Gate opening degree control method and system based on machine learning algorithm

Abstract

The invention relates to the technical field of gate opening control, in particular to a gate opening control method and system based on a machine learning algorithm, a dual-channel LSTM (Long Short Term Memory) prediction model unit adopts dual-channel processing, processes meteorological modal data and hydrological modal data respectively, performs corresponding feature extraction, and adopts an attention mechanism, so as to achieve the control of the gate opening. The influence weight ratio of the meteorological factors and the hydrological factors is dynamically adjusted, a subsequent prediction result is more adaptive to an actual application scene, and the prediction precision is further improved; in the subsequent control instruction generation, an instruction rolling optimization generation adjustment algorithm is adopted, and the scene adaptability is further improved, so that the actual application scene is met in the subsequent control adjustment of the gate opening degree; through three core technologies of double-channel processing, attention fusion and rolling optimization, the industrial problems of meteorological-hydrological response disjunction, insufficient long-time-domain prediction precision and the like in water conservancy control are solved, and reliable technical support is provided for intelligent water conservancy construction.

Description

Technical Field

[0001] This invention relates to the field of gate opening control technology, and more specifically, to a gate opening control method and system based on machine learning algorithms. Background Technology

[0002] Gates are installed on the openings of structures such as spillways, bank spillways, drainage holes, hydraulic tunnels, pumping stations, and sluices. They are used to regulate flow, control upstream and downstream water levels, discharge floodwaters, and remove silt or floating debris. They are an important component of hydraulic structures. In sluice projects, gates are the main part, often occupying a large part of the water-retaining area. The opening control of gates is crucial for the overall flow control and regulation. In existing automatic adjustment methods, such as the pump station gate opening control method and system disclosed in patent publication number CN114237078B, the traditional method of calculating and predicting water level data to achieve automatic control of gate opening is adopted. This prediction method is difficult to adapt to dynamic environments, such as meteorological environments with sudden changes in rainfall. In traditional prediction methods, meteorological data and hydrological data are difficult to dynamically integrate, resulting in low prediction accuracy. The gate opening adjustment cannot meet the actual application scenarios, leading to a sharp increase in the probability of accidents. Therefore, most sluice gates in China have upgraded their gate control technology from manual to remote automated control and connected to a group control system, with staff at the command center coordinating the gate operations. However, because each sluice gate has different energy dissipation, anti-scour, and irrigation / drainage requirements, the operation is quite difficult for staff, especially in the event of sudden extreme weather, which puts even greater pressure on the command center.

[0003] Compared to traditional sluice gate scheduling methods in China, foreign water conservancy projects have a higher level of automation and intelligence. For example, the Maeslantkering storm surge dam at the Nieuwe Waterweg outlet in Rotterdam, Netherlands, is one of the world's largest dam projects and movable floodgates. When extreme high water levels are predicted, it can automatically open and close the gates by relying on advanced computer control and accurate environmental monitoring, ensuring that the flood control system can operate normally even when staff are not present during sudden extreme weather events.

[0004] Therefore, based on the above problems and the current status of sluice gate projects at home and abroad, a gate opening control method and system based on machine learning algorithms is designed. This method can effectively integrate meteorological and hydrological data and combine them with automatic gate control to make the gate opening control more in line with actual application scenarios. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a gate opening control method and system based on machine learning algorithms to solve the problems existing in the background technology.

[0006] The above-mentioned technical objective of this invention is achieved through the following technical solution: a gate opening control method based on machine learning algorithms, comprising: Step A, Real-time Multimodal Data Acquisition and Feature Construction: Construct a multi-source data acquisition layer to collect real-time multimodal data affecting gate opening at several acquisition time steps; the real-time multimodal data includes meteorological modal data and hydrological modal data; Step B: Constructing dual channels and channel preprocessing algorithms: Construct meteorological and hydrological channels, and construct and execute independent data preprocessing algorithms for each channel. Step C: Constructing a dual-channel multidimensional time series: Based on real-time multimodal data, define the feature dimensions of the meteorological and hydrological channels, construct the multidimensional time series matrix of the corresponding channels based on the defined feature dimensions, and construct the sliding window length; Step D: Based on the set neural network hyperparameters and the constructed dual channels, construct a dual-channel LSTM prediction model; Step E, Data Prediction: Based on real-time multimodal data, input into a dual-channel LSTM prediction model to obtain predicted water level data for several future time steps; Step F, Generation and Execution of Dynamic Control Command for Gate Opening: Based on the predicted water level data for several future time steps, a dynamic control command for gate opening is generated, and then the gate opening is adjusted accordingly based on the dynamic control command.

[0007] Optionally, step A is implemented as follows, including: A1. The data acquisition layer includes the electromechanical sensing layer and the meteorological sensing and processing layer; A2. Based on the meteorological sensing and processing layer, the rainfall intensity, temperature, wind speed, reflectivity factor and infrared cloud image of the target sluice gate within several collection time steps are obtained. The spatiotemporal alignment algorithm is used to align and fit the meteorological data collected at different time steps, thereby generating meteorological modal data. A3. Based on the electromechanical sensing layer, acquire hydrological data within several acquisition time steps at the target sluice gate, including water level, flow rate and velocity, and then fit and generate hydrological mode data.

[0008] Optionally, step B is implemented as follows, including: B1. Construct meteorological and hydrological channels; B2. The meteorological channel has a built-in preprocessing algorithm for meteorological modal data; B3. The hydrological channel has a built-in algorithm for preprocessing hydrological modal data.

[0009] Optionally, step C is implemented as follows, including: C1. Define the four-dimensional meteorological channel characteristics, including radar reflectivity, ground rainfall intensity, 6-hour cumulative rainfall, and wind speed. C2. Define four-dimensional hydrological channel characteristics, including upstream water level, downstream water level, flow rate of change, and cumulative rainfall over time. C3. Based on the defined 4-dimensional meteorological channel feature dimension and 4-dimensional hydrological channel feature dimension, construct MxN meteorological multidimensional time series matrix and MxN hydrological multidimensional time series matrix respectively, where M is the number of dimensions, N is the number of data, and N is the number of time-aligned data samples. C4. Based on the meteorological multidimensional time series matrix and hydrological multidimensional time series matrix obtained in step C3, the sliding window method is used for processing and output.

