Power plant multi-source data arrangement and transmission method

Abstract

The invention provides a power plant multi-source data arrangement and transmission method, which comprises the following steps of: constructing a multi-scale data alignment model, performing time axis mapping by using a B-spline curve fitting algorithm, and introducing a quantum genetic algorithm into data with large time deviation to optimize a time window; carrying out data fusion, adopting an improved Dempster-Shafer synthesis rule, combining with the uncertainty of cloud model quantification, and carrying out grey correlation analysis on key parameters to calculate the correlation degree; abnormal data detection and restoration are carried out, an improved isolated forest algorithm is used for detection, and Kalman filtering interpolation restoration is carried out; data transmission optimization is implemented, routing is optimized through an ant colony algorithm, the fountain code technology of LDPC codes is adopted, and real-time data is scheduled through a token bucket algorithm; and constructing an LSTM and ESN fusion model to predict a trend, and sliding a window to update parameters. The method improves the intelligent level of power plant data management, and improves the intelligent level and decision accuracy of power plant data management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a power plant multi-source data arrangement and delivery method. BACKGROUND

[0002] In a traditional power plant, the data acquisition of power generation equipment has a multi-source heterogeneous problem. The data collected by the system automatically and the data collected by manual inspection are often inconsistent in time points, which often leads to data deviation. These data are directly uploaded to the server without effective integration, which damages the accuracy of the final power generation data and even causes system errors. In the prior art, although there are some data fusion methods, the time characteristics and dynamic change rules of the data in the power plant environment are not fully considered, and it is difficult to meet the high-precision data processing requirements. SUMMARY

[0003] The main purpose of the present application is to provide a power plant multi-source data arrangement and delivery method, which solves the problem of inconsistent time points between the data collected by the system automatically and the data collected by manual inspection in the traditional power plant.

[0004] To solve the above technical problems, the technical scheme adopted by the present application is as follows: a power plant multi-source data arrangement and delivery method, which comprises the following steps: S1, a multi-scale data alignment model is constructed, a B-spline curve fitting algorithm is used for time axis mapping, and a control point is used to adjust the curve; for data with large time deviation, a quantum genetic algorithm is introduced to optimize the time window; S2, data fusion processing is performed, an improved Dempster-Shafer synthesis rule is used, including a credibility factor and a conflict redistribution, and a cloud model is used to quantify uncertainty; a grey correlation analysis algorithm is used to calculate the correlation degree of the key parameters; S3, abnormal data detection and repair are carried out, an improved isolated forest algorithm is used and a density peak value clustering is introduced to detect the abnormality, and a local outlier factor is used to quantify the degree of abnormality; the abnormal data are repaired by Kalman filter interpolation; S4, data delivery optimization is implemented, an ant colony algorithm is used to optimize the routing, including considering the data priority, network load and transmission delay; a LDPC code fountain code technology is used to realize redundant transmission; a token bucket algorithm is used to schedule the data with high real-time performance; S5, data trend prediction is performed, an LSTM and ESN fusion prediction model is constructed, the characteristics of the two are combined to predict the trend; and the model parameters are updated through a sliding window mechanism.

[0005] In the preferred scheme, step S1 further comprises: firstly, collecting system automatically collected data and manually collected data, eliminating invalid values, and normalizing time stamps; then, using MATLAB Curve Fitting Toolbox or Python scipy.interpolate module to realize B-spline curve fitting, generating node vectors by setting order and control point number, calculating basis functions and constructing curve equation, and solving control point coordinates by least square method to complete time axis mapping; For data points with large time deviation, quantum genetic algorithm is realized by means of Python QGA library or self-programming, and after initializing population, time window is iteratively optimized by measurement, selection, crossover and mutation, and the optimal value is output and the data is matched again; The overall process can be integrated into a power plant data management system developed based on Java or C#, and the above algorithm modules are called to realize automatic operation, and the data and results in the processing process are stored in combination with the database module.

[0006] In the preferred scheme, S1.1 adopts B-spline curve fitting algorithm for time axis mapping: collecting system automatically collected data and manually collected data, and eliminating invalid values; Set the order k of B-spline curve, determine the number of control points p according to the data distribution density, and generate node vector , wherein the nodes satisfy , , the intermediate nodes adopt uniform distribution or Chebyshev distribution; According to the node vector U, the B-spline basis function is calculated by the recursive formula , wherein when , , when , , wherein when the denominator is 0, the corresponding item is 0; Take the control point coordinates as parameters to construct the B-spline curve equation , solve the control point coordinates by least square method, and then substitute the time stamps of the two types of data into the curve equation to obtain the aligned time-data mapping relationship; S1.2 For data points with large time deviation, quantum genetic algorithm is introduced to optimize the time window: set the candidate range of the time window as , take the time window size T as the optimization parameter, and each time window corresponds to quantum chromosome , wherein and are complex amplitudes and satisfy ; Construct fitness function , wherein r is the proportion of the number of matching pairs of data in the window to the total amount of data, is the sum of squares of deviations of the data of the matching pairs, is the weight; Initialize the population, generate N quantum chromosomes and randomly initialize and , measure the time window T for each chromosome and calculate the fitness, retain the top 20% of chromosomes with the highest fitness as parents, generate offspring through quantum gate operations, iterate until the preset number of iterations or the fitness is stable, and output the optimal time window ; , the data points of the time deviation are re-matched to complete alignment.

[0007] In the preferred embodiment, step S2 further comprises: collecting raw data from different data sources, determining the recognition framework and obtaining the basic probability distribution of each data source, using the PyDSTool library of Python to implement the improved Dempster-Shafer combination rule, introducing a credibility factor through programming for weighted processing, calculating the conflict degree, assigning the conflict according to the rule and completing the combination after calculating the conflict degree; For cloud model quantification of uncertainty, use the scikit-fuzzy library of Python or write a module yourself, calculate the expected value, entropy and hyper-entropy according to the data samples, and generate cloud droplets to realize the quantification of uncertainty; For gray relational analysis of key parameters, use the Grey System Toolbox of MATLAB or the greyrelationalanalysis module of Python, input the reference sequence and comparison sequence, and calculate the correlation coefficient and correlation degree after dimensionless processing; The overall process can be integrated into a power plant data fusion module developed based on C++, and the library in step S1 above and the MySQL database module are called to store the intermediate results and final fusion data, ensuring that each step is automatically executed to complete the data fusion process.

