Intelligent flood gate automatic control system based on deep learning

Abstract

The invention discloses an intelligent flood gate automatic control system based on deep learning, and the system comprises a data collection module which is used for obtaining and preprocessing multi-source monitoring data; the feature extraction module is used for performing multi-layer time sequence convolution and residual connection; the variable relation modeling module is used for constructing a dynamic graph adjacency matrix, performing graph convolution operation and updating a variable dependency relation; the time sequence modeling module is used for transmitting and updating a hidden state in a time step dimension, calculating correlation and performing weighted aggregation; the control instruction module is used for performing weighted fusion and analysis on the flood routing trend prediction result and the real-time monitoring signal; and the execution control module is used for driving the gate and the flood control baffle according to the automatic control instruction and executing operation by combining signals of the front and rear laser correlation sensors and remote control or manual input. According to the method, deep learning modeling and intelligent control are fused, flood gate autonomous scheduling is achieved, and the method has the advantages of being accurate in prediction, safe, reliable and efficient in response.

Description

Technical Field

[0001] This invention relates to the field of flood control automation control, and in particular to an intelligent floodgate automatic control system based on deep learning. Background Technology

[0002] Existing floodgate control systems mostly rely on single or limited monitoring data such as water level, rainfall, and flow velocity, combined with empirical formulas or simple logical rules to generate gate opening and closing commands. While these methods can accomplish basic flood control scheduling to a certain extent, they lack the ability to dynamically model complex flood evolution processes. They struggle to accurately predict flood trends in complex scenarios such as sudden heavy rainfall and uneven upstream inflows, potentially leading to delayed or inappropriate gate control. Some studies have attempted to introduce machine learning methods, but most remain at the shallow feature extraction stage, failing to effectively utilize the temporal correlations and variable dependencies of multi-source monitoring data.

[0003] Meanwhile, existing automated control systems also have shortcomings in terms of execution safety and control flexibility. For example, the lack of a real-time detection mechanism for pedestrians or obstacles poses safety hazards during execution; manual intervention is usually independent of automatic control, lacking a combined coordination mechanism, which reduces the flexibility of emergency response.

[0004] Therefore, how to provide an intelligent floodgate automatic 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

[0005] One objective of this invention is to propose an intelligent floodgate automatic control system based on deep learning. This invention integrates deep learning modeling and intelligent control to achieve autonomous scheduling of floodgates, and has the advantages of accurate prediction, safety and reliability, and high response efficiency.

[0006] According to an embodiment of the present invention, an intelligent floodgate automatic control system based on deep learning includes: The data acquisition module is used to acquire and preprocess multi-source monitoring data in real time. The feature extraction module is used to perform multi-layer temporal convolution and residual connections on standardized monitoring data; The variable relationship modeling module is used to construct a dynamic graph adjacency matrix, perform graph convolution operations on high-dimensional semantic feature sequences, and update variable dependencies. The temporal modeling module is used to pass and update the hidden state of the enhanced feature sequence along the time step dimension, and to calculate the correlation and weighted aggregation by combining the temporal attention mechanism; The control command module is used to perform weighted fusion and analysis of flood evolution trend prediction results and real-time monitoring signals; The execution control module is used to drive the gate and flood control barrier according to the automatic control command, and to perform the operation in combination with the front and rear laser beam sensor signals and remote control or manual input.

[0007] Optionally, modules can be integrated using the following methods: Real-time acquisition of multi-source monitoring data, preprocessing, and generation of standardized monitoring data; Standardized monitoring data is constructed into a time series tensor, which is then input into the multi-layer temporal convolutional unit of the improved MTGNN to perform multi-scale convolution and residual connections, generating a high-dimensional semantic feature sequence. A dynamic graph adjacency matrix is ​​constructed, graph convolution is performed on the high-dimensional semantic feature sequence, and variable dependencies are updated to generate an enhanced feature sequence. The enhanced feature sequence is input into the gated recurrent unit network of the improved MTGNN, which is passed and updated in the time step dimension. The correlation is calculated and weighted aggregation is combined with the time attention mechanism to generate the flood evolution trend prediction result. The flood evolution trend prediction results are weighted, fused and analyzed with real-time monitoring signals to generate automatic control commands and drive the gates and flood control barriers to complete the corresponding actions. During the gate's operation, signals from the front and rear laser beam sensors are collected. When a pedestrian or obstacle is detected, a pause command is output. Remote control or manual input is received to generate human control commands, which are then combined with automatic control commands to generate composite control commands and drive the gate and flood control barrier to perform corresponding operations.

[0008] Optionally, the multi-source monitoring data includes rainfall data, water level data, flow velocity data, and meteorological data, and the preprocessing includes formatting, outlier removal, missing value imputation, and normalization of the multi-source monitoring data.

[0009] Optionally, the generation of the high-dimensional semantic feature sequence includes the following specific steps: The standardized monitoring data are arranged in chronological order. The multi-source monitoring data collected at the same time step are combined into a vector. The vectors corresponding to each time step are stacked in chronological order to obtain a time series tensor. The time series tensor is input into the multi-layer temporal convolutional unit of the improved MTGNN. The multi-scale convolutional kernel group is called to perform one-dimensional convolution operation on the time series tensor along the time step dimension. Within each time window, the value of the monitored variable in the corresponding time window is multiplied by the corresponding convolutional kernel weight parameter and summed. The output of the summation operation is used as the convolution output value of the corresponding time window. The process is repeated for all time steps of the entire time series tensor to obtain the convolution output sequence covering different convolutional kernel scales. All convolution output sequences are concatenated to obtain the convolutional feature sequence. Residual connections are added to each temporal convolutional unit. The input features and output convolutional feature sequences of the corresponding temporal convolutional unit are added point by point at the corresponding time step and the monitoring variable position to generate a residual enhancement sequence. The residual connections are implemented by setting a bypass in the temporal convolutional unit. The residual enhancement features at each time step are stacked as tensors and arranged and combined along the dimension of the monitored variables to generate a high-dimensional semantic feature sequence.

