A plate heat exchanger water temperature stability control method based on time series prediction

Abstract

The application discloses a kind of plate heat exchanger water temperature stable control method based on timing prediction, it is related to industrial thermal process control technical field, including: to the sparse timing self-attention processing of timing feature matrix, identify key historical moment and aggregate timing correlation information, obtain context-enhanced feature matrix;Based on context-enhanced feature matrix and real-time flow estimation thermal hysteresis time delay, construct thermal hysteresis compensation features and carry out multi-step decoding, obtain future multi-step outlet water temperature prediction mean sequence and prediction variance sequence;Based on future multi-step outlet water temperature prediction mean sequence, prediction variance sequence and real-time feedback deviation, calculate feedforward compensation and feedback correction, and carry out valve opening limit and change rate constraint processing, obtain regulating valve opening instruction;The application is estimated thermal hysteresis time delay by context-enhanced feature matrix and real-time flow and constructs thermal hysteresis compensation features and carries out multi-step decoding, improves the effect of outlet water temperature stability.

Description

Technical Field

[0001] This invention relates to the field of industrial thermal process control technology, and in particular to a method for stabilizing water temperature in a plate heat exchanger based on time-series prediction. Background Technology

[0002] Plate heat exchangers are widely used in scenarios such as centralized heating, industrial heat exchange, and building energy stations. The stability of their outlet water temperature usually depends on the coordinated control between the opening of the regulating valve, the state of the primary heat medium, and the secondary water load. With the development of automatic control and data-driven prediction technology, model-based heat exchange control technology is gradually being used to predict future water temperature trends based on historical temperature, flow rate, and setpoint changes, and to adjust the heat exchange process in advance in conjunction with valve actuators.

[0003] In actual operation, the internal flow channel volume of the plate heat exchanger, the heat medium transport process, and the fluctuation of the secondary side load will create a significant thermal hysteresis delay. Existing control methods based on fixed models or conventional feedback regulation usually use the current deviation as the main adjustment basis. In scenarios with rapid changes in primary side flow, sudden changes in secondary side water intake load, or continuous adjustment of setpoint, time misalignment is likely to occur between the prediction results and valve actions. This makes it difficult for the regulating valve opening correction to respond in a timely manner to future changes in outlet water temperature, thereby affecting the continuity and precision of stable outlet water temperature control. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a time-series prediction-based plate heat exchanger water temperature stabilization control method to solve the problem of control valve lag caused by outlet water temperature prediction deviation under thermal hysteresis.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for stable water temperature control of a plate heat exchanger based on time-series prediction. The method includes: collecting inlet and outlet temperatures, flow rates, regulating valve opening, and outlet temperature setpoints on both sides of the plate heat exchanger; constructing a sliding window, expanding differential features, and performing adaptive normalization to obtain a standardized historical feature matrix; performing dilated causal convolutional coding based on the standardized historical feature matrix to extract multi-scale temporal dynamic information, obtaining a temporal feature matrix; performing sparse temporal self-attention processing on the temporal feature matrix to identify key historical moments and aggregate temporal correlation information, obtaining a context-enhanced feature matrix; and based on the above... The following section enhances the feature matrix and real-time flow estimation thermal hysteresis delay, constructs thermal hysteresis compensation features, and performs multi-step decoding to obtain the predicted mean sequence and predicted variance sequence of the outlet water temperature in the future. Based on the predicted mean sequence, predicted variance sequence, and real-time feedback deviation of the outlet water temperature in the future, the feedforward compensation amount and feedback correction amount are calculated, and valve opening limit and rate of change constraint processing are performed to obtain the regulating valve opening command. Based on the regulating valve opening command, the regulating valve of the plate heat exchanger is controlled to operate, and the predicted residual is periodically monitored. The thermal hysteresis compensation decoder is fine-tuned online according to the predicted residual to obtain the outlet water temperature control result.

[0007] As a preferred embodiment of the time-series prediction-based plate heat exchanger water temperature stability control method of the present invention, the steps include collecting the inlet and outlet temperatures, flow rates, regulating valve openings, and outlet temperature setpoints on both sides of the plate heat exchanger, and performing sliding window construction, differential feature expansion, and adaptive normalization to obtain a standardized historical feature matrix. The specific steps are as follows: Collect the secondary side outlet water temperature, secondary side inlet water temperature, primary side heat medium inlet temperature, primary side heat medium outlet temperature, primary side heat medium volumetric flow rate, secondary side cold water volumetric flow rate, previous cycle regulating valve opening, and current outlet water temperature setpoint. The inlet and outlet temperature difference of the primary side heat medium is calculated based on the inlet and outlet temperatures of the primary side heat medium, and the estimated value of the instantaneous heat exchange of the primary side is calculated based on the primary side heat medium flow rate and the inlet and outlet temperature difference of the primary side. A historical feature matrix is ​​constructed based on the secondary side outlet water temperature, secondary side inlet water temperature, primary side heat medium inlet temperature, primary side heat medium outlet temperature, primary side heat medium cumulative flow, secondary side cold water volumetric flow, previous cycle regulating valve opening, current outlet water temperature setpoint, primary side inlet and outlet temperature difference, and estimated primary side instantaneous heat exchange. First-order difference features are added to the historical feature matrix to obtain the augmented historical matrix. The augmented historical matrix is ​​then normalized online using the exponential moving average statistic to obtain the standardized historical feature matrix.

[0008] As a preferred embodiment of the time-series prediction-based plate heat exchanger water temperature stabilization control method of the present invention, the step of obtaining the time-series feature matrix by performing multi-layer expanded causal convolution and residual connection processing through an expanded time-series convolution encoder based on the standardized historical feature matrix is ​​as follows: The standardized historical feature matrix is ​​subjected to input projection processing to obtain the projected feature matrix. Then, according to the hierarchical order of the dilated temporal convolutional encoder, the projected feature matrix is ​​subjected to multi-layer dilated causal convolution processing based on the kernel width, dilation factor, weight matrix and bias to obtain the convolutional features of each layer. The receptive field is expanded layer by layer according to the expansion factor corresponding to each layer of expanded causal convolution, and the multi-scale historical information in the standardized historical feature matrix is ​​covered to obtain multi-scale convolution features. The multi-scale convolutional features and the corresponding layer input features are processed by residual connection to obtain the residual temporal features of each layer. The residual temporal features of the last layer are used as the temporal feature matrix.

[0009] As a preferred embodiment of the time-series prediction-based plate heat exchanger water temperature stability control method of the present invention, the specific steps of performing sparse time-series self-attention processing on the time-series feature matrix, identifying key historical moments and aggregating time-series correlation information to obtain a context-enhanced feature matrix are as follows: Multi-head linear projection is performed on the temporal feature matrix to obtain the query matrix, key matrix and value matrix. The attention score matrix is ​​calculated based on the query matrix and key matrix. The attention score matrix is ​​then sparsified according to the learnable sparse threshold corresponding to each attention head to obtain sparse attention weights. The sparse attention weights are weighted and fused with the value matrix to obtain temporal correlation features. The temporal correlation features are then processed by multi-head concatenation, residual connection and layer normalization to obtain the context enhancement feature matrix.

