Gate real-time opening and closing strategy control system based on deep learning

Abstract

The application discloses a gate real-time opening and closing strategy control system based on deep learning, comprising the following modules: a data acquisition module for generating a real-time water regime sequence; a feature modeling module for inputting the real-time water regime sequence into an improved Perceiver IO model to obtain a disturbance evolution feature vector sequence and an explanation field result; a trend risk calculation module for obtaining a control trend value and an oscillation risk value; an action generation module for generating a preliminary gate opening and closing action; a manifold analysis module for performing manifold curvature singularity dispersion analysis to obtain a water regime fluctuation intensity; an action adjustment module for adjusting the preliminary gate opening and closing action; and a control execution module for generating a gate opening and closing control instruction to perform opening and closing control on a gate actuator. The application realizes water regime disturbance modeling and gate intelligent control by combining the improved Perceiver IO model and the manifold curvature singularity dispersion analysis.

Description

Technical Field

[0001] This invention relates to the field of intelligent gate control technology, and in particular to a gate real-time opening and closing strategy control system based on deep learning. Background Technology

[0002] With the increasing demand for smart water conservancy and refined water resource management, real-time monitoring and intelligent control technologies for water level fluctuations, flow changes, and their coupling relationships during gate operation have received widespread attention. Existing gate opening and closing control systems mainly rely on empirical rule-based control strategies or simple single-variable feedback adjustment methods for gate opening control. However, these systems commonly suffer from the following problems in practical applications: Given the nonlinear, non-stationary, and strongly coupled characteristics of hydrological changes, traditional statistical analysis methods or simple time series prediction models struggle to effectively characterize the dynamic evolution of water level oscillations, especially under sudden disturbances or complex operating conditions, which can easily lead to response lags or control instability. Furthermore, existing methods typically lack interpretability for hydrological disturbances, making it difficult to comprehensively characterize water level oscillation behavior from multiple dimensions such as probability, energy, and gradient changes. This results in control strategies relying on empirical settings and failing to adapt to complex and changing environments. Simultaneously, during gate opening and closing control, the lack of quantitative assessment of the severity of hydrological fluctuations makes it impossible to effectively constrain opening changes when oscillation risks are high, easily triggering secondary water level fluctuations or even amplifying oscillations, thereby affecting the stability and safety of system operation.

[0003] Therefore, how to provide a gate real-time opening and closing strategy control system based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a real-time gate opening and closing strategy control system based on deep learning. This invention combines an improved Perceiver IO model with manifold curvature singularity dispersion analysis to perform unified modeling and feature extraction of multimodal hydrological operation data, thereby achieving accurate characterization of hydrological disturbance evolution characteristics. Furthermore, it adaptively adjusts the gate opening and closing actions based on the severity of hydrological fluctuations, thus realizing intelligent gate control. This system has the advantages of timely response, stable control, and adaptability to complex hydrological changes.

[0005] The deep learning-based gate real-time opening and closing strategy control system according to an embodiment of the present invention includes the following modules: The data acquisition module is used to acquire multimodal hydrological operation data, perform time-series alignment, and generate real-time hydrological sequences. The feature modeling module is used to input real-time hydrological sequences into an improved Perceiver IO model. The improved Perceiver IO model includes an input interaction layer, a frequency field sensing layer, a perturbation coupling and recombination layer, and an interpretation field generation layer to obtain a perturbation evolution feature vector sequence and interpretation field results. The frequency field sensing layer has a built-in frequency field spectralization mechanism. The interpretation field results include a latent variable perturbation probability field, an oscillation sensitive region field, and an instantaneous water level oscillation gradient field. The trend risk calculation module is used to determine the direction of water level change based on the perturbation evolution feature vector sequence, obtain the control trend value, and combine the latent variable perturbation probability field and the oscillation sensitive area field to obtain the oscillation risk value. The action generation module is used to generate preliminary gate opening and closing actions based on control trend values ​​and oscillation risk values; The manifold analysis module is used to map the perturbation evolution feature vector sequence to the water condition feature manifold space constructed by the instantaneous water level oscillation gradient field, perform manifold curvature singularity dispersion analysis, and obtain the degree of water condition fluctuation by calculating the curvature change and singularity distribution density of the current water condition. The action adjustment module is used to adjust the initial gate opening and closing action according to the intensity of water condition fluctuations, so as to obtain the final gate opening and closing action; The control execution module is used to generate gate opening and closing control commands based on the final gate opening and closing action, and to control the gate actuator to open and close.

[0006] Optionally, the data acquisition module specifically comprises: Collect multimodal hydrological operation data, which includes upstream water level data, downstream water level data, upstream flow data, downstream flow data, and gate opening data, and assign time stamps to each data. A unified sampling time sequence is set, and the multimodal hydrological operation data is time-aligned using the unified sampling time sequence as a time reference. Under the unified sampling time sequence, for each sampling time, the corresponding upstream water level data, downstream water level data, inlet flow data, outlet flow data, and gate opening data are extracted to generate the corresponding hydrological data fragment for the sampling time. The hydrological data segments corresponding to each sampling time are arranged sequentially according to time to generate a real-time hydrological sequence.

[0007] Optionally, the input interaction layer specifically comprises: The real-time hydrological sequence is input into a linear mapping structure, and each hydrological data segment is mapped to a hydrological feature vector of a uniform dimension. A corresponding time position encoding vector is generated according to the sequence number of each sampling time in the time series. The time position encoding vector has the same dimension as the corresponding hydrological feature vector. The time location encoding vector and the hydrological feature vector are added element by element along their corresponding dimensions to obtain the input feature vector, and an input feature vector sequence is constructed. According to the set window length and sliding step size, continuous subsequences are selected sequentially on the input feature vector sequence. The mean of the corresponding input feature vectors is calculated within each sliding time window to obtain aggregated feature vectors. The aggregated feature vectors are then used as latent vectors to form a latent array. Using the latent array as a query matrix, the input feature vector sequence is transformed into a key matrix and a value matrix through linear transformation. The query matrix and the key matrix are transposed and then multiplied to obtain the correlation matrix. Divide each element in the correlation matrix by the square root of the feature dimension of the key matrix, and input the result into the Softmax function for normalization calculation to obtain the attention weight matrix; The attention weight matrix and the value matrix are multiplied to obtain the latent feature sequence after interaction. Multiple feature update calculations based on the query matrix, key matrix, and value matrix are performed on the latent feature sequence after the interaction. After each feature update calculation, the updated latent feature sequence is added element by element to the corresponding input, and the addition result is subjected to layer normalization processing according to the feature dimension. After completing all feature update calculations, the final latent representation sequence is output and then input into the frequency field sensing layer.

