A Method for Automating Heating System Scheduling Based on Machine Learning and Predictive Control

Abstract

This invention discloses a method for automating the scheduling of a heating system based on machine learning and predictive control. The method includes: establishing a heat load prediction model for end-users and an energy efficiency prediction model for the heating system using machine learning algorithms; setting up a neural network controller: input variables are state variables, including the heating system's energy efficiency ratio, outdoor meteorological parameters, and end-user heat load; output variables are control variables, including the secondary side water supply temperature, circulating pump control frequency, and electric regulating valve opening; setting the predictive control optimization objective function and equipment state constraints for the heating system; using the neural network controller to perform rolling optimization of the control variables within a finite prediction time domain, with the output being the optimal target value of the control variables; automating the scheduling of the heating system based on the optimal target value of the control variables, and performing feedback correction control based on the scheduling results, so that the heating system is scheduled according to the optimal target value of the control variables.

Description

Technical Field

[0001] This invention belongs to the field of smart heating technology, specifically relating to a method for automating the scheduling of heating systems based on machine learning and predictive control. Background Technology

[0002] Automated scheduling of heating systems involves establishing a remote monitoring, management, and information dissemination platform for heating operations at a dispatch center. Advanced control software is used to comprehensively and automatically monitor, adjust, and intelligently optimize the operating status and parameters of heat sources, heating networks, and various heating stations. With the rapid development of my country's economy, energy consumption is increasing year by year, and the energy crisis is becoming increasingly severe. Residential heating is a significant aspect of energy consumption. How to reduce heat consumption and improve the energy efficiency of heating systems while ensuring normal heating supply has become an important research direction.

[0003] Heating systems are complex nonlinear systems with multiple coupled loops. Traditional PID control parameters are difficult to tune, making the system prone to oscillations under changing operating conditions. Furthermore, PID control methods are difficult to design and debug, resulting in poor control performance and an inability to simultaneously meet user heat load demands and achieve energy savings. In addition, current automated scheduling of heating systems in practical engineering projects is mostly done manually, leading to inefficient management and control and significant energy waste. Therefore, it is essential to research automated control strategies for heating system scheduling to achieve the control objective of maximizing energy savings while meeting user heat load demands.

[0004] Based on the aforementioned technical issues, a new method for automating the scheduling of heating systems based on machine learning and predictive control needs to be designed. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide a method for automating the scheduling of heating systems based on machine learning and predictive control. This method can establish a heat load prediction model for end-users and an energy efficiency prediction model for the heating system through machine learning algorithms, providing a data state variable foundation for the calculation of control variables in the automated scheduling of the heating system. Furthermore, it employs a neural network as a feedback controller to achieve predictive control rolling optimization, combining variational methods and stochastic gradient descent. With the goal of meeting the energy efficiency and heat load requirements of the heating system, a neural network controller is designed to overcome the influence of uncertainties and ensure the system remains stable under high energy efficiency conditions even with constantly changing heat loads. The control variables obtained through predictive control calculations enable automated scheduling of the heating system, allowing the system to be controlled and scheduled according to the control objectives.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] This invention provides a method for automating the scheduling of a heating system based on machine learning and predictive control, comprising:

[0008] A digital twin model of the heating system was established using mechanistic modeling and parameter identification methods.

[0009] Based on the digital twin model of the heating system, machine learning algorithms are used to establish a heat load prediction model for end users of the heating system and an energy efficiency prediction model for the heating system.

[0010] Set up a neural network controller: The input variables are state variables, including at least the heating system energy efficiency ratio, outdoor meteorological parameters, and heat load of end users of the heating system; the output variables are control variables, including at least the secondary side water supply temperature of the heating system, the control frequency of the circulating pump, and the opening degree of the electric regulating valve.

[0011] Under the conditions of meeting the heat load demand of end users and improving the energy efficiency of the heating system, the objective function and equipment state constraints of the predictive control optimization of the heating system are set.

[0012] By using a neural network controller to perform rolling optimization of the control variables within a limited prediction time domain, the output is the optimal target value of the control variables.

[0013] The heating system is automated by scheduling based on the optimal target value of the control variables, and feedback correction control is performed based on the scheduling results, so that the heating system is scheduled according to the optimal target value of the control variables.

[0014] Furthermore, the method of establishing a digital twin model of the heating system using mechanistic modeling and parameter identification includes:

[0015] Physical models, logical models, simulation models, and data models are established for the primary network, secondary network, heating stations, and end-user heat in the heating system. The physical models, logical models, simulation models, and data models are coupled and integrated at multiple levels and scales. After mapping and reconstructing the physical entities in the physical space in the virtual space, a digital twin model of the heating system is established.

[0016] An improved adaptive inertial weighted particle swarm optimization (PSO) algorithm is used to optimize the parameters of a digital twin model of a heating system. This is achieved by combining a nonlinear decreasing update strategy for inertial weights with a mutation operation into the PSO algorithm. The parameters to be identified in the heating system's digital twin model are determined, and the value ranges for each parameter are set. The optimization problem for each parameter is transformed into an optimization problem for particle positions. Position variables are introduced, and the deviation of the heating system's digital twin model is solved using training sample data obtained from the heating system. The improved adaptive inertial weighted PSO algorithm selects the optimal particle positions based on the magnitude of this deviation, thus obtaining the optimal parameters for the heating system's digital twin model. The root mean square error (RMSE) and mean absolute percentage error (MAS) are selected as performance metrics to verify the performance of the heating system's digital twin model.

[0017] Furthermore, the establishment of a heat load prediction model for end users and an energy efficiency prediction model for the heating system based on a digital twin model of the heating system, using machine learning algorithms, includes:

[0018] Based on the digital twin model of the heating system, operational data related to the heat load of end users and operational data related to the energy efficiency of the heating system are obtained;

[0019] The Gabor filter is set as the cloud image feature extraction algorithm to extract the texture direction features and edge information of the cloud layer in different directions, and to obtain the data features of the cloud layer at different scales.

[0020] The data characteristics of the cloud layer at different scales, historical operating data related to the heat load of end users of the heating system, and outdoor temperature, wind direction, and humidity data are used as data samples for the heat load prediction model of end users of the heating system.

[0021] An improved PSO optimization algorithm is used to perform local optimization on the RBF neural network to obtain the optimal initial weights and thresholds, which are then used as the initial values ​​for the LM algorithm to train and obtain the global optimal weights and thresholds of the RBF neural network.

[0022] The data samples of the heating system end-user heat load prediction model are input into the optimized RBF neural network for training to establish the heating system end-user heat load prediction model;

[0023] In addition, the data characteristics of the cloud layer at different scales, the operating data related to the energy efficiency of the heating system, and the outdoor temperature, wind direction, and humidity data are used as data samples for the energy efficiency prediction model of the heating system.