[0010] Optionally, step D is implemented as follows, including: D1. Construct a comprehensive data input layer, which includes a meteorological data input layer and a hydrological data input layer; meteorological modal data is input into the meteorological data input layer through the meteorological channel, and hydrological modal data is input into the hydrological data input layer through the hydrological channel; D2. Construct an LSTM data network based on the data input from the integrated data input layer; the LSTM data network includes a hydrological LSTM data network and a meteorological LSTM data network. D3. Construct a feature fusion layer. Based on the data network architecture of LSTM data network, use the attention fusion mechanism to fuse features and then output the influencing features. D4. Construct a joint prediction layer and predict water level data based on the high-impact features output by the feature fusion layer.

[0011] Optionally, the specific implementation process of step E is as follows, including: E1. Receive real-time multimodal data transmitted by the multi-source data acquisition layer, separate and fit to form hydrological modal data and meteorological modal data; E2. Input the hydrological mode data and meteorological mode data into the integrated data input layer, and input the hydrological mode data into the hydrological channel and the meteorological mode data into the meteorological channel respectively; E3. Based on the established data network architecture of hydrological LSTM data network and meteorological LSTM data network, meteorological modal data and hydrological modal data that have passed through the corresponding data channels are input into the feature fusion layer to fuse features and then output the influencing features. E4. Based on the high-impact features output by the feature fusion layer, the data is input into the joint prediction layer to obtain the predicted water level data at the target sluice gate for several future time steps.

[0012] Optionally, the specific implementation process of step F is as follows, including: F1. Based on the predicted water level data at the target sluice gate in several future time steps, the rolling optimization control algorithm is used for optimization, and then the optimal opening sequence is solved. F2. Based on the optimized optimal opening sequence, a smooth opening change trajectory is generated using a fifth-order polynomial function; F3. Based on the smooth opening change trajectory generated by F2, a PLC instruction conversion algorithm is used to convert it into a gate opening control instruction; F4. Based on the gate opening control command, the gate opening is adjusted accordingly.

[0013] Optionally, step F2 further includes: Based on the generated control commands, a safety constraint algorithm is used to apply control constraints to the corresponding control commands based on the set safety constraint conditions.

[0014] A gate opening control system for executing the above-described gate opening control method based on machine learning algorithms includes: The multimodal data acquisition module is used to acquire and transmit real-time multimodal data affecting the gate opening degree of the target gate during the acquisition time period. The dual-channel LSTM prediction model unit is used to train and update the prediction model based on the real-time multimodal data and historical input data collected by the multimodal data acquisition module, and to construct a dual-channel data input. The data transfer module is used to receive real-time multimodal data transmitted by the multimodal data acquisition module, and separate the meteorological data and hydrological data therein to transmit to the dual-channel LSTM prediction model unit, thereby generating predicted water level data; The instruction generation module is used to generate dynamic control instructions for gate opening based on the predicted water level data generated by the dual-channel LSTM prediction model unit. The safety constraint module, based on the gate opening dynamic control command generated by the command generation module, outputs the adjusted gate opening dynamic control command through the set safety threshold. The instruction execution module controls the target gate accordingly based on the gate opening dynamic control instruction generated by the safety constraint module.

[0015] In summary, the present invention has the following beneficial effects: 1. The real-time multimodal data collected by the multi-source data acquisition layer includes not only hydrological modal data from traditional forecasting methods, but also meteorological modal data. The two types of data collected provide more comprehensive judgments for subsequent forecasting, control and regulation, further improving forecast accuracy and making the regulation of gate opening more suitable for actual application scenarios.

[0016] 2. The dual-channel LSTM prediction model unit employs dual-channel processing to process meteorological and hydrological modal data separately, extracting corresponding features. An attention mechanism is used to dynamically adjust the weighting of meteorological and hydrological factors, making the subsequent prediction results more suitable for real-world application scenarios and further improving prediction accuracy. In the subsequent control command generation, a rolling optimization algorithm is used to further improve scenario adaptability, ensuring that the control and adjustment of gate opening conforms to actual application scenarios. Through these three core technologies—dual-channel processing, attention fusion, and rolling optimization—the model overcomes industry challenges such as the disconnect between meteorological and hydrological responses and insufficient long-term prediction accuracy in water conservancy control, providing reliable technical support for smart water conservancy construction. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the water level prediction process of the present invention; Figure 3 This is a schematic diagram of the data preprocessing flow of the meteorological channel of the present invention; Figure 4 This is a schematic diagram of the data preprocessing flow of the hydrological channel of the present invention; Figure 5 This is a schematic diagram of a system control example of the present invention. Detailed Implementation

[0018] To make the objectives, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein.

[0019] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature.

[0020] In this invention, unless otherwise expressly specified and limited, "above" or "below" a second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of a second feature includes the first feature being directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" of a second feature includes the first feature being directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature. The terms "vertical," "horizontal," "left," "right," "above," "below," and similar expressions are for illustrative purposes only and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed or operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0021] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] This invention provides a gate opening control method and system based on machine learning algorithms, such as... Figure 1 As shown, it includes: Step A, Real-time Multimodal Data Acquisition and Feature Construction: Construct a multi-source data acquisition layer to collect real-time multimodal data affecting gate opening at several acquisition time steps; the real-time multimodal data includes meteorological modal data and hydrological modal data; Step B: Constructing dual channels and channel preprocessing algorithms: Construct meteorological and hydrological channels, and construct and execute independent data preprocessing algorithms for each channel. Step C: Constructing a dual-channel multidimensional time series: Based on real-time multimodal data, define the feature dimensions of the meteorological and hydrological channels, construct the multidimensional time series matrix of the corresponding channels based on the defined feature dimensions, and construct the sliding window length; Step D: Based on the set neural network hyperparameters and the constructed dual channels, construct a dual-channel LSTM prediction model; Step E, Data Prediction: Based on real-time multimodal data, input into a dual-channel LSTM prediction model to obtain predicted water level data for several future time steps; Step F, Generation and Execution of Dynamic Control Command for Gate Opening: Based on the predicted water level data for several future time steps, a dynamic control command for gate opening is generated, and then the gate opening is adjusted accordingly based on the dynamic control command.

[0023] Furthermore, the specific implementation process of step A is as follows, including: A1. The data acquisition layer includes the electromechanical sensing layer and the meteorological sensing and processing layer; A2. Based on the meteorological sensing and processing layer, the rainfall intensity, temperature, wind speed, reflectivity factor and infrared cloud image of the target sluice gate within several collection time steps are obtained. The spatiotemporal alignment algorithm is used to align and fit the meteorological data collected at different time steps, thereby generating meteorological modal data. A3. Based on the electromechanical sensing layer, acquire hydrological data within several acquisition time steps at the target sluice gate, including water level, flow rate and velocity, and then fit and generate hydrological mode data.