[0010] Optionally, compared with the original MTGNN model, the improved MTGNN introduces multi-layer temporal convolutional units and adds residual connections in the feature extraction stage to perform convolution operations and stacking on standardized monitoring data to generate high-dimensional semantic feature sequences. In the variable relationship modeling stage, a dynamic graph adjacency matrix is ​​constructed to perform graph convolution operations on the high-dimensional semantic feature sequences and update variable dependencies to generate enhanced feature sequences. In the temporal correlation modeling stage, a gated recurrent unit network is combined to pass the enhanced feature sequences along the time step dimension and update the hidden states. A time attention mechanism is superimposed to calculate correlations and perform weighted aggregation to generate flood evolution trend prediction results. The multi-layer temporal convolutional unit includes an input end, a multi-scale convolutional kernel group, a convolution operation module, a residual connection module, and an output end; The multi-scale convolutional kernel group includes odd-length convolutional kernels of lengths three, five, and seven, wherein the convolutional kernel of length three extracts local short-term dependencies, the convolutional kernel of length five extracts medium-term dependencies, and the convolutional kernel of length seven extracts long-term dependencies. The extraction of local short-term dependencies specifically includes: using a convolutional kernel of length three, multiplying the value of each monitored variable at the corresponding three time steps with the corresponding convolutional kernel weight parameters one by one and summing the results to obtain the convolutional output sequence of local short-term dependencies. The extraction of intermediate dependencies specifically includes: using a convolutional kernel of length five, multiplying the value of each monitored variable at the corresponding five time steps with the corresponding convolutional kernel weight parameters one by one and summing the results to obtain the convolutional output sequence of intermediate dependencies. The extraction of long-term dependencies specifically includes: using a convolutional kernel of length seven, multiplying the value of each monitored variable at the corresponding seven time steps with the corresponding convolutional kernel weight parameters one by one and summing the results to obtain the convolutional output sequence of long-term dependencies. The gated loop unit network includes an input terminal, an update gate, a reset gate, a candidate hidden state generation module, a hidden state update module, and an output terminal. The candidate hidden state generation module includes an input terminal, a reset gate adjustment unit, a weighted summation unit, and a nonlinear transformation unit; The hidden state update module includes an input terminal, a weighted fusion unit, and an output terminal.

[0011] Optionally, the generation of the enhanced feature sequence includes the following specific steps: The feature vectors corresponding to each monitoring variable in the high-dimensional semantic feature sequence at the same time step are paired up. The feature vectors of each pair are multiplied one by one in the feature dimension and summed to obtain the correlation measure value. The correlation measure value is assigned to the position between the corresponding monitoring variables in the adjacency matrix to obtain the weighted adjacency matrix. The weighted adjacency matrix and the high-dimensional semantic feature sequence are input at different time steps. The monitoring variable features and the edge weights in the adjacency matrix at each time step are multiplied point by point and accumulated to obtain the weighted feature sequence that changes with time step. During the training process, the edge weights in the adjacency matrix are iteratively updated according to the backpropagation algorithm to obtain the dynamic graph adjacency matrix. Matrix multiplication is performed on the high-dimensional semantic feature sequence and the dynamic graph adjacency matrix to obtain the variable-dependent feature sequence. The adjacency matrix is ​​called to perform weighted aggregation of the feature vectors corresponding to each monitored variable at the same time step. Within each time step, the feature values ​​of each monitored variable in the variable-dependent feature sequence are multiplied one by one with the corresponding edge weights in the adjacency matrix and summed in the dimension of the monitored variable. The summation result is used as the aggregated output value of the corresponding monitored variable at the corresponding time step. All time steps of the entire time series tensor are traversed in turn to obtain the updated feature sequence. The high-dimensional semantic feature sequence and the updated feature sequence are concatenated at the corresponding time steps and monitoring variable positions. The feature values ​​of the same dimension are added point by point and weighted averaged on the concatenated sequence to obtain the fused feature sequence. The fused feature sequence is then normalized in the feature dimension to generate the enhanced feature sequence.

[0012] Optionally, the generation of the flood evolution trend prediction results includes the following specific steps: The enhanced feature sequences are input into the gated recurrent unit network of the improved MTGNN one by one in chronological order. At the current time step, the input feature vector corresponding to the enhanced feature sequence and the hidden state of the previous time step are multiplied one by one with the update gate parameter and the reset gate parameter, and the sum is calculated. The sum is input into the Sigmoid nonlinear function for normalization to obtain the update gate value and the reset gate value. The hidden state of the previous time step is multiplied point by point with the reset gate to obtain the hidden state adjusted by the reset gate. The hidden state adjusted by the reset gate and the input feature vector are weighted and summed and then input into the hyperbolic tangent nonlinear function for nonlinear transformation to obtain the candidate hidden state. The candidate hidden state is weighted and fused with the hidden state of the previous time step according to the update threshold to obtain the updated hidden state of the current time step. The updated hidden state is then passed to the next time step until all time steps of the enhanced feature sequence are traversed. All hidden states are arranged in order to obtain the output feature sequence. A temporal attention mechanism is adopted to calculate the correlation score between the output feature vectors at each time step within the same time range of the output feature sequence. The correlation score is then input into an exponential function for normalization to obtain the attention weights. In each time step, the feature values ​​of the output feature sequence are multiplied one by one with the corresponding attention weights and summed to obtain the attention feature sequence, which is then aggregated along the time step dimension to generate the flood evolution trend prediction result.

[0013] Optionally, the generation of the automatic control command includes the following specific steps: Align the flood evolution trend prediction results and real-time monitoring signals along the time step dimension to obtain the prediction sequence and monitoring sequence, respectively; At each time step, the predicted sequence value and the monitored sequence value are multiplied by the weighting coefficients one by one and summed to output the fused signal; The fused signal is input into the control parameter generation module, which performs component analysis on the fused signal. The components related to gate opening and closing are extracted and combined as gate opening and closing control parameters, and the components related to flood control barrier descent are extracted and combined as flood control barrier descent control commands. The gate opening and closing control parameters and flood control barrier descent control commands are spliced ​​together in the time series dimension to generate automatic control commands and drive the gate and flood control barrier to complete the corresponding actions.

[0014] Optionally, the gate includes a remote control switch and laser beam sensors are installed in front of and behind the gate. The front and rear laser beam sensors are detection devices based on the principle of laser beam transmission, consisting of a laser emitter and a receiver. They form a beam channel in the passage path of the gate, generate a beam interruption signal when a pedestrian or obstacle passes by, and output a detection signal.