[0024] After decomposing the data samples of the heating system energy efficiency prediction model using the EMD method to obtain multiple energy efficiency feature components, the Conditional Generative Adversarial Network (CGAN) is then used to train the multiple energy efficiency feature components to establish the heating system energy efficiency prediction model.

[0025] Furthermore, the Gabor filter is set as the cloud image feature extraction algorithm to extract texture direction features and edge information of the cloud layer in different directions, thereby obtaining data features of the cloud layer at different scales, including:

[0026] Based on the different sizes and dimensions of the sky and clouds in the cloud map, and their different spatial locations, a Gabor filter bank with 5 scales and 8 directions is selected to extract the texture direction features and edge information of the clouds in different directions; the direction values ​​of the cloud map are set to θ1, θ2, θ3, θ4, θ5, θ6, θ7, θ8, and the scale values ​​are f1, f2, f3, f4, f5;

[0027] After obtaining the Gabor filter bank, each Gabor filter is convolved with the original contour image. Let the original contour image be I(x,y), and the Gabor filter at a specific scale and orientation be G. f,θ If (x, y), then the filtered image is represented as:

[0028]

[0029] f is the scale ordinal number, f = 1, 2, 3, 4, 5; θ is the direction ordinal number, θ = 1, 2, 3, 4, 5, 6, 7, 8; G f,θ (x,y) R G f,θ (x,y) I These are the real and imaginary parts of the Gabor filter at the corresponding scale and direction, respectively.

[0030] After obtaining multiple Gabor features from the cloud image, an energy-based optimization method is used to optimize these features: the total energy of the filtered image is calculated. and the energy of each filtered image The energy of each filtered image is arranged in order, and a preset number of filtered images are selected as preferred feature images.

[0031] The selected feature images are smoothed and filtered to obtain feature matrices of multiple feature images. These feature matrices are then vectorized and transformed into a dataset X = {X1, X2, ..., XN} containing M×N pixel samples, following a left-to-right and top-to-bottom order. i ,…,X S}, (S=M×N), where the i-th sample X i ={x i1 ,x i2 ,…,x ik ,…,x in} is an n-dimensional vector, x ikIt corresponds to the pixel sample X i The k-th Gabor feature;

[0032] The feature vectors are normalized to zero mean to eliminate the differences in the magnitude of the feature values ​​and obtain the final feature vectors.

[0033] Furthermore, the historical operating data related to the heat load of the end users of the heating system shall include at least the current load value, the load value of the previous day, the load value of the previous week, the secondary water supply temperature, the secondary return water temperature, and the secondary water supply flow rate;

[0034] The improved PSO optimization algorithm is used to perform local optimization on the RBF neural network to obtain the optimal initial weights and thresholds, which are then used as initial values ​​for training the LM algorithm to obtain the globally optimal weights and thresholds of the RBF neural network. This includes:

[0035] Introducing the adaptive weighting method into the PSO optimization algorithm forms an improved PSO optimization algorithm, which is expressed as:

[0036]

[0037] w is the inertia weight; w min and w max These are the minimum and maximum values ​​of the inertia weighting factor, respectively; f is the particle's fitness; f avg and f min These are the average and minimum fitness values, respectively.

[0038] Initialize the initial velocity and position of the particles, and set the learning factor, individual optimal coordinates pbest, and population optimal coordinates gbest;

[0039] Calculate particle fitness values ​​and update velocity and position;

[0040] Compare the fitness of each particle position in the population with the fitness of pbest, and update pbest; similarly, compare pbest with gbest and update gbest.

[0041] The network weights are adjusted according to the adaptive weighting method. When the algorithm meets the termination condition or reaches the maximum number of iterations, the optimal initial weights and thresholds of the RBF neural network are output; otherwise, the particle fitness values ​​are recalculated and the velocity and position are updated.

[0042] Initialize the weights and threshold W0 of the RBF neural network, and set the network training error allowable value ε, the maximum number of training iterations, and the initial scale factor μ and step factor β;

[0043] Calculate the error exponent S(W) of the network in the k-th iteration. k ), W kThe weights and thresholds of the neural network generated in the k-th iteration;

[0044] Calculate the Jacobian matrix and obtain the changes ΔW in the network weights and thresholds, while also obtaining the new weights and thresholds W. k+1 , is represented as:

[0045]

[0046] e i (W k ) represents the difference between the predicted and actual output values ​​of the RBF neural network; N is the dimension; J is the Jacobian matrix; I is the identity matrix; μ is a positive constant.

[0047] If S(W) k If ) < ε, satisfying the error requirement, then stop calculating the globally optimal weights and threshold of the output RBF neural network; otherwise, return to recalculate the error exponent S(W) for the (k+1)th iteration. k+1 );

[0048] Compare S(W) k+1 ) and S(W k If S(W) k+1 ) <S(W k If k = k+1, μ = μ / β, then return to recalculate the error exponent; otherwise, μ = μ·β, return to calculate the weight and threshold changes of the RBF neural network.

[0049] Furthermore, the operating data related to the energy efficiency of the heating system shall include at least: the current load value, the current system energy efficiency ratio and the secondary side water supply temperature, the circulation pump control frequency, the opening degree of the electric regulating valve, and the secondary water supply flow rate;

[0050] The process involves using the EMD method to decompose the data samples of the heating system energy efficiency prediction model to obtain multiple energy efficiency feature components, and then using a conditional generative adversarial network (CGAN) to train these multiple energy efficiency feature components to establish the heating system energy efficiency prediction model. This includes:

[0051] The EMD method is used to decompose the data samples of the heating system energy efficiency prediction model into n-1 intrinsic mode function (IMF) components and one residual (RES) component at multiple time scales.

[0052] The structure of the Conditional Adversarial Network (CGAN) is as follows: Based on the game-theoretic structure of the GAN generator and discriminator, a conditional value y is added as input. Random noise z and the conditional value y are used together as input to the generator. The generator generates samples G(z|y). The discriminator judges whether the generated samples under the corresponding conditions are similar to the real samples x, and feeds back the judgment result D(x|y) to the generator and discriminator through a loss function. The generator and discriminator update their own parameters and optimize according to the feedback loss function to reach Nash equilibrium. The generator consists of n LSTM neural networks. The discriminator consists of multiple convolutional layers and one fully connected layer.

[0053] The n-1 intrinsic mode function (IMF) components and 1 residual (RES) component are concatenated with the conditional value y and input into the n LSTM neural networks of the generator. The outputs of the n LSTMs are then summed to obtain the energy efficiency prediction data of the heating system.

