A method and system for dynamic load matching regulation of a waste heat recovery system

Abstract

The application discloses a kind of waste heat recovery system dynamic load matching regulation method and system, belong to industrial waste heat energy-saving regulation field.The source, load and system multiple operating parameters are collected, and the supply-demand load prediction curve is deduced by physical information neural network and multi-scale time series convolution attention network respectively using double-channel decoupling architecture;An integrated energy efficiency objective function is constructed by combining multiple penalty terms and Lyapunov terminal cost, and a relaxation variable is introduced to handle equipment hard constraints, and collaborative control instructions are obtained to adjust various types of actuators;After regulation, the time series residual dynamic determination threshold is used to fine-tune the deep network of the prediction model for online self-correction.The supporting system includes five modules: data acquisition, double-channel prediction, rolling optimization, collaborative execution and adaptive update, and corresponding computer equipment and storage media are also provided.The application solves the problems of supply-demand load space-time mismatch, prediction drift, optimization easy to have no solution and long-term regulation precision attenuation, and improves the waste heat absorption level and system stability.

Description

Technical Field

[0001] This invention relates to the field of industrial waste heat energy-saving regulation technology, specifically to a dynamic load matching regulation method and system for waste heat recovery systems. Background Technology

[0002] With the continuous advancement of industrial energy conservation and carbon reduction transformation, various process flue gas and circulating medium waste heat recovery systems have been implemented on a large scale. Industrial production processes are intermittent, and the heat flow and temperature of waste heat sources fluctuate continuously and randomly. Meanwhile, the heat load of multiple branch heat terminals in the plant changes synchronously with production schedules and ambient temperatures, resulting in a natural temporal and spatial mismatch between supply and demand heat power. Traditional waste heat recovery often uses fixed threshold PID control, adjusting pumps and valves based solely on the single parameter of water supply temperature in a closed loop. This lacks a unified quantitative prediction and global collaborative scheduling mechanism for both source and load loads. The direct discharge of excess waste heat and the delayed activation of supplementary heating equipment when heat shortages occur frequently, resulting in low waste heat utilization rates. Frequent thermal shocks to the pipeline network also exacerbate fatigue wear on heat exchangers and pipes. Achieving dynamic and precise matching between source-side waste heat supply and load-side heat demand has become a core technical challenge in improving the overall energy efficiency of waste heat recovery systems.

[0003] Currently, both domestic and international academic fields have conducted extensive research on heat load forecasting and waste heat system model predictive control, with related results published in core journals such as *Modern Electric Power* and *Thermal Power Engineering*. Existing literature primarily focuses on modeling heat source generation forecasting or terminal load forecasting separately. Common approaches include using BP neural networks and temporal convolutional networks to construct single-sided load forecasting models, followed by a simplified rolling optimization control framework to adjust heat output. However, these approaches generally suffer from significant shortcomings: the source and load are trained using a single shared forecasting model; the physical conservation constraints of industrial waste heat cannot be embedded in the network loss function; and purely data-driven models exhibit severe prediction drift in sparse data intervals such as equipment start-up and shutdown, and sudden changes in operating conditions. Furthermore, they cannot separately explore the operating mechanisms of waste heat source equipment and the temporal inertial characteristics of heat-using terminals, making it difficult to maintain prediction accuracy while considering the source side. Strong physical constraints and multi-scale temporal variations on the load side; the optimization objective function simply superimposes heating deviation and operating energy consumption, without introducing Mahalanobis distance weighting for supply and demand deviation, control increment constraints, and closed-loop stability terminal costs. There is no relaxation variable fault tolerance mechanism when hard constraints conflict, which easily leads to no optimization solution and control oscillations under actual working conditions; fourth, the model is only trained offline once, lacking a hierarchical adaptive fine-tuning strategy based on real-time prediction residuals. Long-term operation leads to continuous accumulation of prediction errors, and the control accuracy decays rapidly with equipment aging and operating condition deviations, making it difficult to support long-term stable dynamic load matching control in industrial scenarios.

[0004] Existing domestic patents also propose technical solutions for intelligent control of waste heat recovery. For example, publication number CN118189442A discloses a temperature control system for plant factories based on waste heat recovery. This system collects the temperature and flow parameters of the medium on the waste heat side, predicts the waste heat supply through a basic model, and adjusts the opening of the heat exchange branch valves according to the predicted values. This alleviates the problem of lag in traditional set-value control to some extent, but still has several key technical defects: This patent only focuses on the parameter control of the waste heat supply side and does not simultaneously collect demand-side characteristics such as multi-terminal heat load and ambient temperature, making it impossible to quantify the supply and demand load difference of the entire system; The model is constructed using a shallow neural network driven purely by data, without introducing physical constraints on energy conservation. Under highly fluctuating operating conditions, the prediction error increases significantly. The optimized scheduling only sets simple upper and lower temperature constraints, without constructing a multivariate rolling optimization framework aimed at achieving optimal comprehensive energy efficiency, and lacks control increment constraints and constraint relaxation fault tolerance mechanisms. At the same time, the model does not have online adaptive update logic, and the prediction deviation continues to expand after long-term operation. It can only achieve local parameter stability on the waste heat side and cannot complete dynamic load matching on both the source and load sides in the entire time domain. The core problem of high energy consumption for waste heat release and supplementary heating has not been fundamentally solved. Summary of the Invention

[0005] Based on the above-mentioned technical problems, this application discloses a dynamic load matching and control method and system for a waste heat recovery system; the dynamic load matching and control method for a waste heat recovery system includes:

[0006] Real-time collection of heat generation status parameters from the waste heat source side, demand status parameters from the heat load side, and operating status parameters of the waste heat recovery system;

[0007] Based on the heat production state parameters and the demand state parameters, the waste heat supply prediction curve and the heat demand prediction curve within a preset time window are obtained through a pre-set dynamic load prediction model.

[0008] Based on the waste heat supply prediction curve, the heat demand prediction curve, and the operating status parameters, the dynamic matching control command sequence of each adjustable actuator in the waste heat recovery system is generated with the optimal overall system energy efficiency as the objective function; wherein, the objective function is coupled with the heat recovery amount, system energy consumption, and supply-demand deviation penalty term;

[0009] The adjustable execution components are coordinated and adjusted according to the dynamic matching control command sequence, so that the actual recovery load of the waste heat recovery system dynamically tracks the heat demand prediction curve and responds to the fluctuations of the waste heat supply prediction curve.

[0010] Preferably, the dynamic load prediction model is a dual-channel decoupled prediction architecture, including a source-side heat generation dynamic sub-model and a load-side heat demand dynamic sub-model.

[0011] The inputs of the source-side heat generation dynamic sub-model include the operating frequency, medium flow rate, inlet temperature and historical heat generation power sequence of the waste heat source equipment, and the output is the waste heat supply prediction curve within a future preset time window.

[0012] The inputs of the load-side heat demand dynamic sub-model include the valve opening degree of the heat-using terminal, return water temperature, ambient temperature and historical heat demand power sequence, and the output is the heat demand prediction curve within a future preset time window.

[0013] The two sub-models are trained and inferenced independently, and then jointly solved through coupling constraints during the optimization decision-making stage.

