Intelligent speed regulation control method and system for hydroelectric generating set multi-source data fusion

Abstract

The invention provides an intelligent speed regulation control method and system for hydroelectric generating set multi-source data fusion, and relates to the technical field of hydropower station automation control, and the method comprises the steps: S1, collecting multi-source heterogeneous data, including mechanical quantity, hydraulic parameters and electrical quantity, in the operation process of a hydroelectric generating set; s2, performing standardized preprocessing on the multi-source heterogeneous data to obtain preprocessed data; s3, on the basis of the preprocessed data, global deep state feature information and operation trend prediction information of the hydroelectric generating set are extracted through a multi-layer data fusion algorithm, and a fused speed regulation decision basis is obtained; s4, generating a speed regulation control instruction for regulating the rotating speed of the hydroelectric generating set based on the fused speed regulation decision basis; and S5, performing speed regulation control on the hydroelectric generating set according to the speed regulation control instruction. Through multi-layer data fusion and a dynamic self-adaptive control strategy, comprehensive perception and accurate regulation and control of the operation state of the hydroelectric generating set are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of hydropower station automation control, and in particular to an intelligent speed regulation control method and system for multi-source data fusion of a hydropower unit. Background Art

[0002] As complex power system equipment, hydropower units generate a large amount of multi-dimensional, heterogeneous data during operation. This data primarily includes three categories: mechanical quantities (such as speed, guide vane opening, and vibration signals), hydraulic parameters (such as head, flow, volute pressure, and tailwater pressure), and electrical quantities (such as active power, reactive power, voltage, and current). This data, characterized by inconsistent acquisition frequencies, heterogeneous dimensions, and varying signal-to-noise ratios, reflects the operating status and performance changes of hydropower units under different operating conditions and serves as a fundamental source of information for precise control and condition monitoring of hydropower units.

[0003] The speed control system of a hydropower unit is a core component that ensures safe and stable operation. Its primary function is to control turbine speed and power output to adapt to grid load fluctuations and frequency regulation requirements. Traditional hydropower unit speed control primarily uses the PID control algorithm, which measures unit speed deviations and calculates corresponding control signals to adjust the guide vanes or governor opening to achieve unit speed control. With the development of intelligent control technology, intelligent algorithms such as fuzzy control and neural network control have been gradually applied to speed control systems, improving the system's nonlinear adaptability.

[0004] However, existing hydropower unit speed control technology still has several limitations. First, traditional PID control parameters are fixed, making it difficult to adapt to changing operating conditions and achieving a balance between response speed and stability. Second, existing control methods typically only utilize a limited number of key parameters for control decisions, failing to fully tap into and utilize the rich information contained in multi-source data. Finally, the control system's ability to resist disturbances is limited in the face of grid disturbances and sudden load changes, making predictive control and early response difficult. These issues limit the regulation performance and operating efficiency of hydropower units under complex operating conditions. Summary of the Invention

[0005] In light of this, this paper proposes an intelligent speed control method and system for hydropower units that integrates multi-source data. This method aims to overcome the challenges of traditional speed control systems, such as the difficulty in balancing response speed and stability, poor adaptability to operating conditions, and insufficient data utilization. By effectively integrating the multi-source, heterogeneous data generated during hydropower unit operation and establishing a multi-level data fusion and processing architecture, accurate perception and prediction of the hydropower unit's operating status are achieved, thereby improving the accuracy, robustness, and anti-disturbance capabilities of speed control, enabling hydropower units to more stably and efficiently cope with complex and changing operating environments and grid conditions.

[0006] The technical solution of the present invention is achieved as follows:

[0007] In one aspect, the present invention provides an intelligent speed control method for a hydropower unit by integrating multi-source data, comprising:

[0008] S1. Collect multi-source heterogeneous data during the operation of hydropower units, including mechanical quantities, hydraulic parameters, and electrical quantities;

[0009] S2. Perform standardized preprocessing on multi-source heterogeneous data to obtain preprocessed data;

[0010] S3. Based on the pre-processed data, a multi-layer data fusion algorithm is used to extract the global deep state characteristic information and operation trend prediction information of the hydropower unit, and the fused speed regulation decision basis is obtained;

[0011] S4. Based on the integrated speed regulation decision basis, generate a speed regulation control instruction for regulating the speed of the hydropower unit;

[0012] S5. Execute speed control on the hydropower unit according to the speed control instruction.

[0013] Preferably, in step S2, the standardization preprocessing includes at least one operation of data clock synchronization, data resampling, data denoising, data filtering, data normalization and data integrity check to obtain the preprocessed data.

[0014] Preferably, in step S3, the multi-layer data fusion algorithm includes:

[0015] Low-level feature extraction and quality assessment: extracting engineering features from preprocessed data and calculating their dynamic quality factors;

[0016] Mid-level subsystem state estimation: This method estimates the state of key subsystems of the hydropower unit based on the underlying engineering characteristics and dynamic quality factors, and obtains the state estimation value of each subsystem and the global state portrait;

[0017] High-level global deep fusion uses the attention mechanism to extract the global deep state feature information of the hydropower unit from the global state portrait, and predicts future key performance indicators as operation trend prediction information, and then comprehensively forms the fused speed regulation decision basis.

[0018] Preferably, in the underlying feature extraction and quality assessment:

[0019] Extract the time domain features, frequency domain features and statistical features of the preprocessed data as engineering features F x (t);

[0020] Calculate the dynamic quality factor Q for each engineering feature at each sampling time t x (t), which is calculated as:

[0021] Q x (t)=w1·Q x,1 (t)+w2·Q x,2 (t)+w3·Q x,3 (t)

[0022] Among them, Q x,1 (t) is the quality assessment factor based on characteristic stability, Q x,2 (t) is the quality assessment factor based on the characteristic signal-to-noise ratio, Q x,3 (t) is the quality assessment factor based on mutation detection, w1, w2, w3 are the weights of each dimension, and w1+w2+w3=1.

