A deep learning-based control valve flow opening degree optimization method and system

Abstract

The application relates to the technical field of industrial process control, and specifically discloses a regulating valve flow opening degree optimization method and system based on deep learning, which comprises the following steps: collecting multi-source operation data of a split-range control system, analyzing control signals, valve actions and flow responses, judging whether a switching point gain mutation exists, if the mutation exists, predicting future flow demand and estimating a current measurement noise level by using a pre-trained time sequence convolution network and an attention mechanism hybrid model, dynamically calculating an optimal switching point and a dynamic overlap area width according to the flow prediction sequence and the noise level, and adaptively adjusting the dynamic overlap area width according to a flow change rate; and comparing flow characteristics of large and small valves, adopting a linear distribution method if the characteristics are the same, or designing a nonlinear interpolation method to distribute valve opening degrees in the dynamic overlap area, so that the gain mutation problem near the switching point in the split-range control is solved, the adaptive optimization of the flow opening degree is realized, and the system stability and control precision are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, specifically to a method and system for optimizing the flow opening of a control valve based on deep learning. Background Technology

[0002] In process industries such as chemical, petroleum, and power, split-range control is a common control strategy used to expand the adjustable range or handle operating conditions with different flow characteristics by employing two or more parallel regulating valves. A typical split-range control system consists of two regulating valves, one large and one small. The smaller valve is activated at the lower end of the control signal range, while the larger valve is activated at the higher end. A switching point and a dynamic overlap zone are set in between to achieve a smooth transition.

[0003] However, split-range control faces a core challenge in practical applications: abrupt gain changes near the switching point. Due to differences in the size and flow characteristics (such as inherent flow characteristics and installation characteristics) of the two valves, the gain of the total flow rate relative to the opening degree will change drastically when the control signal passes through the switching point. This leads to controller parameter mismatch, causing the control signal to repeatedly cross the switching point, and the two valves to "compete" for control in the dynamic overlap region (i.e., alternating opening degrees), ultimately resulting in continuous oscillation of the flow output, which seriously affects the stability of downstream processes and product quality.

[0004] Traditional solutions often rely on engineers' experience, manually setting fixed switching points and dynamic overlap zones for debugging. This approach is not only time-consuming and labor-intensive, but also unable to adapt to dynamic changes in operating conditions such as load variations, valve wear, and changes in media characteristics. Some existing technologies have proposed adaptive split-range control methods, but these are usually based on simple linear models or rule bases, exhibiting poor robustness to nonlinearity, time-varying characteristics, and measurement noise. Furthermore, they cannot predict changes in flow demand in advance, resulting in limited optimization effectiveness.

[0005] Therefore, this invention provides a method and system for optimizing the flow opening of a control valve based on deep learning. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies and solve at least one of the technical problems mentioned in the background, this invention provides a method and system for optimizing the flow opening of a control valve based on deep learning. Addressing the problems of flow jumps, control oscillations, and frequent valve alternation caused by sudden changes in total flow gain near the switching point in multi-valve split-range control systems, this invention uses deep learning to predict flow demand and noise, dynamically adjusts the overlap width and switching point, and adaptively selects linear or nonlinear allocation methods based on valve characteristics to eliminate flow jumps and oscillations caused by sudden changes in gain at the switching point.

[0007] The objective of this invention can be achieved through the following technical solution: a method for optimizing the flow opening of a control valve based on deep learning, comprising the following steps: Multi-source operating data of the split-range control system of the regulating valve are collected to construct a multivariable time series. The control signal, valve action and flow response in the multivariable time series are analyzed to determine whether there is a sudden change in gain at the switching point between the flow and the control signal. If present, a pre-trained hybrid model of temporal convolutional network and attention mechanism is used, combined with multivariate time series data, to output a traffic demand prediction sequence and calculate the noise level of the current traffic measurement value. Based on the flow demand forecast sequence and the current flow measurement noise level, the optimal switching point and dynamic overlap width of the split-range control are dynamically calculated, and the dynamic overlap width is adaptively adjusted according to the rate of change of the flow demand forecast sequence. The flow characteristics of the large valve and the small valve are compared and analyzed to determine whether the flow characteristics of the large valve and the small valve are the same. If they are the same, the opening degree of the large valve and the small valve is allocated by the linear allocation method in the dynamic overlap region. If they are not the same, a nonlinear interpolation method is designed to allocate the opening degree of the large valve and the small valve in the dynamic overlap region.

[0008] Furthermore, the method for determining whether there is a sudden change in gain at the switching point between the flow rate and the control signal is as follows: The total flow measurement value, upstream pressure value, downstream flow demand signal, small valve current opening command, and large valve current opening command are synchronously collected at the same sampling frequency to form a multivariate time series. Median filtering and Z-score normalization preprocessing are then performed on the multivariate time series. Set a sliding window to monitor the average crossing frequency of the total control signal crossing the preset switching point in real time. If the frequency exceeds the preset frequency, it is recorded as an abnormal control signal. Simultaneously monitor the small valve opening command and the large valve opening command in the window. When the small valve opening increases and the large valve opening decreases, or the small valve opening decreases and the large valve opening increases, and both changes exceed the dead zone, it is recorded as one reverse alternation. If the reverse alternation frequency exceeds the reverse alternation frequency threshold, it is recorded as abnormal valve action. Calculate the peak-to-valley difference of the total flow rate within the sliding window. If the peak-to-valley difference exceeds two percent of the range and the flow waveform shows obvious periodicity, it is recorded as an abnormal flow response. If any two of the control signal, valve action, and flow response are abnormal, it is determined that there is a sudden change in switching point gain between the flow and the control signal.

[0009] Furthermore, the pre-training process of the hybrid model of temporal convolutional network and attention mechanism is as follows: A parallel dual-branch architecture is adopted, including a temporal convolutional network branch and a multi-head self-attention branch. The two branches process the input sequence independently and then fuse the extracted features. The temporal convolutional network branch is composed of multiple residual blocks connected in series. Each residual block contains a dilated convolutional layer, weight normalization, ReLU activation function and Dropout layer, and has residual connections to extract local dynamic features. After adding positional encoding to the input sequence through the multi-head self-attention branch, long-distance dependent features are extracted through the multi-head scaled dot product attention mechanism and the feedforward network, and the feature vector is obtained through global average pooling. The feature vectors output by the temporal convolutional network branch and the multi-head self-attention branch are concatenated end to end to form a fusion vector, which is then input into the prediction head composed of fully connected layers to output a traffic demand prediction sequence for multiple future time periods.

