A water turbine speed regulation adaptive control system based on multi-source data

Abstract

The present application relates to the technical field of intelligent control of hydroelectric power generation, in particular to a water turbine speed regulation adaptive control system based on multi-source data, comprising: data acquisition, feature extraction, coupling relationship analysis, water head prediction, state fusion and instruction generation module. The system collects multiple types of real-time monitoring data of the water turbine, extracts the unit operation stability features by using the harmonic analysis algorithm adaptive to the tracking optimization of the water turbine frequency and its multiple frequency, constructs the hydraulic-mechanical dynamic coupling feature matrix and generates the short-term water head prediction sequence. The multiple features and the prediction sequence are fused to form a comprehensive control state representation, and the speed regulation control instruction is output by the deep reinforcement learning control strategy generator. The system can accurately identify the unit vibration state, match the hydraulic-mechanical coupling change law, adapt to the water head fluctuation trend, make the speed regulation instruction more suitable for the unit real-time working condition, and improve the dynamic adaptation effect of speed regulation.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for hydropower generation, and in particular to an adaptive control system for turbine speed regulation based on multi-source data. Background Technology

[0002] Existing turbine speed control systems mostly rely on single operational monitoring data for regulation and control. Vibration spectrum analysis uses traditional fixed harmonic analysis methods. Hydraulic and mechanical operating parameters are only subjected to simple correlation processing. Head changes rely on real-time monitoring data to directly participate in control. Speed ​​control strategies mostly use conventional proportional-integral-derivative control logic and do not introduce intelligent learning algorithms to achieve strategy optimization.

[0003] Traditional harmonic analysis cannot achieve dynamic tracking based on the actual operating frequency and harmonics of the turbine, resulting in deviations in the extracted unit operation stability characteristics. The coupling relationship analysis of parameters such as volute pressure, guide vane opening, and active power lacks dynamism and cannot truly reflect the hydraulic-mechanical coupling characteristics. Upstream reservoir water level changes are only collected in real time without short-term trend prediction. The control strategy is fixed and singular, unable to adapt to the complex operating conditions of the turbine. The dynamic adaptability of speed regulation is insufficient, making it difficult to accurately control the vibration stability of the unit operation. The matching degree between speed regulation commands and actual operating conditions is low.

[0004] To address the issue that turbine vibration spectrum analysis cannot adaptively track rotational and harmonic frequencies, harmonic analysis algorithms need to be optimized to extract accurate operational stability features. To address the problems of lag in hydraulic-mechanical coupling feature analysis, missing head prediction, and rigid control strategies, a dynamic coupling feature matrix needs to be constructed. Combined with short-term head prediction sequences, this matrix should be used to generate speed control commands adapted to real-time operating conditions using deep reinforcement learning. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and propose a turbine speed regulation adaptive control system based on multi-source data.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a turbine speed regulation adaptive control system based on multi-source data, comprising:

[0007] The data acquisition module collects a multi-source real-time monitoring data set of the turbine, which includes unit vibration spectrum data, spiral casing inlet pressure time series data, guide vane opening time series data, active power output time series data, and upstream reservoir water level change data.

[0008] The feature extraction module performs harmonic analysis algorithm processing on the vibration spectrum data of the unit to extract feature indicators characterizing the unit's operating stability. The harmonic analysis algorithm optimizes traditional harmonic analysis based on the adaptive tracking of the turbine's rotational frequency and its harmonics.

[0009] The coupling relationship module performs multivariate coupling relationship analysis on the time series data of the volute inlet pressure, the time series data of the guide vane opening, and the time series data of the active power output, and generates a dynamic coupling feature matrix that reflects the hydraulic-mechanical coupling relationship of the turbine.

[0010] The head prediction module performs trend prediction on the upstream reservoir water level change data and generates a short-term head prediction sequence.

[0011] The state fusion module inputs the characteristic indicators representing the unit's operational stability, the dynamic coupling characteristic matrix, and the short-term head prediction sequence into the adaptive control model to generate a comprehensive control state representation.

[0012] The instruction generation module calls a control strategy generator based on deep reinforcement learning to process the comprehensive control state representation and outputs an adaptive speed regulation control instruction sequence for the turbine.

[0013] As a further aspect of the present invention, the vibration spectrum data of the unit is processed by a harmonic analysis algorithm to extract characteristic indicators representing the operational stability of the unit, including:

[0014] The raw unit vibration spectrum data is preprocessed, including filtering to remove high-frequency noise and baseline drift correction;

[0015] By applying the aforementioned harmonic analysis algorithm, within a preset fundamental frequency search range, the main operating frequency of the current unit is adaptively determined by calculating the spectral kurtosis.

[0016] Based on the adaptively determined main frequency, calculate the harmonic components within its preset multiples, and simultaneously calculate the energy ratio of each harmonic component relative to the fundamental frequency component to form a harmonic energy distribution vector.

[0017] Identify the abnormal harmonic frequency components in the harmonic energy distribution vector that exceed the normal operating threshold, and combine the frequency, energy ratio, and phase relationship with adjacent components of each abnormal harmonic frequency component to construct the characteristic index that characterizes the unit's operating stability.

[0018] As a further aspect of the present invention, multivariate coupling analysis is performed on the time-series data of the volute inlet pressure, the guide vane opening, and the active power output to generate a dynamic coupling feature matrix reflecting the hydraulic-mechanical coupling relationship of the turbine, including:

[0019] The timing data of the volute inlet pressure, guide vane opening, and active power output are time-aligned and resampled to ensure that the three timing data have the same sampling time and frequency.

[0020] Within the sliding time window, the cross-correlation function between the volute inlet pressure time series data and the guide vane opening time series data is calculated, and the maximum cross-correlation coefficient and the corresponding time delay are extracted as the first coupling feature pair.

[0021] Within the same sliding time window, the transfer function estimates of the guide vane opening timing data and the active power output timing data are calculated, and the gain and phase of the main frequency channels are extracted as the second coupling feature pair.

[0022] Within the same sliding time window, the strength of the nonlinear causal relationship between the volute inlet pressure time series data and the active power output time series data is analyzed as a third coupling feature.