[0054] The predicted energy efficiency data and the actual energy efficiency data of the heating system are input into the discriminator along with the conditional value y. Then, the discrimination results of the actual energy efficiency data and the discrimination results of the predicted energy efficiency data are combined to form the cross-entropy and fed back to the discriminator and the generator for model optimization training, thus establishing the energy efficiency prediction model of the heating system.

[0055] Furthermore, the setting of the neural network controller includes:

[0056] The heating system is a multi-input multi-output nonlinear system, represented as:

[0057] x[k+1] = f(x[k], u[k]);

[0058] x[0] = x0;

[0059] x[k] is the n-dimensional state variable at time k, x[k]=[x1(k),x2(k),x3(k),…,x n (k)];u[k] is the m-dimensional control variable at time k, u[k]=[u1(k),u2(k),u3(k),…,u m [(k)]; x0 is the state variable of the system in the initial state; k is the current time, and its value ranges from 0, 1, 2, ...; k+1 is the next time.

[0060] The neural network controller is defined as follows:

[0061] u[k]=g(x[k],x * [k+1],W);

[0062] W is the weight matrix of the neural network controller; x *[k+1] represents the expected value of the state variable at the next time step; the input layer of the neural network controller consists of 2n+1 neurons, with input variables x[k] and x... * The threshold [k+1] corresponds to -1, the output layer consists of m neurons, and the output variable is u[k].

[0063] The input variables include at least the heating system energy efficiency ratio, outdoor meteorological parameters, and heat load of end users of the heating system; the output variables include at least the secondary side water supply temperature of the heating system, the control frequency of the circulating pump, and the opening degree of the electric regulating valve.

[0064] Furthermore, under the conditions of meeting the heat load demand of end users and improving the energy efficiency of the heating system, the objective function and equipment state constraints for predictive control optimization of the heating system are set, including:

[0065] The objective of predictive control optimization for a heating system is to maximize its energy efficiency while meeting the heat load demands of end users. When a setpoint for the heating system's energy efficiency is given, the objective is to ensure that the system's energy efficiency follows that setpoint, expressed as:

[0066]

[0067] J(EER)[k] is the objective function for predictive control optimization of the heating system; M is the prediction time domain; t1 is the start time of the prediction time domain; EER[k] is the energy efficiency of the heating system at the current moment; EER set [k] represents the current energy efficiency setpoint of the heating system, and the predicted energy efficiency value of the heating system obtained based on the heating system energy efficiency prediction model is used as the energy efficiency setpoint; Q[k] and Q set [k] represents the actual value and the given value of the heat load of the end users of the heating system at the current time. The predicted value of the heat load of the end users of the heating system obtained based on the heat load prediction model of the end users of the heating system is used as the given value of the heat load.

[0068] Set equipment status constraints, including at least the parameter ranges for secondary side water supply temperature, circulating pump control frequency, and electric regulating valve opening.

[0069] Furthermore, the step of using a neural network controller to perform rolling optimization of the control variables within a finite prediction time domain, with the output being the optimal target value of the control variables, includes:

[0070] Step S1: Initialize the parameters of the neural network controller, including setting the initial state variables, the initial weight matrix, and determining the expected values ​​of the state variables, the control period, and the prediction time domain;

[0071] Step S2: Initialize the state variable x[k] = EER[k] and set the expected value of EER x. *[k+1] is input to the neural network controller for calculation to obtain the initial value of the control variable u[k];

[0072] Step S3: Obtain the predicted heat load of end users based on the end-user heat load prediction model of the heating system. Then, input the initial values ​​of the control variable u[k], the initial values ​​of the state variable x[k] = EER[k], and the predicted heat load of end users into the heating system energy efficiency prediction model of the system to obtain the predicted values ​​of the state variables at the next time step.

[0073] Step S4: Calculate the control variables and state variables at each time point within a prediction time domain: Keep the weight matrix of the neural network controller unchanged within the same prediction time domain, including the predicted values ​​of the state variables obtained in the previous step. and the expected value of the state variable x at the next moment * The state variable [k+2] is input to the neural network controller to obtain the control variable u[k+1]. Then, based on the end-user heat load prediction model of the heating system, the predicted value of the end-user heat load Q[k+1] is obtained step by step. The predicted values ​​of the control variable u[k+1] and the state variable are then used to obtain the predicted value of the end-user heat load. The predicted heat load value Q[k+1] of the end users is input into the system's heating system energy efficiency prediction model to obtain the predicted value of the state variables at the next time step. By repeatedly performing step S4, a prediction time domain is completed;

[0074] Step S5: The Lagrange multiplier method is used to optimize the objective function. Lagrange multiplier vectors are introduced and Hamiltonian functions are constructed to form an augmented objective function. Then, according to the regular equation, the Lagrange multiplier vectors λ(k) and q(k) are calculated from back to front.

[0075] Step S6: Based on the calculated value of the Lagrange multiplier q(k), correct the weight matrix of the neural network controller using the stochastic gradient descent method;

[0076] Step S7: Repeat steps S4-S6 to modify the weight matrix of the neural network controller until the neural network converges.

[0077] Step S8: Enter the next sampling cycle and repeat steps S2-S7 to obtain the optimal control variables at each sampling time until the entire control process ends.

[0078] The beneficial effects of this invention are:

[0079] This invention establishes a digital twin model of a heating system using mechanistic modeling and parameter identification. Based on this model, machine learning algorithms are used to establish a heat load prediction model for end-users and an energy efficiency prediction model for the heating system. A neural network controller is set up with input variables as state variables, including at least the heating system energy efficiency ratio, outdoor meteorological parameters, and end-user heat load. Output variables are control variables, including at least the secondary side water supply temperature, circulating pump control frequency, and electric regulating valve opening. Under the conditions of meeting end-user heat load demands and improving heating system energy efficiency, a predictive control optimization objective function and equipment state constraints are set. The neural network controller performs rolling optimization on the control variables within a finite prediction time domain, and the output is the optimal target value of the control variables. Based on the optimal target value of the control variables... The system automates the scheduling of the heating system and performs feedback correction control based on the scheduling results, enabling the heating system to be scheduled according to the optimal target values ​​of the control variables. It can establish a heat load prediction model for end-users and an energy efficiency prediction model for the heating system through machine learning algorithms, providing a data state variable foundation for the calculation of control variables in the automated scheduling of the heating system. Furthermore, it employs a neural network as a feedback controller to achieve predictive control rolling optimization, combining variational methods and stochastic gradient descent. With the goal of meeting the energy efficiency and heat load requirements of the heating system, a neural network controller is designed to overcome the influence of uncertainties, ensuring the system remains stable under high energy efficiency conditions even with constantly changing heat loads. The control variables obtained through predictive control calculations enable automated scheduling of the heating system, allowing it to be controlled and scheduled according to the control objectives.

