Water machine electricity multi-source signal coupling prediction method and system

Abstract

This invention discloses a method and system for predicting multi-source signal coupling in hydropower systems, comprising: performing a signal component decomposition process and obtaining different intrinsic mode components; constructing a multi-source cross-coupling prediction task and predicting related coupling characteristics; constructing a multi-source cross-mixed basis model and obtaining related prediction results; constructing a residual correction model and obtaining coupling prediction results; performing a mutual information analysis process and a transfer entropy calculation process, and quantifying and analyzing the correlation strength and dynamic transfer relationship. This invention solves the problems of low prediction accuracy and poor signal coupling relationship mining in traditional single-model systems, achieving high-precision coupling prediction and intrinsic correlation analysis of hydropower signals, providing reliable technical support for the condition monitoring of hydropower station equipment.

Description

Technical Field

[0001] This invention relates to the field of hydropower system monitoring and prediction technology, and in particular to a method and system for predicting multi-source signals coupled from hydroelectric power systems. Background Technology

[0002] The hydroelectric system of a hydropower station is a complex coupled system composed of a hydraulic system, a mechanical transmission system, and a power output system. The hydraulic signals, mechanical signals, and electrical signals within the system do not exist independently, but are interconnected through energy transfer and force conversion. Accurately predicting the dynamic change trends of these three types of signals and quantifying their interaction and coupling relationship is of great significance for improving the safety, stability, and economy of hydropower station operation.

[0003] In existing prediction methods, single models focus on prediction in a single direction, making it difficult to simultaneously capture the long-range dependence, local features, and sequential memory characteristics of signals. Furthermore, they lack a comprehensive understanding of the cross-coupling relationships between hydraulic, mechanical, and electrical signals, resulting in limited overall prediction accuracy and making it difficult to meet the high-precision prediction requirements in complex coupling scenarios. On the other hand, fusion models adopt a static, fixed-weight weighted averaging strategy, which neither fully considers the differences in prediction accuracy of different base models under different operating conditions and signal types, nor has it achieved dynamic adaptation between the base models and the prediction task. It also fails to conduct quantitative analysis on the correlation strength and dynamic transmission path between signals, thus failing to provide effective data support for a deeper understanding of the coupling mechanism of hydro-mechanical-electric systems. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method and system for multi-source signal coupling prediction in water electromechanical systems.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for multi-source signal coupling prediction in hydroelectric systems, comprising: Based on the preprocessed hydropower station electromechanical multi-source signal data, a signal component decomposition process is performed to obtain the intrinsic mode components corresponding to different hydropower station electromechanical multi-source signal data. A multi-source cross-coupling prediction task is constructed based on the decomposed intrinsic mode components, and the correlation coupling characteristics corresponding to the multi-source signal data of different hydropower stations are predicted. A multi-source cross-hybrid basis model is constructed based on the preset multi-source feature capture model, and the relevant prediction results corresponding to the multi-source cross-coupling prediction task are obtained. A residual correction model is constructed based on the preset hierarchical model fusion method, and the coupling prediction results corresponding to the multi-source cross-mixing basis model are obtained. Based on the preprocessed multi-source signal data of hydropower station electromechanical systems, a mutual information analysis process and a transfer entropy calculation process are performed, and the correlation strength and dynamic transfer relationship of the multi-source signal data of hydropower station electromechanical systems are quantitatively analyzed.

[0006] In some embodiments, the hydropower station's multi-source signal data includes hydraulic operating condition signals, mechanical vibration signals, and electrical operating condition signals.

[0007] In some embodiments, the preprocessing process of the hydropower station's multi-source signal data includes missing value removal, outlier handling, time-series signal alignment, standardization and normalization, and data quality verification.

[0008] In some embodiments, the missing value removal includes: Identifying global outliers based on the Laida criterion; Identification of local outliers based on the sliding window quartile method; Global and local outliers are corrected through a moving average filtering process.

[0009] In some embodiments, the signal component decomposition process preferentially uses the variational mode decomposition process, and when the variational mode decomposition process fails, the fitting filter decomposition process is used.

[0010] In some embodiments, the multi-source cross-coupling prediction task includes a hydraulic condition signal coupling prediction task, a power condition signal coupling prediction task, and a mechanical vibration signal coupling prediction task.

[0011] In some embodiments, the hydraulic condition signal coupling prediction task uses the intrinsic mode components of the hydraulic condition signal as input to predict the eigenvalues ​​of the mechanical vibration signal and the eigenvalues ​​of the power condition signal, respectively. The power condition signal coupling prediction task uses the intrinsic mode components of the power condition signal as input to predict the eigenvalues ​​of the hydraulic condition signal and the eigenvalues ​​of the mechanical vibration signal, respectively. The mechanical vibration signal coupling prediction task uses the intrinsic mode components of the mechanical vibration signal as input to predict the eigenvalues ​​of the hydraulic operating condition signal and the eigenvalues ​​of the power operating condition signal, respectively.