[0024] Specifically, in the actual data acquisition process, missing values ​​will occur due to the accumulation of sampling errors from the data acquisition unit and sensors in the acquisition instrument. To ensure the accuracy and effectiveness of the acquired data, the electromechanical sensing layer and the meteorological sensing and processing layer use linear interpolation to process the missing values. The calculation method is as follows: For outliers in the collected data that are clearly inconsistent with the average water level, such as values ​​that are several times higher than the average water level, the normal distribution can be used. The principle for removing outliers is as follows: ,in Represents the average value of the sample. The standard deviation of the sample; Furthermore, the specific implementation process of step B is as follows, including: B1. Construct meteorological and hydrological channels; B2. The meteorological channel has a built-in preprocessing algorithm for meteorological modal data; B3. The hydrological channel has a built-in algorithm for preprocessing hydrological modal data.

[0025] Specifically, data processing in meteorological channels first involves normalization. This normalization process standardizes the meteorological modal data to unify the time series. An exponential transform + standardization normalization algorithm is used. Represented as a zero-prevention value, , Where R represents the original rainfall intensity and N represents the total number of historical data points. It is represented as a weighted historical average (with more recent data having higher weight). Expressed as historical standard deviation, This is expressed as a time decay constant, typically 24 hours; the logarithmic transformation method can effectively solve the problem of long-tailed rainfall distribution. Represented as the current timestamp, Represented as the timestamp of the i-th data point; The data processing methods for the meteorological channel also include Butterworth low-pass filtering, specifically expressed as:

[0026]

[0027] The coefficients are calculated as follows: ; The corresponding coefficients are defined as follows: s represents the Laplace variable, i.e., the complex frequency; Represented as the cutoff angular frequency, and the cutoff frequency The relationship is z represents the Z-transform variable; This is represented as the dynamic cutoff frequency. This is expressed as the sampling frequency: , , , , These are represented as the coefficients of a digital filter. Wavelet thresholding is also used for denoising, as follows:

[0028] The threshold is calculated as follows: , It is represented as the wavelet coefficients at scale K; N represents the number of wavelet coefficients. The adaptive filtering algorithm for dynamically adjusting the cutoff frequency within the meteorological channel is expressed as follows: ,in =0.01Hz, which is the base cutoff frequency. Expressed as the standard deviation of the rate of change of rainfall intensity, This is expressed as an adjustment coefficient, with a value of 0.05 Hz / (mm / h / min), which is then adjusted using actual values ​​to improve data reliability. An algorithm function can also be set to improve the signal-to-noise ratio of meteorological modal data, expressed as follows: ;in, It is expressed as the standard deviation of the signal. This is expressed as the standard deviation of noise before preprocessing. This is expressed as the standard deviation of the noise after preprocessing; it improves the signal-to-noise ratio of the meteorological channel by 15dB and effectively suppresses the impulse noise of the tipping bucket rain gauge.

[0029] Specifically, the data processing in the hydrological channel adopts a piecewise linear transformation normalization method, expressed as:

[0030] in, This represents the lowest water level in history. This indicates the normal water level. This is indicated as a warning water level to adapt to hydrological conditions at different times; The hydrological channel employs Kalman filtering, and its state-space model is represented as follows: The state vector is defined as:

[0031] Kalman gain update is represented as: ; in, This is represented as process noise, typically Gaussian white noise; This is represented as observation noise, typically Gaussian white noise; H is represented as an observation vector, such as the water level value; H is represented as the observation matrix, which maps the state to the observation. R is denoted as the prior estimation error covariance matrix; R is denoted as the observation noise covariance matrix. Furthermore, Kalman noise covariance adaptation was employed to further remove noise from the data, as shown below:

[0032] in The value is 1e-5, representing the noise in the water level measurement. The value of is 1e-4, representing first-order derivative noise. The value of is 1e-3, which represents second-order derivative noise. The value is 0.5, which represents the sensitivity to change; after preprocessing the data from the hydrological channel, the water level filtering error of the hydrological channel is ≤0.02m, and the delay is <10 seconds; Expressed as the rate of change of water level; Meanwhile, outlier removal is used to preprocess upstream and downstream water levels, thereby eliminating sudden disturbances such as ship passage; differential calculation is used to enhance the water level gradient direction information and preprocess the flow rate change. Time-accumulated rainfall is acquired through meteorological data and accumulated at multiple time scales. Due to its data characteristics, it directly drives hydrological responses, such as runoff formation and soil saturation. Breaking the inertia of traditional grouping by data source, it is now grouped by physical meaning. Spatial interpolation processing algorithms are used, and then the time-accumulated rainfall data is dynamically standardized to enhance its characteristics.

[0033] Furthermore, the specific implementation process of step C is as follows, including: C1. Define the four-dimensional meteorological channel characteristics, including radar reflectivity, ground rainfall intensity, 6-hour cumulative rainfall, and wind speed. C2. Define four-dimensional hydrological channel characteristics, including upstream water level, downstream water level, flow rate of change, and cumulative rainfall over time. C3. Based on the defined 4-dimensional meteorological channel feature dimension and 4-dimensional hydrological channel feature dimension, construct MxN meteorological multidimensional time series matrix and MxN hydrological multidimensional time series matrix respectively, where M is the number of dimensions, N is the number of data, and N is the number of time-aligned data samples. C4. Based on the meteorological multidimensional time series matrix and hydrological multidimensional time series matrix obtained in step C3, set the sliding window length L, that is, manually divide the correspondence between input and output, where L is the number of input quantities and the output quantity is water level.