[0014] Preferably, the source-side heat generation dynamic sub-model is constructed using a physical information neural network, and its loss function is a weighted sum of a data-driven error term and a physical conservation residual term, as shown in the formula: ,in To predict the heat production power of the model, To measure the actual heat production capacity, For the density of the medium, For specific heat capacity, For volumetric flow rate, , These are the inlet and outlet temperatures of the heat exchanger, For physically conserved residual weights, For gradient regularization weights, For network learnable parameters, The Frobenius norm is used; the physical conservation residual term forces the model output to satisfy the energy conservation equation, and the gradient regularization term suppresses overfitting oscillations of the model in sparse data regions.

[0015] Preferably, the load-side heat demand dynamic sub-model is constructed using a multi-scale temporal convolutional attention network, and the output of its core feature extraction layer is calculated using the following formula:

[0016]

[0017]

[0018] in For the first Layer at time The output characteristics, To dilate the weights of the causal convolution kernel, The time delay corresponds to the expansion factor. For input to the previous layer, For activation function, For element-wise multiplication, For channel attention module, For global average pooling, , For the parameters of the fully connected layer of the attention module, The sigmoid function is used; the dilated causal convolution is used to capture the heat-demanding inertia features at different time scales, and the channel attention module is used to adaptively enhance the feature channels that are sensitive to load mutations.

[0019] Preferably, the formula for the objective function is: ,in To control the incremental sequence, To predict the time domain, , These are the predicted recovery volume and the demand volume, respectively. The Mahalanobis distance weight matrix represents the supply-demand imbalance. The time-varying deviation penalty coefficient and it varies with... Increasing monotonically decreasing, The total power consumption of the system is To control the increment, To control the incremental weighting matrix, , These are the weighting coefficients. For the terminal cost function, The terminal weights; the Mahalanobis distance weight matrix The terminal cost function is obtained by inverse estimation of the historical supply and demand deviation covariance matrix and is used to normalize the deviation under different operating conditions. This is an approximation of the Lyapunov function obtained offline, used to ensure the closed-loop stability of the rolling optimization.

[0020] Preferably, when solving the objective function, slack variables are introduced to handle the infeasibility of hard constraints, and the constraints are rewritten as follows:

[0021]

[0022]

[0023] in For the first The primitive inequality constraint functions, For the corresponding slack variables, The upper limit for cumulative relaxation is set; the penalty term of the relaxation variable is added to the objective function in a quadratic form, and the penalty coefficient is greater than the sum of all other weight coefficients in the objective function, ensuring that relaxation is only enabled when there is a constraint conflict.

[0024] Preferably, after the step of coordinating the adjustment of each adjustable execution component according to the dynamic matching control command sequence, the method further includes an online adaptive model update step:

[0025] After each sampling period, calculate the prediction residual at the current time. ;

[0026] When continuous The absolute value of the residual for each sampling period exceeds the adaptive threshold. When this occurs, local parameter fine-tuning is triggered; the adaptive threshold... The formula is dynamically adjusted based on the sliding standard deviation of recent residuals:

[0027]

[0028] in This is the sensitivity coefficient. The length of the sliding window. To prevent zero constants, the local parameter fine-tuning only updates the parameters of the last two layers of the dynamic load prediction model, freezes the feature extraction parameters of the previous layers, and uses the most recent samples for fine-tuning. The input / output pairs for each sampling period, with a fine-tuning iteration count not exceeding [number missing]. Second-rate.

[0029] A dynamic load matching and control system for waste heat recovery systems includes a data acquisition module, a dual-channel prediction module, a rolling optimization module, a collaborative execution module, and an adaptive update module, specifically:

[0030] The data acquisition module collects heat generation status parameters from the waste heat source side, demand status parameters from the heat load side, and operating status parameters of the waste heat recovery system in real time.

[0031] The dual-channel prediction module obtains the waste heat supply prediction curve and the heat demand prediction curve within a preset time window through the source-side heat generation dynamic sub-model and the load-side heat demand dynamic sub-model, respectively; wherein, the source-side heat generation dynamic sub-model is a physical information neural network, and the load-side heat demand dynamic sub-model is a multi-scale temporal convolutional attention network.

[0032] The rolling optimization module takes the optimal overall energy efficiency of the system as the objective function and combines the constraints to generate a dynamic matching control command sequence for each adjustable execution component.

[0033] The collaborative execution module coordinates the adjustment of each adjustable execution component according to the dynamic matching control command sequence;

[0034] The adaptive update module dynamically adjusts the trigger threshold based on the prediction residual and performs local parameter fine-tuning on the prediction model when the trigger condition is met.

[0035] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a dynamic load matching and control method for a waste heat recovery system as described in any one of the claims.

[0036] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any one of the methods for dynamic load matching and control of a waste heat recovery system.

[0037] Compared with the prior art, the technical solution of this application has the following technical effects:

[0038] This invention employs a dual-channel decoupled prediction architecture to construct source-side and load-side sub-models. Both sub-models utilize network structures adapted to their respective operating conditions to independently complete training and inference, with joint constraint solving only occurring during the optimization phase. The source-side model incorporates energy conservation constraints based on a physical information neural network, aligning with the objective physical laws governing waste heat generation. The load-side model leverages a multi-scale temporal convolutional attention network to capture multi-dimensional load change characteristics, continuously improving the accuracy of future supply-demand curve predictions and providing a reliable computational foundation for subsequent coordinated regulation.

[0039] This invention constructs a system-wide optimal energy efficiency objective function coupled with multiple penalty terms. During the optimization calculation, it considers the total amount of waste heat recovery, overall equipment power consumption, and supply-demand load deviation, while introducing terminal costs to ensure stable closed-loop operation of the control system. The objective function sets differentiated weights for supply-demand deviations at different times and normalizes various operating condition deviations based on Mahalanobis distance, fully covering the core optimization dimensions in the control process. This ensures that the resulting control commands align with the overall requirements for long-term energy-saving operation of the system.

[0040] This invention incorporates relaxation variables in the objective function solution process to handle various hard constraints on equipment operation. By setting the range and cumulative upper limit of relaxation variable values, it avoids situations where optimization is unsolvable, and only activates relaxation adjustment when there are conflicting operating conditions. Relaxation-related penalty terms have higher weights, preventing the relaxation mechanism from arbitrarily interfering with normal control processes, ensuring that the operating parameters of each execution component are always within the equipment's safe range, and improving the feasibility and adaptability of the multi-variable collaborative optimization solution process.

[0041] This invention dynamically generates and adjusts the trigger threshold by predicting residuals in real time, only performing local fine-tuning of network parameters outside the shallow layers of the prediction model when the residuals continuously exceed the limit. This method eliminates the need to completely retrain the entire network, reducing the computational overhead of parameter adjustments, continuously correcting prediction biases generated by the model over long-term operation, maintaining the accuracy of the source-load prediction curve, and ensuring the continuous and stable effect of dynamic load matching and regulation throughout the entire cycle.

[0042] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0043] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0045] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0046] Figure 1 A schematic diagram of the complete closed-loop timing execution process of a dynamic load matching and control method for a waste heat recovery system;

[0047] Figure 2 The overall architecture diagram for input processing and joint optimization of the dual-channel decoupled parallel load forecasting model;

[0048] Figure 3 The fitting comparison diagram shows the fitting effect of four existing models with the source-load two-sided thermal power time series prediction of this application;

[0049] Figure 4 A long-term convergence trend curve of various operational indicators for a 30-day full-process closed-loop simulation.

[0050] Figure 5 A diagram illustrating the effect of dynamic load tracking and control of waste heat supply and demand and heat storage charging and discharging on a single day throughout the entire time sequence.