[0010] As a preferred embodiment of the time-series prediction-based plate heat exchanger water temperature stabilization control method of the present invention, the steps of constructing thermal hysteresis compensation features based on context-enhanced feature matrix and real-time flow estimation thermal hysteresis delay, and performing multi-step decoding to obtain the predicted mean sequence and predicted variance sequence of future multi-step outlet water temperature are as follows: The primary side heat medium volume flow rate and the secondary side cold water volume flow rate are collected, and the thermal hysteresis delay is estimated by combining the effective flow channel volume of the heat exchanger and the empirical correction coefficient. The thermal hysteresis delay is discretized into time steps, and the context representation of the corresponding historical moment is read from the context enhancement feature matrix to obtain the thermal hysteresis compensation feature. The context representation of the latest time step in the context-enhanced feature matrix is ​​concatenated with the thermal hysteresis compensation feature to obtain the decoding input feature. The decoding input feature is then input into the thermal hysteresis compensation decoder for multi-step decoding to obtain the mean sequence and variance sequence of the predicted outlet water temperature for the next multiple steps.

[0011] As a preferred embodiment of the time-series prediction-based plate heat exchanger water temperature stabilization control method of the present invention, the method further includes joint training of the extended temporal convolutional encoder, sparse temporal self-attention processing, and thermal hysteresis compensation decoder before performing water temperature stabilization control. The specific steps are as follows: Read the future actual outlet water temperature corresponding to the prediction step size from the historical operation data, and calculate the negative log-likelihood loss based on the prediction mean sequence, prediction variance sequence and future actual outlet water temperature of the future multi-step outlet water temperature. Based on the temperature change in adjacent prediction steps and the maximum allowable single-step change in outlet temperature in the future multi-step outlet water temperature prediction mean sequence, the temperature change rate constraint term is calculated, and the negative log-likelihood loss and the temperature change rate constraint term are weighted and combined to obtain the water temperature prediction training loss. The sparse threshold regularization term is calculated based on the learnable sparse threshold and the target sparsity. The water temperature prediction training loss and the sparse threshold regularization term are combined to obtain the total training loss. The parameters of the dilated temporal convolutional encoder, the sparse temporal self-attention processing parameters, and the thermal hysteresis compensation decoder parameters are updated based on the total training loss.

[0012] As a preferred embodiment of the time-series prediction-based plate heat exchanger water temperature stabilization control method of the present invention, the specific steps for calculating the feedforward compensation and feedback correction based on the predicted mean sequence, predicted variance sequence, and real-time feedback deviation of the future multi-step outlet water temperature are as follows: The predicted temperature deviation sequence is calculated based on the mean sequence of future multi-step outlet water temperature prediction and the future outlet temperature setpoint. The predicted temperature deviation sequence is then weighted and summed according to the decay weight corresponding to the prediction step size to obtain the weighted prediction deviation. The feedforward gain is dynamically estimated based on the temperature difference of the primary heat medium to obtain the adaptive feedforward gain, and the feedforward compensation amount is calculated based on the weighted prediction deviation, rated flow rate and adaptive feedforward gain. The real-time feedback deviation is calculated based on the current outlet water temperature setpoint and the secondary outlet water temperature. The comprehensive uncertainty measure is calculated based on the predicted variance sequence. The proportional gain and integral gain are adjusted according to the comprehensive uncertainty measure to obtain the proportional gain and integral gain of the current control cycle. The feedback correction amount is calculated based on the real-time feedback deviation, the proportional gain of the current control cycle, and the integral gain of the current control cycle.

[0013] As a preferred embodiment of the time-series prediction-based plate heat exchanger water temperature stability control method of the present invention, the specific steps for performing valve opening limit and rate of change constraint processing to obtain the regulating valve opening command are as follows: The feedforward compensation and feedback correction are superimposed to obtain the initial valve opening control quantity. The initial valve opening control quantity is then subjected to upper and lower limit constraints on the valve opening to obtain the amplitude-limited valve opening control quantity. The valve opening control quantity of the limiting valve is constrained by the rate of change of adjacent control cycles to obtain the constrained valve opening control quantity. When the constrained valve opening control quantity touches the upper or lower limit of the valve opening, the integrator accumulation stops, and the regulating valve opening command is obtained based on the constrained valve opening control quantity.

[0014] As a preferred embodiment of the time-series prediction-based plate heat exchanger water temperature stabilization control method of the present invention, the step of controlling the operation of the regulating valve of the plate heat exchanger based on the regulating valve opening command and periodically monitoring the prediction residual is as follows: The control valve opening command is sent to the control valve actuator of the plate heat exchanger. The control valve actuator adjusts the control valve opening according to the control valve opening command and collects the secondary side outlet water temperature after the control valve opening is adjusted. The predicted residuals are calculated based on the predicted average sequence of multi-step outlet water temperature and the secondary outlet water temperature after the adjustment of the regulating valve opening. The root mean square of the predicted residuals within the sliding window is calculated according to the detection cycle.

[0015] As a preferred embodiment of the time-series prediction-based plate heat exchanger water temperature stabilization control method of the present invention, the specific steps of performing online fine-tuning of the thermal hysteresis compensation decoder based on the prediction residual to obtain the outlet water temperature control result are as follows: The root mean square of the predicted residuals is compared with a set threshold. When the root mean square of the predicted residuals exceeds the set threshold, the parameters of the dilated temporal convolutional encoder and the sparse temporal self-attention processing are frozen. The parameters of the thermal hysteresis compensation decoder are fine-tuned online based on the most recent running samples to obtain the updated thermal hysteresis compensation decoder parameters. Based on the updated thermal hysteresis compensation decoder parameters, the system continues to generate the predicted mean sequence, predicted variance sequence, and regulating valve opening command for future multi-step outlet water temperature. Based on the regulating valve opening command, the system controls the operation of the regulating valve of the plate heat exchanger to obtain the outlet water temperature control result.

[0016] The beneficial effects of this invention are as follows: by using context-enhanced feature matrix and real-time flow to estimate thermal hysteresis delay and constructing thermal hysteresis compensation features for multi-step decoding, the corresponding fusion of historical key moment information and heat exchange hysteresis process is realized, and a predictive basis with timing advance is provided for feedforward compensation and feedback correction, thereby improving the accuracy of regulating valve opening command generation and the stability of outlet water temperature. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a time-series prediction-based method for stabilizing water temperature in plate heat exchangers.

[0019] Figure 2 The flowchart for outputting the standardized historical feature matrix.

[0020] Figure 3 A flowchart for generating a context-enhanced feature matrix.

[0021] Figure 4 A flowchart for generating the mean and variance sequences of future multi-step outlet water temperature predictions. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for stabilizing the water temperature of a plate heat exchanger based on time-series prediction, comprising the following steps: S1. Collect the inlet and outlet temperatures, flow rates, regulating valve opening, and outlet temperature setpoints on both sides of the plate heat exchanger, and perform sliding window construction, differential feature expansion, and adaptive normalization to obtain a standardized historical feature matrix.

[0026] Collect the secondary side outlet water temperature, secondary side inlet water temperature, primary side heat medium inlet temperature, primary side heat medium outlet temperature, primary side heat medium flow rate, secondary side cold water volume flow rate, previous cycle regulating valve opening, and current outlet water temperature setpoint.