[0034] Specifically, the dimensions can be defined in greater depth according to the complexity of the actual application scenario, including but not limited to temperature, humidity, air pressure, wind speed, wind direction, precipitation, etc. In this embodiment, the meteorological channel dimension is defined as a 4-dimensional meteorological channel feature dimension. This 4-dimensional feature is one of the most important meteorological factors affecting the gate opening, and is expressed as follows:

[0035] Among them, radar reflectivity reflects cloud water vapor density, ground rainfall intensity reflects real-time precipitation intensity, 6-hour cumulative rainfall reflects soil saturation, and wind speed reflects the speed at which the rainstorm moves. When the original data is input, the feature values ​​of the corresponding feature dimensions are calculated, including: Radar reflectivity can be obtained directly from meteorological station data or calculated using a formula, expressed as:

[0036] This represents complex sampling of the radar echo signal, where N represents the number of sampling points within the scanning period; it is calculated in real time by a Doppler radar DSP processor; in the preferred embodiment, to ensure the accuracy of the actual data, data is shared with a weather station; the output data unit is dBz; Surface rainfall intensity can be directly transmitted from meteorological station data, or calculated using a formula, expressed as:

[0037] This calculation method is derived from data from rainwater measuring stations. This indicates the number of times the rainwater measuring bucket flips. This represents the volume of the rainwater measuring bucket. This is indicated as wind speed correction. This represents the sampling time interval, with an effective data acquisition frequency of 1 minute interval; the output data unit is... ; The 6-hour cumulative rainfall is transmitted directly from the weather station data; Based on the above features, a meteorological multidimensional time series matrix of M x N is constructed, represented as:

[0038] Where M represents the number of features of the meteorological channel, with a value of 4, and N represents the number of time-aligned data samples, i.e., the number of time steps; the definition in the matrix is: Expressed as radar reflectivity, Indicated as surface rainfall intensity, Expressed as cumulative rainfall over 6 hours. Expressed as wind speed; The dynamics of rainstorm formation and dissipation are reflected through the characteristic quantities of meteorological channels and the meteorological multidimensional time series matrix.

[0039] Specifically, the four-dimensional hydrological channel feature dimensions are defined as follows:

[0040] Among them, the physical meaning of upstream water level is water storage capacity, the physical meaning of downstream water level is water flow energy, the physical meaning of flow rate change rate is flood formation speed, and the physical meaning of cumulative rainfall over time is to reflect the impact of rainfall on water level and represent the historical rainfall load of the basin; all of the above characteristic values ​​can be obtained directly through sensor data. The meaning of time-cumulative rainfall differs from that of 6-hour cumulative rainfall in meteorological data. Meteorological data tends to represent the amount within a region, which is a macroscopic representation, i.e., the long-term impact. Hydrological data, on the other hand, is the data collected at the target gate, which is more specific and more in line with practical applications, i.e., the short-term impact. Based on the above features, an MxN hydrological multidimensional time series matrix is ​​constructed, represented as:

[0041] Where M represents the number of features of the hydrological channel, with a value of 4, and N represents the number of time-aligned data samples, i.e., the number of time steps; the definition in the matrix is: This represents the upstream water level. This represents the downstream water level. Expressed as the rate of change of flow rate, Represented as cumulative rainfall over time; The time-dependent inertial characteristics of water flow are reflected by the characteristic quantities of the hydrological channel and the multidimensional time series matrix of the hydrology. Based on the above calculation results, a sliding window is constructed, and the input and output quantities are defined as follows:

[0042]

[0043] The input quantities are represented as meteorological multi-dimensional time series matrices and hydrological multi-dimensional time series matrices. The input window length L is 72, which is represented as 6 hours. The reason is that if the window length L is too short, it will not be able to capture the flood formation process, and if it is too long, it will introduce noise and increase the computational burden. This value is usually determined by the watershed confluence time. The output is represented as predicted water level data, and the number of output windows is K = 12, based on the gate response cycle, which is expressed as 1 hour. The process of constructing a sliding window is as follows: take adjacent... L Each sample is used as a sliding window, mapping it to a water level data point, and this sliding window... L It should contain complete information controlling the acquisition cycle. Specifically, the method for constructing time series data in a single dimension is as follows: Table 1. Methods for constructing time series data (showing only a single dimension)

[0044] Furthermore, the specific implementation process of step D is as follows, including: D1. Construct a comprehensive data input layer, which includes a meteorological data input layer and a hydrological data input layer; meteorological modal data is input into the meteorological data input layer through the meteorological channel, and hydrological modal data is input into the hydrological data input layer through the hydrological channel; D2. Construct an LSTM data network based on the data input from the integrated data input layer; the LSTM data network includes a hydrological LSTM data network and a meteorological LSTM data network. D3. Construct a feature fusion layer. Based on the data network architecture of LSTM data network, use the attention fusion mechanism to fuse features and then output the influencing features. D4. Construct a joint prediction layer and predict water level data based on the high-impact features output by the feature fusion layer.

[0045] Specifically, the integrated data input layer is constructed as follows:

[0046] in Represented as the Cartesian product of the two, it is used to represent a dual-channel parallel input structure. The features are represented as follows: meteorological feature space, containing 72 time steps × 4 features; hydrological feature space, containing 72 time steps × 4 features, where L represents the time step size, with a value of 72, indicating coverage of 6 hours of historical data, and each 5 minutes is used as the data interval. Based on the meteorological channel data input from the integrated data input layer, a meteorological LSTM network is constructed, and the network structure is represented as follows:

[0047]

[0048]

[0049]

[0050]

[0051] The input state is defined as follows: This is represented as the meteorological input vector at time step t. This is represented as the input weight matrix of the input gate. This is represented as the input weight matrix of the forget gate. This is represented as the input weight matrix of the output gate. The input weight matrix is ​​represented as the cell state. The hidden state is defined as follows: Represented as the hidden state at time step t−1, Represented as the hidden weight matrix of the input gate, Represented as the hidden weight matrix of the forget gate, This is represented as the hidden weight matrix of the output gate. The hidden weight matrix is ​​represented as the cell state. The symbol for the bias term is defined as follows: It is represented as the bias vector of the input gate. This is represented as the bias vector of the forget gate. This is represented as the bias vector of the output gate. Represented as a bias vector of cell states; The symbol for a gating mechanism is defined as follows: Represented as the input gate activation vector, Represented as the forget gate activation vector, Represented as the output gate activation vector; The state update symbol is defined as: The cell state at time step t is used to represent the long-term memory unit of an LSTM. This represents the hidden state at time step t, used to represent the short-term memory output of the LSTM; Based on the hydrological channel data input from the integrated data input layer, a hydrological LSTM network is constructed. The network structure is represented as follows:

[0052]

[0053]

[0054]

[0055]

[0056]