[0080] Other features and advantages will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0081] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0082] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0083] Figure 1This is a flowchart of a method for automating the scheduling of a heating system based on machine learning and predictive control, according to the present invention.

[0084] Figure 2 This is a block diagram illustrating the principle of automated scheduling of a heating system based on machine learning and predictive control, as described in this invention. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0086] Example 1

[0087] Figure 1 This is a flowchart of a method for automating the scheduling of a heating system based on machine learning and predictive control, which is involved in this invention.

[0088] Figure 2 This is a block diagram illustrating the principle of automated scheduling of a heating system based on machine learning and predictive control, as described in this invention.

[0089] like Figure 1-2 As shown in the figure, this embodiment 1 provides a method for automating the scheduling of a heating system based on machine learning and predictive control, which includes:

[0090] A digital twin model of the heating system was established using mechanistic modeling and parameter identification methods.

[0091] Based on the digital twin model of the heating system, machine learning algorithms are used to establish a heat load prediction model for end users of the heating system and an energy efficiency prediction model for the heating system.

[0092] Set up a neural network controller: The input variables are state variables, including at least the heating system energy efficiency ratio, outdoor meteorological parameters, and heat load of end users of the heating system; the output variables are control variables, including at least the secondary side water supply temperature of the heating system, the control frequency of the circulating pump, and the opening degree of the electric regulating valve.

[0093] Under the conditions of meeting the heat load demand of end users and improving the energy efficiency of the heating system, the objective function and equipment state constraints of the predictive control optimization of the heating system are set.

[0094] By using a neural network controller to perform rolling optimization of the control variables within a limited prediction time domain, the output is the optimal target value of the control variables.

[0095] The heating system is automated by scheduling based on the optimal target value of the control variables, and feedback correction control is performed based on the scheduling results, so that the heating system is scheduled according to the optimal target value of the control variables.

[0096] In this embodiment, the establishment of a digital twin model of the heating system using mechanistic modeling and parameter identification methods includes:

[0097] Physical models, logical models, simulation models, and data models are established for the primary network, secondary network, heating stations, and end-user heat in the heating system. The physical models, logical models, simulation models, and data models are coupled and integrated at multiple levels and scales. After mapping and reconstructing the physical entities in the physical space in the virtual space, a digital twin model of the heating system is established.

[0098] An improved adaptive inertial weighted particle swarm optimization (PSO) algorithm is used to optimize the parameters of a digital twin model of a heating system. This is achieved by combining a nonlinear decreasing update strategy for inertial weights with a mutation operation into the PSO algorithm. The parameters to be identified in the heating system's digital twin model are determined, and the value ranges for each parameter are set. The optimization problem for each parameter is transformed into an optimization problem for particle positions. Position variables are introduced, and the deviation of the heating system's digital twin model is solved using training sample data obtained from the heating system. The improved adaptive inertial weighted PSO algorithm selects the optimal particle positions based on the magnitude of this deviation, thus obtaining the optimal parameters for the heating system's digital twin model. The root mean square error (RMSE) and mean absolute percentage error (MAS) are selected as performance metrics to verify the performance of the heating system's digital twin model.

[0099] In this embodiment, the step of establishing a heat load prediction model for end users and an energy efficiency prediction model for the heating system based on a digital twin model of the heating system and using machine learning algorithms includes:

[0100] Based on the digital twin model of the heating system, operational data related to the heat load of end users and operational data related to the energy efficiency of the heating system are obtained;

[0101] The Gabor filter is set as the cloud image feature extraction algorithm to extract the texture direction features and edge information of the cloud layer in different directions, and to obtain the data features of the cloud layer at different scales.

[0102] The data characteristics of the cloud layer at different scales, historical operating data related to the heat load of end users of the heating system, and outdoor temperature, wind direction, and humidity data are used as data samples for the heat load prediction model of end users of the heating system.

[0103] An improved PSO optimization algorithm is used to perform local optimization on the RBF neural network to obtain the optimal initial weights and thresholds, which are then used as the initial values ​​for the LM algorithm to train and obtain the global optimal weights and thresholds of the RBF neural network.

[0104] The data samples of the heating system end-user heat load prediction model are input into the optimized RBF neural network for training to establish the heating system end-user heat load prediction model;

[0105] In addition, the data characteristics of the cloud layer at different scales, the operating data related to the energy efficiency of the heating system, and the outdoor temperature, wind direction, and humidity data are used as data samples for the energy efficiency prediction model of the heating system.

[0106] After decomposing the data samples of the heating system energy efficiency prediction model using the EMD method to obtain multiple energy efficiency feature components, the Conditional Generative Adversarial Network (CGAN) is then used to train the multiple energy efficiency feature components to establish the heating system energy efficiency prediction model.

[0107] In this embodiment, the Gabor filter is set as a cloud image feature extraction algorithm to extract texture direction features and edge information of the cloud layer in different directions, thereby obtaining data features of the cloud layer at different scales, including:

[0108] Based on the different sizes and dimensions of the sky and clouds in the cloud map, and their different spatial locations, a Gabor filter bank with 5 scales and 8 directions is selected to extract the texture direction features and edge information of the clouds in different directions; the direction values ​​of the cloud map are set to θ1, θ2, θ3, θ4, θ5, θ6, θ7, θ8, and the scale values ​​are f1, f2, f3, f4, f5;

[0109] After obtaining the Gabor filter bank, each Gabor filter is convolved with the original contour image. Let the original contour image be I(x,y), and the Gabor filter at a specific scale and orientation be G. f,θ If (x, y), then the filtered image is represented as:

[0110]

[0111] f is the scale ordinal number, f = 1, 2, 3, 4, 5; θ is the direction ordinal number, θ = 1, 2, 3, 4, 5, 6, 7, 8; G f,θ (x,y) R G f,θ (x,y) I These are the real and imaginary parts of the Gabor filter at the corresponding scale and direction, respectively.

[0112] After obtaining multiple Gabor features from the cloud image, an energy-based optimization method is used to optimize these features: the total energy of the filtered image is calculated. and the energy of each filtered image The energy of each filtered image is arranged in order, and a preset number of filtered images are selected as preferred feature images.

