established using the relationship that pressure difference ≈ drag coefficient × flow rate squared. The system generates Sigma points through unscented transformation, substitutes these Sigma points into the nonlinear equation for propagation, and then estimates the state values ​​at the missing time points based on the propagated mean and covariance.

[0247] Step 5, Mask generation and export

[0248] For the repaired values ​​obtained by the KF, LSTM or UKF algorithms, the corresponding position in matrix M is marked as 1.

[0249] For example, if the main pressure P is missing in rows 520 to 569 of an Excel file, then after data repair, these 50 positions will be filled with Kalman filter estimates, and the corresponding 50 positions in the mask matrix will be marked as 1. If the circulating water flow rate Q is missing in rows 730 to 735, then LSTM predictions will be filled in, and the corresponding mask positions will be marked as 1.

[0250] During the export phase, the repaired complete data matrix is ​​fused with the mask matrix M. For cells with a value of 0, the exported file retains the default background, indicating that the value is the original measurement; for cells with a value of 1, the exported file sets the corresponding cell to a light yellow background or other preset highlight style, indicating that the value is the algorithm-filled value. The final generated Excel audit report not only contains complete repaired data but also clearly shows which data comes from the original sensors and which data comes from algorithm estimation, thus facilitating review by engineers and subsequent accountability.

[0251] As can be seen from the above examples, the method in this application does not simply apply uniform interpolation to all missing values. Instead, it first classifies the data based on the missing rate, consecutive missing length, and physical properties of different sensor columns, then adaptively selects repair models such as KF, LSTM, or UKF, and uses a mask matrix to achieve visual auditing of the repair behavior. Therefore, this method is more suitable for batch processing of time series data from various types of industrial sensors, such as temperature, pressure, flow rate, liquid level, valve opening, and pump frequency.

[0252] This application also provides a data repair system capable of performing the data repair steps in the above-described device control method.

[0253] Figure 2This is a schematic diagram of a data repair system architecture according to an embodiment of this application. Figure 2 As shown, the data repair system includes a core algorithm layer, a logic layer, and an interaction layer.

[0254] The core algorithm layer mainly includes a multimodal algorithm library, which contains KF kernel, LSTM kernel and UKF kernel. Kalman filter is used to process linear Gaussian systems, long short-term memory network is used to capture nonlinear long-term dependencies, and unscented Kalman filter is used to deal with strongly nonlinear scenarios.

[0255] The logical layer further includes a diagnostic analysis engine, a batch processing scheduler, and an audit export module.

[0256] The diagnostic analysis engine is used for data perception and feature diagnosis. In the process of data perception and feature diagnosis, the system does not just read the Excel file stream. Instead, considering that industrial data often contains time indexes and multidimensional physical quantities, a missing feature extraction process is introduced. That is, it traverses all numerical columns, calculates the missing rate index and the distribution of consecutive missing lengths, and uses these two statistical features to construct a health profile of the data. The data columns are divided into different states such as complete, discrete missing, and consecutive block missing, providing a decision basis for subsequent algorithm matching.

[0257] Batch processing schedulers manage and schedule computational task queues. Facing the demands of processing massive industrial-scale data, the scheduler automatically prioritizes repair tasks based on their missing data rates, allowing for the utilization of high-missing-rate columns to leverage their inherent temporal autocorrelation or provide a complete foundation for subsequent multivariate joint repair. In terms of task scheduling, the scheduler employs an adaptive routing mapping mechanism, automatically routing the sequences to be repaired to the optimal algorithm kernel based on the data's physical properties (such as the gradual variability of temperature or the abrupt changes in flow) and missing data characteristics. This ensures that the repair process conforms to both statistical data patterns and respects physical trends.

[0258] The core of the audit export module is the mask audit mechanism: a Boolean mask matrix is ​​maintained in memory that is isomorphic to the original data, recording the coordinates of each filling operation; during the export stage, the mask is mapped to a visual highlighting style (such as a light yellow background), so that the repair results have WYSIWYG traceability.

[0259] The interaction layer contains interactive interface modules, which are used to display the GUI interface, file loading button, and log window, etc.

[0260] This application also provides an equipment control device for a power system, such as... Figure 7 As shown, the device includes:

[0261] The data feature extraction unit 1 is suitable for real-time acquisition of time-series data from industrial sensors and for statistical analysis of the data characteristics of each data column, including the missing rate and the maximum consecutive missing length.

[0262] Repair model matching unit 2 is adapted to match a repair model for each data column according to the statistical characteristics and physical properties of each data column. The repair model includes a Kalman filter model suitable for linear data columns containing Gaussian noise, a long short-term memory network model suitable for nonlinear and long time-series dependent data columns, and an unscented Kalman filter model suitable for strongly nonlinear data columns.