[0057] Where W represents the weight matrix, and in the meteorological LSTM network, it takes the value of . In the hydrological LSTM network, the value is taken as , is used to represent the input transformation, and its structure is similar to that of the meteorological LSTM network, so it will not be described in detail here; U represents the state matrix, and in the meteorological LSTM network, it takes the value of _____. In the hydrological LSTM network, the value is taken as , used to indicate state propagation; b represents the bias vector, which takes the value in the meteorological LSTM network. In the hydrological LSTM network, the value is taken as , used to indicate offset adjustment; It is represented by the sigmoid activation function, and t−1 represents the time step, i.e. the input / hidden state at the previous time step; The meteorological LSTM network has 64 hidden units, while the hydrological LSTM network has 128 hidden units. The feature fusion layer employs an attention fusion mechanism and is represented as follows:

[0058]

[0059] in, This is represented as an attention bias term. The value is 64, representing the dimension of the attention space; Represented as a meteorological feature transformation matrix, Represented as a hydrological feature transformation matrix, both are represented as linear transformation matrices, used to map meteorological and hydrological features to the same attention space, enabling them to be compared and fused; This is represented as the preliminary fusion of meteorological and hydrological characteristics at time step jj; Represented as an activation function, it is used to introduce nonlinearity; Represented as a weight vector, the fused 64-dimensional vector is compressed into a scalar. ; Attention weights are calculated through normalization, where Represented as an exponential function, this represents the attention energy. Convert to a positive number; It is represented as the sum of the exponential energy over all time steps, which serves as a normalization function; its calculated value is... Let the attention weights be represented as those at the j-th time step; all weights satisfy... ,pass It enables the focus on key moments, such as the time step when a rainstorm begins or the inflection point of the water level, which will be given higher weight. By interactively weighting and fusing hydrological and meteorological features, the fused features are output, represented as follows:

[0060] in, It is a weighted average fusion feature that captures important information across time steps and across modes. This corresponds to critical moments, such as sudden increases in heavy rainfall or turning points in water levels; This means that the 64-dimensional meteorological features and 128-dimensional hydrological features at time step jj are concatenated into a vector to obtain a 192-dimensional vector; This is represented by scaling the concatenated feature vector with attention weights, amplifying the features of important time steps and suppressing the features of unimportant time steps; finally, the fused features are obtained by summing, which serve as the final weighted average fused feature vector. The vector already contains information from all time steps within the entire time window and reflects the model's dynamic focus on key time steps. Based on the fusion features, high-impact features are extracted, represented as follows:

[0061] in , ; , This represents the weights and biases of the fully connected layer. Its function is to perform a one-step nonlinear transformation and dimensionality fine-tuning on the fused feature z to extract the high-impact features most effective for prediction. This is represented as layer normalization, used to stabilize the training process and accelerate convergence; When extracting high-impact features, the attention weights of hydrological and meteorological features are dynamically adjusted when the following situations occur; For example, during periods dominated by meteorological conditions, such as typhoons or severe weather like rain and snow, meteorological features receive higher attention weight, as represented by:

[0062] in This is represented as a definition of association; Expressed as the rate of change in rainfall intensity, Represented as radar reflectivity; The sensitivity coefficient is expressed as the rate of change in rainfall. The sensitivity coefficient, expressed as radar reflectivity, The coefficients are determined through grid search to ensure that the weighting of meteorological features is within a reasonable range; For example, when identifying water temperature as the dominant factor, such as during flood season or when upstream water levels rise significantly, hydrological features receive higher attention weight, as represented by:

[0063] in This is represented as a definition of association; It is expressed as the rate of change of flow, and is the first derivative; It is expressed as the acceleration of the flow rate change, and is the second derivative; A value of 1 indicates the sensitivity coefficient for the rate of change in flow rate. A value of 0.5 represents the flow acceleration sensitivity coefficient. The value is relatively small because the second derivative has greater noise. For example, when meteorological and hydrological factors interact synergistically, such as a typhoon bringing abundant rainfall that causes water levels to rise, the resulting mutual influence can be represented as:

[0064] in This is represented as an association definition; It is expressed as the product of cumulative rainfall and water level gradient; It is expressed as the product of wind speed and the rate of change of flow rate; A value of 0.7 represents the rainfall-water level synergy coefficient. A value of 0.3 represents the wind speed-flow rate coordination coefficient. This allows for dynamic adaptation to real-world application scenarios; The joint prediction layer receives By using a fully connected network nonlinear transformation, the predicted water level for the next K time steps is output. First, feature transformation and nonlinear activation are performed. The feature transformation is expressed as follows:

[0065]

[0066] Nonlinear activation is performed using the GeLU activation function, expressed as:

[0067] Then, a multi-step prediction output is performed, represented as:

[0068] Specifically, the multi-step prediction output is represented as follows:

[0069] Then, the predicted water level data for the next k time steps is output; During the training update process, the dual-channel LSTM prediction model is initialized, and the model architecture is represented as follows:

[0070]

[0071]

[0072]

[0073] in, Represented as a meteorological LSTM hidden state sequence, initialized with a Glorot uniform distribution; Represented as a sequence of hidden states in a hydrological LSTM, initialized with orthogonal initialization; The weights of the feature fusion layer are represented by the initial values ​​following a He normal distribution. Represented as output layer weights, initialized to zero; During training, historical data is used as input to the model. This data is then fed into the model via forward propagation and attention fusion to predict the output value. The predicted value is then compared with the historical data to calculate the loss. The loss function is expressed as:

[0074] Based on the computational structure of the loss calculation function, dynamic weight adjustment is performed, as follows:

[0075] This enables the training of the dual-channel LSTM prediction model using historical data; The trained model can also be updated online. The update trigger condition is:

[0076] The online learning algorithm is represented as:

[0077]

[0078] Gradient pruning is also employed to ensure that the online learning samples better match real-world application scenarios, as shown below:

[0079] This allows for online model updates, further adapting the model to real-world application scenarios.

[0080] Optionally, step D4 may also include: D4.1 The joint prediction layer is configured with a loss function, which is a dual-channel weighted NRMSE. The feature weights are then dynamically adjusted based on the calculation results of the loss function. D4.2 The joint prediction layer also uses hyperparameter configuration, with parameters being Dropout ratio, learning rate, batch size, and time step.