[0023] The first coupling feature pair, the second coupling feature pair, and the third coupling feature obtained from different time windows are arranged in chronological order to form the dynamic coupling feature matrix.

[0024] As a further aspect of the present invention, trend prediction is performed on the upstream reservoir water level change data to generate a short-term head prediction sequence, including:

[0025] Differential processing is performed on historical and current upstream reservoir water level change data to obtain a water level change rate sequence;

[0026] An autoregressive moving average model incorporating periodic components is established, and the parameters of the autoregressive moving average model are trained using the water level change rate sequence.

[0027] The trained autoregressive moving average model is used to make multiple predictions in the future to obtain the predicted water level change rate at multiple consecutive time points.

[0028] The predicted water level change rate at multiple consecutive future times is integrated and combined with the current water level to obtain the predicted water level at the corresponding future time.

[0029] Subtracting the preset tailwater level from the predicted water level at future times yields the predicted water head values ​​for multiple consecutive future times, forming the short-term water head prediction sequence.

[0030] As a further aspect of the present invention, the characteristic indicators representing the unit's operational stability, the dynamic coupling characteristic matrix, and the short-term head prediction sequence are input into the adaptive control model to generate a comprehensive control state representation, including:

[0031] The adaptive control model includes a stability feature encoder, a coupling feature encoder, and a head prediction encoder.

[0032] The stability feature encoder receives the feature index characterizing the unit's operational stability and maps it into a stability state vector through a fully connected neural network.

[0033] The coupling feature encoder receives the dynamic coupling feature matrix, extracts its temporal dependency features through a temporal convolutional network, and outputs a coupling state vector.

[0034] The head prediction encoder receives the short-term head prediction sequence, encodes its changing trend through a recurrent neural network, and outputs a head state vector.

[0035] The stability state vector, coupling state vector, and head state vector are concatenated, and the concatenated vector is reweighted using a fusion attention layer to output the comprehensive control state representation.

[0036] As a further aspect of the present invention, the step of calling a control policy generator based on deep reinforcement learning to process the comprehensive control state representation and outputting an adaptive speed control command sequence for the turbine includes:

[0037] The control strategy generator includes a strategy network and a value evaluation network;

[0038] The integrated control state characterization is input into the strategy network, and the strategy network outputs the probability distribution of the adjustment amount and adjustment rate of the turbine guide vane target opening under the current state through multi-layer nonlinear transformation.

[0039] By sampling from the probability distribution, a specific guide vane target opening adjustment action and the corresponding adjustment rate are obtained;

[0040] The integrated control state characterization is combined with the sampled guide vane target opening adjustment action and adjustment rate, and then input into the value evaluation network to evaluate the long-term expected return corresponding to the action state of the current combination.

[0041] Based on the long-term expected return, the parameters of the policy network are iteratively updated to make it tend to output actions with higher expected returns;

[0042] The trained policy network directly outputs a deterministic sequence of adaptive speed control commands for the turbine based on the input comprehensive control state representation.

[0043] As a further aspect of the present invention, the adaptive speed control command sequence of the turbine includes set values ​​for guide vane opening, blade angle, and governor control parameters, and its generation process includes:

[0044] The target adjustment amount and adjustment rate of the guide vane opening are extracted from the actions output by the strategy network.

[0045] Based on the current guide vane opening, a sequence of guide vane opening setpoints for multiple future control cycles is calculated.

[0046] If the turbine is a propeller type, then based on the head-output-guide vane opening coordination model, the corresponding blade angle setpoint sequence can be calculated from the guide vane opening setpoint and the current active power.

[0047] Based on the dynamic characteristics of the system reflected by the dynamic coupling feature matrix, the proportional gain and integral gain parameters of the speed governor are dynamically adjusted to form a sequence of speed governor parameter setpoints.

[0048] The guide vane opening setting value sequence, the blade angle setting value sequence, and the governor parameter setting value sequence are aligned according to the timestamps to form the adaptive speed control command sequence.

[0049] As a further aspect of the present invention, the step of calculating the cross-correlation function between the volute inlet pressure time series data and the guide vane opening time series data within the sliding time window, and extracting the maximum cross-correlation coefficient and the corresponding time delay as the first coupling feature pair, includes:

[0050] Set a fixed-length sliding time window and extract the corresponding data segment within the current window from the aligned volute inlet pressure time series data and guide vane opening time series data;

[0051] The extracted volute inlet pressure data segment and guide vane opening data segment were preprocessed to zero mean.

[0052] Calculate the sequence of cross-correlation function values ​​of the volute inlet pressure data segment and the guide vane opening data segment under different time delays after zero-mean normalization;

[0053] Within a preset physically feasible time delay range, search for the maximum absolute value of the cross-correlation function value sequence, and record the maximum absolute value as the maximum cross-correlation number of the current time window;

[0054] Obtain the time delay parameter corresponding to the maximum absolute value, and record the time delay parameter as the characteristic time delay of the current time window;

[0055] The maximum cross-correlation coefficient is combined with the feature time delay to form the first coupled feature pair.

[0056] As a further aspect of the present invention, the system further includes an instruction conversion module, used for:

[0057] The adaptive speed control command sequence is converted into the underlying servo drive command of the turbine speed control system, specifically including:

[0058] Analyze the guide vane opening setpoint sequence in the adaptive speed control command sequence;

[0059] Based on the displacement-opening calibration curve of the servo mechanism of the speed control system, the guide vane opening setpoint sequence is converted into the target displacement sequence of the servo cylinder.

[0060] Based on the performance parameters of the servo driver, the target displacement sequence is decomposed into pulse control commands for motor speed, direction of rotation, and running time;

[0061] The pulse control command is encapsulated according to the speed regulation system communication protocol, and a timestamp and command check code are added to form the underlying servo drive command;

[0062] The underlying servo drive command is sent to the actuator of the speed control system.

[0063] As a further aspect of the present invention, the step of decomposing the target displacement sequence into pulse control commands for motor speed, direction, and running time based on the performance parameters of the servo driver includes:

[0064] Based on the difference in target displacement between adjacent time points, the required displacement change of the servo cylinder in each control cycle is calculated.