[0263] The queue construction and repair unit 3 is adapted to sort all columns to be repaired according to the missing rate, generate a computation task queue, and predict and fill the missing positions of the corresponding columns using a matching repair model based on the computation task queue, thereby generating a full repair data matrix.

[0264] Mask matrix construction unit 4 is suitable for constructing a Boolean mask matrix with the same structure as the original data column data matrix, used to record the data coordinates where the filling operation occurs;

[0265] Data fusion unit 5 is adapted to fuse the full repair data matrix with the mask matrix to obtain a fusion matrix;

[0266] The control unit 6 is adapted to generate a state vector based on the fusion matrix, generate control commands based on the state vector, and input the control commands to the device to control the operation of the device.

[0267] In a preferred embodiment of this application, the above-described apparatus further includes:

[0268] A rendering unit is adapted to drive a rendering engine using the mask matrix to apply visual saliency markers to data units in the repaired regions of the fusion matrix.

[0269] As a preferred embodiment of this application, the process noise covariance of the Kalman filter model for:

[0270]

[0271] For the first The fundamental process noise covariance of physical quantities of each sensor; This is the missing rate weighting coefficient; This represents the missing rate of the data column; The weighting coefficient is the length of consecutive missing characters. Indicates as of the date The current consecutive missing length at each moment; Indicates the preset reference missing length.

[0272] As a preferred embodiment of this application, the final repair value of the Kalman filter model for:

[0273]

[0274] in, and The first The minimum and maximum allowable values ​​of the physical quantities corresponding to each industrial sensor. For the first The sensor at the first The repair value at each moment.

[0275] As a preferred embodiment of this application, the input feature vector of the Long Short-Term Memory network model for:

[0276]

[0277] in, This represents the historical valid values ​​or repaired values ​​prior to the missing points. This indicates the equipment control quantity at the previous moment. This indicates the length of consecutive missing data up to the previous time step. For missing indicator variables, A value of 1 indicates that the previous time step was a valid measured value. A value of 0 indicates that the value was missing in the previous time step; For the first The attribute encoding of physical quantities corresponding to each sensor.

[0278] As a preferred embodiment of this application,

[0279] The Long Short-Term Memory (LSTM) network model incorporates a physical constraint loss function during training and model updates. :

[0280]

[0281] in, For the prediction error of effective observation points, Indicates the first The sensor at the first The measured values ​​at each sampling time; To control response constraints; Indicates the effect of the change in control quantity on the first The expected impact of each physical quantity; For boundary penalty terms; and These are the weighting coefficients; This represents the total number of sampling points in the parameter column; This is the boundary penalty function.

[0282] In a preferred embodiment of this application, the nonlinear state-space model of the unscented Kalman filter model is expressed as follows:

[0283]

[0284]

[0285] Subscript express Time, subscript Indicates the first One sensor, Indicates the first The sensor at the first The state vector of the physical quantity at each moment. This represents a nonlinear state transition function, used to describe how the device state changes with time and control input; It is a nonlinear observation function used to describe the relationship between sensor measurements and internal states; for Time-based control input; This is process noise; To observe noise; Indicates the first One sensor in The actual sampled value at that moment.

[0286] In a preferred embodiment of this application, the control unit is further adapted to generate a control confidence coefficient based on the mask matrix and the current consecutive missing length. The control command is adjusted based on the control reliability coefficient to obtain the final control command. :

[0287]

[0288] in, This indicates the control commands generated based on the repair data. This indicates the security control strategy.

[0289] The equipment control device of the aforementioned power system can execute the processing of each step of the equipment control method of the aforementioned power system. Its principle and effect are the same as those of the equipment control method of the power system, and will not be described again here.

[0290] This application also provides an electronic device. For example... Figure 8As shown, the electronic device includes: a memory 7; a processor 8; and a computer program 9; wherein the computer program 9 is stored in the memory 7 and configured to be executed by the processor 8 to implement the device control method of the power system according to the embodiments of this application. A bus 10 is used for communication between the memory 7 and the processor 8.

[0291] The various techniques described herein can be implemented in combination with hardware or software, or a combination thereof. Thus, the methods and apparatus of the present invention, or certain aspects or portions thereof, can take the form of program code (i.e., instructions) embedded in a tangible medium, such as a removable hard disk, USB flash drive, floppy disk, CD-ROM, or any other machine-readable storage medium, wherein when the program is loaded into and executed by a machine such as a computer, the machine becomes an apparatus for practicing the present invention.