[0010] Furthermore, the calculation process for the noise level of the current flow measurement value is as follows: The residual between the predicted flow demand sequence and the actual flow measurement is calculated in real time to obtain the residual sequence. The mean square error is statistically analyzed using a sliding window on the residual sequence, and the obtained mean square error value is used as the noise level of the current flow measurement value.

[0011] Furthermore, the method for dynamically calculating the optimal switching point and dynamic overlap width of the split-range control is as follows: Calculate the average rate of change and instantaneous rate of change of flow based on the flow demand forecast sequence; Set a base overlap width and introduce a predicted rate of change correction term and a noise correction term. The dynamic overlap width is equal to the base overlap width plus the predicted rate of change correction term plus the noise correction term. The predicted rate of change correction term is positively correlated with the average flow rate change rate, and the noise correction term is positively correlated with the noise level of the current flow measurement value. Obtain the opening-flow function curves of the large valve and the small valve and derive them to obtain the gain curves. Within the intersection of the two gain curves of the large valve and the small valve, find the point that minimizes the absolute value of the gain difference between the two valves, and take the corresponding total control signal value as the optimal switching point.

[0012] Furthermore, the method for determining whether the flow characteristics of the large valve and the small valve are the same is as follows: Data from the individual operating regions of the small valve and the large valve are filtered out from the sliding window, and the least squares method is used to fit polynomials to obtain the flow characteristic curves of the small valve and the large valve. The flow characteristic curves of the small valve and the large valve are normalized respectively. Within the common opening range of the dynamic overlap zone, the average value of the normalized flow absolute deviation is calculated to obtain the average absolute deviation. If the average absolute deviation is less than the preset deviation, the large valve and the small valve are judged to have the same flow characteristics; otherwise, they are judged to be different.

[0013] Furthermore, the process of allocating the opening degree of the large valve and the small valve using a linear allocation method within the dynamic overlap region is as follows: Based on the optimal switching point and the width of the dynamic overlap area, determine the left boundary of the dynamic overlap area as optimal switching point - width of dynamic overlap area / 2 and the right boundary as optimal switching point + width of dynamic overlap area / 2. For the current total control signal, when the total control signal is located within the interval [left boundary, right boundary], calculate the relative position of the total control signal in the dynamic overlap area = (total control signal - left boundary) / (right boundary - left boundary); The valve opening is allocated according to a linear relationship: small valve opening = 1 - the relative position of the total control signal in the dynamic overlap area, and large valve opening = the relative position of the total control signal in the dynamic overlap area.

[0014] Furthermore, the process of allocating the opening degrees of the large and small valves within the dynamic overlap region using the nonlinear interpolation method described above is as follows: The complete range of the total control signal is divided into the small valve individual action area, the dynamic overlap area, and the large valve individual action area by using the dynamic overlap area parameter; Within the dynamic overlap region, a nonlinear allocation function based on the logic stipulation function is used to decompose the total control signal into small valve opening commands and large valve opening commands. Independent steepness parameters are set for the large valve and the small valve respectively, and the optimal steepness parameters are obtained by using a numerical optimization algorithm with the goal of minimizing the error between the actual total flow and the ideal linear total flow in the dynamic overlap area. During online operation, the opening commands of the small valve and the large valve are calculated by substituting the normalized position of the current total control signal in the dynamic overlap area into the optimized allocation function.

[0015] Furthermore, the design method of the nonlinear interpolation method is as follows: An opening allocation function is constructed based on the logistic function. The small valve opening function is set to 1 minus the logistic function value, and the large valve opening function is set to the logistic function value. The small valve opening function and the large valve opening function are defined based on the normalized position of the total control signal in the dynamic overlap area and the steepness parameters of the small valve and the large valve. The optimal steepness parameter is obtained by using the gradient descent method to find the sum of squares of the differences between the actual total flow and the ideal linear total flow at multiple sampling points within the dynamic overlap area as the optimization objective.

[0016] A deep learning-based system for optimizing the flow opening of a control valve includes the following modules: Gain mutation detection module: Collects multi-source operating data of the control valve split-range control system to construct a multivariable time series, and analyzes the control signal, valve action and flow response in the multivariable time series to determine whether there is a gain mutation at the switching point between the flow and the control signal; Deep learning prediction module: If it exists, pre-train a hybrid model of temporal convolutional network and attention mechanism, and combine it with multivariate time series to output a traffic demand prediction sequence and calculate the noise level of the current traffic measurement value; Dynamic parameter optimization module: Based on the traffic demand prediction sequence and the current traffic measurement noise level, dynamically calculate the optimal switching point and dynamic overlap width of the split control, and adaptively adjust the dynamic overlap width according to the rate of change of the traffic demand prediction sequence. Valve characteristic comparison and opening allocation module: The flow characteristics of the large valve and the small valve are compared and analyzed to determine whether the flow characteristics of the large valve and the small valve are the same. If they are the same, the linear allocation method is used to allocate the opening of the large valve and the small valve in the dynamic overlap area. If they are not the same, a nonlinear interpolation method is designed to allocate the opening of the large valve and the small valve in the dynamic overlap area.

[0017] The beneficial effects of this invention are as follows: By comprehensively analyzing the characteristics of three dimensions—control signal crossover frequency, valve reverse alternation frequency, and flow periodic oscillation—online, accurate, and quantifiable intelligent diagnosis of gain mutations at switching points is achieved, avoiding the subjectivity and lag of traditional methods that rely on human experience. By introducing a hybrid model of temporal convolutional network and attention mechanism, dynamic patterns of time series can be effectively learned from historical multivariate data. This model can accurately predict future traffic demand, provide feedforward information for dynamic adjustment, and estimate measurement noise level online, thus improving robustness under complex working conditions. Based on the predicted flow rate and real-time noise level, the width of the dynamic overlap zone is dynamically adjusted, achieving an adaptive control effect that narrows during stable conditions to improve economy and widens during drastic conditions to enhance stability. Simultaneously, the optimal switching point is calculated based on the minimum gain difference principle, fundamentally optimizing the structural parameters of the split-range control. By comparing the actual flow characteristics of large and small valves, a linear or nonlinear opening allocation strategy is intelligently selected. In particular, for valves with different characteristics, a nonlinear interpolation method based on the logistic function is designed. By optimizing the steepness parameter, the nonlinearity of the valve's inherent characteristics is perfectly compensated, and the total flow is linearized with the control signal, eliminating abrupt changes in gain. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1This is a flowchart of a deep learning-based method for optimizing the flow opening of a control valve, as described in Embodiment 1 of the present invention. Figure 2 This is a logic judgment diagram of a deep learning-based method for optimizing the flow opening of a control valve in Embodiment 1 of the present invention. Figure 3 This is a functional block diagram of a deep learning-based regulating valve flow opening optimization system in Embodiment 2 of the present invention. Detailed Implementation

[0020] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0021] Example 1: Please refer to Figures 1-2 As shown in the embodiment of the present invention, a method for optimizing the flow opening of a control valve based on deep learning specifically includes the following steps: Step 1: Collect multi-source operating data of the control valve split-range control system to construct a multivariable time series, and analyze the control signal, valve action and flow response in the multivariable time series to determine whether there is a sudden change in gain at the switching point between the flow and the control signal.

