A method and system for predicting and actively suppressing water hammer pressure

Abstract

This invention relates to the field of fluid pipeline system safety and intelligent control, providing a method and system for water hammer pressure prediction and active suppression. The method includes: acquiring a multidimensional digital signal stream, generating a window data matrix sequence through sliding window segmentation and dimensionless transformation; image-encoding into a spatiotemporal feature sequence, processing through a dual-branch spatiotemporal prediction network, outputting a predicted feature map sequence and decoding it into a predicted waveform matrix; extracting future risk indicators and morphological features, concatenating them with the multidimensional digital signal stream to form a predicted enhanced state vector, inputting it into a strategy network to generate valve action control parameters; decoding the target valve opening curve, obtaining a composite control signal through feedforward-feedback composite control; acquiring actual response data streams and encapsulating historical data, and updating the prediction network and strategy network online through hybrid priority filtering. This invention achieves forward-looking and adaptive water hammer prediction and active suppression, providing insight into the pressure evolution trend of pipeline systems and offering decision support for the safe operation of fluid pipeline networks.

Description

Technical Field

[0001] This invention relates to the field of fluid pipeline system safety and intelligent control technology, and in particular to a method and system for predicting and actively suppressing water hammer pressure. Background Technology

[0002] In fluid pipeline systems of critical sectors such as municipal water supply and energy and chemical industries, sudden opening and closing of valves or starting and stopping of water pumps often causes destructive water hammer, seriously threatening the safe operation of the pipeline system. Traditional protective devices, due to their fixed parameters, are difficult to adapt to complex and changing operating conditions. Therefore, developing an intelligent control technology that can predict and actively suppress water hammer has become an important issue that urgently needs to be addressed in the field of active safety protection for pipeline systems.

[0003] Chinese patent application CN113011636A discloses a method and apparatus for predicting water hammer parameters in urban hot water heating networks. The method includes: setting pipe information and equipment information in the network, and determining network parameters based on the pipe information and equipment information, wherein the network parameters are used to reflect the working status of each pipe and each piece of equipment in the network; determining the water hammer parameters of the network based on the network parameters, wherein the water hammer parameters are used to reflect the flow rate data and head data of the network; and visualizing the flow rate data and the head data to obtain display information of the flow rate data and the head data, so as to achieve the prediction of the water hammer parameters.

[0004] However, current technology still faces many challenges. In large fluid transmission networks, when encountering sudden situations such as emergency trips of downstream pump units, existing control systems lack the ability to predict complex water hammer waveforms. The valve closing actions performed to isolate the fault often rely on preset rates or simple hysteresis feedback control. This control strategy is not based on proactive avoidance of future pressure evolution and cannot achieve optimal suppression. This suboptimal control behavior can induce secondary pressure transients, which, combined with the main pressure wave caused by pump tripping, create extreme peak pressures far exceeding expectations. At this point, physical protection devices with fixed parameters, such as water hammer eliminators, serving as the final line of defense, often fail due to exceeding their design limits, potentially leading to pipeline joint rupture, valve and pump damage, or even pipeline bursts, resulting in significant economic losses and safety accidents. Summary of the Invention

[0005] To achieve the above objectives, this invention provides a method for predicting and actively suppressing water hammer pressure, the specific technical solution of which is as follows:

[0006] The multidimensional digital signal stream characterizing the dynamic response of the pipeline system is collected and sequentially segmented and transformed using a sliding window mechanism to generate a window data matrix sequence.

[0007] The window data matrix sequence is image-encoded and a spatiotemporal feature sequence is constructed. The spatiotemporal feature sequence is processed by a dual-branch spatiotemporal prediction network to output a prediction feature map sequence. The prediction feature map sequence is then decoded to obtain the prediction waveform matrix.

[0008] The future risk index vector and morphological feature vector are extracted from the predicted waveform matrix and concatenated with the multidimensional digital signal stream at the current decision moment to form a predicted enhanced state vector. The predicted enhanced state vector is then input into the reinforcement learning strategy network to generate valve action control parameters. At the same time, the real-time risk coefficient is calculated based on the future maximum pressure peak and the preset pressure threshold, and the reward weights of each sub-item are dynamically adjusted to generate an adaptive reward value.

[0009] The valve action control parameters are decoded into the target valve opening curve. A feedforward-feedback composite control architecture is used to obtain a composite control signal to drive the valve to execute. The actual response data stream is collected synchronously. The actual response data stream and historical experience data are encapsulated into a quadruple. The spatiotemporal prediction network and policy network are filtered and updated online through a hybrid priority mechanism to obtain the updated spatiotemporal prediction network and policy network.

[0010] Furthermore, the method for generating the window data matrix sequence includes:

[0011] Synchronously acquire multidimensional physical quantities from key measuring points of the pipeline system to construct a multidimensional digital signal stream;

[0012] A sliding window mechanism is used to serialize and segment the multidimensional digital signal stream according to the preset window length and sliding step size, and to construct a data segment matrix sequence.

[0013] To address the dimensional differences among multiple physical quantity channels in a data fragment matrix sequence, a linear mapping is used to make the data dimensionless, generating a window data matrix sequence with a unified numerical domain.

[0014] Furthermore, the method for obtaining the predicted waveform matrix includes:

[0015] The time series values ​​of each physical quantity channel in a single window data matrix in the window data matrix sequence are mapped to polar coordinates to generate corresponding angle values. A two-dimensional image matrix is ​​generated based on the angle values ​​and trigonometric functions. The two-dimensional image matrices corresponding to all physical quantity channels are stacked along the channel dimension to construct a multi-channel state feature map sequence.

[0016] Based on the preset historical sequence step size, the multi-channel state feature map sequence is stacked along the time dimension to construct a spatiotemporal feature sequence;

[0017] Based on the spatiotemporal feature sequences, long-time sequences and short-time sequences are constructed respectively. A dual-branch spatiotemporal prediction network is used to extract and encode them in parallel to generate a fusion context vector. The fusion context vector is then input into the decoder to obtain the predicted feature map sequence.

[0018] After performing symmetry structure correction and feature denoising refinement preprocessing on the predicted feature map sequence, the predicted waveform matrix is ​​obtained by combining Gram angle field inverse transformation and inverse normalization operations.

[0019] Furthermore, the method for generating the valve action control parameters includes:

[0020] The future risk index vector is extracted from the predicted waveform matrix, and a morphological feature vector is generated by compression through a convolutional autoencoder. The multidimensional digital signal stream at the current decision moment, the future risk index vector, and the morphological feature vector are concatenated to form a prediction enhancement state vector.

[0021] The predicted enhanced state vector is input into the reinforcement learning policy network to generate proposed action parameters. The safety action envelope is calculated based on the future risk index vector and fluid dynamics principles. The proposed action parameters are then processed in parallel within the safety action envelope by a decision function to generate valve action control parameters.

[0022] Furthermore, the method for generating the adaptive reward value includes: calculating a real-time risk coefficient based on the maximum peak value of future pressure in the future risk indicator vector and a preset pressure threshold, dynamically adjusting the weight coefficient based on the risk coefficient, and using the weight coefficient to perform a weighted summation of the four sub-item rewards to generate an adaptive reward value.