[0008] Optionally, the frequency field sensing layer receives the final latent representation sequence, performs a frequency field spectralization mechanism, divides the final latent representation sequence into multiple time segment sequences according to time order, and performs a discrete Fourier transform operation on each time segment sequence to convert the corresponding time segment sequence from a time domain representation into a frequency spectrum sequence in complex form. Each frequency component in the frequency spectrum sequence is represented in complex form, wherein each frequency component includes a corresponding real part and an imaginary part. The real and imaginary parts of each frequency component are squared, and the summation and square root operations are performed on the squared results to obtain the amplitude of the corresponding frequency component. Based on the amplitudes of all frequency components within each time segment sequence, construct the corresponding amplitude spectrum sequence; The amplitude of each frequency component in the amplitude spectrum sequence is accumulated to obtain the frequency perturbation intensity value of the corresponding time segment sequence; Based on the frequency perturbation intensity values ​​of all time segment sequences, a sequence of frequency perturbation intensity values ​​arranged in chronological order is obtained; The frequency perturbation intensity value sequence is concatenated with the final potential representation sequence according to the feature dimension to generate a frequency-enhanced representation sequence; The perturbation coupling recombination layer receives the frequency enhancement representation sequence, performs a linear transformation operation on the frequency enhancement representation sequence, and performs element-wise multiplication with the frequency perturbation intensity value sequence to generate a perturbation coupling feature vector sequence corresponding to the frequency enhancement representation sequence in the time dimension. The perturbation coupling feature vector sequence is subjected to a difference operation along the time dimension by adjacent perturbation coupling feature vectors to obtain a change vector sequence; Based on the perturbation coupling feature vector sequence and the change vector sequence, the change vector corresponding to each time position is concatenated with the perturbation coupling feature vector of the corresponding time position to obtain the perturbation evolution feature vector of the corresponding time position. The perturbation evolution feature vectors corresponding to all time positions are arranged in chronological order to generate the perturbation evolution feature vector sequence.

[0009] Optionally, the interpretation field generation layer receives the perturbation evolution feature vector sequence, inputs the perturbation evolution feature vectors corresponding to each time position into the fully connected linear transformation structure, performs the Sigmoid function operation on the output result of the fully connected linear transformation structure to obtain the latent variable perturbation probability value corresponding to each time position, and arranges the latent variable perturbation probability values ​​corresponding to all time positions in chronological order to form the latent variable perturbation probability field. The perturbation evolution feature vectors corresponding to each time position are squared element by element, and the squared results are accumulated element by element to obtain the energy value corresponding to each time position. The energy values ​​corresponding to all time locations are summed and divided by the total number of time locations to obtain the global energy baseline value. The energy values ​​corresponding to time positions with energy values ​​greater than the global energy reference value are retained, while the energy values ​​corresponding to time positions with energy values ​​less than or equal to the global energy reference value are set to 0, and arranged in chronological order to form an oscillation-sensitive region field. Perform element-wise subtraction on the perturbation evolution feature vectors at adjacent time positions in the perturbation evolution feature vector sequence to obtain the difference vector; The absolute value of each element in the difference vector is taken and then accumulated to obtain the change intensity value corresponding to each time position; Subtract the change intensity values ​​corresponding to adjacent time positions to obtain the change intensity gradient value, and arrange all change intensity gradient values ​​in time order to form the instantaneous water level oscillation gradient field; The hidden variable perturbation probability field, the oscillation sensitive region field, and the instantaneous water level oscillation gradient field are output as the explanatory field results.

[0010] Optionally, the trend risk calculation module specifically comprises: Perform element-wise subtraction on the perturbation evolution feature vectors at adjacent time positions in the perturbation evolution feature vector sequence to obtain the difference vector; An element-wise summation operation is performed on each element in the difference vector to obtain the direction change value corresponding to each time position. The direction of water level change is determined according to the sign of the direction change value. The time position where the direction change value is greater than 0 is determined to be a rising water level and assigned a value of 1. The time position where the direction change value is equal to 0 is determined to be a level water level and assigned a value of 0. The time position where the direction change value is less than 0 is determined to be a falling water level and assigned a value of -1, thus obtaining a sequence of water level change directions. The values ​​at each time position in the sequence of water level change directions are accumulated to obtain the cumulative value of the direction; Divide the cumulative value of the direction by the total number of time positions to obtain the control trend value; The probability values ​​of hidden variable perturbations at each time position in the probability field of hidden variable perturbations are multiplied element-wise with the energy values ​​at the corresponding time positions in the field of oscillation sensitive region to obtain the oscillation intensity values ​​at the corresponding time positions. Based on the oscillation intensity values ​​at all time points, an oscillation intensity value sequence is constructed; An element-by-element summation operation is performed on the oscillation intensity value sequence to obtain the oscillation risk value.

[0011] Optionally, the action generation module specifically comprises: The gate opening and closing direction is determined based on the sign of the control trend value. When the control trend value is greater than 0, it is determined to be the opening direction; when the control trend value is equal to 0, it is determined to maintain the current state; and when the control trend value is less than 0, it is determined to be the closing direction. The absolute value of the control trend value is taken as the basic opening change. The adjustment amount of the opening change is obtained by subtracting the basic opening change from the oscillation risk value. When the change in the adjustment opening is less than 0, the change in the adjustment opening is set to 0; The change in gate opening is assigned a sign according to the gate opening and closing direction. When the gate opening and closing direction is the opening direction, the change in gate opening is positive; when the gate opening and closing direction is the closing direction, the change in gate opening is negative; and when the gate opening and closing direction is to maintain the current state, the change in gate opening is 0. The marked adjustment opening change is used as the gate opening change to generate the preliminary gate opening and closing action, which includes the gate opening and closing direction and the gate opening change.

[0012] Optionally, the manifold analysis module specifically comprises: The change intensity gradient value corresponding to each time position in the instantaneous water level oscillation gradient field is concatenated with the time index of the corresponding time position to construct the water condition characteristic manifold space. The perturbation evolution feature vector corresponding to each time position in the perturbation evolution feature vector sequence is multiplied element-wise with the change intensity gradient value at the corresponding time position to obtain the mapping feature vector. All the mapping feature vectors are arranged in chronological order to form the trajectory feature sequence in the hydrological state feature manifold space. Perform manifold curvature singularity dispersion analysis on the trajectory feature sequence, and perform element-wise subtraction on the mapping feature vectors of adjacent time positions to obtain the first-order difference vector; The element-wise subtraction operation is performed again on adjacent time positions in the first-order difference vector to obtain the second-order difference vector. The absolute value of each element in the second-order difference vector is then taken and accumulated to obtain the curvature value corresponding to each time position. The curvature values ​​corresponding to all time positions are summed and then divided by the total number of time positions to obtain the degree of curvature change. The maximum value among all curvature values ​​corresponding to all time positions is taken as the curvature reference value, and the degree of curvature change is divided by the curvature reference value to obtain the curvature normalization value; Count the time locations where the curvature value is greater than the degree of curvature change, and divide the count by the total number of time locations to obtain the singularity distribution density; The degree of water condition fluctuation is obtained by adding the normalized curvature value to the singularity distribution density.

