Heat exchange station control method and device based on neural network decoupling and medium

Abstract

The invention discloses a heat exchange station control method and device based on neural network decoupling and a medium, and relates to the field of neural networks, and the method comprises the steps: carrying out the feedforward decoupling through a trained time-delay recurrent neural network, and outputting a decoupling control instruction; performing prediction through the trained prediction model, and outputting a prediction parameter output value at a future moment; outputting a frequency conversion increment through a rolling optimizer according to the predicted parameter output value; determining an output difference value between the second actual parameter output value and the predicted parameter output value; and adjusting model parameters and / or variable frequency increment of the prediction model. Feedforward decoupling is carried out by using a time-delay recurrent neural network, and dynamic characteristics of the system are learned through a data driving mode without depending on an accurate mathematical model. Compared with the traditional feedforward decoupling and frequency conversion decoupling strategies, the coupling relationship between the quality regulation channel and the quantity regulation channel can be processed more thoroughly, and the problem that multi-variable dynamic response is difficult to coordinate by single pressure difference control is solved.

Description

Technical Field

[0001] This application relates to the field of neural networks, specifically to a control method, equipment, and medium for heat exchange stations based on neural network decoupling. Background Technology

[0002] Heat exchange stations are an important component of heating networks and serve as the primary control points for heating systems. Due to the time-varying, time-delay, and coupling characteristics of heat exchange station systems, conventional control strategies often fail to meet their required control accuracy.

[0003] In traditional solutions, an improved ant colony algorithm can be used to optimize the fuzzy PID parameters of the heat exchange station system, resulting in a shorter system adjustment time and higher steady-state accuracy. However, this approach does not address the time-varying and strongly coupled characteristics of the heat exchange station system. The ant colony algorithm has limited adaptability to dynamic parameter changes and struggles to handle the time delay problem of the quality control channel in real time.

[0004] It can also be combined with the heat exchange station process to decouple temperature and flow rate by feedforward, and introduce adaptive control to regulate the temperature difference in the quality regulation channel to compensate for the impact of the lag characteristics of the heat station. The system's anti-disturbance and robustness are improved to a certain extent. However, its decoupling is not thorough and relies on an accurate mathematical model. The actual operating conditions of the heat exchange station are complex and variable, which can easily lead to model mismatch. Moreover, the adaptive control has limited ability to compensate for time-delay links and it is difficult to dynamically correct the multivariate coupling relationship.

[0005] A variable frequency decoupling strategy for heat exchange station pumps can also be proposed, which decouples the system hydraulically based on pressure difference and takes into account time factors to make heating more intelligent; however, it only decouples from the perspective of hydraulic balance and ignores thermodynamic coupling: there is strong coupling between quality regulation (temperature) and quantity regulation (flow rate), and it is difficult for single pressure difference control to coordinate the dynamic response of the two, while not solving the control lag problem caused by time delay. Summary of the Invention

[0006] To address the aforementioned issues, this application proposes a heat exchange station control method based on neural network decoupling, comprising:

[0007] The preset parameter settings and the first actual parameter output values ​​at historical moments are used as inputs. The system performs feedforward decoupling through a trained time-delay recurrent neural network and outputs decoupling control commands.

[0008] The decoupling control command is taken as input, and the prediction model is used to make predictions, outputting the predicted parameter values ​​for future times.

[0009] The rolling optimizer outputs the frequency conversion increment based on the predicted parameter output value, and controls the heat exchange station system based on the frequency conversion increment.

[0010] Obtain the second actual parameter output value at the future time after the heat exchange station system is controlled, and determine the output difference between the second actual parameter output value and the predicted parameter output value;

[0011] Based on the output difference, the model parameters of the prediction model and / or the frequency conversion increment are adjusted.

[0012] In one example, preset parameter settings and the first actual parameter output values ​​at historical moments are used as inputs. A trained time-delay recurrent neural network is used for feedforward decoupling, outputting decoupling control commands, specifically including:

[0013] The preset parameter settings include the booster pump frequency conversion setting and the circulation pump frequency setting, and the first actual parameter output values ​​at multiple historical moments include the first supply and return water temperature difference measured value sequence and the first secondary network flow measured value sequence.

[0014] Feedforward decoupling is performed using a trained time-delay recurrent neural network, and decoupling control commands are generated through internal weight calculation; the decoupling control commands include temperature difference channel decoupling commands and flow rate channel decoupling commands.