[0023] Furthermore, the four sub-rewards are peak suppression reward, oscillation suppression reward, duration penalty reward, and high-frequency energy penalty reward, respectively.

[0024] The peak suppression reward is a penalty term for the maximum time series value of the pipeline outlet pressure, used to guide the policy network to learn to suppress overpressure peaks;

[0025] The oscillation suppression reward is a penalty term for the integral of the square of the second time derivative of the pipeline outlet pressure, which is used to guide the policy network to learn to suppress severe pressure oscillations.

[0026] The duration penalty reward is a penalty item for the total duration of the action;

[0027] The high-frequency energy penalty reward is a penalty term for spectral energy in the pipeline outlet pressure that is higher than the cutoff frequency, used to guide the policy network to learn to suppress high-frequency vibrations.

[0028] Furthermore, the method for obtaining the updated spatiotemporal prediction network and policy network includes:

[0029] Valve action control parameters are divided into: An ordered subset of parameters is used to generate A local target opening curve is obtained. Position continuity constraints and velocity continuity constraints are applied at the splicing points of the local target opening curves. The curves are spliced ​​sequentially in time order and decoded into a target valve opening curve.

[0030] Based on the target valve opening curve, a feedforward control signal is generated using the valve actuator inverse dynamics model, and a feedback control signal is generated by combining the tracking error between the actual valve position and the target position. The feedforward control signal and the feedback control signal are superimposed to form a composite control signal to drive the valve to perform the action, and the actual response data stream is collected synchronously.

[0031] Based on the assessment of unexpectedness and risk, a hybrid priority is calculated for the experience quadruples stored in the experience replay pool, and historical experience data is extracted according to the hybrid priority. The spatiotemporal prediction network and policy network are then updated using the historical experience data to obtain the updated spatiotemporal prediction network and policy network.

[0032] Furthermore, the positional continuity constraint is achieved by constraining the last control point of the preceding curve segment and the first control point of the following curve segment to be the same point; the velocity continuity constraint is achieved by constraining the penultimate control point of the preceding curve segment, the splicing point, and the second control point of the following curve segment to be collinear.

[0033] A water hammer pressure prediction and active suppression system, used to implement the above-mentioned water hammer pressure prediction and active suppression method, includes: a dynamic data acquisition module, an intelligent waveform prediction module, a valve action control module, and an adaptive closed-loop control module;

[0034] The dynamic data acquisition module is used to acquire multidimensional digital signal streams characterizing the dynamic response of the pipeline system, and generates a window data matrix sequence by sequentially segmenting and dimensionless transformation through a sliding window mechanism.

[0035] The intelligent waveform prediction module is used to image-encode the window data matrix sequence and construct a spatiotemporal feature sequence, process the spatiotemporal feature sequence using a dual-branch spatiotemporal prediction network, output a prediction feature map sequence, and decode the prediction feature map sequence to obtain the prediction waveform matrix.

[0036] The valve action control module is used to extract future risk index vectors and morphological feature vectors from the predicted waveform matrix, and splice them with the multidimensional digital signal stream at the current decision moment to form a predicted enhanced state vector. The predicted enhanced state vector is then input into the reinforcement learning strategy network to generate valve action control parameters. At the same time, it calculates the real-time risk coefficient based on the future maximum pressure peak and the preset pressure threshold, and dynamically adjusts the reward weights of each sub-item to generate an adaptive reward value.

[0037] The adaptive closed-loop control module is used to decode the valve action control parameters into the target valve opening curve, obtain the composite control signal using a feedforward-feedback composite control architecture to drive the valve to execute, synchronously collect the actual response data stream, encapsulate the actual response data stream and historical experience data into a quadruple, and filter and update the spatiotemporal prediction network and policy network online through a hybrid priority mechanism to obtain the updated spatiotemporal prediction network and policy network.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] This invention performs high-frequency synchronous acquisition and structured preprocessing of multidimensional physical quantities of pipeline systems, and quantitatively characterizes key dynamic details such as the steep rising edge of water hammer pressure waves and high-frequency oscillations that are easily ignored by traditional low-frequency monitoring systems. This improves the problem that prediction models cannot accurately capture complex waveforms due to the lack of key transient information.

[0040] This invention addresses the problem that traditional single-scale prediction models, which struggle to balance both aspects, lead to underestimation of the peak value or inaccurate overall trend judgment of complex water hammer waveforms by encoding one-dimensional time-series signals into spatiotemporal feature sequences and using a dual-branch network to extract their high-frequency transient details and low-frequency long-term trends in parallel.

[0041] This invention employs a hybrid decision-making framework that integrates reinforcement learning strategies and physical safety models. It combines data-driven performance optimization with safety boundaries based on fluid dynamics principles, thus avoiding the problem that traditional single reinforcement learning networks may generate dangerous control commands that exceed the physical limits of the pipeline system when pursuing optimal performance.

[0042] This invention constructs a closed-loop iterative mechanism that integrates high-fidelity command execution, real response acquisition, and hybrid priority experience playback. It intelligently filters the experience data from each physical interaction and uses it to update the prediction and decision-making model online. This improves the problem that traditional static models cannot detect the drift of operating conditions such as pipeline aging and medium changes, which leads to a gradual decline in their long-term operating performance. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating the principle of a water hammer pressure prediction and active suppression method according to the present invention.

[0045] Figure 2 This is a functional block diagram of a water hammer pressure prediction and active suppression system according to the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example 1

[0048] Please see Figure 1 As shown, this embodiment provides a method for predicting and actively suppressing water hammer pressure, including:

[0049] S1000, acquires multi-dimensional digital signal streams characterizing the dynamic response of pipeline systems. The data is then segmented using a sliding window mechanism and transformed using a dimensionless transformation to generate a sequence of window data matrices. .in, For time variables, For window indexing, For the number of windows, This is a window data matrix.

[0050] Specifically, this step aims to process the multi-dimensional digital signal stream acquired by field sensors, which characterizes the dynamic response of the pipeline system. Through a series of signal processing operations, it is transformed into a structured window data matrix sequence. This enables an effective connection between raw physical measurement data and model input features, providing a data foundation for subsequent steps.

[0051] Further, step S1000 includes:

[0052] Step S1100: Synchronously acquire multi-dimensional physical quantities of key measuring points in the pipeline system to construct a multi-dimensional digital signal stream. .

[0053] Specifically, this step aims to conduct high-frequency and synchronous dynamic testing of key physical quantities in the pipeline system, transforming the invisible, transient, and complex hydraulic state within the pipeline system into a quantifiable, multi-dimensional digital signal flow. .

[0054] In the specific implementation process, in order to comprehensively monitor the transient hydraulic response of the pipeline system under the excitation of actuator action, this step deploys distributed and high-precision data acquisition (DAQ) modules at key measuring points of the pipeline system. These key measuring points preferably include, but are not limited to, locations upstream and downstream of valves, pump outlets, pipeline bends, and pipeline ends.

[0055] All data acquisition modules operate at a synchronous sampling frequency of at least 10kHz. The system performs parallel acquisition of time-series data for multiple key physical quantities. The sampling frequency is set to ensure distortion-free capture of the steep rising edges, high-frequency oscillations, and complex waveform details formed by multiple reflections and superpositions of water hammer pressure waves. These high-frequency components are information that traditional low-frequency monitoring systems easily overlook, but is crucial for prediction models. The sampling frequency... This indicates the number of samples collected per second by the data acquisition module, measured in Hertz (Hz), and is used to characterize the resolution of the data in the time dimension.