[0051] Figure 6 This is a closed-loop diagram of the data flow of the five functional modules of the waste heat recovery dynamic load matching and control system. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0053] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0054] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0055] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0056] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0057] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0058] Example 1 mainly describes a dynamic load matching and control method for a waste heat recovery system, such as... Figure 1 As shown, it includes the following steps:

[0059] Real-time collection of heat generation status parameters from the waste heat source side, demand status parameters from the heat load side, and operating status parameters of the waste heat recovery system;

[0060] Based on the heat production state parameters and the demand state parameters, the waste heat supply prediction curve and the heat demand prediction curve within a preset time window are obtained through a pre-set dynamic load prediction model.

[0061] Based on the waste heat supply prediction curve, the heat demand prediction curve, and the operating status parameters, the dynamic matching control command sequence of each adjustable actuator in the waste heat recovery system is generated with the optimal overall system energy efficiency as the objective function; wherein, the objective function is coupled with the heat recovery amount, system energy consumption, and supply-demand deviation penalty term;

[0062] The adjustable execution components are coordinated and adjusted according to the dynamic matching control command sequence, so that the actual recovery load of the waste heat recovery system dynamically tracks the heat demand prediction curve and responds to the fluctuations of the waste heat supply prediction curve.

[0063] Furthermore, the collected parameters are divided into three independent yet coupled parameter sets. The first set is the set of heat generation status parameters on the waste heat source side. This set includes a continuous sampling sequence of the operating frequency of the waste heat source driving equipment, a time series array of the real-time volumetric flow rate of the heat exchange medium, a time series array of the inlet temperature of the heat exchanger medium, a time series array of the outlet temperature of the heat exchanger medium, and a continuous time series vector of historical heat generation power. All parameters are synchronously collected according to a uniform sampling step size, and all physical quantities are uniformly converted to international standard units and then encapsulated into a source-side input feature matrix. The matrix dimension is determined by the sampling window length and the number of feature types. Each row in the matrix corresponds to all source-side feature values ​​at a single sampling time.

[0064] The second category is the set of demand status parameters on the heat load side. This set includes the real-time opening time sequence of all heat terminal regulating valves, the time sequence array of heating return water temperature for each branch, the continuous sampling sequence of ambient dry-bulb temperature, and the time sequence vector of historical actual heat demand power. All parameters undergo synchronous dimensional conversion and are encapsulated into a load-side input feature matrix. The matrix row and column dimensions are consistent with the source side feature matrix, ensuring that the time axes of the temporal features on both sides are fully aligned, thus eliminating prediction calculation errors caused by temporal misalignment.

[0065] The third category is the set of global operating status parameters for the waste heat recovery system. This set includes the real-time charging and discharging power of the heat storage device, the current temperature range of the heat storage medium, the operating power consumption of the circulating water pump, the start / stop status of the auxiliary heating equipment, the upper and lower limits of the current opening of each branch valve, and the upper and lower limits of the allowable medium flow rate of the heat exchange equipment. All parameters in this category are used as the constraint boundary conditions input for subsequent rolling optimization solutions, and a separate constraint parameter vector is constructed. Each element in the vector corresponds to a hard boundary indicator for device operation.

[0066] After the three sets of parameters are collected, they are uniformly processed by outlier filtering, missing value linear interpolation to complete, and time series alignment preprocessing. The three sets of parameter vectors after preprocessing are completely input into the dual-channel decoupled dynamic load prediction model without deleting any collected dimension parameters, ensuring that the model input feature dimensions are complete and without missing values.

[0067] like Figure 2 As shown, the dynamic load prediction model adopts a fully decoupled dual-channel parallel architecture. Channel 1 is the source-side heat generation dynamic sub-model, and Channel 2 is the load-side heat demand dynamic sub-model. The hardware inference processes of the two sub-models are independent of each other. The training dataset, network weight parameters, and loss calculation equations are completely separated. The numerical correlation is only established through the global coupling constraint matrix during the rolling optimization solution stage. There is no parameter sharing or loss function merging calculation during the training stage.

[0068] The source-side model uses a physical information neural network, which has a three-layer structure: the first layer is the feature-normalized input layer, the second layer is a multi-layer fully connected hidden layer, and the third layer is the heat generation power output layer. All learnable parameters of the network are uniformly collected into a parameter set. , This represents the total number of all weights and bias parameters in the network.

[0069] The formula for the total loss function during the model training phase is: ;

[0070] The loss function is split into three independent terms. The first term is the data-driven fitting error term, and the complete expanded calculation form is as follows:

[0071]

[0072] Within the formula This represents the total number of training samples in a single batch. For the network Each sample outputs a predicted heat production scalar. To correspond to the actual measured heat generation power at the sampling time, this item quantifies the deviation between the predicted value and the actual measured value through squaring. The arithmetic mean of the errors of all samples within the batch is taken as the value of this sub-item loss.

[0073] The second term is the physical conservation residual constraint term, which fully expands the internal energy conservation calculation formula:

[0074]

[0075] Within the formula Represents a constant density scalar value for the heat exchange medium. Represents the scalar value of the specific heat capacity at constant pressure of the medium. Representing the Real-time volumetric flow rate of the medium at the time of sample collection. The inlet medium temperature of the heat exchanger. The outlet medium temperature of the heat exchanger. By pre-setting physical constraint weight constants, this term forces the network output heat generation power to strictly match the medium heat transfer energy conservation equation, eliminating prediction results from purely data-driven models that violate fundamental thermodynamic laws.

[0076] The third term is the gradient regularization term for suppressing overfitting, fully expanded to the Frobenius norm calculation form:

[0077]

[0078] Within the formula The gradient matrix representing the predicted heat production power is obtained with respect to all network parameters. The square root of the sum of the squares of all elements within the matrix yields the Frobenius norm. This is the gradient penalty weight constant. This term constrains the range of gradient values ​​for network parameters, preventing drastic oscillations in network parameters under sparse working conditions from causing distortions in prediction results.

[0079] The entire network training process uses a gradient descent optimizer to update the parameter set via backpropagation. In each iteration, the three losses mentioned above are calculated synchronously and summed with weights. All network weights are adjusted synchronously based on the gradient of the total loss. After training, the network parameters are fixed, and the complete feature matrix from the source side is input during the inference phase. It outputs discrete time-series predicted values ​​within a preset future time window at each time step, and splices all discrete predicted values ​​in chronological order to generate a continuous waste heat supply prediction curve. .

[0080] The load submodel employs a composite network structure combining multi-scale dilated causal temporal convolution with channel attention modules. The network performs forward propagation computation layer by layer. Complete two-layer feature output calculation formula:

[0081]

[0082]

[0083] The first layer convolutional feature calculation formula is completely decomposed and expanded. The weighted summation of the convolution is expanded as follows:

[0084]

[0085] In the formula Represents the total length of a single-level dilated convolution kernel. Representing the Layer convolution Each positional weight scalar This represents the temporal delay value corresponding to the convolution stride. For the upper layer network in Output feature vector at each time step. The summation result of the summation of the global bias constant vector for this layer is input into the activation function. Complete the nonlinear mapping.

[0086] The element-wise multiplication operation of the tensor is used to multiply the convolution output feature vector with the weight vector output by the channel attention module position by position, thereby achieving adaptive adjustment of the strength of the feature channels.