[0008] In the preferred embodiment, the improved Dempster-Shafer combination rule is used to process data: determine the recognition framework containing possible hypotheses such as equipment operating state , collect the basic probability distribution of each hypothesis in the recognition framework from different data sources ; Introduce a credibility factor , and the weighted processing of the basic probability distribution is , is a subset of ; Calculate the conflict degree between data ​A, B, ..., N are A subset of the set whose pairwise intersection is empty; when At that time, according to a certain subset The conflicting part of the weighted support allocation, namely ; Synthetic final basic probability assignment ,in A subset corresponding to each data source; Quantifying data uncertainty using cloud models: for each data sample Determine the digital characteristics of the cloud model, where the expected value is... , Entropy is the number of samples. hyperentropy , The local entropy for each sample, The average entropy is used to quantify the randomness and fuzziness of the data by generating cloud droplets; The correlation degree of key parameters such as temperature and pressure was calculated using grey relational analysis algorithm to determine the reference sequence. and comparison sequences ,in The value is the dimensionless value. The original value; The initialization method is used for dimensionless processing, that is... ;in The value is the dimensionless value. The original value; Calculate the correlation coefficient , The resolution coefficient, ; Calculate the correlation The degree of correlation is used as the weighting basis for data fusion.

[0009] In the preferred scheme, the multi-source data of the power plant is first preprocessed to remove duplicate and formatted data to obtain the dataset to be detected; the IsolationForest module in the scikit-learn library of Python is used as the basis, and the density peak clustering idea is introduced by combining custom programming to calculate the local density and minimum distance of the samples to determine the initial candidate points of anomalies, adjust the isolation tree construction process, and call the LocalOutlierFactor module in the same library to calculate the local outlier factor to quantify the degree of anomaly; For outlier data, the Kalman filter module in Python's filterpy library is used to implement Kalman filter interpolation repair. By configuring the state transition matrix, observation matrix and noise covariance parameters, the prediction and update steps are completed to obtain the optimal estimate to replace the outlier data. The whole process can be integrated into a Python-based power plant data processing system, combining the pandas module for data preprocessing, and the MySQL database module for storing detection results and repaired data.

[0010] In the preferred scheme, the data detection and repair includes the following steps: using the improved isolation forest algorithm and introducing density peak clustering for anomaly detection: pre-processing the power plant multi-source data to obtain the data set to be detected ; Calculate the local density of each sample , , the Euclidean distance between samples , the cutoff distance , the indicator function and when , otherwise 0; Calculate the minimum distance of each sample to the sample with higher local density than itself , the of the sample with the highest local density; take the maximum value of all ; Select the sample with larger and as the initial abnormal candidate point, and preferentially put it into the root node of the isolation tree; , the average path length of the sample in all isolation trees, , the average path length constant when the sample size is n, greater than the set threshold value is marked as abnormal data; Quantify the abnormal degree by the local outlier factor, calculate , where is the k , distance from sample y to its kth nearest neighbor; Repair abnormal data using Kalman filter interpolation: for the marked abnormal data, establish a system state space model, the state equation is , the observation equation is , is the system state at time t, A is the state transition matrix, is the process noise and obeys the Gaussian distribution with mean 0 and covariance Q, is the observation value, H is the observation matrix, is the observation noise and obeys the Gaussian distribution with mean 0 and covariance R; Through Kalman filter iteration, the prediction step calculates and​ , , , , , , , , , , ,

[0011] In the preferred embodiment, step S4 further comprises: first constructing a power plant data transmission network topology model, defining nodes and links, using the ACO-Pathfinding library of Python or programming an ant colony algorithm by oneself, obtaining the optimal route through iteration by configuring the link weight function parameters, setting the number of ants, pheromone parameters, etc. For the fountain code technology of LDPC code, with the help of the LDPC encoding module and the fountain code generation tool in the Communications Toolbox of MATLAB, the check block and the encoding package are generated after the original data is blocked, and the redundant transmission is realized; for the data with high real-time performance, the token bucket algorithm is configured by using the traffic control tool of the Linux system, the token bucket capacity and the generation rate are set, or the flow scheduling is realized by programming the token-bucket library of Python; The overall process is integrated into the power plant data transmission management system based on C++, and the network state is monitored in combination with the libpcap library, and the routing information and transmission log are stored in the MySQL database.

[0012] In the preferred embodiment, the ant colony algorithm is used to optimize the route: a network topology model is constructed, the nodes are regarded as vertices , and the links are regarded as edges ; a link weight function is defined , P is the data priority, , is the network load rate, , is the transmission delay, ; The number of ants N, the initial pheromone concentration , the evaporation coefficient , the heuristic factor and the factor are initialized; the ants construct the path according to the transition probability ; The pheromone is updated as , wherein , the optimal route is iteratively outputted; A fountain code technology using an LDPC code is used to realize redundancy transmission: original data D is divided into blocks ; a check matrix H is generated based on the LDPC code, and a check block is generated, satisfying ; an encoded packet is generated through random linear combination , wherein is a data block or a check block, , and a transmission end continuously transmits to a receiving end to recover the original data; The token bucket algorithm is used to schedule data with high real-time performance: the token bucket capacity B and the token generation rate r are configured; tokens are added to the bucket at a rate of r, and each unit of data consumes one token when transmitting real-time data; if the tokens are sufficient, the transmission is performed, otherwise, the transmission is waited or discarded; r is dynamically adjusted to adapt to the flow change and ensure the priority transmission of critical data.

[0013] In the preferred scheme, the historical operation data of the power plant is collected and normalized, an LSTM submodel is constructed using the TensorFlow or Keras library of Python, the number of hidden units and the parameters of the gating mechanism are configured, and the Adam optimizer is trained to minimize the mean square error; An ESN submodel is built with the pyESN library of Python, the size of the reservoir and the sparsity parameters are set, and the output weight is solved by ridge regression; The weighted fusion of LSTM and ESN is realized through self-defined programming, and the fusion weight is optimized based on the validation set; for the sliding window mechanism, the pandas library of Python is used to process the sliding update of time series data, the window size and update period are set, and the incremental learning module of scikit-learn is used to fine-tune the LSTM weight and recalculate the ESN output weight; the overall process is integrated into a power plant data prediction system based on Python, and a MySQL database is used to store historical data and model parameters.