[0056] The raw data from multiple channels acquired at any given time t constitutes a multidimensional real-time data vector, i.e., a multidimensional digital signal stream. The multidimensional digital signal stream As a digital representation of the instantaneous dynamic state of the pipeline system, it specifically includes five key physical quantities, namely inlet pressure. Export pressure Instantaneous flow Actuator control quantity and water pump speed .

[0057] Among them, inlet pressure This represents the instantaneous pressure value on the upstream side of the pipeline system monitoring section, measured in megapascals (MPa). It is used to characterize the dynamic changes in upstream boundary conditions and provides a basis for the source and propagation of pressure waves; outlet pressure. This indicates the instantaneous pressure value on the downstream side of the monitored section of the pipeline system, measured in megapascals (MPa), used to quantify the intensity and morphology of water hammer phenomena; instantaneous flow rate... This represents the instantaneous fluid volumetric flow rate at a specific cross-section within a pipe, expressed in cubic meters per second. Used to reflect the motion state of fluid mass; actuator control quantity The control state of a valve can be represented by two physically equivalent quantities: the normalized valve opening and the control electrical signal of the valve actuator; pump speed. This represents the instantaneous rotational speed of the water pump, used to reflect fluctuations in the pump's operating status. This physical quantity can be measured in different units, namely angular velocity and rotational speed, depending on the type of sensor deployed, but all represent the same physical meaning.

[0058] The normalized opening degree of the valve is used to describe the physical position of the valve core, and its value is a dimensionless real number in the interval [0, 1], where 0 and 1 correspond to the valve being fully closed and fully open, respectively. The control electrical signal of the valve actuator represents the command signal given to the valve positioner. The units of the angular velocity and rotational speed are degrees per second, respectively. The speed is measured in revolutions per minute (rpm), and the two can be converted to each other using a fixed coefficient.

[0059] Step S1200: A sliding window mechanism is adopted to process the multidimensional digital signal stream according to the preset window length and sliding step size. Perform serialization segmentation to construct a data fragment matrix sequence. .in, This is a matrix of data fragments.

[0060] Specifically, this step aims to process the continuous multidimensional digital signal stream acquired in step S1100. A sliding window mechanism is used for structured processing, forming a series of discrete and fixed-size data segment matrix sequences. Construct a standardized input sequence.

[0061] In the specific implementation process, in order to perform structured analysis on continuous dynamic response signals and ensure that subsequent deep neural network models can receive inputs of uniform size, this step uses a fixed-length sliding window to analyze the multidimensional digital signal stream. Periodic truncation is performed. This process is implemented using two key parameters: window length L and sliding step size S. The window length L defines the number of sampling points contained in each data segment, and its value must be selected to ensure complete coverage of the main oscillation period and decay process of a typical water hammer event. The sliding step size S is the number of sampling points between the starting points of two adjacent windows, and its value is a positive integer less than the window length L, used to control the overlap rate between windows.

[0062] Through the aforementioned sliding window mechanism, multidimensional digital signal streams The data is divided into a series of finite-length data segments. The segmentation process is shown in the following formula:

[0063] ;

[0064] in, The matrix representing the i-th data segment is composed of L consecutively acquired data vectors. A 3D matrix; i represents the window index, a monotonically increasing positive integer starting from 1, ranging from 1 to the total number of windows that the data stream can be divided into. , used to uniquely identify the data segment matrix in the data segment matrix sequence; K represents the total number of physical quantities; This represents the starting time of the i-th window, and is the matrix of the i-th data segment. The timestamp of the first included data vector is determined based on the window index i, the sliding step size S, and the sampling time interval. The calculated nonnegative real number This is used to precisely anchor the current data segment matrix across the entire time axis; The sampling time interval is the time difference between two adjacent data sampling points, and its value is the sampling frequency. The reciprocal of.

[0065] Finally, all the sequentially generated data fragment matrices are combined into an ordered sequence of data fragment matrices. .

[0066] Step S1300, for the data fragment matrix sequence The dimensional differences of multiple physical quantity channels are addressed by performing dimensionless data transformation through linear mapping, generating a window data matrix sequence with a unified numerical domain. .

[0067] Specifically, this step aims to sequence the data fragment matrix formed in step S1200. Different physical quantity channels are normalized to a unified numerical range through a linear mapping method, generating a window data matrix sequence with a unified numerical domain. This is to eliminate the impact of dimensional differences on subsequent steps.

[0068] In the specific implementation process, this step involves processing each data segment matrix. Each channel within the array undergoes Min-Max Normalization to linearly map the time-series values ​​of each physical quantity to the interval [-1, 1], resulting in the corresponding window data matrix. This operation eliminates the impact of differences in dimensions, units, and numerical ranges among various physical quantity channels on the convergence and prediction accuracy of subsequent models, ensuring comparability under a unified scale.

[0069] After normalization, all the sequentially generated window data matrices form an ordered sequence of window data matrices. .

[0070] S2000, sequence of window data matrix Perform image encoding and construct spatiotemporal feature sequences A dual-branch spatiotemporal prediction network is used to analyze the spatiotemporal feature sequence. Process the data to output a predicted feature map sequence. For the predicted feature map sequence Decoding is performed to obtain the predicted waveform matrix. .

[0071] Specifically, this step aims to process the window data matrix sequence constructed in step S1300. A time-series image-based approach is used for depth prediction, encoding one-dimensional time-series data into two-dimensional image features to capture the spatiotemporal dynamics of the pipeline system and outputting a predicted waveform matrix. This enables the prediction of the future state of a nonlinear dynamic system like water hammer.

[0072] Further, step S2000 includes:

[0073] Step S2100: Sequence of window data matrix Single window data matrix Time series values ​​of each physical quantity channel within the system Perform polar coordinate angle mapping to generate corresponding angle values. Based on the angle value Generate a two-dimensional image matrix using trigonometric functions The two-dimensional image matrix corresponding to all physical quantity channels Stack along the channel dimension to construct a multi-channel state feature map sequence. .in, This is a multi-channel state feature map.

[0074] Specifically, this step aims to sequence a one-dimensional window data matrix. Encoded as a two-dimensional multi-channel state feature map sequence This enables the representation of image structured texture features that reflect the intrinsic correlations of time-series signals.

[0075] In the specific implementation process, this step involves the window data matrix sequence. Single window data matrix Each physical quantity channel within is encoded, and this encoding process consists of the following two stages:

[0076] The first stage involves the perspective mapping of time series values. This involves normalizing the time series values ​​to the interval [-1, 1]. By mapping the signal amplitude at each time point to the angle interval [0, π] using the inverse cosine function, the signal amplitude is converted into its angle value on the unit circle in the polar coordinate system. .in, This represents the normalized value of the k-th physical quantity channel at the j-th time point, i.e., the time series value; k represents the channel index, used to identify the current physical quantity time series being processed, and its value range is... K is the total number of physical quantities defined in step S1100, for example Corresponding inlet pressure, Corresponding to the export pressure; j represents the time point index, used to identify the time point position or sampling point number within the window, and its value range is... , where L is the window length defined in step S1200.