[0015] The decoupling control command is output to the prediction model.

[0016] In one example, the training process of the time-delay recurrent neural network includes:

[0017] Obtain multiple preset actual parameter settings, and for each actual parameter setting, obtain the first actual parameter output value when the heat exchange station system control reaches steady state based on that actual parameter setting value;

[0018] For each first actual parameter output value, the ideal parameter setting value under the undecoupled state is obtained by back-reasoning through an independent single-loop channel expression;

[0019] The ideal parameter settings are used as input, and the corresponding actual parameter settings are used as output to train a time-delay recurrent neural network.

[0020] In one example, the decoupling control command is taken as input, and a trained prediction model is used to predict the parameters for future time moments, specifically including:

[0021] For the temperature difference channel decoupling instruction and flow channel decoupling instruction included in the decoupling control instruction, determine the corresponding booster pump frequency conversion instruction and circulating pump frequency instruction respectively;

[0022] The booster pump frequency conversion command is taken as input, and the trained prediction model is used to predict and output a sequence of predicted values ​​for the supply and return water temperature difference at future times.

[0023] The prediction model is a controlled autoregressive moving average integral model compensated by a neural network decoupler.

[0024] In one example, the rolling optimizer outputs frequency conversion increments based on the predicted parameter output values, and the heat exchange station system is controlled based on these frequency conversion increments, specifically including:

[0025] A target function is constructed, which includes a prediction tracking term and a control penalty term; the prediction tracking term is determined based on the predicted value sequence of the supply and return water temperature difference and the variable frequency setting value of the booster pump; the control penalty term is determined based on the variable frequency increment of the booster pump.

[0026] The rolling optimizer performs a minimization calculation based on the objective function and outputs a frequency conversion incremental sequence for the booster pump.

[0027] Based on the frequency increment of the booster pump in the frequency increment sequence at the current moment, an actual control command is generated, and the heat exchange station system is controlled through the actual control command.

[0028] In one example, the frequency conversion command of the booster pump is taken as input, and a trained prediction model is used to predict and output a sequence of predicted supply and return water temperature differences for future times, specifically including:

[0029] Based on the trained prediction model, an implicit prediction equation is established; the implicit prediction equation includes a booster pump frequency conversion increment sequence and a parameter vector; the parameter vector includes a dynamic response matrix and a prediction vector; the dynamic response matrix is ​​used to establish the dynamic relationship between the booster pump frequency conversion increment and the predicted value of the supply and return water temperature difference; the prediction vector is used to represent the control-independent component in the predicted value of the supply and return water temperature difference.

[0030] Based on the implicit prediction equation, a sequence of predicted supply and return water temperature differences for future times is output.

[0031] In one example, the model parameters of the prediction model are adjusted based on the output difference, specifically including:

[0032] The dynamic response matrix is ​​solved, the intermediate covariance matrix is ​​determined according to the preset forgetting factor, and the adaptive gain matrix is ​​determined according to the intermediate covariance matrix.

[0033] The parameter vector of the prediction model is updated based on the output difference and the adaptive gain matrix.

[0034] And update the covariance matrix based on the adaptive gain matrix;

[0035] The prediction vector is solved, and the prediction vector is updated based on the output difference and the predicted value of the supply and return water temperature difference at historical time.

[0036] In one example, the frequency increment of the frequency conversion increment is adjusted based on the output difference, specifically including:

[0037] Based on the dynamic response matrix and the prediction vector, the corresponding frequency conversion increment sequence is determined;

[0038] Update the corresponding frequency increment based on the frequency increment sequence.

[0039] On the other hand, this application also proposes a heat exchange station control device based on neural network decoupling, comprising:

[0040] At least one processor; and,

[0041] A memory communicatively connected to the at least one processor; wherein,

[0042] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the neural network-based decoupling heat exchange station control method as described in any of the above examples.

[0043] On the other hand, this application also proposes a non-volatile computer storage medium storing computer-executable instructions configured to implement the heat exchange station control method based on neural network decoupling as described in any of the above examples.

[0044] The heat exchange station control method based on neural network decoupling proposed in this application can bring the following beneficial effects:

[0045] 1. Using a time-delay recurrent neural network for feedforward decoupling, the system's dynamic characteristics are learned through a data-driven approach, without relying on a precise mathematical model. Compared to traditional feedforward decoupling and frequency conversion decoupling strategies, this approach more thoroughly addresses the coupling relationship between the quality control channel and the quantity control channel, solving the problem of single differential pressure control's inability to coordinate the dynamic response of multiple variables.