[0027] Specifically, within each control cycle, the secondary side outlet water temperature, secondary side inlet water temperature, primary side heat medium inlet temperature, primary side heat medium outlet temperature, primary side heat medium volumetric flow rate, secondary side cold water volumetric flow rate, previous cycle regulating valve opening, and current outlet water temperature setpoint are read from the plate heat exchanger. These data are then recorded according to the current control cycle to form the basic data for the current control cycle. Among these data, the secondary side outlet water temperature is used as the controlled variable, the previous cycle regulating valve opening is used as the control feedback, and the current outlet water temperature setpoint is used as the benchmark for calculating subsequent prediction deviations and real-time feedback deviations.

[0028] The inlet and outlet temperature difference of the primary side heat medium is calculated based on the inlet and outlet temperatures of the primary side heat medium, and the estimated value of the instantaneous heat exchange of the primary side is calculated based on the accumulated flow rate of the primary side heat medium and the inlet and outlet temperature difference of the primary side.

[0029] Specifically, the expression for estimating the instantaneous heat exchange on the primary side is: ; in, This represents the estimated instantaneous heat exchange value on the primary side; Indicates the density of the heat transfer medium; Indicates the specific heat capacity of the heat transfer medium; Indicates the primary heat transfer fluid flow rate; Indicates the primary side heat medium inlet temperature; This indicates the primary side heat medium outlet temperature.

[0030] A historical feature matrix is ​​constructed based on the secondary side outlet water temperature, secondary side inlet water temperature, primary side heat medium inlet temperature, primary side heat medium outlet temperature, primary side heat medium cumulative flow rate, secondary side cold water volumetric flow rate, previous cycle regulating valve opening, current outlet water temperature setpoint, primary side inlet and outlet temperature difference, and estimated primary side instantaneous heat exchange rate.

[0031] Specifically, according to the order of features, the original feature vector of the current control cycle is obtained by combining the secondary side outlet water temperature, secondary side inlet water temperature, primary side heat medium inlet temperature, primary side heat medium outlet temperature, primary side heat medium flow rate, secondary side cold water volume flow rate, previous cycle regulating valve opening, current outlet water temperature setpoint, primary side inlet and outlet temperature difference, and primary side instantaneous heat exchange estimate within the same control cycle. Using the current control cycle as the end point of the window, the original feature vectors corresponding to the continuous control cycles are read and arranged in chronological order to obtain the historical feature matrix.

[0032] First-order difference features are added to the historical feature matrix to obtain the augmented historical matrix. The augmented historical matrix is ​​then normalized online using the exponential moving average statistic to obtain the standardized historical feature matrix.

[0033] Specifically, following the chronological order of the original feature vectors in the historical feature matrix, the original feature vectors corresponding to adjacent control cycles are read one by one. The original feature vector of the current control cycle is then compared with the original feature vector of the previous control cycle using corresponding dimension differences to obtain the first-order difference feature of the current control cycle. The expression for the first-order difference feature is as follows: ; Where δr(t) represents the first-order differential eigenvector corresponding to the current control cycle; r(t) represents the original eigenvector of the current control cycle; and r(t-1) represents the original eigenvector of the previous control cycle. Indicates the meaning of difference; The time index variable represents the current control cycle; This represents the original feature vector.

[0034] The original feature vector of each control cycle is concatenated with the corresponding first-order difference feature according to the feature dimension to obtain the expanded feature vector. The expanded feature vectors of multiple consecutive control cycles are then arranged in chronological order to obtain the augmented history matrix.

[0035] Read the augmented history matrix 3D features, and based on the exponential moving average of the previous control period and the 3D feature of the current control period. 3D eigenvalue update The exponential moving mean of the dimensional feature, the th The expression for the exponential moving average of the dimensional feature is: ; in, This represents the exponential moving average of the j-th dimension feature during the current control period; This represents the exponential moving average; It represents the exponential moving average of the j-th dimension feature of the previous control period; The smoothing coefficient of the exponential moving average statistic is given, and 0 < 0. <1, used to control the weight of historical mean and current feature value in mean update; 1- This represents the weight of the j-th eigenvalue in the mean update during the current control period; This represents the eigenvalue of the j-th dimension feature in the augmented history matrix of the current control cycle.

[0036] Based on the exponential moving variance of the previous control period and the current control period's... Update the feature deviation of the first dimension The exponential moving variance of the dimensional feature, the 3rd The expression for the exponential moving variance of the dimensional feature is: ; in, Variance is used to reflect the degree to which the j-th eigenvalue fluctuates around the mean. Represents the exponential moving variance of the j-th dimension feature in the current control period; This represents the exponential moving variance of the j-th dimension feature from the previous control period.

[0037] Based on the The exponentially moving mean and exponentially moving variance of the dimensional feature, for the 3rd dimension The 3-dimensional features are standardized. The standardized expression for the dimensional feature is: ; in, This represents the standardized eigenvalue of the j-th dimension feature in the current control cycle after online normalization. This represents the exponentially moving standard deviation of the j-th dimension feature in the current control period, which is composed of the exponentially moving variance of the j-th dimension feature. The result is obtained by taking the root. It is a numerically stable term.

[0038] The features of each dimension in the augmented history matrix are normalized online in the manner described above to obtain the standardized history feature matrix.

[0039] S2. Based on the standardized historical feature matrix, dilated causal convolutional coding is performed to extract multi-scale temporal dynamic information and obtain the temporal feature matrix.

[0040] The standardized historical feature matrix is ​​subjected to input projection processing to obtain the projected feature matrix. Then, according to the hierarchical order of the dilated temporal convolution encoder, multi-layer dilated causal convolution processing is performed on the projected feature matrix based on the kernel width, dilation factor, weight matrix and bias to obtain the convolutional features of each layer.

[0041] Specifically, the feature vectors corresponding to each time position in the standardized historical feature matrix are read, and the feature vectors at each time position are mapped to the hidden layer representation space through the input projection weight matrix and input projection bias to obtain the projected feature matrix. Following the hierarchical order of the dilated temporal convolutional encoder, the features of the previous layer at the current time position and historical time positions are read, and the interval of historical time positions is determined according to the dilation factor corresponding to the current layer. Based on the convolution kernel width, dilation factor, weight matrix, and bias, dilated causal convolution is performed on the features of the previous layer, and the calculation result is activated to obtain the first... Layer convolution features, the first Layer The expression for the convolutional features at each time point is: ; in, Indicates the first The layer in the first Convolutional features output at each time point; Represents temporal characteristics; Indicates the layer number of the dilated causal convolution; This represents the index of the current time position being calculated; Indicates the position number within the convolution kernel; The width of the convolution kernel. For the first Layer expansion factor ( ); This is the weight matrix; For bias; for Activation function.

[0042] The projected feature matrix is ​​subjected to multi-layer dilated causal convolution processing in the manner described above to obtain the convolutional features of each layer.

[0043] The receptive field is expanded layer by layer according to the expansion factor corresponding to each layer of expanded causal convolution, and the multi-scale historical information in the standardized historical feature matrix is ​​covered to obtain multi-scale convolution features.