[0015] The beneficial effects of this invention are: This invention presents a deep learning-based intelligent floodgate automatic control system that fully integrates multi-source monitoring data, including rainfall, water level, flow velocity, and meteorological data. Through standardized processing and time-series construction, it enables multi-dimensional feature extraction within a unified data framework. Compared to existing models that rely on single or limited parameters, this invention introduces multi-layer temporal convolutional units and adds residual connections during the feature extraction stage. By using multi-scale convolutional kernels, it extracts different levels of temporal dependencies in the short, medium, and long term, ensuring comprehensive modeling of the dynamic characteristics of the flood evolution process and effectively improving feature representation capabilities and model robustness. This not only avoids the problem of insufficient prediction accuracy in complex scenarios caused by shallow feature extraction but also enhances the ability to capture nonlinear flood evolution patterns under extreme weather conditions.

[0016] In the variable relationship modeling stage, this invention constructs a dynamic graph adjacency matrix, performs graph convolution operations on high-dimensional semantic feature sequences, and dynamically updates variable dependencies, enabling the system to characterize the complex interactions between monitored variables. Through this mechanism, the model can not only reflect the static correlation between rainfall, water level, flow velocity, and meteorological conditions, but also capture dynamic dependencies that change over time, further improving the accuracy of flood evolution trend prediction. Building on this, this invention combines a gated recurrent unit network to pass and update hidden states along the time step dimension, and superimposes a time attention mechanism to weighted aggregate output features, achieving focused attention on key time steps. This enhances the model's memory and correlation capabilities in long-term series prediction, thereby generating more accurate flood evolution trend prediction results.

[0017] In the control execution phase, this invention generates automatic control commands by weightedly fusing and analyzing flood evolution trend prediction results with real-time monitoring signals. This enables precise operation of gates and flood barriers based on both real-time and predictive information, effectively improving the foresight and reliability of flood control scheduling. Simultaneously, the invention introduces a front and rear laser beam sensor detection mechanism in the execution phase, enabling real-time acquisition of channel environmental signals during gate operation. When pedestrians or obstacles are detected, a pause command is promptly output to ensure operational safety. Furthermore, the system allows for receiving remote control or manual input to generate human control commands, which are then combined with automatic control commands to form a composite control mechanism. This allows for flexible switching between automation and manual intervention, improving operational controllability and system adaptability in emergency situations. In summary, this invention significantly outperforms existing technologies in terms of prediction accuracy, control rationality, and execution safety, possessing strong practical value and promotional significance. Attached Figure Description

[0018] 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:

[0019] Figure 1 This is a flowchart of a method for an intelligent floodgate automatic control system based on deep learning proposed in this invention. Figure 2 This is a schematic diagram of an improved MTGNN structure for an intelligent floodgate automatic control system based on deep learning proposed in this invention. Detailed Implementation

[0020] 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.

[0021] refer to Figure 1 and Figure 2 A deep learning-based intelligent floodgate automatic control system includes: The data acquisition module is used to acquire and preprocess multi-source monitoring data in real time. The feature extraction module is used to perform multi-layer temporal convolution and residual connections on standardized monitoring data; The variable relationship modeling module is used to construct a dynamic graph adjacency matrix, perform graph convolution operations on high-dimensional semantic feature sequences, and update variable dependencies. The temporal modeling module is used to pass and update the hidden state of the enhanced feature sequence along the time step dimension, and to calculate the correlation and weighted aggregation by combining the temporal attention mechanism; The control command module is used to perform weighted fusion and analysis of flood evolution trend prediction results and real-time monitoring signals; The execution control module is used to drive the gate and flood control barrier according to the automatic control command, and to perform the operation in combination with the front and rear laser beam sensor signals and remote control or manual input.

[0022] This invention constructs a complete deep learning prediction and control process. Multi-source monitoring data is input into an improved MTGNN, which sequentially performs feature extraction, variable relationship modeling, temporal correlation modeling, and control command generation. Combined with sensor detection and manual input for execution, this constitutes a data-driven, end-to-end intelligent automatic floodgate control system. Compared to traditional flood control facilities that rely on manual operation or single threshold triggering, this invention can analyze flood evolution trends in advance and automatically issue control commands based on the prediction results. It balances automated operation with manual intervention, improving prediction accuracy while ensuring the safety and flexibility of system operation.

[0023] In this embodiment, the modules are connected through the following method: Real-time acquisition of multi-source monitoring data, preprocessing, and generation of standardized monitoring data; Standardized monitoring data is constructed into a time series tensor, which is then input into the multi-layer temporal convolutional unit of the improved MTGNN to perform multi-scale convolution and residual connections, generating a high-dimensional semantic feature sequence. A dynamic graph adjacency matrix is ​​constructed, graph convolution is performed on the high-dimensional semantic feature sequence, and variable dependencies are updated to generate an enhanced feature sequence. The enhanced feature sequence is input into the gated recurrent unit network of the improved MTGNN, which is passed and updated in the time step dimension. The correlation is calculated and weighted aggregation is combined with the time attention mechanism to generate the flood evolution trend prediction result. The flood evolution trend prediction results are weighted, fused and analyzed with real-time monitoring signals to generate automatic control commands and drive the gates and flood control barriers to complete the corresponding actions. During the gate's operation, signals from the front and rear laser beam sensors are collected. When a pedestrian or obstacle is detected, a pause command is output. Remote control or manual input is received to generate human control commands, which are then combined with automatic control commands to generate composite control commands and drive the gate and flood control barrier to perform corresponding operations.

[0024] This invention establishes a step-by-step processing flow between modules, tightly integrating data acquisition, feature extraction, variable modeling, time-series modeling, and control execution, thus avoiding the problems of fragmented modules and delayed response in traditional methods. Multi-source monitoring data is directly fed into an improved MTGNN for deep learning analysis after acquisition. The prediction results are fused with real-time signals to generate control commands, which are then transmitted to the edge execution unit to achieve real-time control of gates and flood barriers. This invention enables the entire system to operate with higher closed-loop performance and automation, reducing reliance on manual intervention and significantly improving flood control response speed and reliability.

[0025] In this embodiment, the multi-source monitoring data includes rainfall data, water level data, flow velocity data, and meteorological data. The preprocessing includes formatting, outlier removal, missing value imputation, and normalization of the multi-source monitoring data.

[0026] This invention significantly improves the quality of input data by incorporating multi-source monitoring data, including rainfall, water level, flow velocity, and meteorological data, and performing formatting, outlier removal, missing value imputation, and normalization during the preprocessing stage. Compared to traditional methods relying on a single water level signal, this invention can integrate more key factors influencing flood evolution during the input stage, while eliminating noise and defects in the monitoring data through preprocessing, thus ensuring the integrity and consistency of the input data. This approach effectively enhances the accuracy of subsequent feature extraction and prediction, making the model closer to actual flood scenarios and improving the robustness of the overall system.