[0012] In some embodiments, the multi-source feature capture model includes a self-attention mechanism model, a long short-term memory network model, and a temporal convolutional network model. The decomposed multiple sets of intrinsic modal components are used to construct a time series sequence according to time steps and to divide the training set and test set as input data for the multi-source cross-hybrid base model. The training task of the multi-source cross-hybrid base model is to construct multiple sets of multi-source cross-coupled prediction tasks.

[0013] In some embodiments, the construction of a residual correction model based on a preset hierarchical model fusion method and the acquisition of the coupling prediction results corresponding to the multi-source cross-mixing basis model include: Dynamic weights are calculated based on the prediction loss of the multi-source cross-mixed basis model on the validation set, and adaptive weighting of the multi-source cross-mixed basis model is performed. The prediction results and weighted average results of the multi-source cross-mixing basis model are extracted as a fusion feature set, and a residual correction model is constructed. The residual correction model is built based on the fusion feature input to obtain the residual output result, and the corrected coupled prediction result is obtained by summing the weighted average prediction result and the residual output result.

[0014] In a second aspect, the present invention also provides a hydroelectric multi-source signal coupling prediction system for operating the hydroelectric multi-source signal coupling prediction method as described in the first aspect, the prediction system comprising: The data processing module is used to preprocess, analyze mutual information, and calculate the transfer entropy of multi-source signals from hydroelectric turbines in hydropower stations. The task building module is used to construct multi-source cross-coupled prediction tasks; The model building module is used to build multi-source cross-mixed basis models and residual correction models; The visualization module is used to generate signal decomposition curves, base model training loss curves, reconstruction prediction comparison charts, fusion model residual distribution charts, and mutual information heatmaps.

[0015] The present invention has the following beneficial effects: This invention employs variational mode decomposition to perform component decomposition on preprocessed multi-source hydropower and electromechanical signal data. It then constructs a multi-source cross-coupling prediction task and a multi-source cross-mixed basis model to achieve cross-coupling prediction between hydraulic, mechanical, and electrical signals. A hierarchical model fusion method is used to construct a residual correction model to correct and optimize the output results, obtaining the corresponding coupling prediction results. Simultaneously, through mutual information analysis and transfer entropy calculation, the correlation strength and dynamic transfer relationships of the multi-source hydropower and electromechanical signal data are quantitatively analyzed. This invention solves the problems of low prediction accuracy and poor signal coupling relationship mining in traditional single-model approaches, achieving high-precision coupling prediction and intrinsic correlation analysis of hydropower and electromechanical signals, providing reliable technical support for hydropower station equipment condition monitoring. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the multi-source signal coupling prediction method for hydroelectric systems proposed in this invention. Figure 1 ; Figure 2 This is a flowchart illustrating the multi-source signal coupling prediction method for hydroelectric systems proposed in this invention. Figure 2 ; Figure 3 This is a data curve diagram of variational mode decomposition provided in an embodiment of the present invention; Figure 4 The training loss curve of the self-attention mechanism model provided in the embodiments of the present invention; Figure 5 The training loss curve of the Long Short-Term Memory network model provided in the embodiments of the present invention; Figure 6 The training loss curve of the temporal convolutional network model provided in the embodiments of the present invention; Figure 7 A comparison chart of final reconstruction predictions provided for embodiments of the present invention; Figure 8 The residual distribution diagram of the fusion model provided in the embodiments of the present invention; Figure 9 Mutual information diagram provided for embodiments of the present invention; Figure 10 Information transmission diagram provided for embodiments of the present invention; Figure 11 This is a schematic diagram of the principle of the multi-source signal coupling prediction system for water electromechanical systems proposed in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This application provides a method and system for predicting the coupling of multi-source signals from hydropower and electromechanical systems, solving the problems of low prediction accuracy and poor signal coupling relationship mining in traditional single-model methods. This application uses variational mode decomposition to decompose the preprocessed multi-source signal data from hydropower and electromechanical systems, then constructs a multi-source cross-coupling prediction task and a multi-source cross-mixed basis model to achieve cross-coupling prediction between hydraulic, mechanical, and electrical signals respectively. A hierarchical model fusion method is used to construct a residual correction model to correct and optimize the output results, obtaining the corresponding coupling prediction results. Simultaneously, through mutual information analysis and transfer entropy calculation, the correlation strength and dynamic transfer relationship of the multi-source signal data from hydropower and electromechanical systems are quantitatively analyzed, achieving high-precision coupling prediction and intrinsic correlation analysis of hydropower and electromechanical signals, providing reliable technical support for hydropower station equipment condition monitoring.