[0046] 2. Recurrent neural network structures can memorize historical states and, combined with the rolling optimization mechanism of predictive models, achieve dynamic compensation for time-delay components. Compared to the static parameter optimization of ant colony algorithms and adaptive control, this approach can more effectively alleviate control lag problems caused by time delays.

[0047] 3. By comparing the difference between the predicted output and the actual output in real time, the prediction model parameters and frequency conversion increments are dynamically corrected to form a closed-loop optimization. This mechanism enables the system to autonomously adapt to changes in operating conditions and model mismatch, overcoming the shortcomings of traditional methods that rely on fixed models and have insufficient disturbance rejection.

[0048] 4. The combination of the predictive model and the rolling optimizer simultaneously optimizes temperature and flow control targets on the basis of decoupling, avoiding the control limitations caused by a single optimization index and improving the overall control accuracy of the system. Attached Figure Description

[0049] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0050] Figure 1 This is a flowchart illustrating the heat exchange station control method based on neural network decoupling in the embodiments of this application;

[0051] Figure 2 This is a schematic diagram illustrating the interconnected coupling of a battery swapping station system under one scenario in an embodiment of this application.

[0052] Figure 3 This is a schematic diagram of the structure of a recurrent neural network in one embodiment of this application;

[0053] Figure 4 This is a schematic diagram of the decoupling structure of a time-delay recurrent neural network in one scenario of an embodiment of this application;

[0054] Figure 5 This is a schematic diagram of a feedforward decoupling structure under one scenario in an embodiment of this application;

[0055] Figure 6 This is a schematic diagram of the heat exchange station control device based on neural network decoupling in the embodiments of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0057] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0058] like Figure 1 As shown, this application provides a heat exchange station control method based on neural network decoupling, including:

[0059] S101: The preset parameter settings and the first actual parameter output values ​​at historical moments are used as inputs. The pre-trained time-delay recurrent neural network is used for feedforward decoupling, and the decoupling control command is output.

[0060] A heat exchange station system is a nonlinear, large-time-delay, and strongly coupled controlled object. The input response and control processes typically have long delays, and this time-delay characteristic easily causes fluctuations in various parameters. The heat exchange station system has many controlled objects that are heavily interdependent and coupled; adjusting one input variable can affect other controlled objects, increasing the control complexity. Figure 2 As shown, the input quantities can include the frequency converter of the primary network booster pump and the frequency of the secondary network circulation pump. The frequency converter of the primary network booster pump controls the temperature difference between the secondary network supply and return water through the quality adjustment channel, and the frequency of the secondary network circulation pump controls the flow rate of the secondary network circulation water through the quantity adjustment channel.

[0061] Among them, the quality regulation channel and the quantity regulation channel are two core control dimensions, independently regulating the temperature and flow rate of heat transfer, respectively. The quality regulation channel changes the quality of heat delivered to users by adjusting the temperature of the primary network heat source. The quantity regulation channel changes the quantity of heat delivered to users by adjusting the flow rate of the secondary network circulating water. Due to the existence of coupling channel 1 and coupling channel 2, there is a coupling phenomenon between the quality regulation channel and the quantity regulation channel.

[0062] The overall decoupled control implementation process consists of a time-delay recurrent neural network, a predictive model, a rolling optimizer, model decoupling, and feedback correction. Time delay and time-varying characteristics are the main reasons for the difficulty in controlling the quality regulation loop of the heat exchange station system. Therefore, an iterative generalized predictive control (IGPC) strategy is adopted for the quality regulation channel, while a proportional-integral-derivative (PID) controller is used for the quantity regulation channel.

[0063] During the feedforward decoupling process, preset parameter settings are determined, including the booster pump frequency conversion setting x1(t) (used to control the temperature difference in the quality regulation channel) and the circulation pump frequency setting x2(t) (used to control the flow rate in the quantity regulation channel). Multiple historical time-based first actual parameter output values ​​are also determined, including a sequence of measured first supply and return water temperature differences and a sequence of measured first secondary network flow rates. Since subsequent schemes also require determining the second actual parameter output values ​​for future timeframes, the historical first supply and return water temperature difference and the future second supply and return water temperature difference are both referred to as y1(t), and the historical first secondary network flow rate and the future second secondary network flow rate are both referred to as y2(t).

