Multivariable multi-step prediction method and system for complex electromechanical system

CN119066426APending Publication Date: 2024-12-03XI AN JIAOTONG UNIV
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202411236154.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-04
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The problems of low prediction accuracy and low computational efficiency of complex electromechanical systems are especially insufficient prediction performance in the long-term range.

Method used

The cross-CNN network model is adopted to train through normalized data sets, and time-series features are extracted using downsampling and reconstruction techniques, and cross-calculation is performed in horizontal and vertical directions. Combined with local and global feature extraction and information interaction, multivariate multi-variable Step prediction.

Benefits of technology

It improves the multi-step prediction accuracy of complex electromechanical systems, reduces model complexity, improves generalization capabilities, and is suitable for long-term operation status prediction of complex electromechanical systems, providing a general solution.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066426A_ABST
    Figure CN119066426A_ABST
Patent Text Reader

Abstract

The invention discloses a multivariable multi-step prediction method and system for a complex electromechanical system, and belongs to the field of complex electromechanical system monitoring variable analysis and prediction, and the method comprises the steps: obtaining multi-dimensional monitoring DCS historical data in the operation process of the complex electromechanical system, and carrying out the normalization processing of the obtained historical data, and obtaining a normalized data set; training a cross CNN network model based on the normalized data set to obtain a multivariable multi-step prediction model; the current multi-dimensional monitoring DCS data of the complex electromechanical system to be monitored are predicted through the multivariable multi-step prediction model, and a direct multi-step prediction result of the complex electromechanical system in the future time step is obtained. The method is suitable for accurate classification, identification and prediction tasks of the operation state of the complex electromechanical system, and a general solution is provided for the problem of direct multi-step prediction of the time sequence of the complex electromechanical system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of complex electromechanical system monitoring variable analysis and prediction, and specifically relates to a complex electromechanical system multi-variable multi-step prediction method and system. Background Art

[0002] Complex electromechanical systems are essentially composed of many large-scale power machinery, electrical equipment, and multi-media network coupling. Due to the complex structure, function, and mechanism of such systems, unplanned downtime may occur during service, leading to safety accidents and causing serious economic losses. The prediction method based on operation monitoring data can analyze the operating status of complex electromechanical systems under service conditions. It is an effective technical means for preventive maintenance and health status management, which is crucial to the effective operation and maintenance of complex electromechanical systems.

[0003] The problem of multivariate analysis and prediction of complex electromechanical systems is essentially a data prediction problem for complex nonlinear systems. Unlike classic time series prediction models, deep learning methods do not require a lot of data preprocessing. Simple models based on deep learning can directly obtain the next prediction value of the time series, with high computational efficiency and low complexity. However, the monitoring data that can reflect the operating status in complex electromechanical systems has high-frequency characteristics, and single-step prediction methods cannot solve the demand for long-term prediction in actual industrial processes. Effective multi-step prediction methods are more suitable for dealing with monitoring variable analysis and prediction problems in complex electromechanical systems.

[0004] Methods for implementing multi-step prediction can be divided into two categories: (1) direct multi-step prediction; (2) indirect multi-step prediction. Indirect multi-step prediction is to obtain new multi-step prediction results through continuous iteration based on single-step prediction. However, the accumulation of iterative errors will significantly reduce the prediction performance of the model in long-term prediction. Direct multi-step prediction method refers to a prediction method that can predict all future time steps at once through a trained prediction model. The training model establishes a dynamic mapping relationship between the input time series and the future data series, avoiding the error accumulation problem of single-step iteration. In addition, there is a coupling influence relationship between the multidimensional monitoring variables of complex electromechanical systems. Therefore, it is difficult to accurately analyze the future operating status of complex electromechanical systems based on independent predictions of single variables. Multivariate time series prediction considering correlation is more suitable for multivariate analysis and prediction of complex electromechanical systems. However, as the complexity of the model increases, it also requires corresponding complex training strategies, which affects the computational efficiency and generalization ability of the model. Summary of the invention