[0022] In step one, the process of whether there is a sudden change in gain at the switching point includes: synchronously collecting total flow measurement value, upstream pressure value, downstream flow demand signal, small valve current opening command, and large valve current opening command at the same sampling frequency to form a multivariate time series.

[0023] The downstream flow demand signal refers to the expected total flow value calculated by the upstream controller (such as a PID controller) based on the process target. It serves as the total control signal input for the split-range controller. The downstream flow demand signal has the same physical meaning as the total control signal and is used for gain change detection and model input to reflect the system's real-time flow demand.

[0024] Preprocessing of multivariate time series includes: performing median filtering (window width of 3), eliminating impulse noise, and performing Z-score normalization on each feature.

[0025] Set a sliding window, for example, with a window length of 1-5 minutes, monitor the total control signal of the split controller within the sliding window in real time, and count the number of times the total control signal crosses a preset switching point within the sliding window to obtain the average crossing frequency. Crossing the preset switching point is defined as changing from below the preset switching point to above the preset switching point, or changing from above the preset switching point to below the preset switching point. If the average crossing frequency within the sliding window exceeds the preset frequency, it is recorded as a control signal abnormality.

[0026] Simultaneously monitor the small valve opening command and the large valve opening command within the sliding window. Each time the small valve opening increases and the large valve opening decreases, and the change exceeds the dead zone, or the small valve opening decreases and the large valve opening increases, it is recorded as one reverse alternation. If the frequency of reverse alternation within the sliding window exceeds the reverse alternation frequency threshold, it is recorded as an abnormal valve action.

[0027] It should be noted that, in order to avoid false counting caused by noise, the dead zone is set to 0.5% (that is, a change in opening exceeding 0.5% is considered a valid action). Referring to the dead zone range of common positioners (0.2%-1%), the reverse alternation frequency threshold is 5 times per minute. Under normal coordinated control, the two valves will not frequently reverse. If it exceeds 5 times, it indicates that the allocation strategy in the dynamic overlap area is improper or that the gain change causes contention.

[0028] Within the sliding window, the difference between the maximum and minimum total flow rate is calculated to obtain the peak-to-valley difference. At the same time, the periodicity of the flow rate curve is analyzed. If the peak-to-valley difference exceeds two percent of the range and the waveform shows obvious periodicity, it is recorded as an abnormal flow response.

[0029] Specifically, the process for determining if the waveform exhibits obvious periodicity is as follows: Obtain the flow measurement value within the most recent sliding window to obtain the flow sequence; calculate the average value of the flow sequence; subtract the average value from each flow value in the flow sequence to obtain a zero-mean fluctuation sequence; copy the sequence and gradually shift it; for the shift amount k, k=1,2,3,...,N / 2, where N is the number of elements in the flow sequence; calculate the correlation coefficient R(T) between the shifted sequence and the original sequence; if there exists a shift amount T (T corresponds to at least 3 sampling points) such that R(T) is greater than 0.5 and R(2T) is greater than 0.3, and R(T) is a local peak, i.e., R(T) is greater than R(T-1) and R(T+1), and decreasing local peaks appear at shift amounts 2T and 3T, i.e., R(2T) is a local peak near this position and R(3T) is a local peak near this position, then the waveform is determined to have obvious periodicity and the flow response is abnormal.

[0030] It should be noted that the basis for the peak-to-valley difference exceeding two percent of the range is that, in industry, it is generally believed that flow fluctuations exceeding 2% will affect downstream processes.

[0031] If any two of the three indicators—control signal, valve action, and flow response—are abnormal, it is determined that there is a sudden change in switching point gain between the flow rate and the control signal.

[0032] Understandably, the logic for determining a sudden change in switching point gain is as follows: a sudden change in switching point gain will directly lead to: the controller repeatedly adjusting near the switching point - the control signal frequently crossing the switching point, the two valves competing for gain due to gain mismatch in the dynamic overlap area - the valve openings alternating in opposite directions, the closed-loop system generating continuous oscillations - the flow exhibiting periodic fluctuations. These three phenomena are physically causally related. Therefore, when any two of them occur simultaneously, it can be determined with high confidence that a sudden change in switching point gain does exist in the system.

[0033] It should be noted that the purpose of determining whether there is a sudden change in gain at the switching point between the flow rate and the control signal is to quickly identify the root cause of the problem by analyzing the control signal crossover frequency, the valve reverse alternation frequency, and the flow oscillation characteristics, and to decide whether to initiate subsequent optimization.

[0034] Step 2: If it exists, pre-train a hybrid model of temporal convolutional network and attention mechanism, and combine it with multivariate time series to output the traffic demand prediction sequence, and calculate the noise level of the current traffic measurement value.

[0035] In step two, the pre-training process of the hybrid model of temporal convolutional network and attention mechanism includes: the model adopts a parallel dual-branch architecture, and at the same time utilizes the sensitivity of temporal convolutional network to local dynamic changes and the ability of multi-head self-attention mechanism to capture long-distance dependencies. The two branches process the input sequence independently, and then the extracted features are fused to finally output the traffic demand prediction values ​​for multiple future time periods, and estimate the statistical characteristics of measurement noise online.

[0036] The preprocessed multivariate time series are used to construct overlapping sliding window samples with a fixed window length and prediction step size. Each sample contains the time series of all features within the past window length.

[0037] Specifically, set the window length and prediction step size. For example, the window length is 50 sampling points, corresponding to 5 seconds of historical data, and the prediction step size is 10 sampling points, corresponding to 1 second of future traffic demand.

[0038] The input for each sample is an input matrix consisting of all features within the past window length time steps, with the shape (window length, number of features). The corresponding output label is the sequence of actual flow values ​​for the future prediction step length time steps. The window moves forward by one step each time (e.g., 1 sampling point).