[0064] Feedforward decoupling is performed using a trained time-delay recurrent neural network. Decoupling control commands are generated through internal weight calculations; these commands include decoupling commands for the temperature difference channel and the flow rate channel. The decoupling control commands are then output to the prediction model.

[0065] like Figure 3 As shown, a recurrent neural network (RNN) model structure is provided. The topology of the RNN can effectively solve the nonlinearity and strong coupling problems existing in the control system. The heating station system exhibits a time delay; the state of the input values ​​at time point k and previous historical times both affect the output value at the current time point k. Based on this, the input at the current moment and the output values ​​of the previous k historical moments are combined as the input to the time-delay RNN. This allows past outputs to be considered, thereby realizing the dynamic characteristics of the heating station system.

[0066] The mathematical expression for solving time series problems in heat exchange stations using a recurrent neural network model is shown in Formula 1:

[0067]

[0068] Where h is the state value, σ is the activation function, U is the input weight at time t, and V is the state weight at time t-1; O i For the decoupling instructions of the i-th output channel (including the booster pump frequency conversion instruction and the circulating pump frequency instruction), f(h) t ) is the computation function for the recurrent neural network model.

[0069] Considering the characteristics of the heat exchange station, a time delay link (TDL) is added before network training. The structure of the time-delay recurrent neural network decoupler is as follows: Figure 4 As shown. The output of the decoupler at time n can be described by formula two:

[0070] u ij (k)=N(x ij (k),xij (k-1),…x ij (kn),u ij (k-1),u ij (k-2),…u ij (kn))

[0071] Formula 2;

[0072] Among them, u ij (k) represents the decoupling control command at time k, N is the neural network mapping function, and x ij (k) represents the parameter setting value at time k.

[0073] By employing a feedforward compensation approach, the coupling between individual loop channels (including quality loop channels and quantity loop channels) is reduced. For example, Figure 5 As shown, the entire system is constructed using time-delay recurrent neural network decouplers and the controlled process. Multiple time-delay recurrent neural network decouplers (NDDs) are included. The time-delay recurrent neural network decouplers NND11, NND21, NND12, and NND22 are made to approximate the ideal decouplers G11, G21, G12, and G22, respectively, thereby eliminating the coupling effect of the coupled system while preserving the main channel characteristics of the coupled system. This can be illustrated by Equation 3:

[0074]

[0075] For time-delay recurrent neural networks, the training process includes:

[0076] Multiple preset actual parameter settings are obtained, and for each actual parameter setting, the first actual parameter output value when the heat exchange station system reaches steady state is obtained based on that actual parameter setting value. That is, given an array x = (x1, x2) consisting of a set of booster pump frequency conversion setting value x1(t) and circulation pump frequency setting value x2(t), it is input into the coupled channel model to obtain an array y = (y1, y2) of the first actual parameter output value when the system reaches steady state under this setting value, including the measured value of the first supply and return water temperature difference and the measured value of the first secondary network flow.

[0077] For each first actual parameter output value, the ideal parameter setting value under undecoupled conditions is obtained by back-calculating using an independent single-loop channel expression. Substituting the output array into the mass loop channel expression and quantity loop channel expression under undecoupled conditions, the input array x ~ = (x3, x4) corresponding to array y under undecoupled conditions is obtained. Specifically, during the back-calculation, the measured value y1 of the first supply and return water temperature difference is input into the independent mass loop channel expression, outputting the ideal booster pump frequency value x3 under undecoupled conditions; the measured value y2 of the first and second secondary network flow rates is input into the independent quantity loop channel expression, outputting the ideal circulating pump frequency value x4 under undecoupled conditions.

[0078] The ideal parameter setting value x ~ = (x3, x4) is used as input, and the corresponding actual parameter setting value x = (x1, x2) is used as output. At the same time, the hidden nodes of NND11 and NND22 are set to 6, the hidden nodes of NND12 are set to 9, and the hidden nodes of NND21 are set to 6, so as to train the time-delay recurrent neural network.

[0079] S102: The decoupling control command is taken as input, and the prediction is performed by the trained prediction model to output the predicted parameter output value at the future time.

