Curing furnace energy efficiency optimization method and system based on AI intelligent prediction of temperature field

Abstract

The application discloses a curing furnace energy efficiency optimization method and system based on AI intelligent temperature field prediction, relates to the related technical field of energy efficiency optimization, and comprises the following steps: performing multi-dimensional real-time collection based on the identification of a plurality of spatial positions of a curing furnace, obtaining a multi-dimensional time series data set, performing AI intelligent temperature field space-time distribution calculation, and obtaining a temperature field prediction tensor; determining target heating power configuration parameters; adjusting the power of the curing furnace in multiple regions, continuously monitoring synchronous operation data to perform error calculation, and constructing a prediction error sequence; and incrementally learning to construct an energy efficiency optimization strategy to perform adaptive energy efficiency optimization control on the curing furnace. The application solves the technical problems of temperature field representation distortion and prediction lag under dynamic working conditions, heating power configuration dependence on experience, and lack of global optimization in the prior art, and achieves the technical effects of improving the global uniformity control precision of the temperature field, the heating power utilization efficiency, and the adaptive control capability in the face of working condition drift.

Description

Technical Field

[0001] This invention relates to the field of energy efficiency optimization technology, specifically to a method and system for optimizing the energy efficiency of curing ovens based on AI-based intelligent temperature field prediction. Background Technology

[0002] The essence of the curing oven process is to induce the desired physical or chemical cross-linking reaction in materials by precisely controlling the temperature field distribution. However, curing oven energy consumption typically accounts for 15% to 30% of the total energy consumption of a factory. Traditional curing oven temperature control largely relies on setpoint-feedback regulation (such as PID control), which uses thermocouples placed at a limited number of points within the oven chamber to obtain temperature signals and drive heaters for power compensation. This approach has many drawbacks. Due to airflow disturbances, differences in workpiece placement, and thermal inertia, significant temperature gradients often exist in different areas within the oven, and thermocouples cannot accurately describe the complex three-dimensional temperature field distribution within the oven chamber. Feedback regulation relies on passive correction after errors occur, making it difficult to cope with sudden disturbances such as oven door opening and closing, ambient temperature changes, or workpiece batch switching, resulting in overheating or underheating and directly reducing energy efficiency. Utilization rate; and there is a strong coupling effect between thermal radiation and convection in different areas of the curing oven. Independent loop control cannot coordinate the thermal interference between areas, often causing a "one-sided increase" in power consumption; in addition, temperature field simulation based on finite element analysis (FEA) or computational fluid dynamics (CFD) and temperature prediction model established by neural network cannot reverse the power limit to deduce the temperature reachable domain. As a result, the configuration parameters given are often not the optimal solution for global energy efficiency. Furthermore, when the oven body ages, thermocouples drift, or process parameters change, the prediction error continues to accumulate, causing the energy efficiency optimization effect to degrade sharply over time.

[0003] Therefore, current technologies suffer from technical problems such as distortion and lag in temperature field characterization under dynamic operating conditions, dependence on experience in heating power configuration, and lack of global optimization. Summary of the Invention

[0004] This application provides a curing oven energy efficiency optimization method and system based on AI intelligent temperature field prediction. It solves the technical problems of temperature field characterization distortion and prediction lag under dynamic operating conditions, heating power configuration relying on experience and lacking global optimization in the prior art. It achieves the technical effects of improving the global uniformity control accuracy of temperature field, heating power utilization efficiency, and adaptive control capability in the face of operating condition drift.

[0005] This application provides a method for optimizing the energy efficiency of a curing oven based on AI-powered intelligent temperature field prediction. The method includes: performing multi-dimensional real-time data acquisition based on multiple spatial locations identified by the curing oven to obtain a multi-dimensional time-series dataset; performing AI-powered intelligent spatiotemporal distribution calculation of the temperature field to obtain a temperature field prediction tensor; performing bidirectional target search based on the temperature field prediction tensor to determine target heating power configuration parameters; adjusting the power of the curing oven in multiple regions according to the target heating power configuration parameters; continuously monitoring synchronous operation data to calculate errors and constructing a prediction error sequence; and performing incremental learning by backtracking from the prediction error sequence to the temperature field prediction tensor to construct an energy efficiency optimization strategy for adaptive energy efficiency optimization control of the curing oven.

[0006] In a possible implementation, multi-dimensional real-time acquisition is performed based on multiple spatial locations identified by the curing oven to obtain a multi-dimensional time-series dataset for AI-powered intelligent temperature field spatiotemporal distribution calculation, thereby obtaining a temperature field prediction tensor. The method includes: incorporating the geometric features of the curing oven for three-dimensional analysis to identify a first density spatial location; locating the heating zone layout parameters of the curing oven for three-dimensional analysis to identify a second density spatial location; non-uniformly distributing the curing oven based on the first and second density spatial locations to determine multiple spatial locations; performing multi-dimensional real-time acquisition based on the multiple spatial locations to obtain a multi-dimensional dataset, performing spatiotemporal alignment to generate a multi-dimensional time-series dataset; constructing a spatiotemporal graph attention network, synchronizing the multi-dimensional time-series dataset to the spatiotemporal graph attention network for AI-powered intelligent temperature field spatiotemporal distribution calculation, and generating the temperature field prediction tensor.