[0039] The first point to clarify is that the construction of the temporal convolutional network is as follows: First, the input matrix is ​​passed through a 1×1 convolutional layer to map the feature dimension to a unified working dimension within the model. The working dimension is used to unify the feature representation dimension of the temporal convolutional network branches and the multi-head self-attention branches. The value ranges from 32 to 256, with a default of 64 dimensions. The specific value is determined based on the number of input features, sequence length, and computational resources. Under the premise of ensuring the model's prediction accuracy, the higher the working dimension, the stronger the model's expressive power, but the computational cost also increases accordingly. In order to achieve a balance between the model's expressive power and the computational efficiency of edge devices, a working dimension of 64 dimensions is preferred.

[0040] The temporal convolutional network branch is composed of multiple residual blocks connected in series. Each residual block contains: a dilated convolutional layer with a kernel size of 3. The dilation rate increases exponentially with the block index. For example, the first block has a dilation rate of 1, the second block has a dilation rate of 2, the third block has a dilation rate of 4, and the fourth block has a dilation rate of 8.

[0041] Weight normalization: The kernel weights of the dilated convolutional layers in each residual block are normalized (i.e., the kernel weight matrix is ​​normalized by row or by channel) to accelerate model training convergence and improve training stability.

[0042] ReLU activation function.

[0043] The Dropout layer is enabled during model training with a dropout rate of 0.2; during online inference, the Dropout layer is disabled, meaning all neurons participate in computation without random dropout.

[0044] Residual connection: The input of the block is directly added to the convolution output (if the input and output dimensions are different, they are matched by a 1×1 convolution).

[0045] After all residual blocks, the feature vector of the last time step is taken as the global feature representation extracted by the temporal convolutional network branch. The dimension of the vector is equal to the working dimension of the model.

[0046] Secondly, it should be noted that the construction of the multi-head self-attention branch is as follows: Since the multi-head self-attention mechanism itself does not contain sequence order information, it is necessary to explicitly add position encoding. A sine / cosine function is used to generate a fixed position encoding, the encoding dimension is the same as the working dimension of the model, and it is added to the input features.

[0047] The input matrix is ​​first mapped to the working dimension of the model through a linear layer (without bias), and then the corresponding positional encoding is added to obtain the embedding sequence with order information.

[0048] Set the number of attention heads (e.g., 8 heads), calculate the query (Q), key (K), and value (V) matrices for the embedded sequence, calculate the attention weight of each time step to other time steps using the scaled dot product attention formula, and obtain the output by weighted summation.

[0049] A two-layer fully connected network is followed by the self-attention layer (the hidden dimension in the middle is usually 4 times the working dimension, and the activation function is ReLU). Each sub-layer (self-attention layer and feedforward network) adopts a structure of residual connection and layer normalization to alleviate the gradient vanishing phenomenon in deep network training and accelerate the model convergence speed.

[0050] We obtain an output sequence with the same length as the input. We then perform global average pooling on the temporal dimension of the sequence to obtain the feature representation of the multi-head attention branch, which is also the working dimension.

[0051] Thirdly, it should be noted that parallel fusion specifically involves concatenating the output feature vectors of the temporal convolutional network branches with the output feature vectors of the multi-head attention branches to form a fusion vector with a dimension twice that of the working dimension.

[0052] In step two, the output process of the traffic demand prediction sequence includes: the fused feature vector is input into a prediction head composed of several fully connected layers (preferably two fully connected layers, specifically the first fully connected layer maps the fused vector to 128 dimensions, and the second fully connected layer maps the 128 dimensions to T_pred dimensions, i.e., the traffic demand values ​​at the future T_pred time steps). The output layer does not use an activation function and directly outputs the traffic demand prediction sequence. The prediction head structure is: the first fully connected layer maps the fused vector to 128 dimensions with ReLU as the activation function; the second fully connected layer maps the 128 dimensions to T_pred dimensions (i.e., the traffic demand values ​​at the future T_pred time steps); the output layer does not use an activation function and directly outputs the traffic demand prediction sequence.

[0053] Where T_pred is the prediction step size, which represents the number of future time steps output by the model. The value ranges from 1 to 20 sampling points. The specific value of T_pred is determined according to the response speed of the controlled object and the sampling period. A larger value is used for systems with slower response speeds, and a smaller value is used for systems with faster response speeds. Under the condition of a sampling frequency of 10Hz, T_pred is preferably 10 sampling points, corresponding to the future flow demand of 1 second.

[0054] After the model is trained, it is put into online operation. During training, the mean absolute error (MAE) is used as the main loss function. The historical dataset is divided into training set, validation set and test set in chronological order. During the model training process, after each training round, the mean absolute error loss value is calculated on the validation set, and the trend of the validation set loss value is monitored. If the validation set loss value of 10 consecutive training rounds is not lower than the lowest validation set loss value recorded in all previous rounds, that is, the validation set loss value has not decreased further, it is determined that the model performance has not improved. At this time, training is stopped in advance to prevent overfitting, and the model parameters are rolled back to the model parameters corresponding to the round with the lowest validation set loss value as the final model parameters.

[0055] Finally, a hybrid model of temporal convolutional network and attention mechanism is obtained. The sliding window samples are input into the hybrid model of temporal convolutional network and attention mechanism, and the output is a traffic demand prediction sequence for a specified future time.

[0056] In step two, the noise level calculation process of the current flow measurement value includes: calculating the residual between the predicted flow value and the current flow measurement value in real time to obtain the residual sequence, and using a sliding window to calculate the root mean square error of the residual sequence to obtain the noise level of the current flow measurement value.

[0057] It is understandable that the physical meaning of the noise level of the current flow measurement is: representing the intensity of the measurement noise, a statistical measure of sensor random error or fluid pulsation.

[0058] It should be noted that the role of traffic demand forecasting and noise level calculation is to: predict future traffic demand and estimate noise levels, extract time-series features using deep learning models, provide feedforward information for dynamic parameter adjustment, and quantify measurement uncertainty.

[0059] Step 3: Based on the flow demand forecast sequence and the current flow measurement noise level, dynamically calculate the optimal switching point and dynamic overlap width of the split-range control, and adaptively adjust the dynamic overlap width according to the rate of change of the flow demand forecast sequence.