[0080] As mentioned above, at the channel level, decoupling control commands include temperature difference channel decoupling commands and flow rate channel decoupling commands. At the equipment level, the temperature difference is achieved through adjustments to the booster pump, and the flow rate is achieved through adjustments to the circulation pump. Therefore, at the equipment level, the temperature difference channel decoupling command and the flow rate channel decoupling command correspond to the booster pump frequency conversion command and the circulation pump frequency command, respectively. The content of the temperature difference channel decoupling command and the booster pump frequency conversion command, as well as the flow rate channel decoupling command and the circulation pump frequency command, is the same; only the emphasis of their descriptions differs.

[0081] For the frequency command of the circulating pump, traditional PID control can be used. Here, we mainly consider the processing of the frequency conversion command of the booster pump.

[0082] Specifically, the frequency conversion command of the booster pump is taken as input, and a trained prediction model is used to predict the future supply and return water temperature difference sequence. The prediction model is a controlled autoregressive and integrated moving-average model (CARIMA) compensated by a neural network decoupler, and its expression is shown in Equation 4:

[0083] A(z -1 )y(k)=B(z -1 Formula 4: u(kd) + ε(k) / Δ

[0084] Where Δ=1-z -1 Let u(kd) be the frequency conversion command of the booster pump at time d lag, y(k) be the actual output value of the supply and return water temperature difference at the current time (time k), and A(z) be the differential operator. -1 B(z) is the system dynamic characteristic polynomial, describing the autoregressive characteristics of the predicted supply and return water temperature difference y(k). -1 Let be the control response polynomial, describing the dynamic response relationship between the booster pump frequency conversion command u(k) and the predicted supply and return water temperature difference y(k), where d is the system time delay step number, and z is the control response polynomial. -1 It is the unit delay operator, where ε(k) represents the external random disturbance at the current time.

[0085] The prediction model has been described by Equation 4; however, its solution complexity is high, resulting in low real-time control performance. Therefore, based on the trained prediction model, an implicit prediction equation is established. By solving this implicit prediction equation, the predicted parameter output value, which is the predicted value of the supply and return water temperature difference, is obtained.

[0086] Specifically, conventional generalized predictive control algorithms use Diophantine equations to predict future output values. However, the recursive solution process of Diophantine equations is computationally cumbersome and has a slow response speed, which is not conducive to practical engineering implementation. Therefore, an implicit algorithm is adopted to directly identify the corresponding output by using the frequency conversion of the primary network pressurization pump valves and the output secondary network temperature difference data of the heat exchange station system.

[0087] Based on the optimal output prediction matrix of the conventional generalized predictive control algorithm, n parallel predictors are written as shown in Formula 5:

[0088]

[0089] Where y(k+j) is the predicted temperature difference between the supply and return water at time j, Δu(k+i) is the frequency conversion increment of the booster pump at time i, and g i Let f(k+i) be the dynamic response coefficient at time i in the future, f(k+i) be the free response component at time i in the future, ε(k+i) be the external random disturbance at time i in the future, and E be the dynamic response coefficient at time i in the future. j This is the noise gain coefficient.

[0090] At this point, as shown in Formula 5, its last line contains all elements of the dynamic response matrix G within the optimal control rate Δu. Therefore, calculating only the y(k+n) equation yields the dynamic response matrix G. The dynamic response matrix G is derived from g. i composition.

[0091] At this point, the implicit prediction equation includes the booster pump frequency conversion increment sequence and parameter vector; the parameter vector includes the dynamic response matrix and prediction vector; the dynamic response matrix is ​​used to establish the dynamic relationship between the booster pump frequency conversion increment and the predicted value of the supply and return water temperature difference; the prediction vector is used to represent the component in the predicted value of the supply and return water temperature difference that is not related to control.

[0092] The implicit prediction equation can be written as shown in Formula 6:

[0093] y(k+n)=X(k)θ(k)+E n Formula 6 for ε(k+n);

[0094] Where X(k) = [Δu(k), Δu(k+1), ..., Δu(k+n-1), 1], is the frequency conversion increment sequence of the booster pump, which includes the frequency conversion increment of the booster pump from time k (the current time) to the (k+n-1)th time in the future, and sets the last bit to 1, forming a 1*n matrix; θ(k) = [g n-1 ,g n-2 ,…,g0,f(k+n)] T It is a parameter vector, including the dynamic response coefficients at each time step, forming an n*1 matrix; E n Let ε(k+n) be the noise gain coefficient at time n in the future, and let ε(k+n) be the external random disturbance at time n in the future.