[0027] In this embodiment, the generation of the high-dimensional semantic feature sequence includes the following specific steps: Standardized monitoring data are arranged in chronological order. Multi-source monitoring data collected at the same time step are combined into a vector. The vectors corresponding to each time step are stacked in chronological order to obtain a time series tensor. The time series tensor includes a time step dimension and a monitoring variable dimension. The monitoring variables are various data components in the multi-source monitoring data. The time series tensor is input into the multi-layer temporal convolutional unit of the improved MTGNN. The multi-scale convolutional kernel group is called to perform one-dimensional convolution operation on the time series tensor along the time step dimension. Within each time window, the value of the monitored variable in the corresponding time window is multiplied by the corresponding convolutional kernel weight parameter and summed. The output of the summation operation is used as the convolution output value of the corresponding time window. The process is repeated for all time steps of the entire time series tensor to obtain the convolution output sequence covering different convolutional kernel scales. All convolution output sequences are concatenated to obtain the convolutional feature sequence. The weight parameters are randomly initialized and iteratively updated during training by combining the backpropagation algorithm. After training, they are fixed as the convolution coefficients in the inference stage. Residual connections are added to each temporal convolutional unit. The input features and output convolutional feature sequences of the corresponding temporal convolutional unit are added point by point at the corresponding time step and the monitoring variable position to generate a residual enhancement sequence. The residual connections are implemented by setting a bypass in the temporal convolutional unit. The residual enhancement features at each time step are stacked as tensors and arranged and combined along the dimension of the monitored variables to generate a high-dimensional semantic feature sequence.

[0028] This invention employs multi-layer temporal convolutional units during feature extraction and utilizes multi-scale convolutional kernels to process time-series tensors. It also introduces residual connection structures to prevent information from gradually decaying within the stacked convolutional layers. Compared to traditional single convolutional or shallow networks, this invention can more completely capture the short-term, medium-term, and long-term dependency features of monitoring data, while preserving the integrity of the input features in the convolutional results. This improvement enhances the expressive power of features and improves the stability of training convergence, making the generated high-dimensional semantic feature sequences more representative and providing a solid foundation for subsequent modeling.

[0029] In this embodiment, compared with the original MTGNN model, the improved MTGNN introduces multi-layer temporal convolutional units and adds residual connections in the feature extraction stage to perform convolution operations and stacking on standardized monitoring data to generate a high-dimensional semantic feature sequence. In the variable relationship modeling stage, a dynamic graph adjacency matrix is ​​constructed to perform graph convolution operations on the high-dimensional semantic feature sequence and update variable dependencies to generate an enhanced feature sequence. In the temporal correlation modeling stage, a gated recurrent unit network is combined to pass the enhanced feature sequence along the time step dimension and update the hidden state. A time attention mechanism is superimposed to calculate the correlation and perform weighted aggregation to generate flood evolution trend prediction results. The multi-layer temporal convolutional unit includes an input end, a multi-scale convolutional kernel group, a convolution operation module, a residual connection module, and an output end; The multi-scale convolutional kernel group includes odd-length convolutional kernels of lengths three, five, and seven, wherein the convolutional kernel of length three extracts local short-term dependencies, the convolutional kernel of length five extracts medium-term dependencies, and the convolutional kernel of length seven extracts long-term dependencies. The extraction of local short-term dependencies specifically includes: using a convolutional kernel of length three, multiplying the value of each monitored variable at the corresponding three time steps with the corresponding convolutional kernel weight parameters one by one and summing the results to obtain the convolutional output sequence of local short-term dependencies. The extraction of intermediate dependencies specifically includes: using a convolutional kernel of length five, multiplying the value of each monitored variable at the corresponding five time steps with the corresponding convolutional kernel weight parameters one by one and summing the results to obtain the convolutional output sequence of intermediate dependencies. The extraction of long-term dependencies specifically includes: using a convolutional kernel of length seven, multiplying the value of each monitored variable at the corresponding seven time steps with the corresponding convolutional kernel weight parameters one by one and summing the results to obtain the convolutional output sequence of long-term dependencies. The gated loop unit network includes an input terminal, an update gate, a reset gate, a candidate hidden state generation module, a hidden state update module, and an output terminal. The candidate hidden state generation module includes an input terminal, a reset gate adjustment unit, a weighted summation unit, and a nonlinear transformation unit; The hidden state update module includes an input terminal, a weighted fusion unit, and an output terminal.

[0030] This invention introduces multi-layer temporal convolutional units in the feature extraction stage. Compared to traditional feature extraction methods based on single convolutions, it can simultaneously utilize multi-scale convolutional kernels to process time-series tensors, resulting in superior performance in capturing short-term, medium-term, and long-term dependencies. Through this structure, standardized monitoring data is fully parsed into multi-dimensional temporal features, effectively avoiding the shortcomings of single-scale convolutions that ignore local or long-term dependencies. This allows high-dimensional semantic feature sequences to more completely represent the multi-level dynamic characteristics of flood evolution, providing more accurate input for subsequent variable relationship modeling.

[0031] This invention adds residual connections between temporal convolutional units, which effectively alleviates the gradient vanishing and information decay problems that occur in deep structures compared to the traditional layer-by-layer convolution stacking method. The residual connections directly add the input features to the output features point-by-point through a bypass, allowing information to flow across layers and thus preserving the integrity of the input features after convolution stacking. This structure not only enhances the expressive power of convolutional features but also improves the convergence speed and stability of the network, making the generated high-dimensional semantic feature sequences more robust and ensuring that the temporal features of the input data are not lost as the number of layers increases during flood prediction.

[0032] This invention constructs a dynamic graph adjacency matrix during the variable relationship modeling stage. Compared to the traditional fixed adjacency matrix modeling method, it can continuously update the dependencies between monitored variables based on the backpropagation algorithm during training. By calculating the correlation metric of eigenvectors at the same time step and assigning it to the adjacency matrix, this invention dynamically captures the interaction strength between monitored variables. As time progresses, the edge weights of the adjacency matrix are iteratively updated, thereby generating a dynamic graph structure that reflects the dependencies of variables at different time periods. This approach avoids the rigidity of fixed dependency graphs, making the flood evolution trend prediction results more consistent with actual hydrological dynamics.