Claims

1. A method for controlling equipment in a power system, characterized in that, include: Real-time acquisition of time-series data from industrial sensors, and statistical analysis of the data characteristics of each data column, including the missing rate and the maximum consecutive missing length; Based on the statistical characteristics and physical properties of each data column, a repair model is matched for each data column. The repair models include a Kalman filter model suitable for linear data columns containing Gaussian noise, a long short-term memory network model suitable for nonlinear and long time-series dependent data columns, and an unscented Kalman filter model suitable for strongly nonlinear data columns. All columns to be repaired are sorted according to the missing rate to generate a computation task queue. Based on the computation task queue, the missing positions of the corresponding columns are predicted and filled using a matching repair model to generate a full repair data matrix. Construct a Boolean mask matrix with the same structure as the original data column matrix to record the coordinates of the data where the filling operation occurred; The full repair data matrix is ​​fused with the mask matrix to obtain a fused matrix; A state vector is generated based on the fusion matrix, a control command is generated based on the state vector, and the control command is input to the device to control the device operation.

2. The method as described in claim 1, characterized in that, Also includes: The rendering engine is driven by the mask matrix to apply visual saliency markers to the data units of the repaired region in the fusion matrix.

3. The method as described in claim 1, characterized in that, The process noise covariance of the Kalman filter model for: ； in, For the first The fundamental process noise covariance of physical quantities of each sensor; This is the missing rate weighting coefficient; This represents the missing rate of the data column; The weighting coefficient is the length of consecutive missing characters. Indicates as of the date The current consecutive missing length at each moment; Indicates the preset reference missing length.

4. The method as described in claim 3, characterized in that, The final repair value of the Kalman filter model for: ； in, and The first The minimum and maximum allowable values ​​of the physical quantities corresponding to each industrial sensor. For the first The sensor at the first The repair value at each moment.

5. The method as described in claim 1, characterized in that, The input feature vector of the Long Short-Term Memory network model for: ； in, This represents the historical valid values ​​or repaired values ​​prior to the missing points. This indicates the equipment control quantity at the previous moment. This indicates the length of consecutive missing data up to the previous time step. For missing indicator variables, A value of 1 indicates that the previous time step was a valid measured value. A value of 0 indicates that the value was missing in the previous time step; For the first The attribute encoding of physical quantities corresponding to each sensor.

6. The method as described in claim 5, characterized in that, The Long Short-Term Memory (LSTM) network model incorporates a physical constraint loss function during training and model updates. : ； in, For the prediction error of effective observation points, Indicates the first The sensor at the first The measured values ​​at each sampling time; To control response constraints; Indicates the effect of the change in control quantity on the first The expected impact of each physical quantity; For boundary penalty terms; and These are the weighting coefficients; This represents the total number of sampling points in the parameter column; This is the boundary penalty function.

7. The method as described in claim 1, characterized in that, The nonlinear state-space model of the unscented Kalman filter model is expressed as: ； ； Subscript express Time, subscript Indicates the first One sensor, Indicates the first The sensor at the first The state vector of the physical quantity at each moment. This represents a nonlinear state transition function, used to describe how the device state changes with time and control input; It is a nonlinear observation function used to describe the relationship between sensor measurements and internal states; for Time-based control input; This is process noise; To observe noise; Indicates the first One sensor in The actual sampled value at that moment.

8. The method as described in claim 1, characterized in that, Also includes: Generate the first [number] based on the mask matrix and the current consecutive missing length. A sensor or state variable in Reliability of Time Control The control commands are adjusted based on the control confidence level to obtain the final control commands. : ； in, This indicates the control commands generated based on the repair data. This indicates the security control strategy.

9. A power system equipment control device, comprising: The data feature extraction unit is suitable for real-time acquisition of time-series data from industrial sensors and for statistical analysis of the data characteristics of each data column, including the missing rate and the maximum consecutive missing length. The repair model matching unit is adapted to match a repair model for each data column based on the statistical characteristics and physical properties of each data column. The repair model includes a Kalman filter model suitable for linear data columns containing Gaussian noise, a long short-term memory network model suitable for nonlinear and long time-series dependent data columns, and an unscented Kalman filter model suitable for strongly nonlinear data columns. The queue construction and repair unit is adapted to sort all columns to be repaired according to the missing rate, generate a computation task queue, and, based on the computation task queue, use a matching repair model to predict and fill in the missing positions of the corresponding columns to generate a full repair data matrix. The mask matrix construction unit is suitable for constructing a Boolean mask matrix with the same structure as the original data column matrix, and is used to record the data coordinates where the filling operation occurs. A data fusion unit is adapted to fuse the full-scale repair data matrix with the mask matrix to obtain a fusion matrix; The control unit is adapted to generate a state vector based on the fusion matrix, generate control commands based on the state vector, and input the control commands to the device to control the operation of the device.

10. An electronic device, characterized in that, include: Memory; processor; as well as Computer programs; The computer program is stored in the memory and configured to be executed by the processor to implement the method as described in any one of claims 1 to 8.

Images