[0023] Preferably, in the state estimation of the middle-level subsystem:

[0024] Divide the hydropower unit into several key subsystems, including the governor hydraulic servo system, turbine water diversion and hydraulic disturbance system, generator and its excitation system;

[0025] Build a nonlinear state-space model for each subsystem:

[0026] x j,k =f j (x j,k-1 ,u j,k-1 )+w j,k-1

[0027] z j,k =h j (x j,k )+v j,k

[0028] Among them, x j,k is the state vector of subsystem j at time k, u j,k-1 is the control input, z j,k is the characteristic observation vector associated with subsystem j, f j (·) is the nonlinear state transfer function, h j (·) is the nonlinear observation function, w j,k-1 and v j,k are process noise and measurement noise, respectively;

[0029] The state estimation is performed using an extended Kalman filter based on dynamic weighted covariance, where the i-th diagonal element of the measurement noise covariance matrix is dynamically adjusted as follows:

[0030]

[0031] Among them, R base,x,i is the reference measurement noise covariance, α is the adjustment factor, Qx,i (k) is the dynamic quality factor of the corresponding feature at time k, and ∈ is a small positive constant to prevent the denominator from being zero.

[0032] Preferably, in the high-level global deep fusion:

[0033] Combine the outputs of the mid-level subsystem state estimates into a global state profile X seq (k);

[0034] Use a deep neural network based on a multi-head self-attention mechanism to process the global state portrait X seq (k), extract context-aware sequence representation C seq (k);

[0035] From context-aware sequence representation C seq (k) extracts the global deep state feature vector GSF(k) and predicts the key performance indicators PF(k+1,...,k+M) for the next M time steps;

[0036] Calculate the global state fusion confidence C based on the attention weight distribution and prediction results global (k).

[0037] Preferably, step S4 includes:

[0038] S41. Construct a robust PID controller with dynamic adaptive parameter adjustment. The controller output is:

[0039]

[0040] Among them, e(t) is the speed or frequency deviation, u(t) is the control output, K p ,K i ,K d are the PID controller parameters;

[0041] S42, according to the global deep state feature vector GSF(k) and the global state fusion confidence C global (k) Dynamically adjust PID controller parameters:

[0042] K p (k) = K p0 α p (GSF(k),C global (k))

[0043] K i (k) = K i0 α i (GSF(k),C global (k))

[0044] K d(k) = K d0 α d (GSF(k),C global (k))

[0045] Among them, K p0 , K i0 and K d0 is the benchmark PID parameter; α p , α i and α d is the adjustment factor;

[0046] S43. Based on the key performance indicators PF(k+1,...,k+M) of the next M time steps and the global deep state feature vector GSF(k), calculate the feedforward compensation:

[0047] u ff (k) = f ffc (PF(k+1,...,k+M),GSF(k))

[0048] Among them, f ffc It is the feedforward compensation calculation function;

[0049] S44. Combine the PID controller output and the feedforward compensation amount to obtain the final speed control instruction.

[0050] Preferably, in step S42, α p , α i and α d Computed via a small feedforward neural network:

[0051] [α p ,α i ,α d ]=NN PID (GSF(k),C global (k));

[0052] Among them, the neural network NN PID The design is a three-layer feedforward network, the input layer receives the concatenated GSF(k) vector and a single C global (k) value, the hidden layer uses the ReLU activation function, and the output layer uses the Sigmoid function and is scaled to the range of [0.5, 2.0].

[0053] Preferably, the method further comprises:

[0054] Real-time monitoring of safety-related parameters during the operation of hydropower units;

[0055] Determine whether safety-related parameters exceed preset safety boundaries;

[0056] When it is detected that any safety-related parameter exceeds its corresponding preset safety boundary, a preset protection strategy is automatically triggered, and the protection strategy includes at least one of an alarm prompt, a load adjustment rate limit, and an emergency shutdown.

[0057] Another method, the present invention also provides an intelligent speed control system for a hydropower unit with multi-source data fusion, the system being used to implement any of the above methods, the system comprising:

[0058] Data acquisition module, used to collect multi-source heterogeneous data during the operation of hydropower units, including mechanical quantities, hydraulic parameters and electrical quantities;

[0059] The data preprocessing module is used to perform standardized preprocessing on multi-source heterogeneous data to obtain preprocessed data;

[0060] The multi-layer data fusion module is used to extract the global deep state characteristic information and operation trend prediction information of the hydropower unit based on the pre-processed data through the multi-layer data fusion algorithm, and obtain the fused speed regulation decision basis;

[0061] A speed control instruction generation module is used to generate a speed control instruction for adjusting the speed of the hydropower unit based on the integrated speed control decision basis;

[0062] The speed control execution module is used to execute speed control on the hydropower unit according to the speed control instruction.

[0063] The present invention has the following beneficial effects compared to the prior art:

[0064] (1) The present invention achieves comprehensive perception and precise control of the operating status of hydropower units through multi-layer data fusion and dynamic adaptive control strategies. This method effectively integrates multi-source heterogeneous data, extracts deep state characteristics, and improves the adaptability of the speed control system to complex working conditions, thereby improving control accuracy, reducing the oscillation amplitude during the transition process, and shortening the adjustment time. This effectively solves the technical problem of the traditional speed control system that is difficult to balance response speed and stability.

[0065] (2) The multi-layer progressive data fusion architecture adopted by the present invention enables the system to extract high-quality state features from disordered multi-source data through hierarchical processing of bottom-level feature extraction and quality assessment, middle-level subsystem state estimation, and high-level global deep fusion;

[0066] (3) The present invention combines the extended Kalman filter with a dynamic weight adjustment strategy to perform subsystem state estimation, and constructs a measurement noise covariance matrix through a real-time calculated dynamic quality factor, thereby improving the accuracy and robustness of state estimation;

[0067] (4) The high-level fusion module based on the multi-head self-attention mechanism of the present invention can effectively capture the temporal dependencies and nonlinear characteristics in the operation data of hydropower units, extract the global deep state feature vector and predict key performance indicators;

[0068] (5) The control strategy of the present invention, which combines dynamic adaptive PID control of parameters with feedforward compensation and safety protection mechanism, not only achieves precise regulation of the hydropower unit, but also ensures the safety of system operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0070] Figure 1 is a flow chart of the method of the present invention;

[0071] Figure 2 It is a technical implementation diagram of the present invention;

[0072] Figure 3 Schematic diagram of the system module of the present invention. DETAILED DESCRIPTION

[0073] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described 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 making creative efforts are within the scope of protection of the present invention.