[0044] Specifically, following the principle that the dilation factor increases with the number of layers in the dilated causal convolution, the historical reading range corresponding to the convolutional features of each layer is expanded layer by layer, and short-term temperature fluctuation information, medium-range flow rate change information, and long-term thermal inertia information are read from the standardized historical feature matrix respectively; when the total number of layers in the dilated causal convolution is At that time, the expression for the total receptive field of the extended temporal convolutional encoder is: ; in, This represents the total receptive field of the extended temporal convolutional encoder; This represents the total number of layers in the dilated causal convolution; Indicates in =3 and The total receptive field calculated when =6, i.e., the output feature at the current time position, can at most cover the current control cycle and the preceding cycles. Historical information for each control cycle.

[0045] Based on the receptive field corresponding to each layer of expanded causal convolution, historical features at different time scales are extracted by convolution to obtain multi-scale convolution features.

[0046] The multi-scale convolutional features and the corresponding layer input features are processed by residual connection to obtain the residual temporal features of each layer. The residual temporal features of the last layer are used as the temporal feature matrix.

[0047] Specifically, the multi-scale convolutional features output from each layer of dilated causal convolution are read, and the corresponding layer input features are read. Residual concatenation processing is performed on the multi-scale convolutional features and the corresponding layer input features to obtain the residual temporal features of each layer. The residual temporal features of each layer are passed layer by layer according to the hierarchical order of the dilated temporal convolutional encoder, and the residual temporal features of the last layer are arranged according to their time positions to obtain the temporal feature matrix.

[0048] S3. Perform sparse temporal self-attention processing on the temporal feature matrix to identify key historical moments and aggregate temporal correlation information to obtain a context-enhanced feature matrix.

[0049] Multi-head linear projection is performed on the temporal feature matrix to obtain the query matrix, key matrix, and value matrix. The attention score matrix is ​​calculated based on the query matrix and key matrix, and the attention score matrix is ​​sparsified according to the learnable sparse threshold corresponding to each attention head to obtain sparse attention weights.

[0050] Specifically, the time series feature matrix is ​​denoted as... And linearly project the temporal feature matrix according to the attention head number to obtain the query matrix, key matrix and value matrix corresponding to each attention head; where, the calculation of the first... The query matrix corresponding to each attention head Key matrix Sum matrix The expression is: ; ; ; in, Indicates the first The query projection weight matrix corresponding to each attention head; Indicates the first The key projection weight matrix corresponding to each attention head; Indicates the first The values ​​corresponding to each attention head are projected onto the weight matrix; This indicates the attention header number.

[0051] Based on the The original attention score matrix is ​​calculated from the query matrix and key matrix corresponding to each attention head. The expression for calculating the original attention score matrix is ​​as follows: ; in, Indicates the first The original attention score matrix corresponding to each attention head; The feature dimensions of the query matrix and the key matrix; Represents the attention score matrix; This represents the raw score without sparse threshold filtering and normalization. This indicates that the original attention score matrix belongs to OK A column of real numbers.

[0052] Compare the attention scores at each historical moment in the original attention score matrix with the first... The learnable sparse thresholds corresponding to each attention head are compared, and attention scores below the learnable sparse thresholds are masked to obtain a sparse attention score matrix. Based on the original attention score matrix and the learnable sparse thresholds, the sparse attention score matrix is ​​calculated using the following expression: ; ; in, Indicates the first The i-th row and i-th column of the sparse attention score matrix corresponding to each attention head Column elements; Indicates the first The original attention score matrix corresponding to the attention head is the first one line, number Column elements; This represents the index indicating the current time position being queried; This represents the index of the historical time location being followed; Indicates the attention head number; Indicates the first The learnable sparse threshold corresponding to each attention head is automatically learned through backpropagation and initialized to . ; Indicates the first The sparse attention weight matrix corresponding to each attention head; (·) denotes the normalization exponential function, used to convert the sparse attention score matrix into an attention weight matrix.

[0053] The sparse attention weights are weighted and fused with the value matrix to obtain temporal correlation features. The temporal correlation features are then processed by multi-head concatenation, residual connection and layer normalization to obtain the context enhancement feature matrix.

[0054] Specifically, the first The sparse attention weights corresponding to the attention heads are weighted and fused with the value matrix to obtain the th attention head. Temporal correlation features corresponding to each attention head; the temporal correlation features are calculated based on sparse attention weights and value matrices. The expression for calculating the temporal correlation features is as follows: ; in, Indicates the first The temporal correlation feature matrix corresponding to each attention head; express belong OK A column of real numbers.

[0055] The temporal correlation features corresponding to each attention head are concatenated and linearly mapped using the output weight matrix to obtain the multi-head fused temporal features. The multi-head fused temporal features are then calculated based on the temporal correlation features corresponding to each attention head and the output weight matrix. The expression for calculating the multi-head fused temporal features is as follows: ; in, This represents the multi-head fusion temporal feature matrix; Indicates a splicing operation; Indicates the first to the second The temporal correlation feature matrices output by each attention head; This represents the output weight matrix; Represents the multi-head fusion temporal feature matrix For L row A column of real numbers.

[0056] The multi-head fused temporal features and the temporal feature matrix are residually concatenated and then subjected to layer normalization to obtain the context enhancement feature matrix. The context enhancement feature matrix is ​​then calculated based on the multi-head fused temporal features and the temporal feature matrix. The expression for calculating the context enhancement feature matrix is ​​as follows: ; in, Represents the context-enhanced feature matrix; (·) represents the layer normalization function; C represents the temporal feature matrix; C represents the multi-head fusion temporal feature matrix. Representation of context-enhanced feature matrix for OK A column of real numbers.

[0057] S4. Based on the context-enhanced feature matrix and real-time flow estimation thermal hysteresis delay, construct thermal hysteresis compensation features and perform multi-step decoding to obtain the mean sequence and variance sequence of future multi-step outlet water temperature prediction.

[0058] The primary side heat medium volumetric flow rate and the secondary side cold water volumetric flow rate are collected. The thermal hysteresis delay is estimated by combining the effective flow channel volume of the heat exchanger and the empirical correction coefficient. The thermal hysteresis delay is discretized into time steps, and the context representation of the corresponding historical moment is read from the context enhancement feature matrix to obtain the thermal hysteresis compensation feature.

[0059] Specifically, the primary side heat medium volumetric flow rate and the secondary side cold water volumetric flow rate of the current control cycle are read, along with the effective flow channel volume of the heat exchanger and the empirical correction coefficient. The effective flow channel volume of the heat exchanger is obtained from the plate heat exchanger equipment parameters, and the empirical correction coefficient is obtained through offline calibration. Based on the primary side heat medium volumetric flow rate, the secondary side cold water volumetric flow rate, the effective flow channel volume of the heat exchanger, and the empirical correction coefficient, the thermal hysteresis delay is estimated. The expression for calculating the thermal hysteresis delay is as follows: ; in, This indicates the effective flow channel volume of the heat exchanger (m³). This is an empirical correction factor; Indicates the current control cycle Thermal hysteresis delay; Indicates the current control cycle The next side heat medium accumulation flow; Indicates the current control cycle The volumetric flow rate of the secondary cooling water below.