[0019] Please refer to the following examples for details: Reference Figures 1-10 An embodiment of the multi-source signal coupling prediction method for hydroelectric systems provided by the present invention includes: The signal component decomposition process is performed based on the preprocessed multi-source signal data of hydropower station turbines, and the intrinsic mode components corresponding to the multi-source signal data of different hydropower station turbines are obtained. Based on the decomposed intrinsic mode components, a multi-source cross-coupling prediction task is constructed, and the correlation coupling characteristics of multi-source signal data of hydroelectric power stations are predicted. A multi-source cross-mixing basis model is constructed based on the pre-defined multi-source feature capture model, and the relevant prediction results corresponding to the multi-source cross-coupling prediction task are obtained. A residual correction model is constructed based on the pre-defined hierarchical model fusion method, and the coupling prediction results corresponding to the multi-source cross-mixed basis model are obtained. Based on the preprocessed multi-source signal data of hydropower station electromechanical systems, a mutual information analysis process and a transfer entropy calculation process are performed, and the correlation strength and dynamic transfer relationship of the multi-source signal data of hydropower station electromechanical systems are quantitatively analyzed.

[0020] For example, firstly, multi-source operating signals such as hydraulic operating signals, mechanical vibration signals, and electrical operating signals of the hydropower station are acquired, and the entire process data is preprocessed. Then, Variational Mode Decomposition (VMD) is used to adaptively decompose the preprocessed hydro-turbine-electric signals, thereby extracting the Intrinsic Mode Function (IMF) components of each signal to ensure the integrity of the signal features. Next, based on the coupling characteristics of hydro-turbine-electricity systems, three types of multi-source cross-coupling prediction tasks are constructed, and multi-source cross-hybrid basis models are built. Model training and independent prediction for the three types of coupling prediction tasks are completed respectively, and the prediction results of each basis model are output. Then, extreme gradient boosting (XGBoost) and stacking fusion strategies are used to dynamically weight and correct the residuals of the hybrid basis model prediction results, outputting the final optimal result of the hydro-turbine-electricity signal coupling prediction. Finally, mutual information analysis is used to quantify the correlation strength between hydro-turbine-electric signals, and the dynamic transmission path and causal relationship of the signals are analyzed by combining transfer entropy calculation, completing the analysis of the inherent coupling mechanism of the signals.

[0021] In some embodiments, three core multi-source monitoring signals of hydraulic, mechanical vibration and power are collected during the stable operation of actual hydropower station units, and all signals are synchronously collected data under the same time series. The acquired hydropower station's multi-source signal data includes hydraulic operating condition signals, mechanical vibration signals, and electrical operating condition signals.

[0022] For example, the hydraulic operating condition signals of a hydraulic system include guide vane opening ( Peak-to-peak value of pressure pulsation in the top cover ( ), Peak-to-peak value of pressure pulsation at tailrace manhole ( ), peak-to-peak value of tailrace pipe pressure pulsation ( ) and peak-to-peak value of pressure pulsation at the tail of the volute ( ); The mechanical vibration signal of the mechanical system includes the stator frame + X-axis vibration peak-to-peak value ( Stator frame-Y direction vibration peak value ( Stator core + X-axis vibration peak value ( Stator core-Y direction vibration peak value ( ), Peak-to-peak value of horizontal + X-axis vibration of the upper frame ( ) and peak-to-peak value of horizontal-Y direction vibration of the upper frame ( ); Electrical system power condition signals include active power ( ), reactive power ( ), excitation voltage ( ) and excitation current ( ); In some embodiments, the preprocessing process for the collected hydropower station multi-source signal data includes missing value removal, outlier handling, time-series signal alignment, standardization and normalization, and data quality verification.

[0023] For example, when completing the acquisition of raw signals and high-standard data preprocessing, noise and interference in the raw signals can be eliminated through multi-dimensional data optimization methods, ensuring the accuracy and effectiveness of subsequent signal decomposition, model training and correlation analysis. The preprocessing stage is a progressive full-process processing with no blind spots in data processing.

[0024] In some embodiments, during the missing value identification and removal process, since the original collected water turbine electrical signals contain a small number of discrete missing values ​​and continuous missing segments due to factors such as sensor failure and transmission interruption, a classification processing rule is used to complete the missing value optimization. For discrete missing values ​​at a single time step, the proportion is ≤0.5%. Therefore, this application uses linear interpolation for completion, and the interpolation formula is: ; For missing segments that span more than 5 consecutive time steps, the proportion is ≤0.1%, so the entire segment of data is directly removed to avoid the accumulation of errors introduced by interpolation completion; When the missing data of a single type of signal accounts for more than 1% of the total length, the signal data under that operating condition will be re-acquired to ensure data integrity.

[0025] In some embodiments, during outlier identification and correction, the original water turbine electrical signal contains outliers such as spikes, outliers, and drift values ​​caused by sudden sensor interference and instantaneous fluctuations in unit operating conditions. These outliers have no actual physical meaning and therefore severely affect the signal feature extraction effect. This application employs a dual outlier detection method to achieve accurate identification and correction, balancing the accuracy and comprehensiveness of the detection. First, global outliers are identified based on the Laida criterion (3σ criterion), that is, for time series of single-type signals. Calculate the mean and standard deviation When a certain data point satisfies When this occurs, it is determined to be a global outlier; Subsequently, local outliers were identified using the sliding window quartile method (IQR), which involves setting the sliding window length to 100 time steps and calculating the lower quartile of the data within each window. Upper quartiles Interquartile range When the data points within the window satisfy or When this occurs, it is determined to be a local outlier; Finally, for both types of outliers mentioned above, a moving average filtering process is used for correction. This involves replacing the outlier with the average of the five normal data points before and after the given data point. The correction formula is as follows: ,in, This is normal data.