[0087] The channel attention module performs step-by-step computation, with the first step being global average pooling for complete computation:

[0088]

[0089] For characteristic time series length, The pooling process outputs a one-dimensional global feature vector, representing the total number of feature channels. The second step involves a linear transformation of the first fully connected layer.

[0090]

[0091] The weight matrix of the fully connected layer is used to reduce the dimension, and the output vector is input to the ReLU activation function to perform nonlinear filtering; the second fully connected layer performs dimension upscaling transformation:

[0092]

[0093] To increase the dimension of the fully connected weight matrix, the output vector is input to the Sigmoid activation function. The output is a channel weight vector with values ​​ranging from zero to one, which is the final output of the SE module.

[0094] Multi-layer network stacking enables multi-scale temporal feature extraction with different inflation factors. Capturing instantaneous load fluctuations on a short timescale and inertial load changes during a long production cycle on a long timescale, the network inference stage inputs a complete feature matrix from the load side. It outputs discrete predicted values ​​of future heat demand power in the future time domain at each time step, and splices all discrete time series points to generate a continuous heat demand prediction curve. .

[0095] The joint coupling constraints of the dual-channel model are as follows:

[0096] The inference operations of the source sub-model and the load sub-model are completely parallel and independent, each generating a complete-length prediction time-domain curve. After both sets of curves are output to the rolling optimization calculation unit, a coupling equation constraint matrix is ​​added to establish a bidirectional numerical correlation. The complete mathematical expression of the coupling constraint is as follows:

[0097]

[0098]

[0099] In the formula To determine the actual heat power recovered by the system after regulation, To maximize the utilization of waste heat, To assist the output of heat power of the supplementary heating equipment, this coupling constraint associates the independent prediction curves on both sides into a unified optimization variable, which serves as the boundary condition for the rolling optimization solution, enabling the dual-channel prediction results to participate in the global load matching calculation in a coordinated manner.

[0100] Furthermore, the rolling optimization unit receives the waste heat supply forecast curve, the heat demand forecast curve, and the system operation constraint parameter vector, and performs a constrained solution with the goal of minimizing the overall operating cost. The optimization variables are unified as the control increment sequence over the entire forecast time domain. The complete objective function formula: ;

[0101] Complete decomposition and expansion formula of objective function

[0102] The first term is the cost component of time-varying weighted supply and demand load deviation, calculated using the complete Mahalanobis distance expansion formula:

[0103]

[0104] The complete formula for solving the covariance matrix of historical supply and demand deviations is as follows: , The total duration of the historical sample. For the first Historical moment supply-demand load difference vector The vector of all historical differences is obtained by inverting the covariance matrix. To achieve deviation normalization and quantification under different fluctuation amplitude operating conditions; The time-varying penalty coefficient sequence satisfies the monotonically decreasing constraint:

[0105]

[0106] The closer to the current sampling time, the higher the load deviation penalty weight value, prioritizing the accuracy of the recent time-domain load matching.

[0107] The second item is the total system operating power consumption cost, which is fully explained below: , To predict the total power consumption of the entire system at every moment in the time domain, including the power consumption of the circulating pump, regulating valve drive mechanism, and auxiliary heating equipment, By fixing the power consumption penalty weight constant, the optimization process is continuously constrained to reduce the overall power consumption of the system.

[0108] The third term is the component constrained by the control increment stationarity, with the weighted norm fully expanded: , To control the incremental positive definite weighted matrix, the diagonal elements of the matrix correspond to the adjustment amplitude constraint weights of different actuators. To control the fluctuation penalty constant, suppress large jumps in control commands within a short period of time, and avoid frequent large-scale actions of the actuator.

[0109] The fourth term is the component representing the predicted time-domain terminal stability cost, expressed in full by the Lyapunov approximation terminal function:

[0110]

[0111] To obtain the positive definite constant matrix from solving the algebraic Lyapunov equations offline, the fundamental Lyapunov equations are:

[0112]

[0113] Here is the linear state transition matrix of the waste heat recovery system. It is a positive definite weight matrix. The weight constant for scaling the terminal cost is used to ensure the global asymptotic stability of the rolling optimization iteration closed-loop system and eliminate long-distance time-domain control oscillations.

[0114] Complete merged form of the overall objective function:

[0115]

[0116] The solution process is as follows To optimize the variables, find the value that maximizes the total cost within all constraint boundaries. The control increment sequence with the smallest numerical value.

[0117] During the optimization process, the original equipment operation inequality hard constraints pose a risk of an empty feasible region. Non-negative relaxation variables are introduced to transform the constraints. The basic transformation formula is as follows:

[0118]

[0119]

[0120]

[0121] This represents the total number of hard constraints for the operation of all equipment. For the first Original constraint mapping function, For corresponding constraints The relaxation scalar of time. This represents the upper limit of the total amount of relaxation allowed for this constraint within a single prediction time domain.

[0122] Formula for a quadratic function with a slack variable penalty term: , The relaxation penalty weight constant is numerically greater than the value inside the objective function. The sum of all weight values, plus the relaxation penalty term, is added to the overall objective function to form a fault-tolerant complete optimization objective:

[0123]

[0124] The relaxation variable is greater than zero only when there is no feasible solution to the original hard constraint. Under normal operating conditions, the relaxation variable is always equal to zero and will not change the original optimization control logic. At the same time, it ensures that there is a feasible control increment solution under any operating condition, thus eliminating the problem of control interruption caused by no solution to the optimization.

[0125] Furthermore, after the rolling optimization solution is completed, a complete time-series control increment sequence is output. The current baseline control value of the actuator is superimposed to generate a dynamic matching control command sequence. The instruction sequence is sent sequentially to all adjustable execution components of the system according to the sampling time, including the waste heat recovery heat exchange regulating valve, the circulating water pump frequency converter, the heat storage device charging and discharging regulating mechanism, and the auxiliary heat replenishment equipment power regulating unit. All execution components adjust their operating conditions synchronously according to the instructions at the corresponding time. After the adjustment is completed, a new round of system-wide parameters are collected, and the model adaptive update calculation process is entered.

[0126] The adaptive update fully expands all mathematical calculation rules, and the basic formula for calculating the prediction residual of a single sampling period is as follows:

[0127]

[0128] for Real-time on-site measurement of actual heat load values. This represents the one-step forward prediction output of the model for the current load in the previous sampling period, and the absolute value of the residual. As a quantitative indicator of model prediction bias.

[0129] Fully expanded formula for adaptive dynamic threshold calculation: , The residual sliding statistical window has a fixed length, and continuous data is stored within the window. For each historical sampling period of residual values, first calculate the arithmetic mean of the squared residuals within the window, and then take the square root to obtain the moving standard deviation of the residuals; This is the threshold sensitivity proportionality constant. To prevent the threshold from returning to zero when the residual standard deviation is zero, a very small positive constant is used to prevent the threshold from returning to zero. The threshold is updated synchronously after each sampling cycle is completed. .

[0130] Complete logic formula for triggering local parameter fine-tuning:

[0131]

[0132] Only when consecutive When the absolute value of the residuals from all adjacent sampling periods is greater than the dynamic threshold at the corresponding time, the local parameter fine-tuning process of the network is initiated. During the fine-tuning execution phase, all shallow feature extraction network parameters of the dual-channel sub-model are frozen, and only the output layer weights and bias parameters of the last two layers of the network are allowed to participate in the iterative update. The training dataset is truncated to the most recent data. The input and output time-series samples are taken over each sampling period, and a maximum iteration count is set. The loss function in the iterative process follows the original training loss formula of the model. Fine-tuning is terminated when the number of iterations reaches the upper limit or the local loss converges. After fine-tuning is completed, the model stores the weight parameters and enters the next round of global data collection cycle, continuously iterating and optimizing the output accuracy of the dual-channel prediction model in a closed loop.