[0077] The second stage involves image encoding of angular relationships. Based on the angle values ​​generated in the previous stage, angle values ​​at different time points are calculated. Using trigonometric functions to generate a two-dimensional image matrix .in, Represents the time series value of the k-th physical quantity channel The generated dimension is A two-dimensional image matrix. The two-dimensional image matrix... The pixel values ​​in the m-th row and n-th column are respectively determined by the angle values ​​corresponding to the m-th and n-th time points. and The pixel value is defined by the sum of the angle cosine and the sum of the two time points, and it encodes the temporal correlation between these two time points. Here, m and n represent the two-dimensional image matrix. The row and column pixel indices are used in the two-dimensional image matrix. To locate a pixel, its value range is... .

[0078] After completing the above two stages, generate a two-dimensional image matrix from all physical quantity channels. Stack them along the channel dimension to form a Multi-channel state feature map This ultimately forms a multi-channel state feature map sequence. .

[0079] Step S2200: Based on the preset historical sequence step size For multi-channel state feature map sequences Stacking along the time dimension to construct a spatiotemporal feature sequence .

[0080] Specifically, this step aims to transform the time-discrete static multi-channel state feature map sequence output by step S2100 into a series of time-discrete static multi-channel state feature map sequences. They are combined along the time dimension to form a spatiotemporal feature sequence containing dynamic evolutionary information. It enables the transformation of data structures from static image sequences to dynamic video, and can simultaneously characterize spatial patterns and temporal evolution features.

[0081] In the specific implementation process, this step sets the historical sequence step size. From multi-channel state feature map sequences Extracting consecutive elements sequentially Multi-channel state feature map The multi-channel state feature map corresponding to window index i. Starting from, continuously take until They are then stacked along a new time dimension to construct a four-dimensional spatiotemporal feature sequence. This allows for the capture of the spatiotemporal dynamic evolution and time dependence of the water hammer effect in pipeline systems. The spatiotemporal feature sequence... Structurally equivalent to a segment containing A video with frames, each having K physical quantity channels, is a dimensional video. The four-dimensional tensor.

[0082] Step S2300, based on spatiotemporal feature sequences Construct long time series respectively and short time series Using a dual-branch spatiotemporal prediction network Parallel extraction and encoding to generate fused context vectors The fusion context vector The input is fed into the decoder to obtain the predicted feature map sequence. .

[0083] Specifically, this step aims to transform the spatiotemporal feature sequence containing historical evolutionary information constructed in step S2200. The data is processed through a dual-branch spatiotemporal prediction network, which outputs a sequence of predicted feature maps representing future dynamic information. This enables the prediction of nonlinear dynamic systems.

[0084] In its implementation, this step abandons the traditional approach of using a single fixed historical length for prediction and proposes a spatiotemporal prediction network that can simultaneously capture transient details and long-term trends. The network is a dual-branch structure that integrates long-term and short-term time scales. Its specific structure and processing flow are as follows:

[0085] The first stage involves multi-scale sequence construction. For each current time point to be predicted, the spatiotemporal feature sequence is used... Extract two subsequences of different historical lengths from the data: one with a historical length of... short time series And a history length of Long time series Among them, long-term time series Length of history Larger than short time series Length of history .

[0086] The second stage involves dual-branch parallel encoding. The two sequences are input into two independent encoders with identical structures but different parameters for parallel feature extraction. For the short-time encoder, the input short-time sequence is processed. It focuses on extracting and encoding high-frequency transient features such as steep rising edges and sharp peaks in water hammer waveforms to generate short-time context vectors. This encoder captures transient details that determine the maximum stress in the piping system by analyzing a small number of recent data frames, contributing sharpness and realism of waveform details, i.e., local features, to the prediction.

[0087] For long-time encoders, the input long-time sequence is processed. This study focuses on extracting and encoding low-frequency, long-term dependent features of water hammer pressure waves, such as macroscopic oscillation periods, attenuation patterns, and baseline drift, to generate long-term time-series context vectors. This encoder analyzes historical data across a wider range to grasp the overall macroscopic patterns of the system, contributing to the accuracy of the overall waveform structure and the rationality of trends in prediction.

[0088] The third stage is multi-scale feature fusion. This involves fusing the short temporal context vectors output by the two encoders. and long-term context vector The process involves fusion to generate a fused context vector that incorporates both transient details and long-term trends. This fusion can be achieved through vector concatenation, which enhances the correlation between features at different time scales, thus preserving both local dynamic responses and overall evolutionary patterns.

[0089] The fourth stage is sequence decoding and generation. This involves fusing the context vector. The input is fed to a shared decoder, which progressively generates a sequence of predicted feature maps for future H frames in an autoregressive manner. .in, Indicates the moment immediately following the end of historical observation. Initially, an ordered list of consecutive H frames is used to characterize the dynamic evolution of water hammer pressure over a future time period; This represents a single-frame predicted feature map at a future time, with dimensions of... , used to characterize the estimated state in a future frame; H represents the number of future time steps to predict, which is a preset positive integer used to determine the length of the future time for a single prediction, i.e., the predicted feature map sequence. The total number of future feature maps included.

[0090] Step S2400, for the predicted feature map sequence After performing symmetry structure correction and feature denoising as refinement preprocessing steps, the predicted waveform matrix is ​​obtained by combining Gram angle field inverse transform and inverse normalization operations. .

[0091] Specifically, this step aims to transform the predicted feature map sequence generated in step S2300. Decoding is performed to restore the high-dimensional image feature space to a one-dimensional time-series waveform, in order to construct a predictive waveform matrix that can be directly used for active suppression control decisions. .

[0092] In the specific implementation process, this step abandons the prediction of feature map sequences. The traditional direct decoding approach adds a feature refinement stage to eliminate minor artifacts and noise that may exist in the prediction. This process refines the predicted feature map sequence. Predicted feature maps for each frame Perform the following operations in sequence:

[0093] The first stage involves refining the predicted feature map. To improve the physical realism and signal quality of the waveform reconstruction, this step refines the predicted feature map before decoding. Each layer within the image matrix predicts the Gramian Angular Field (GAF). Structure preservation and feature denoising are performed to obtain a refined feature map. .

[0094] The structure is maintained by utilizing the inherent symmetry of the GAF matrix to predict GAF image matrices that may have slight asymmetries. Forced symmetry processing is performed to obtain the symmetric predicted GAF ​​image matrix. This is done to correct structural biases. The feature-based noise reduction is applied to the symmetrically predicted GAF ​​image matrix. Singular Value Decomposition (SVD) is performed to retain the primary singular values ​​that characterize the core dynamics of the signal, while noise-related secondary singular values ​​are set to zero. This process reconstructs the refined feature map after noise reduction. .

[0095] The second stage involves the Gram angle field GAF ​​inverse transform. This refines the characteristic spectrum. Perform the Gram angular field GAF ​​inverse transform by extracting the pixel values ​​on its diagonal. And by performing a square root operation, the corresponding normalized predicted value is decoded and reconstructed. The specific process formula is as follows:

[0096] ;

[0097] in, The normalized prediction value represents the dimensionless prediction value of the k-th physical quantity channel at the j-th future time point, reconstructed after Gram angle field GAF ​​inverse transform; k represents the channel index, whose value ranges from... K is the total number of physical quantities defined in step S1100; j represents the time point index, used to identify the time point position or sampling point number within the window, and its value range is... , where L is the window length defined in step S1200.