[0013] Optionally, the motion adjustment module specifically comprises: The absolute value of the gate opening change during the initial gate opening and closing action is taken to obtain the initial opening change amplitude; Multiply the initial opening change amplitude by the severity of the water condition fluctuation to obtain the fluctuation impact. Subtract the fluctuation effect from the initial opening change amplitude to obtain the adjusted opening change amplitude; When the adjusted opening change amplitude is less than 0, the adjusted opening change amplitude is set to 0; Based on the gate opening and closing direction in the initial gate opening and closing action, the amplitude of the adjusted opening change is assigned a sign. When the gate opening and closing direction is the opening direction, it is a positive value; when the gate opening and closing direction is the closing direction, it is a negative value; and when the gate opening and closing direction is to maintain the current state, it is 0. The adjusted opening change amplitude after assigning a sign is taken as the final gate opening change, and combined with the gate opening and closing direction to obtain the final gate opening and closing action.

[0014] Optionally, the control execution module specifically comprises: Obtain the gate opening / closing direction and the final gate opening change during the final gate opening / closing action; The final gate opening change is added to the current gate opening value to obtain the target gate opening value; The target gate opening value is combined with the gate opening and closing direction to generate a gate opening and closing control command; The gate opening and closing control command is sent to the gate actuator to drive the gate actuator to perform opening and closing control.

[0015] The beneficial effects of this invention are: This invention addresses the problems of inconsistent timing of hydrological data, difficulty in characterizing oscillation characteristics, and reliance on experience in existing gate control systems. It introduces an improved Perceiver by modeling and analyzing real-time hydrological sequences. The IO model achieves joint acquisition of the perturbation evolution feature vector sequence and the interpretation field results, and enhances the ability to express the periodic characteristics of water level oscillation through the frequency field spectralization mechanism in the frequency field perception layer. Simultaneously, it constructs a latent variable perturbation probability field, an oscillation-sensitive region field, and an instantaneous water level oscillation gradient field in conjunction with the interpretation field generation layer, enabling comprehensive analysis of water level perturbations in terms of probability, energy, and changing trends. Furthermore, based on the manifold space of water level state characteristics, it quantifies the intensity of water level fluctuations through manifold curvature singularity dispersion analysis, transforming water level changes from a single time series analysis into a comprehensive analysis method with spatial structural characteristics. On this basis, the intensity of water level fluctuations is introduced into the gate opening and closing action adjustment process, enabling the control results to adaptively adjust with water level changes. This ensures adjustment efficiency when water levels are stable and suppresses oscillation amplification during severe fluctuations, effectively improving the accuracy and stability of water level perturbation identification and enhancing the adaptive capability and operational safety of the gate control strategy. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the gate real-time opening and closing strategy control system based on deep learning proposed in this invention. Figure 2 This is a schematic diagram of the improved Perceiver IO model structure proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 andFigure 2 The deep learning-based real-time gate opening and closing strategy control system includes the following modules: The data acquisition module is used to acquire multimodal hydrological operation data, perform time-series alignment, and generate real-time hydrological sequences. The feature modeling module is used to input real-time hydrological sequences into the improved Perceiver IO model. The improved Perceiver IO model includes an input interaction layer, a frequency field sensing layer, a perturbation coupling and recombination layer, and an interpretation field generation layer to obtain a perturbation evolution feature vector sequence and interpretation field results. The frequency field sensing layer has a built-in frequency field spectralization mechanism, and the interpretation field results include the latent variable perturbation probability field, the oscillation sensitive area field, and the instantaneous water level oscillation gradient field. The trend risk calculation module is used to determine the direction of water level change based on the perturbation evolution feature vector sequence, obtain the control trend value, and combine the latent variable perturbation probability field and the oscillation sensitive area field to obtain the oscillation risk value. The action generation module is used to generate preliminary gate opening and closing actions based on control trend values ​​and oscillation risk values; The manifold analysis module is used to map the perturbation evolution feature vector sequence to the water condition feature manifold space constructed by the instantaneous water level oscillation gradient field, perform manifold curvature singularity dispersion analysis, and obtain the degree of water condition fluctuation by calculating the curvature change and singularity distribution density of the current water condition. The action adjustment module is used to adjust the initial gate opening and closing action according to the intensity of water condition fluctuations, so as to obtain the final gate opening and closing action; The control execution module is used to generate gate opening and closing control commands based on the final gate opening and closing action, and to control the gate actuator to open and close.

[0019] In this embodiment, the data acquisition module specifically comprises: Collect multimodal hydrological operation data, including upstream water level data, downstream water level data, upstream flow data, downstream flow data, and gate opening data, and assign time stamps to each data. A unified sampling time sequence is set up, and the unified sampling time sequence is used as a time reference to perform time-series alignment processing on multimodal hydrological operation data; Under a unified sampling time sequence, for each sampling time, the corresponding upstream water level data, downstream water level data, inlet flow data, outlet flow data, and gate opening data are extracted to generate the corresponding hydrological data fragment for the sampling time. The hydrological data segments corresponding to each sampling time are arranged sequentially according to time to generate a real-time hydrological sequence; By applying unified time stamping and time-series alignment to multimodal hydrological operation data and constructing continuous hydrological data segments under a unified sampling time series, the time deviation and sampling inconsistency between different data sources are effectively eliminated, improving the temporal consistency and completeness of the input data. At the same time, by arranging the data in chronological order to form a real-time hydrological sequence, the improved Perceiver IO model can accurately characterize the dynamic evolution relationship between water level, flow rate, and gate opening, thereby improving the accuracy and stability of disturbance identification and control decisions.

[0020] In this embodiment, the input interaction layer is specifically as follows: The real-time hydrological sequence is input into a linear mapping structure, which maps each hydrological data segment into a hydrological feature vector of a uniform dimension. The corresponding time position encoding vector is generated according to the sequence number of each sampling time in the time series. The time position encoding vector has the same dimension as the corresponding hydrological feature vector. The time location encoding vector and the hydrological feature vector are added element by element along their corresponding dimensions to obtain the input feature vector, and an input feature vector sequence is constructed. According to the set window length and sliding step size, continuous subsequences are selected sequentially on the input feature vector sequence. The mean of the corresponding input feature vectors is calculated within each sliding time window to obtain aggregated feature vectors. The aggregated feature vectors are then used as latent vectors to form a latent array. The latent array is used as the query matrix. The input feature vector sequence is transformed into a key matrix and a value matrix through linear transformation. The query matrix and the key matrix are transposed and then multiplied to obtain the correlation matrix. Divide each element in the correlation matrix by the square root of the feature dimension of the key matrix, and input the result into the Softmax function for normalization calculation to obtain the attention weight matrix; The attention weight matrix and the value matrix are multiplied to obtain the latent feature sequence after interaction. Multiple feature update calculations based on the query matrix, key matrix, and value matrix are performed on the latent feature sequence after the interaction. After each feature update calculation, the updated latent feature sequence is added element by element to the corresponding input, and the addition result is processed by layer normalization according to the feature dimension. After completing all feature update calculations, the final latent representation sequence is output and then input into the frequency field sensing layer. In this invention, the input interaction layer first converts each hydrological data segment in the real-time hydrological sequence into a 128-dimensional hydrological feature vector through linear mapping. Then, it generates a corresponding time position encoding vector based on the sampling time's sequence number in the time series. This time position encoding vector also has a 128-dimensional dimension and is element-wise added to the corresponding hydrological feature vector. Subsequently, the input feature vector sequence is processed using a sliding time window with a window length of 8 and a sliding step size of 2. Within each sliding time window, the mean of 8 consecutive input feature vectors is calculated to obtain an aggregated feature vector, and a latent array containing 64 latent vectors is constructed. This latent array is then used as a query matrix to input the feature vectors. The vector sequence generates key and value matrices through three sets of linear transformations, each with a dimension of 128. A correlation matrix is ​​calculated by multiplying the query matrix with the transposed key matrix. Each element of the correlation matrix is ​​then divided by the square root of the key matrix's feature dimension, and the result is input into a Softmax function to calculate the attention weight matrix. This attention weight matrix is ​​then multiplied with the value matrix to obtain the interacting latent feature sequence. Subsequently, four feature update calculations are performed on the latent feature sequence. Each update is based on a new query matrix, key matrix, and value matrix. After each update, the result is added element-wise to the input, and then the result is normalized according to the feature dimension to obtain the final latent representation sequence. This structural configuration enables the input interaction layer to achieve efficient information compression and key feature extraction of multimodal hydrological sequences within a limited computational scale, improving the stability and accuracy of disturbance identification.