[0074] like Figure 1 As shown, the present invention provides an intelligent speed control method for a hydropower unit by fusion of multi-source data, comprising:

[0075] S1. Collect multi-source heterogeneous data during the operation of hydropower units, including mechanical quantities, hydraulic parameters, and electrical quantities;

[0076] S2. Perform standardized preprocessing on multi-source heterogeneous data to obtain preprocessed data;

[0077] S3. Based on the pre-processed data, a multi-layer data fusion algorithm is used to extract the global deep state characteristic information and operation trend prediction information of the hydropower unit, and the fused speed regulation decision basis is obtained;

[0078] S4. Based on the integrated speed regulation decision basis, generate a speed regulation control instruction for regulating the speed of the hydropower unit;

[0079] S5. Execute speed control on the hydropower unit according to the speed control instruction.

[0080] like Figure 2 As shown, the technical implementation process of the present invention is as follows: first, multi-source heterogeneous data (including mechanical quantities, hydraulic parameters and electrical quantities) during the operation of the hydropower unit are collected, and standardized pre-processing such as clock synchronization, resampling, and denoising is performed; then a three-layer progressive data fusion algorithm is used for processing: the bottom layer extracts and evaluates the features through dynamic quality factor evaluation, the middle layer uses the extended Kalman filter to estimate the subsystem state, and the top layer extracts the global deep state feature vector based on the multi-head self-attention mechanism and predicts the key performance indicators; finally, a control strategy combining parameter dynamic adaptive PID control and feedforward compensation is implemented based on the fused information, and a safety protection mechanism is set at the same time, thereby achieving precise regulation and safe operation of the hydropower unit. The present invention overcomes the technical problems of traditional speed control systems such as difficulty in balancing response speed and stability and poor adaptability to working conditions through the full mining and fusion of multi-source data.

[0081] Specifically, in one embodiment of the present invention, step S1 includes:

[0082] Through the sensor network set up at key parts of the hydropower unit, three types of multi-source heterogeneous data are collected during the operation of the hydropower unit:

[0083] Mechanical quantities include the unit speed signal ω, which is collected through a magnetoelectric speed sensor or a photoelectric speed sensor and is used to monitor the real-time speed of the unit; the guide vane opening signal α, which is collected through a guide vane opening sensor or a displacement sensor and reflects the opening degree of the turbine guide vanes; the main pressure regulating valve stroke signal, which is collected through a displacement sensor and is used to monitor the position of the speed control system actuator; the vibration signal, which collects the vibration acceleration or displacement of key parts such as the top cover, water guide bearing, thrust bearing, and generator base, and is used to monitor the dynamic characteristics of the mechanical system; and the bearing temperature, which is collected through a thermocouple or PT100 temperature sensor and is used to monitor the thermal state of the bearing.

[0084] Hydraulic parameters, including volute pressure P sc , collected by pressure sensor, reflects the hydraulic conditions at the turbine inlet; the tailwater pipe pressure P dt Or water level, collected through pressure sensors or water level gauges to monitor outlet water pressure or water level; upstream water level (head H), collected through water level gauges to monitor available hydraulic resources.

[0085] Electrical quantities, including generator active power P e , collected through the power monitoring device, reflecting the output power of the unit; reactive power Qe , collected through the power monitoring device, reflecting the reactive power regulation of the unit; the grid frequency f, obtained through a frequency meter or from the grid monitoring system, monitors the grid frequency changes; the machine-end voltage U, collected through a voltage transformer and measuring equipment, monitors the generator output voltage.

[0086] Specifically, in one embodiment of the present invention, in step S2, the standardization preprocessing includes: at least one operation of data clock synchronization, data resampling, data denoising, data filtering, data normalization and data integrity check to obtain the preprocessed data.

[0087] In a specific example, the normalization preprocessing process is as follows:

[0088] Clock synchronization and resampling: A high-precision GPS clock signal is used as a synchronization reference to ensure millisecond-level timestamp accuracy across different data sources. For sensor data with different sampling frequencies, cubic spline interpolation is used for resampling, uniformly resampling all data to a consistent baseline sampling rate for time alignment. High-frequency signals (such as vibration) are sampled at 2kHz; intermediate-frequency signals (such as pressure) are sampled at 500Hz; and low-frequency signals (such as temperature and water level) are sampled below 10Hz. Anti-aliasing filtering is employed during resampling to ensure spectral distortion.

[0089] Denoising and filtering: Adaptive filtering methods are used according to different signal characteristics:

[0090] Vibration signal: Adaptive denoising is performed using variational mode decomposition (VMD). VMD decomposes the signal into a series of intrinsic mode functions (IMFs) using the following formula: where u k is the decomposed IMF component, ω k is the center frequency of each component, and f is the original signal. Denoising is achieved by removing high-frequency IMF components.

[0091] Pressure signal: The Savitzky-Golay filter is used for smoothing, which can effectively preserve the peak characteristics of the signal: where c j is the convolution coefficient, which is determined by polynomial fitting, and m is the half-width of the filter window.

[0092] Electrical signal: A combination of low-pass filtering and median filtering is used to remove high-frequency noise and sudden interference.

[0093] Normalization: Use the Z-score normalization method to normalize data of different dimensions to a unified range: where μ X and σ Xare the mean and standard deviation of signal X. For signals with strong non-Gaussian distribution, piecewise linear normalization or quantile-based normalization is used.

[0094] Data integrity check: Using window sliding technology, real-time detection of data loss, sudden jumps and other anomalies, and marking data quality indicators. Calculation of data integrity rate C_r:

[0095]

[0096] Among them, N valid is the effective data volume, N total is the total data volume, when the completeness rate C r When the value is lower than the preset threshold, the data completion mechanism is triggered to complete the missing data through linear interpolation or forward filling.