[0026] For example, the three types of signals acquired—hydraulic operating condition signals, mechanical vibration signals, and electrical operating condition signals—need to have their timestamps strictly aligned.

[0027] In some embodiments, all acquired signal sequences are precisely matched according to their timestamps to ensure that at any given time... All three types of signals—hydraulic operating condition signal, mechanical vibration signal, and electrical operating condition signal—can be extracted simultaneously, laying a foundation for time-series consistency in subsequent cross-coupling prediction.

[0028] For example, during data standardization and normalization, since hydraulic operating condition signals, mechanical vibration signals and power operating condition signals have different dimensions and large differences in numerical magnitude, directly inputting them into the model will lead to an imbalance of feature weights, slow model convergence and reduced prediction accuracy. Therefore, Z-Score standardization is performed on all preprocessed signal data.

[0029] In some embodiments, all data are mapped to a standard normal distribution interval with a mean of 0 and a standard deviation of 1, and the standardization formula is: ,in, The original data values ​​for a single type of signal. The mean of all data for this type of signal. The standard deviation of all data for this type of signal. These are the standardized dimensionless data values.

[0030] In some embodiments, during the data quality verification after preprocessing, it is necessary to perform quality verification on the final hydraulic condition signal, mechanical vibration signal, and power condition signal data. The verification criteria are: missing value ratio ≤ 0.05%, outlier ratio ≤ 0.1%, timing alignment error ≤ 0.01s, and all data are dimensionless and the numerical range is within [-3,3]. Data that passes the verification is used as input data for subsequent steps, while data that fails the verification is returned to the corresponding stage for reprocessing.

[0031] In some embodiments, the signal component decomposition process preferentially uses the variational mode decomposition process, and the fitting filter decomposition process is used when the variational mode decomposition process fails.

[0032] For example, variational mode decomposition (VMD) is used to decompose the preprocessed hydropower station multi-source signal data into components to obtain the intrinsic mode components (IMF) corresponding to each signal. In some embodiments, variational mode decomposition (VMD) is used to adaptively decompose the preprocessed hydraulic, mechanical vibration, and electrical operating condition signals one by one. This decomposes the non-stationary and nonlinear original signals into several intrinsic mode components (IMFs) with uniform frequency distribution and independent characteristics. This effectively separates the signal's trend term, high-frequency noise term, and effective feature term, ultimately improving the accuracy of subsequent model predictions. Some decomposition results are shown below. Figure 3 .

[0033] In some embodiments, with the decomposition parameters fixed, the penalty factor Noise tolerance Decomposition of mode numbers DC component suppression Initialization method Convergence threshold ; In some embodiments, the original signal is adaptively decomposed into three intrinsic mode components by iteratively optimizing and solving the constrained variational problem. , , ,satisfy ,in The original signal; In some embodiments, after decomposition, a variational mode decomposition curve is generated, which intuitively displays the time-domain waveform and frequency characteristics of each type of original signal and its corresponding three intrinsic mode components, facilitating feature analysis and verification. In some embodiments, when the variational mode decomposition process fails due to insufficient signal length or other reasons, Savitzky-Golay fitting filtering is used as an alternative decomposition method to ensure the integrity of the data decomposition. After the decomposition is completed, a variational mode decomposition (VMD) curve is generated, which can intuitively display the characteristics of each mode component. For example, based on the hydroelectric coupling physical characteristics of the water diversion system-turbine-generator in a hydropower station, the three types of signals—hydraulic, mechanical, and electrical—are not independent but have an inherent relationship of bidirectional causality, mutual coupling, and mutual influence. Therefore, this application abandons the traditional single-signal self-prediction mode and constructs a three-type multi-source cross-coupling prediction task with full dimensions and no blind spots to achieve deep coupling prediction of hydroelectric signals. All three types of tasks use the intrinsic mode components (IMF) obtained by decomposition as input features.

[0034] In some embodiments, based on the coupling characteristics of hydroelectric signals, three types of multi-source cross-coupling prediction tasks are constructed, including hydraulic operating condition signal coupling prediction task (hydraulic → mechanical / electric), electrical operating condition signal coupling prediction task (electric → hydraulic / mechanical), and mechanical vibration signal coupling prediction task (mechanical → hydraulic / electric). Through the above three types of cross-coupling tasks, all coupling relationships between hydroelectric signals are fully covered, achieving full-link prediction and conforming to the actual operating coupling law of hydropower station units.