[0065] Based on the lead screw pitch of the servo cylinder, the displacement change is converted into the number of revolutions required by the servo motor.

[0066] Calculate the average speed required by the servo motor based on the duration of the control cycle;

[0067] Based on the electronic gear ratio of the servo drive and the encoder resolution, the required number of rotations and the average speed are converted into the corresponding number of pulses and pulse frequency;

[0068] The direction of the motor is determined by the sign of the displacement change, and a corresponding direction control level signal is generated.

[0069] The pulse quantity, pulse frequency, and direction control level signal together constitute the pulse control command.

[0070] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0071] Based on the adaptive tracking of turbine rotation frequency and its harmonics, traditional harmonic analysis is optimized. The harmonic analysis algorithm is applied to the unit vibration spectrum data, which can dynamically follow the changes in rotation frequency and harmonics during turbine operation, accurately identify harmonic components in the vibration spectrum related to the unit's operational stability, filter out invalid interference information in the spectrum data, and extract feature indicators that can truly reflect the stable state of the unit's operation. It avoids the feature extraction bias caused by the fixed frequency band analysis of traditional harmonic analysis, clearly presents the subtle changes in the unit's vibration state, and the feature indicators have a higher degree of fit with the actual operating state of the unit.

[0072] Multivariate coupling relationship analysis is performed on the time-series data of spiral casing inlet pressure, guide vane opening, and active power to form a dynamic coupling feature matrix. Short-term head prediction sequence is obtained by predicting the trend of upstream reservoir water level changes. The above features and sequences are fused and input into the adaptive control model. The speed control command sequence is processed by the control strategy generator through deep reinforcement learning. It can match the dynamic change law of hydraulic-mechanical coupling relationship in real time and adapt to the short-term change trend of upstream head. The control strategy can be dynamically adjusted with the comprehensive operating status of the turbine. The speed control command is more adaptable to the real-time operating conditions of the unit, avoiding the problem that fixed control strategies cannot adapt to changes in operating conditions. The dynamic response of speed regulation is more in line with the actual operating needs of the turbine. The fusion of coupling features and head prediction information makes the control state more comprehensive, and the generation of speed control commands is more in line with the unit operating status reflected by multi-source data. Attached Figure Description

[0073] Figure 1 This is a timing diagram of a turbine speed regulation adaptive control system based on multi-source data as described in this invention.

[0074] Figure 2 A flowchart for adaptive harmonic analysis of vibration spectrum and extraction of characteristic indicators of unit operation stability;

[0075] Figure 3 The flowchart illustrates the process of analyzing the multivariable coupling relationship between hydraulic and mechanical systems and generating the dynamic coupling feature matrix. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0077] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0078] See Figure 1 This invention provides a turbine speed regulation adaptive control system based on multi-source data, specifically comprising:

[0079] The data acquisition module continuously collects multi-source real-time monitoring data from the turbine, encompassing unit vibration spectrum data, spiral casing inlet pressure time-series data, guide vane opening time-series data, active power output time-series data, and upstream reservoir water level change data. The feature extraction module processes the unit vibration spectrum data using a harmonic analysis algorithm to extract characteristic indicators representing the unit's operational stability. This harmonic analysis algorithm optimizes traditional harmonic analysis based on adaptive tracking of the turbine's rotational frequency and its harmonics. The coupling relationship module performs multivariate coupling relationship analysis on the spiral casing inlet pressure time-series data, guide vane opening time-series data, and active power output time-series data, generating a dynamic coupling feature matrix reflecting the turbine's hydraulic-mechanical coupling relationship. The head prediction module performs trend prediction on the upstream reservoir water level change data, generating a short-term head prediction sequence. The state fusion module inputs the characteristic indicators representing the unit's operational stability, the dynamic coupling feature matrix, and the short-term head prediction sequence into the adaptive control model to generate a comprehensive control state representation. The instruction generation module calls a control strategy generator based on deep reinforcement learning to process the comprehensive control state representation and outputs an adaptive speed control instruction sequence for the turbine.

[0080] In one embodiment of the present invention, see [reference] Figure 2 The original unit vibration spectrum data is preprocessed, including filtering to remove high-frequency noise and baseline drift correction. A harmonic analysis algorithm is applied to adaptively determine the current operating frequency of the unit within a preset fundamental frequency search range by calculating spectral kurtosis. Using the adaptively determined primary frequency as a reference, the harmonic components within preset multiples are calculated, and the energy ratio of each harmonic component relative to the fundamental frequency component is simultaneously calculated, forming a harmonic energy distribution vector. Abnormal harmonic components exceeding the normal operating threshold are identified in the harmonic energy distribution vector. The frequency, energy ratio, and phase relationship with adjacent components of each abnormal harmonic component are combined to construct a characteristic index characterizing the unit's operational stability.

[0081] In specific implementations, preprocessing is performed on the original unit vibration spectrum data. This preprocessing includes two steps: filtering to remove high-frequency noise and baseline drift correction. High-frequency noise removal is achieved using a low-pass filter with a cutoff frequency higher than a predetermined multiple of the unit's rated operating frequency. Baseline drift correction is performed by calculating the moving average of the vibration signal and subtracting this average from the original signal, making the preprocessed spectrum data more prominent in terms of periodic components related to unit operation. In some embodiments, when applying the harmonic analysis algorithm, the preset fundamental frequency search range is set based on a combination of the turbine runner blade passage frequency and the inherent frequency range of the unit's shaft system. In specific implementations, the main operating frequency of the current unit is adaptively determined by calculating spectral kurtosis. The principle is that the main operating frequency of the unit during actual operation will form a significant peak cluster in the spectrum, exhibiting a high kurtosis value. The algorithm traverses the candidate frequencies within the fundamental frequency search range, calculates the spectral kurtosis in a small neighborhood centered on each candidate frequency, and determines the candidate frequency that achieves the maximum kurtosis as the main operating frequency of the current unit.