[0081] In the specific implementation process, in order to reduce the impact of the amplitude of the predicted value, the normalized root mean square error (NRMSE) of the normalized RMSE value is used as the loss function to evaluate the accuracy of the prediction model, thereby dynamically adjusting the feature weights, as follows:

[0082] The hyperparameters of the corresponding LSTM neural network are set as shown in Table 2, including: Table 2 Hyperparameters of LSTM Neural Network

[0083] Furthermore, the specific implementation process of step E is as follows, including: E1. Receive real-time multimodal data transmitted by the multi-source data acquisition layer, separate and fit to form hydrological modal data and meteorological modal data; E2. Input the hydrological mode data and meteorological mode data into the integrated data input layer, and input the hydrological mode data into the hydrological channel and the meteorological mode data into the meteorological channel respectively; E3. Based on the established data network architecture of hydrological LSTM data network and meteorological LSTM data network, meteorological modal data and hydrological modal data that have passed through the corresponding data channels are input into the feature fusion layer to fuse features and then output the influencing features. E4. Based on the high-impact features output by the feature fusion layer, the data is input into the joint prediction layer to obtain the predicted water level data at the target sluice gate for several future time steps.

[0084] Specifically, in the actual prediction process, hydrological mode data and meteorological mode data are respectively input through the hydrological and meteorological channels of the integrated data input layer. Based on the constructed hydrological LSTM data network and meteorological LSTM data network, the current state vector is constructed through the LSTM data network, represented as:

[0085]

[0086]

[0087]

[0088]

[0089]

[0090] This is followed by a multi-step prediction loop to predict water level data for several future time steps, while incorporating physical constraints to make it conform to practical application scenarios, represented as follows: ,in Represented as traffic sensitivity, Expressed as water level difference coefficient, Represented as the gate characteristic constant; The system will further update the status, iteratively update the meteorological channel forecast data and the hydrological channel prediction values, and predict the water level data based on the prediction values. The prediction results will be stored and the next iteration will be carried out until the prediction results for the set number of time steps are predicted. The prediction results will be output as the predicted water level data for several time steps.

[0091] Furthermore, the specific implementation process of step F is as follows, including: F1. Based on the predicted water level data at the target sluice gate in several future time steps, the rolling optimization control algorithm is used for optimization, and then the optimal opening sequence is solved. F2. Based on the optimized optimal opening sequence, a smooth opening change trajectory is generated using a fifth-order polynomial function; F3, based on the smooth opening change trajectory generated by F3, is converted into a gate opening control command using a PLC instruction conversion algorithm; F4. Based on the gate opening control command, the gate opening is adjusted accordingly.

[0092] Specifically, based on the predicted water level data for several future time steps, a rolling optimization control algorithm is used for optimization, as follows:

[0093] The optimization objective is an optimization problem that includes three objectives: water level deviation, gate opening change rate, and flow tracking, along with the corresponding gate safety constraints, which are expressed as follows:

[0094]

[0095]

[0096]

[0097]

[0098] in, This is represented as the gate opening at step k; This is expressed as the predicted upstream water level. This is represented as the target water level. The target value is set based on the scheduling plan and can be dynamically adjusted according to factors such as the actual date and flood season. This is expressed as the maximum permissible change in opening degree; Represented as traffic forecast value; This represents the target traffic, which is set according to the actual application scenario; , , These represent the weight coefficients for each optimization objective. + + =1; This is represented as a dynamic velocity constraint function, with a value ranging from 0.5 to 1; This is expressed as a safety margin coefficient; It is expressed as the standard deviation of the forecast, reflecting the uncertainty of the forecast; Based on optimization of three different objectives, this method addresses the problem that traditional sluice gate control methods, which only consider water level deviation, lead to frequent changes in opening degree and accelerated equipment wear. It also balances control accuracy, equipment lifespan, and energy efficiency. The rolling optimization method re-optimizes within each control cycle, enabling timely responses to system changes and external disturbances. Based on the opening sequence adjusted within the control cycle, a smooth opening change trajectory is generated using a fifth-order polynomial function. Simultaneously, the boundary conditions of the trajectory are dynamically adjusted according to real-time operating conditions, as expressed below:

[0099] The boundary conditions for dynamically adjusting the trajectory based on real-time operating conditions are expressed as follows:

[0100]

[0101]

[0102]

[0103]

[0104]

[0105] in, ~ Represented as polynomial coefficients, they are solved using boundary conditions; Boundary conditions are divided into three categories, including location boundary conditions. This represents the opening degree at the start of the trajectory, used to indicate the current actual opening position of the gate; This represents the opening at the end of the trajectory, used to indicate the target opening value obtained through optimization calculation; This is represented as the currently measured aperture value; This is represented as the target opening degree to be optimized; Including velocity boundary conditions, It represents the rate of change of the gate opening at the start of the trajectory, and is used to reflect the actual speed of the gate's movement at the current moment; It is expressed as the rate of change of opening at the end of the trajectory, used to reflect the target end velocity of the plan; This is expressed as the rate of change of the current opening degree; It is represented as the terminal velocity coefficient, used to indicate the magnitude of the control terminal velocity; Including acceleration boundary conditions, The acceleration at the start of the trajectory is used to reflect the current acceleration during control, and is generally set to 0. Used to represent the acceleration at the end of the trajectory, and to reflect the target acceleration at the end of the planned trajectory; Other parameters are decision parameters, including: T represents the trajectory execution time, which is dynamically adjusted based on the distance during the actual adjustment, reflecting the total time to complete the gate opening adjustment; This is represented as predicted water level data; This is represented as the target water level, used to indicate the water level value required by the dispatch plan; This is expressed as the water level deviation threshold, reflecting the allowable deviation in water level control. It is represented as a water level gradient, reflecting the trend and direction of water level changes; It is represented as a function of water level change direction, reflecting the trend of water level rising or falling; Substituting the boundary conditions into the fifth-degree polynomial and its derivative, the solution is obtained. ~ The above solution yields the change trajectory, and the boundary conditions effectively limit and optimize the predictable changes from the opening to the closing of the gate, enabling further adaptation to the allocation plan set at the gate; effectively avoiding the impact of step changes on the gate's electromechanical equipment, while improving the stability of water level control and adapting to different positioning strategies under different working conditions. After obtaining the optimal opening sequence and change trajectory, the above data is converted into PLC instructions with safety constraints and timing information. Depending on the actual working conditions, such as normal conditions and emergency situations, the absolute position mode or trajectory tracking mode is adopted respectively. In absolute position mode, which is used for general adjustment, the upper limit of speed and acceleration can be adjusted in real time according to changes in water level and rainfall, ensuring response speed while avoiding equipment overload; the command parameters are dynamically configured and expressed as follows:

[0106] The formula represents the optimal opening value obtained directly from optimization calculation at that control moment;

[0107] This formula represents the optimal gate opening or closing speed. This represents the maximum permissible speed of the equipment. Represented as the influence function of water level gradient, This indicates that the faster the water level changes, the higher the allowable adjustment speed.