[0033] This invention introduces a gated recurrent unit network (GRN) in the temporal correlation modeling stage. Compared to traditional recurrent neural networks, it effectively solves the problem of long-term dependency information being easily lost during propagation through its update and reset gate control mechanism. When processing enhanced feature sequences, the GRN can propagate and update hidden states along the time step dimension, enabling continuous capture of the dynamic changes in the flood evolution process. Simultaneously, the generation of candidate hidden states and the iterative update of hidden states make the model more stable in complex time series predictions, balancing information preservation and new feature learning, thereby ensuring the continuity and accuracy of the predicted sequence.

[0034] This invention superimposes a temporal attention mechanism onto the feature sequence output by a gated recurrent unit network. Compared to the traditional uniform weighting method, it can allocate attention weights based on the relevance scores at different time steps. By calculating the correlation between the output feature vectors at each time step and normalizing it into weights, this invention can highlight key time segments in flood evolution and weaken irrelevant or redundant parts. After weighted aggregation, the generated attention feature sequence can more accurately reflect the flood evolution trend. This improvement not only enhances the flexibility and accuracy of prediction but also strengthens the model's adaptability to complex time-series scenarios, making the prediction results closer to real-world water situation changes.

[0035] In this embodiment, the generation of the enhanced feature sequence includes the following specific steps: The feature vectors corresponding to each monitoring variable in the high-dimensional semantic feature sequence at the same time step are paired up. The feature vectors of each pair are multiplied one by one in the feature dimension and summed to obtain the correlation measure value. The correlation measure value is assigned to the position between the monitoring variables in the adjacency matrix to obtain the weighted adjacency matrix. The adjacency matrix is ​​a two-dimensional matrix, and the rows and columns of the matrix correspond to the monitoring variables. The elements in the matrix are the edge weights between the monitoring variables. The weighted adjacency matrix and the high-dimensional semantic feature sequence are input at different time steps. The monitoring variable features and the edge weights in the adjacency matrix at each time step are multiplied point by point and accumulated to obtain the weighted feature sequence that changes with time step. During the training process, the edge weights in the adjacency matrix are iteratively updated according to the backpropagation algorithm to obtain the dynamic graph adjacency matrix. The monitoring variable features are the feature vectors corresponding to each monitoring variable in the high-dimensional semantic feature sequence. Matrix multiplication is performed on the high-dimensional semantic feature sequence and the adjacency matrix of the dynamic graph to obtain the variable-dependent feature sequence. The matrix multiplication formula is: the output matrix is ​​equal to the sum of the products of each element value in the adjacency matrix and the corresponding variable value in the high-dimensional semantic feature sequence, where the element values ​​of the adjacency matrix are training parameters representing the strength of the variable relationship. The adjacency matrix is ​​called to perform weighted aggregation of the feature vectors corresponding to each monitored variable at the same time step. Within each time step, the feature values ​​of each monitored variable in the variable-dependent feature sequence are multiplied one by one with the corresponding edge weights in the adjacency matrix and summed in the dimension of the monitored variable. The summation result is used as the aggregated output value of the corresponding monitored variable at the corresponding time step. All time steps of the entire time series tensor are traversed in turn to obtain the updated feature sequence. The high-dimensional semantic feature sequence and the updated feature sequence are concatenated at the corresponding time steps and monitoring variable positions. The feature values ​​of the same dimension are added point by point and weighted averaged on the concatenated sequence to obtain the fused feature sequence. The fused feature sequence is then normalized in the feature dimension to generate the enhanced feature sequence.

[0036] This invention introduces a dynamic graph adjacency matrix in the variable relationship modeling stage, breaking through the limitations of traditional fixed dependency graphs. By calculating the correlation metric of monitored variables in a high-dimensional semantic feature sequence, the edge weights in the adjacency matrix are dynamically updated, and graph convolution is used to aggregate the features, resulting in an enhanced feature sequence. This invention can adaptively adjust the relationships between monitored variables over time, making the variable dependencies more consistent with the dynamic characteristics of flood evolution, avoiding the rigidity of static graph structures, thereby improving the accuracy and flexibility of predictions and significantly enhancing the ability to model complex flood processes.

[0037] In this embodiment, the generation of the flood evolution trend prediction result includes the following specific steps: The enhanced feature sequences are sequentially input into the gated recurrent unit network of the improved MTGNN. At the current time step, the input feature vector corresponding to the enhanced feature sequence is multiplied by the hidden state of the previous time step and the update gate parameter and reset gate parameter, respectively, and the sum is calculated. The sum is input into the Sigmoid nonlinear function for normalization to obtain the update gate value and reset gate value. The hidden state is the state vector output by the gated recurrent unit network at the previous time step, which is initialized to a zero vector at the first time step. The update gate parameter and reset gate parameter are the training weight parameters and bias parameters used in the gated recurrent unit network to calculate the update gate value and reset gate value. They are randomly initialized before the model training begins, and the gradient is calculated according to the loss function during the training process and iteratively updated through the backpropagation algorithm. During the inference phase, they are fixed as the coefficients of convolution calculation and gate operation. The hidden state of the previous time step is multiplied point by point with the reset gate to obtain the hidden state adjusted by the reset gate. The hidden state adjusted by the reset gate and the input feature vector are weighted and summed and then input into the hyperbolic tangent nonlinear function for nonlinear transformation to obtain the candidate hidden state. The candidate hidden state is weighted and fused with the hidden state of the previous time step according to the update threshold to obtain the updated hidden state of the current time step. The updated hidden state is then passed to the next time step until all time steps of the enhanced feature sequence are traversed. All hidden states are arranged in order to obtain the output feature sequence. A temporal attention mechanism is adopted to calculate the correlation score between the output feature vectors at each time step within the same time range of the output feature sequence. The correlation score is then input into an exponential function for normalization to obtain the attention weights. The calculation of the relevance score specifically includes: multiplying each output feature vector in the output feature sequence with the query mapping parameter matrix, the key mapping parameter matrix, and the value mapping parameter matrix one by one and summing them, using the multiplication and summation results as the output of the linear transformation to obtain the corresponding query vector, key vector, and value vector, respectively; multiplying the query vector and key vector of the output feature vector at any two time steps within the same time range one by one and summing them to obtain the relevance score. The query mapping parameter matrix, key mapping parameter matrix, and value mapping parameter matrix are all training weight parameters that are iteratively updated through the backpropagation algorithm during the training of the gated recurrent unit network, and are fixed as the coefficients of the vector mapping during the inference phase. In each time step, the feature values ​​of the output feature sequence are multiplied one by one with the corresponding attention weights and summed to obtain the attention feature sequence, which is then aggregated along the time step dimension to generate the flood evolution trend prediction result.