[0097] Specifically, in one embodiment of the present invention, in step S3, the multi-layer data fusion algorithm includes:

[0098] Low-level feature extraction and quality assessment: extracting engineering features from preprocessed data and calculating their dynamic quality factors;

[0099] Mid-level subsystem state estimation: This method estimates the state of key subsystems of the hydropower unit based on the underlying engineering characteristics and dynamic quality factors, and obtains the state estimation value of each subsystem and the global state portrait;

[0100] High-level global deep fusion uses the attention mechanism to extract the global deep state feature information of the hydropower unit from the global state portrait, and predicts future key performance indicators as operation trend prediction information, and then comprehensively forms the fused speed regulation decision basis.

[0101] In this embodiment, step S3 uses a three-layer progressive data fusion architecture, from bottom-level feature extraction to high-level deep fusion, to gradually refine the value of data. This multi-layer data fusion algorithm includes three levels: bottom-level feature extraction and quality assessment, mid-level subsystem state estimation, and high-level global deep fusion. It then extracts global deep state feature information and operational trend prediction information of the hydropower unit, forming the basis for the fused speed regulation decision.

[0102] S3.1 Low-level feature extraction and quality assessment

[0103] Engineering features are extracted from the preprocessed data, and dynamic quality factors are calculated in real time for each feature to evaluate its reliability and effectiveness.

[0104] From each pre-processed data stream, a set of predefined engineering features that can reflect the operating status of different aspects of the unit are extracted. M): For example, the guide vane opening change rate, the main shaft specific frequency band vibration energy, the water guide bearing oil temperature change gradient. Hydraulic characteristics (set F H ): For example, the amplitude of volute pressure fluctuation, the main peak of tailwater pressure frequency, and the change of unit reference flow rate per unit time. Electrical characteristics (set F E ): For example, the total harmonic distortion (THD) of the stator current, the excitation voltage ripple coefficient, and the acceleration of active power change. Feature extraction uses a variety of methods such as time domain features (such as mean, change rate, and fluctuation range), frequency domain features (such as spectrum peaks and specific frequency band energy), and statistical features (such as skewness, kurtosis, and entropy) to form a comprehensive feature vector F x (t), which can be specifically expressed as a time series: F M,i (t),F H,j (t),F E,k (t).

[0105] For each extracted feature F x (t) Calculate a dynamic quality factor Q at each sampling time t x (t), the value range is (0,1], the larger the value, the higher the quality.

[0106] Calculation method:

[0107] Dimension 1: Quality assessment based on feature stability. Calculate feature F x (t) Volatility within a short time window: σ 2 (F x [tw:t]) is the variance of the feature in the window [tw:t], μ 2 (F x [tw:t]) is the square of the mean, ∈ is a small positive number to prevent the denominator from being zero, and w is the window length.

[0108] Dimension 2: Quality assessment based on characteristic signal-to-noise ratio. Use methods such as wavelet decomposition or empirical mode decomposition (EMD) to estimate the signal-to-noise ratio: P signal (t) and P noise (t) are the estimated power of signal and noise respectively.

[0109] Dimension 3: Quality assessment based on mutation detection. Cumulative sum (CUSUM) is used to detect F x (t) Whether a mutation occurs. If a mutation occurs, temporarily reduce its quality factor. Q x,3 (t) is a penalty factor based on the mutation detection result (e.g., 1 when there is no mutation and a value between 0.1-0.5 when there is mutation).

[0110] Comprehensive quality factor: Q x(t)=w1·Q x,1 (t)+w2·Q x,2 (t)+w3·Q x,3 (t), where w1, w2, w3 are the weights of each dimension, ∑w i =1.

[0111] The bottom fusion module outputs each feature sequence F x (t) and its corresponding dynamic quality factor sequence Q x (t), serves as the input for the state estimation of the middle-level subsystem.

[0112] S3.2 Middle-level subsystem state estimation

[0113] Based on the underlying engineering characteristics and dynamic quality factors, the state of the key subsystems of the hydropower unit is estimated to obtain the state estimation value of each subsystem and the global state portrait.

[0114] The hydropower unit is divided into three key subsystems: the governor hydraulic servo system S1; the turbine water diversion and hydraulic disturbance system S2; and the generator and its excitation system S3.

[0115] Each subsystem S j Formulate the discrete-time nonlinear state equation:

[0116] x j,k =f j (x j,k-1 ,u j,k-1 )+w j,k-1

[0117] And the observation equation:

[0118] z j,k =h j (x j,k )+v j,k

[0119] Among them, x j,k is the state vector of subsystem j at time k, u j,k-1 is the control input, z j,k is the characteristic observation vector associated with subsystem j, f j (·) is the nonlinear state transfer function, h j (·) is the nonlinear observation function, w j,k-1 and v j,k are process noise and measurement noise, respectively;

[0120] For each subsystem S j , the extended Kalman filter (EKF) is applied for state estimation.

[0121] For the observation vector z j,kThe i-th observation feature F in x,i (k)(its dynamic quality factor is Q x,i (k), which is in R j,k The corresponding diagonal elements in R j,k (i,i) is calculated as follows:

[0122]

[0123] Among them, R base,x,i is the reference measurement noise covariance, α is the adjustment factor, Q x,i (k) is the dynamic quality factor of the corresponding feature at time k, and ∈ is a small positive constant to prevent the denominator from being zero.

[0124] This method of dynamically adjusting the measurement noise covariance enables EKF to adaptively adjust the trust in different observations according to the data quality: when the data quality is high (Q x,i When (k) is close to 1), the corresponding observation noise covariance decreases, and the filter trusts the observation more; when the data quality is low (Q x,i When (k) is close to 0), the corresponding observation noise covariance increases, and the filter tends to trust the model prediction more.

[0125] In this embodiment, the EKF execution process is as follows:

[0126] a. Prediction step: (where A j,k is f j The Jacobian matrix, W j,k is the process noise Jacobian, Q j,k-1 is the process noise covariance).