[0005] The purpose of the present invention is to provide a complex electromechanical system multi-variable multi-step prediction method and system to overcome the problems of low multi-variable prediction accuracy and low computational efficiency in the prior art for complex electromechanical systems.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A multi-variable multi-step prediction method for a complex electromechanical system comprises the following steps:

[0008] Acquire multi-dimensional monitoring DCS historical data during the operation of complex power systems, and normalize the acquired historical data to obtain a normalized data set;

[0009] Based on the normalized data set, the cross CNN network model is trained to obtain a multivariate multi-step prediction model;

[0010] The multivariable multi-step prediction model is used to predict the current multi-dimensional monitoring DCS data of the complex electrical system to be monitored, and the direct multi-step prediction results of the complex electromechanical system in the future time step are obtained.

[0011] Preferably, the training of the multivariate multi-step prediction model is as follows: taking the normalized data set as input, using downsampling and reconstruction techniques, extracting the time series features in a single time series in the normalized data set through horizontal and vertical cross-direction calculations, and obtaining a two-dimensional tensor for each monitored variable.

[0012] Preferably, local features and global features are extracted from the two-dimensional tensor of each monitoring variable, and then information interaction is performed on the local features and global features of different monitoring variables; for each monitoring variable, the one-dimensional sequences under different subsequence lengths are summed to obtain the fusion features of the monitoring variable and then they are added to obtain the prediction result.

[0013] Preferably, a continuous downsampling strategy is applied to the i-th dimension time series, the time series data of length L is converted into multiple non-overlapping subsequences of length P, and the subsequences are reconstructed into a two-dimensional tensor.

[0014] Preferably, a CNN convolution kernel of size (1×1) is used to realize information interaction between different variables and quantify correlation features.

[0015] A multi-variable multi-step prediction system for complex electromechanical systems, including a preprocessing module, a training module and a prediction module:

[0016] The preprocessing module acquires the multi-dimensional monitoring DCS historical data during the operation of the complex power system and normalizes the acquired historical data to obtain a normalized data set;

[0017] The training module trains the cross-CNN network model based on the normalized data set to obtain a multivariate multi-step prediction model;

[0018] The prediction module uses a multivariable multi-step prediction model to predict the current multi-dimensional monitoring DCS data of the complex electrical system to be monitored, and obtains the direct multi-step prediction results of the complex electromechanical system in the future time step.

[0019] Preferably, the training of the multivariate multi-step prediction model is as follows: taking the normalized data set as input, using downsampling and reconstruction techniques, extracting the time series features in a single time series in the normalized data set through horizontal and vertical cross-direction calculations, and obtaining a two-dimensional tensor for each monitored variable.

[0020] Preferably, local features and global features are extracted from the two-dimensional tensor of each monitoring variable, and then information interaction is performed on the local features and global features of different monitoring variables; for each monitoring variable, the one-dimensional sequences under different subsequence lengths are summed to obtain the fusion features of the monitoring variable and then they are added to obtain the prediction result.

[0021] Preferably, a continuous downsampling strategy is applied to the i-th dimension time series, the time series data of length L is converted into multiple non-overlapping subsequences of length P, and the subsequences are reconstructed into a two-dimensional tensor.

[0022] Preferably, a CNN convolution kernel of size (1×1) is used to realize information interaction between different variables and quantify correlation features.

[0023] Compared with the prior art, the present invention has the following beneficial technical effects:

[0024] The present invention discloses a multivariable multi-step prediction method for a complex electromechanical system. The method comprises the following steps: obtaining multidimensional monitoring DCS historical data during the operation of the complex electrical system, and performing normalization processing on the acquired historical data to obtain a normalized data set; training a cross CNN network model based on the normalized data set to obtain a multivariable multi-step prediction model; and using the multivariable multi-step prediction model to predict the current multidimensional monitoring DCS data of the complex electrical system to be monitored, so as to obtain a direct multi-step prediction result of the complex electromechanical system in the future time step. The present invention can flexibly and effectively process the analysis problem of high-dimensional time series monitoring data, and is suitable for the tasks of accurately classifying, identifying and predicting the operation status of the complex electromechanical system, and provides a general solution to the problem of direct multi-step prediction of the time series of the complex electromechanical system.