[0133] This embodiment accurately extrapolates supply and demand loads through a dual-channel decoupled prediction architecture, integrates physical constraints and multi-scale attention to improve prediction accuracy, and achieves optimal energy efficiency scheduling through multi-objective rolling optimization with relaxation variables. It also continuously reduces load deviations through residual-driven model adaptive correction, thereby reducing waste heat release and heat replenishment consumption throughout the process and stably adapting to various fluctuating industrial waste heat conditions.

[0134] Example 2: To fully verify the feasibility, prediction accuracy, and global control optimization performance of the dynamic load matching and control method for waste heat recovery systems proposed in this application, a unified simulation verification environment was built based on two publicly available and traceable industrial heat load time-series datasets. One dataset is the "Time-Series Dataset of Waste Heat Utilization Rate in China's Petroleum Refineries" published by the National Basic Science Public Science Data Center. This dataset has a sampling interval of 15 minutes and a continuous collection period of 92 days. It includes 12 time-series features: refinery heating furnace waste heat source medium flow rate, inlet and outlet temperatures, equipment operating frequency, hourly heat generation power, process heat terminal return water temperature, ambient temperature, and hourly required heat power. The total effective sample size is 88,320 sets. The data was generated by AspenPlus process simulation coupled with on-site SCADA measurement calibration and possesses complete two-sided real fluctuation characteristics of industrial waste heat source load. The second dataset is the open-source Hourly Thermal Load Profiles for US Manufacturing Subsectors from the Geothermal Data Repository. The industrial heat load dataset for the manufacturing sector, with a sampling interval of 1 hour, covers the heat generation and consumption time series data of 12 types of manufacturing processes, with an effective sample size of 141,960 sets. This dataset is used to expand the sparse sample range under varying operating conditions. All data underwent standardized preprocessing, including linear interpolation for missing values, outlier removal using the 3σ criterion, and feature standardization. The dataset is divided into training, testing, and validation sets in a 7:2:1 ratio, and a unified prediction time domain is used. The sampling period step was uniformly aligned to 15 minutes. The simulation hardware was equipped with an NVIDIA RTX A6000 graphics card and 64GB of RAM. The simulation software adopted a hybrid programming environment of MATLAB R2024b and PyTorch 2.1. The rolling optimization solver was the Gurobi 10.0 linear programming solver. The hyperparameter comparison benchmark was fixed throughout the process to eliminate experimental bias caused by hardware, software and dataset partitioning. A two-level progressive simulation was set up for verification.

[0135] This embodiment conducts independent simulation tests on the source-side heat generation prediction sub-model and the load-side heat demand prediction sub-model, without connecting to the back-end rolling optimization and control link. The core objective is to quantitatively verify the differences in prediction error, generalization ability, and sparse condition stability of the dual-channel PINN source-side sub-model and multi-scale temporal convolutional attention load-side sub-model of this application compared with existing mainstream temporal prediction networks. Four widely used existing benchmark models in the current field of industrial heat load prediction are selected for horizontal comparison with the same dataset, the same training rounds, and the same parameter scale. The benchmark models are a single-layer fully connected deep neural network (DNN), a two-layer gated long short-term memory network (LSTM), a basic dilated temporal convolutional network (TCN), and a single-channel fusion attention temporal network (TC-Att). The input feature dimensions, total number of network parameters, and training iteration rounds of all benchmark models are specified. Batch sample size The early stopping threshold is completely unified. The training loss function of the source sub-model strictly adopts the weighted loss formula containing the energy conservation physical residual mentioned above. The load sub-model is fully constructed with multi-scale dilated causal convolution combined with SE channel attention hierarchical structure. The validation set data is frozen during the training phase and does not participate in gradient backpropagation. Three types of industry-standard prediction error quantification indicators, MAE (mean absolute error) and RMSE (root mean square error), are selected. (Goodness of fit) completes the performance quantification, and the index calculation formula is as follows: , ,in This represents the total number of samples in the test set. To predict thermal power for the model, The dataset contains actual measured thermal power. To test the actual load mean, three sets of indicators were independently calculated for source-side heat generation prediction and load-side heat demand prediction, respectively. Table 1, a quantitative comparison table of model prediction accuracy, was generated. In the table, "Model" represents the model type: Src-MAE, Src-RMSE, Src- These correspond to three types of indicators for predicting heat generation on the source side: Dem-MAE, Dem-RMSE, and Dem- These correspond to three categories of indicators for load-side heat demand forecasting, with the unit uniformly set in MW. The data is dimensionless, and all data in the tables are taken from the average statistical results of all samples in the test set.

[0136] Table 1. Quantitative Comparison of Source-Load Two-Sided Prediction Accuracy of Different Time-Series Prediction Models

[0137]

[0138] After completing the statistical analysis of the indicator values, based on the complete time-series data of 192 consecutive sampling times in the test set, as follows: Figure 3 As shown, Figure 3(a) Showing the fitting effect of heat generation power on the waste heat source side, Figure 3 (b) The fitting effect of the heat load side heat demand power is shown. The two sub-plotting areas share a unified horizontal coordinate axis, which represents the time series sampling number and is set to a value range of 1 to 192. The vertical coordinate axis matches the load power range on both sides respectively. The vertical coordinate value range of the source side is 0MW to 12MW, and the vertical coordinate value range of the load side is 0MW to 9MW. Six sets of time series curves are plotted simultaneously in the plotting area, including the actual load baseline curve measured on site, the DNN prediction curve, the LSTM prediction curve, the TCN prediction curve, the TC-Att prediction curve, and the prediction curve of the dual-channel sub-model of this application. They are distinguished by different line styles and colors. The actual load baseline curve is marked with a black thick dashed line, all four types of traditional comparison models are marked with light-colored thin dashed lines, and the output curve of the dual-channel model of this application is marked with a dark blue thick solid line. The entire time series includes four typical operating condition intervals: a steady-state high-load interval, a low-load sparse interval caused by the shutdown of heat source equipment, a load step increase interval caused by production process switching, and a load drop interval caused by the simultaneous shutdown of multiple terminals. In the low-load sparse interval and the load change interval, the curves of the four existing models, namely DNN, LSTM, TCN, and TC-Att, all show visual characteristics of significant deviation from the true value and large curve oscillations. The curve of the model in this application closely follows the true load benchmark curve throughout the entire process, without significant deviation or violent jitter, intuitively demonstrating the effect of physical information constraints and multi-scale attention structure on improving prediction accuracy. The simulation process extracts the statistical error of the sparse operating condition interval where the load is 20% below the rated value. The Src-RMSE of the DNN sparse interval increases to 1.342MW, the Src-RMSE of the LSTM sparse interval increases to 0.975MW, and the Src-RMSE of the TCN and TC-Att sparse intervals increases to 0.783MW and 0.619MW, respectively. The Src-RMSE of the PINN source-side sub-model in this application is only 0.241MW. It relies on the energy conservation loss term to suppress the prediction drift under sparse data. The load-side multi-scale convolutional attention network captures the long-period load inertia based on multi-expansion factor convolution. The channel attention adaptively amplifies the weights of load change-related features, continuously reducing the time-series prediction deviation of the demand side. It outputs the training convergence curve of the dual-channel sub-model, the time-series error data of the test set, and the sparse operating condition error statistical report. It confirms that the prediction model of this application has stable and higher prediction accuracy under steady-state and sparse variable operating conditions, and can provide a precise time-series input basis for back-end rolling optimization and control.