[0113] The selected feature images are smoothed and filtered to obtain feature matrices of multiple feature images. These feature matrices are then vectorized and transformed into a dataset X = {X1, X2, ..., XN} containing M×N pixel samples, following a left-to-right and top-to-bottom order. i ,…,X S}, (S=M×N), where the i-th sample X i ={x i1 ,x i2 ,…,x ik ,…,x in} is an n-dimensional vector, x ik It corresponds to the pixel sample X i The k-th Gabor feature;

[0114] The feature vectors are normalized to zero mean to eliminate the differences in the magnitude of the feature values ​​and obtain the final feature vectors.

[0115] It's important to note that cloud maps not only reflect current weather conditions but also predict weather changes over a future period. For routine weather changes, cloud maps can supplement the weather evolution information missing from structured weather data, providing an early indication of current weather trends. For transitional weather changes, cloud maps can promptly correct significant discrepancies between structured meteorological data and actual weather conditions, expressing the meteorological characteristics of sudden weather changes. Therefore, by fusing structured and unstructured meteorological data, lagging unstructured meteorological data reflecting weather conditions can be corrected, yielding accurate parameters characterizing the current weather state. Using the integrated meteorological information obtained from this fusion, prediction models can more objectively establish complex causal relationships between meteorological data and heat load and energy efficiency, significantly improving the accuracy of heat load and energy efficiency predictions.

[0116] In this embodiment, the historical operating data related to the heat load of the end users of the heating system includes at least the current load value, the load value of the previous day, the load value of the previous week, the secondary water supply temperature, the secondary return water temperature, and the secondary water supply flow rate.

[0117] The improved PSO optimization algorithm is used to perform local optimization on the RBF neural network to obtain the optimal initial weights and thresholds, which are then used as initial values ​​for training the LM algorithm to obtain the globally optimal weights and thresholds of the RBF neural network. This includes:

[0118] Introducing the adaptive weighting method into the PSO optimization algorithm forms an improved PSO optimization algorithm, which is expressed as:

[0119]

[0120] w is the inertia weight; w min and w max These are the minimum and maximum values ​​of the inertia weighting factor, respectively; f is the particle's fitness; f avg and f min These are the average and minimum fitness values, respectively.

[0121] Initialize the initial velocity and position of the particles, and set the learning factor, individual optimal coordinates pbest, and population optimal coordinates gbest;

[0122] Calculate particle fitness values ​​and update velocity and position;

[0123] Compare the fitness of each particle position in the population with the fitness of pbest, and update pbest; similarly, compare pbest with gbest and update gbest.

[0124] The network weights are adjusted according to the adaptive weighting method. When the algorithm meets the termination condition or reaches the maximum number of iterations, the optimal initial weights and thresholds of the RBF neural network are output; otherwise, the particle fitness values ​​are recalculated and the velocity and position are updated.

[0125] Initialize the weights and threshold W0 of the RBF neural network, and set the network training error allowable value ε, the maximum number of training iterations, and the initial scale factor μ and step factor β;

[0126] Calculate the error exponent S(W) of the network in the k-th iteration. k ), W k The weights and thresholds of the neural network generated in the k-th iteration;

[0127] Calculate the Jacobian matrix and obtain the changes ΔW in the network weights and thresholds, while also obtaining the new weights and thresholds W. k+1 , is represented as:

[0128]

[0129] e i (W k ) represents the difference between the predicted and actual output values ​​of the RBF neural network; N is the dimension; J is the Jacobian matrix; I is the identity matrix; μ is a positive constant.

[0130] If S(W) kIf ) < ε, satisfying the error requirement, then stop calculating the globally optimal weights and threshold of the output RBF neural network; otherwise, return to recalculate the error exponent S(W) for the (k+1)th iteration. k+1 );

[0131] Compare S(W) k+1 ) and S(W k If S(W) k+1 ) <S(W k If k = k+1, μ = μ / β, then return to recalculate the error exponent; otherwise, μ = μ·β, return to calculate the weight and threshold changes of the RBF neural network.

[0132] In this embodiment, the operating data related to the energy efficiency of the heating system includes at least: the current load value, the current system energy efficiency ratio and the secondary side water supply temperature, the circulation pump control frequency, the electric regulating valve opening degree, and the secondary water supply flow rate;

[0133] The process involves using the EMD method to decompose the data samples of the heating system energy efficiency prediction model to obtain multiple energy efficiency feature components, and then using a conditional generative adversarial network (CGAN) to train these multiple energy efficiency feature components to establish the heating system energy efficiency prediction model. This includes:

[0134] The EMD method is used to decompose the data samples of the heating system energy efficiency prediction model into n-1 intrinsic mode function (IMF) components and one residual (RES) component at multiple time scales.

[0135] The structure of the Conditional Adversarial Network (CGAN) is as follows: Based on the game-theoretic structure of the GAN generator and discriminator, a conditional value y is added as input. Random noise z and the conditional value y are used together as input to the generator. The generator generates samples G(z|y). The discriminator judges whether the generated samples under the corresponding conditions are similar to the real samples x, and feeds back the judgment result D(x|y) to the generator and discriminator through a loss function. The generator and discriminator update their own parameters and optimize according to the feedback loss function to reach Nash equilibrium. The generator consists of n LSTM neural networks. The discriminator consists of multiple convolutional layers and one fully connected layer.

[0136] The n-1 intrinsic mode function (IMF) components and 1 residual (RES) component are concatenated with the conditional value y and input into the n LSTM neural networks of the generator. The outputs of the n LSTMs are then summed to obtain the energy efficiency prediction data of the heating system.

[0137] The predicted energy efficiency data and the actual energy efficiency data of the heating system are input into the discriminator along with the conditional value y. Then, the discrimination results of the actual energy efficiency data and the discrimination results of the predicted energy efficiency data are combined to form the cross-entropy and fed back to the discriminator and the generator for model optimization training, thus establishing the energy efficiency prediction model of the heating system.

[0138] In this embodiment, setting the neural network controller includes:

[0139] The heating system is a multi-input multi-output nonlinear system, represented as:

[0140] x[k+1] = f(x[k], u[k]);

[0141] x[0] = x0;

[0142] x[k] is the n-dimensional state variable at time k, x[k]=[x1(k),x2(k),x3(k),…,x n (k)];u[k] is the m-dimensional control variable at time k, u[k]=[u1(k),u2(k),u3(k),…,u m [(k)]; x0 is the state variable of the system in the initial state; k is the current time, and its value ranges from 0, 1, 2, ...; k+1 is the next time.

[0143] The neural network controller is defined as follows:

[0144] u[k]=g(x[k],x * [k+1],W);

[0145] W is the weight matrix of the neural network controller; x * [k+1] represents the expected value of the state variable at the next time step; the input layer of the neural network controller consists of 2n+1 neurons, with input variables x[k] and x... * The threshold [k+1] corresponds to -1, the output layer consists of m neurons, and the output variable is u[k].