[0025] The present invention establishes a cross CNN model based on the CNN structure, performs calculations in the horizontal and vertical cross directions, and can simultaneously extract the multi-scale time dependency within a single time series and the information interaction between different time series.

[0026] The present invention adopts fewer parameters and a relatively simple network structure, which can not only ensure high prediction performance, but also improve the generalization ability of the model, and can give full play to the advantages of deep learning methods, reduce model complexity, and improve the accuracy of prediction results. The present invention is easy to implement, can achieve direct multi-step prediction, and is suitable for the actual industrial needs of complex electromechanical systems that need to predict the future long-term operating status, and provides a reference basis for multivariate data prediction and state prediction in other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of the process flow of the multi-variable multi-step prediction method for complex electromechanical systems in an embodiment of the present invention.

[0028] Figure 2 2 is a working process diagram of multi-scale downsampling and reconstruction in an embodiment of the present invention.

[0029] Figure 3 It is a working process diagram of the information interaction module in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0032] like Figure 1 As shown, the present invention provides a multi-variable multi-step prediction method for a complex electromechanical system, which is used to realize direct multi-step prediction of multiple variables of a complex electromechanical system, and is applicable to actual industrial needs of complex electromechanical systems that need to predict future long-term operating states, and provides a reference basis for multi-variable data prediction and state prediction in other fields, and specifically includes the following steps:

[0033] Acquire multi-dimensional monitoring DCS historical data during the operation of complex power systems, and normalize the acquired historical data to obtain a normalized data set;

[0034] Based on the normalized data set, the cross CNN network model is trained to obtain a multivariate multi-step prediction model;

[0035] The multivariable multi-step prediction model is used to predict the current multi-dimensional monitoring DCS data of the complex electrical system to be monitored, and the direct multi-step prediction results of the complex electromechanical system in the future time step are obtained.

[0036] Specifically, in the embodiment of the present application, the training of the multivariate multi-step prediction model takes the normalized data set as input, adopts downsampling and reconstruction technology, extracts the time series features in a single time series in the normalized data set through horizontal and vertical cross-direction calculations, represents the multi-scale time dependency in a single time series, and obtains a two-dimensional tensor of each monitoring variable;

[0037] The local and global features of the two-dimensional tensor of each monitoring variable are extracted to capture and mine the dynamic time features in the time series. Then, the local and global features of different monitoring variables are interacted to quantify the correlation between the time series of different monitoring variables in the complex electromechanical system. For each monitoring variable, the one-dimensional sequences of different subsequence lengths need to be summed to obtain the fusion features of the monitoring variable, and then they are added to obtain the prediction results.

[0038] For the multi-dimensional monitoring DCS historical data of complex electromechanical systems collected, an original data set is established, in which the monitoring variables are parameters that can reflect the operating status of the complex electromechanical system, with single-variable time series characteristics and multi-variable correlation characteristics. The normalization method is used to normalize the monitoring variables and model training.

[0039] The cross-CNN network model first applies the continuous downsampling strategy to the i-th dimension time series, converts the time series data of length L into multiple non-overlapping subsequences of length P, and reconstructs the subsequences into a two-dimensional tensor. The correlation between different variables is quantified, and the correlation characteristics between multiple monitoring variables of complex electromechanical systems are extracted. Given the hyperparameters of the cross-CNN network model, the experimental environment, and the indicators used to evaluate the performance of the cross-CNN network model, the prediction accuracy and analysis efficiency of the model are measured and analyzed in different dimensions.