[0060] Based on the thermal hysteresis delay and sampling period, the thermal hysteresis delay is discretized into time steps. Based on the corresponding time steps of the thermal hysteresis delay, the context representation of the corresponding historical moment is read from the context enhancement feature matrix, and the thermal hysteresis compensation feature is calculated. The expression for calculating the thermal hysteresis compensation feature is as follows: ; in, This represents the thermal hysteresis compensation feature extracted under the current control period t; The context-enhanced feature matrix Z represents the first... The context representation corresponding to each historical time position.

[0061] The context representation of the latest time step in the context-enhanced feature matrix is ​​concatenated with the thermal hysteresis compensation feature to obtain the decoding input feature. The decoding input feature is then input into the thermal hysteresis compensation decoder for multi-step decoding to obtain the mean sequence and variance sequence of the predicted outlet water temperature for the next multiple steps.

[0062] Specifically, the latest context representation in the context enhancement feature matrix is ​​read, and this latest context representation is concatenated with the thermal hysteresis compensation features to obtain the decoded input features. These decoded input features are then input into the thermal hysteresis compensation decoder and decoded through a fully connected decoding layer to obtain the decoded hidden layer features. Based on the decoded input features, the fully connected weight matrix, and the fully connected bias, the decoded hidden layer features are calculated. The expression for calculating the decoded hidden layer features is as follows: ; in, This represents the hidden layer features of the decoded output of the thermal hysteresis-compensated decoder; (·) denotes the Gaussian error linear unit activation function; This represents the weight matrix of the first fully connected layer of the thermal hysteresis-compensated decoder; This represents the weight matrix of the second fully connected layer of the thermal hysteresis-compensated decoder; This represents the bias term of the first fully connected layer of the thermal hysteresis-compensated decoder; This represents the bias term of the second fully connected layer of the thermal hysteresis-compensated decoder; This indicates that the latest time-time context will be represented. With thermal hysteresis compensation characteristics The decoded input features are obtained by concatenating them according to the feature dimensions.

[0063] The expression for calculating the mean sequence of predicted outlet water temperature over multiple steps is as follows: Based on the decoded hidden layer features, the mean sequence of predicted outlet water temperature over multiple steps is calculated. ; in, This represents the sequence of predicted mean values ​​of the secondary side outlet water temperature over multiple steps, output by the thermal hysteresis-compensated decoder. This represents the output layer weight matrix in the thermal hysteresis compensation decoder, which is used to output the future multi-step predicted mean sequence of outlet water temperature. This represents the output layer bias term.

[0064] The predicted variance sequence is calculated based on the decoded hidden layer features. The expression for calculating the predicted variance sequence is as follows: ; in, This represents the variance sequence of future multi-step secondary side outlet water temperature predictions output by the thermal hysteresis-compensated decoder; (·) represents the activation function, which maps the output to a non-negative value, thereby ensuring that the prediction variance is positive; This represents the variance output layer weight matrix used to output the predicted variance sequence in the thermal hysteresis compensated decoder; This represents the variance output layer bias term.

[0065] It should also be noted that, before implementing water temperature stabilization control, the extended temporal convolutional encoder, sparse temporal self-attention processing, and thermal hysteresis compensation decoder are jointly trained.

[0066] Read the future actual outlet water temperature corresponding to the prediction step size from the historical operation data, and calculate the negative log-likelihood loss based on the predicted mean sequence, predicted variance sequence and future actual outlet water temperature of the future multi-step outlet water temperature.

[0067] Specifically, the future true outlet water temperature corresponding to the prediction step size is read from the historical operating data of the plate heat exchanger, and the future true outlet water temperature is compared step by step with the prediction mean sequence of the future multi-step outlet water temperature output by the thermal hysteresis compensation decoder. At the same time, the prediction variance sequence is read as the basis for uncertainty calibration to obtain the negative log-likelihood loss.

[0068] Based on the temperature change in adjacent prediction steps and the maximum allowable single-step change in outlet temperature in the future multi-step outlet water temperature prediction mean sequence, the temperature change rate constraint term is calculated, and the negative log-likelihood loss and the temperature change rate constraint term are weighted and combined to obtain the water temperature prediction training loss.

[0069] Specifically, the mean value of the predicted outlet water temperature in the next multiple predicted steps is read from the predicted mean value sequence of the outlet water temperature in the next multiple steps, and the temperature change between the next predicted steps is calculated; the temperature change is compared with the maximum allowable single-step change constraint of the outlet temperature to obtain the temperature change rate constraint term; the negative log-likelihood loss and the temperature change rate constraint term are weighted and combined to obtain the water temperature prediction training loss.

[0070] The sparse threshold regularization term is calculated based on the learnable sparse threshold and the target sparsity. The water temperature prediction training loss and the sparse threshold regularization term are combined to obtain the total training loss. The parameters of the dilated temporal convolutional encoder, the sparse temporal self-attention processing parameters, and the thermal hysteresis compensation decoder parameters are updated based on the total training loss.

[0071] Specifically, the learnable sparse thresholds corresponding to each attention head in the sparse temporal self-attention processing are read, and the actual sparsity rate corresponding to each attention head is calculated; the actual sparsity rate is compared with the target sparsity rate to obtain the sparsity threshold regularization term; the expression for calculating the sparsity threshold regularization term is: ; in, This represents the sparse threshold regularization term; This indicates an indicator function, which takes the value 1 when the value is 1, and when the value is 1 .... When ≤0, the value is 0, which is used to count the number of valid attention connections retained in the attention head; This represents the preset target sparsity.

[0072] Furthermore, the water temperature prediction training loss is weighted and combined with the sparse threshold regularization term to obtain the total training loss, which is expressed as follows: ; in, Indicates the total training loss; This indicates training loss in water temperature prediction; This represents the sparse threshold regularization term; This represents the weight coefficient corresponding to the sparse threshold regularization term, which is used to adjust the proportion of sparse threshold constraints in the total training loss.

[0073] Backpropagation is performed based on the total training loss to update the parameters of the dilated temporal convolutional encoder, the sparse temporal self-attention processing parameters, and the thermal hysteresis compensated decoder parameters.

[0074] S5. Based on the future multi-step outlet water temperature prediction mean sequence, prediction variance sequence and real-time feedback deviation, calculate the feedforward compensation amount and feedback correction amount, and perform valve opening limit and change rate constraint processing to obtain the regulating valve opening command.

[0075] The predicted temperature deviation sequence is calculated based on the mean sequence of future multi-step outlet water temperature predictions and the future outlet temperature setpoint. The predicted temperature deviation sequence is then weighted and summed according to the decay weight corresponding to the prediction step size to obtain the weighted prediction deviation.

[0076] Specifically, the system reads the sequence of predicted average outlet water temperatures over multiple future steps and the future outlet temperature setpoint, and calculates the deviation between the future outlet temperature setpoint and the corresponding predicted average outlet water temperature for each prediction step, thus obtaining the predicted temperature deviation sequence; the system then calculates the... The expression for the predicted temperature deviation in each prediction step is: , ; in, Indicates the current control cycle Next The predicted temperature deviation corresponding to each prediction step; Indicates the first The outlet water temperature setpoint corresponding to the predicted future time; The model predicts the first... The predicted average of the secondary outlet water temperature at each future predicted time.