[0082] In practical implementation, the adaptively determined main frequency is used as the benchmark, and the harmonic components within its preset multiples are calculated. These preset multiples are determined based on the vibration modes of the unit's mechanical structure and the potential upper frequency limit of the hydraulic excitation force. In practical implementation, the energy ratio of each harmonic component relative to the fundamental frequency component is calculated simultaneously. The fundamental frequency component's energy is the square of the amplitude at the corresponding frequency point of the main frequency, and the energy of a certain harmonic component is the square of the amplitude at the corresponding frequency point. The energy ratio calculation formula is as follows:

[0083]

[0084] in: Indicates the first The energy ratio of the second harmonic component to the fundamental frequency component. Representing the The energy of the first harmonic component, This represents the energy of the fundamental frequency component. In practice, the energy ratios of each harmonic are arranged in order of order to form a harmonic energy distribution vector, the dimension of which is equal to the harmonic order being considered.

[0085] In practical implementation, abnormal harmonic components exceeding the normal operating threshold in the harmonic energy distribution vector are identified. The normal operating threshold is set based on the percentile of the harmonic energy distribution obtained through long-term statistical analysis of the unit under steady-state conditions. In practice, the frequency, energy ratio, and phase relationship with adjacent components of each abnormal harmonic component are combined to construct a characteristic index characterizing the unit's operational stability. The frequency of the abnormal harmonic component can be directly obtained from the spectral data, the energy ratio comes from the harmonic energy distribution vector, and the phase relationship with adjacent components is determined by calculating the phase difference between the abnormal harmonic component and its preceding and following harmonic components in the complex spectrum.

[0086] In one embodiment of the present invention, see [reference] Figure 3 Time alignment and resampling were performed on the time-series data of volute inlet pressure, guide vane opening, and active power output to ensure that the three time-series data had the same sampling time and frequency. Within a sliding time window, the cross-correlation function of the volute inlet pressure time-series data and the guide vane opening time-series data was calculated, and the maximum cross-correlation coefficient and corresponding time delay were extracted as the first coupling feature pair. Within the same sliding time window, the transfer function estimation of the guide vane opening time-series data and the active power output time-series data was calculated, and the gain and phase of the main frequency channels were extracted as the second coupling feature pair. Within the same sliding time window, the strength of the nonlinear causal relationship from the volute inlet pressure time-series data to the active power output time-series data was analyzed as the third coupling feature. The first coupling feature pair, the second coupling feature pair, and the third coupling feature obtained from different time windows were arranged in chronological order to form a dynamic coupling feature matrix.

[0087] A fixed-length sliding time window is set, and the corresponding data segment within the current window is extracted from the aligned volute inlet pressure time series data and guide vane opening time series data. The extracted volute inlet pressure data segment and guide vane opening data segment are preprocessed to zero mean. The cross-correlation function value sequences of the zero-mean volute inlet pressure data segment and guide vane opening data segment under different time delays are calculated. Within a preset physically feasible time delay range, the maximum absolute value of the cross-correlation function value sequence is searched, and this maximum absolute value is recorded as the maximum cross-correlation coefficient of the current time window. The time delay parameter corresponding to this maximum absolute value is obtained, and this time delay parameter is recorded as the characteristic time delay of the current time window. The maximum cross-correlation coefficient and the characteristic time delay are combined to form the first coupling feature pair.

[0088] In practice, the time-series data of the volute inlet pressure, guide vane opening, and active power output are time-aligned and resampled to ensure that the three time-series data have the same sampling time and frequency. This process is completed through a unified timestamp index and a linear interpolation algorithm to eliminate phase errors caused by the acquisition delay of different sensors. In practice, time alignment is based on the clock source of the data acquisition module, correcting the timestamps of the volute inlet pressure, guide vane opening, and active power output time-series data to the same point in time. Subsequently, each set of data is resampled using a fixed new sampling interval, so that the three sets of data are completely synchronized on the time axis.

[0089] In some embodiments, the cross-correlation function of the volute inlet pressure time series data and the guide vane opening time series data is calculated within a sliding time window, and the maximum cross-correlation coefficient and corresponding time delay are extracted as the first coupling feature pair. In a specific implementation, a fixed-length sliding time window is set, and the corresponding data segment within the current window is extracted from the aligned volute inlet pressure time series data and guide vane opening time series data. The length of the sliding time window is set according to the main time constant of the turbine hydraulic transient process. In a specific implementation, the extracted volute inlet pressure data segment and guide vane opening data segment are respectively subjected to zero-mean preprocessing, that is, the arithmetic mean of the data segment is subtracted from each data segment to eliminate the interference of the DC component on the correlation calculation. In a specific implementation, the cross-correlation function value sequence of the zero-mean volute inlet pressure data segment and guide vane opening data segment under different time delays is calculated. The discrete calculation formula of the cross-correlation function is:

[0090]

[0091] in: Indicates time delay The cross-correlation function values ​​under the following conditions This is the zero-mean value of the volute inlet pressure data segment. This is the zero-mean normalized guide vane opening data segment. The number of data points within the sliding time window. The time delay is an integer. Within the preset physically feasible time delay range, the maximum absolute value of the cross-correlation function value sequence is searched, and the maximum absolute value is recorded as the maximum cross-correlation number of the current time window; the time delay parameter corresponding to the maximum absolute value is obtained, and the time delay parameter is recorded as the characteristic time delay of the current time window; the maximum cross-correlation number and the characteristic time delay are combined to form the first coupling feature pair.

[0092] Within the same sliding time window, the transfer function estimates of the guide vane opening time-series data and the active power output time-series data are calculated, and the gain and phase of the main frequency channels are extracted as the second coupling feature pair. Optionally, the transfer function estimation adopts frequency domain analysis techniques based on the Welch method. The frequency response function is obtained by calculating the cross power spectral density of the input signal (guide vane opening time-series data) and the output signal (active power output time-series data) and the auto-power spectral density of the input signal. In specific implementation, the main frequency channel refers to the frequency range located near the unit rotation frequency and the passing frequency of the turbine blades. Within this range, several frequency points with the largest amplitude are selected, and their amplitude ratio (gain) and phase angle (phase) in the frequency response function are recorded to form the second coupling feature pair.