[0095] At this point, predictions can be made based on the implicit prediction equation, outputting a sequence of predicted supply and return water temperature differences for future times. In the actual prediction process, it is necessary to solve the implicit prediction equation, which means solving for the dynamic response matrix and prediction vector in the parameter vector. This will be discussed in more detail later when updating the parameter vector.

[0096] S103: The rolling optimizer outputs the frequency conversion increment based on the predicted parameter output value, and controls the heat exchange station system based on the frequency conversion increment.

[0097] After obtaining the frequency conversion increment (i.e., the booster pump frequency conversion increment) corresponding to the predicted parameter output value (i.e., the predicted value of the supply and return water temperature difference), the heat exchange station system can be controlled accordingly.

[0098] Specifically, an objective function is constructed, which includes a prediction and tracking term and a control penalty term. The prediction and tracking term is determined based on the predicted sequence of supply and return water temperature differences and the frequency conversion setpoint of the booster pump. The control penalty term is determined based on the frequency conversion increment of the booster pump. The objective function can be shown in Formula 7.

[0099]

[0100] in, For the prediction tracking term, calculate the sum of squares of the deviations between the predicted supply and return water temperature difference y(k+j) and the booster pump frequency conversion setpoint ω(k+j) over the next n time points from the current time. To control the penalty term, the weighted sum of squares of the frequency converter increment Δu(k+j-1) of the booster pump over the next m time steps from the current time is calculated; m and n represent the control length and prediction accuracy of the algorithm, respectively; to make the system control characteristics smoother, a reference trajectory (including multiple booster pump frequency converter setpoints) is introduced as a guiding transition process, λ k To control the weighting constant.

[0101] The rolling optimizer performs a minimization calculation based on the objective function and outputs a booster pump frequency conversion increment sequence. Based on the frequency conversion increment at the current moment in the booster pump frequency conversion increment sequence, actual control commands are generated, and the heat exchange station system is controlled through the actual control commands.

[0102] The rolling optimization solution is performed, taking the predicted value of the supply and return water temperature difference y(k+j) as input, minimizing the objective function J, solving for the optimal frequency conversion increment sequence for the next m steps, and then extracting the optimal frequency conversion increment Δu(k) at the current moment to calculate the actual control command of the booster pump frequency conversion.

[0103] S104: Obtain the second actual parameter output value at the future time after the heat exchange station is controlled, and determine the output difference between the second actual parameter output value and the predicted parameter output value.

[0104] After executing the obtained optimal frequency conversion increment Δu(k), the corresponding second actual parameter output value can be collected. Here, it mainly refers to the measured value of the second supply and return water temperature difference, so as to determine the output difference between it and the predicted value of the supply and return water temperature difference.

[0105] S105: Adjust the model parameters of the prediction model and / or the frequency conversion increment based on the output difference.

[0106] When updating the model parameters, it is necessary to solve for the dynamic response matrix and the prediction vector.

[0107] When solving the dynamic response matrix, based on the implicit prediction equation in Formula 6, the noise term is treated as having a mean of 0, thus yielding the predicted value of the supply and return water temperature difference at the current time (time k) n steps (the future time n), as shown in Formula 8:

[0108] y(k|kn)=X(kn)θ(k) Formula 8.

[0109] At this point, the output difference is determined, and the intermediate covariance matrix is ​​determined according to the preset forgetting factor. The adaptive gain matrix is ​​then determined according to the intermediate covariance matrix. The parameter vector of the model parameters of the prediction model is updated according to the output difference and the adaptive gain matrix. The covariance matrix is ​​then updated according to the adaptive gain matrix.

[0110] Specifically, the parameter vector θ(k) is estimated using the Forgetting Factor Recursive Least Squares (FFRLS) method, as shown in Equation 9:

[0111]

[0112] Where F is the output difference, representing the deviation between the measured value y(k) of the second supply and return water temperature difference at the current moment and the predicted value X(k-1)θ(k-1) of the supply and return water temperature difference at the previous moment, λ1 is the forgetting factor, M is the intermediate covariance matrix, K(k) is the adaptive gain matrix, θ(k) is the updated parameter vector, and P(k) is the covariance matrix.

[0113] Thus, updating the parameter vector θ(k) completes the update of the dynamic response matrix G. Since the parameter vector θ(k) contains the long-term free response component f(k+n), it is also necessary to calculate the short-term free response component f(k+1).