[0021] In this embodiment, the frequency field sensing layer receives the final latent representation sequence, executes the frequency field spectralization mechanism, divides the final latent representation sequence into multiple time segment sequences according to the time order, and performs discrete Fourier transform operation on each time segment sequence to convert the corresponding time segment sequence from the time domain representation into a frequency spectrum sequence in complex form. Each frequency component in the frequency spectrum sequence is represented as a complex number, where each frequency component includes a corresponding real part and an imaginary part. The real and imaginary parts of each frequency component are squared, and the summation and square root operations are performed on the squared results to obtain the amplitude of the corresponding frequency component. Based on the amplitudes of all frequency components within each time segment sequence, construct the corresponding amplitude spectrum sequence; The amplitudes of each frequency component in the amplitude spectrum sequence are accumulated to obtain the frequency perturbation intensity value of the corresponding time segment sequence; Based on the frequency perturbation intensity values ​​of all time segment sequences, a sequence of frequency perturbation intensity values ​​arranged in chronological order is obtained; The frequency perturbation intensity value sequence and the final latent representation sequence are concatenated along the feature dimension to generate the frequency-enhanced representation sequence; The perturbation coupling recombination layer receives the frequency enhancement representation sequence, performs a linear transformation operation on the frequency enhancement representation sequence, and performs element-wise multiplication with the frequency perturbation intensity value sequence to generate a perturbation coupling feature vector sequence corresponding to the frequency enhancement representation sequence in the time dimension. By performing a difference operation on adjacent perturbation coupling feature vectors along the time dimension, a change vector sequence is obtained; Based on the perturbation coupling feature vector sequence and the change vector sequence, the change vector corresponding to each time position is concatenated with the perturbation coupling feature vector of the corresponding time position to obtain the perturbation evolution feature vector of the corresponding time position. The perturbation evolution feature vectors corresponding to all time positions are arranged in chronological order to generate the perturbation evolution feature vector sequence. In this invention, the frequency field sensing layer receives the final latent representation sequence output by the input interaction layer and processes it based on the frequency field spectralization mechanism. Specifically, the final latent representation sequence is divided into a 16-time-segment sequence according to time order, and a discrete Fourier transform operation is performed on each time-segment sequence to obtain the corresponding complex frequency spectrum sequence. The real and imaginary parts of each frequency component are squared, and the sum of the squared results is taken as the square root to obtain the amplitude of the corresponding frequency component, thereby constructing an amplitude spectrum sequence. The amplitudes of all frequency components in each time segment are accumulated to obtain a frequency perturbation intensity value in the form of a single scalar. The same processing is performed on all time segments to form a frequency perturbation intensity value sequence arranged in time order. Subsequently, the frequency perturbation intensity value sequence and the final latent representation sequence are concatenated according to the feature dimension to generate a frequency-enhanced representation sequence.

[0022] The perturbation coupling recombination layer performs a 128×128 linear transformation on the frequency enhancement representation sequence and multiplies the transformation result element-wise with the frequency perturbation intensity value sequence, so that the features at each time step are modulated by the corresponding frequency perturbation intensity. Furthermore, a first-order difference operation is performed on adjacent perturbation coupling feature vectors to obtain a change vector sequence. This change vector sequence is then combined with the perturbation coupling feature vector sequence to obtain the perturbation evolution feature vector sequence. Through this processing, frequency information participates in temporal feature modulation in scalar intensity form, avoiding the introduction of redundant frequency domain features and improving the stability and response sensitivity of perturbation identification.

[0023] In this embodiment, the interpretation field generation layer receives the perturbation evolution feature vector sequence, inputs the perturbation evolution feature vector corresponding to each time position into the fully connected linear transformation structure, performs the Sigmoid function operation on the output result of the fully connected linear transformation structure to obtain the latent variable perturbation probability value corresponding to each time position, and arranges the latent variable perturbation probability values ​​corresponding to all time positions in chronological order to form the latent variable perturbation probability field. The perturbation evolution feature vectors corresponding to each time position are squared element by element, and the squared results are accumulated element by element to obtain the energy value corresponding to each time position. The energy values ​​corresponding to all time locations are summed and divided by the total number of time locations to obtain the global energy baseline value. The energy values ​​corresponding to time positions with energy values ​​greater than the global energy reference value are retained, while the energy values ​​corresponding to time positions with energy values ​​less than or equal to the global energy reference value are set to 0, and arranged in chronological order to form an oscillation-sensitive region field. Perform element-wise subtraction on the perturbation evolution feature vectors at adjacent time positions in the perturbation evolution feature vector sequence to obtain the difference vector; After taking the absolute value of each element in the difference vector, an accumulation operation is performed to obtain the change intensity value corresponding to each time position; Subtract the change intensity values ​​corresponding to adjacent time positions to obtain the change intensity gradient value, and arrange all change intensity gradient values ​​in time order to form the instantaneous water level oscillation gradient field; The hidden variable perturbation probability field, the oscillation sensitive region field, and the instantaneous water level oscillation gradient field are output as the explanatory field results; In this invention, the interpretation field generation layer directly uses the perturbation evolution feature vector sequence as input. Through a fully connected linear transformation combined with the Sigmoid function, it calculates the latent variable perturbation probability field, achieving a continuous and quantitative expression of the probability of perturbation occurrence at each time point. Simultaneously, by performing element-wise squaring and summation operations on the perturbation evolution feature vectors, it obtains energy values ​​and compares them with the global energy benchmark value, constructing an oscillation-sensitive region field. This accurately reflects the key time points of abnormal energy concentration in the system. Furthermore, by performing difference operations and calculating the gradient value of change intensity on the perturbation evolution feature vectors of adjacent time points, it forms an instantaneous water level oscillation gradient field, used to characterize the severity and trend of water level changes. These three types of interpretation fields are generated on a unified data basis, avoiding complex intermediate modeling processes and improving computational stability and feasibility. At the same time, each interpretation field characterizes water disturbances from three dimensions: probability, energy, and gradient, enhancing the multi-dimensional expression of gate operation status, thereby improving the accuracy and reliability of opening and closing strategy decisions.