[0093] Within the same sliding time window, the strength of the nonlinear causal relationship between the volute inlet pressure time series data and the active power output time series data is analyzed and used as the third coupling feature. It can be understood that the calculation of the nonlinear causal relationship strength is based on an extension of Granger causality, measuring the degree of nonlinear influence of volute inlet pressure on active power by comparing the prediction errors of a prediction model that includes historical information about volute inlet pressure with one that does not. In specific implementation, the first coupling feature pair, the second coupling feature pair, and the third coupling feature calculated in different time windows are arranged in chronological order to form a dynamic coupling feature matrix. Optionally, the rows of the dynamic coupling feature matrix correspond to different sliding time windows, and the columns correspond to the maximum cross-correlation coefficient and characteristic delay in the first coupling feature pair, the gain and phase in the second coupling feature pair, and the third coupling feature, thus forming a multidimensional time series feature tensor.

[0094] In one embodiment of the present invention, historical and current upstream reservoir water level change data are differentially processed to obtain a water level change rate sequence. An autoregressive moving average model incorporating periodic components is established, and the parameters of this model are trained using the water level change rate sequence. The trained autoregressive moving average model is then used to perform multi-step predictions to obtain predicted water level change rates for multiple consecutive future time periods. These predicted rates are then integrated and combined with the current water level to obtain the predicted water level for the corresponding future time period. Finally, a preset tailwater level is subtracted from the predicted water level for each future time period to obtain the predicted head for multiple consecutive future time periods, forming a short-term head prediction sequence.

[0095] The adaptive control model comprises a stability feature encoder, a coupling feature encoder, and a head predictive encoder. The stability feature encoder receives feature indices characterizing the unit's operational stability and maps them to a stability state vector using a fully connected neural network. The coupling feature encoder receives a dynamic coupling feature matrix, extracts its temporal dependency features using a temporal convolutional network, and outputs a coupling state vector. The head predictive encoder receives a short-term head prediction sequence, encodes its changing trend using a recurrent neural network, and outputs a head state vector. The stability state vector, coupling state vector, and head state vector are concatenated, and a fusion attention layer reweights the concatenated vector to output a comprehensive control state representation.

[0096] In practice, historical and current upstream reservoir water level change data are differentially processed to obtain a water level change rate sequence. The order and interval of the differential processing are set according to the seasonal pattern of reservoir water level changes and the scheduling cycle. Furthermore, an autoregressive moving average model incorporating periodic components is established. The parameters of the autoregressive moving average model are trained using the water level change rate sequence. The introduction of periodic components into the model aims to capture the recurring patterns in reservoir water levels caused by the influence of rainfall-runoff cycles and power generation scheduling cycles.

[0097] In practice, the trained autoregressive moving average model is used to perform multi-step predictions to obtain predicted water level change rates for multiple consecutive future time periods. The prediction step size is consistent with the length of the short-term head prediction sequence. In practice, the predicted water level change rates for multiple consecutive future time periods are integrated and combined with the current water level to obtain the predicted water level for the corresponding future time period. The integration operation uses an accumulation method, with the initial value being the current water level, and each step adding the product of the predicted water level change rate for the corresponding time period and the time interval. In practice, the predicted water level for future time periods is subtracted from the preset tailwater level to obtain the predicted head for multiple consecutive future time periods, forming the short-term head prediction sequence. The tailwater level is a fixed value or a dynamically queried value determined based on the water level-discharge relationship of the downstream river channel and the tailwater characteristics of the power station. Referring to Table 1, a structural diagram of a short-term head prediction sequence is given. The table lists the predicted head values ​​for five consecutive future time periods, which constitute the actual data carrier of the short-term head prediction sequence.

[0098] Table 1: Schematic diagram of the structure of the periodic head prediction sequence

[0099]

[0100] The adaptive control model comprises a stability feature encoder, a coupling feature encoder, and a head predictive encoder. In implementation, the stability feature encoder receives feature indices characterizing the unit's operational stability and maps them to a stability state vector using a fully connected neural network; the number of layers and neurons in the fully connected neural network is configured according to the dimension and complexity of the feature indices. In implementation, the coupling feature encoder receives a dynamically coupled feature matrix, extracts its temporal dependency features using a temporal convolutional network, and outputs a coupled state vector; the kernel size and dilation coefficients of the temporal convolutional network are set according to the temporal resolution of the dynamically coupled feature matrix and the temporal correlation between features.

[0101] In practical implementation, the head predictive encoder receives short-term head prediction sequences, encodes their changing trends using a recurrent neural network (RNN), and outputs a head state vector. The RNN can be a gated recurrent unit (GRU) or a long short-term memory (LSTM) network, with the hidden layer state dimension determined based on the length and magnitude of the short-term head prediction sequence. In practice, the stability state vector, coupled state vector, and head state vector are concatenated, and a fusion attention layer reweights the concatenated vector to output a comprehensive control state representation. The weight allocation mechanism of the fusion attention layer allows the model to automatically focus on state components more important for the current control decision.

[0102] In one embodiment of the present invention, the control strategy generator includes a strategy network and a value evaluation network. A comprehensive control state representation is input into the strategy network, which, through multi-layer nonlinear transformations, outputs a probability distribution of the adjustment amount and rate of the turbine guide vane target opening under the current state. A specific guide vane target opening adjustment action and its corresponding adjustment rate are sampled from this probability distribution. The comprehensive control state representation is combined with the sampled guide vane target opening adjustment action and adjustment rate, and input into the value evaluation network to evaluate the long-term expected return corresponding to the current combination of actions. Based on the long-term expected return, the parameters of the strategy network are iteratively updated to tend to output actions with higher expected returns. The trained strategy network, based on the input comprehensive control state representation, directly outputs a deterministic adaptive speed control command sequence for the turbine.