[0014] The power plant multi-source data arrangement and delivery method realizes high-precision alignment and fusion of multi-source data, effectively solves the data deviation problem caused by inconsistent time; through the abnormality detection and repair mechanism, the data quality is improved; the optimized data delivery strategy ensures the reliability and real-time performance of data transmission; the prediction model provides forward-looking support for data processing, and overall improves the intelligent level and decision accuracy of power plant data management. BRIEF DESCRIPTION OF DRAWINGS

[0015] The application will be further described below in combination with the drawings and examples: Figure 1It is a power plant multi-source data arrangement and delivery method. DETAILED DESCRIPTION

[0016] Embodiment 1 As shown in the figure, a power plant multi-source data arrangement and delivery method, the method comprising: Figure 1 S1, constructing a multi-scale data alignment model, using a B-spline curve fitting algorithm for time axis mapping, adjusting the curve through control points; for data with large time deviation, introducing quantum genetic algorithm to optimize the time window; S2, data fusion processing, using improved Dempster-Shafer synthesis rule, including credibility factor and conflict redistribution, quantifying uncertainty with cloud model; for key parameters, use grey correlation analysis algorithm to calculate correlation degree; S3, abnormal data detection and repair, using improved isolation forest algorithm and introducing density peak clustering to detect anomalies, quantifying the degree of abnormality by local outlier factor; abnormal data is repaired by Kalman filter interpolation; S4, data delivery optimization, using ant colony algorithm to optimize routing, including considering data priority, network load, and transmission delay; using LDPC code to realize redundant transmission; for data with high real-time performance, use token bucket algorithm for scheduling; S5: data trend prediction, construct a LSTM and ESN fusion prediction model, combine the characteristics of both to predict trends; update model parameters through sliding window mechanism. Embodiment 2

[0017] Further illustrated in combination with Embodiment 1, step S1 further comprises: first collecting system automatically collected data and manually collected data, eliminating invalid values, and normalizing time stamps; then using MATLAB Curve Fitting Toolbox or Python scipy.interpolate module to realize B-spline curve fitting, generating node vector by setting order and control point number, calculating basis function and constructing curve equation, and solving control point coordinates to complete time axis mapping; For data points with large time deviation, use Python QGA library or self-programming to realize quantum genetic algorithm, initialize the population, and optimize the time window through measurement, selection, crossover, and mutation iteration, output the optimal value and match the data again; The overall process can be integrated into a power plant data management system developed based on Java or C#, and the above algorithm modules can be called to realize automatic operation, and the database module can be used to store data and results during processing.

[0018] ​In the preferred scheme, S1.1 uses a B-spline curve fitting algorithm for time axis mapping: the data acquisition system automatically collects data and manual inspection collects data, and invalid values ​​are removed; Define the order k of the B-spline curve, determine the number p of control points based on the data distribution density, and generate the node vector. The nodes satisfy , The intermediate nodes adopt a uniform distribution or a Chebyshev distribution; Based on the node vector U, the B-spline basis function is calculated using a recursive formula. , among which when hour, ,when hour, When the denominator is 0, the corresponding term is 0; Using control point coordinates Construct the B-spline curve equation using parameters. The coordinates of the control points are solved by the least squares method, and then the timestamps of the two types of data are substituted into the curve equation to obtain the aligned time-data mapping relationship. S1.2 For data points with large time deviations, a quantum genetic algorithm is introduced to optimize the time window: the candidate range of the time window is set as follows. The time window size T is used as the parameter to be optimized, and each time window corresponds to a quantum chromosome. ,in and It is a complex amplitude and satisfies ; Constructing the fitness function Where r is the proportion of the number of matching pairs within the window to the total amount of data. The sum of squared deviations of the matched pairs of data. As weight; Initialize the population, generate N quantum chromosomes and initialize them randomly. and For each chromosome, a time window T is measured and its fitness is calculated. The top 20% of chromosomes in terms of fitness are retained as parents, and offspring are generated through quantum gate operations. The process is repeated until a preset number of iterations or when the fitness stabilizes, at which point the optimal time window is output. ; Time deviation Data points, using Rematch to complete the alignment.

[0019] The specific steps are as follows: A multi-scale data alignment model was constructed, and a B-spline curve fitting algorithm was used for time axis mapping. The system automatically collects data (referred to as datasets). , containing timestamp and corresponding monitoring value ) and manual inspection data collection (denoted as dataset , containing timestamp and corresponding monitoring value ), where n, m are the sample sizes of the two types of data respectively. Clean the data, remove invalid values and duplicate records.

[0020] Map the system collection timestamp and the manual collection timestamp to the [0, 1] interval, denoted as , , where is the minimum value of all timestamps, is the maximum value, ensuring consistent time axis scale.

[0021] Select a 3-order B-spline curve and set the number of control points p. Generate the node vector , where the first nodes and the last nodes take 0 and 1 (i.e. , ), and the middle nodes are linearly distributed ( ).

[0022] Based on the node vector U, calculate the basis function ( ) through the recursive formula, which is the basic building block of the curve.

[0023] Construct the B-spline curve equation , where is the monitoring value parameter of the control point. To minimize the deviation between the curve predicted value and the original data, solve (i.e. minimize the error function ) by least squares method. Substitute the timestamps of the two types of data into the fitted curve to obtain the aligned time-monitoring value mapping relationship, realizing the time axis unification of data with different collection frequencies.

[0024] B-spline basis function recursive formula: when , ; when , (the denominator is 0, this term is 0), define the basic form of B-spline curve, realize the smoothness and local controllability of the curve through the linear combination of basis functions, adapt to the different collection frequency characteristics of system and manual data.

[0025] Control point parameter solving formula: error function E to Taking the partial derivatives and setting them to zero, we obtain the system of linear equations. ,in: ; ; Solving for the given information ; The optimal control point parameters are determined by the least squares method to ensure that the B-spline curve fits the original data to the maximum extent and improve the accuracy of time axis mapping.