[0024] The improved Perceiver IO model in this invention inherits the core idea of ​​the existing Perceiver IO model in its overall architecture. That is, it compresses information and extracts features from high-dimensional input sequences by constructing a latent array and using the interaction mechanism between queries, keys, and values. It still uses a cross-attention mechanism to realize the information interaction between the input feature vector sequence and the latent vector, and gradually optimizes the latent representation through multiple feature updates.

[0025] Building upon this foundation, this invention makes targeted improvements to address the dynamic characteristics of hydrological time-series data: In the input interaction stage, a mean aggregation mechanism based on a sliding time window is introduced to construct a latent array, allowing the latent vectors to directly originate from the local statistical information of the original time-series features, thus enhancing the ability to characterize temporal continuity. In the model, a frequency field sensing layer is used to introduce a frequency field spectralization mechanism to perform a discrete Fourier transform on the latent representation sequence and extract the frequency perturbation intensity, achieving explicit modeling of the periodic characteristics of water level oscillations. Furthermore, a perturbation coupling and recombination layer couples the frequency perturbation intensity with the time-domain features element-wise, and combines this with difference operations to construct a perturbation evolution feature vector sequence, enabling the model to simultaneously express both time-domain changes and frequency-domain perturbations. Finally, an interpretation field generation layer is introduced at the model output to construct a latent variable perturbation probability field, an oscillation-sensitive region field, and an instantaneous water level oscillation gradient field, achieving a structured interpretation of the model results.

[0026] Through the above improvements, the model is transformed from a simple feature encoding structure into a multi-layer linkage structure with time domain, frequency domain, and interpretation field capabilities. This not only improves the identification accuracy of complex water disturbances but also enhances the interpretability and engineering applicability of the model output, thereby improving the reliability and stability of the gate opening and closing control strategy.

[0027] In this embodiment, the trend risk calculation module specifically includes: Perform element-wise subtraction on the perturbation evolution feature vectors at adjacent time positions in the perturbation evolution feature vector sequence to obtain the difference vector; Perform element-wise summation on each element in the difference vector to obtain the direction change value corresponding to each time position. Determine the direction of water level change based on the sign of the direction change value. Time positions with a direction change value greater than 0 are determined to be water level rising and assigned a value of 1. Time positions with a direction change value equal to 0 are determined to be water level remaining flat and assigned a value of 0. Time positions with a direction change value less than 0 are determined to be water level falling and assigned a value of -1, thus obtaining the water level change direction sequence. The values ​​at each time position in the sequence of water level change directions are accumulated to obtain the cumulative value of the direction; Divide the cumulative direction value by the total number of time positions to obtain the control trend value; The probability values ​​of hidden variable perturbations at each time position in the probability field of hidden variable perturbations are multiplied element-wise with the energy values ​​at the corresponding time positions in the field of oscillation sensitive region to obtain the oscillation intensity values ​​at the corresponding time positions. Based on the oscillation intensity values ​​at all time points, an oscillation intensity value sequence is constructed; Perform an element-by-element summation operation on the oscillation intensity value sequence to obtain the oscillation risk value; In this invention, by performing element-wise differencing and accumulation operations on the perturbation evolution feature vector sequence, the directional change values ​​at each time position are extracted. Based on the positive and negative relationships of the directional change values, the water level change state is uniformly mapped into a numerical water level change direction sequence, thereby achieving a quantitative expression of the water level change trend. Furthermore, by accumulating the water level change direction sequence, a directional accumulation value is obtained. Dividing the directional accumulation value by the total number of time positions yields the control trend value, transforming the overall water level change trend from a discrete state into a continuously calculable indicator, improving the stability and consistency of trend judgment. Simultaneously, the oscillation intensity value is obtained by combining the latent variable perturbation probability field and the oscillation sensitive area field. By performing element-wise accumulation operations on the oscillation intensity value sequence, an oscillation risk value is generated, achieving a coupled measurement of perturbation probability and energy concentration. Through the above processing, the water level change trend and oscillation risk are quantitatively expressed within a unified computational framework, avoiding the problem of relying on empirical thresholds in traditional methods, improving the objectivity and repeatability of water condition judgment, and thus enhancing the accuracy and reliability of gate opening and closing strategies.

[0028] In this embodiment, the action generation module specifically comprises: The gate opening and closing direction is determined by the sign of the control trend value. When the control trend value is greater than 0, it is determined to be the opening direction; when the control trend value is equal to 0, it is determined to maintain the current state; and when the control trend value is less than 0, it is determined to be the closing direction. The absolute value of the control trend value is taken as the basic opening change. Subtract the base opening change from the oscillation risk value to obtain the adjusted opening change. When the change in opening is less than 0, set the change in opening to 0. The change in gate opening is assigned a sign according to the gate opening and closing direction. When the gate opening and closing direction is the opening direction, the change in gate opening is positive; when the gate opening and closing direction is the closing direction, the change in gate opening is negative; and when the gate opening and closing direction is to maintain the current state, the change in gate opening is 0. The marked adjustment opening change is used as the gate opening change to generate the initial gate opening and closing action. The initial gate opening and closing action includes the gate opening and closing direction and the gate opening change. In this invention, a gate opening and closing action is generated using a control trend value and an oscillation risk value as inputs. The gate opening and closing direction is determined based on the sign of the control trend value, thus identifying the direction of water level change trends. Subsequently, the absolute value of the control trend value is taken to obtain the basic gate opening change, which is then subtracted from the oscillation risk value to obtain the adjusted gate opening change, thereby incorporating water level oscillation risk into the gate opening adjustment process. Furthermore, by applying a non-negative constraint to the adjusted gate opening change and assigning a positive or negative sign based on the opening and closing direction, the gate opening change possesses both directionality and amplitude. Finally, the combination of the gate opening and closing direction and the gate opening change forms a well-defined and directly executable preliminary gate opening and closing action. Through this process, a unified expression of trend-driven and risk-suppressive mechanisms is achieved, avoiding the reliance on empirical rules in traditional methods and improving the stability and adaptability of gate control.