[0103] From the actions output by the strategy network, the target adjustment amount and adjustment rate of the guide vane opening are extracted. Combined with the current guide vane opening, a sequence of guide vane opening setpoints for multiple future control cycles is calculated. If the turbine is a propeller type, based on the head-output-guide vane opening coordination model, the corresponding blade angle setpoint sequence is calculated from the guide vane opening setpoint and the current active power. Based on the system dynamic characteristics reflected by the dynamic coupling feature matrix, the proportional gain and integral gain parameters of the governor are dynamically adjusted to form a governor parameter setpoint sequence. The guide vane opening setpoint sequence, blade angle setpoint sequence, and governor parameter setpoint sequence are aligned according to timestamps to form an adaptive speed control command sequence.

[0104] In practical implementation, the control strategy generator includes a strategy network and a value evaluation network. The strategy network adopts a multilayer perceptron structure, and the value evaluation network is also built based on a deep neural network. In practical implementation, the comprehensive control state representation is input into the strategy network. The strategy network, through multilayer nonlinear transformation, outputs the probability distribution of the adjustment amount and adjustment rate of the turbine guide vane target opening under the current state. This probability distribution is assumed to be a multivariate Gaussian distribution, and its mean and variance are calculated by the strategy network based on the input comprehensive control state representation.

[0105] In practice, a specific guide vane target opening adjustment action and its corresponding adjustment rate are obtained by sampling from a probability distribution. The sampling process is achieved by generating random numbers from a standard normal distribution and performing an affine transformation. In practice, the comprehensive control state representation is combined with the sampled guide vane target opening adjustment action and adjustment rate, and input into a value assessment network to evaluate the long-term expected return corresponding to the action state of the current combination. The output of the value assessment network is a scalar value representing the estimated value of the cumulative discounted reward that can be obtained after performing a given action from the current state.

[0106] In practical implementation, the parameters of the policy network are iteratively updated based on the long-term expected return, making it inclined to output actions with higher expected returns. The update process is based on the policy gradient method, adjusting the network weights by calculating the gradient of the expected return with respect to the policy network parameters. In practical implementation, the trained policy network directly outputs a deterministic adaptive speed control command sequence for the turbine based on the comprehensive control state representation input. At this time, the policy network no longer outputs a probability distribution, but directly outputs the deterministic action value that maximizes the expected return. Referring to Table 2, a structure of the adaptive speed control command sequence for the turbine is shown. The sequence contains various setpoints for the next five control cycles, which together constitute the complete control command sequence.

[0107] Table 2: Structure of Adaptive Speed ​​Control Command Sequence for Hydropower Turbines

[0108]

[0109] In practical implementation, the target adjustment amount and adjustment rate of the guide vane opening are extracted from the actions output by the strategy network. These two values ​​correspond to specific dimensions of the action vector. In practice, combined with the current guide vane opening, a sequence of guide vane opening setpoints for multiple future control cycles is calculated. The calculation process involves using the current guide vane opening as the starting value and sequentially adding the product of the adjustment amount and adjustment rate for each control cycle to obtain the setpoints for the entire sequence.

[0110] In some embodiments, if the turbine is a propeller type, the corresponding blade angle setpoint sequence is calculated based on the head-output-guide vane opening coordination model, using the guide vane opening setpoint and the current active power. The coordination model is a mapping function established based on the turbine characteristic curve, which can determine the optimal blade angle under given head, guide vane opening, and active power. In specific implementations, the proportional gain and integral gain parameters of the governor are dynamically adjusted according to the system dynamic characteristics reflected by the dynamic coupling feature matrix, forming a governor parameter setpoint sequence. The adjustment rules for proportional gain and integral gain are based on the system damping characteristics and response speed requirements. When the dynamic coupling feature matrix shows a large system coupling strength, the gain is appropriately reduced to improve stability. In specific implementations, the guide vane opening setpoint sequence, blade angle setpoint sequence, and governor parameter setpoint sequence are aligned according to timestamps to form an adaptive speed control command sequence. The timestamps are arranged incrementally according to the control cycle interval, with the current time as the reference, ensuring that all setpoints are completely synchronized in the time dimension. It is understandable that the probability distribution output by the policy network during the training phase reflects the tendency to explore different actions, while the deterministic instructions output during the deployment phase are a direct application of the learned optimal policy.

[0111] In one embodiment of the present invention, converting the adaptive speed control command sequence into the underlying servo drive command of the turbine speed control system specifically includes: parsing the guide vane opening setpoint sequence in the adaptive speed control command sequence; converting the guide vane opening setpoint sequence into the target displacement sequence of the servo cylinder based on the displacement-opening calibration curve of the servo mechanism of the speed control system; decomposing the target displacement sequence into pulse control commands for motor speed, direction, and running time based on the performance parameters of the servo driver; encapsulating the pulse control commands according to the speed control system communication protocol, adding timestamps and command check codes to form the underlying servo drive command; and sending the underlying servo drive command to the actuator of the speed control system.

[0112] Based on the difference in target displacement between adjacent time points, the required displacement change of the servo cylinder within each control cycle is calculated. Based on the servo cylinder's lead screw pitch, the displacement change is converted into the number of revolutions required by the servo motor. Based on the duration of the control cycle, the required average speed of the servo motor is calculated. According to the servo driver's electronic gear ratio and encoder resolution, the required number of revolutions and average speed are converted into the corresponding number of pulses and pulse frequency. The sign of the displacement change determines the motor's direction of rotation, generating a corresponding direction control level signal. The number of pulses, pulse frequency, and direction control level signal together constitute the pulse control command.

[0113] In practical implementation, the adaptive speed control command sequence is converted into the underlying servo drive command of the turbine speed control system. Specifically, this includes parsing the guide vane opening setpoint sequence from the adaptive speed control command sequence. This sequence is a list of values ​​extracted by field from the control command structure output by the command generation module. In practice, based on the displacement-opening calibration curve of the speed control system's servo mechanism, the guide vane opening setpoint sequence is converted into the target displacement sequence of the servo cylinder. The displacement-opening calibration curve is a mapping relationship obtained by measuring the servo cylinder piston position corresponding to different guide vane openings during the equipment commissioning phase and fitting a polynomial function. This mapping relationship is used to convert the percentage-form opening value into a displacement value in millimeters.