[0108] This formula represents the optimal gate opening acceleration or closing acceleration. This is expressed as the maximum permissible acceleration of the equipment; This is expressed as the influence function of cumulative rainfall. This is used to indicate that, for safety reasons, the greater the rainfall, the stricter the acceleration limit; In trajectory tracking mode, it is more suitable for high-precision scenarios, such as those requiring minute adjustments to balance upstream and downstream water levels or reservoir storage and release. It directly uses optimized calculated velocity and acceleration values, but limits them to the maximum allowable range to ensure safety. Command parameters are dynamically configurable, as shown below:

[0109] Wherein represents the optimal opening value obtained directly from optimization calculation at that control moment;

[0110] in This represents the velocity amplitude of the optimized trajectory; coefficient 2 is a constant coefficient used to represent a safety margin of 2 times; it is used to ensure that the speed limit does not exceed twice the optimized value and the maximum value of the equipment.

[0111] in This represents the acceleration amplitude of the optimized trajectory. The coefficient 1.5 is a constant coefficient used to represent a safety margin of 1.5 times; it is used to limit the acceleration to no more than 1.5 times the optimized value and the maximum value of the equipment. The aforementioned safety margin calculation is to determine the basic safety margin during the working process, expressed as:

[0112] The calculation of safety margin involves two normal influencing factors, including environmental risk increment and equipment health increment. In this embodiment, two normal influencing factors are selected. If the environmental complexity at the target gate is too great, several increments can be dynamically added to further adapt to the actual scenario. In the calculation of risk increment, it is expressed as: ;

[0113] in and These are expressed as weighting coefficients, α=0.6, β=0.4; ∈[0,1], is used to represent the comprehensive water level safety margin and rainfall risk, and then to calculate the final risk increment; In the calculation of device health increment, it is represented as: ;

[0114] in =0.4, =0.3, =0.3, indicating the weighting coefficient of each index; , and This represents various indices of the device, obtained by checking the logs; and then the specific safety margin is calculated. For instructions that are about to be generated, timestamp synchronization is performed, represented as follows:

[0115] in, This represents a control time timestamp. This represents the processing delay, set to 50ms. This represents the transmission delay, set to 20ms. This is represented as a buffer time, set to 30ms; synchronization via timestamps ensures timing consistency for multi-device collaborative control; timestamps are bound to control commands. During the sending and execution process, CRC verification is required. Verification is performed before sending to ensure the integrity of the instruction, and after receiving to confirm the correctness of the instruction. This enables the control of the target sluice gate's opening degree according to the PLC instructions.

[0116] In other embodiments, a digital twin verification and simulation engine is also included, which performs digital twin prediction of execution results before PLC instruction execution for further implementation reference, in order to further solve the existing technical pain point of "blind execution of control instructions" in traditional methods. The digital twin engine comprises a physical model layer, a data-driven layer, and a fusion prediction layer; The physical model layer includes hydraulic and mechanical dynamics models; the data-driven layer includes LSTM prediction models and real-time data assimilation devices; the fusion prediction layer is used to fuse various models, perform joint simulations, and output results. In the specific implementation process, the instruction to be executed is substituted into the fusion prediction, and multiple models are used for parallel simulation. The output results of the multiple models are fused to output the simulation results, and the simulation results are evaluated and the confidence level is calculated. Based on the evaluation results and the confidence level of the evaluation results, a decision is made to determine whether to execute the PLC instruction. The introduction of a digital twin engine enables a paradigm shift from "execute first, observe later" to "verify first, execute later," while filling the technological gap in virtual debugging. In traditional water conservancy control systems, where effective virtual debugging methods are lacking, digital twins provide a realistic simulation environment and testing platform, supporting the pre-verification and optimization of control strategies.

[0117] Optionally, step F2 further includes: Based on the generated control commands, a safety constraint algorithm is used to apply control constraints to the corresponding control commands based on the set safety constraint conditions.

[0118] Specifically, a three-level safety constraint system is adopted, represented as follows:

[0119] Among them, constraint type one is the water level safety boundary, and the constraint factor is... When the predicted water level exceeds the water level safety boundary, the response strategy is to limit the opening width [0.3, 0.7]. Constraint type two is for rainstorm protection, with a constraint factor of [missing information]. When the predicted rainfall intensity has reached the level of a rainstorm, the response strategy is to reduce the opening by 20%. Constraint type three is flow mutation suppression, with a constraint factor of 1. The flow rate has exceeded the safety threshold, and the response strategy is to limit the opening change rate.

[0120] A gate opening control system for executing the above-described gate opening control method based on machine learning algorithms includes: The multimodal data acquisition module is used to acquire and transmit real-time multimodal data affecting the gate opening during the acquisition time period. The dual-channel LSTM prediction model unit is used to train and update the prediction model based on the real-time multimodal data and historical input data collected by the multimodal data acquisition module, and to construct a dual-channel data input. The data transfer module is used to receive real-time multimodal data transmitted by the multimodal data acquisition module, and separate the meteorological data and hydrological data therein to transmit to the dual-channel LSTM prediction model unit, thereby generating predicted water level data; The instruction generation module is used to generate dynamic control instructions for gate opening based on the predicted water level data generated by the dual-channel LSTM prediction model unit. The instruction execution module controls the gate according to the gate opening dynamic control instruction generated by the instruction generation module.

[0121] This invention discloses a gate opening control method and system based on machine learning algorithms. The dual-channel LSTM prediction model unit employs dual-channel processing to process meteorological and hydrological modal data separately, extracting corresponding features. An attention mechanism is used to dynamically adjust the weighting of meteorological and hydrological factors, making subsequent prediction results more suitable for actual application scenarios and further improving prediction accuracy. In the subsequent control command generation, a rolling optimization generation and adjustment algorithm is used to further improve scenario adaptability, ensuring that the gate opening control and adjustment conforms to actual application scenarios. Through three core technologies—dual-channel processing, attention fusion, and rolling optimization—this invention overcomes industry challenges in water conservancy control, such as the disconnect between meteorological and hydrological responses and insufficient long-term prediction accuracy, providing reliable technical support for smart water conservancy construction.