[0139] The validated dual-channel PINN-TCAN prediction sub-model was reused as the prediction input unit for global control and connected to the control link. It includes a global multi-source parameter acquisition and preprocessing module, a dual-channel prediction inference unit, a multi-objective rolling optimization solution unit with relaxed variable fault-tolerant constraints, an actuator collaborative control output unit, and an online adaptive update unit for the prediction model. A complete closed-loop simulation system was built, with a simulation cycle of 2880 consecutive sampling periods, each sampling period lasting 15 minutes, for a total simulation duration of 30 days. The time-series parameters from two publicly available datasets were imported synchronously and updated hourly to update the simulation system input. A complete dynamic mathematical model of the waste heat recovery system equipment was built within the simulation, covering four types of adjustable actuators: waste heat exchangers, variable frequency circulating water pumps, thermal storage charging and discharging tanks, and gas-fired auxiliary heating units. Sixteen hard constraint boundaries were set for the equipment, corresponding to the constraint functions mentioned earlier. Each constraint is configured with an independent slack variable. With the upper limit of cumulative relaxation Relaxation penalty coefficient Setting it to 1200 is greater than the target function. The sum of all weighted coefficients, the rolling optimization objective function fully superimposed with the Mahalanobis distance component of supply and demand deviation, the system power consumption component, the control increment smoothing component, the Lyapunov terminal stability cost component, and the time-varying deviation penalty coefficient. Values ​​are assigned according to the linear monotonically decreasing rule for prediction in the time domain. The control time domain is set to 8, and the solver single-cycle optimization convergence threshold is set. Adaptive update module parameters uniformly configure sliding window length Length of continuous residual exceeding the standard Sensitivity coefficient , prevent zero constant Fine-tuning the sample length Maximum number of iterations in a single round of fine-tuning Gradient fine-tuning was performed only on the last two layers of the dual-channel sub-model, and the weights of the shallow feature extraction network were frozen. Throughout the process, six core operational indicators were recorded synchronously for each sampling period: system supply and demand load difference, total waste heat recovery, total system power consumption, auxiliary heating gas consumption, slack variable activation frequency, and mean prediction residual. The average values ​​were then summarized according to daily duration to generate Table 2, a quantitative statistical table of the entire process control performance. In the table, Day represents the number of consecutive days of simulation operation. Represents the absolute value of the daily average supply-demand load deviation, in MW. Represents the total daily waste heat recovery, in MWh. This represents the average daily system power consumption, in kW. Represents the daily auxiliary heating gas consumption, in units , This represents the total number of constraint conflicts triggered by slack variables on a single day. The value represents the average one-step forward prediction residual for a single day, in MW. All values ​​are the arithmetic mean of 96 sampling periods within 24 hours of the day. Data from the simulation on days 5, 10, 15, 20, 25, and 30 are continuously extracted to form a complete time-series statistical sample.

[0140] Table 2. Daily Performance Quantitative Statistics of the Full-Process Closed-Loop Control Simulation

[0141]

[0142] Based on daily statistical data from a 30-day complete simulation, such as Figure 4 As shown, five independent variation curves are plotted within the region, each corresponding to the daily average supply and demand load deviation. Total daily waste heat recovery Average daily power consumption of the system Daily gas consumption Daily average forecast residual The daily number of slack variable triggers is displayed using light gray bar charts. The entire convergence pattern can be intuitively observed from the trend of the curve. As the number of simulation days continues to increase, the online adaptive update module continuously corrects the weight parameters of the prediction model. The average prediction residual curve shows a continuous and stable downward trend. The average deviation between supply and demand loads gradually narrows in sync. The total daily waste heat recovery of the system steadily increases and slowly tends to stabilize. The average operating power consumption and auxiliary heating gas consumption of the system continue to decrease in sync. The number of equipment constraint conflict triggers continues to decrease. All indicators show no oscillation or rebound phenomenon, which intuitively verifies the stable optimization effect of the adaptive fine-tuning mechanism and the multi-objective optimization objective function under long-term operation.

[0143] Extract the complete time-series data of 96 sampling periods for day 30 of the simulation, such as... Figure 5As shown, the time-series curves for waste heat supply forecast, heat demand forecast, and actual recovered load after system regulation are displayed. Light blue blocks mark the power range during the heat storage and charging period, and light orange blocks mark the power range during the heat storage and releasing period. The daily time-series includes three typical periods of drastic load fluctuation. In the first period, the waste heat source's heat production power increases by 3.6MW in a single step, supplying heat exceeding the total heat demand of all terminals. The regulation system prioritizes activating the heat storage units to store the excess heat, and the actual recovered load consistently matches the demand forecast curve, with no excess waste heat being released. In the second period, all heat-consuming terminals simultaneously reduce production, resulting in a sharp drop in demand of 2.9MW. The heat storage units continuously absorb excess heat, preventing the system from releasing excess heat. In the third period, the waste heat source equipment operates at low load, with the supply power consistently lower than the terminal heat demand. The heat storage units continuously release stored heat to fill the load gap, only slightly activating auxiliary heating equipment after the heat storage and releasing power reaches the equipment's upper limit. Throughout the entire time series, the actual recovered load curve follows the changes in the heat demand forecast curve. Relying on the thermal storage unit to dynamically smooth the fluctuations in the waste heat supply on the source side, there will be no significant deviation from the demand curve. This fully demonstrates the dynamic load matching capability of this application with rolling optimization and relaxed fault-tolerant control logic.

[0144] Example 3 describes in detail a dynamic load matching and control system for a waste heat recovery system, such as... Figure 6 As shown, it includes a data acquisition module, a dual-channel prediction module, a rolling optimization module, a collaborative execution module, and an adaptive update module, specifically:

[0145] The data acquisition module collects heat generation status parameters from the waste heat source side, demand status parameters from the heat load side, and operating status parameters of the waste heat recovery system in real time.

[0146] The dual-channel prediction module obtains the waste heat supply prediction curve and the heat demand prediction curve within a preset time window through the source-side heat generation dynamic sub-model and the load-side heat demand dynamic sub-model, respectively; the source-side heat generation dynamic sub-model is a physical information neural network, and the load-side heat demand dynamic sub-model is a multi-scale temporal convolutional attention network.

[0147] The rolling optimization module takes the optimal overall energy efficiency of the system as the objective function and combines the constraints to generate a dynamic matching control command sequence for each adjustable execution component.

[0148] The collaborative execution module adjusts each adjustable execution component in a coordinated manner according to the dynamic matching control command sequence; the adaptive update module dynamically adjusts the trigger threshold according to the prediction residual, and performs local parameter fine-tuning on the prediction model when the trigger condition is met.