[0035] In some embodiments, the hydraulic condition signal coupling prediction task uses the intrinsic mode components of the hydraulic condition signal as input to predict the eigenvalues ​​of the mechanical vibration signal and the eigenvalues ​​of the power condition signal, respectively. In some embodiments, the power condition signal coupling prediction task uses the intrinsic mode components of the power condition signal as input to predict the eigenvalues ​​of the hydraulic condition signal and the eigenvalues ​​of the mechanical vibration signal, respectively. In some embodiments, the mechanical vibration signal coupling prediction task uses the intrinsic mode components of the mechanical vibration signal as input to predict the eigenvalues ​​of the hydraulic operating condition signal and the eigenvalues ​​of the power operating condition signal, respectively.

[0036] For example, multi-source feature capture models include self-attention mechanism models, long short-term memory network models, and temporal convolutional network models. Hybrid base models can be constructed using self-attention mechanism (Transformer) models, long short-term memory (LSTM) network models, and temporal convolutional network (TCN) models. In some embodiments, the three types of models focus on capturing different characteristics of the signal: the Transformer model captures long-range dependencies through self-attention, the Long Short-Term Memory (LSTM) model enhances sequence memory capabilities through gating units, and the Temporal Convolutional Network (TCN) model extracts local features through dilated convolutions. The three types of models each have their own technical advantages and are highly complementary. They capture different characteristics of water, electromechanical, and electrical time-series signals to achieve synergy. Finally, the three types of coupled prediction tasks are trained and predicted independently, and the prediction results of each base model are output.

[0037] In some embodiments, the input to all models is the intrinsic mode components obtained by decomposition, and a time-series input sequence of length 48 is constructed according to time steps. The ratio of training set to test set is 8:2. Then, two types of multi-source cross-hybrid base models are trained to complete multiple sets of multi-source cross-coupling prediction tasks, and the prediction results of the corresponding multi-source cross-hybrid base models are obtained. During the training process, the corresponding base model training loss curve is generated to monitor the convergence of the training process.

[0038] In some embodiments, the self-attention mechanism (Transformer) model in the hybrid base model can accurately capture the long-range temporal dependencies of the hydroelectric signals through a multi-head self-attention mechanism. The configuration parameters are: feature dimension. Attention count The encoder has 4 layers, and the feedforward network has a dimension of 4. Position coding can be added to compensate for the lack of time-awareness; its calculation formula is as follows:

[0039]

[0040]

[0041] in, Q, K, and V are the query, key, and value matrices, respectively, and dk is the dimension of the key vector. ";" indicates concatenation in the column direction. The input is mapped to a 128-dimensional feature space through an embedding layer. After adding position encoding, it is input into a 4-layer custom encoder layer. Each layer contains a multi-head self-attention mechanism and a 512-dimensional feedforward network, which can model the long-distance temporal dependence of hydro-generator unit operation data.

[0042] In some embodiments, the Long Short-Term Memory (LSTM) network model in the hybrid base model can enhance the sequential memory characteristics of the hydroelectric signal through a gated unit structure. The configuration parameters are: hidden layer dimension of 64, LSTM unit layer of 3, and Dropout regularization coefficient of 0.3. This effectively suppresses model overfitting and adapts to the short-term memory characteristics of the signal. The calculation formula is as follows:

[0043]

[0044] in, (Forgotten Gate) (Input Gate) (Candidate cell status) (Core memory unit) (Output Gate) (Output features) are the gating / state variables of the LSTM; The signal characteristics of the previous time step. This is the current water turbine electrical signal. This is the activation function.

[0045] In some embodiments, the Temporal Convolutional Network (TCN) model in the hybrid basis model can efficiently extract local fine features and frequency features of the hydroelectric signal through dilated causal convolution and residual connections. The configuration parameters are: increasing channel number [32, 64, 128], kernel size 5, and weight normalization is used to improve training stability. The convolution transformation formula is as follows: ,in, As the expansion factor, The kernel size is used; the temporal convolutional network branch can serve as an extractor of high-frequency local dynamic features, compensating for the smoothing effect of long short-term memory networks on high-frequency features.

[0046] In some embodiments, all three models employ the AdamW optimizer with a learning rate of [missing information]. The loss function is mean squared error (MSE). Training is performed for 60 epochs with early stopping enabled (patience=10). During training, a training loss curve for the base model is generated, and the model convergence status is monitored in real time. After training, independent prediction results for the Transformer model, Long Short-Term Memory (LSTM) model, and Temporal Convolutional Network (TCN) model are output respectively. A model training loss curve is generated after prediction is completed; see the curve below. Figure 4 .

[0047] For example, a residual correction model is constructed based on a pre-defined hierarchical model fusion method, and the coupled prediction results corresponding to the multi-source cross-mixing basis model are obtained, including: Dynamic weights are calculated based on the prediction loss of the multi-source cross-mixed basis model on the validation set, and adaptive weighting of the multi-source cross-mixed basis model is performed. The prediction results and weighted average results of the multi-source cross-mixing basis model are extracted as a fusion feature set, and a residual correction model is constructed. The residual correction model is built based on the fusion feature input to obtain the residual output result, and the corrected coupled prediction result is obtained by summing the weighted average prediction result and the residual output result.