[0060] In step three, the calculation process of the optimal switching point and dynamic overlap width of the split-range control includes: First, it should be noted that the rate of change of the predicted flow is calculated, specifically: the flow demand prediction sequence output by the model is obtained, and the average rate of change is calculated. Specifically, the flow prediction values ​​of the earliest and the latest times in the flow demand prediction sequence are taken, and the difference between the flow prediction values ​​of the earliest and the latest times is calculated. Then, the average flow rate of change is calculated proportionally with the prediction time span to reflect the overall trend of future load changes. The absolute value of the rate of change between two adjacent points in the prediction sequence is calculated, and the maximum value is taken as the representative value of the instantaneous rate of change to capture step or rapid fluctuations.

[0061] Secondly, it should be noted that the basic value and correction item for determining the dynamic overlap width are as follows: The dynamic calculation of the dynamic overlap width follows the principle of narrowing when the situation is stable and widening when the situation is drastic. Specifically, it includes the following components: Basic overlap width: a preset minimum value to ensure that there is at least a certain dynamic overlap area under any operating condition to avoid hard switching. It is usually set according to the inherent response characteristics of the two valves and field experience.

[0062] The basic overlap width is typically set at 5%-10%, designed to ensure minimal overlap between the two valves under any operating condition, avoiding hard switching. Too small an overlap results in an unsmooth transition, while too large an overlap increases energy loss. Adjustments are made through on-site testing, generally using the valve response time matching value.

[0063] Predicted rate of change correction term: The average flow rate of change (absolute value) is multiplied by a predicted rate of change correction term coefficient. The larger the rate of change, the larger the predicted rate of change correction term coefficient, thereby increasing the width of the dynamic overlap zone and reserving more space for a smooth transition.

[0064] It is understandable that the physical meaning of the predicted rate of change correction term coefficient is to control the sensitivity of the dynamic overlap area to load changes. The larger the predicted rate of change correction term coefficient is, the faster the dynamic overlap area widens. The typical value is 0.5-2.0 (unit: seconds / %), which can be tuned through simulation or experience. The purpose is to temporarily widen the dynamic overlap area to 20%-30% during rapid step changes to provide sufficient buffer.

[0065] Noise correction term: The noise standard deviation estimate is multiplied by the noise correction term coefficient. The greater the measurement noise, the larger the noise correction term coefficient, in order to avoid frequent switching caused by false crossings due to noise.

[0066] Understandably, the noise correction factor is usually set to 0.5-1.5. The greater the noise, the wider the dynamic overlap area, which helps to avoid crossing jitter. The noise correction factor is determined based on the curve of the effect of measured noise on the crossing frequency of the switching point. You can first set the noise correction factor to 1 and observe the actual switching frequency to make adjustments.

[0067] The dynamic overlap width is equal to the base overlap width plus the predicted rate of change correction term, plus the noise correction term.

[0068] Fourthly, it should be noted that determining the optimal switching point is as follows: the optimal switching point is the position where the flow gain of the small valve and the large valve is as equal as possible when the total control signal is in that position. The specific calculation steps are as follows: obtain the opening degree-flow function curves of the two valves, differentiate the opening degree-flow function of each valve to obtain the curve of gain changing with opening degree, and plot the gain curve of the small valve and the gain curve of the large valve in the same opening degree coordinate system. The two form an intersection interval on the opening degree axis.

[0069] If the gain curves of the two valves intersect on the opening axis, then find the point within the intersection interval that minimizes the absolute value of the difference between the gain of the smaller valve and the gain of the larger valve. The corresponding total control signal value is the optimal switching point. If there are multiple points within the intersection interval with the same absolute value of the difference and all of them being the minimum, then select the total control signal value corresponding to the point located at the midpoint of the intersection interval as the optimal switching point.

[0070] If the gain curves of the two valves do not intersect on the opening axis, the total control signal value corresponding to the point where the absolute value of the difference between the gain of the small valve and the gain of the large valve is the smallest (i.e., the endpoints where the two curves are closest) is selected as the optimal switching point, and the residual gain mismatch caused by this is compensated by increasing the width of the dynamic overlap area.

[0071] It should be noted that the purpose of dynamically calculating the optimal switching point and the dynamic overlap width is to adaptively adjust the split parameters based on the predicted rate of change and noise level, so that the dynamic overlap widens when there are drastic changes and narrows when things are stable, thus avoiding sudden changes in gain.

[0072] Step 4: Compare and analyze the flow characteristics of the large valve and the small valve to determine whether their flow characteristics are the same. If they are the same, use a linear distribution method to allocate the opening of the large valve and the small valve in the dynamic overlap area. If they are not the same, design a nonlinear interpolation method to allocate the opening of the large valve and the small valve in the dynamic overlap area.

[0073] In step four, the process of determining whether the flow characteristics of the large valve and the small valve are the same includes: filtering the data of the small valve's independent operation area and the large valve's independent operation area from the sliding window, specifically including: small valve's independent operation area: the total control signal is less than the switching point minus half the width of the dynamic overlap area. At this time, the opening of the large valve is zero, and the flow rate is completely determined by the small valve.

[0074] Large valve operating zone: The total control signal is greater than the switching point plus half the width of the dynamic overlap zone. At this time, the small valve is fully open, and the flow rate is completely determined by the large valve. (Note: The small valve is usually fully open in the large valve operating zone, which needs to be confirmed according to the process. If the small valve is closed at the right boundary, the opening degree of the small valve in the large valve operating zone is zero, and the flow rate is determined only by the large valve.)

[0075] For the large valve and the small valve, with the opening command as the independent variable and the flow rate as the dependent variable, the least squares method is used to fit polynomials. The fitting order is 3 to 5, preferably 3rd order polynomial, to achieve a balance between fitting accuracy and anti-overfitting. The flow characteristic curves of the large valve and the small valve are obtained respectively. The flow characteristic curves of the large valve and the small valve are normalized respectively, that is, divided by their respective flow rates when fully open, to obtain the normalized flow rates.

[0076] To determine the common opening range that may be involved in the dynamic overlap area, specifically, the openings of both the small valve and the large valve are normalized to the [0,1] interval. Several points are uniformly selected within the common opening range [0,1] (for example, 21 points with a 5% step size). The normalized flow rate of the small valve at each normalized opening value is read from the characteristic curve of the small valve to obtain the characteristic flow rate of the small valve at each sampling point. The normalized flow rate of the large valve at each normalized opening value is read from the characteristic curve of the large valve to obtain the characteristic flow rate of the large valve at each sampling point. The average absolute deviation is calculated by taking the average absolute deviation of the absolute difference between the characteristic flow rates of the small valve and the large valve at all sampling points. If the average absolute deviation is less than the preset deviation, the flow characteristics of the two valves are determined to be the same; otherwise, they are determined to be different.