[0077] The predicted temperature deviation sequence is weighted and summed according to the decay weight corresponding to the prediction step size to obtain the weighted prediction deviation; the expression for calculating the weighted prediction deviation is: ; ; in, Indicates the current control cycle Weighted prediction bias; Indicates the first The decay weights corresponding to each prediction step; decay coefficients .

[0078] The feedforward gain is dynamically estimated based on the temperature difference of the primary heat medium to obtain the adaptive feedforward gain, and the feedforward compensation is calculated based on the weighted prediction bias, rated flow rate and adaptive feedforward gain.

[0079] Specifically, the inlet and outlet temperature difference of the primary side, the primary side temperature difference under rated operating conditions, and the nominal feedforward gain are read. The feedforward gain is dynamically estimated based on the inlet and outlet temperature difference of the primary side to obtain the adaptive feedforward gain. The expression for calculating the adaptive feedforward gain is as follows: ; in, Indicates the current control cycle Adaptive feedforward gain under these conditions; This indicates the nominal feedforward gain, which can be preset based on rated operating conditions or historical operating data. Indicates the current control cycle The temperature difference between the inlet and outlet of the primary side can be obtained by subtracting the inlet temperature of the primary side heat medium from the outlet temperature of the primary side heat medium. This indicates the temperature difference between the primary side inlet and outlet under rated operating conditions. .

[0080] The feedforward compensation is calculated based on the weighted prediction bias, secondary side cooling water volumetric flow rate, rated flow rate, and adaptive feedforward gain. The expression for calculating the feedforward compensation is: ; in, This represents the adaptive feedforward compensation amount under the current control period t; This represents the secondary side cooling water volumetric flow rate under the current control period t; This indicates the volumetric flow rate of the secondary cooling water under rated operating conditions.

[0081] The real-time feedback deviation is calculated based on the current outlet water temperature setpoint and the secondary outlet water temperature. The comprehensive uncertainty measure is calculated based on the predicted variance sequence. The proportional gain and integral gain are adjusted according to the comprehensive uncertainty measure to obtain the proportional gain and integral gain of the current control cycle. The feedback correction amount is calculated based on the real-time feedback deviation, the proportional gain of the current control cycle, and the integral gain of the current control cycle.

[0082] Specifically, the current outlet water temperature setpoint and the secondary outlet water temperature are read to obtain the real-time feedback deviation; the expression for calculating the real-time feedback deviation is: ; in, Indicates the current control cycle Real-time feedback deviation; Indicates the current control cycle The outlet water temperature setpoint is below; Indicates the current control cycle The actual secondary side outlet water temperature collected below.

[0083] Before adjusting the proportional gain and integral gain based on the predicted variance sequence, a comprehensive uncertainty measure is calculated based on the predicted variance sequence of the outlet water temperature over multiple future steps. The proportional gain and integral gain are then adjusted according to the comprehensive uncertainty measure to obtain the proportional gain and integral gain for the current control cycle. The expression for calculating the proportional gain for the current control cycle is as follows: ; in, Indicates the current control cycle Adaptive proportional gain to uncertainty; This indicates the nominal proportional gain, which can be preset based on rated operating conditions or historical commissioning data. (·) denotes an exponential function with the natural constant e as its base; This represents the uncertainty suppression coefficient, used to adjust the attenuation of prediction uncertainty relative to the proportional gain; This represents the overall uncertainty measure under the current control period t, which can be calculated from the variance sequence of future multi-step outlet water temperature predictions.

[0084] The expression for calculating the integral gain of the current control cycle is: ; in, This represents the adaptive integral gain of the uncertainty under the current control period t; This indicates the nominal integral gain, which can be pre-set based on rated operating conditions, historical operating data, or on-site commissioning results.

[0085] The feedback correction is calculated based on the real-time feedback deviation, the proportional gain of the current control cycle, and the integral gain of the current control cycle. The feedforward compensation and feedback correction are superimposed to obtain the initial valve opening control quantity. The initial valve opening control quantity is then subjected to upper and lower limit constraints to obtain the amplitude-limited valve opening control quantity.

[0086] Specifically, the valve opening degree of the previous cycle is read, and the valve opening degree of the previous cycle, the feedforward compensation amount, and the feedback correction amount are superimposed to obtain the initial valve opening control amount; the expression for calculating the initial valve opening control amount is: ; in, This represents the initial valve opening control value under the current control cycle t; This indicates the valve opening degree in the previous control cycle; This represents the adaptive feedforward compensation amount under the current control period t, which is calculated from the weighted prediction deviation, adaptive feedforward gain, and secondary side flow correction term. This represents the feedback correction amount under the current control period t.

[0087] First, constrain the rate of change of the initial valve opening control quantity relative to the valve opening of the previous control cycle. Then, constrain the valve opening control quantity after the rate of change constraint by applying upper and lower limits to the valve opening.

[0088] The initial valve opening control quantity is subject to rate-of-change limitation and valve opening amplitude limit processing; the expression for calculating the valve opening change is: ; in, This represents the change in valve opening after the rate of change constraint during the current control period t. (·) represents a limiting function, used to restrict the input value to a preset upper and lower limit range; This represents the initial valve opening control value under the current control cycle t; This indicates the change in the valve opening control amount at the beginning of the current control cycle relative to the valve opening in the previous control cycle. This indicates the maximum allowable change in valve opening between two adjacent control cycles.

[0089] The valve opening control quantity of the limiting valve is constrained by the rate of change of adjacent control cycles to obtain the constrained valve opening control quantity. When the constrained valve opening control quantity touches the upper or lower limit of the valve opening, the integrator accumulation stops, and the regulating valve opening command is obtained based on the constrained valve opening control quantity.

[0090] Specifically, the valve opening degree of the previous cycle is superimposed with the valve opening change, and upper and lower limits of the valve opening are constrained to obtain the valve opening degree of the current control cycle; the expression for calculating the valve opening degree of the current control cycle is: ; in, This indicates the final valve opening command obtained under the current control cycle t; (·) represents a limiting function, used to restrict the input value to a preset upper and lower limit range; This represents the valve opening control quantity after the rate of change constraint during the current control cycle.

[0091] When the valve opening in the current control cycle is inconsistent with the initial valve opening control value, it is determined that the valve opening in the current control cycle is affected by the rate of change constraint or the upper and lower limits of the valve opening, and the integrator accumulation is stopped or rolled back.

[0092] S6. Control the operation of the regulating valve of the plate heat exchanger based on the regulating valve opening command, and periodically monitor the predicted residual. Make online fine adjustments to the thermal hysteresis compensation decoder based on the predicted residual to obtain the outlet water temperature control result.

[0093] The control valve opening command is sent to the control valve actuator of the plate heat exchanger. The control valve actuator adjusts the control valve opening according to the control valve opening command and collects the secondary side outlet water temperature after the control valve opening is adjusted.