[0026] The step of optimizing the time window using a quantum genetic algorithm is as follows: Calculate the deviation of timestamps in the original data. Set a regular threshold (e.g., 5 seconds), when Data points marked as having large time deviations need to be rematched by optimizing the time window.

[0027] Time window search range: (Based on historical data deviation statistics settings, such as...) , ).

[0028] Population size: (quantum chromosome number).

[0029] Quantum chromosome encoding: Each chromosome is represented by L qubits, and the k-th qubit is... ,satisfy The binary code is obtained through measurement and then converted into a candidate value T for the time window.

[0030] For candidate time windows T, calculate: matching rate ; Matching deviation ( (for matching pairs). The fitness function is defined as follows: ( For weights, such as A higher value indicates a better time window.

[0031] Quantum genetic iterative optimization: Measurement and selection: For each quantum chromosome, the time window T is measured, the fitness is calculated, and the top 20% of high-fitness chromosomes are retained as parents.

[0032] Quantum crossover: Crossover is achieved by applying a Hadamard gate to the parent chromosome, updating the complex amplitude. , .

[0033] Quantum mutation: in terms of probability Apply Pauli-X gate and swap complex amplitudes , .

[0034] Termination condition: output the time window with the highest fitness after 100 iterations .

[0035] Data alignment based on the optimal time window: for the "data points with large time deviation" marked, re-matching is adopted, that is, when , it is considered as a valid match, and the time axis alignment is completed. Mathematical formula and use: quantum chromosome measurement probability formula: for a quantum bit

[0036] , the probability of the measurement result being "0" , and the probability of being "1" , the measurement value is determined by random number generation; realize the mapping of quantum chromosome to time window candidate value, and ensure the diversity of population.

[0037] Fitness function standardization formula ( is the minimum and maximum value of the current population fitness) Normalize the fitness value to the interval [0, 1] to facilitate the comparison and selection operation of individuals in the population.

[0038] Through the above B-spline curve fitting algorithm, the time axis smoothing mapping of data with different collection frequencies is realized, and the time window optimized by quantum genetic algorithm solves the matching problem of data points with large time deviation, which together constructs a multi-scale data alignment model and lays a foundation for subsequent data fusion.

[0039] Embodiment 3 Further illustrated in combination with Embodiment 1, step S2 further includes: first collecting original data of different data sources, determining the recognition framework and obtaining the basic probability distribution of each data source, using the PyDSTool library of Python to realize the improved Dempster-Shafer synthesis rule, introducing the credibility factor through programming for weighted processing, calculating the conflict degree, and then distributing the conflict according to the rule and completing the synthesis; For cloud model quantization of uncertainty, with the help of Python's scikit-fuzzy library or self-written module, calculate the expected value, entropy and hyperentropy, etc. Digital features according to data samples, generate cloud droplets to realize uncertainty quantization; For gray relational analysis of key parameters, use MATLAB's Grey System Toolbox or Python's greyrelationalanalysis module, input reference sequence and comparison sequence, calculate correlation coefficient and correlation degree after dimensionless processing; ​The whole process can be integrated into a power plant data fusion module developed based on C++, and the intermediate results and final fusion data are stored by calling the library in step S1 and the MySQL database module, to ensure automatic execution of each step to complete the data fusion process.

[0040] In the preferred scheme, the improved Dempster-Shafer combination rule is used to process data: determine the identification framework containing possible assumptions such as the operating state of the equipment , collect the basic probability distribution of each hypothesis in the identification framework from different data sources ; Introduce a credibility factor , and weight the basic probability distribution to obtain , is a subset of ; Calculate the degree of conflict between data , A, B,..., N are subsets of with an empty intersection; when , the conflict part is allocated according to the weighted support of the subset , that is ; and the final basic probability distribution is synthesized , where is the subset corresponding to each data source; Quantify the uncertainty of data using a cloud model: for each data sample , determine the digital characteristics of the cloud model, where the expectation , is the number of samples, the entropy , and the hyper entropy , is the local entropy of each sample, is the average entropy, and the randomness and fuzziness of the data are quantified by generating cloud droplets; Calculate the correlation degree using the grey correlation analysis algorithm for temperature and pressure key parameters: determine the reference sequence and the comparison sequence , where is the dimensionless value, is the original value; Use the initial value method for dimensionless processing, that is ; where is the dimensionless value, is the original value; Calculate the correlation coefficient , is the resolution coefficient, ; Calculate the correlation degree , and use the correlation degree as the weight basis for data fusion.

[0041] S2, data fusion processing is performed, and the specific steps are as follows: The improved Dempster-Shafer combination rule is used to process data: First, determine the identification framework , which contains all possible hypotheses. For power plant multi-source data, the hypothesis may be the equipment operating state reflected by the data.

[0042] Collect the basic probability assignment (BPA) of each hypothesis in the identification framework from different data sources (system automatically collected data and manually collected data), denoted as , where is the number of data sources.

[0043] Introduce a credibility factor to measure the reliability of each data source, The value of ranges from 0 to 1, which can be determined according to the historical performance of the data source, the accuracy of the collection equipment, etc. If a data source has long-term stable performance and high accuracy of the collection equipment, the value of

[0044] Weighted processing of the basic probability assignment of each data source is performed to obtain the weighted basic probability assignment , where is a subset of the identification framework.

[0045] Calculate the conflict degree between data , the formula is , where A, B,..., N are subsets of the identification framework and intersect with each other.

[0046] When , conflict redistribution is performed. The conflict part is distributed according to the weighted support of each subset, and the conflict distribution obtained by a subset is .

[0047] Synthesize the final basic probability assignment , the formula is , where is the subset corresponding to each data source. The use of this formula is to avoid the failure of the traditional D-S combination rule under high conflict by using the credibility factor and conflict redistribution strategy, making the synthesis result more reliable.

[0048] 2. Quantify the uncertainty of data using cloud model: For each data sample , determine its corresponding cloud model of three digital features: the expected Ex, entropy En and hyper-entropy He. The expected Ex is the central tendency of the data, which can be estimated by the sample mean, that is , where is the sample size; the entropy En reflects the fuzziness and randomness of the data, which can be calculated according to the degree of dispersion of the data, such as ; the hyper-entropy He is the uncertainty measure of the entropy, which can be estimated by the standard deviation of the entropy, that is , where is the local entropy corresponding to each sample, is the average entropy.