[0127] b. Update step: calculate the Kalman gain K j,k , update the state estimate Update the estimated covariance P j,k∣k =(IK j,k H j,k )(where H j,k It is h j Jacobian matrix).

[0128] The output of the middle-level fusion module includes: the optimized state estimation vector sequence of each subsystem and the corresponding state estimation covariance matrix sequence P j,k∣k The global state profile X formed by integrating the states of all subsystems seq (k).

[0129] S3.3 High-Level Global Deep Fusion

[0130] The attention mechanism is used to extract the global deep state feature information of the hydropower unit from the global state portrait, and predict future key performance indicators as operation trend prediction information.

[0131] The state estimation vectors of each subsystem output by the middle-level subsystem state estimation are (forj=1...N) in the most recent time window W L (For example, the data within the past L sampling points) are spliced and sorted to form a multivariate time series matrix Among them D state is the total dimension of the state vectors of all subsystems.

[0132] Optionally, the estimated covariance P of each state can be j,k∣k The diagonal elements (variance) or their determinants are used as uncertainty measures, and X seq (k) Parallel input or fusion to the embedding layer. For example, the embedding layer: X seq (k) Mapping to a higher-dimensional feature space through a learnable embedding layer (such as a fully connected layer or a 1D convolutional layer) to obtain an embedded sequence

[0133] A deep neural network based on the Multi-head Self-Attention (MHSA) mechanism is used to process the global state portrait X. seq (k) (If the initial global state portrait is embedded, the embedded sequence E seq (k) is the input sequence X at this time seq (k)), extracting context-aware sequence representation C seq (k), the specific process is as follows:

[0134] First, the input sequence X seq (k) is input to the multi-head self-attention module and mapped into query (Query), key (Key) and value (Value) through linear transformation:

[0135] Q=X seq W Q

[0136] K=X seq W K

[0137] V=X seq W

[0138] Among them, W Q 、W K and W V is the learnable weight matrix.

[0139] Then, the attention weights and weighted outputs are calculated:

[0140]

[0141] Among them, Q, K, V are query matrix, key matrix and value matrix respectively, d k is the dimension of the key vector, which is used to scale the dot product result. The softmax function converts the attention score into a probability distribution.

[0142] Multi-head attention executes the above process multiple times in parallel and then concatenates the results:

[0143] MultiHead(X)=Concat(head1,head2,...,head h )W O

[0144] where head i =Attention(XW i Q ,XW i K ,XW i V )

[0145] Among them, W O and W i Q 、W i K 、W i V are all learnable parameters.

[0146] MHSA performs weighted aggregation of sequence information by calculating the correlation (attention weight) between features at each time step within the sequence. This allows the model to focus on the historical information segments and state combinations that are most important to the current overall state and future trends. Multiple attention heads learn different dependency patterns in parallel.

[0147] The attention mechanism can go beyond simple time series averaging or filtering, capture complex nonlinear dependencies and key dynamic patterns, thereby "enhancing" the information content of the original state estimation sequence and generating more representative deep features.

[0148] Output: After MHSA and subsequent feedforward network layers and layer normalization, a context-aware sequence representation is obtained.

[0149] From C seq (k) The output vector of the last time step (or C seq(k) Performing pooling operations, such as average pooling or maximum pooling, and mapping through one or more fully connected layers to obtain a global deep state feature vector that represents the overall operating state of the current unit This vector is a highly abstract and condensed representation of the original multidimensional state.

[0150] C seq (k)(or C last (k) is input into another independent prediction head (composed of one or more fully connected layers) to predict key performance indicators (KPIs) for the next M time steps, such as the unit speed deviation series and active power output series, to obtain the prediction feature vector PF(k+1,...,k+M), which includes: future speed deviation prediction, future power fluctuation prediction, and future vibration level prediction.

[0151] Based on the entropy of the attention weight distribution output by MHSA (the smaller the entropy, the more focused the attention, and the higher the confidence). Based on the variance of the predicted feature sequence PF or the width of the prediction interval (the smaller the variance or the narrower the interval, the higher the confidence). Combining the above indicators, through weighted summation or fuzzy logic rules, the global state fusion confidence C is obtained. global (k)∈[0,1]:C global (k) = β1·(1-H norm (A w ))+β2·C pred (k), where H norm (A w ) is the normalized attention weight entropy, which indicates the certainty of attention distribution; C pred (k) is the prediction consistency index, which is based on the output consistency of multiple prediction heads; β1 and β2 are weight coefficients, and β1+β2=1.

[0152] The output of the high-level fusion module is: global deep state feature vector GSF(k); key performance indicator prediction sequence PF(k+1,...,k+M); confidence evaluation C of the global state vector global (k).

[0153] In this embodiment, the training and updating of the attention mechanism adopts an offline / online hybrid strategy:

[0154] Offline: The attention network parameters are trained through supervised learning using historical operational data. The loss function includes state estimation error and prediction error terms. Online: Incremental learning is used to fine-tune the network parameters based on real-time feedback (such as control effect evaluation and prediction accuracy). Updates are performed every 100-1000 samples (approximately 5-50 seconds) to ensure computational efficiency.

[0155] Through the above-mentioned three-layer progressive data fusion architecture, the present invention enables the system to extract high-quality status features from disordered multi-source heterogeneous data, thereby achieving accurate perception of the operating status of hydropower units and scientific prediction of future trends.

[0156] Specifically, in one embodiment of the present invention, step S4 is based on the fusion decision basis output from step S3 (global deep state feature vector GSF(k), key performance indicator prediction sequence PF(k+1,...,k+M) and global state fusion confidence C global (k)) Generates speed control instructions for adjusting the speed of the hydropower unit. This step uses a strategy that combines parameter dynamic adaptive PID control with feedforward compensation, and sets a safety protection mechanism to achieve precise control of the hydropower unit speed. It includes:

[0157] S41. Construct a robust PID controller with dynamic adaptive parameter adjustment. Traditional speed regulators typically use fixed-parameter PID controllers, which cannot adapt to the dynamic characteristics of hydropower units under different operating conditions. This embodiment uses a dynamic adaptive PID controller with parameters based on global deep state characteristics, and its control law is as follows:

[0158]

[0159] The corresponding discrete implementation is:

[0160]

[0161] Among them, e(t) or e(k) is the speed or frequency deviation, that is, the difference between the reference speed and the actual speed; u(t) or u PID (k) is the control output, corresponding to the guide vane opening control signal; T s is the control period; K p ,K i ,K d are PID controller parameters, representing proportional, integral, and differential parameters, respectively. Correspondingly, K p (k), K i (k), K d (k) is the time-varying PID control parameter.