[0040] In a specific implementation manner of the present application, the operation process of a complex electromechanical system represented by a steam turbine is monitored, specifically including the following steps:

[0041] S1, multi-dimensional monitoring DCS historical data collection during the operation of complex electrical systems: collect DCS monitoring data of various points during the operation of complex electromechanical systems represented by steam turbines, and establish original data sets, in which the monitoring variables are used as parameters that can reflect the operating status of steam turbines, with single variable time series characteristics and multivariate correlation characteristics. The original data sets obtained are normalized to obtain normalized data sets for cross-CNN network model training;

[0042] Step 1.1) Multi-dimensional monitoring DCS historical data collection: collect N monitoring variables that can reflect the operating status of the turbine. All historical data of each monitoring variable contains n time points. All historical data are expressed as:

[0043]

[0044] Where: X 1:n ——All historical data; ——All historical data of the ith monitoring variable; ——The value of the i-th monitored variable at the j-th time point.

[0045] Assume the window length is L, the sliding step is 1, and the time window slides along the time sequence on the historical data to obtain samples at different times. The sample at time t can be recorded as X t-L+1:t (t>L). The input data dimension of the multi-step prediction of the present invention is N, the output dimension is N, and the prediction step length is H. Its digital expression is:

[0046]

[0047] Step 1.2) Normalize the training data: Normalize the N-dimensional monitoring time series of the steam turbine unit obtained in step 1.1) Reversible instance normalization (RevIN) is used to process the data form required by the cross-CNN network structure. First, each monitoring variable is calculated The mean and variance of :

[0048]

[0049]

[0050] Then, RevIN reversible instance normalization is performed on each monitoring variable, and the calculation formula is as follows:

[0051]

[0052] in: ——The result after normalizing the value of the ith monitoring variable at time j; γ i ,β i——The learnable affine transformation parameter of the ith monitoring variable; ε——An extremely small constant value that can avoid the denominator being zero.

[0053] S2, based on S1, a normalized data set of a complex electromechanical system represented by a steam turbine is obtained. Downsampling and reconstruction techniques are used to obtain multi-scale time dependencies in a single time series through horizontal and vertical cross-direction calculations:

[0054] Step 2.1) Select subsequence length: For the normalized data set in step 1.2), downsample each dimensional time series with the set subsequence length to construct multiple non-repeating subsequences. Select the one-dimensional time series X i ,Right now Firstly, the Fast Fourier Transform (FFT) method is used to determine the appropriate subsequence length, and FFT is used for each dimensional time series to calculate the amplitude of each frequency component.

[0055] The specific process is as follows: Let the length of the time series be L and the length of the subsequence be P. If L is divisible by P, we can obtain non-repeating subsequences; if L cannot be divided by P, the original time series can be padded with zeros first, and then continuously downsampled to obtain non-repeating subsequences, where Indicates rounding up. On the contrary, if L is divisible by P, by continuously downsampling the original time series, we get non-repeating subsequences; if L cannot be divided by P, first fill the original time series with zeros, and then continuously downsample to obtain non-repeating subsequences, where Indicates rounding up. Therefore, the calculation process to determine the length of the subsequence is as follows:

[0056]

[0057] Among them: A i ——the amplitude of all frequencies of the i-th monitoring variable sequence; X i ——The i-th monitoring variable; FFT(·) represents the FFT operation; Amp(·) represents the amplitude corresponding to each frequency component; ——The amplitude of the j-th frequency component of the ith monitored variable; N——The number of monitored variables.

[0058] The amplitude corresponding to each frequency component indicates the strength of the frequency component. In order to simplify the calculation and reduce the interference of noise components, only the first k items with the largest amplitude are considered. K can be set to different values ​​as needed. The largest k frequencies are calculated as follows:

[0059]

[0060] {f1,…,f k} = argTopk(A)

[0061] Where: A f=j ——the average amplitude of the frequency components of all monitored variables j; A——the average amplitude of all frequencies of all monitored variables; {f1,...,f k}——the k frequencies required; argTopk(·) represents the frequencies corresponding to the largest k amplitudes.