[0114] In practical implementation, based on the performance parameters of the servo drive, the target displacement sequence is decomposed into pulse control commands for motor speed, direction of rotation, and running time. The servo drive's performance parameters include the electronic gear ratio, encoder lines, and maximum pulse frequency; these parameters determine how physical displacement is converted into digital pulses executable by the drive. In practical implementation, the pulse control commands are encapsulated according to the speed control system's communication protocol, and timestamps and command check codes are added to form the underlying servo drive commands. The timestamp identifies the execution time of the command, and the command check code is generated using the CRC16 algorithm to ensure the integrity of data transmission. In practical implementation, the underlying servo drive commands are sent to the actuators of the speed control system. The sending process is completed through an industrial Ethernet or fieldbus communication interface. After receiving the commands, the servo drive in the actuator parses them and drives the motor to move.

[0115] Based on the difference in target displacement between adjacent moments, the required displacement change of the servo cylinder within each control cycle is calculated. This displacement change is a signed value; positive values ​​indicate the extension direction, and negative values ​​indicate the retraction direction. Based on the servo cylinder's lead screw pitch, the displacement change is converted into the number of rotations required by the servo motor. The lead screw pitch is an inherent parameter of the servo cylinder's mechanical structure, representing the distance the piston moves per rotation of the lead screw. Based on the duration of the control cycle, the required average speed of the servo motor is calculated. The control cycle duration is a fixed time interval of the system, and the speed calculation considers the influence of motor acceleration and deceleration times. Based on the servo driver's electronic gear ratio and encoder resolution, the required number of rotations and average speed are converted into the corresponding number of pulses and pulse frequency. The electronic gear ratio defines the ratio between the internal pulses of the driver and the motor's rotation angle, and the encoder resolution determines the number of feedback pulses per revolution. The formula for calculating the number of pulses is:

[0116]

[0117] in: Indicates the number of pulses. It represents the amount of displacement change. Indicates the lead screw pitch. The numerator representing the electronic gear ratio, The denominator represents the electronic gear ratio. This indicates the number of command pulses required per revolution of the servo drive. The pulse frequency is calculated based on the number of pulses and the control cycle duration. The direction of the motor is determined by the sign of the displacement change, generating a corresponding direction control level signal; a high level corresponds to positive displacement, and a low level corresponds to negative displacement. The number of pulses, pulse frequency, and direction control level signal together constitute the pulse control command.

[0118] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A water turbine speed regulation adaptive control system based on multi-source data, characterized in that, include: The data acquisition module collects a set of multi-source real-time monitoring data of the turbine. The set of multi-source real-time monitoring data includes unit vibration spectrum data, spiral casing inlet pressure time series data, guide vane opening time series data, active power output time series data, and upstream reservoir water level change data. The feature extraction module performs harmonic analysis algorithm processing on the vibration spectrum data of the unit to extract feature indicators characterizing the unit's operating stability. The harmonic analysis algorithm optimizes traditional harmonic analysis based on the adaptive tracking of the turbine's rotational frequency and its harmonics. The coupling relationship module performs multivariate coupling relationship analysis on the time series data of the volute inlet pressure, the time series data of the guide vane opening, and the time series data of the active power output, and generates a dynamic coupling feature matrix that reflects the hydraulic-mechanical coupling relationship of the turbine. The head prediction module performs trend prediction on the upstream reservoir water level change data and generates a short-term head prediction sequence. The state fusion module inputs the characteristic indicators representing the unit's operational stability, the dynamic coupling characteristic matrix, and the short-term head prediction sequence into the adaptive control model to generate a comprehensive control state representation. The instruction generation module calls a control strategy generator based on deep reinforcement learning to process the comprehensive control state representation and outputs an adaptive speed regulation control instruction sequence for the turbine.

2. The adaptive control system for hydro-turbine speed regulation based on multi-source data according to claim 1, characterized in that, The vibration spectrum data of the unit is processed by a harmonic analysis algorithm to extract characteristic indicators representing the operational stability of the unit, including: The raw unit vibration spectrum data is preprocessed, including filtering to remove high-frequency noise and baseline drift correction; By applying the aforementioned harmonic analysis algorithm, within a preset fundamental frequency search range, the main operating frequency of the current unit is adaptively determined by calculating the spectral kurtosis. Based on the adaptively determined main frequency, calculate the harmonic components within its preset multiples, and simultaneously calculate the energy ratio of each harmonic component relative to the fundamental frequency component to form a harmonic energy distribution vector. Identify the abnormal harmonic components in the harmonic energy distribution vector that exceed the normal operating threshold, and combine the frequency, energy ratio, and phase relationship with adjacent components of each abnormal harmonic component to construct the characteristic index that characterizes the unit's operational stability.

3. The adaptive control system for hydro-turbine speed regulation based on multi-source data according to claim 1, wherein, Multivariate coupling analysis was performed on the time-series data of the volute inlet pressure, guide vane opening, and active power output to generate a dynamic coupling characteristic matrix reflecting the hydraulic-mechanical coupling relationship of the turbine, including: The timing data of the volute inlet pressure, guide vane opening, and active power output are time-aligned and resampled to ensure that the three timing data have the same sampling time and frequency. Within the sliding time window, the cross-correlation function between the volute inlet pressure time series data and the guide vane opening time series data is calculated, and the maximum cross-correlation coefficient and the corresponding time delay are extracted as the first coupling feature pair. Within the same sliding time window, the transfer function estimates of the guide vane opening timing data and the active power output timing data are calculated, and the gain and phase of the main frequency channels are extracted as the second coupling feature pair. Within the same sliding time window, the strength of the nonlinear causal relationship between the volute inlet pressure time series data and the active power output time series data is analyzed as a third coupling feature. The first coupling feature pair, the second coupling feature pair, and the third coupling feature obtained from different time windows are arranged in chronological order to form the dynamic coupling feature matrix.