[0098] The third stage is denormalization. This involves reconstructing the normalized predicted values. Perform an inverse normalization linear transformation operation, using the historical maximum value. and historical minimum By restoring it to its original physical dimensions and numerical range, the predicted value of the physical quantity at a single point can be obtained. The specific process formula is as follows:

[0099] ;

[0100] in, This represents the predicted value of a single-point physical quantity, which is the predicted value of the k-th physical quantity channel at the j-th time point in the future after inverse normalization. The historical maximum value is the maximum value of the data sequence of the k-th physical quantity channel that is recorded or used when normalizing the data of the k-th physical quantity channel in step S1300. The historical minimum value is the minimum value of the historical data sequence of the k-th physical quantity channel that is recorded or used when the data of the k-th physical quantity channel is normalized in step S1300.

[0101] Finally, the predicted values ​​for all physical quantity channels are... By combining these components, a prediction waveform matrix with high physical realism is obtained. .

[0102] S3000, for the predicted waveform matrix Extracting future risk indicator vectors and morphological feature vectors and the current decision-making moment Multidimensional digital signal stream Concatenate to form the prediction enhancement state vector The predicted enhanced state vector Input reinforcement learning policy network Generate valve action control parameters Simultaneously, a real-time risk coefficient is calculated based on the future maximum pressure peak and the preset pressure threshold. It also dynamically adjusts the reward weights of each sub-item to generate adaptive reward values. .

[0103] Specifically, this step aims to construct a reinforcement learning-based decision-making module that incorporates the multidimensional digital signal stream from step S1100. and the predicted waveform matrix output by step S2400 These are collectively mapped to valve action control parameters. This step constructs a forward-looking state representation, integrating current state and future prediction information to achieve a shift from passive, delayed response to proactive, pre-intervention, suppressing potential destructive water hammer in its nascent stage.

[0104] Further, step S3000 includes:

[0105] Step S3100, from the predicted waveform matrix Extracting future risk indicator vectors The morphological feature vector is generated by compression using a convolutional autoencoder. The current decision moment Multidimensional digital signal stream The future risk indicator vector and morphological feature vectors The vectors are concatenated to form the predicted enhanced state vector. .

[0106] Specifically, this step aims to construct a state representation with condensed information and future vision for the reinforcement learning decision-making module, i.e., to predict the augmented state vector. The predicted enhanced state vector Integrating the current decision-making moment Multidimensional digital signal stream and predicted waveform matrix .

[0107] In the specific implementation process, this step proposes a guided state representation construction method based on feature engineering. This method, while fusing the current state, deeply refines and structurally expresses the future prediction waveform. Specifically, it includes the following three stages:

[0108] The first stage involves extracting future risk indicators. This includes analyzing the predicted waveform matrix. Calculations are performed to extract key statistical features for quantifying future risks, forming a future risk indicator vector. The future risk indicator vector This includes the future maximum pressure peak, the time to reach the maximum peak, the future pressure integral, and the energy or variance of the future waveform oscillation.

[0109] The second stage involves future waveform compression. To understand the overall shape of the predicted waveform, the predicted waveform matrix is... The input is a pre-trained convolutional autoencoder, through which the predicted waveform matrix is ​​generated by its encoder part. Compressed into low-dimensional morphological feature vectors .

[0110] The third stage involves final state vector concatenation. This involves combining the multidimensional digital signal stream of the current state. Future risk indicator vector and morphological feature vectors The vectors are then concatenated to form the final predicted augmentation state vector. .

[0111] Step S3200, predict the enhanced state vector Input reinforcement learning policy network To generate proposed action parameters And based on future risk indicator vectors And the fluid dynamics principle is used to calculate the safety action envelope, and the proposed action parameters are determined by the adjudication function within the safety action envelope. Parallel processing is performed to generate valve action control parameters. .

[0112] Specifically, this step aims to enhance the predicted state vector output from step S3100 using a hybrid decision framework. Mapped to optimal and safe valve action control parameters .

[0113] In its implementation, this step abandons the traditional approach of relying on a single black-box deep reinforcement learning network for decision-making, and proposes a hybrid decision-making framework that integrates performance-driven and security-constrained approaches. This hybrid decision-making framework collaboratively operates two core components, as follows:

[0114] The first core component is the optimal policy proposal based on reinforcement learning. This involves predicting the augmented state vector. Input to reinforcement learning policy network For example, the Actor network in the Deep Deterministic Policy Gradient (DDPG) algorithm. This network, acting as a performance constraint, learns, through massive training, to map any state to a proposed action parameter that maximizes the composite reward function. The composite reward function is defined in step S3300 below. Furthermore, to ensure the smoothness of valve operation, the proposed action parameters... The coordinates of the control points are parameterized as a curve.

[0115] Core component two: Dynamic calculation of security boundaries based on a physical model. In parallel, the physical security module, serving as a security constraint, enhances the state vector based on predictions. Medium-key future risk indicator vector In particular, by predicting the maximum future pressure peak, and combining fundamental principles of fluid mechanics, such as the Joukowsky theorem, a safe operating envelope is dynamically calculated. This envelope defines the valve operating curve constraints that ensure the pipeline water hammer pressure never exceeds a preset safety threshold, such as the maximum permissible closing speed and the minimum closing time.

[0116] Finally, the proposed action parameters for the policy network are determined by the adjudication function based on the constraints of the safety action envelope. Project or correct the parameters to generate valve action control parameters that strictly adhere to safety red lines. .in, Indicates the first Each control point constitutes a valve action control parameter. The basic elements of a set; The index of a control point is a non-negative integer used to locate a specific control point, and its value ranges from 1 to 2. ; It is a non-negative integer that determines the complexity and degrees of freedom of the curve. The total number of control points, used to define the control parameters that constitute the valve's action. The upper limit of the number of control points.

[0117] Step S3300, based on the future risk indicator vector The real-time risk coefficient is calculated by comparing the future maximum pressure peak with the preset pressure threshold. And based on the aforementioned risk coefficient Dynamically adjust weighting coefficients Using the weighting coefficients The four sub-rewards are weighted and summed to generate an adaptive reward value. .

[0118] Specifically, this step aims to define a comprehensive and adaptively adjustable evaluation system, based on the predicted augmented state vector output in step S3100 during the reinforcement learning training phase. and valve action control parameters from step S3200 Calculate the adaptive reward value The quality of actions is quantitatively evaluated to guide the policy network to update towards the optimal goal under different operating conditions.

[0119] In the specific implementation process, this step abandons the traditional composite reward function with fixed weight allocation and extracts a composite reward function based on risk state adaptive weighting. This function can adjust the optimization priority of various performance indicators in real time according to the predicted future risks.