[0062] Because the period and frequency are reciprocals of each other, we can find the period corresponding to k frequencies. Taking each period as the length of the subsequence, each monitoring variable sequence can be converted into k two-dimensional tensors:

[0063]

[0064] Where: f l ——lth frequency; p l ——The period corresponding to the lth frequency; ——The i-th detection variable in the subsequence length is p l The two-dimensional tensor obtained by continuous downsampling.

[0065] Step 2.2) Multi-scale continuous downsampling: Based on the FFT-based selection of subsequence length in step 2.1), the amplitudes of all monitored variables at each frequency are calculated. Here, the periods corresponding to the two amplitudes with the largest mean are taken as the two subsequence lengths.

[0066] The periods corresponding to the two amplitudes with the largest mean are used as the subsequence lengths, and each input monitoring time series is continuously downsampled. At this point, each monitoring variable is converted into a two-dimensional tensor.

[0067] After continuous downsampling, the original one-dimensional time series is converted into a two-dimensional matrix. Each column in the matrix is ​​a continuous time point, so convolution in the column direction can extract local time series features; each row in the matrix is ​​an equidistant time point in the original time series, so convolution in the row direction can extract global time series features.

[0068] S3, time series feature extraction of monitoring time series: Based on the two-dimensional tensor converted from S2, the time series feature extraction module is used to extract local and global features, capture and mine dynamic time features in the time series:

[0069] Step 3.1) Extract local temporal features: For each two-dimensional matrix converted in step 2.2), each column represents a continuous time point in a cycle, and convolution kernels of different sizes are used to perform convolution operations with the matrix to extract local temporal features. A CNN structure consisting of multiple CNN filters with different receptive fields, namely the inception block, is used to extract local and global temporal features from the two-dimensional tensor. For the two-dimensional tensor converted by the i-th monitoring variable, The calculation process of extracting local features is expressed as:

[0070]

[0071] in: ——The i-th monitoring variable in the subsequence length is p l 2D tensor obtained by continuous downsampling; Inception_local(·) — performs column-wise convolution on each sequence to extract local features; LeakyReLU(·) — a nonlinear activation function; ——local characteristics of the ith monitored variable;

[0072] Step 3.2) Extract global feature parameters: The inception block used for local feature extraction described in step 3.1) is applied to global feature extraction. For the two-dimensional tensor transformed by the i-th monitoring variable, The calculation process of extracting global features is expressed as:

[0073]

[0074] Inception(X)=Average(W i *X+b i ),i∈{1,…,q}

[0075] in: ——local features of the ith monitoring variable; Inception_global(·)——perform convolution operation on each row of the subsequence to extract long-range temporal features; *——convolution operation; W i ——The i-th convolution kernel;

[0076] After calculation, the output of the initial module (inception block) is the average of the convolution results of different kernel sizes; for different two-dimensional tensors, the parameters of the initial modules (inception blocks) of the same type are shared, thereby improving the calculation efficiency. So far, the feature extraction of a single time series has been achieved.

[0077] S4, quantification of correlation between multiple monitoring time series: Based on the single monitoring variable time series feature extraction model of S3, an information interaction module based on CNN convolution calculation is established to quantify the correlation between multiple monitoring variable time series of complex electromechanical systems:

[0078] Step 4.1) Correlation feature extraction: According to the two-dimensional tensor obtained by continuous downsampling of the monitoring variable time series as described in step 2.1), in the information interaction module, a CNN convolution kernel of size (1×1) is used to realize information interaction between different variables and quantify the correlation features.