[0094] Specifically, within the current control cycle, the regulating valve opening command is sent to the regulating valve actuator of the plate heat exchanger, and the regulating valve actuator adjusts the target side flow according to the valve opening corresponding to the regulating valve opening command. After the regulating valve opening is adjusted, the secondary side outlet water temperature after the regulating valve opening is adjusted is collected by the secondary side outlet water temperature sensor, and the secondary side outlet water temperature after the regulating valve opening is adjusted is used as the data source for subsequent prediction residual calculation.

[0095] The predicted residuals are calculated based on the predicted average sequence of multi-step outlet water temperature and the secondary outlet water temperature after the adjustment of the regulating valve opening. The root mean square of the predicted residuals within the sliding window is calculated according to the detection cycle.

[0096] Specifically, the system reads the sequence of predicted average outlet water temperatures for future multiple steps output by the thermal hysteresis compensation decoder in the historical control cycle, and determines the prediction step number corresponding to the current actual sampling time based on the current actual sampling time; it reads the currently collected secondary side outlet water temperature, calculates the difference between the currently collected secondary side outlet water temperature and the predicted average outlet water temperature at the same prediction time generated in the previous corresponding control cycle, and obtains the prediction residual; it continuously reads the prediction residual within the sliding window according to the detection cycle, and performs root mean square calculation on the prediction residual within the sliding window to obtain the root mean square of the prediction residual.

[0097] Before performing online drift detection, the allowable control error of the secondary side outlet water temperature is determined based on the offline validation set or historical samples from the stable operation phase, and the root mean square threshold of the predicted residual is set according to the allowable control error.

[0098] The root mean square of the predicted residuals is compared with a set threshold. When the root mean square of the predicted residuals exceeds the set threshold, the parameters of the dilated temporal convolutional encoder and the sparse temporal self-attention processing are frozen. The parameters of the thermal hysteresis compensation decoder are fine-tuned online based on the most recently run samples to obtain the updated thermal hysteresis compensation decoder parameters.

[0099] Specifically, the root mean square of the predicted residuals within the sliding window is compared with a set threshold. When the root mean square of the predicted residuals does not exceed the set threshold, the parameters of the extended temporal convolutional encoder, the sparse temporal self-attention processing, and the thermal hysteresis compensation decoder remain unchanged. When the root mean square of the predicted residuals exceeds the set threshold, it is determined that the plate heat exchanger has long-term operating condition drift, and the parameters of the extended temporal convolutional encoder and the sparse temporal self-attention processing are frozen. The most recently run samples are read, and the parameters of the thermal hysteresis compensation decoder are fine-tuned online based on the most recently run samples to obtain the updated thermal hysteresis compensation decoder parameters.

[0100] Based on the updated thermal hysteresis compensation decoder parameters, the system continues to generate the predicted mean sequence, predicted variance sequence, and regulating valve opening command for future multi-step outlet water temperature. Based on the regulating valve opening command, the system controls the operation of the regulating valve of the plate heat exchanger to obtain the outlet water temperature control result.

[0101] Specifically, the updated thermal hysteresis compensation decoder parameters are written into the thermal hysteresis compensation decoder, and multi-step decoding is performed based on the context-enhanced feature matrix and thermal hysteresis compensation features in subsequent control cycles to generate new future multi-step outlet water temperature prediction mean sequence and prediction variance sequence. Based on the new future multi-step outlet water temperature prediction mean sequence, prediction variance sequence and real-time feedback deviation, the feedforward compensation amount and feedback correction amount are recalculated, and valve opening limit and rate of change constraint processing are performed to obtain a new regulating valve opening command. Based on the new regulating valve opening command, the regulating valve of the plate heat exchanger is controlled to obtain the outlet water temperature control result.

[0102] In summary, this invention achieves multi-step decoding by using context-enhanced feature matrices and real-time flow estimation to estimate thermal hysteresis delay and constructing thermal hysteresis compensation features. This enables the corresponding fusion of historical key moment information and heat exchange hysteresis process, and provides a predictive basis with timing advance for feedforward compensation and feedback correction, thereby improving the accuracy of regulating valve opening command generation and the stability of outlet water temperature.

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for stabilizing water temperature in a plate heat exchanger based on time-series prediction, characterized in that, include: The inlet and outlet temperatures, flow rates, regulating valve openings, and outlet temperature setpoints of the plate heat exchanger are collected. A sliding window is constructed, differential features are expanded, and adaptive normalization is performed to obtain a standardized historical feature matrix. Dilated causal convolutional coding is performed based on the standardized historical feature matrix to extract multi-scale temporal dynamic information and obtain the temporal feature matrix; Sparse temporal self-attention processing is applied to the temporal feature matrix to identify key historical moments and aggregate temporal correlation information, resulting in a context-enhanced feature matrix. Based on the context-enhanced feature matrix and real-time flow estimation thermal hysteresis delay, thermal hysteresis compensation features are constructed and multi-step decoding is performed to obtain the mean sequence and variance sequence of future multi-step outlet water temperature prediction. Based on the future multi-step outlet water temperature prediction mean sequence, prediction variance sequence and real-time feedback deviation, the feedforward compensation and feedback correction are calculated, and valve opening limit and rate of change constraint processing are performed to obtain the regulating valve opening command. The control valve of the plate heat exchanger is operated based on the control valve opening command, and the prediction residual is periodically monitored. The thermal hysteresis compensation decoder is fine-tuned online based on the prediction residual to obtain the outlet water temperature control result.

2. The plate heat exchanger water temperature stabilization control method based on time-series prediction as described in claim 1, characterized in that, The system collects the inlet and outlet temperatures, flow rates, regulating valve openings, and outlet temperature setpoints on both sides of the plate heat exchanger. It then performs sliding window construction, differential feature expansion, and adaptive normalization to obtain a standardized historical feature matrix. The specific steps are as follows: Collect the secondary side outlet water temperature, secondary side inlet water temperature, primary side heat medium inlet temperature, primary side heat medium outlet temperature, primary side heat medium volumetric flow rate, secondary side cold water volumetric flow rate, previous cycle regulating valve opening, and current outlet water temperature setpoint. The inlet and outlet temperature difference of the primary side heat medium is calculated based on the inlet and outlet temperatures of the primary side heat medium, and the estimated value of the instantaneous heat exchange of the primary side is calculated based on the primary side heat medium flow rate and the inlet and outlet temperature difference of the primary side. A historical feature matrix is ​​constructed based on the secondary side outlet water temperature, secondary side inlet water temperature, primary side heat medium inlet temperature, primary side heat medium outlet temperature, primary side heat medium cumulative flow, secondary side cold water volumetric flow, previous cycle regulating valve opening, current outlet water temperature setpoint, primary side inlet and outlet temperature difference, and estimated primary side instantaneous heat exchange. First-order difference features are added to the historical feature matrix to obtain the augmented historical matrix. The augmented historical matrix is ​​then normalized online using the exponential moving average statistic to obtain the standardized historical feature matrix.