[0122] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A gate opening control method based on machine learning algorithms, characterized in that, include: Step A, Real-time Multimodal Data Acquisition and Feature Construction: Construct a multi-source data acquisition layer to collect real-time multimodal data affecting gate opening at several acquisition time steps; the real-time multimodal data includes meteorological modal data and hydrological modal data; Step B: Constructing dual channels and channel preprocessing algorithms: Construct meteorological and hydrological channels, and construct and execute independent data preprocessing algorithms for each channel. Step C: Constructing a dual-channel multidimensional time series: Based on real-time multimodal data, define the feature dimensions of the meteorological and hydrological channels, construct the multidimensional time series matrix of the corresponding channels based on the defined feature dimensions, and construct the sliding window length; Step D: Based on the set neural network hyperparameters and the constructed dual channels, construct a dual-channel LSTM prediction model; Step E, Data Prediction: Based on real-time multimodal data, input into a dual-channel LSTM prediction model to obtain predicted water level data for several future time steps; Step F, Generation and Execution of Dynamic Control Command for Gate Opening: Based on the predicted water level data for several future time steps, a dynamic control command for gate opening is generated, and then the gate opening is adjusted accordingly based on the dynamic control command.

2. The gate opening control method based on machine learning algorithm according to claim 1, characterized in that, The specific implementation process of step A is as follows, including: A1. The data acquisition layer includes the electromechanical sensing layer and the meteorological sensing and processing layer; A2. Based on the meteorological sensing and processing layer, the rainfall intensity, temperature, wind speed, reflectivity factor and infrared cloud image of the target sluice gate within several collection time steps are obtained. The spatiotemporal alignment algorithm is used to align and fit the meteorological data collected at different time steps, thereby generating meteorological modal data. A3. Based on the electromechanical sensing layer, acquire hydrological data within several acquisition time steps at the target sluice gate, including water level, flow rate and velocity, and then fit and generate hydrological mode data.

3. The gate opening control method based on machine learning algorithm according to claim 1, characterized in that, The specific implementation process of step B is as follows, including: B1. Construct meteorological and hydrological channels; B2. The meteorological channel has a built-in preprocessing algorithm for meteorological modal data; B3. The hydrological channel has a built-in algorithm for preprocessing hydrological modal data.

4. The gate opening control method based on machine learning algorithm according to claim 3, characterized in that, The specific implementation process of step C is as follows, including: C1. Define the four-dimensional meteorological channel characteristics, including radar reflectivity, ground rainfall intensity, 6-hour cumulative rainfall, and wind speed. C2. Define four-dimensional hydrological channel characteristics, including upstream water level, downstream water level, flow rate of change, and cumulative rainfall over time. C3. Based on the defined 4-dimensional meteorological channel feature dimension and 4-dimensional hydrological channel feature dimension, construct MxN meteorological multidimensional time series matrix and MxN hydrological multidimensional time series matrix respectively, where M is the number of dimensions, N is the number of data, and N is the number of time-aligned data samples. C4. Based on the meteorological multidimensional time series matrix and hydrological multidimensional time series matrix obtained in step C3, the sliding window method is used for processing and output.

5. The gate opening control method based on machine learning algorithm according to claim 1, characterized in that, The specific implementation process of step D is as follows, including: D1. Construct a comprehensive data input layer, which includes a meteorological data input layer and a hydrological data input layer; meteorological modal data is input into the meteorological data input layer through the meteorological channel, and hydrological modal data is input into the hydrological data input layer through the hydrological channel; D2. Construct an LSTM data network based on the data input from the integrated data input layer; the LSTM data network includes a hydrological LSTM data network and a meteorological LSTM data network. D3. Construct a feature fusion layer. Based on the data network architecture of LSTM data network, use the attention fusion mechanism to fuse features and then output the influencing features. D4. Construct a joint prediction layer and predict water level data based on the high-impact features output by the feature fusion layer.

6. The gate opening control method based on machine learning algorithm according to claim 5, characterized in that, The specific implementation process of step E is as follows, including: E1. Receive real-time multimodal data transmitted by the multi-source data acquisition layer, separate and fit to form hydrological modal data and meteorological modal data; E2. Input the hydrological mode data and meteorological mode data into the integrated data input layer, and input the hydrological mode data into the hydrological channel and the meteorological mode data into the meteorological channel respectively; E3. Based on the established data network architecture of hydrological LSTM data network and meteorological LSTM data network, meteorological modal data and hydrological modal data that have passed through the corresponding data channels are input into the feature fusion layer to fuse features and then output the influencing features. E4. Based on the high-impact features output by the feature fusion layer, the data is input into the joint prediction layer to obtain the predicted water level data at the target sluice gate for several future time steps.

7. The gate opening control method based on machine learning algorithm according to claim 6, characterized in that, The specific implementation process of step F is as follows, including: F1. Based on the predicted water level data at the target sluice gate in several future time steps, the rolling optimization control algorithm is used for optimization, and then the optimal opening sequence is solved. F2. Based on the optimized optimal opening sequence, a smooth opening change trajectory is generated using a fifth-order polynomial function; F3. Based on the smooth opening change trajectory generated by F2, a PLC instruction conversion algorithm is used to convert it into a gate opening control instruction; F4. Based on the gate opening control command, the gate opening is adjusted accordingly.

8. The gate opening control method based on machine learning algorithm according to claim 7, characterized in that, Step F3 further includes: Based on the generated control commands, a safety constraint algorithm is used to apply control constraints to the corresponding control commands based on the set safety constraint conditions.

9. A gate opening control system for executing the gate opening control method based on machine learning algorithm according to any one of claims 1-8, characterized in that, include: The multimodal data acquisition module is used to acquire and transmit real-time multimodal data affecting the gate opening degree of the target gate during the acquisition time period. The dual-channel LSTM prediction model unit is used to train and update the prediction model based on the real-time multimodal data and historical input data collected by the multimodal data acquisition module, and to construct a dual-channel data input. The data transfer module is used to receive real-time multimodal data transmitted by the multimodal data acquisition module, and separate the meteorological data and hydrological data therein to transmit to the dual-channel LSTM prediction model unit, thereby generating predicted water level data; The instruction generation module is used to generate dynamic control instructions for gate opening based on the predicted water level data generated by the dual-channel LSTM prediction model unit. The safety constraint module, based on the gate opening dynamic control command generated by the command generation module, outputs the adjusted gate opening dynamic control command through the set safety threshold. The instruction execution module controls the target gate accordingly based on the gate opening dynamic control instruction generated by the safety constraint module.

Images