[0079] Step 4.1.1) Stacking three-dimensional tensors: naturally stack the two-dimensional tensors corresponding to the monitoring variables described in step 4.1) to obtain a three-dimensional tensor N is the number of monitoring variables, that is, the number of input channels. For each value in the two-dimensional tensor of the monitoring variable, it is obtained by adaptively weighting all monitoring time series;

[0080] Step 4.1.2) Information interaction module calculation: For the three-dimensional tensor described in step 4.1), the input three-dimensional tensor is convolved with a (1×1) convolution kernel to obtain a three-dimensional tensor of the same size as the input. The calculation process of information interaction is expressed as:

[0081] h info,l =LeakyRelu(W*X 2D,l )

[0082] Where: *——convolution operation; W——(1×1) convolution kernel;

[0083] Step 4.2) Convolution and feeding: Step 4.1) gives the output of the information interaction module. In the last step of the cross-CNN calculation, the output h of each information interaction module is info,l , (l=1,…,k) for reconstruction. Since a variety of subsequence lengths were used for downsampling before, for each monitoring variable, it is necessary to sum the one-dimensional sequences under different subsequence lengths to obtain the fusion features of the monitoring variable and then add them and feed them into the linear layer to obtain the prediction result.

[0084] S5, cross-CNN network model training: According to the cross-CNN network model calculation process in S2 to S4, the normalized DCS multivariable monitoring data is used as the input of the cross-CNN network structure model, and the complex electromechanical system represented by the steam turbine is directly predicted in the future time step to obtain the prediction results of multiple monitoring variables.

[0085] Step 5.1) Model initialization: According to the cross-CNN network model described in steps 2) to 4), the training process of the multi-variable direct multi-step prediction model of a complex electromechanical system represented by a steam turbine mainly includes the following steps:

[0086] Step 5.1.1) Configure the training environment: The model is built based on the PyTorch framework and executed with the support of NVIDIA 3080 GPU and Intel Xeon(R) Platinum 8255C CPU@2.50GHz;

[0087] Step 5.1.2) Set model parameters: set the lookback length L to 30, and the prediction step length H to 5, 10, and 15 respectively. The hyperparameters of the cross-CNN network model include the number of subsequence lengths, the number of feature extraction modules, and the initial learning rate. Grid search is used to adjust the key hyperparameters of the model to achieve the best performance;

[0088] Step 5.2) Model prediction and performance evaluation: Based on the initialization conditions described in step 5.1), the model is trained and predicted using actual data, and the model performance is evaluated in multiple dimensions.

[0089] Step 5.2.1) Divide the data set: Divide the data into three segments in a ratio of 7:1:2 in chronological order, corresponding to the training set, validation set, and test set respectively;

[0090] Step 5.2.2) Multivariate direct multi-step prediction: 5, 10 and 15-step prediction results are obtained on the test set. Real-time monitoring data is obtained from the DCS monitoring system of complex electromechanical systems represented by steam turbines and normalized, and the multivariate time series is used as the input of the cross-CNN network model; the parameters of each model are optimized through training, and the model parameters with the best performance on the validation set are saved; each model is validated on the test set and the results are recorded;

[0091] Step 5.2.3) Metric evaluation: By comparing with other models, the model performance is evaluated from multiple dimensions such as prediction accuracy, number of training parameters, and training and computing efficiency. At the same time, in order to avoid randomness of the results and eliminate random errors, the MSE and MAE values ​​of the model prediction accuracy represent the final results obtained by taking the average of 20 experiments under the same hyperparameters.

[0092] Example

[0093] Based on the cross-CNN network structure of CNN network, direct multi-step prediction of multi-variable monitoring time series of complex electromechanical system is carried out. A method of direct multi-step prediction of multi-variable complex electromechanical system based on cross-CNN is established, which includes the following steps:

[0094] Collection of raw data for complex electromechanical system status monitoring. Collect DCS monitoring data from various points during the operation of complex electromechanical systems represented by steam turbines, and establish raw data sets, in which monitoring variables are used as parameters that can reflect the operating status of steam turbines, and have single variable time series characteristics and multivariate correlation characteristics;

[0095] The obtained original data set is normalized using the reversible instance normalization (RevIN) method to obtain a normalized data set for model training.