[0029] In this embodiment, the manifold analysis module specifically comprises: The change intensity gradient value corresponding to each time position in the instantaneous water level oscillation gradient field is concatenated with the time index of the corresponding time position to construct the water condition characteristic manifold space; The perturbation evolution feature vector corresponding to each time position in the perturbation evolution feature vector sequence is multiplied element-wise with the change intensity gradient value at the corresponding time position to obtain the mapping feature vector. All the mapping feature vectors are arranged in chronological order to form a trajectory feature sequence in the hydrological state feature manifold space. Perform manifold curvature singularity dispersion analysis on the trajectory feature sequence, and perform element-wise subtraction on the mapping feature vectors of adjacent time positions to obtain the first-order difference vector; The element-wise subtraction operation is performed again on adjacent time positions in the first-order difference vector to obtain the second-order difference vector. The absolute value of each element in the second-order difference vector is then taken and accumulated to obtain the curvature value corresponding to each time position. The curvature values ​​corresponding to all time positions are summed and then divided by the total number of time positions to obtain the degree of curvature change. The maximum value of curvature values ​​corresponding to all time positions is taken as the curvature reference value, and the degree of curvature change is divided by the curvature reference value to obtain the curvature normalization value; Count the time locations where the curvature value is greater than the degree of curvature change, and divide the count by the total number of time locations to obtain the singularity distribution density; Adding the curvature normalization value to the singularity distribution density yields the degree of intensity of water condition fluctuations; In this invention, based on the instantaneous water level oscillation gradient field, a water condition characteristic manifold space is constructed by concatenating the gradient value of the change intensity with the time index, transforming the water condition change from a single time series into an expression with spatial structure characteristics. Furthermore, the perturbation evolution feature vector is multiplied element-wise with the corresponding gradient value to achieve gradient modulation of the original perturbation information, forming a trajectory feature sequence, thereby characterizing the water condition evolution path in the water condition characteristic manifold space. By performing first-order and second-order difference operations on the trajectory feature sequence, curvature values ​​are obtained and the degree of curvature change is calculated. Simultaneously, the maximum curvature value is selected as the curvature benchmark value for normalization, eliminating the influence of differences in curvature dimensions at different time scales. Based on this, the singularity distribution density is obtained by statistically analyzing the proportion of time positions where curvature exceeds the overall level, thus characterizing the concentration of localized violent fluctuations. Finally, the normalized curvature value is added to the singularity distribution density to obtain the degree of intensity of water condition fluctuations. The above methods enable a unified quantitative expression of the overall trend and local anomalies of water condition changes, improve the ability to characterize complex nonlinear water level fluctuations, and thus enhance the response capability and control accuracy of gate control strategies to sudden oscillations.

[0030] In this embodiment, the motion adjustment module specifically comprises: The absolute value of the gate opening change during the initial gate opening and closing action is taken to obtain the initial opening change amplitude; Multiply the initial opening change amplitude by the severity of the water level fluctuation to obtain the fluctuation impact. Subtract the fluctuation effect from the initial opening change amplitude to obtain the adjusted opening change amplitude; When the adjusted opening change amplitude is less than 0, set the adjusted opening change amplitude to 0; Based on the gate opening and closing direction in the initial gate opening and closing action, assign a sign to the amplitude of the adjusted opening degree change. When the gate opening and closing direction is the opening direction, it is a positive value; when the gate opening and closing direction is the closing direction, it is a negative value; and when the gate opening and closing direction is to maintain the current state, it is 0. The adjusted opening change amplitude after assigning a sign is taken as the final gate opening change, and combined with the gate opening and closing direction to obtain the final gate opening and closing action. In this invention, the intensity of water level fluctuations is used as an adjustment factor to continuously correct the opening change in the initial gate opening and closing action. The fluctuation impact is obtained by multiplying the initial opening change amplitude with the intensity of water level fluctuations, and then adaptively reducing the opening change amplitude through subtraction, allowing the gate adjustment range to dynamically change with the intensity of water level fluctuations. Furthermore, non-negative constraint processing prevents abnormal reverse amplification of the opening change amplitude, and the adjustment result is assigned a sign based on the gate opening and closing direction, achieving an integrated expression of direction and amplitude. The final gate opening change and opening / closing direction together constitute the final gate opening and closing action. Through these processes, the gate control maintains a large adjustment capacity when the water level is stable, and automatically converges the adjustment amplitude during periods of severe water level fluctuations, thereby effectively suppressing the risk of oscillation amplification and improving the stability and safety of the system operation.

[0031] In this embodiment, the control execution module specifically comprises: Obtain the gate opening / closing direction and the final gate opening change during the final gate opening / closing action; The final gate opening change is added to the current gate opening value to obtain the target gate opening value; The target gate opening value is combined with the gate opening and closing direction to generate a gate opening and closing control command; The gate opening and closing control command is sent to the gate actuator to drive the gate actuator to perform opening and closing control.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to a gate water level control scenario. In this scenario, the gate is mainly used to regulate the upstream and downstream water levels and flow distribution. During actual operation, the incoming water exhibits obvious unstable characteristics, with large flow fluctuations in a short period, leading to frequent oscillations of rising, falling, and then rising water levels. Traditional control methods typically employ rule-based control based on water level thresholds. That is, when the water level is higher than the set upper limit, the gate is opened; when the water level is lower than the set lower limit, the gate is closed, and adjustments are made according to a fixed opening adjustment amount. This method is simple in structure, but it cannot identify water level change trends and potential disturbance risks, and is prone to repeated adjustments when the water level approaches the threshold, leading to amplified oscillations and frequent gate opening and closing.

[0033] This embodiment applies the system of the present invention to a gate water level control scenario. The system acquires upstream and downstream water level data, upstream and downstream flow data, and gate opening data through a data acquisition module. These data are then processed with unified time stamping and time sequence alignment to form a continuous real-time hydrological sequence. Subsequently, the real-time hydrological sequence is input into a feature modeling module, which improves the perceiver... The input interaction layer, frequency field perception layer, disturbance coupling and reorganization layer, and interpretation field generation layer in the IO model are processed to obtain the disturbance evolution feature vector sequence, and generate the latent variable disturbance probability field, oscillation sensitive region field, and instantaneous water level oscillation gradient field. Based on this, the trend risk calculation module calculates the control trend value based on the disturbance evolution feature vector sequence, and obtains the oscillation risk value by combining the latent variable disturbance probability field and the oscillation sensitive region field. The action generation module generates the initial gate opening and closing action based on the control trend value and the oscillation risk value. Subsequently, the manifold analysis module maps the disturbance evolution feature vector sequence to the water condition characteristic manifold space, and obtains the intensity of water condition fluctuations through manifold curvature singularity dispersion analysis. The action adjustment module corrects the initial gate opening and closing action according to the intensity of water condition fluctuations. Finally, the control execution module generates the gate opening and closing control command to achieve fine control of the gate actuator.

[0034] In actual operation, data from multiple consecutive scheduling cycles were selected for statistical analysis. Under the same inflow conditions, a comparative experiment was conducted between the rule-based control method based on water level thresholds and the method of this invention. The results are shown in Table 1 below.

[0035] Table 1 Comparative Analysis of Gate Control Performance As can be seen from the data in Table 1 above, compared with the rule-based control method based on water level thresholds, the system of this invention shows significant advantages in several key operational indicators. Regarding water level control effectiveness, the maximum water level fluctuation amplitude decreased from 0.46m to 0.26m, and the average water level fluctuation amplitude decreased from 0.30m to 0.17m, indicating that this invention can effectively suppress large fluctuations in water level, making water condition changes more stable and avoiding the oscillation amplification problem caused by lag adjustment in traditional methods. In terms of control behavior, the number of gate actions decreased from 17 to 8, and the maximum number of continuous adjustments decreased from 6 to 3, indicating that this invention significantly reduces frequent opening and closing phenomena, making the adjustment process more stable and continuous, which helps reduce equipment wear and improve operational reliability. Meanwhile, in terms of adjustment precision, the average opening adjustment amplitude decreased from 0.12m to 0.08m, and the standard deviation of opening change decreased from 0.095 to 0.052, indicating that this invention can achieve smaller amplitude and more uniform opening changes during the adjustment process, avoiding the abrupt changes caused by fixed adjustment amounts in traditional methods.