[0077] The core physical meaning of this engineering trade-off, which balances measurement noise robustness and computational efficiency while ensuring control linearity requirements, is that a complex nonlinear interpolation method is only employed when the average difference between the normalized flow characteristics of the two valves exceeds the system's allowable tolerance for abrupt gain changes. In practical applications, 5% can be used as the default threshold, and further optimized through on-site debugging or adaptive algorithms.

[0078] In step four, the process of allocating the opening of the large valve and the small valve in the dynamic overlap area using the linear allocation method includes: determining the left boundary L and the right boundary R of the dynamic overlap area based on the calculated optimal switching point S and the width W of the dynamic overlap area: left boundary L = SW / 2, right boundary R = S + W / 2. When the total control signal u is located in the interval [L, R], it is in the dynamic overlap area.

[0079] For the current total control signal u, calculate the relative position of u in the dynamic overlap region t=(uL) / (RL). The range of t is from 0 to 1. When u is equal to the left boundary, t=0. When u is equal to the right boundary, t=1.

[0080] The opening degree of the small valve and the opening degree of the large valve are calculated using the normalized position t according to a linear relationship. Specifically, the opening degree of the small valve = 1 - t, and the opening degree of the large valve = t.

[0081] In step four, the process of designing a nonlinear interpolation method to allocate the opening degrees of the large and small valves within the dynamic overlap region includes: First, it should be noted that determining the segmented intervals of the control signal is specifically as follows: The total control signal range is divided into three intervals using the dynamic overlap area parameter: Small valve independent operation area: The total control signal ranges from the minimum value to the switching point minus half the width of the dynamic overlap area. Within the small valve independent operation area, the small valve opening changes with the control signal, while the large valve remains closed (opening is zero).

[0082] Dynamic overlap zone: The total control signal ranges from the switching point minus half the width of the dynamic overlap zone to the switching point plus half the width of the dynamic overlap zone. Within this range, both valves operate simultaneously, and their respective opening degrees are determined by a nonlinear interpolation function.

[0083] Large valve independent operation zone: The total control signal ranges from the switching point plus half the width of the dynamic overlap zone to the maximum value. Within this range, the opening of the large valve changes with the control signal, while the small valve remains fully open (or gradually closes according to process requirements, usually remaining fully open to reduce resistance).

[0084] Secondly, it should be noted that the nonlinear interpolation function is designed as follows: In the dynamic overlap region of split-range control, when the total control signal changes continuously, the signal needs to be decomposed into small valve opening command and large valve opening command, so that the sum of the actual flow of the two valves has a linear relationship with the total control signal (i.e., the total flow changes uniformly with the control signal), thereby eliminating the sudden change in gain caused by the nonlinearity of valve characteristics. For this purpose, a nonlinear allocation function is designed to meet the following requirements: at the left boundary of the dynamic overlap region, the small valve should be in the fully open state and the large valve should be in the closed state.

[0085] At the right boundary of the dynamic overlap zone, the small valve should remain fully open (or gradually close according to process requirements), while the large valve should be fully open.

[0086] Within the dynamic overlap region, the opening changes of the two valves should be smooth, with the rate of change slowing down at the boundary and faster in the middle.

[0087] It can adapt to situations where the flow characteristics of the two valves are asymmetrical, meaning that the small valve and the large valve can use different steepness parameters.

[0088] Using the sigmoid function as the basic form, the function's range is between 0 and 1, with a steep change in the middle region and a gradual change at both ends. This naturally satisfies the requirement of slow change at the boundaries of the dynamic overlap region. The standard form is: f(x) = 1 / (1+e^(-1 / 2)) -k×（x-x0） ), where x is the independent variable, x0 is the inflection point (the point where the function value is 0.5), and k is the steepness parameter, k>0, the larger k is, the steeper the curve is near the inflection point.

[0089] To satisfy the boundary conditions of the dynamic overlap region (small valve fully open and large valve closed at the left boundary; small valve closed and large valve fully open at the right boundary), the small valve opening function is defined as equal to one minus the value of the logic style function, and the large valve opening function is equal to the logic style function itself. Thus, when the total control signal is at the left boundary, the small valve opening is one (fully open) and the large valve opening is zero (closed); when the total control signal is at the right boundary, the small valve opening is zero and the large valve opening is one (fully open). The sum of the two valve openings is not necessarily equal to one, but they smoothly alternate within the dynamic overlap region.

[0090] Because the flow characteristics of the large valve and the small valve are asymmetrical, different steepness parameters are set for the large valve and the small valve respectively. The steepness parameter of the small valve controls how quickly the small valve descends from fully open to closed, and the steepness parameter of the large valve controls how quickly the large valve rises from closed to fully open. Both steepness parameters are positive numbers and can be adjusted independently.

[0091] Specifically, multiple total control signal points are selected within the dynamic overlap zone (e.g., ten points at equal intervals). For each total control signal point, the opening degree of the small valve and the large valve is calculated based on the current steepness parameter. Then, the corresponding flow rate is found based on their respective opening-flow curves, and the total flow rate is obtained by summing them. At the same time, the ideal linear total flow rate corresponding to the total control signal point (i.e., the linear interpolation from the total flow rate at the left boundary to the total flow rate at the right boundary of the dynamic overlap zone) is calculated. The sum of the squares of the differences between the actual total flow rate and the ideal total flow rate at all selected points is calculated as the error function. Numerical optimization algorithms (such as gradient descent) are used to adjust the steepness parameter to minimize the error function. The final steepness parameter obtained is the optimal value.

[0092] Thirdly, it should be noted that the allocation of the opening degree of the large valve and the small valve within the dynamic overlap zone is specifically as follows: The current total control signal value is obtained, typically ranging from the minimum value (corresponding to the valve being fully closed) to the maximum value (corresponding to the valve being fully open). The current total control signal is then compared with the pre-calculated left and right boundaries. If the total control signal is less than the left boundary, it belongs to the small valve's independent operation zone; if it is greater than the right boundary, it belongs to the large valve's independent operation zone; otherwise, it belongs to the dynamic overlap zone.

[0093] In the small valve's independent operating zone: the small valve's opening degree is linearly mapped from the total control signal to the small valve's opening degree range (from minimum to maximum value), and the large valve's opening degree is set to zero.

[0094] In the main valve's independent operating range: the main valve opening degree is linearly mapped from the total control signal to the main valve's opening degree range, and the small valve opening degree is set to the maximum value (fully open).