[0162] S42, according to the global deep state feature vector GSF(k) and the global state fusion confidence C global (k) Dynamically adjust PID controller parameters:

[0163] K p (k) = K p0 α p (GSF(k),C global (k))

[0164] Ki (k) = K i0 α i (GSF(k),C global (k))

[0165] K d (k) = K d0 α d (GSF(k),C global (k))

[0166] Among them, K p0 , K i0 and K d0 is the benchmark PID parameter; α p , α i and α d is the adjustment factor, ranging from [0.5, 2.0], calculated by a small feedforward neural network: [α p ,α i ,α d ]=NN PID (GSF(k),C global (k)); where the neural network NN PID The design is a three-layer feedforward network, the input layer receives the concatenated GSF(k) vector and a single C global (k) value, the hidden layer uses the ReLU activation function, and the output layer uses the Sigmoid function and is scaled to the range of [0.5, 2.0]. PID Training goal: The optimized network should be able to distinguish more subtle working conditions, early imbalance states, etc. based on more refined local states GSF(k) and information C global (k) to output a more accurate PID adjustment factor to achieve better control performance (such as smaller overshoot, faster response, and stronger adaptability to working conditions).

[0167] Specifically, to avoid rapid fluctuations in parameters, a smooth update mechanism is introduced:

[0168] K p (k)=(1-λ(C global (k)))·K p (k-1)+λ(C global (k))·K p0 α p (k)

[0169] K i (k)=(1-λ(C global (k)))·K i (k-1)+λ(C global (k))·K i0 α i (k)

[0170] K d (k)=(1-λ(C global (k)))·K d (k-1)+λ(C global (k))·K d0 α d (k)

[0171] Among them, λ is the smoothing coefficient, which is a global (k), such as λ(C global (k))=λ base ·C global (k), λ base is the basic smoothing coefficient, such as 0.1-0.3. global When (k) is low, λ is automatically reduced, and the parameter update is more conservative, giving priority to ensuring stability; when C global When (k) is higher, λ increases and the parameter adjustment is more sensitive.

[0172] This embodiment also designs anti-integral saturation and introduces a conditional integral mechanism. When the control output approaches saturation, the integral effect is automatically reduced: Among them, σ is the Sigmoid function, u threshold is the saturation threshold.

[0173] S43. Based on the key performance indicators PF(k+1,...,k+M) of the next M time steps and the global deep state feature vector GSF(k), calculate the feedforward compensation:

[0174] u ff (k) = f ffc (PF(k+1,...,k+M),GSF(k))

[0175] Among them, f ffc is the feedforward compensation calculation function; the goal of this function is to calculate the optimal feedforward control action based on the serialized prediction of future disturbance trends and the current deep state GSF(k) of the unit (used to judge the unit's tolerance and response characteristics to disturbances).

[0176] In a specific example, f ffc The following situations are mainly considered:

[0177] Predicted speed deviation compensation: is the predicted future speed deviation, γ j is the time attenuation weight coefficient, satisfying And γ1>γ2>…>γ M .

[0178] Compensation for predicted load changes: It is the power prediction value of the most recent step. This compensation mainly responds to rapid load changes.

[0179] Predicted hydraulic disturbance compensation: u ff,H (k) = K ff,H ΔH pred (k+j)| j＝1 , ΔH pred (k+j)| j＝1 It is the predicted head change, used to deal with the disturbance caused by head fluctuation.

[0180] The actual application ratio of feedforward compensation is related to the global state fusion confidence:

[0181] β ff (k) = β ff,max ·(C global (k)) γ

[0182] Among them, β ff,max is the maximum feedforward compensation ratio, γ is the adjustment coefficient, when the prediction confidence is low, β ff (k) is reduced, which reduces the proportion of feedforward compensation in the total control amount.

[0183] S44, combine the PID controller output and the feedforward compensation to obtain the final speed control instruction: u total (k) = (1-β ff (k))·u PID (k)+β ff (k) u ff (k).

[0184] Specifically, a dynamic limiting mechanism is introduced: ff (k)|≤u ff,max ·g(C global (k),Confidence PF )

[0185] Among them, u ff,max is the maximum allowable feedforward compensation, g(·) is a function that comprehensively considers the global state fusion confidence C global (k) and the confidence of the prediction sequence PF itself (if the high-level fusion can output a confidence assessment of the prediction sequence, such as the expected variance of the prediction error). If the prediction confidence of the high-level fusion is high and the global state is credible, a larger feedforward compensation is allowed; otherwise, the feedforward effect is limited to prevent incorrect predictions from causing control degradation.

[0186] Based on the global deep state feature vector GSF(k) and the global state fusion confidence C global (k),Dynamically adjust the preference of control strategy.

[0187] Mode definition: For example, define efficiency priority mode, stability priority mode, quick response mode, etc.

[0188] The objective function is:

[0189] J efficiency =w1·J freq +w2·J stability +w3·J eff

[0190] In the stability priority mode, the control objective function is: J stability =w1′·J freq +w2′·J stability +w3′·J eff

[0191] Among them, J freq 、J stability and J eff They are frequency control performance, system stability and efficiency indicators respectively; the weight coefficient w i and w i ′ reflects preferences under different modes.

[0192] In one example, the adjustment logic in different modes can be designed as a function or rule set based on GSF(k). For example, when GSF(k) indicates that the unit is in a certain inefficient operation area but is still stable, J can be appropriately increased. eff When GSF(k) indicates that the unit is close to the unstable boundary, significantly increase J stability The weight of .