[0036] This embodiment demonstrates that the present invention, by constructing an improved Perceiver IO model, achieves unified modeling of multimodal hydrological operation data. By combining a frequency field sensing layer and a disturbance coupling and reorganization layer, it collaboratively expresses the time-domain and frequency-domain characteristics of water level changes, thereby accurately characterizing the evolution of hydrological disturbances. Simultaneously, by constructing a latent variable disturbance probability field, an oscillation-sensitive region field, and an instantaneous water level oscillation gradient field through an interpretation field generation layer, it achieves a multi-dimensional quantitative description of the hydrological state, enabling the system to possess a refined perception capability of water level change trends and oscillation risks. Furthermore, through the linkage of the trend risk calculation module and the manifold analysis module, hydrological changes are transformed from traditional experience-based judgments into calculable structured indicators. Adaptive adjustment of gate opening and closing actions is achieved through the action generation module and action adjustment module, ensuring that the control process simultaneously considers trend response and oscillation suppression. Thus, gate control shifts from a single threshold-driven approach to a trend and risk-driven approach, not only improving the stability and continuity of water level control but also enhancing the system's adaptability to complex nonlinear hydrological changes, thereby improving overall control accuracy and operational reliability.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A gate real-time opening and closing strategy control system based on deep learning, characterized in that, Includes the following modules: The data acquisition module is used to acquire multimodal hydrological operation data, perform time-series alignment, and generate real-time hydrological sequences. The feature modeling module is used to input real-time hydrological sequences into the improved Perceiver IO model. The improved Perceiver IO model includes an input interaction layer, a frequency field sensing layer, a perturbation coupling and recombination layer, and an interpretation field generation layer to obtain a perturbation evolution feature vector sequence and interpretation field results. The frequency field sensing layer has a built-in frequency field spectralization mechanism. The interpretation field results include a latent variable perturbation probability field, an oscillation sensitive region field, and an instantaneous water level oscillation gradient field. The trend risk calculation module is used to determine the direction of water level change based on the perturbation evolution feature vector sequence, obtain the control trend value, and combine the latent variable perturbation probability field and the oscillation sensitive area field to obtain the oscillation risk value. The action generation module is used to generate preliminary gate opening and closing actions based on control trend values ​​and oscillation risk values; The manifold analysis module is used to map the perturbation evolution feature vector sequence to the water condition feature manifold space constructed by the instantaneous water level oscillation gradient field, perform manifold curvature singularity dispersion analysis, and obtain the degree of water condition fluctuation by calculating the curvature change and singularity distribution density of the current water condition. The action adjustment module is used to adjust the initial gate opening and closing action according to the intensity of water condition fluctuations, so as to obtain the final gate opening and closing action; The control execution module is used to generate gate opening and closing control commands based on the final gate opening and closing action, and to control the gate actuator to open and close.

2. The gate real-time opening and closing strategy control system based on deep learning according to claim 1, characterized in that, The data acquisition module is specifically: Collect multimodal hydrological operation data, which includes upstream water level data, downstream water level data, upstream flow data, downstream flow data, and gate opening data, and assign time stamps to each data. A unified sampling time sequence is set, and the multimodal hydrological operation data is time-aligned using the unified sampling time sequence as a time reference. Under the unified sampling time sequence, for each sampling time, the corresponding upstream water level data, downstream water level data, inlet flow data, outlet flow data, and gate opening data are extracted to generate the corresponding hydrological data fragment for the sampling time. The hydrological data segments corresponding to each sampling time are arranged sequentially according to time to generate a real-time hydrological sequence.

3. The gate real-time opening and closing strategy control system based on deep learning according to claim 1, characterized in that, The input interaction layer is specifically as follows: The real-time hydrological sequence is input into a linear mapping structure, and each hydrological data segment is mapped to a hydrological feature vector of a uniform dimension. A corresponding time position encoding vector is generated according to the sequence number of each sampling time in the time series. The time position encoding vector has the same dimension as the corresponding hydrological feature vector. The time location encoding vector and the hydrological feature vector are added element by element along their corresponding dimensions to obtain the input feature vector, and an input feature vector sequence is constructed. According to the set window length and sliding step size, continuous subsequences are selected sequentially on the input feature vector sequence. The mean of the corresponding input feature vectors is calculated within each sliding time window to obtain aggregated feature vectors. The aggregated feature vectors are then used as latent vectors to form a latent array. Using the latent array as a query matrix, the input feature vector sequence is transformed into a key matrix and a value matrix through linear transformation. The query matrix and the key matrix are transposed and then multiplied to obtain the correlation matrix. Divide each element in the correlation matrix by the square root of the feature dimension of the key matrix, and input the result into the Softmax function for normalization calculation to obtain the attention weight matrix; The attention weight matrix and the value matrix are multiplied to obtain the latent feature sequence after interaction. Multiple feature update calculations based on the query matrix, key matrix, and value matrix are performed on the latent feature sequence after the interaction. After each feature update calculation, the updated latent feature sequence is added element by element to the corresponding input, and the addition result is subjected to layer normalization processing according to the feature dimension. After completing all feature update calculations, the final latent representation sequence is output and then input into the frequency field sensing layer.

4. The gate real-time opening and closing strategy control system based on deep learning according to claim 1, characterized in that, The frequency field sensing layer receives the final latent representation sequence, executes the frequency field spectralization mechanism, divides the final latent representation sequence into multiple time segment sequences according to time order, and performs discrete Fourier transform operation on each time segment sequence to convert the corresponding time segment sequence from time domain representation into a frequency spectrum sequence in complex form. Each frequency component in the frequency spectrum sequence is represented in complex form, wherein each frequency component includes a corresponding real part and an imaginary part. The real and imaginary parts of each frequency component are squared, and the summation and square root operations are performed on the squared results to obtain the amplitude of the corresponding frequency component. Based on the amplitudes of all frequency components within each time segment sequence, construct the corresponding amplitude spectrum sequence; The amplitude of each frequency component in the amplitude spectrum sequence is accumulated to obtain the frequency perturbation intensity value of the corresponding time segment sequence; Based on the frequency perturbation intensity values ​​of all time segment sequences, a sequence of frequency perturbation intensity values ​​arranged in chronological order is obtained; The frequency perturbation intensity value sequence is concatenated with the final potential representation sequence according to the feature dimension to generate a frequency-enhanced representation sequence; The perturbation coupling recombination layer receives the frequency enhancement representation sequence, performs a linear transformation operation on the frequency enhancement representation sequence, and performs element-wise multiplication with the frequency perturbation intensity value sequence to generate a perturbation coupling feature vector sequence corresponding to the frequency enhancement representation sequence in the time dimension. The perturbation coupling feature vector sequence is subjected to a difference operation along the time dimension by adjacent perturbation coupling feature vectors to obtain a change vector sequence; Based on the perturbation coupling feature vector sequence and the change vector sequence, the change vector corresponding to each time position is concatenated with the perturbation coupling feature vector of the corresponding time position to obtain the perturbation evolution feature vector of the corresponding time position. The perturbation evolution feature vectors corresponding to all time positions are arranged in chronological order to generate the perturbation evolution feature vector sequence.