[0095] If within the dynamic overlap zone: First, calculate the normalized variable, i.e., (total control signal minus left boundary) divided by the width of the dynamic overlap zone. Then, substitute the normalized variable into the small valve opening function and the large valve opening function. The small valve opening function is: 1 minus the logistic function (the independent variable is the normalized variable minus 0.5, multiplied by the small valve steepness parameter). The large valve opening function is: the logistic function (the independent variable is the normalized variable minus 0.5, multiplied by the large valve steepness parameter). Two opening values ​​are obtained and used as the opening commands for the small valve and the large valve, respectively.

[0096] It should be noted that the function of characteristic comparison and opening allocation is to select a linear or nonlinear allocation method based on the degree of matching of the characteristics of the two valves, and smoothly decompose the total control signal into the opening commands of the two valves in the dynamic overlap area, thereby realizing the linearization of the total flow.

[0097] The technical solution and advantages of this application embodiment are as follows: Multi-source operating data of the control valve split-range control system are collected to construct a multivariate time series. The control signal, valve action, and flow response in the multivariate time series are analyzed to determine if there is a sudden change in gain at the switching point between the flow rate and the control signal. If so, a hybrid model of a temporal convolutional network and an attention mechanism is pre-trained and combined with the multivariate time series to output a flow demand prediction sequence, and the noise level of the current flow measurement value is calculated. Based on the flow demand prediction sequence and the current flow measurement noise level, the optimal switching point and dynamic overlap width of the split-range control are dynamically calculated, and the dynamic overlap width is adaptively adjusted according to the rate of change of the flow demand prediction sequence. The flow characteristics of the large valve and the small valve are compared and analyzed to determine whether the flow characteristics of the large valve and the small valve are the same. If they are the same, the opening degree of the large valve and the small valve is allocated using a linear allocation method within the dynamic overlap region. If they are not the same, a nonlinear interpolation method is designed to allocate the opening degree of the large valve and the small valve within the dynamic overlap region. This invention collects multi-source operating data from a split-range control system, analyzes control signals, valve actions, and flow responses to determine if there are abrupt gain changes at the switching point. If such changes exist, a pre-trained hybrid model combining a temporal convolutional network and an attention mechanism is used to predict future flow demand and estimate the current measurement noise level. Based on the flow prediction sequence and noise level, the optimal switching point and dynamic overlap region width are dynamically calculated, and the dynamic overlap region width is adaptively adjusted according to the flow rate of change. The flow characteristics of the large and small valves are compared; if the characteristics are the same, a linear allocation method is used; otherwise, a nonlinear interpolation method is designed to allocate valve openings within the dynamic overlap region. This solves the problem of abrupt gain changes near the switching point in split-range control, achieves adaptive optimization of flow opening, and improves system stability and control accuracy.

[0098] Example 2: Please refer to Figure 3 As shown in the embodiment of the present invention, a deep learning-based control valve flow opening optimization system includes the following modules: Gain mutation detection module: Collects multi-source operating data of the control valve split-range control system to construct a multivariable time series, and analyzes the control signal, valve action and flow response in the multivariable time series to determine whether there is a gain mutation at the switching point between the flow and the control signal; Deep learning prediction module: If it exists, pre-train a hybrid model of temporal convolutional network and attention mechanism, and combine it with multivariate time series to output a traffic demand prediction sequence and calculate the noise level of the current traffic measurement value; Dynamic parameter optimization module: Based on the traffic demand prediction sequence and the current traffic measurement noise level, dynamically calculate the optimal switching point and dynamic overlap width of the split control, and adaptively adjust the dynamic overlap width according to the rate of change of the traffic demand prediction sequence. Valve characteristic comparison and opening allocation module: The flow characteristics of the large valve and the small valve are compared and analyzed to determine whether the flow characteristics of the large valve and the small valve are the same. If they are the same, the linear allocation method is used to allocate the opening of the large valve and the small valve in the dynamic overlap area. If they are not the same, a nonlinear interpolation method is designed to allocate the opening of the large valve and the small valve in the dynamic overlap area.

[0099] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A method for optimizing the flow opening of a control valve based on deep learning, characterized in that: Includes the following steps: Multi-source operating data of the split-range control system of the regulating valve are collected to construct a multivariable time series. The control signal, valve action and flow response in the multivariable time series are analyzed to determine whether there is a sudden change in gain at the switching point between the flow and the control signal. If present, a pre-trained hybrid model of temporal convolutional network and attention mechanism is used, combined with multivariate time series data, to output a traffic demand prediction sequence and calculate the noise level of the current traffic measurement value. Based on the flow demand forecast sequence and the current flow measurement noise level, the optimal switching point and dynamic overlap width of the split-range control are dynamically calculated, and the dynamic overlap width is adaptively adjusted according to the rate of change of the flow demand forecast sequence. The flow characteristics of the large valve and the small valve are compared and analyzed to determine whether the flow characteristics of the large valve and the small valve are the same. If they are the same, the opening degree of the large valve and the small valve is allocated by the linear allocation method in the dynamic overlap region. If they are not the same, a nonlinear interpolation method is designed to allocate the opening degree of the large valve and the small valve in the dynamic overlap region.

2. The method for optimizing the flow opening of a control valve based on deep learning according to claim 1, characterized in that: The method for determining whether there is a sudden change in gain at the switching point between the flow rate and the control signal is as follows: The total flow measurement value, upstream pressure value, downstream flow demand signal, small valve current opening command, and large valve current opening command are synchronously collected at the same sampling frequency to form a multivariate time series. Median filtering and Z-score normalization preprocessing are then performed on the multivariate time series. Set a sliding window to monitor the average crossing frequency of the total control signal crossing the preset switching point in real time. If the frequency exceeds the preset frequency, it is recorded as an abnormal control signal. Simultaneously monitor the small valve opening command and the large valve opening command in the window. When the small valve opening increases and the large valve opening decreases, or the small valve opening decreases and the large valve opening increases, and both changes exceed the dead zone, it is recorded as one reverse alternation. If the reverse alternation frequency exceeds the reverse alternation frequency threshold, it is recorded as abnormal valve action. Calculate the peak-to-valley difference of the total flow rate within the sliding window. If the peak-to-valley difference exceeds two percent of the range and the flow waveform shows obvious periodicity, it is recorded as an abnormal flow response. If any two of the control signal, valve action, and flow response are abnormal, it is determined that there is a sudden change in switching point gain between the flow and the control signal.