[0049] According to the digital features of the cloud model, cloud droplets are generated, and the distribution of the cloud droplets is used to quantify the uncertainty of the data, and the randomness and fuzziness of the data are uniformly represented. The use of this digital feature calculation is to extract the key features from the data to quantify the uncertainty of the data and provide more comprehensive information for subsequent data fusion.

[0050] 3. Calculate the correlation degree of the key parameters using the grey correlation analysis algorithm: Determine the reference sequence and the comparison sequence. The reference sequence is the standard sequence or ideal sequence of the actual physical quantity; the comparison sequence is the sequence of key parameters collected from different data sources, , is the number of data sources, is the data dimension.

[0051] Non-dimensional processing is performed on the reference sequence and the comparison sequence to eliminate the influence of dimension. The initial value method is used, that is , where is the non-dimensional value, is the original value.

[0052] Calculate the correlation coefficient , the formula is , where is the resolution coefficient, with a value range of [0, 1], usually taking 0.5; is the two-level minimum difference, is the two-level maximum difference. The use of this formula is to measure the correlation degree of the comparison sequence and the reference sequence at a certain time.

[0053] Calculate the correlation degree , the formula is , the larger the correlation degree , the closer the correlation between the key parameters of the data source and the actual physical quantity. The use of this formula is to comprehensively consider the correlation coefficients at each time to obtain the overall correlation degree of the comparison sequence and the reference sequence, which provides the weight basis for data fusion, and the data source with a larger correlation degree is given a higher weight in the fusion.

[0054] Example 4 To further illustrate with reference to Example 1, in step S3, the multi-source data of the power plant is first preprocessed to remove duplicate and formatted data, resulting in the dataset to be detected. Based on the IsolationForest module in the scikit-learn library of Python, the density peak clustering idea is introduced by combining custom programming to calculate the local density and minimum distance of the samples to determine the initial candidate points of anomalies. The isolation tree construction process is adjusted, and the LocalOutlierFactor module in the same library is called to calculate the local outlier factor to quantify the degree of anomaly. For outlier data, the Kalman filter module in Python's filterpy library is used to implement Kalman filter interpolation repair. By configuring the state transition matrix, observation matrix and noise covariance parameters, the prediction and update steps are completed to obtain the optimal estimate to replace the outlier data. The entire process can be integrated into a Python-based power plant data processing system, using the pandas module for data preprocessing and the MySQL database module to store the detection results and the repaired data.

[0055] In the preferred scheme, data detection and repair include the following steps: using an improved isolated forest algorithm and introducing density peak clustering for anomaly detection; preprocessing multi-source data from the power plant to obtain the dataset to be detected. ; Calculate each sample Local density , Let Euclidean distance be the distance between samples. To cut off the distance, For indicator functions and hour Otherwise, it is 0; Calculate each sample Minimum distance to samples with a local density higher than itself The sample with the highest local density Take all The maximum value; select and Samples with larger average values ​​are used as initial anomaly candidates and are preferentially placed near the root node of the isolated tree. Construct an improved isolation forest and calculate sample anomaly scores. ( The average path length of the sample across all isolated trees. Let n be the average path length constant when the sample size is n. Data exceeding a set threshold is marked as abnormal. The degree of anomaly is quantified by local outlier factors, and the calculation is performed. wherein, is k , is the distance from the sample y to its kth nearest neighbor; Repairing abnormal data by Kalman filter interpolation: for the marked abnormal data, a system state space model is established, the state equation is , and the observation equation is , is the system state at time t, A is the state transition matrix, is the process noise and obeys the Gaussian distribution with mean 0 and covariance Q, is the observation value, H is the observation matrix, is the observation noise and obeys the Gaussian distribution with mean 0 and covariance R; Through Kalman filter iteration, the prediction step calculates and , is the optimal estimation at time t, is the estimation covariance at time t, the update step calculates the Kalman gain , the optimal estimation and the update covariance , I is the identity matrix; Replace the abnormal data point with the optimal estimation to achieve repair.

[0056] S3, carry out abnormal data detection and repair, the specific steps are as follows: 1. Use the improved isolation forest algorithm and introduce density peak clustering for abnormal detection: first, preprocess the power plant multi-source data to remove duplicate data and obviously format error data, and obtain the data set to be detected wherein is the ith data sample, and N is the total number of samples.

[0057] The idea of density peak clustering is introduced to determine the initial abnormal candidate points of the isolation forest. The local density of each sample is calculated , the formula is wherein is the Euclidean distance between samples and , is the cutoff distance, is the indicator function, when , , otherwise The formula is used to measure the density of data around the sample. The sample with low local density is more likely to be an abnormal candidate point.

[0058] Calculate the average path length of each sample to the sample with local density higher than itself The formula is For the sample with the highest local density, Take the maximum value of all . Reflect the distance of the sample from the high-density area, The larger the sample, the more likely it is to be far from the normal data distribution.

[0059] Select the sample with both and larger as the initial abnormal candidate point, and put it in the root node of the isolation tree first, to improve the sensitivity of the isolation forest to abnormal data.

[0060] Construct an improved isolation forest: randomly select part of the features and samples, recursively divide the data to construct multiple isolation trees, and calculate the abnormal score of each sample The formula is Where is the average path length of sample in all isolation trees, is the average path length constant when the sample size is n. When is greater than the set threshold (such as 0.7), it is marked as abnormal data. This formula quantifies the abnormality of the sample by path length. The higher the score, the more likely the sample is abnormal.

[0061] Introduce the local outlier factor (LOF) to further quantify the abnormality. For the detected abnormal data, calculate its local outlier factor The formula is Where is the k th nearest neighbor of sample , is the distance from sample y to its kth nearest neighbor. The larger the value, the higher the abnormality of the sample, which is used to distinguish different severity of abnormalities.