5. The gate real-time opening and closing strategy control system based on deep learning according to claim 1, characterized in that, The interpretation field generation layer receives the perturbation evolution feature vector sequence, inputs the perturbation evolution feature vectors corresponding to each time position into the fully connected linear transformation structure, performs the Sigmoid function operation on the output result of the fully connected linear transformation structure to obtain the latent variable perturbation probability value corresponding to each time position, and arranges the latent variable perturbation probability values ​​corresponding to all time positions in chronological order to form the latent variable perturbation probability field. The perturbation evolution feature vectors corresponding to each time position are squared element by element, and the squared results are accumulated element by element to obtain the energy value corresponding to each time position. The energy values ​​corresponding to all time locations are summed and divided by the total number of time locations to obtain the global energy baseline value. The energy values ​​corresponding to time positions with energy values ​​greater than the global energy reference value are retained, while the energy values ​​corresponding to time positions with energy values ​​less than or equal to the global energy reference value are set to 0, and arranged in chronological order to form an oscillation-sensitive region field. Perform element-wise subtraction on the perturbation evolution feature vectors at adjacent time positions in the perturbation evolution feature vector sequence to obtain the difference vector; The absolute value of each element in the difference vector is taken and then accumulated to obtain the change intensity value corresponding to each time position; Subtract the change intensity values ​​corresponding to adjacent time positions to obtain the change intensity gradient value, and arrange all change intensity gradient values ​​in time order to form the instantaneous water level oscillation gradient field; The hidden variable perturbation probability field, the oscillation sensitive region field, and the instantaneous water level oscillation gradient field are output as the explanatory field results.

6. The gate real-time opening and closing strategy control system based on deep learning according to claim 1, characterized in that, The trend risk calculation module is specifically as follows: Perform element-wise subtraction on the perturbation evolution feature vectors at adjacent time positions in the perturbation evolution feature vector sequence to obtain the difference vector; An element-wise summation operation is performed on each element in the difference vector to obtain the direction change value corresponding to each time position. The direction of water level change is determined according to the sign of the direction change value. The time position where the direction change value is greater than 0 is determined to be a rising water level and assigned a value of 1. The time position where the direction change value is equal to 0 is determined to be a level water level and assigned a value of 0. The time position where the direction change value is less than 0 is determined to be a falling water level and assigned a value of -1, thus obtaining a sequence of water level change directions. The values ​​at each time position in the sequence of water level change directions are accumulated to obtain the cumulative value of the direction; Divide the cumulative value of the direction by the total number of time positions to obtain the control trend value; The probability values ​​of hidden variable perturbations at each time position in the probability field of hidden variable perturbations are multiplied element-wise with the energy values ​​at the corresponding time positions in the field of oscillation sensitive region to obtain the oscillation intensity values ​​at the corresponding time positions. Based on the oscillation intensity values ​​at all time points, an oscillation intensity value sequence is constructed; An element-by-element summation operation is performed on the oscillation intensity value sequence to obtain the oscillation risk value.

7. The gate real-time opening and closing strategy control system based on deep learning according to claim 1, characterized in that, The action generation module is specifically: The gate opening and closing direction is determined based on the sign of the control trend value. When the control trend value is greater than 0, it is determined to be the opening direction; when the control trend value is equal to 0, it is determined to maintain the current state; and when the control trend value is less than 0, it is determined to be the closing direction. The absolute value of the control trend value is taken as the basic opening change. The adjustment amount of the opening change is obtained by subtracting the basic opening change from the oscillation risk value. When the change in the adjustment opening is less than 0, the change in the adjustment opening is set to 0; The change in gate opening is assigned a sign according to the gate opening and closing direction. When the gate opening and closing direction is the opening direction, the change in gate opening is positive; when the gate opening and closing direction is the closing direction, the change in gate opening is negative; and when the gate opening and closing direction is to maintain the current state, the change in gate opening is 0. The marked adjustment opening change is used as the gate opening change to generate the preliminary gate opening and closing action, which includes the gate opening and closing direction and the gate opening change.

8. The gate real-time opening and closing strategy control system based on deep learning according to claim 1, characterized in that, The manifold analysis module specifically includes: The change intensity gradient value corresponding to each time position in the instantaneous water level oscillation gradient field is concatenated with the time index of the corresponding time position to construct the water condition characteristic manifold space. The perturbation evolution feature vector corresponding to each time position in the perturbation evolution feature vector sequence is multiplied element-wise with the change intensity gradient value at the corresponding time position to obtain the mapping feature vector. All the mapping feature vectors are arranged in chronological order to form the trajectory feature sequence in the hydrological state feature manifold space. Perform manifold curvature singularity dispersion analysis on the trajectory feature sequence, and perform element-wise subtraction on the mapping feature vectors of adjacent time positions to obtain the first-order difference vector; The element-wise subtraction operation is performed again on adjacent time positions in the first-order difference vector to obtain the second-order difference vector. The absolute value of each element in the second-order difference vector is then taken and accumulated to obtain the curvature value corresponding to each time position. The curvature values ​​corresponding to all time positions are summed and then divided by the total number of time positions to obtain the degree of curvature change. The maximum value among all curvature values ​​corresponding to all time positions is taken as the curvature reference value, and the degree of curvature change is divided by the curvature reference value to obtain the curvature normalization value; Count the time locations where the curvature value is greater than the degree of curvature change, and divide the count by the total number of time locations to obtain the singularity distribution density; The degree of water condition fluctuation is obtained by adding the normalized curvature value to the singularity distribution density.

9. The gate real-time opening and closing strategy control system based on deep learning according to claim 1, characterized in that, The motion adjustment module is specifically as follows: The absolute value of the gate opening change during the initial gate opening and closing action is taken to obtain the initial opening change amplitude; Multiply the initial opening change amplitude by the severity of the water condition fluctuation to obtain the fluctuation impact. Subtract the fluctuation effect from the initial opening change amplitude to obtain the adjusted opening change amplitude; When the adjusted opening change amplitude is less than 0, the adjusted opening change amplitude is set to 0; Based on the gate opening and closing direction in the initial gate opening and closing action, the amplitude of the adjusted opening change is assigned a sign. When the gate opening and closing direction is the opening direction, it is a positive value; when the gate opening and closing direction is the closing direction, it is a negative value; and when the gate opening and closing direction is to maintain the current state, it is 0. The adjusted opening change amplitude after assigning a sign is taken as the final gate opening change, and combined with the gate opening and closing direction to obtain the final gate opening and closing action.

10. The gate real-time opening and closing strategy control system based on deep learning according to claim 1, characterized in that, The control execution module is specifically: Obtain the gate opening / closing direction and the final gate opening change during the final gate opening / closing action; The final gate opening change is added to the current gate opening value to obtain the target gate opening value; The target gate opening value is combined with the gate opening and closing direction to generate a gate opening and closing control command; The gate opening and closing control command is sent to the gate actuator to drive the gate actuator to perform opening and closing control.

Images