[0048] In some embodiments, an XGBoost residual correction model is constructed by employing a fusion strategy of extreme gradient boosting (XGBoost) and stacking. This model optimizes and integrates the prediction results corresponding to the output self-attention mechanism (Transformer) model, long short-term memory network (LSTM) model, and temporal convolutional network (TCN) model, thereby overcoming the limitations of single-model prediction. Through dynamic weight allocation and residual correction, the accuracy and stability of coupled prediction are further improved, ultimately outputting the optimal hydroelectric signal coupled prediction result.

[0049] In some embodiments, dynamic weights are calculated based on the prediction loss of the multi-source cross-mixed base model on the validation set, following the principle that the smaller the loss, the greater the weight. A low loss value indicates that the model has a stronger predictive ability on the coupled task, and a higher weight ratio needs to be assigned. Adaptive weighting of the multi-source cross-mixed base model is also performed. In some embodiments, the independent prediction results and weighted average results of the multi-source cross-mixing basis model are extracted as a fusion feature set to enrich the dimensions of input information and avoid the loss of information from a single feature. In some embodiments, an extreme gradient boosting (XGBoost) and stacking fusion strategy is adopted to construct an XGBoost residual correction model as a residual corrector. The model is trained with fused features as input and the residual (true value - weighted average predicted value) as output, and the optimal model parameters are fixed as follows: Learning rate Decision tree depth L1 regularization coefficient L2 regularization coefficient ; In some embodiments, a residual correction model is built based on the fused feature input to obtain the final residual output result, and the corrected coupled prediction result is obtained by summing the weighted average prediction result and the residual output result. The correction formula is as follows: , For the average prediction results, Output the residual result; In some embodiments, after fusion is completed, a final reconstruction prediction comparison map, a fusion model residual distribution map, and an input component importance map are generated. The input component importance map is obtained by calculating the error change rate through an occlusion experiment, thereby quantifying the contribution of each modal component to the prediction result.

[0050] In some embodiments, mutual information analysis and transmission entropy calculation are used to quantitatively explore the correlation strength and dynamic transmission relationship between hydroelectric signals.

[0051] Table 1 Results of Prediction Indicators for Hydro-Mechanical-Electric Coupling

[0052] Table 2 Results of Prediction Indicators for Hydro-Mechanical-Electric Coupling

[0053] Table 3 Results of Prediction Indicators for Hydro-Mechanical-Electric Coupling

[0054] Based on the experimental data in Tables 1-3, the model performance will be analyzed from the perspective of different prediction tasks: In the electrical / mechanical task of hydraulic prediction, taking water-active power prediction as an example, the fused model has a MAE of 3.69 and an RMSE of 5.52, which is significantly better than the individual Transformer (MAE: 4.55), LSTM (MAE: 4.89), and TCN (MAE: 5.79), thus improving accuracy. The fused model has an R² of over 0.99 on multiple metrics, indicating that the model can explain most of the variance in the signal and has a better fit. In mechanical prediction of hydraulic / electrical tasks, taking the prediction of peak-to-peak pressure fluctuations between the machine and the top cover as an example, the fusion model achieved an R² of 0.9316, which is about 6.99% higher than the worst-performing base model. For the complex prediction of the machine-excitation current, all models maintained high accuracy (R²>0.97), but the fusion model still ranked first with the lowest MAE of 36.92. In the electrical prediction of hydraulic / mechanical tasks, taking the prediction of electric guide vane opening as an example, the MAE of the fusion model is only 1.2312, which is a significant improvement over Transformer (1.6638) and LSTM (1.6780); Understandably, the experimental data strongly demonstrate the effectiveness of the decomposition and multi-model fusion strategy. In the prediction of all tested physical variables (such as pressure pulsation, peak-to-peak vibration, and power), the fused model comprehensively outperformed the single basis model in all three metrics: MAE, RMSE, and R². This indicates that by correcting the prediction residuals of the basis model using XGBoost, the nonlinear characteristics of hydropower unit operation can be captured more accurately.

[0055] Transformer is relatively robust when dealing with variables with global correlation, while TCN has a competitive advantage in predicting signals with drastic local fluctuations. Since each base model has a different focus in capturing features, the fusion model can effectively complement each other's strengths, thereby improving overall robustness.

[0056] In summary, the prediction method based on variational mode decomposition and base model fusion proposed in this application has extremely high accuracy and reliability in multi-parameter prediction tasks for hydropower units, providing scientific data support for unit condition monitoring and fault early warning.

[0057] For example, by analyzing mutual information (MI) to quantify the correlation strength between water turbine electrical signals, and by combining transfer entropy (TE) to calculate and analyze the dynamic transmission path and causal relationship of the signals, the intrinsic coupling mechanism of the signals can be mined.

[0058] In some embodiments, as a quantitative mining step of the hydro-mechanical-electric coupling mechanism, it is carried out in parallel with the model prediction step. Through two classic information theory methods, mutual information analysis and transfer entropy calculation, the static correlation strength and dynamic transfer relationship among the three types of signals of hydro-mechanical-electricity are quantitatively analyzed. Without subjective experience intervention, the analysis results are objective and accurate, providing theoretical support for coupling prediction.