[0146] The input variables include at least the heating system energy efficiency ratio, outdoor meteorological parameters, and heat load of end users of the heating system; the output variables include at least the secondary side water supply temperature of the heating system, the control frequency of the circulating pump, and the opening degree of the electric regulating valve.

[0147] In this embodiment, the step of setting the objective function and equipment state constraints for predictive control optimization of the heating system, under the conditions of meeting the heat load demand of end users and improving the energy efficiency of the heating system, includes:

[0148] The objective of predictive control optimization for a heating system is to maximize its energy efficiency while meeting the heat load demands of end users. When a setpoint for the heating system's energy efficiency is given, the objective is to ensure that the system's energy efficiency follows that setpoint, expressed as:

[0149]

[0150] J(EER)[k] is the objective function for predictive control optimization of the heating system; M is the prediction time domain; t1 is the start time of the prediction time domain; EER[k] is the energy efficiency of the heating system at the current moment; EER set [k] represents the current energy efficiency setpoint of the heating system, and the predicted energy efficiency value of the heating system obtained based on the heating system energy efficiency prediction model is used as the energy efficiency setpoint; Q[k] and Q set [k] represents the actual value and the given value of the heat load of the end users of the heating system at the current time. The predicted value of the heat load of the end users of the heating system obtained based on the heat load prediction model of the end users of the heating system is used as the given value of the heat load.

[0151] Set equipment status constraints, including at least the parameter ranges for secondary side water supply temperature, circulating pump control frequency, and electric regulating valve opening.

[0152] In this embodiment, the step of using a neural network controller to perform rolling optimization of the control variables within a finite prediction time domain, with the output being the optimal target value of the control variables, includes:

[0153] Step S1: Initialize the parameters of the neural network controller, including setting the initial state variables, the initial weight matrix, and determining the expected values ​​of the state variables, the control period, and the prediction time domain;

[0154] Step S2: Initialize the state variable x[k] = EER[k] and set the expected value of EER x. * [k+1] is input to the neural network controller for calculation to obtain the initial value of the control variable u[k];

[0155] Step S3: Obtain the predicted heat load of end users based on the end-user heat load prediction model of the heating system. Then, input the initial values ​​of the control variable u[k], the initial values ​​of the state variable x[k] = EER[k], and the predicted heat load of end users into the heating system energy efficiency prediction model of the system to obtain the predicted values ​​of the state variables at the next time step.

[0156] Step S4: Calculate the control variables and state variables at each time point within a prediction time domain: Keep the weight matrix of the neural network controller unchanged within the same prediction time domain, including the predicted values ​​of the state variables obtained in the previous step. and the expected value of the state variable x at the next moment * The state variable [k+2] is input to the neural network controller to obtain the control variable u[k+1]. Then, based on the end-user heat load prediction model of the heating system, the predicted value of the end-user heat load Q[k+1] is obtained step by step. The predicted values ​​of the control variable u[k+1] and the state variable are then used to obtain the predicted value of the end-user heat load. The predicted heat load value Q[k+1] of the end users is input into the system's heating system energy efficiency prediction model to obtain the predicted value of the state variables at the next time step. By repeatedly performing step S4, a prediction time domain is completed;

[0157] Step S5: The Lagrange multiplier method is used to optimize the objective function. Lagrange multiplier vectors are introduced and Hamiltonian functions are constructed to form an augmented objective function. Then, according to the regular equation, the Lagrange multiplier vectors λ(k) and q(k) are calculated from back to front.

[0158] Step S6: Based on the calculated value of the Lagrange multiplier q(k), correct the weight matrix of the neural network controller using the stochastic gradient descent method;

[0159] Step S7: Repeat steps S4-S6 to modify the weight matrix of the neural network controller until the neural network converges.

[0160] Step S8: Enter the next sampling cycle and repeat steps S2-S7 to obtain the optimal control variables at each sampling time until the entire control process ends.

[0161] It should be noted that the general expression for the objective function of predictive control optimization is:

[0162]

[0163] J is the optimization objective function, also known as the optimization performance index; N is the control period; φ[X(N),N] is the terminal objective function; k is the current time; L[X(k),U(k),k] is the process optimization objective function in the prediction time domain;

[0164] Predictive control uses a rolling optimization approach to find the optimal sequence of control variables in the prediction time domain. The goal is to minimize J in the prediction time domain. Therefore, the objective function of predictive control can be rewritten as follows:

[0165]

[0166] φ t1+M [X(t1+M),t1+M] represents the final value optimization objective function in the prediction time domain; M represents the prediction time domain; t1 represents the start time of the prediction time domain.

[0167] Using the Lagrange multiplier method, we introduce undetermined Lagrange multiplier vectors and construct an augmented optimization objective function, expressed as:

[0168]

[0169] Let λ and q be n-dimensional and m-dimensional Lagrange multiplier vectors, respectively; and let f(k) and g(k,W) represent the right-hand sides of the prediction model and the neural network controller, respectively. Then, the Hamiltonian function can be obtained, expressed as:

[0170] H(k)=H(x,x * ,u,λ,q,k,W)=L(x[k],u[k],k)+λ T [k+1]f(k)+q T [k]g(k,W);

[0171] Assuming the variational values ​​of the state variables, control variables, and neural network controller weight matrix are δx[k], δu[k], and δW, respectively, for any variational vectors δx[k], δu[k], and δW, minimize the augmented optimization objective function J. a The necessary condition is δJ a =0.

[0172] To minimize the augmented optimization objective function, the following condition must be met:

[0173] Hamiltonian canonical equation:

[0174]

[0175]

[0176] Cross section condition:

[0177]

[0178] Extreme value conditions:

[0179]

[0180] In practical applications, neural networks are used as the optimization feedback controller. The objective function of the control system optimization, namely satisfying the optimal heat load of the end users and the optimal energy efficiency of the heating system, is used as the optimization performance index of the neural network. Based on the Lagrange algorithm and the stochastic gradient descent method, the weights of the neural network controller are rolled for optimization, which can handle the uncertainty problems of disturbances and model inaccuracy. Moreover, the calculation of control variables based on the neural network controller has the advantages of small computational load and small storage space.