[0062] 2. Abnormal data is repaired by Kalman filter interpolation: for the data marked as abnormal, it is regarded as missing value, and Kalman filter is used for interpolation repair. Establish the system state space model, the state equation is The observation equation is Where is the system state at time t, A is the state transition matrix, The process noise (follows a Gaussian distribution with mean 0 and covariance Q). Let H be the observation value at time t, and H be the observation matrix. The observed noise follows a Gaussian distribution with a mean of 0 and a covariance of R. This model is used to describe the temporal variation of the data and provides a foundation for Kalman filtering.

[0063] Kalman filter iterative process: Prediction Steps: Calculate the predicted state Predicting covariance ,in for The optimal estimate of the time. for The estimated covariance at time.

[0064] Update steps: Calculate Kalman gain Optimal estimation Update covariance , where I is the identity matrix.

[0065] For outlier data points, the optimal estimate obtained using Kalman filtering Replacement is used to interpolate and repair abnormal data, restoring the continuity and rationality of the data sequence.

[0066] Example 5 In conjunction with Example 1, step S4 further includes: first constructing a power plant data transmission network topology model, identifying nodes and links, using Python's ACO-Pathfinding library or implementing an ant colony algorithm through self-programming, configuring link weight function parameters, setting the number of ants, pheromone parameters, etc., and iteratively obtaining the optimal route; For fountain code technology of LDPC codes, the LDPC encoding module and fountain code generation tool in MATLAB's Communications Toolbox are used to divide the original data into blocks and generate check blocks and encoded packets to achieve redundant transmission. For data with high real-time requirements, the token bucket algorithm is configured using the traffic control tool of the Linux system, and the token bucket capacity and generation rate are set, or traffic scheduling is implemented through Python's token-bucket library. The entire process is integrated into a C++-based power plant data transmission management system, using the libpcap library to monitor network status and a MySQL database to store routing information and transmission logs.

[0067] In the preferred scheme, the ant colony algorithm is used to optimize routing: a network topology model is constructed, and nodes are regarded as vertices. The link is regarded as an edge Define link weight function P represents data priority. For network load rate, For transmission delay, ; Initialize the number of ants N and the initial pheromone concentration. Evaporation coefficient Heuristic factors and factors Ants according to transfer probability Build path; Pheromones updated to ,in Iterate and output the optimal route; Redundant transmission is achieved using fountain code technology based on LDPC codes: the original data D is divided into blocks. Generate a parity check matrix H based on the LDPC code, and generate a parity block. ,satisfy ; Generate coded packets through random linear combinations ,in, For data blocks or check blocks, The transmitting end continuously sends data to the receiving end to restore the original data; S4. Implement data transmission optimization, the specific steps are as follows: 1. Optimize routing using the ant colony algorithm, taking into account data priority, network load, and transmission latency: Construct a power plant data transmission network topology model, treating the nodes in the network as vertices. The communication links between nodes are considered as edges. ,in Indicates from node arrive The link.

[0068] Define link weight function Taking into account data priority, network load, and transmission latency, the formula is: Where P is the priority of the currently transmitted data. For link The current network load rate (between 0 and 1, where 1 indicates full load). For link Transmission delay (unit: ms). The weighting coefficients are satisfied. This formula is used to quantify the overall transmission cost of a link, providing a basis for route selection.

[0069] Initialize ant colony algorithm parameters: Set the number of ants N and the initial pheromone concentration for each link. (constant), pheromone evaporation coefficient , heuristic factor , pheromone importance factor and heuristic importance factor .

[0070] Ant path construction: each ant starts from the source node, and selects the next hop node according to the transition probability , where is the set of nodes not visited by ant k. When the ant reaches the destination node, record its path and total weight.

[0071] Pheromone update: after all ants complete path construction, calculate the total weight of each path , the better the path, the greater the pheromone increment. The pheromone update formula is , where , Q is the pheromone constant. Through iterative updating, the pheromone concentration of high-quality paths gradually increases, guiding subsequent ants to select better routes.

[0072] Iteration termination: when the number of iterations reaches the preset value or the path weight tends to be stable, output the optimal routing path.

[0073] 2. Implement redundancy transmission using LDPC code-based fountain code technology: Block the original data that needs to be transmitted, let the original data be , divide it into $ k$ data blocks .

[0074] Generate a check matrix H based on the LDPC code, encode the data blocks through the check matrix, and generate check blocks , so that the original data blocks and the check blocks satisfy .

[0075] Redundancy packaging using fountain code technology: based on the original data blocks and the check blocks, generate an infinite number of encoding packets through random linear combination , where is the i-th original data block or check block, is a random coefficient, and t is the encoding packet number.

[0076] The transmitting end continuously sends the encoding packets until the receiving end feedbacks that the original data has been successfully recovered. After receiving a sufficient number of encoding packets, the receiving end solves the linear equations through Gaussian elimination method to recover the original data blocks, realizing the redundancy transmission of data, and ensuring the data integrity when the network packet is lost.

[0077] 3. Use token bucket algorithm to schedule data with high real-time performance: Configure token bucket parameters for high real-time data (such as device fault alarms and real-time monitoring data of key parameters): token bucket capacity B (maximum number of tokens that can be cached), token generation rate r (number of tokens generated per unit time, matched with data transmission rate).

[0078] Token bucket working mechanism: the system continuously adds tokens to the bucket at a rate of r, and no longer adds when the number of tokens in the bucket reaches B. When there is real-time data to be transmitted, each unit of data consumes one token, and if the bucket has enough tokens, transmission is allowed; if the tokens are insufficient, the data waits or is discarded (according to priority settings).

[0079] Dynamic adjustment of parameters: dynamically adjust the token generation rate r according to the flow changes of real-time data, appropriately increase r during peak flow period to ensure timely transmission of key data, and reduce r during low flow period to avoid waste of network resources. Through the token bucket algorithm, the flow control of high real-time data is realized to ensure its priority transmission and not to exceed the network bandwidth limit.

[0080] Example 6 Further illustrated in Example 1, for high real-time data, use token bucket algorithm for scheduling: configure token bucket capacity B and token generation rate r; add tokens to the bucket at a rate of r, consume one token per unit of data when transmitting real-time data, transmit if the tokens are sufficient, or wait or discard if the tokens are insufficient; dynamically adjust r to adapt to flow changes to ensure priority transmission of key data.