[0059] In some embodiments, mutual information (MI) analysis quantifies the static correlation strength between any two types of signals, including hydraulic, mechanical, and electrical signals, without directional attributes. The larger the value, the stronger the correlation between the signals. The joint probability distribution and marginal probability distribution of the signals are estimated by the two-dimensional histogram method, and the mutual information values ​​between each pair of signals are calculated by substituting them into the mutual information formula. A mutual information heatmap is then constructed to visually display the ranking of the correlation between the signals. When calculating transfer entropy (TE), it quantifies the direction and intensity of dynamic information transmission between any two types of signals in hydraulic, mechanical, and electrical systems. It possesses a clear causal attribute; a larger value indicates a stronger unidirectional driving effect of the signal. The calculation formula is as follows when the signal is discretized: ,in, , The time delay parameter of the signal. This method uses the number of discretized bins to represent the continuous hydroelectric signals. By quantifying the information transmission between signals, it can accurately model the dynamic causal dependencies of hydraulic, mechanical, and electrical signals of a hydroelectric generator unit. Discretization is employed to map the signals to a finite state space. Delay parameters k=1, l=1, and the number of bins n_bins=10 are set. The transmission entropy between signals is calculated, a causal network diagram of signal transmission is constructed, and the dynamic influence paths between signals are analyzed.

[0060] Table 4. Results of Electrical Signal Transmission for Water Turbines

[0061] Table 5. Results of water turbine electrical signal transmission 2

[0062] Table 6. Results of water turbine electrical signal transmission 3

[0063] Table 7 Results of Water Turbine Electrical Signal Transmission 4

[0064] Table 8. Results of water turbine electrical signal transmission 5

[0065] Based on the experimental data in Tables 4-8, the quantitative analysis system of the coupling relationship between the electrical, hydraulic and mechanical fields of the hydropower unit quantifies the intensity of dynamic causal influence and information flow between various physical quantities through the transfer entropy (TE) index.

[0066] The electro-hydraulic correlation analysis revealed the causal chain between the excitation system (Uf, If) and power (P, Q) and the regulation system and hydraulic pressure fluctuations (GV, Psc, etc.). The results showed that the electrical fluctuations have a significant driving effect on the hydraulic field. The electromechanical correlation analysis (X / Y direction) includes vibration data of the stator frame, stator core and upper and lower frames in two horizontal dimensions (X and Y directions), quantifying the information transmission between electromagnetic force (represented by excitation and power) and the mechanical structure response of the unit; The hydraulic-mechanical correlation analysis (X / Y direction) analyzed the dynamic coupling between the guide vane opening and pressure pulsation at various parts and the vibration of each bearing and frame of the unit, demonstrating how water pressure fluctuations are converted into mechanical energy and induce multi-dimensional structural vibrations; The above data not only provides the magnitude of the entropy transfer between various parameters, but also clarifies the master-slave relationship of cause and effect by calculating the net information content, laying a data foundation for in-depth tracing of unit operating status and multi-field feature fusion prediction.

[0067] Reference Figure 11The present invention also provides an embodiment of a multi-source signal coupling prediction system for hydroelectric systems, used to run the multi-source signal coupling prediction method for hydroelectric systems described in the above embodiment. The prediction system includes: The data processing module is used to preprocess, analyze mutual information, and calculate the transfer entropy of multi-source signals from hydroelectric turbines in hydropower stations. The task building module is used to construct multi-source cross-coupled prediction tasks; The model building module is used to build multi-source cross-mixed basis models and residual correction models; The visualization module is used to generate signal decomposition curves, base model training loss curves, reconstruction prediction comparison charts, fusion model residual distribution charts, and mutual information heatmaps.

[0068] For example, the data processing module performs standardization, denoising, and completion preprocessing operations on the collected raw data of hydraulic, mechanical vibration, and power signals from the hydropower station, and completes the quantification of the correlation strength between the multi-source signals and the mining of dynamic transmission relationships through mutual information analysis and transfer entropy calculation; the task construction module, based on the coupling characteristics of the hydroelectric system, specifically constructs multi-source cross-coupling prediction tasks for hydraulic signals to predict mechanical vibration and power signals, power signals to predict hydraulic and mechanical vibration signals, and mechanical vibration signals to predict hydraulic and power signals; the model construction module constructs a multi-source cross-hybrid basis model that integrates the advantages of multiple models, and... The system constructs a residual correction model based on a fusion strategy, completes the training, optimization, and coupled prediction tasks of the two types of models, and outputs the prediction results of each base model and the fusion model. The visualization module visualizes the processing and analysis results of each stage of the system, and can automatically generate signal decomposition curves, base model training loss curves, reconstruction prediction comparison charts, fusion model residual distribution charts, and mutual information heatmaps. It intuitively displays the signal decomposition effect, model training process, prediction accuracy, residual distribution characteristics, and coupling correlation laws between multi-source signals, providing visualized data charts for analyzing the operating status and coupling mechanism of the hydro-mechanical system.