[0181] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can also be implemented in other ways. The system embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0182] Furthermore, the functional modules in the various embodiments of this invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. If the function is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods in the various embodiments of this invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0183] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for automating the scheduling of a heating system based on machine learning and predictive control, characterized in that, It includes: A digital twin model of the heating system was established using mechanistic modeling and parameter identification methods. Based on the digital twin model of the heating system, machine learning algorithms are used to establish a heat load prediction model for end users and an energy efficiency prediction model for the heating system, including: Based on the digital twin model of the heating system, operational data related to the heat load of end users and operational data related to the energy efficiency of the heating system are obtained; The Gabor filter is set as the cloud image feature extraction algorithm to extract the texture direction features and edge information of the cloud layer in different directions, and to obtain the data features of the cloud layer at different scales. The data characteristics of the cloud layer at different scales, historical operating data related to the heat load of end users of the heating system, and outdoor temperature, wind direction, and humidity data are used as data samples for the heat load prediction model of end users of the heating system. An improved PSO optimization algorithm is used to perform local optimization on the RBF neural network to obtain the optimal initial weights and thresholds, which are then used as the initial values ​​for the LM algorithm to train and obtain the global optimal weights and thresholds of the RBF neural network. The data samples of the heating system end-user heat load prediction model are input into the optimized RBF neural network for training to establish the heating system end-user heat load prediction model. In addition, the data characteristics of the cloud layer at different scales, the operating data related to the energy efficiency of the heating system, and the outdoor temperature, wind direction, and humidity data are used as data samples for the energy efficiency prediction model of the heating system. After decomposing the data samples of the heating system energy efficiency prediction model using the EMD method to obtain multiple energy efficiency feature components, the conditional generative adversarial network CGAN is then used to train the multiple energy efficiency feature components to establish the heating system energy efficiency prediction model. Set up a neural network controller: The input variables are state variables, including at least the heating system energy efficiency ratio, outdoor meteorological parameters, and heat load of end users of the heating system; the output variables are control variables, including at least the secondary side water supply temperature of the heating system, the control frequency of the circulating pump, and the opening degree of the electric regulating valve. Under the conditions of meeting the heat load demand of end users and improving the energy efficiency of the heating system, the objective function and equipment state constraints of the predictive control optimization of the heating system are set. By using a neural network controller to perform rolling optimization of the control variables within a limited prediction time domain, the output is the optimal target value of the control variables. The heating system is automated by scheduling based on the optimal target value of the control variables, and feedback correction control is performed based on the scheduling results, so that the heating system is scheduled according to the optimal target value of the control variables.

2. The method for automating the scheduling of a heating system according to claim 1, characterized in that, The establishment of a digital twin model of the heating system using mechanistic modeling and parameter identification methods includes: Physical models, logical models, simulation models, and data models are established for the primary network, secondary network, heating stations, and end-user heat in the heating system. The physical models, logical models, simulation models, and data models are coupled and integrated at multiple levels and scales. After mapping and reconstructing the physical entities in the physical space in the virtual space, a digital twin model of the heating system is established. An improved adaptive inertial weighted particle swarm optimization (PSO) algorithm is used to optimize the parameters of a digital twin model of a heating system. This is achieved by combining a nonlinear decreasing update strategy for inertial weights with a mutation operation into the PSO algorithm. The parameters to be identified in the heating system's digital twin model are determined, and the value ranges for each parameter are set. The optimization problem for each parameter is transformed into an optimization problem for particle positions. Position variables are introduced, and the deviation of the heating system's digital twin model is solved using training sample data obtained from the heating system. The improved adaptive inertial weighted PSO algorithm selects the optimal particle positions based on the magnitude of this deviation, thus obtaining the optimal parameters for the heating system's digital twin model. The root mean square error (RMSE) and mean absolute percentage error (MAS) are selected as performance metrics to verify the performance of the heating system's digital twin model.

3. The method for automating the scheduling of a heating system according to claim 1, characterized in that, The Gabor filter is set as the cloud image feature extraction algorithm to extract texture direction features and edge information of the cloud layer in different directions, thereby obtaining data features of the cloud layer at different scales, including: Based on the different sizes and dimensions of the sky and clouds in the cloud image, and their different spatial locations, a Gabor filter bank with 5 scales and 8 directions was selected to extract the texture direction features and edge information of the clouds in different directions; the direction value of the cloud image was set as... , , , , , , , The scale value is , , , , ; After obtaining the Gabor filter bank, each Gabor filter is convolved with the original contour image. Let the original contour image be... Gabor filters of specific scales and orientations are The filtered image is represented as: ； For scale ordinal numbers, ; The direction ordinal number, ; , These are the real and imaginary parts of the Gabor filter at the corresponding scale and direction, respectively; After obtaining multiple Gabor features from the cloud image, an energy-based selection method is used to select from these features: the total energy of the filtered image is calculated. and the energy of each filtered image The energy of each filtered image is arranged in order, and a preset number of filtered images are selected as the selected feature images. The selected feature images are smoothed and filtered to obtain feature matrices from multiple feature images. These feature matrices are then vectorized, transformed into a matrix containing multiple feature images in a left-to-right and top-to-bottom order. Dataset of pixel samples , The i-th sample It is an n-dimensional vector. It corresponds to the pixel sample. The k-th Gabor feature; The feature vectors are normalized to zero mean to eliminate the differences in the magnitude of the feature values ​​and obtain the final feature vectors.

4. The method for automating the scheduling of a heating system according to claim 1, characterized in that, The historical operating data related to the heat load of the end users of the heating system shall include at least the current load value, the load value of the previous day, the load value of the previous week, the secondary water supply temperature, the secondary return water temperature, and the secondary water supply flow rate; The improved PSO optimization algorithm is used to perform local optimization on the RBF neural network to obtain the optimal initial weights and thresholds, which are then used as initial values ​​for training the LM algorithm to obtain the globally optimal weights and thresholds of the RBF neural network. This includes: Introducing the adaptive weighting method into the PSO optimization algorithm forms an improved PSO optimization algorithm, which is expressed as: ； Inertial weights; and These are the minimum and maximum values ​​of the inertia weighting factor, respectively. For the fitness of particles; and These are the average and minimum fitness values, respectively. Initialize the initial velocity and position of the particles, and set the learning factor, individual optimal coordinates pbest, and population optimal coordinates gbest; Calculate particle fitness values ​​and update velocity and position; Compare the fitness of each particle position in the population with the fitness of pbest, and update pbest; similarly, compare pbest with gbest and update gbest. The network weights are adjusted according to the adaptive weighting method. When the algorithm meets the termination condition or reaches the maximum number of iterations, the optimal initial weights and thresholds of the RBF neural network are output; otherwise, the particle fitness values ​​are recalculated and the velocity and position are updated. Initialize the weights and thresholds of the RBF neural network And set the allowable value for network training error. Maximum number of training iterations and initial scaling factor and step factor ; Calculate the error exponent of the network in the kth iteration. , The weights and thresholds of the neural network generated in the k-th iteration; Calculate the Jacobian matrix and obtain the changes in network weights and thresholds. At the same time, new weights and thresholds are obtained. , represented as: ； This represents the difference between the predicted output value and the actual output value of the RBF neural network. Dimension; It is a Jacobian matrix; For unit array; like If the error requirement is met, stop calculating the globally optimal weights and thresholds of the output RBF neural network; otherwise, return to recalculate the error exponent for the (k+1)th iteration. ; Compare and ,like < Then k = k + 1, If the error exponent is not found, return and recalculate it; otherwise, Returns the calculated changes in the weights and thresholds of the RBF neural network.