4. The adaptive control system for hydro-turbine speed regulation based on multi-source data according to claim 1, wherein, Perform trend prediction on the upstream reservoir water level change data to generate a short-term head prediction sequence, including: Differential processing is performed on historical and current upstream reservoir water level change data to obtain a water level change rate sequence; An autoregressive moving average model incorporating periodic components is established, and the parameters of the autoregressive moving average model are trained using the water level change rate sequence. The trained autoregressive moving average model is used to make multiple predictions in the future to obtain the predicted water level change rate at multiple consecutive time points. The predicted water level change rate at multiple consecutive future times is integrated and combined with the current water level to obtain the predicted water level at the corresponding future time. Subtracting the preset tailwater level from the predicted water level at future times yields the predicted water head values ​​for multiple consecutive future times, forming the short-term water head prediction sequence.

5. The adaptive speed control system for a water turbine based on multi-source data according to claim 1, characterized in that, The characteristic indicators representing the unit's operational stability, the dynamic coupling characteristic matrix, and the short-term head prediction sequence are input into the adaptive control model to generate a comprehensive control state representation, including: The adaptive control model includes a stability feature encoder, a coupling feature encoder, and a head prediction encoder. The stability feature encoder receives the feature index characterizing the unit's operational stability and maps it into a stability state vector through a fully connected neural network. The coupling feature encoder receives the dynamic coupling feature matrix, extracts its temporal dependency features through a temporal convolutional network, and outputs a coupling state vector. The head prediction encoder receives the short-term head prediction sequence, encodes its changing trend through a recurrent neural network, and outputs a head state vector. The stability state vector, coupling state vector, and head state vector are concatenated, and the concatenated vector is reweighted using a fusion attention layer to output the comprehensive control state representation.

6. The adaptive speed control system for a water turbine based on multi-source data according to claim 1, characterized in that, The process of calling a control policy generator based on deep reinforcement learning to process the comprehensive control state representation and outputting an adaptive speed control command sequence for the turbine includes: The control strategy generator includes a strategy network and a value evaluation network; The integrated control state characterization is input into the strategy network, and the strategy network outputs the probability distribution of the adjustment amount and adjustment rate of the turbine guide vane target opening under the current state through multi-layer nonlinear transformation. By sampling from the probability distribution, a specific guide vane target opening adjustment action and the corresponding adjustment rate are obtained; The integrated control state characterization is combined with the sampled guide vane target opening adjustment action and adjustment rate, and then input into the value evaluation network to evaluate the long-term expected return corresponding to the action state of the current combination. Based on the long-term expected return, the parameters of the policy network are iteratively updated to make it tend to output actions with higher expected returns; The trained policy network directly outputs a deterministic sequence of adaptive speed control commands for the turbine based on the input comprehensive control state representation.

7. The adaptive control system for hydro-turbine speed regulation based on multi-source data according to claim 6, wherein, The adaptive speed control command sequence of the turbine includes set values ​​for guide vane opening, blade angle, and governor control parameters, and its generation process includes: The target adjustment amount and adjustment rate of the guide vane opening are extracted from the actions output by the strategy network. Based on the current guide vane opening, a sequence of guide vane opening setpoints for multiple future control cycles is calculated. If the turbine is a propeller type, then based on the head-output-guide vane opening coordination model, the corresponding blade angle setpoint sequence can be calculated from the guide vane opening setpoint and the current active power. Based on the dynamic characteristics of the system reflected by the dynamic coupling feature matrix, the proportional gain and integral gain parameters of the speed governor are dynamically adjusted to form a sequence of speed governor parameter setpoints. The guide vane opening setting value sequence, the blade angle setting value sequence, and the governor parameter setting value sequence are aligned according to the timestamps to form the adaptive speed control command sequence.

8. The adaptive control system for hydro-turbine speed regulation based on multi-source data according to claim 3, wherein, Within the sliding time window, the cross-correlation function between the volute inlet pressure time series data and the guide vane opening time series data is calculated, and the maximum cross-correlation coefficient and corresponding time delay are extracted as the first coupling feature pair, including: Set a fixed-length sliding time window and extract the corresponding data segment within the current window from the aligned volute inlet pressure time series data and guide vane opening time series data; The extracted volute inlet pressure data segment and guide vane opening data segment were preprocessed to zero mean. Calculate the sequence of cross-correlation function values ​​of the volute inlet pressure data segment and the guide vane opening data segment under different time delays after zero-mean normalization; Within a preset physically feasible time delay range, search for the maximum absolute value of the cross-correlation function value sequence, and record the maximum absolute value as the maximum cross-correlation number of the current time window; Obtain the time delay parameter corresponding to the maximum absolute value, and record the time delay parameter as the characteristic time delay of the current time window; The maximum cross-correlation coefficient is combined with the feature time delay to form the first coupled feature pair.

9. The adaptive control system for hydro-turbine speed regulation based on multi-source data according to claim 1, wherein, The system also includes an instruction conversion module, used for: The adaptive speed control command sequence is converted into the underlying servo drive command of the turbine speed control system, specifically including: Analyze the guide vane opening setpoint sequence in the adaptive speed control command sequence; Based on the displacement-opening calibration curve of the servo mechanism of the speed control system, the guide vane opening setpoint sequence is converted into the target displacement sequence of the servo cylinder. Based on the performance parameters of the servo driver, the target displacement sequence is decomposed into pulse control commands for motor speed, direction of rotation, and running time; The pulse control command is encapsulated according to the speed regulation system communication protocol, and a timestamp and command check code are added to form the underlying servo drive command; The underlying servo drive command is sent to the actuator of the speed control system.

10. The adaptive control system for hydro-turbine speed regulation based on multi-source data according to claim 9, wherein, The step of decomposing the target displacement sequence into pulse control commands for motor speed, direction, and running time based on the performance parameters of the servo driver includes: Based on the difference in target displacement between adjacent time points, the required displacement change of the servo cylinder in each control cycle is calculated. Based on the lead screw pitch of the servo cylinder, the displacement change is converted into the number of revolutions required by the servo motor. Calculate the average speed required by the servo motor based on the duration of the control cycle; Based on the electronic gear ratio of the servo drive and the encoder resolution, the required number of rotations and the average speed are converted into the corresponding number of pulses and pulse frequency; The direction of the motor is determined by the sign of the displacement change, and a corresponding direction control level signal is generated. The pulse quantity, pulse frequency, and direction control level signal together constitute the pulse control command.

Images