[0193] Strategy selection decision function:

[0194]

[0195] Among them, J s is the comprehensive evaluation function under strategy s, whose calculation depends on the current deep state and information confidence. When the confidence is high, it is more inclined to execute the optimal strategy based on the refined judgment of GSF(k); when the confidence is low, it may revert to a more conservative or general strategy.

[0196] Specifically, in one embodiment of the present invention, the method further includes:

[0197] Real-time monitoring of safety-related parameters during the operation of hydropower units;

[0198] Determine whether safety-related parameters exceed preset safety boundaries;

[0199] When it is detected that any safety-related parameter exceeds its corresponding preset safety boundary, a preset protection strategy is automatically triggered, and the protection strategy includes at least one of an alarm prompt, a load adjustment rate limit, and an emergency shutdown.

[0200] In a specific example, a multi-level safety protection mechanism was designed to ensure the safe operation of a hydropower unit:

[0201] Define a set of security-related parameters SP = sp1, sp2, ..., sp L , including guide vane opening, turbine surge parameters, bearing temperature, vibration amplitude, etc., to monitor in real time whether these parameters are close to the preset safety boundaries:

[0202]

[0203] Among them, S threshold,l is the safety threshold of the lth safety-related parameter, S range,l Its normal range.

[0204] When d safety,l When it is close to 1 (that is, close to the safety boundary), the safety constraint is triggered:

[0205] u constrained (t)=clip(u final (t),u min (t),u max (t))

[0206] Among them, u min (t) and u max (t) is the dynamically adjusted control output limit:

[0207]

[0208]

[0209] Δu safe is the safety change step, according to the safety distance d safety,l Dynamic Adjustment:

[0210]

[0211] Where Δu normal is the normal change step size, σ is the Sigmoid function, and k is the sensitivity parameter.

[0212] Emergency protection strategy: When a serious anomaly is detected (such as C global (k) Very low or d safety,l is negative), it automatically switches to protection mode: reduces control gain, increases stability margin; limits guide vane opening change rate; triggers emergency shutdown procedure when necessary.

[0213] Specifically, in one embodiment of the present invention, step S5 includes:

[0214] Convert digital control instructions into analog or digital signals that can be recognized by the actuator and transmit them to the execution unit via an industrial communication network (such as PROFIBUS-DP or OPC UA).

[0215] The control signal drives the main pressure regulating valve in the hydraulic system, generating the corresponding oil flow and direction, pushing the guide vane servo cylinder to move, changing the turbine guide vane opening, thereby adjusting the flow through the turbine and achieving speed control.

[0216] The actual guide vane opening is monitored in real time by a position sensor installed on the guide vane opening mechanism, and this information is fed back to the control system to form a closed-loop control.

[0217] When abnormal operating conditions are detected, such as excessive vibration, excessive bearing temperature or abnormal water flow, corresponding safety protection measures are implemented, such as limiting the guide vane opening change rate, locking the guide vane position or executing the emergency shutdown procedure.

[0218] The system monitors speed response, power output and stability indicators in real time, evaluates the speed regulation effect, and provides a basis for the next round of control optimization.

[0219] On the other hand, Figure 3 As shown, the present invention also provides an intelligent speed control system for hydropower units with multi-source data fusion, which is used to implement any of the above methods, characterized in that the system includes:

[0220] Data acquisition module, used to collect multi-source heterogeneous data during the operation of hydropower units, including mechanical quantities, hydraulic parameters and electrical quantities;

[0221] The data preprocessing module is used to perform standardized preprocessing on multi-source heterogeneous data to obtain preprocessed data;

[0222] The multi-layer data fusion module is used to extract the global deep state characteristic information and operation trend prediction information of the hydropower unit based on the pre-processed data through the multi-layer data fusion algorithm, and obtain the fused speed regulation decision basis;

[0223] A speed control instruction generation module is used to generate a speed control instruction for adjusting the speed of the hydropower unit based on the integrated speed control decision basis;

[0224] The speed control execution module is used to execute speed control on the hydropower unit according to the speed control instruction.

[0225] In this embodiment, the speed control system synchronously collects heterogeneous data in the three major fields of mechanics, hydraulics, and electricity through a high-precision sensor network, and adopts adaptive filtering and data consistency verification technology to ensure data quality. Its key innovation lies in the use of a three-layer progressive data fusion architecture, including bottom-level feature extraction and quality assessment, middle-level subsystem state estimation, and high-level global deep fusion, which can effectively extract deep state features hidden in multi-source data. In terms of control decision-making, the system combines parameter dynamic adaptive PID with feedforward compensation strategy, dynamically adjusts control parameters according to the fusion results, and predictively responds to system disturbances. The speed control execution module adopts a multi-level closed-loop control structure, combined with a safety protection mechanism based on fuzzy reasoning, to ensure the execution accuracy while ensuring the safety of the unit operation. The entire system is continuously optimized through closed-loop feedback to achieve dynamic response performance and steady-state control accuracy far exceeding that of traditional speed regulators.

[0226] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent speed control method for hydropower units based on multi-source data fusion, characterized in that: include: S1. Collect multi-source heterogeneous data during the operation of hydropower units, including mechanical quantities, hydraulic parameters, and electrical quantities; S2. Perform standardized preprocessing on multi-source heterogeneous data to obtain preprocessed data; S3. Based on the pre-processed data, a multi-layer data fusion algorithm is used to extract the global deep state characteristic information and operation trend prediction information of the hydropower unit, and the fused speed regulation decision basis is obtained; S4. Based on the integrated speed regulation decision basis, generate a speed regulation control instruction for regulating the speed of the hydropower unit; S5. Execute speed control on the hydropower unit according to the speed control instruction.

2. The intelligent speed control method for hydropower units based on multi-source data fusion according to claim 1 is characterized in that: In step S2, the standardization preprocessing includes: at least one operation of data clock synchronization, data resampling, data denoising, data filtering, data normalization and data integrity check to obtain the preprocessed data.