5. The method for automating the scheduling of a heating system according to claim 1, characterized in that, The energy efficiency-related operating data of the heating system includes at least the following: current load value, current system energy efficiency ratio and secondary side water supply temperature, circulating pump control frequency, electric regulating valve opening degree, and secondary water supply flow rate. The process involves using the EMD method to decompose the data samples of the heating system energy efficiency prediction model to obtain multiple energy efficiency feature components, and then using a conditional generative adversarial network (CGAN) to train these multiple energy efficiency feature components to establish the heating system energy efficiency prediction model. This includes: The EMD method is used to decompose the data samples of the heating system energy efficiency prediction model into n-1 intrinsic mode function (IMF) components and one residual (RES) component at multiple time scales. The structure of the Conditional Adversarial Network (CGAN) is set up as follows: Based on the game-theoretic structure of the GAN generator and discriminator, a conditional value y is added as input. Random noise z, along with the conditional value y, serves as input to the generator, which then generates samples. The discriminator distinguishes between generated samples and real samples under corresponding conditions. Whether they are similar, and the result of the judgment. The loss function is fed back to the generator and discriminator; the generator and discriminator update their own parameters and optimize themselves according to the fed-back loss function to achieve Nash equilibrium; the generator consists of n LSTM neural networks; the discriminator consists of multiple convolutional layers and 1 fully connected layer. The n-1 intrinsic mode function (IMF) components and 1 residual (RES) component are concatenated with the conditional value y and input into the n LSTM neural networks of the generator. The outputs of the n LSTMs are then summed to obtain the energy efficiency prediction data of the heating system. The predicted energy efficiency data and the actual energy efficiency data of the heating system are input into the discriminator along with the conditional value y. Then, the discrimination results of the actual energy efficiency data and the discrimination results of the predicted energy efficiency data are combined to form the cross-entropy and fed back to the discriminator and the generator for model optimization training, thus establishing the energy efficiency prediction model of the heating system.

6. The method for automating the scheduling of a heating system according to claim 1, characterized in that, The setting of the neural network controller includes: The heating system is a multi-input multi-output nonlinear system, represented as: ； ； Let n be the state variables at time k. ; Let m be the control variables at time k. ; These are the state variables of the system in its initial state. For the current moment, its value range is: ; For the next moment; The neural network controller is defined as follows: ； This is the weight matrix of the neural network controller; The expected value of the state variable at the next time step; the input layer of the neural network controller consists of 2n+1 neurons, and the input variable is... , With a threshold of -1, the output layer consists of m neurons, and the output variable is... ; The input variables include at least the heating system energy efficiency ratio, outdoor meteorological parameters, and heat load of end users of the heating system; the output variables include at least the secondary side water supply temperature of the heating system, the control frequency of the circulating pump, and the opening degree of the electric regulating valve.

7. The method for automating the scheduling of a heating system according to claim 1, characterized in that, The objective function and equipment state constraints for predictive control optimization of the heating system are set under the conditions of meeting the heat load demand of end users and improving the energy efficiency of the heating system, including: The objective of predictive control optimization for a heating system is to maximize its energy efficiency while meeting the heat load demands of end users. When a setpoint for the heating system's energy efficiency is given, the objective is to ensure that the system's energy efficiency follows that setpoint, expressed as: ； Optimize the objective function for predictive control of the heating system; For prediction in the time domain; It is the start time of the predicted time domain; The current energy efficiency of the heating system; The current energy efficiency setpoint for the heating system is set using the predicted energy efficiency value of the heating system obtained based on the energy efficiency prediction model of the heating system. and These are the actual and given values ​​of the heat load of the end users of the heating system at the current moment, respectively. The predicted value of the heat load of the end users of the heating system obtained based on the heat load prediction model of the end users of the heating system is used as the given value of the heat load. Set equipment status constraints, including at least the parameter ranges for secondary side water supply temperature, circulating pump control frequency, and electric regulating valve opening.

8. The method for automating the scheduling of a heating system according to claim 1, characterized in that, The method of using a neural network controller to perform rolling optimization of control variables within a finite prediction time domain, with the output being the optimal target value of the control variables, includes: Step S1: Initialize the parameters of the neural network controller, including setting the initial state variables, the initial weight matrix, and determining the expected values ​​of the state variables, the control period, and the prediction time domain; Step S2: Transfer the state variables Initial value, expected value of EER The input is fed into the neural network controller for calculation to obtain the control variables. The initial value; Step S3: Obtain the predicted heat load values ​​for end users based on the end-user heat load prediction model of the heating system, and then adjust the control variables. initial values ​​and state variables The initial values ​​and the predicted values ​​of the end-user heat load are input into the system's heating system energy efficiency prediction model to obtain the predicted values ​​of the state variables at the next time step. ; Step S4: Calculate the control variables and state variables at each time point within a prediction time domain: Keep the weight matrix of the neural network controller unchanged within the same prediction time domain, including the predicted values ​​of the state variables obtained in the previous step. and the expected value of the state quantity at the next moment The state variables are input to the neural network controller to obtain the control variables. Then, based on the end-user heat load prediction model of the heating system, the predicted values ​​of the end-user heat load are gradually obtained. Control variables Predicted values ​​of state variables and end-user heat load forecast The energy efficiency prediction model of the heating system is input into the system to obtain the predicted values ​​of the state variables at the next time step. By repeatedly performing step S4, a prediction time domain is completed. Step S5: Apply the Lagrange multiplier method to the objective function, introduce the Lagrange multiplier vector, and construct the Hamiltonian function to form the augmented objective function. Then, calculate the Lagrange multiplier vector from back to front according to the canonical equation. and ; Step S6: Based on the calculated Lagrange multipliers The value of is used to correct the weight matrix of the neural network controller based on the stochastic gradient descent method; Step S7: Repeat steps S4-S6 to modify the weight matrix of the neural network controller until the neural network converges. Step S8: Enter the next sampling cycle and repeat steps S2-S7 to obtain the optimal control variables at each sampling time until the entire control process ends.

Images