[0096] See also Figure 2 As shown in the figure, when performing multi-scale downsampling and reconstruction of the time series, the downsampling length is first determined by using FFT, and the amplitude of all monitoring variables at each frequency is calculated. Then, the amplitude mean at different frequencies is calculated, and the periods corresponding to the two amplitudes with the largest mean are taken as the lengths of the two subsequences. When the time series length and the sampling length do not have an integer divisibility relationship, the time series is padded with zeros. The periods corresponding to the two amplitudes with the largest mean are taken as the subsequence lengths, and each input monitoring time series is continuously downsampled, so that each monitoring variable is converted into two two-dimensional tensors.

[0097] See also Figure 3 As shown, the information interaction module is used to process three-dimensional tensors. The two-dimensional tensors corresponding to each monitoring variable are naturally stacked into a three-dimensional tensor, calculated with a (1×1) convolution kernel, and finally the outputs are added and fed into the linear layer.

[0098] Real-time monitoring data is obtained from the DCS monitoring system of complex electromechanical systems represented by steam turbines and normalized. The normalized DCS multivariate monitoring data is divided into three segments of 7:1:2 in chronological order, corresponding to the training set, validation set and test set respectively, and the multivariate time series is used as the input of the cross CNN network structure model.

[0099] Through training, the parameters of each model are optimized and the model parameters with the best performance on the validation set are saved; each model is validated on the test set and the results are recorded;

[0100] Direct multi-step prediction of complex electromechanical systems represented by steam turbines in future time steps to obtain prediction results of multiple monitoring variables;

[0101] By comparing with other models, the model performance is evaluated from multiple dimensions such as prediction accuracy, number of training parameters, and training and computing efficiency. At the same time, in order to avoid randomness of results and eliminate random errors, the MSE and MAE values ​​of the model prediction accuracy represent the final results obtained by averaging 20 experiments under the same hyperparameters.

[0102] In one embodiment of the present invention, a multi-variable multi-step prediction system for a complex electromechanical system is provided, comprising a preprocessing module, a training module and a prediction module:

[0103] The preprocessing module acquires the multi-dimensional monitoring DCS historical data during the operation of the complex power system and normalizes the acquired historical data to obtain a normalized data set;

[0104] The training module trains the cross-CNN network model based on the normalized data set to obtain a multivariate multi-step prediction model;

[0105] The prediction module uses a multivariable multi-step prediction model to predict the current multi-dimensional monitoring DCS data of the complex electrical system to be monitored, and obtains the direct multi-step prediction results of the complex electromechanical system in the future time step.

[0106] Preferably, the training of the multivariate multi-step prediction model is as follows: taking the normalized data set as input, using downsampling and reconstruction techniques, extracting the time series features in a single time series in the normalized data set through horizontal and vertical cross-direction calculations, and obtaining a two-dimensional tensor for each monitored variable.

[0107] Preferably, local features and global features are extracted from the two-dimensional tensor of each monitoring variable, and then information interaction is performed on the local features and global features of different monitoring variables; for each monitoring variable, the one-dimensional sequences under different subsequence lengths are summed to obtain the fusion features of the monitoring variable and then they are added to obtain the prediction result.

[0108] Preferably, a continuous downsampling strategy is applied to the i-th dimension time series, the time series data of length L is converted into multiple non-overlapping subsequences of length P, and the subsequences are reconstructed into a two-dimensional tensor.

[0109] Preferably, a CNN convolution kernel of size (1×1) is used to realize information interaction between different variables and quantify correlation features.

[0110] The hyperparameters of the cross-CNN network model include the number of subsequence lengths, the number of feature extraction modules, and the initial learning rate. Grid search is used to adjust the key hyperparameters of the model to achieve optimal performance. The model is trained and predicted using actual data, and the model performance is evaluated in multiple dimensions.

[0111] The cross-CNN structure achieves high prediction performance with fewer parameters and a relatively simple network structure. The cross-CNN network model provides a solution to the problem of direct multi-step prediction of time series of complex electromechanical systems, and provides a general basic framework for data prediction and state analysis problems in complex electromechanical systems and other fields.