[0149] The data acquisition module, as the front-end data input unit, integrates multiple analog signal acquisition channels, pulse signal acquisition channels, and an industrial bus communication interface at the hardware level. It can connect to field sensing devices such as temperature transmitters, electromagnetic flowmeters, frequency acquisition devices, valve opening feedback sensors, ambient temperature and humidity sensors, heat storage tank temperature sensors, and water pump power transmitters. Internally, the module has a built-in preprocessing subroutine that receives three completely independent sets of acquisition parameters. The first set is the heat generation status parameters on the waste heat source side, including the operating frequency of the waste heat source equipment, real-time volumetric flow rate of the heat exchange medium, heat exchanger inlet temperature, heat exchanger outlet temperature, and historical heat generation power time series. The second set is the demand status parameters on the heat load side, covering the real-time opening of regulating valves at each heat-consuming terminal, return water temperature of the heating branch, ambient temperature, and historical heat demand power time series. The third set is the global operating status parameters of the waste heat recovery system, including the upper and lower limits of the heat storage device's charging and discharging power, the rated power range of the circulating water pump, the power adjustment range of the auxiliary heating unit, the opening boundaries of valves in each branch, and the medium flow constraint threshold of the heat exchange equipment. The module performs unified standardization processing, including 3σ criterion outlier removal, linear interpolation to fill missing time series data, unified conversion of all physical quantities to international standard units, and time series length alignment. After preprocessing, three sets of structured feature vectors are generated: source-side feature vector, load-side feature vector, and system constraint parameter vector. These are transmitted in real time to the back-end dual-channel prediction module via Ethernet. At the same time, all the original acquisition data of the most recent W sampling periods are stored in the local cache window, providing a historical residual calculation data source for the subsequent adaptive update module.

[0150] The dual-channel prediction module is the core computing unit for the system's time-series inference. It utilizes an independent GPU chip for parallel inference, and internally comprises two completely decoupled independent inference channels. The hardware computing resources and storage cache of these two channels are independent, with data only transmitted to the rolling optimization module after the output prediction curve. There is no parameter sharing or loss function merging during the training phase. The first channel incorporates a dynamic sub-model of heat generation on the source side. The model's underlying layer is a physical information neural network structure, including a standardized input layer, multiple fully connected hidden layers, and a heat generation power output layer. The module's storage unit stores the complete set of learnable parameters θ for this sub-model and includes its own loss function calculation logic. This loss function integrates data-driven error terms, energy conservation physical residual terms, and gradient regularization overfitting suppression terms, fully reusing the previously mentioned mathematical formulas. This channel receives the source-side feature vector output from the data acquisition module, performs forward inference calculations through the network, and outputs discrete heat generation power values ​​within a preset future time window, which are then stitched together to form a complete waste heat supply prediction curve. The second channel incorporates a dynamic sub-model of load-side heat demand, employing a multi-scale dilated causal temporal convolutional network architecture superimposed with SE channel attention. The module's memory stores convolutional kernel weights, fully connected layer parameters of the attention module, and bias constants for each layer, along with a complete set of formulas for convolutional feature extraction and channel weight calculation. This channel receives load-side feature vectors, performs parallel inference, and outputs discrete values ​​of all future time-domain heat demand power, which are then concatenated to generate a continuous heat demand prediction curve. The dual-channel inference operations are performed simultaneously, and the two generated time-series prediction curves are simultaneously pushed to the rolling optimization module as the input for optimization calculations.

[0151] The rolling optimization module is the core unit for solving the multi-objective optimal control of the system. It has a built-in commercial linear programming solver package, and it has solidified the complete system comprehensive energy efficiency optimal objective function and the complete set of mathematical formulas for hard constraint relaxation transformation. The module receives the waste heat supply prediction curve and heat demand prediction curve output by the dual-channel prediction module, and synchronously reads the system operation constraint parameter vector transmitted by the data acquisition module. The module first constructs a set of original equipment operation inequality constraints, then introduces non-negative relaxation variables to complete the constraint transformation, sets an upper limit for the accumulation of relaxation variables, and superimposes the quadratic terms of relaxation penalties onto the original objective function to form a complete optimization objective with fault tolerance. The optimization variables are defined as the control increment sequence of all adjustable components in the entire prediction time domain. The module has built-in logic for the monotonically decreasing time-varying deviation penalty coefficient assignment, an offline solution program for the Mahalanobis distance covariance matrix, and logic for calculating the Lyapunov terminal cost function. The solver iterates with the goal of minimizing the total comprehensive operating cost until the convergence threshold is met, and then outputs a complete time-series dynamic matching control command sequence. The command sequence contains the adjustment increment values ​​corresponding to the waste heat exchange regulating valve, variable frequency circulating water pump, heat storage charging and discharging mechanism, and auxiliary heating equipment at each sampling time. After the solution is completed, the entire set of time-series commands is sent to the collaborative execution module, and the measured load time-series data at the current time is output to the adaptive update module.

[0152] The collaborative execution module, serving as the system control command output execution unit, integrates multiple inverter drive interfaces, electric regulating valve control loops, thermal storage charging and discharging actuator drive ports, and gas-fired supplementary heating unit power regulation communication interfaces, allowing direct connection to all adjustable actuators on-site. The module incorporates a command timing parsing subroutine, receiving the complete timing control command sequence output by the rolling optimization module. It then breaks down the single-step control increment according to the system's unified sampling period, superimposing the current baseline operating parameters of the equipment to obtain real-time adjustment commands. These commands are synchronously distributed to each type of actuator through different channels to complete the coordinated adjustment of operating conditions. Simultaneously, the module collects the opening degree, power, and flow feedback signals after each actuator's adjustment, transmitting them back to the front-end data acquisition module to form a closed-loop control system. This ensures that the equipment's operating status is entered into the system parameter pool in real-time after each adjustment action, providing the latest operating condition data support for the next round of prediction and optimization.

[0153] The adaptive update module is responsible for the online dynamic correction of the dual-channel prediction model. The module retrieves historical measured load data cached locally by the data acquisition module and one-step advanced predicted load data output by the dual-channel prediction module, and solves the predicted residual value for each sampling period according to the residual calculation formula. The module has built-in dynamic adaptive threshold calculation logic, reads all historical residual values ​​within a residual sliding window of length W, calculates the sliding standard deviation, and generates a real-time judgment threshold by combining the sensitivity coefficient and the zero constant. The module is internally configured with continuous exceedance judgment logic. When the absolute value of the residuals for L consecutive sampling periods exceeds the dynamic threshold at the corresponding time, the local parameter fine-tuning process of the model is automatically triggered. During the fine-tuning process, a parameter freeze command is issued to the dual-channel prediction module to lock all shallow feature extraction network parameters of the two types of sub-models. Only the weights and bias parameters of the last two output layers of the network are opened to participate in the iterative update. The fine-tuning training samples are extracted from the input and output time-series data of the most recent B sampling periods in the buffer. The maximum number of iterations is set to limit the computational overhead. After the iteration converges or the iteration limit is reached, the updated network weights are written back to the model storage unit of the dual-channel prediction module to complete the online self-correction of the prediction model. If there is no continuous residual exceedance, the fine-tuning process is not started. Only the residual update judgment threshold is continuously calculated. The entire process does not occupy the parallel inference computing power of the dual-channel prediction module and does not interfere with the normal load prediction and control process of the system.

[0154] This application also provides a computer device that stores a complete control program in its built-in memory. When the processor retrieves the program and runs it, it can sequentially execute all steps of parameter acquisition, load prediction, rolling optimization, coordinated control, and model adaptive update. Based on the hardware computing power, it completes the iterative solution of each model formula, stably outputs matching control instructions, and completely reproduces all the logic of the aforementioned waste heat load matching control method, ensuring the continuous operation of the industrial waste heat system in real-time dynamic control.