[0120] Based on the predicted enhanced state vector Future forecast information, especially future risk indicator vectors The maximum future pressure peak is calculated, and a standardized real-time risk coefficient is determined. This is used to quantify the degree of danger of the pipeline system under its current condition. The specific process is as follows: Preset a safe operating pressure threshold. and critical safety pressure threshold Two pressure thresholds are used. The safe operating pressure threshold is lower than the pipeline design pressure and represents the upper limit of safety during normal operation. The critical safe pressure threshold is the absolute maximum pressure that the pipeline system can withstand, or the pressure set by the safety valve. The future maximum pressure peak is then linearly normalized within the range defined by the two preset pressure thresholds, transforming the predicted pressure value in specific physical units into a dimensionless real-time risk coefficient between [0, 1]. .

[0121] If the maximum future pressure peak is lower than the safe operating pressure threshold The real-time risk coefficient A value of 0 indicates that the pipeline system is in a safe zone; if the future maximum pressure peak exceeds the critical safe pressure threshold... The real-time risk coefficient A value of 1 indicates that the pipeline system is in a dangerous zone; if the future maximum pressure peak falls between these two values, the real-time risk coefficient is... A smooth transition between 0 and 1 indicates that the piping system is in the warning zone.

[0122] The composite reward function consists of four sub-rewards with different performance metrics, and each sub-reward has a weighting coefficient. Based on real-time risk coefficient Dynamically adjusted, rather than a fixed constant. The sub-rewards include: peak suppression reward. Oscillation suppression reward Duration-based penalties and rewards and high-frequency energy penalty rewards The specific process formula is as follows:

[0123] ;

[0124] in, Indicates an adaptive reward value; These are the dynamic weighting coefficients for each sub-item reward, used to adjust the priority of optimization objectives under different operating conditions.

[0125] The peak suppression reward The pressure at the pipeline outlet The penalty term for the maximum value of the time series, which takes the value of a negative real number or zero, is used to guide the policy network to learn to suppress overvoltage peaks and ensure safety; the oscillation suppression reward... The pressure at the pipeline outlet The penalty term for the integral of the square of the second-order time derivative, which takes the value of a negative real number or zero, is used to guide the policy network to learn to suppress violent stress oscillations and improve stability; the duration penalty reward... This is a penalty term for the total action duration, with a negative real number as its value. It is used to guide the policy network to seek more efficient operations while satisfying safety constraints. The total action duration is the time consumed to execute a complete valve action, and its value is determined by the valve action control parameters. The determined positive real number; the high-frequency energy penalty reward The pressure at the pipeline outlet The penalty term for spectral energy above the cutoff frequency, which takes the value of a negative real number or zero, is used to guide the policy network to learn to suppress high-frequency vibrations and improve fatigue resistance. The cutoff frequency is a preset frequency value based on the physical characteristics of the pipeline system, rather than a dynamically changing variable during training, and its value is a positive real number.

[0126] S4000, valve action control parameters Decoded into the target valve opening curve The composite control signal is obtained by adopting a feedforward-feedback composite control architecture. To drive the valve to perform the operation, the actual response data stream is collected synchronously. The actual response data stream Encapsulate historical experience data into quadruples The spatiotemporal prediction network is selected and updated online through a hybrid priority mechanism. With policy network The updated spatiotemporal prediction network is obtained. With policy network .in, This is the actual reward value. Enhance the state vector for the new prediction;

[0127] Specifically, this step aims to process the valve action control parameters output from step S3200. The data is translated into control commands for physical actuators and executed. Simultaneously, real feedback data is collected, and the prediction and decision models are continuously iterated and updated based on this data, outputting an updated spatiotemporal prediction network. With policy network This enables the organic closed-loop coupling of intelligent predictive decision-making with the pipeline physical system.

[0128] Further, step S4000 includes:

[0129] Step S4100: Set valve action control parameters Divided into An ordered subset of parameters is used to generate Local target opening curve In the local target opening curve At the splicing points, positional continuity constraints and velocity continuity constraints are applied, and the splices are performed sequentially in time order to decode the target valve opening curve. .

[0130] Specifically, this step aims to adjust the valve action control parameters. By using a trajectory stitching method optimized by continuity constraints, the target trajectory is decoded into a high-resolution and globally smooth continuous target trajectory that can be directly tracked and executed by the underlying controller, i.e., the target valve opening curve. .

[0131] In the specific implementation process, to overcome the shortcomings of traditional single-curve description methods in balancing global efficiency and fine-grained local control, this step proposes a piecewise composite curve generation method that satisfies high-order continuity constraints. This method includes the following three stages:

[0132] The first stage involves segmented parameterization of the valve's action control parameters. Divided into An ordered set of parameter subsets Each parameter subset It is a set used to define the first The curve control points required for the local curve of each action segment. Among them, Indicates the first The parameter subset of the i-th action segment, containing the i-th action segment All the information required for a segment of a local curve, its value is a combination of... Control points The ordered set constituted Indicates the first A subset of parameters for each action segment; The index representing the action segment is a positive integer used to identify the order of the action segments, and its value ranges from 1 to 2. ; Indicates the number of action segments, defining how many independent stages the entire action is divided into; Indicates the first The first segment of the curve The control point constitutes the first control point. The parameter subset of each action segment The basic elements; The index of a control point is a non-negative integer used to identify the sequence number of a control point within a specific segment. Its value range is... , Indicates the first The first segment of the curve One control point; Indicates the first The order of a curve segment is a non-negative integer that determines the complexity of that curve segment. This represents the total number of control points for this curve segment.

[0133] The second stage involves segmented curve generation and splicing. This is based on segmenting each action. The corresponding control point parameter subset Calculate the local target aperture curve within this segment. The specific process formula is as follows:

[0134] ;

[0135] in, Indicates the time interval The first defined within Segment local target opening curve; These represent the binomial coefficients used to determine each control point. Influence weight; This represents a time variable, with values ​​being continuous real numbers throughout the entire action duration interval. It is used to query the target valve opening at any given time, when a specific segment is called. At that time, the value of the variable t falls within the effective time interval of that segment. Inside; They represent the first The start and end times of each action segment are the boundary points that divide the duration of that action segment. Indicates the first The duration of each action segment, whose value is... .

[0136] The third stage involves higher-order continuity constraints. This applies to all... Local target opening curve The components are pieced together sequentially in chronological order to form a complete target valve opening curve. At the junction, two types of continuity constraints are preferably applied to ensure that the valve action has no sudden speed changes during segmented transitions, achieving global smoothness. Specifically, this includes the following:

[0137] First, positional continuity constraint. At any splicing point, the first... The endpoint of the curve segment must be at the same position as the first segment. The starting points of the curve segments completely coincide. The continuity constraint at this position is achieved by constraining the first... The last control point of the segment curve With the The first control point of the segment curve To achieve the same point.

[0138] Second, velocity continuity constraint. At any splicing point, the first... The tangent at the endpoint of the curve segment must be parallel to the first... The tangents at the starting points of the curve segments are completely consistent. The tangent refers to the instantaneous velocity. This velocity continuity constraint is achieved by constraining the first... The second to last control point of the segment curve splicing points and the The second control point of the segment curve This is achieved by having three points collinear.

[0139] Step S4200, based on the target valve opening curve Feedforward control signals are generated using the inverse dynamics model of the valve actuator. And in combination with the actual position of the valve and target location Tracking error Generate feedback control signal The feedforward control signal and feedback control signal Superimposed to form a composite control signal To drive the valve to perform actions, the actual response data stream is collected simultaneously. .