[0081] In the preferred scheme, first collect the historical operation data of the power plant and normalize the data, use TensorFlow or Keras library of Python to build LSTM sub-model, configure the number of hidden units and gate mechanism parameters, and train through Adam optimizer to minimize mean square error; ESN sub-model is built with pyESN library of Python, reserve pool size and sparsity parameters are set, and output weight is solved by ridge regression; The weighted fusion of LSTM and ESN is realized by self-defined programming, and the fusion weight is optimized based on the validation set; for the sliding window mechanism, the pandas library of Python is used to process the sliding update of time series data, the window size and update period are set, and the incremental learning module of scikit-learn is used to fine-tune the LSTM weight and recalculate the ESN output weight; the overall process is integrated into the Python-based power plant data prediction system, and the MySQL database is used to store historical data and model parameters.

[0082] The token bucket algorithm scheduling can guarantee the priority transmission of real-time data, the dynamic adjustment of the token generation rate can adapt to the flow change, and the waste of network resources is avoided; the prediction model of the fusion of LSTM and ESN combines the characteristics of both, improves the prediction accuracy of data trend, and the sliding window mechanism dynamically updates the data and parameters to make the model adapt to the change of the power plant operation condition, guarantee the timeliness and reliability of the prediction, and the overall scheme improves the effectiveness of the power plant data transmission and the accuracy of the data prediction, and provides strong support for the power plant data management.

[0083] The above embodiments are only preferred technical solutions of the present application, and should not be regarded as a limitation of the present application. The protection scope of the present application should be based on the technical solutions recited in the claims, including equivalent replacement solutions of the technical features recited in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present application.

Claims

1. A method for collating and delivering multi-source data of a power plant, characterized in that: The method comprises: S1, constructing a multi-scale data alignment model, using a B-spline curve fitting algorithm for time axis mapping, adjusting the curve through control points; for data with large time deviation, introducing a quantum genetic algorithm to optimize the time window; S2, data fusion processing, using improved Dempster-Shafer synthesis rule, including credibility factor and conflict redistribution, quantifying uncertainty with cloud model; for key parameters, use grey correlation analysis algorithm to calculate correlation degree; S3, abnormal data detection and repair, use improved isolation forest algorithm and introduce density peak clustering to detect anomalies, quantify the degree of anomaly by local outlier factor; abnormal data is repaired by Kalman filter interpolation; S4, data transmission optimization, using ant colony algorithm to optimize routing, including considering data priority, network load, transmission delay; use LDPC code to realize redundant transmission; for real-time data, use token bucket algorithm to schedule; S5: data trend prediction, build LSTM and ESN fusion prediction model, combine the characteristics of the two to predict trends; update model parameters through sliding window mechanism.

2. The method of claim 1, wherein the method further comprises: Step S1 further comprises: First, collect system automatically collected data and manual inspection collected data, eliminate invalid values, and normalize the timestamp; then use MATLAB Curve Fitting Toolbox or Python scipy.interpolate module to realize B-spline curve fitting, generate node vector by setting order and control point number, calculate basis function and construct curve equation, and solve control point coordinates to complete time axis mapping; For data points with large time deviation, use Python QGA library or self-programming to implement quantum genetic algorithm, initialize the population, and optimize the time window through measurement, selection, crossover, and mutation iteration, output the optimal value and match the data again; The overall process can be integrated into a power plant data management system developed based on Java or C#, and the above algorithm modules can be called to realize automatic operation, and the database module can be used to store data and results during processing.

3. The power plant multi-source data processing and transmission method according to claim 2, characterized in that: S1.1, using B-spline curve fitting algorithm for time axis mapping: collecting system automatically collected data and manual inspection collected data, eliminating invalid values; Setting the order k of the B-spline curve, determining the number p of control points according to the data distribution density, and generating a node vector wherein the nodes satisfy , the intermediate nodes adopt uniform distribution or Chebyshev distribution; The B-spline basis functions are computed by a recursive formula from the node vector U where when , when , where the corresponding term is taken as 0 when the denominator is 0. With the control point coordinates as parameters, the B-spline curve equation is constructed , the control point coordinates are solved by the least square method, and the timestamps of the two types of data are respectively substituted into the curve equation to obtain the aligned time-data mapping relationship. S1.2 For data points with large time deviation, introduce quantum genetic algorithm to optimize time window: set the candidate range of time window as , and take the time window size T as the optimization parameter, and each time window corresponds to a quantum chromosome , where , and is the complex amplitude and satisfies ; Constructing fitness function where r is the proportion of matching pairs in the window, is the sum of squared deviations of the matching pairs, is the weight; Initialize the population, generate N quantum chromosomes and randomly initialize And For each chromosome, measure the time window T and calculate the fitness, keep the top 20% of chromosomes as parents, generate offspring through quantum gate operation, iterate until the preset number of times or the fitness is stable, output the optimal time window ; Time deviation Data points, using Rematch to complete the alignment.

4. The method of claim 1, wherein the method further comprises: Step S2 further comprises: first collect raw data from different data sources, determine the recognition framework and obtain the basic probability assignment of each data source, use Python PyDSTool library to realize improved Dempster-Shafer synthesis rule, introduce credibility factor through programming for weighted processing, calculate the conflict degree, and distribute the conflict according to the rule and complete the synthesis; For cloud model quantifying uncertainty, use Python scikit-fuzzy library or self-programming module to calculate the expected value, entropy and hyperentropy of the data sample, and generate cloud droplets to quantify uncertainty; For the grey correlation analysis of key parameters, use MATLAB's Grey System Toolbox or Python's greyrelationalanalysis module to input the reference sequence and comparison sequence, and calculate the correlation coefficient and correlation degree after dimensionless processing; The overall process can be integrated into a power plant data fusion module developed based on C++, and the intermediate results and final fusion data can be stored by calling the library in step S1 and the MySQL database module to ensure automatic execution of each step to complete data fusion processing.