3. The intelligent speed control method for hydropower units based on multi-source data fusion according to claim 1 is characterized in that: In step S3, the multi-layer data fusion algorithm includes: Low-level feature extraction and quality assessment: extracting engineering features from preprocessed data and calculating their dynamic quality factors; Mid-level subsystem state estimation: This method estimates the state of key subsystems of the hydropower unit based on the underlying engineering characteristics and dynamic quality factors, and obtains the state estimation value of each subsystem and the global state portrait; High-level global deep fusion uses the attention mechanism to extract the global deep state feature information of the hydropower unit from the global state portrait, and predicts future key performance indicators as operation trend prediction information, and then comprehensively forms the fused speed regulation decision basis.

4. The intelligent speed control method for hydropower units based on multi-source data fusion according to claim 3 is characterized in that: In the underlying feature extraction and quality assessment: Extract the time domain features, frequency domain features and statistical features of the preprocessed data as engineering features F x (t); Calculate the dynamic quality factor Q for each engineering feature at each sampling time t x (t), which is calculated as: Q x (t)=w1·Q x,1 (t)+w2·Q x,2 (t)+w3·Q x,3 (t) Among them, Q x,1 (t) is the quality assessment factor based on characteristic stability, Q x,2 (t) is the quality assessment factor based on the characteristic signal-to-noise ratio, Q x,3 (t) is the quality assessment factor based on mutation detection, w1, w2, w3 are the weights of each dimension, and w1+w2+w3=1.

5. The intelligent speed control method for hydropower units based on multi-source data fusion according to claim 3 is characterized in that: In the state estimation of the middle-level subsystem: Divide the hydropower unit into several key subsystems, including the governor hydraulic servo system, turbine water diversion and hydraulic disturbance system, generator and its excitation system; Build a nonlinear state-space model for each subsystem: x j,k =f j (x j,k-1 ,u j,k-1 )+w j,k-1 z j,k =h j (x j,k )+v j,k Among them, x j,k is the state vector of subsystem j at time k, u j,k-1 is the control input, z j,k is the characteristic observation vector associated with subsystem j, f j (·) is the nonlinear state transfer function, h j (·) is the nonlinear observation function, w j,k-1 and v j,k are process noise and measurement noise, respectively; The state estimation is performed using an extended Kalman filter based on dynamic weighted covariance, where the i-th diagonal element of the measurement noise covariance matrix is dynamically adjusted as follows: Among them, R base,x,i is the reference measurement noise covariance, α is the adjustment factor, Q x,i (k) is the dynamic quality factor of the corresponding feature at time k, and ∈ is a small positive constant to prevent the denominator from being zero.

6. The intelligent speed control method for hydropower units based on multi-source data fusion according to claim 3 is characterized in that: In the high-level global deep fusion: Combine the outputs of the mid-level subsystem state estimates into a global state profile X seq (k); Use a deep neural network based on a multi-head self-attention mechanism to process the global state portrait X seq (k), extract context-aware sequence representation C seq (k); From context-aware sequence representation C seq (k) extracts the global deep state feature vector GSF(k) and predicts the key performance indicators PF(k+1,...,k+M) for the next M time steps; Calculate the global state fusion confidence C based on the attention weight distribution and prediction results global (k).

7. The intelligent speed control method for hydropower units based on multi-source data fusion according to claim 6 is characterized in that: Step S4 includes: S41. Construct a robust PID controller with dynamic adaptive parameter adjustment. The controller output is: Among them, e(t) is the speed or frequency deviation, u(t) is the control output, K p ,K i ,K d are the PID controller parameters; S42, according to the global deep state feature vector GSF(k) and the global state fusion confidence C global (k) Dynamically adjust PID controller parameters: K p (k)=K p0 ·α p (GSF(k),C global (k)) K i (k)=K i0 ·α i (GSF(k),C global (k)) K d (k)=K d0 ·α d (GSF(k),C global (k)) Among them, K p0 , K i0 and K d0 is the benchmark PID parameter; α p , α i and α d is the adjustment factor; S43. Calculate the feedforward compensation based on the key performance indicators PF(k+1,...,k+M) of the next M time steps and the global deep state feature vector GSF(k): u ff (k)=f ffc (PF(k+1,...,k+M),GSF(k)) Among them, f ffc It is the feedforward compensation calculation function; S44. Combine the PID controller output and the feedforward compensation amount to obtain the final speed control instruction.

8. The intelligent speed control method for hydropower units based on multi-source data fusion according to claim 7 is characterized in that: In step S42, α p , α i and α d Computed via a small feedforward neural network: [α p ,α i ,α d ]=NN PID (GSF(k),C global (k)); Among them, the neural network NN PID The design is a three-layer feedforward network, the input layer receives the concatenated GSF(k) vector and a single C global (k) value, the hidden layer uses the ReLU activation function, and the output layer uses the Sigmoid function and is scaled to the range of [0.5, 2.0].

9. The intelligent speed control method for hydropower units based on multi-source data fusion according to claim 1 is characterized in that: The method further comprises: Real-time monitoring of safety-related parameters during the operation of hydropower units; Determine whether safety-related parameters exceed preset safety boundaries; When it is detected that any safety-related parameter exceeds its corresponding preset safety boundary, a preset protection strategy is automatically triggered, and the protection strategy includes at least one of an alarm prompt, a load adjustment rate limit, and an emergency shutdown.

10. An intelligent speed control system for a hydropower unit with multi-source data fusion, used to implement the method according to any one of claims 1 to 9, characterized in that: The system comprises: The data acquisition module is used to collect multi-source heterogeneous data during the operation of the hydropower unit, including mechanical quantities, hydraulic parameters and electrical quantities; The data preprocessing module is used to perform standardized preprocessing on multi-source heterogeneous data to obtain preprocessed data; The multi-layer data fusion module is used to extract the global deep state characteristic information and operation trend prediction information of the hydropower unit based on the pre-processed data through the multi-layer data fusion algorithm, and obtain the fused speed regulation decision basis; A speed control instruction generation module is used to generate a speed control instruction for adjusting the speed of the hydropower unit based on the integrated speed control decision basis; The speed control execution module is used to execute speed control on the hydropower unit according to the speed control instruction.

Images