[0069] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting multi-source signals coupled in hydroelectric systems, characterized in that, include: Based on the preprocessed hydropower station electromechanical multi-source signal data, a signal component decomposition process is performed to obtain the intrinsic mode components corresponding to different hydropower station electromechanical multi-source signal data. A multi-source cross-coupling prediction task is constructed based on the decomposed intrinsic mode components, and the correlation coupling characteristics corresponding to the multi-source signal data of different hydropower stations are predicted. A multi-source cross-hybrid basis model is constructed based on the preset multi-source feature capture model, and the relevant prediction results corresponding to the multi-source cross-coupling prediction task are obtained. A residual correction model is constructed based on the preset hierarchical model fusion method, and the coupling prediction results corresponding to the multi-source cross-mixing basis model are obtained. Based on the preprocessed multi-source signal data of hydropower station electromechanical systems, a mutual information analysis process and a transfer entropy calculation process are performed, and the correlation strength and dynamic transfer relationship of the multi-source signal data of hydropower station electromechanical systems are quantitatively analyzed.

2. The hydroelectric multi-source signal coupling prediction method according to claim 1, characterized in that, The hydropower station's multi-source signal data includes hydraulic operating condition signals, mechanical vibration signals, and electrical operating condition signals.

3. The hydroelectric multi-source signal coupling prediction method according to claim 1, characterized in that, The preprocessing process for the multi-source signal data of hydropower station includes missing value removal, outlier handling, time-series signal alignment, standardization and normalization, and data quality verification.

4. The hydroelectric multi-source signal coupling prediction method according to claim 1, characterized in that, The removal of missing values ​​includes: Identifying global outliers based on the Laida criterion; Identification of local outliers based on the sliding window quartile method; Global and local outliers are corrected through a moving average filtering process.

5. The hydroelectric multi-source signal coupling prediction method according to claim 1, characterized in that, The signal component decomposition process preferentially uses variational mode decomposition. When the variational mode decomposition process fails, the fitting filter decomposition process is used.

6. The hydroelectric multi-source signal coupling prediction method according to claim 1, characterized in that, The multi-source cross-coupling prediction task includes hydraulic condition signal coupling prediction task, power condition signal coupling prediction task and mechanical vibration signal coupling prediction task.

7. The hydroelectric multi-source signal coupling prediction method according to claim 6, characterized in that, The hydraulic operating condition signal coupling prediction task uses the intrinsic mode components of the hydraulic operating condition signal as input to predict the eigenvalues ​​of the mechanical vibration signal and the eigenvalues ​​of the power operating condition signal, respectively. The power condition signal coupling prediction task uses the intrinsic mode components of the power condition signal as input to predict the eigenvalues ​​of the hydraulic condition signal and the eigenvalues ​​of the mechanical vibration signal, respectively. The mechanical vibration signal coupling prediction task uses the intrinsic mode components of the mechanical vibration signal as input to predict the eigenvalues ​​of the hydraulic operating condition signal and the eigenvalues ​​of the power operating condition signal, respectively.

8. The hydroelectric multi-source signal coupling prediction method according to claim 1, characterized in that, The multi-source feature capture model includes a self-attention mechanism model, a long short-term memory network model, and a temporal convolutional network model. The decomposed multiple sets of intrinsic modal components are used to construct a time series according to time steps and are divided into training and test sets as input data for the multi-source cross-hybrid basis model. The training task of the multi-source cross-hybrid basis model is to construct multiple sets of multi-source cross-coupled prediction tasks.

9. The hydroelectric multi-source signal coupling prediction method according to claim 1, characterized in that, The method for constructing a residual correction model based on a pre-defined hierarchical model fusion method and obtaining the coupling prediction results corresponding to the multi-source cross-mixing basis model includes: Dynamic weights are calculated based on the prediction loss of the multi-source cross-mixed basis model on the validation set, and adaptive weighting of the multi-source cross-mixed basis model is performed. The prediction results and weighted average results of the multi-source cross-mixing basis model are extracted as a fusion feature set, and a residual correction model is constructed. The residual correction model is built based on the fusion feature input to obtain the residual output result, and the corrected coupled prediction result is obtained by summing the weighted average prediction result and the residual output result.

10. A multi-source signal coupling prediction system for hydroelectric systems, characterized in that, The prediction system is used to run the hydroelectric multi-source signal coupling prediction method as described in any one of claims 1 to 9, the prediction system comprising: The data processing module is used to preprocess, analyze mutual information, and calculate the transfer entropy of multi-source signals from hydroelectric turbines in hydropower stations. The task building module is used to construct multi-source cross-coupled prediction tasks; The model building module is used to build multi-source cross-mixed basis models and residual correction models; The visualization module is used to generate signal decomposition curves, base model training loss curves, reconstruction prediction comparison charts, fusion model residual distribution charts, and mutual information heatmaps.

Images