[0140] Specifically, this step aims to act as a bridge connecting digital control commands and physical execution actions, executing step S4100 to decode the target valve opening curve. Simultaneously, it collects the actual response data stream of the system's real feedback triggered by the action. .

[0141] In the specific implementation process, to overcome the inherent delay and error of traditional feedback controllers in high-speed, nonlinear trajectory tracking, this step proposes a feedforward-feedback control architecture based on an inverse dynamics model. This includes the following three stages:

[0142] The first stage is the feedforward prediction stage. This stage constructs an inverse dynamics model of the valve actuator. This model will determine the target valve opening curve. Using its time derivative as input, and based on the inertia, damping, friction, and other characteristics of the valve actuator, the main control quantities required to achieve the trajectory, such as motor current or torque, are pre-calculated to generate a feedforward control signal. The time derivative mentioned here is the first-order time derivative, i.e., the target velocity. and the second time derivative, i.e., the target acceleration. .

[0143] The second stage is the feedback correction stage. A low-level feedback controller is used to continuously monitor the actual valve position. With the target location Tracking error between And generate a feedback control signal based on the tracking error. This is used to compensate for nonlinear characteristics and external disturbances not covered by the model. The actual position of the valve is... It is collected in real time by a position sensor installed on the valve actuator to form a time series, which is used to calculate the tracking error.

[0144] The third stage is the composite control signal output stage. This involves outputting the aforementioned feedforward control signal. With feedback control signal The signals are superimposed to form a composite control signal that is ultimately applied to the valve actuator. .

[0145] This feedforward-feedback control architecture proactively and predictively provides the majority of the control energy in the feedforward prediction stage, while the feedback correction stage only needs to compensate for minor residual errors. This collaborative working mode ensures that upper-level intelligent decision-making can be realized to the greatest extent in the physical world. Ultimately, this results in the formation of a composite control signal. Subsequently, the distributed data acquisition (DAQ) modules deployed in step S1100 synchronously monitor all key physical quantities under this control strategy, forming an actual response data stream. .

[0146] Step S4300: Based on the assessment of unexpectedness and risk, generate the experience quadruple stored in the experience replay pool. Calculate the hybrid priority, extract historical experience data based on the hybrid priority, and use the historical experience data to perform gradient updates on the spatiotemporal prediction network and the policy network to obtain the updated spatiotemporal prediction network. With policy network .

[0147] In the specific implementation process, this step abandons the inefficient and easily overlooked traditional model update method and proposes an experience replay mechanism based on hybrid priorities. This experience replay mechanism can intelligently select the most valuable historical experience data for focused learning. Specifically, it includes the following three stages:

[0148] The first stage involves the encapsulation and storage of empirical quadruples. This includes the predicted enhanced state vector output from step S3100. Valve action control parameters output by step S3200 The actual response data stream monitored in step S4200 The actual reward value is obtained by substituting the real data into the composite reward function defined in step S3300. and based on actual response data stream The final time measurement value is substituted into step S3100 to generate a new predicted enhanced state vector. Together they are encapsulated into an empirical quadruple. And store them in a limited-capacity experience replay pool.

[0149] The second phase involves experience-based sampling with mixed priorities. When sampling from the experience replay pool, a mixed priority score is calculated for each experience quadruple, and sampling is prioritized based on this score. The mixed priority is weighted by the following unexpectedness assessment and hazard assessment.

[0150] The unexpectedness assessment, based on temporal differential error, measures the difference between the evaluator network's pre-emptive value prediction of the "state-action" pair and its post-emptive value assessment after experiencing the actual reward and state transition. The larger the absolute value of this difference, the more the experience exceeds expectations, the higher its value for the evaluator network's update, and therefore the higher its priority. The risk assessment is based on the actual response data recorded in the experience quadruple, combined with the proximity of its actual maximum stress peak to the stress threshold, to calculate its objective post-emptive risk level. Experiences with extremely high risk are still given high priority even if their unexpectedness assessment is low.

[0151] The third stage involves gradient updates of the model parameters. Based on the aforementioned mixed priority, a small batch of historical empirical data is extracted from the empirical replay pool, and this high-value empirical data is used to update the spatiotemporal prediction network in step S2300. and the reinforcement learning policy network of step S3200 The network parameters are updated using gradients to obtain the updated spatiotemporal prediction network. With policy network .

[0152] Specifically, as the network learns from online data, the spatiotemporal prediction network is updated. It can capture slowly changing physical characteristics in pipeline systems caused by factors such as fluid temperature, composition changes, or pipe aging. For example, when the friction coefficient increases due to scaling on the inner wall of the pipe, it learns and corrects its predictions through "execution-feedback" experience, resulting in a more accurate predicted waveform matrix. Compared to the future actual waveform, i.e., the actual response data stream The error between them remains at an extremely low level over a long period, ensuring the reliability of the predictive decisions. Meanwhile, the updated policy network... By focusing on key historical cases of high risk or inaccurate predictions, the system will exhibit a clear risk aversion tendency when encountering similar high-risk situations in the future, generating more conservative and safer valve action control parameters. .

[0153] Example 2

[0154] This embodiment, based on Embodiment 1, provides a water hammer pressure prediction and active suppression system, such as... Figure 2 As shown, it includes: a dynamic data acquisition module, an intelligent waveform prediction module, a valve action control module, and an adaptive closed-loop control module;

[0155] The dynamic data acquisition module is used to acquire multi-dimensional digital signal streams characterizing the dynamic response of the pipeline system. The data is then segmented using a sliding window mechanism and transformed using a dimensionless transformation to generate a sequence of window data matrices. .

[0156] The intelligent waveform prediction module is used to sequence the window data matrix. Perform image encoding and construct spatiotemporal feature sequences A dual-branch spatiotemporal prediction network is used to analyze the spatiotemporal feature sequence. Process the data to output a predicted feature map sequence. For the predicted feature map sequence Decoding is performed to obtain the predicted waveform matrix. .

[0157] The valve action control module is used to process the predicted waveform matrix. Extracting future risk indicator vectors and morphological feature vectors and the current decision-making moment Multidimensional digital signal stream Concatenate to form the prediction enhancement state vector The predicted enhanced state vector Input reinforcement learning policy network Generate valve action control parameters Simultaneously, a real-time risk coefficient is calculated based on the future maximum pressure peak and the preset pressure threshold. It also dynamically adjusts the reward weights of each sub-item to generate adaptive reward values. .

[0158] The adaptive closed-loop control module is used to process valve action control parameters. Decoded into the target valve opening curve The composite control signal is obtained by adopting a feedforward-feedback composite control architecture. To drive the valve to perform the operation, the actual response data stream is collected synchronously. The actual response data stream Encapsulate historical experience data into quadruples The spatiotemporal prediction network is selected and updated online through a hybrid priority mechanism. With policy network The updated spatiotemporal prediction network is obtained. With policy network .