[0114] When solving for the prediction vector, the prediction vector is updated based on the output difference and the predicted supply and return water temperature difference at historical times, as shown in Formula 10:

[0115]

[0116] Where f(k+1) is the updated prediction vector, y(k+j|n) is the predicted value of the supply and return water temperature difference at time n, and e(k+1)=y(k+1)-y(k+1|k) represents the difference between the measured value of the second supply and return water temperature difference at time k+1 and the predicted value of the supply and return water temperature difference at time k+1, which is the output difference.

[0117] When updating the frequency converter increment, the corresponding frequency converter increment sequence is determined based on the dynamic response matrix and the prediction vector, as shown in Formula 11:

[0118] ΔU=(G T G+λ k I) -1 G T (Wf) Formula Eleven;

[0119] Where ΔU is the frequency conversion increment sequence, G is the dynamic response matrix, and λk To control the weighting coefficients, I is the identity matrix, W is the reference trajectory which includes the expected sequence of future booster pump frequency conversion setpoints, and f is the prediction vector.

[0120] Update the corresponding frequency increment based on the frequency increment sequence, letting (G) T G+λ k I) -1 G T The first line is g T Then it can be represented as shown in Formula Twelve:

[0121] u(k)=u(k-1)+g T (Wf) Formula Twelve;

[0122] Where u(k) is the booster pump frequency conversion command at the current time (time k), u(k-1) is the booster pump frequency conversion command at the previous time, and g T (Wf) is the update of the frequency conversion increment.

[0123] 1. Using a time-delay recurrent neural network for feedforward decoupling, the system's dynamic characteristics are learned through a data-driven approach, without relying on a precise mathematical model. Compared to traditional feedforward decoupling and frequency conversion decoupling strategies, this approach more thoroughly addresses the coupling relationship between the quality control channel and the quantity control channel, solving the problem of single differential pressure control's inability to coordinate the dynamic response of multiple variables.

[0124] 2. Recurrent neural network structures can memorize historical states and, combined with the rolling optimization mechanism of predictive models, achieve dynamic compensation for time-delay components. Compared to the static parameter optimization of ant colony algorithms and adaptive control, this approach can more effectively alleviate control lag problems caused by time delays.

[0125] 3. By comparing the difference between the predicted output and the actual output in real time, the prediction model parameters and frequency conversion increments are dynamically corrected to form a closed-loop optimization. This mechanism enables the system to autonomously adapt to changes in operating conditions and model mismatch, overcoming the shortcomings of traditional methods that rely on fixed models and have insufficient disturbance rejection.

[0126] 4. The combination of the predictive model and the rolling optimizer simultaneously optimizes temperature and flow control targets on the basis of decoupling, avoiding the control limitations caused by a single optimization index and improving the overall control accuracy of the system.

[0127] like Figure 6 As shown in the figure, this application provides a heat exchange station control device based on neural network decoupling, including:

[0128] At least one processor; and,

[0129] A memory communicatively connected to the at least one processor; wherein,

[0130] The memory stores instructions that can be executed by the at least one processor, which, when executed, enable the at least one processor to perform the heat exchange station control method based on neural network decoupling as described in any of the above embodiments.

[0131] This application provides a non-volatile computer storage medium storing computer-executable instructions configured to implement the heat exchange station control method based on neural network decoupling as described in any of the above embodiments.

[0132] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0133] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0134] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A heat exchange station control method based on neural network decoupling, characterized in that, include: The preset parameter settings and the first actual parameter output values ​​at historical moments are used as inputs. The system performs feedforward decoupling through a trained time-delay recurrent neural network and outputs decoupling control commands. The decoupling control command is taken as input, and the prediction model is used to make predictions, outputting the predicted parameter values ​​for future times. The rolling optimizer outputs the frequency conversion increment based on the predicted parameter output value, and controls the heat exchange station system based on the frequency conversion increment. Obtain the second actual parameter output value at the future time after the heat exchange station system is controlled, and determine the output difference between the second actual parameter output value and the predicted parameter output value; Based on the output difference, the model parameters of the prediction model and / or the frequency conversion increment are adjusted.