3. The plate heat exchanger water temperature stabilization control method based on time-series prediction as described in claim 2, characterized in that, The process of obtaining the temporal feature matrix based on the standardized historical feature matrix involves multi-layer dilated causal convolution and residual connection processing using an dilated temporal convolutional encoder. The specific steps are as follows: The standardized historical feature matrix is ​​subjected to input projection processing to obtain the projected feature matrix. Then, according to the hierarchical order of the dilated temporal convolutional encoder, the projected feature matrix is ​​subjected to multi-layer dilated causal convolution processing based on the kernel width, dilation factor, weight matrix and bias to obtain the convolutional features of each layer. The receptive field is expanded layer by layer according to the expansion factor corresponding to each layer of expanded causal convolution, and the multi-scale historical information in the standardized historical feature matrix is ​​covered to obtain multi-scale convolution features. The multi-scale convolutional features and the corresponding layer input features are processed by residual connection to obtain the residual temporal features of each layer. The residual temporal features of the last layer are used as the temporal feature matrix.

4. The plate heat exchanger water temperature stabilization control method based on time-series prediction as described in claim 3, characterized in that, The process of performing sparse temporal self-attention processing on the temporal feature matrix to identify key historical moments and aggregate temporal correlation information to obtain a context-enhanced feature matrix is ​​as follows: Multi-head linear projection is performed on the temporal feature matrix to obtain the query matrix, key matrix and value matrix. The attention score matrix is ​​calculated based on the query matrix and key matrix. The attention score matrix is ​​then sparsified according to the learnable sparse threshold corresponding to each attention head to obtain sparse attention weights. The sparse attention weights are weighted and fused with the value matrix to obtain temporal correlation features. The temporal correlation features are then processed by multi-head concatenation, residual connection and layer normalization to obtain the context enhancement feature matrix.

5. The plate heat exchanger water temperature stabilization control method based on time-series prediction as described in claim 4, characterized in that, The process involves constructing thermal hysteresis compensation features based on the context-enhanced feature matrix and real-time flow estimation thermal hysteresis delay, and performing multi-step decoding to obtain the mean sequence and variance sequence of future multi-step outlet water temperature predictions. The specific steps are as follows: The primary side heat medium volume flow rate and the secondary side cold water volume flow rate are collected, and the thermal hysteresis delay is estimated by combining the effective flow channel volume of the heat exchanger and the empirical correction coefficient. The thermal hysteresis delay is discretized into time steps, and the context representation of the corresponding historical moment is read from the context enhancement feature matrix to obtain the thermal hysteresis compensation feature. The context representation of the latest time step in the context-enhanced feature matrix is ​​concatenated with the thermal hysteresis compensation feature to obtain the decoding input feature. The decoding input feature is then input into the thermal hysteresis compensation decoder for multi-step decoding to obtain the mean sequence and variance sequence of the predicted outlet water temperature for the next multiple steps.

6. The plate heat exchanger water temperature stabilization control method based on time-series prediction as described in claim 5, characterized in that, Before implementing water temperature stabilization control, the process also includes joint training of the extended temporal convolutional encoder, sparse temporal self-attention processing, and thermal hysteresis compensation decoder. The specific steps are as follows: Read the future actual outlet water temperature corresponding to the prediction step size from the historical operation data, and calculate the negative log-likelihood loss based on the prediction mean sequence, prediction variance sequence and future actual outlet water temperature of the future multi-step outlet water temperature. Based on the temperature change in adjacent prediction steps and the maximum allowable single-step change in outlet temperature in the future multi-step outlet water temperature prediction mean sequence, the temperature change rate constraint term is calculated, and the negative log-likelihood loss and the temperature change rate constraint term are weighted and combined to obtain the water temperature prediction training loss. The sparse threshold regularization term is calculated based on the learnable sparse threshold and the target sparsity. The water temperature prediction training loss and the sparse threshold regularization term are combined to obtain the total training loss. The parameters of the dilated temporal convolutional encoder, the sparse temporal self-attention processing parameters, and the thermal hysteresis compensation decoder parameters are updated based on the total training loss.

7. The plate heat exchanger water temperature stabilization control method based on time-series prediction as described in claim 6, characterized in that, The calculation of feedforward compensation and feedback correction based on the future multi-step outlet water temperature prediction mean sequence, prediction variance sequence, and real-time feedback deviation is as follows: The predicted temperature deviation sequence is calculated based on the mean sequence of future multi-step outlet water temperature prediction and the future outlet temperature setpoint. The predicted temperature deviation sequence is then weighted and summed according to the decay weight corresponding to the prediction step size to obtain the weighted prediction deviation. The feedforward gain is dynamically estimated based on the temperature difference of the primary heat medium to obtain the adaptive feedforward gain, and the feedforward compensation amount is calculated based on the weighted prediction deviation, rated flow rate and adaptive feedforward gain. The real-time feedback deviation is calculated based on the current outlet water temperature setpoint and the secondary outlet water temperature. The comprehensive uncertainty measure is calculated based on the predicted variance sequence. The proportional gain and integral gain are adjusted according to the comprehensive uncertainty measure to obtain the proportional gain and integral gain of the current control cycle. The feedback correction amount is calculated based on the real-time feedback deviation, the proportional gain of the current control cycle, and the integral gain of the current control cycle.

8. The plate heat exchanger water temperature stabilization control method based on time-series prediction as described in claim 7, characterized in that, The valve opening limit and rate of change constraint processing is performed to obtain the regulating valve opening command. The specific steps are as follows: The feedforward compensation and feedback correction are superimposed to obtain the initial valve opening control quantity. The initial valve opening control quantity is then subjected to upper and lower limit constraints on the valve opening to obtain the amplitude-limited valve opening control quantity. The valve opening control quantity of the limiting valve is constrained by the rate of change of adjacent control cycles to obtain the constrained valve opening control quantity. When the constrained valve opening control quantity touches the upper or lower limit of the valve opening, the integrator accumulation stops, and the regulating valve opening command is obtained based on the constrained valve opening control quantity.

9. The plate heat exchanger water temperature stabilization control method based on time-series prediction as described in claim 8, characterized in that, The specific steps for controlling the operation of the regulating valve of the plate heat exchanger based on the regulating valve opening command and periodically monitoring and predicting the residual are as follows: The control valve opening command is sent to the control valve actuator of the plate heat exchanger. The control valve actuator adjusts the control valve opening according to the control valve opening command and collects the secondary side outlet water temperature after the control valve opening is adjusted. The predicted residuals are calculated based on the predicted average sequence of multi-step outlet water temperature and the secondary outlet water temperature after the adjustment of the regulating valve opening. The root mean square of the predicted residuals within the sliding window is calculated according to the detection cycle.

10. The plate heat exchanger water temperature stabilization control method based on time-series prediction as described in claim 9, characterized in that, The process of fine-tuning the thermal hysteresis compensation decoder online based on the predicted residual to obtain the outlet water temperature control result is as follows: The root mean square of the predicted residuals is compared with a set threshold. When the root mean square of the predicted residuals exceeds the set threshold, the parameters of the dilated temporal convolutional encoder and the sparse temporal self-attention processing are frozen. The parameters of the thermal hysteresis compensation decoder are fine-tuned online based on the most recent running samples to obtain the updated thermal hysteresis compensation decoder parameters. Based on the updated thermal hysteresis compensation decoder parameters, the system continues to generate the predicted mean sequence, predicted variance sequence, and regulating valve opening command for future multi-step outlet water temperature. Based on the regulating valve opening command, the system controls the operation of the regulating valve of the plate heat exchanger to obtain the outlet water temperature control result.

Images