Claims

1. A multivariable multi-step prediction method for complex electromechanical systems, characterized in that: The following steps are involved: Acquire multi-dimensional monitoring DCS historical data during the operation of complex power systems, and normalize the acquired historical data to obtain a normalized data set; Based on the normalized data set, the cross CNN network model is trained to obtain a multivariate multi-step prediction model; The multivariable multi-step prediction model is used to predict the current multi-dimensional monitoring DCS data of the complex electrical system to be monitored, and the direct multi-step prediction results of the complex electromechanical system in the future time step are obtained.

2. A multivariable multi-step prediction method for complex electromechanical systems according to claim 1, characterized in that: Training of multivariate multi-step prediction model: Taking the normalized data set as input, downsampling and reconstruction techniques are used to extract the time series features in a single time series in the normalized data set through horizontal and vertical cross-direction calculations to obtain a two-dimensional tensor for each monitored variable.

3. A multi-variable multi-step prediction method for complex electromechanical systems according to claim 2, characterized in that: The local and global features of the two-dimensional tensor of each monitoring variable are extracted, and then the local and global features of different monitoring variables are interacted with each other; for each monitoring variable, the one-dimensional sequences of different subsequence lengths are summed to obtain the fusion features of the monitoring variable and then they are added to obtain the prediction result.

4. A multi-variable multi-step prediction method for complex electromechanical systems according to claim 2, characterized in that: Specifically, the continuous downsampling strategy is applied to the i-th dimension time series, the time series data of length L is converted into multiple non-overlapping subsequences of length P, and the subsequences are reconstructed into a two-dimensional tensor.

5. The multivariable multi-step prediction method for complex electromechanical systems according to claim 3, characterized in that: A CNN convolution kernel of size (1×1) is used to realize the information interaction between different variables and quantify the correlation characteristics.

6. A multivariable multi-step prediction system for complex electromechanical systems, characterized in that: Including preprocessing module, training module and prediction module: The preprocessing module acquires the multi-dimensional monitoring DCS historical data during the operation of the complex power system and normalizes the acquired historical data to obtain a normalized data set; The training module trains the cross-CNN network model based on the normalized data set to obtain a multivariate multi-step prediction model; The prediction module uses a multivariable multi-step prediction model to predict the current multi-dimensional monitoring DCS data of the complex electrical system to be monitored, and obtains the direct multi-step prediction results of the complex electromechanical system in the future time step.

7. A complex electromechanical system multivariable multi-step prediction system according to claim 6, characterized in that: Training of multivariate multi-step prediction model: Taking the normalized data set as input, downsampling and reconstruction techniques are used to extract the time series features in a single time series in the normalized data set through horizontal and vertical cross-direction calculations to obtain a two-dimensional tensor for each monitored variable.

8. A complex electromechanical system multivariable multi-step prediction system according to claim 7, characterized in that: The local and global features of the two-dimensional tensor of each monitoring variable are extracted, and then the local and global features of different monitoring variables are interacted with each other; for each monitoring variable, the one-dimensional sequences of different subsequence lengths are summed to obtain the fusion features of the monitoring variable and then they are added to obtain the prediction result.

9. A complex electromechanical system multivariable multi-step prediction system according to claim 7, characterized in that: Specifically, the continuous downsampling strategy is applied to the i-th dimension time series, the time series data of length L is converted into multiple non-overlapping subsequences of length P, and the subsequences are reconstructed into a two-dimensional tensor.

10. A complex electromechanical system multivariable multi-step prediction system according to claim 8, characterized in that: A CNN convolution kernel of size (1×1) is used to realize the information interaction between different variables and quantify the correlation characteristics.

Citation Information

Patent Citations

  • Wind power forecasting method based on genetic algorithm optimization BP neural network

    CN101706335A

  • Multi-dimension and multi-step life forecasting method based on support vector machine

    CN105528621A

  • Correlation-based time sequence data multi-step prediction method and system

    CN107274009A

  • Parallel time series data prediction method

    CN107392307A

  • Low-complexity and high-precision time sequence multi-step prediction method and system

    CN115310650A