2. The heat exchange station control method based on neural network decoupling according to claim 1, characterized in that, The preset parameter settings and the first actual parameter output values ​​at historical moments are used as inputs. A trained time-delay recurrent neural network is used for feedforward decoupling, outputting decoupling control commands, specifically including: The preset parameter settings include the booster pump frequency conversion setting and the circulation pump frequency setting, and the first actual parameter output values ​​at multiple historical moments include the first supply and return water temperature difference measured value sequence and the first secondary network flow measured value sequence. Feedforward decoupling is performed using a trained time-delay recurrent neural network, and decoupling control commands are generated through internal weight calculation; the decoupling control commands include temperature difference channel decoupling commands and flow rate channel decoupling commands. The decoupling control command is output to the prediction model.

3. The heat exchange station control method based on neural network decoupling according to claim 2, characterized in that, The training process of the time-delay recurrent neural network includes: Obtain multiple preset actual parameter settings, and for each actual parameter setting, obtain the first actual parameter output value when the heat exchange station system control reaches steady state based on that actual parameter setting value; For each first actual parameter output value, the ideal parameter setting value under the undecoupled state is obtained by back-reasoning through an independent single-loop channel expression; The ideal parameter settings are used as input, and the corresponding actual parameter settings are used as output to train a time-delay recurrent neural network.

4. The heat exchange station control method based on neural network decoupling according to claim 2, characterized in that, Taking the decoupled control command as input, the trained prediction model makes predictions and outputs the predicted parameter values ​​for future times, specifically including: For the temperature difference channel decoupling instruction and flow channel decoupling instruction included in the decoupling control instruction, determine the corresponding booster pump frequency conversion instruction and circulating pump frequency instruction respectively; The booster pump frequency conversion command is taken as input, and the trained prediction model is used to predict and output a sequence of predicted values ​​for the supply and return water temperature difference at future times. The prediction model is a controlled autoregressive moving average integral model compensated by a neural network decoupler.

5. The heat exchange station control method based on neural network decoupling according to claim 4, characterized in that, The rolling optimizer outputs frequency conversion increments based on the predicted parameter output values, and controls the heat exchange station system based on these frequency conversion increments, specifically including: A target function is constructed, which includes a prediction tracking term and a control penalty term; the prediction tracking term is determined based on the predicted value sequence of the supply and return water temperature difference and the variable frequency setting value of the booster pump; the control penalty term is determined based on the variable frequency increment of the booster pump. The rolling optimizer performs a minimization calculation based on the objective function and outputs a frequency conversion incremental sequence for the booster pump. Based on the frequency increment of the booster pump in the frequency increment sequence at the current moment, an actual control command is generated, and the heat exchange station system is controlled through the actual control command.

6. The heat exchange station control method based on neural network decoupling according to claim 5, characterized in that, The booster pump frequency conversion command is taken as input, and a trained prediction model is used to predict and output a sequence of predicted supply and return water temperature differences for future times, specifically including: Based on the trained prediction model, an implicit prediction equation is established; the implicit prediction equation includes a booster pump frequency conversion increment sequence and a parameter vector; the parameter vector includes a dynamic response matrix and a prediction vector; the dynamic response matrix is ​​used to establish the dynamic relationship between the booster pump frequency conversion increment and the predicted value of the supply and return water temperature difference; the prediction vector is used to represent the control-independent component in the predicted value of the supply and return water temperature difference. Based on the implicit prediction equation, a sequence of predicted supply and return water temperature differences for future times is output.

7. The heat exchange station control method based on neural network decoupling according to claim 6, characterized in that, Based on the output difference, the model parameters of the prediction model are adjusted, specifically including: The dynamic response matrix is ​​solved, the intermediate covariance matrix is ​​determined according to the preset forgetting factor, and the adaptive gain matrix is ​​determined according to the intermediate covariance matrix. The parameter vector of the prediction model is updated based on the output difference and the adaptive gain matrix. And update the covariance matrix based on the adaptive gain matrix; The prediction vector is solved, and the prediction vector is updated based on the output difference and the predicted value of the supply and return water temperature difference at historical time.

8. The heat exchange station control method based on neural network decoupling according to claim 6, characterized in that, Based on the output difference, the frequency increment of the frequency conversion increment is adjusted, specifically including: Based on the dynamic response matrix and the prediction vector, the corresponding frequency conversion increment sequence is determined; Update the corresponding frequency increment based on the frequency increment sequence.

9. A heat exchange station control device based on neural network decoupling, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the heat exchange station control method based on neural network decoupling as described in any one of claims 1 to 8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to implement the heat exchange station control method based on neural network decoupling as described in any one of claims 1 to 8.

Images