[0038] This invention introduces a gated recurrent unit (ROU) network and superimposes a temporal attention mechanism during the temporal modeling stage to enhance the transmission and updating of hidden states across the time step dimension of the feature sequence. The gated ROU controls information flow through update and reset gates, solving the problem of information loss that traditional ROUNCs rely on for a long time. Meanwhile, the temporal attention mechanism highlights key time segments in the output sequence and weakens redundant parts. This combination not only improves the accuracy of flood evolution trend prediction but also enhances its adaptability to complex temporal changes, making the prediction results closer to reality.

[0039] In this embodiment, the generation of the automatic control command includes the following specific steps: Align the flood evolution trend prediction results and real-time monitoring signals along the time step dimension to obtain the prediction sequence and monitoring sequence, respectively; The alignment in the time step dimension specifically includes: establishing a correspondence between the flood evolution trend prediction results and the real-time monitoring signals according to a unified time index; at each time step, arranging the values ​​corresponding to the flood evolution trend prediction results and the values ​​corresponding to the real-time monitoring signals at the same time position; and combining the corresponding values ​​of all time steps in chronological order to form a continuous prediction sequence and a monitoring sequence, respectively. The prediction sequence consists of the flood evolution trend prediction result values ​​at each time step, and the monitoring sequence consists of the real-time monitoring signal values ​​at each time step. Within each time step, the predicted sequence value and the monitoring sequence value are multiplied by the weighting coefficients one by one and summed to output a fused signal. The predicted sequence value is the value of the flood evolution trend prediction result at the corresponding time step, and the monitoring sequence value is the value of the real-time monitoring signal at the corresponding time step. The weighting coefficients are iteratively updated through the backpropagation algorithm during training and are fixed as the coefficients for fusion operation during the inference phase. The fused signal is input to the control parameter generation module, which performs component analysis on the fused signal. The components related to gate opening and closing are extracted and combined as gate opening and closing control parameters, and the components related to flood control barrier descent are extracted and combined as flood control barrier descent control commands. The gate opening and closing control parameters and flood control barrier descent control commands are spliced ​​together in the time series dimension to generate automatic control commands and drive the gate and flood control barrier to complete the corresponding actions. The gate opening and closing control parameters include opening degree parameters, opening and closing speed parameters, and opening and closing direction parameters. The flood control barrier descent control commands include descent height commands, descent speed commands, and reset commands.

[0040] This invention aligns flood evolution trend prediction results with real-time monitoring signals during the result fusion stage and performs weighted fusion along the time step dimension, combining predicted information with real-time data. The resulting data is then parsed into specific control commands by the control parameter generation module. Compared to traditional methods that rely solely on real-time signals or single predictions, this invention balances predictive foresight with real-time responsiveness, effectively reducing biases caused by a single information source. This invention ensures the accuracy and timeliness of control commands, thereby enhancing the automatic control capabilities of floodgates and flood barriers in complex environments.

[0041] In this embodiment, the gate includes a remote control switch and laser beam sensors are installed in front of and behind the gate. The front and rear laser beam sensors are detection devices based on the principle of laser beam transmission. They consist of a laser emitter and a receiver, forming a beam channel in the passage path of the gate. When a pedestrian or obstacle passes by, a beam interruption signal is generated and a detection signal is output.

[0042] This invention deploys laser beam sensors before and after the gate, combined with a remote control switch, enabling the system to provide both automatic control and safety protection capabilities, as well as manual intervention. The laser beam sensors create a beam channel along the passageway; when a pedestrian or obstacle passes, they generate an interruption signal and output a detection result, triggering the system to pause operation and thus avoiding the risk of accidental injury or collision. Simultaneously, this invention combines manual input with automatic control to form composite control commands, ensuring continuous automatic operation in emergency situations while allowing for manual intervention, thus enhancing safety and reliability.

[0043] Example 1: To verify the feasibility of this invention in practice, it was applied to the intelligent management of flood control gates in a riverside city in southern China. Located in a low-lying area at the confluence of rivers, this city is prone to rapid water level rises during the annual flood season due to the combined effects of continuous rainfall and upstream water flow. Previously, the city's flood control gates relied primarily on manual inspections and traditional timed opening and closing methods, which suffered from problems such as delayed response, imprecise scheduling, and inability to intervene in real-time during emergencies. Especially at night or during heavy rainfall, the excessive reliance on manual labor made it impossible to handle complex hydrological changes in a short time, leading to the risk of backflow and flooding in some areas under extreme conditions, seriously affecting residents' travel and safety.

[0044] In practical applications, this invention continuously collects real-time data by deploying multi-source monitoring sensors for rainfall, water level, flow velocity, and meteorological data around the gate. This data undergoes preprocessing steps, including formatting, outlier removal, missing value imputation, and normalization, to ensure data continuity and reliability. Subsequently, the preprocessed data is input into an improved MTGNN model. This model first uses multi-layer temporal convolutional units and residual connections to extract features from the input data, transforming the time-series patterns and implicit relationships between multi-source data into high-dimensional semantic feature sequences. Next, the model constructs a dynamic graph adjacency matrix to model and update the dependencies between monitoring variables, and iteratively learns the correlations between variables through graph convolution to obtain enhanced feature sequences that better reflect real hydrological patterns.

[0045] After feature enhancement, the model further transmits and updates the hidden state over time through a gated recurrent unit network, and introduces a temporal attention mechanism to weight the output features at different time steps, generating flood evolution prediction results that better reflect trend changes. This prediction result is fused with real-time monitoring signals, parsed by the control command generation module, to obtain gate opening and closing parameters and flood control barrier descent commands, which are then transmitted to the edge execution control unit to drive the actual device to complete the operation. Simultaneously, during gate operation, the front and rear laser beam sensors deployed on both sides of the gate channel continuously monitor the surroundings. If a pedestrian or obstacle is detected, the system immediately outputs a pause command to prevent safety accidents during device operation. The system also retains a manual intervention channel, allowing staff to input commands via remote control or manual switch. These commands, combined with automatic control commands, form a composite control sequence, ensuring that manual judgment and automatic control can work together in emergency situations.