[0155] This application also provides a computer-readable storage medium that contains program code that can be read and executed by a processor. The code is fully equipped with all algorithm logics for dual-channel prediction, multi-objective optimization, and residual adaptive fine-tuning. Any hardware with processing capabilities can read the program in the medium and implement any control method flow, freeing it from the limitations of dedicated hardware and adapting it to deployment and use in various types of industrial control terminals.

[0156] This embodiment utilizes layered data acquisition, parallel prediction, optimization solution, equipment execution, and model self-correction functions. Data is transmitted synchronously between modules, enabling independent completion of the entire process of calculation and control without manual intervention. It maintains stable load matching effects during long-term operation and is suitable for multi-specification waste heat recovery site deployments.

[0157] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for dynamic load matching and control of a waste heat recovery system, characterized in that, Includes the following steps: Real-time collection of heat generation status parameters from the waste heat source side, demand status parameters from the heat load side, and operating status parameters of the waste heat recovery system; Based on the heat production state parameters and the demand state parameters, the waste heat supply prediction curve and the heat demand prediction curve within a preset time window are obtained through a pre-set dynamic load prediction model. Based on the waste heat supply prediction curve, the heat demand prediction curve, and the operating status parameters, the dynamic matching control command sequence of each adjustable actuator in the waste heat recovery system is generated with the optimal overall system energy efficiency as the objective function; wherein, the objective function is coupled with the heat recovery amount, system energy consumption, and supply-demand deviation penalty term; The adjustable execution components are coordinated and adjusted according to the dynamic matching control command sequence, so that the actual recovery load of the waste heat recovery system dynamically tracks the heat demand prediction curve and responds to the fluctuations of the waste heat supply prediction curve.

2. The dynamic load matching and control method for a waste heat recovery system according to claim 1, characterized in that, The dynamic load prediction model is a dual-channel decoupled prediction architecture, including a source-side heat generation dynamic sub-model and a load-side heat demand dynamic sub-model. The inputs of the source-side heat generation dynamic sub-model include the operating frequency, medium flow rate, inlet temperature and historical heat generation power sequence of the waste heat source equipment, and the output is the waste heat supply prediction curve within a future preset time window. The inputs of the load-side heat demand dynamic sub-model include the valve opening degree of the heat-using terminal, return water temperature, ambient temperature and historical heat demand power sequence, and the output is the heat demand prediction curve within a future preset time window. The two sub-models are trained and inferenced independently, and then jointly solved through coupling constraints during the optimization decision-making stage.

3. The dynamic load matching and control method for a waste heat recovery system according to claim 2, characterized in that, The source-side heat generation dynamic sub-model is constructed using a physical information neural network. Its loss function is a weighted sum of a data-driven error term and a physical conservation residual term, as shown in the formula: ,in To predict the heat production power of the model, To measure the actual heat production capacity, For the density of the medium, For specific heat capacity, For volumetric flow rate, , These are the inlet and outlet temperatures of the heat exchanger, For physically conserved residual weights, For gradient regularization weights, For network learnable parameters, The Frobenius norm is used; the physical conservation residual term forces the model output to satisfy the energy conservation equation, and the gradient regularization term suppresses overfitting oscillations of the model in sparse data regions.

4. The dynamic load matching and control method for a waste heat recovery system according to claim 3, characterized in that, The load-side heat demand dynamic sub-model is constructed using a multi-scale temporal convolutional attention network, and the output of its core feature extraction layer is calculated using the following formula: in For the first Layer at time The output characteristics, To dilate the weights of the causal convolution kernel, The time delay corresponds to the expansion factor. For input to the previous layer, For activation function, For element-wise multiplication, For channel attention module, For global average pooling, , For the parameters of the fully connected layer of the attention module, The sigmoid function is used; the dilated causal convolution is used to capture the heat-demanding inertia features at different time scales, and the channel attention module is used to adaptively enhance the feature channels that are sensitive to load mutations.

5. The dynamic load matching and control method for a waste heat recovery system according to claim 1, characterized in that, The formula for the objective function is: ,in To control the incremental sequence, To predict the time domain, , These are the predicted recovery volume and the demand volume, respectively. The Mahalanobis distance weight matrix represents the supply-demand imbalance. The time-varying deviation penalty coefficient and it varies with... Increasing monotonically decreasing, The total power consumption of the system is To control the increment, To control the incremental weighting matrix, , These are the weighting coefficients. For the terminal cost function, The terminal weights; the Mahalanobis distance weight matrix The terminal cost function is obtained by inverse estimation of the historical supply and demand deviation covariance matrix and is used to normalize the deviation under different operating conditions. This is an approximation of the Lyapunov function obtained offline, used to ensure the closed-loop stability of the rolling optimization.

6. The dynamic load matching and control method for a waste heat recovery system according to claim 5, characterized in that, When solving the objective function, slack variables are introduced to handle the infeasibility of hard constraints, and the constraints are rewritten as follows: in For the first The original inequality constraint functions, For the corresponding slack variables, This is the upper limit of cumulative relaxation. The penalty term of the relaxation variable is added to the objective function in a quadratic form, and the penalty coefficient is greater than the sum of all other weight coefficients in the objective function, ensuring that relaxation is only enabled when there is a constraint conflict.

7. The dynamic load matching and control method for a waste heat recovery system according to claim 1, characterized in that, After the coordinated adjustment of each adjustable execution component according to the dynamic matching control command sequence, the method further includes an online adaptive model update step: After each sampling period, calculate the prediction residual at the current time. ; When continuous The absolute value of the residual for each sampling period exceeds the adaptive threshold. When this occurs, local parameter fine-tuning is triggered; the adaptive threshold... The formula is dynamically adjusted based on the sliding standard deviation of recent residuals: in This is the sensitivity coefficient. The length of the sliding window. To prevent zero constants, the local parameter fine-tuning only updates the parameters of the last two layers of the dynamic load prediction model, freezes the feature extraction parameters of the previous layers, and uses the most recent samples for fine-tuning. The input / output pairs for each sampling period, with a fine-tuning iteration count not exceeding [number missing]. Second-rate.

8. A dynamic load matching and control system for a waste heat recovery system, used to implement the dynamic load matching and control method for a waste heat recovery system as described in any one of the claims, characterized in that, It includes a data acquisition module, a dual-channel prediction module, a rolling optimization module, a collaborative execution module, and an adaptive update module, specifically: The data acquisition module collects heat generation status parameters from the waste heat source side, demand status parameters from the heat load side, and operating status parameters of the waste heat recovery system in real time. The dual-channel prediction module obtains the waste heat supply prediction curve and the heat demand prediction curve within a preset time window through the source-side heat generation dynamic sub-model and the load-side heat demand dynamic sub-model, respectively; wherein, the source-side heat generation dynamic sub-model is a physical information neural network, and the load-side heat demand dynamic sub-model is a multi-scale temporal convolutional attention network. The rolling optimization module takes the optimal overall energy efficiency of the system as the objective function and combines the constraints to generate a dynamic matching control command sequence for each adjustable execution component. The collaborative execution module coordinates the adjustment of each adjustable execution component according to the dynamic matching control command sequence; The adaptive update module dynamically adjusts the trigger threshold based on the prediction residual and performs local parameter fine-tuning on the prediction model when the trigger condition is met.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a dynamic load matching and control method for a waste heat recovery system as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a dynamic load matching and control method for a waste heat recovery system as described in any one of claims 1 to 7.

Images