3. The method for optimizing the flow opening of a control valve based on deep learning according to claim 1, characterized in that: The pre-training process of the hybrid model of temporal convolutional network and attention mechanism is as follows: A parallel dual-branch architecture is adopted, including a temporal convolutional network branch and a multi-head self-attention branch. The two branches process the input sequence independently and then fuse the extracted features. The temporal convolutional network branch is composed of multiple residual blocks connected in series. Each residual block contains a dilated convolutional layer, weight normalization, ReLU activation function and Dropout layer, and has residual connections to extract local dynamic features. After adding positional encoding to the input sequence through the multi-head self-attention branch, long-distance dependent features are extracted through the multi-head scaled dot product attention mechanism and the feedforward network, and the feature vector is obtained through global average pooling. The feature vectors output by the temporal convolutional network branch and the multi-head self-attention branch are concatenated end to end to form a fusion vector, which is then input into the prediction head composed of fully connected layers to output a traffic demand prediction sequence for multiple future time periods.

4. The method for optimizing the flow opening of a control valve based on deep learning according to claim 3, characterized in that: The calculation process for the noise level of the current flow measurement value is as follows: The residual between the predicted flow demand sequence and the actual flow measurement is calculated in real time to obtain the residual sequence. The mean square error is statistically analyzed using a sliding window on the residual sequence, and the obtained mean square error value is used as the noise level of the current flow measurement value.

5. The method for optimizing the flow opening of a control valve based on deep learning according to claim 1, characterized in that: The method for dynamically calculating the optimal switching point and dynamic overlap width of the split-range control is as follows: Calculate the average rate of change and instantaneous rate of change of flow based on the flow demand forecast sequence; Set a base overlap width and introduce a predicted rate of change correction term and a noise correction term. The dynamic overlap width is equal to the base overlap width plus the predicted rate of change correction term plus the noise correction term. The predicted rate of change correction term is positively correlated with the average flow rate change rate, and the noise correction term is positively correlated with the noise level of the current flow measurement value. Obtain the opening-flow function curves of the large valve and the small valve and derive them to obtain the gain curves. Within the intersection of the two gain curves of the large valve and the small valve, find the point that minimizes the absolute value of the gain difference between the two valves, and take the corresponding total control signal value as the optimal switching point.

6. The method for optimizing the flow opening of a control valve based on deep learning according to claim 1, characterized in that: The method for determining whether the flow characteristics of the large valve and the small valve are the same is as follows: Data from the individual operating regions of the small valve and the large valve are filtered out from the sliding window, and the least squares method is used to fit polynomials to obtain the flow characteristic curves of the small valve and the large valve. The flow characteristic curves of the small valve and the large valve are normalized respectively. Within the common opening range of the dynamic overlap zone, the average value of the normalized flow absolute deviation is calculated to obtain the average absolute deviation. If the average absolute deviation is less than the preset deviation, the large valve and the small valve are judged to have the same flow characteristics; otherwise, they are judged to be different.

7. The method for optimizing the flow opening of a control valve based on deep learning according to claim 6, characterized in that: The process of allocating the opening degrees of the large valve and the small valve using a linear allocation method within the dynamic overlap region is as follows: Based on the optimal switching point and the width of the dynamic overlap area, determine the left boundary of the dynamic overlap area as optimal switching point - width of dynamic overlap area / 2 and the right boundary as optimal switching point + width of dynamic overlap area / 2. For the current total control signal, when the total control signal is located within the interval [left boundary, right boundary], calculate the relative position of the total control signal in the dynamic overlap area = (total control signal - left boundary) / (right boundary - left boundary); The valve opening is allocated according to a linear relationship: small valve opening = 1 - the relative position of the total control signal in the dynamic overlap area, and large valve opening = the relative position of the total control signal in the dynamic overlap area.

8. The method for optimizing the flow opening of a control valve based on deep learning according to claim 7, characterized in that: The process of allocating the opening degrees of the large and small valves within the dynamic overlap region using the nonlinear interpolation method described above is as follows: The complete range of the total control signal is divided into the small valve individual action area, the dynamic overlap area, and the large valve individual action area by using the dynamic overlap area parameter; Within the dynamic overlap region, a nonlinear allocation function based on the logic stipulation function is used to decompose the total control signal into small valve opening commands and large valve opening commands. Independent steepness parameters are set for the large valve and the small valve respectively, and the optimal steepness parameters are obtained by using a numerical optimization algorithm with the goal of minimizing the error between the actual total flow and the ideal linear total flow in the dynamic overlap area. During online operation, the opening commands of the small valve and the large valve are calculated by substituting the normalized position of the current total control signal in the dynamic overlap area into the optimized allocation function.

9. The method for optimizing the flow opening of a control valve based on deep learning according to claim 8, characterized in that: The design method of the nonlinear interpolation method is as follows: An opening allocation function is constructed based on the logistic function. The small valve opening function is set to 1 minus the logistic function value, and the large valve opening function is set to the logistic function value. The small valve opening function and the large valve opening function are defined based on the normalized position of the total control signal in the dynamic overlap area and the steepness parameters of the small valve and the large valve. The optimal steepness parameter is obtained by using the gradient descent method to find the sum of squares of the differences between the actual total flow and the ideal linear total flow at multiple sampling points within the dynamic overlap area as the optimization objective.

10. A deep learning-based control valve flow opening optimization system, the system being used to execute a deep learning-based control valve flow opening optimization method as described in any one of claims 1-9, characterized in that: Includes the following modules: Gain mutation detection module: Collects multi-source operating data of the control valve split-range control system to construct a multivariable time series, and analyzes the control signal, valve action and flow response in the multivariable time series to determine whether there is a gain mutation at the switching point between the flow and the control signal; Deep learning prediction module: If it exists, pre-train a hybrid model of temporal convolutional network and attention mechanism, and combine it with multivariate time series to output a traffic demand prediction sequence and calculate the noise level of the current traffic measurement value; Dynamic parameter optimization module: Based on the traffic demand prediction sequence and the current traffic measurement noise level, dynamically calculate the optimal switching point and dynamic overlap width of the split control, and adaptively adjust the dynamic overlap width according to the rate of change of the traffic demand prediction sequence. Valve characteristic comparison and opening allocation module: The flow characteristics of the large valve and the small valve are compared and analyzed to determine whether the flow characteristics of the large valve and the small valve are the same. If they are the same, the linear allocation method is used to allocate the opening of the large valve and the small valve in the dynamic overlap area. If they are not the same, a nonlinear interpolation method is designed to allocate the opening of the large valve and the small valve in the dynamic overlap area.

Images