[0159] The parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0160] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for water hammer pressure prediction and active mitigation, characterized by, The method comprises the following steps: Collecting a multi-dimensional digital signal stream representing the dynamic response of the pipeline system, and sequentially segmenting and non-dimensionalizing the multi-dimensional digital signal stream through a sliding window mechanism to generate a window data matrix sequence; Image coding and constructing a space-time feature sequence from the window data matrix sequence, processing the space-time feature sequence using a double-branch space-time prediction network, outputting a predicted feature spectrum sequence, and decoding the predicted feature spectrum sequence to obtain a predicted waveform matrix; Extracting a future risk indicator vector and a morphological feature vector from the predicted waveform matrix, and concatenating them with the multi-dimensional digital signal stream at the current decision-making time to form a prediction-enhanced state vector, inputting the prediction-enhanced state vector into a reinforcement learning strategy network to generate a valve action control parameter; simultaneously calculating a real-time risk coefficient according to the maximum future pressure peak value and a preset pressure threshold value, and dynamically adjusting the reward weight of each sub-item to generate an adaptive reward value; Decoding the valve action control parameter into a target valve opening curve, obtaining a composite control signal using a feedforward-feedback composite control architecture to drive the valve to execute, and synchronously collecting actual response data stream, encapsulating the actual response data stream and historical experience data into a four-tuple, filtering and online updating the space-time prediction network and the strategy network through a hybrid priority mechanism to obtain an updated space-time prediction network and strategy network.

2. The method of claim 1, wherein, The method for generating the window data matrix sequence comprises: Synchronously collecting multi-dimensional physical quantities of key measuring points of the pipeline system to construct a multi-dimensional digital signal stream; Using a sliding window mechanism, the multi-dimensional digital signal stream is sequentially segmented according to a preset window length and sliding step to construct a data segment matrix sequence; In view of the dimensional difference of the multi-physical quantity channels in the data segment matrix sequence, data non-dimension is performed through linear mapping to generate a window data matrix sequence with uniform numerical domain.

3. The method of claim 1, wherein the method further comprises: The method for obtaining the predicted waveform matrix comprises: Performing polar coordinate angle mapping on the time sequence values of each physical quantity channel in a single window data matrix in the window data matrix sequence to generate corresponding angle values, generating a two-dimensional image matrix based on the angle values and a trigonometric function, stacking the two-dimensional image matrices corresponding to all physical quantity channels along the channel dimension to construct a multi-channel state feature spectrum sequence; Stacking the multi-channel state feature spectrum sequence along the time dimension according to a preset historical sequence step to construct a space-time feature sequence; Based on the space-time feature sequence, a long time sequence and a short time sequence are constructed respectively, and a double-branch space-time prediction network is used to extract and encode in parallel to generate a fusion context vector, which is input into a decoder to obtain a predicted feature spectrum sequence; After the predicted feature spectrum sequence is sequentially refined and pretreated through symmetry structure correction and feature denoising, a predicted waveform matrix is obtained through Gram angle field inverse transformation and inverse normalization operation.

4. The method of claim 1, wherein, The method for generating the valve action control parameter comprises: Extracting a future risk indicator vector from the predicted waveform matrix, generating a morphological feature vector through a convolutional autoencoder, and concatenating the multi-dimensional digital signal stream at the current decision-making time, the future risk indicator vector, and the morphological feature vector to form a prediction-enhanced state vector; The predicted enhanced state vector is input into a reinforcement learning strategy network to generate proposed action parameters, and a safety action envelope is calculated based on a future risk indicator vector and fluid mechanics principles, the proposed action parameters are processed in parallel within the safety action envelope by a decision function to generate valve action control parameters.

5. The method of water hammer pressure prediction and active mitigation as claimed in claim 1, wherein, The adaptive reward value generation method comprises: calculating a real-time risk coefficient according to a maximum future pressure peak value in the future risk indicator vector and a preset pressure threshold, dynamically adjusting a weight coefficient according to the risk coefficient, and performing weighted summation on four sub-item rewards by using the weight coefficient to generate an adaptive reward value.

6. The method of water hammer pressure prediction and active mitigation of claim 5, wherein, The four sub-item rewards are peak suppression reward, oscillation suppression reward, time length penalty reward, and high-frequency energy penalty reward. The peak suppression reward is a penalty term for the maximum value of the time series of the pipeline outlet pressure. The oscillation suppression reward is a penalty term for the integral of the square of the second-order time derivative of the pipeline outlet pressure. The time length penalty reward is a penalty term for the total action time length. The high-frequency energy penalty reward is a penalty term for the spectral energy of the pipeline outlet pressure higher than a cutoff frequency.

7. The method of water hammer pressure prediction and active mitigation as claimed in claim 1, wherein, The method for obtaining the updated spatiotemporal prediction network and policy network comprises: The valve action control parameters are divided into an ordered parameter subset to generate a local target opening curve, position continuity constraints and velocity continuity constraints are imposed at the splicing points of the local target opening curve, spliced in time sequence, and decoded into a target valve opening curve; According to the target valve opening curve, a feedforward control signal is generated by using a valve actuator inverse dynamics model, a feedback control signal is generated by combining the tracking error of the actual position and the target position of the valve, and the feedforward control signal and the feedback control signal are superimposed to form a composite control signal to drive the valve to act, and the actual response data stream is synchronously collected; Based on the unexpectedness and risk assessment, a mixed priority is calculated for the experience quadruplets stored in the experience replay pool, and historical experience data is extracted according to the mixed priority, and the historical experience data is used to update the gradients of the spatiotemporal prediction network and the policy network to obtain the updated spatiotemporal prediction network and the policy network.

8. The method of water hammer pressure prediction and active mitigation of claim 7, wherein, The position continuity constraint is realized by constraining the last control point of the previous curve and the first control point of the next curve to be the same point; and the velocity continuity constraint is realized by constraining the penultimate control point of the previous curve, the splicing point, and the second control point of the next curve to be collinear.

9. A water hammer pressure prediction and active mitigation system for implementing a water hammer pressure prediction and active mitigation method according to any one of claims 1 to 8, characterized in that, The system comprises a dynamic data acquisition module, an intelligent waveform prediction module, a valve action control module, and an adaptive closed-loop control module. The dynamic data acquisition module is configured to acquire a multi-dimensional digital signal stream representing the dynamic response of the pipeline system, and sequentially perform windowing and dimensionless transformation to generate a window data matrix sequence. The intelligent waveform prediction module is configured to image encode the window data matrix sequence and construct a spatiotemporal feature sequence, process the spatiotemporal feature sequence by using a double-branch spatiotemporal prediction network, output a predicted feature spectrum sequence, and decode the predicted feature spectrum sequence to obtain a predicted waveform matrix. The valve action control module is configured to extract a future risk index vector and a shape feature vector from the predicted waveform matrix, splice the future risk index vector and the shape feature vector with a multi-dimensional digital signal stream at a current decision moment to form a prediction enhanced state vector, input the prediction enhanced state vector into a reinforcement learning strategy network to generate a valve action control parameter, calculate a real-time risk coefficient according to a future pressure maximum peak value and a preset pressure threshold value, and dynamically adjust a reward weight of each sub-item to generate an adaptive reward value. The adaptive closed-loop control module is configured to decode the valve action control parameter into a target valve opening curve, obtain a composite control signal by using a feedforward-feedback composite control architecture to drive the valve to execute, synchronously collect an actual response data stream, encapsulate the actual response data stream and historical experience data into a quadruple, filter and update an online time-space prediction network and a strategy network by using a hybrid priority mechanism to obtain an updated time-space prediction network and strategy network.

Images