[0046] Table 1. Performance Comparison of the Invention and Traditional Intelligent Flood Control Gate Automatic Control Method

[0047] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0048] Regarding the accuracy of flood evolution trend prediction, the traditional method achieves a prediction accuracy of 82.4%, while the method of this invention reaches 91.7%. This difference indicates that the method of this invention has greater advantages in multi-source monitoring data modeling and deep learning processing. Through the combined effect of the improved MTGNN's multi-layer temporal convolutional units, dynamic graph adjacency matrix, and gated recurrent unit network, this invention can better capture the temporal characteristics and variable relationships of flood evolution, thus significantly outperforming traditional methods in trend prediction accuracy.

[0049] Regarding the gate opening and closing response delay, the traditional method has a response delay of 15.6 seconds, while the method of this invention reduces the response delay to 6.8 seconds. This significant reduction is attributed to the weighted fusion mechanism of the prediction results and real-time monitoring signals in the method of this invention, which makes the generation of control commands more direct and effective. Simultaneously, the introduction of the edge execution control unit reduces data transmission and computation delays, enabling the gate to act quickly when sudden changes in water level or weather are detected, thereby improving overall control efficiency.

[0050] Regarding the accuracy of flood control barrier operation, the traditional method achieves an accuracy of 80.3%, while the method of this invention improves the accuracy to 92.1%. This improvement is attributed to the analytical capabilities of the control parameter generation module. This invention can precisely analyze relevant components into specific commands such as barrier descent height, descent speed, and reset, based on the fusion values ​​of flood evolution trend prediction results and real-time signals. Because the commands are more refined and intelligent, the flood control barrier maintains high consistency and accuracy during operation, significantly outperforming traditional experience-driven or manually set control methods.

[0051] Regarding the accuracy of anomaly detection, the traditional method achieves an accuracy of 77.9%, while the method of this invention improves it to 90.6%. This improvement mainly stems from the continuous monitoring and automatic pause mechanism of the front and rear laser beam sensors in this invention, as well as the composite control method combining remote control, manual input, and automatic commands. Through these technical means, this invention can more accurately identify pedestrians or obstacles, avoiding omissions and delays caused by reliance on manual observation in traditional methods, and greatly enhancing the safety of the system during operation.

[0052] Regarding overall system stability, the traditional method achieves a stability of 81.2%, while the method of this invention reaches 93.4%. This advantage stems from the adaptive characteristics of the deep learning model and the fusion mechanism of multi-source monitoring data. This invention can not only cope with complex and ever-changing hydrological scenarios, but also maintain the continuity and consistency of control during long-term operation, avoiding instability caused by signal fluctuations or human error.

[0053] 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 deep learning-based intelligent floodgate automatic control system, characterized in that, include: The data acquisition module is used to acquire and preprocess multi-source monitoring data in real time. The feature extraction module is used to perform multi-layer temporal convolution and residual connections on standardized monitoring data; The variable relationship modeling module is used to construct a dynamic graph adjacency matrix, perform graph convolution operations on high-dimensional semantic feature sequences, and update variable dependencies. The temporal modeling module is used to pass and update the hidden state of the enhanced feature sequence along the time step dimension, and to calculate the correlation and weighted aggregation by combining the temporal attention mechanism; The control command module is used to perform weighted fusion and analysis of flood evolution trend prediction results and real-time monitoring signals; The execution control module is used to drive the gate and flood control barrier according to the automatic control command, and to perform the operation in combination with the signals of the front and rear laser beam sensors and remote control or manual input.

2. The intelligent floodgate automatic control system based on deep learning according to claim 1, characterized in that, The modules are connected in the following way: Real-time acquisition of multi-source monitoring data, preprocessing, and generation of standardized monitoring data; Standardized monitoring data is constructed into a time series tensor, which is then input into the multi-layer temporal convolutional unit of the improved MTGNN to perform multi-scale convolution and residual connections, generating a high-dimensional semantic feature sequence. A dynamic graph adjacency matrix is ​​constructed, graph convolution is performed on the high-dimensional semantic feature sequence, and variable dependencies are updated to generate an enhanced feature sequence. The enhanced feature sequence is input into the gated recurrent unit network of the improved MTGNN, which is passed and updated in the time step dimension. The correlation is calculated and weighted aggregation is combined with the time attention mechanism to generate the flood evolution trend prediction result. The flood evolution trend prediction results are weighted, fused and analyzed with real-time monitoring signals to generate automatic control commands and drive the gates and flood control barriers to complete the corresponding actions. During the gate's operation, signals from the front and rear laser beam sensors are collected. When a pedestrian or obstacle is detected, a pause command is output. Remote control or manual input is received to generate human control commands, which are then combined with automatic control commands to generate composite control commands and drive the gate and flood control barrier to perform corresponding operations.

3. The intelligent floodgate automatic control system based on deep learning according to claim 2, characterized in that, The multi-source monitoring data includes rainfall data, water level data, flow velocity data, and meteorological data. The preprocessing includes formatting, outlier removal, missing value imputation, and normalization of the multi-source monitoring data.

4. The intelligent floodgate automatic control system based on deep learning according to claim 2, characterized in that, The generation of the high-dimensional semantic feature sequence includes the following specific steps: The standardized monitoring data are arranged in chronological order. The multi-source monitoring data collected at the same time step are combined into a vector. The vectors corresponding to each time step are stacked in chronological order to obtain a time series tensor. The time series tensor is input into the multi-layer temporal convolutional unit of the improved MTGNN. The multi-scale convolutional kernel group is called to perform one-dimensional convolution operation on the time series tensor along the time step dimension. Within each time window, the value of the monitored variable in the corresponding time window is multiplied by the corresponding convolutional kernel weight parameter and summed. The output of the summation operation is used as the convolution output value of the corresponding time window. The process is repeated for all time steps of the entire time series tensor to obtain the convolution output sequence covering different convolutional kernel scales. All convolution output sequences are concatenated to obtain the convolutional feature sequence. Residual connections are added to each temporal convolutional unit. The input features and output convolutional feature sequences of the corresponding temporal convolutional unit are added point by point at the corresponding time step and the monitoring variable position to generate a residual enhancement sequence. The residual connections are implemented by setting a bypass in the temporal convolutional unit. The residual enhancement features at each time step are stacked as tensors and arranged and combined along the dimension of the monitored variables to generate a high-dimensional semantic feature sequence.