5. The method of claim 4, wherein the method further comprises: using improved Dempster-Shafer combination rules to process the data: determining a recognition framework containing possible hypotheses such as the state of operation of the device , collecting basic probability assignments of each hypothesis in the recognition framework from different data sources ; Introducing a credibility factor , the basic probability assignment is weighted to obtain , is a subset of ​ Computing a degree of conflict between data A, B,..., N are subsets whose intersection is empty; when , the conflict part is assigned according to the weighted support of the subset , i.e. ; Synthesizing final basic probability assignments wherein is a subset corresponding to each data source; quantifying the uncertainty of data with cloud model: for each data sample , determine the cloud model digital features, wherein the expectation , is the number of samples, the entropy , the hyper entropy , is the local entropy of each sample, is the average entropy, quantifying the randomness and fuzziness of data by generating cloud drops; The correlation degree of the key parameters of temperature and pressure is calculated by using a grey correlation analysis algorithm: determining a reference sequence and a comparison sequence wherein is a dimensionless value, is an original value; The initial value method is used for non-dimensionalization, i.e. ; wherein is the non-dimensionalized value, is the original value. correlation coefficient , for resolving coefficients, ; Computing the correlation degree The correlation degree is used as the weight basis for data fusion.

6. The method of claim 1, wherein the method further comprises: Step S3 also includes: first, pre-processing the power plant multi-source data to remove duplicate and format error data, and obtaining the detected data set; using the IsolationForest module in the scikit-learn library of Python as the basis, combined with the introduction of density peak clustering idea by custom programming, calculating the local density and minimum distance of the sample to determine the initial abnormal candidate point, adjusting the construction process of the isolation tree, and calling the LocalOutlierFactor module in the library to calculate the local outlier factor to quantify the degree of abnormality; For abnormal data, Kalman filter interpolation repair is realized by using the KalmanFilter module in the filterpy library of Python, and the optimal estimation value is obtained by completing the prediction and update steps to replace the abnormal data by configuring the state transition matrix, observation matrix and noise covariance parameters; The overall process can be integrated into a power plant data processing system based on Python, combined with the pandas module for data preprocessing, and the MySQL database module for storing detection results and repaired data.

7. The method of claim 6, wherein the method further comprises: The data detection and repair includes the following steps: An improved isolation forest algorithm is adopted and density peak clustering is introduced for anomaly detection: a power plant multi-source data preprocessing is performed to obtain a to-be-detected data set ; a local density of each sample is calculated , an Euclidean distance between samples is , a truncation distance is , an indicator function is , and 0 otherwise. Compute each sample The minimum distance to the sample with higher local density than itself The local density of the sample with the highest local density Take the maximum of all Select the sample with the largest and Both are larger as the initial abnormal candidate point, and prefer to put in the root node of the isolated tree Construct an improved Isolation Forest, compute sample anomaly scores is the average path length of the sample in all isolation trees, is the average path length constant for a sample size of n, is flagged as anomalous data if it is greater than a set threshold.​ By quantifying the degree of abnormality through local outlier factor, compute where, is k , is the distance of sample y to its kth nearest neighbor; Kallman filter interpolation is used to repair abnormal data: for the marked abnormal data, the system state space model is established, the state equation is , the observation equation is , , A is the state transition matrix, , the process noise is subject to Gaussian distribution with mean 0 and covariance Q, , H is the observation matrix, , the observation noise is subject to Gaussian distribution with mean 0 and covariance R; The prediction step is calculated through Kalman filtering iteration. and , for Time-optimal estimation for Estimate covariance at each step, and calculate Kalman gain in the update step. Optimal estimation and update covariance I is the identity matrix; with optimal estimates Repair is accomplished by replacing the outlier data points.

8. The method of claim 1, wherein the method further comprises: Step S4 also includes: first, constructing a power plant data transmission network topology model to determine the nodes and links, using the ACO-Pathfinding library of Python or self-programming to realize the ant colony algorithm, configuring the link weight function parameters, setting the number of ants, pheromone parameters, etc., and iterating to obtain the optimal routing; For the fountain code technology of LDPC code, use the LDPC encoding module and fountain code generation tool in MATLAB's Communications Toolbox to generate check blocks and encoding packets after block processing of the original data, and realize redundant transmission; for high real-time data, use the traffic control tool of Linux system to configure the token bucket algorithm, set the token bucket capacity and generation rate, or realize traffic scheduling by programming with Python's token-bucket library; The overall process is integrated into a power plant data transmission management system based on C++, combined with the libpcap library to monitor the network state, and the MySQL database to store routing information and transmission logs.

9. The method of claim 8, wherein the method further comprises: Ant colony algorithm is used to optimize routing: construct network topology model, take nodes as vertices , and links as edges ; define link weight function , P is data priority, is network load rate, is transmission delay, ; Initialize the number of ants N, initial pheromone concentration , evaporation coefficient , heuristic factor and factor ; ants construct paths according to transition probability ; The pheromone is updated as wherein and iteratively outputting the optimal route; The LDPC code is adopted to realize the redundant transmission of the fountain code technology: the original data D is divided into blocks ; a check matrix H is generated based on the LDPC code, and a check block is generated, which satisfies ; an encoded package is generated through random linear combination , wherein is a data block or a check block, , and the transmitting end continuously transmits to the receiving end to recover the original data. For data with high real-time performance, use token bucket algorithm for scheduling: configure token bucket capacity B and token generation rate r; add tokens to the bucket at a rate of r, consume 1 token per unit of data when transmitting real-time data, transmit if tokens are sufficient, or wait or discard if not; dynamically adjust r to adapt to traffic changes and ensure that critical data is transmitted first.

10. The method of claim 1, wherein the method further comprises: Step S5 also includes: first collecting power plant historical operation data and normalizing, using Python's TensorFlow or Keras library to build an LSTM sub-model, configuring the number of hidden units and gate mechanism parameters, and training with the Adam optimizer to minimize the mean square error; With the help of Python's pyESN library, build an ESN sub-model, set the reserve pool size and sparsity parameters, and use ridge regression to solve the output weight; Through self-defined programming, realize the weighted fusion of LSTM and ESN, optimize the fusion weight based on the validation set; for the sliding window mechanism, use Python's pandas library to process time series data sliding update, set the window size and update period, and fine-tune the LSTM weight and recalculate the ESN output weight using the incremental learning module of scikit-learn; integrate the overall process into a Python-based power plant data prediction system, and store historical data and model parameters in a MySQL database.

Images