5. The intelligent floodgate automatic control system based on deep learning according to claim 4, characterized in that, Compared to the original MTGNN model, the improved MTGNN introduces multi-layer temporal convolutional units and adds residual connections in the feature extraction stage to perform convolution operations and stacking on standardized monitoring data to generate high-dimensional semantic feature sequences. In the variable relationship modeling stage, a dynamic graph adjacency matrix is ​​constructed to perform graph convolution operations on the high-dimensional semantic feature sequences and update variable dependencies to generate enhanced feature sequences. In the temporal correlation modeling stage, a gated recurrent unit network is combined to pass the enhanced feature sequences along the time step dimension and update the hidden states. A time attention mechanism is also superimposed to calculate correlations and perform weighted aggregation to generate flood evolution trend prediction results. The multi-layer temporal convolutional unit includes an input end, a multi-scale convolutional kernel group, a convolution operation module, a residual connection module, and an output end; The multi-scale convolutional kernel group includes odd-length convolutional kernels of lengths three, five, and seven, wherein the convolutional kernel of length three extracts local short-term dependencies, the convolutional kernel of length five extracts medium-term dependencies, and the convolutional kernel of length seven extracts long-term dependencies. The extraction of local short-term dependencies specifically includes: using a convolutional kernel of length three, multiplying the value of each monitored variable at the corresponding three time steps with the corresponding convolutional kernel weight parameters one by one and summing the results to obtain the convolutional output sequence of local short-term dependencies. The extraction of intermediate dependencies specifically includes: using a convolutional kernel of length five, multiplying the value of each monitored variable at the corresponding five time steps with the corresponding convolutional kernel weight parameters one by one and summing the results to obtain the convolutional output sequence of intermediate dependencies. The extraction of long-term dependencies specifically includes: using a convolutional kernel of length seven, multiplying the value of each monitored variable at the corresponding seven time steps with the corresponding convolutional kernel weight parameters one by one and summing the results to obtain the convolutional output sequence of long-term dependencies. The gated loop unit network includes an input terminal, an update gate, a reset gate, a candidate hidden state generation module, a hidden state update module, and an output terminal. The candidate hidden state generation module includes an input terminal, a reset gate adjustment unit, a weighted summation unit, and a nonlinear transformation unit; The hidden state update module includes an input terminal, a weighted fusion unit, and an output terminal.

6. The intelligent floodgate automatic control system based on deep learning according to claim 2, characterized in that, The generation of the enhanced feature sequence includes the following specific steps: The feature vectors corresponding to each monitoring variable in the high-dimensional semantic feature sequence at the same time step are paired up. The feature vectors of each pair are multiplied one by one in the feature dimension and summed to obtain the correlation measure value. The correlation measure value is assigned to the position between the corresponding monitoring variables in the adjacency matrix to obtain the weighted adjacency matrix. The weighted adjacency matrix and the high-dimensional semantic feature sequence are input at different time steps. The monitoring variable features and the edge weights in the adjacency matrix at each time step are multiplied point by point and accumulated to obtain the weighted feature sequence that changes with time step. During the training process, the edge weights in the adjacency matrix are iteratively updated according to the backpropagation algorithm to obtain the dynamic graph adjacency matrix. Matrix multiplication is performed on the high-dimensional semantic feature sequence and the dynamic graph adjacency matrix to obtain the variable-dependent feature sequence. The adjacency matrix is ​​called to perform weighted aggregation of the feature vectors corresponding to each monitored variable at the same time step. Within each time step, the feature values ​​of each monitored variable in the variable-dependent feature sequence are multiplied one by one with the corresponding edge weights in the adjacency matrix and summed in the dimension of the monitored variable. The summation result is used as the aggregated output value of the corresponding monitored variable at the corresponding time step. All time steps of the entire time series tensor are traversed in turn to obtain the updated feature sequence. The high-dimensional semantic feature sequence and the updated feature sequence are concatenated at the corresponding time steps and monitoring variable positions. The feature values ​​of the same dimension are added point by point and weighted averaged on the concatenated sequence to obtain the fused feature sequence. The fused feature sequence is then normalized in the feature dimension to generate the enhanced feature sequence.

7. The intelligent floodgate automatic control system based on deep learning according to claim 2, characterized in that, The generation of the flood evolution trend prediction results includes the following specific steps: The enhanced feature sequences are input into the gated recurrent unit network of the improved MTGNN one by one in chronological order. At the current time step, the input feature vector corresponding to the enhanced feature sequence and the hidden state of the previous time step are multiplied one by one with the update gate parameter and the reset gate parameter, and the sum is calculated. The sum is input into the Sigmoid nonlinear function for normalization to obtain the update gate value and the reset gate value. The hidden state of the previous time step is multiplied point by point with the reset gate to obtain the hidden state adjusted by the reset gate. The hidden state adjusted by the reset gate and the input feature vector are weighted and summed and then input into the hyperbolic tangent nonlinear function for nonlinear transformation to obtain the candidate hidden state. The candidate hidden state is weighted and fused with the hidden state of the previous time step according to the update threshold to obtain the updated hidden state of the current time step. The updated hidden state is then passed to the next time step until all time steps of the enhanced feature sequence are traversed. All hidden states are arranged in order to obtain the output feature sequence. A temporal attention mechanism is adopted to calculate the correlation score between the output feature vectors at each time step within the same time range of the output feature sequence. The correlation score is then input into an exponential function for normalization to obtain the attention weights. In each time step, the feature values ​​of the output feature sequence are multiplied one by one with the corresponding attention weights and summed to obtain the attention feature sequence, which is then aggregated along the time step dimension to generate the flood evolution trend prediction result.

8. The intelligent floodgate automatic control system based on deep learning according to claim 2, characterized in that, The generation of the automatic control command includes the following specific steps: Align the flood evolution trend prediction results and real-time monitoring signals along the time step dimension to obtain the prediction sequence and monitoring sequence, respectively; At each time step, the predicted sequence value and the monitored sequence value are multiplied by the weighting coefficients one by one and summed to output the fused signal; The fused signal is input into the control parameter generation module, which performs component analysis on the fused signal. The components related to gate opening and closing are extracted and combined as gate opening and closing control parameters, and the components related to flood control barrier descent are extracted and combined as flood control barrier descent control commands. The gate opening and closing control parameters and flood control barrier descent control commands are spliced ​​together in the time series dimension to generate automatic control commands and drive the gate and flood control barrier to complete the corresponding actions.

9. The intelligent floodgate automatic control system based on deep learning according to claim 2, characterized in that, The gate includes a remote control switch and laser beam sensors are installed in front of and behind the gate. The front and rear laser beam sensors are detection devices based on the principle of laser beam transmission. They consist of a laser emitter and a receiver, forming a beam channel in the passage path of the gate. When a pedestrian or obstacle passes by, a beam interruption signal is generated and a detection signal is output.

Images