[0007] In a possible implementation, the construction process of the spatiotemporal graph attention network includes: performing heat conduction analysis on the curing oven based on the multiple spatial locations to generate heat conduction reachable paths; connecting the multiple spatial locations as nodes and the heat conduction reachable paths as edges to construct a spatial topology graph of the curing oven; performing spatial attention analysis based on the spatial topology graph of the curing oven to calculate multiple attention coefficients; weighting and aggregating multiple nodes according to the multiple attention coefficients to obtain spatial encoding features; performing temporal attention analysis based on the spatial topology graph of the curing oven to calculate multiple local temporal features; performing self-attention long-range dependency analysis based on the multiple local temporal features to obtain temporal encoding features; fusing the spatial encoding features and the temporal encoding features to generate initial predicted temperature field data; applying physical loss constraints based on the initial predicted temperature field to obtain a physical loss term; and jointly training the physical loss term and the initial predicted temperature field to construct the spatiotemporal graph attention network.

[0008] In a possible implementation, a bidirectional target search is performed based on the temperature field prediction tensor to determine the target heating power configuration parameters. The method includes: extracting prediction period data based on the temperature field prediction tensor; performing power accumulation analysis based on the prediction period data to set an energy consumption target; performing temperature accumulation analysis based on the prediction period data to set a temperature field uniformity target; locating the starting point based on the prediction period data and extracting power to obtain initial heating power configuration parameters; performing an energy consumption-first search based on the initial heating power configuration parameters and the energy consumption target to obtain an energy consumption boundary solution; performing a uniformity-first search based on the initial heating power configuration parameters and the temperature field uniformity target to obtain a uniformity boundary solution; constraining the energy consumption boundary solution and the uniformity boundary solution to generate multiple candidate power configuration parameters; traversing the multiple candidate power configuration parameters to perform energy consumption calculations to obtain multiple energy consumption values; traversing the multiple candidate power configuration parameters to perform temperature uniformity calculations to obtain multiple uniformity values, where the multiple energy consumption values, the multiple uniformity values, and the multiple candidate power configuration parameters have a corresponding relationship; and performing a comprehensive evaluation based on the multiple energy consumption values ​​and the multiple uniformity values ​​to determine the target heating power configuration parameters.

[0009] In a possible implementation, an energy-priority search is performed based on the initial heating power configuration parameters according to the energy consumption target to obtain the energy consumption boundary solution. The method includes: using the initial heating power configuration parameters as a starting point, calculating multiple local energy consumption gradient parameters according to the energy consumption target and heating zone layout parameters; performing reverse updates according to the multiple local energy consumption gradient parameters, setting a first power update direction vector; introducing a real-time power value, and calculating a trial power configuration parameter by superimposing the first power update direction vector and the real-time power value; performing power constraint determination based on the trial power configuration parameters according to the heating zone layout parameters; and performing power amplitude analysis based on the multiple local energy consumption gradient parameters. The process involves: setting power amplitude constraints; performing change analysis based on the multiple local energy consumption gradient parameters; setting power change rate constraints; conducting feasibility analysis according to the power amplitude constraints and power change rate constraints; defining the feasible region boundary; determining whether the trial power configuration parameters exceed the upper or lower power limits of the heating zone layout parameters; when the trial power configuration parameters exceed the upper or lower power limits of the heating zone layout parameters, projecting the trial power configuration parameters onto the feasible region boundary to obtain feasible power configuration parameters; using the feasible power configuration parameters as real-time power configuration parameters, repeatedly performing projection correction until the energy consumption target decrease is lower than the convergence threshold, and constructing the energy consumption boundary solution.

[0010] In a possible implementation, the curing oven undergoes multi-zone power adjustment based on the target heating power configuration parameters, and synchronous operation data is continuously monitored for error calculation to construct a prediction error sequence. The method includes: parsing the heating zone layout parameters according to the target heating power configuration parameters to generate a time-series power command, which corresponds to the heating zone layout parameters; executing the time-series power command to synchronously adjust the heating zone power; continuously collecting measured temperature parameters from multiple spatial locations to construct multiple temperature field measurement matrices, which contain multiple control times; matching the temperature field prediction tensors according to the multiple control times to extract multiple predicted temperature values; performing spatiotemporal point-by-point comparisons between the multiple predicted temperature values ​​and the multiple temperature field measurement matrices to calculate multiple temperature prediction residuals; and aggregating the multiple temperature prediction residuals to construct multiple residual vectors arranged in temporal order to construct the prediction error sequence.

[0011] In a possible implementation, after constructing the prediction error sequence, the method includes: constructing a multi-scale sliding window, which includes a short-term sliding window and a long-term sliding window; performing statistical analysis on the prediction error sequence through the short-term sliding window to calculate the residual mean data; performing statistical analysis on the prediction error sequence through the long-term sliding window to calculate the residual trend slope; generating a performance drift signal when the residual mean data exceeds a preset noise threshold and the residual trend slope remains positive; and generating a sudden fault alarm signal when the residual mean data exceeds a preset impact threshold.

[0012] In one possible implementation, incremental learning is performed by backtracking from the predicted error sequence to the temperature field prediction tensor to construct an energy efficiency optimization strategy for adaptive energy efficiency optimization control of the curing furnace. The method includes: performing multi-scale statistical monitoring of the predicted error sequence to identify temperature field prediction performance degradation patterns and generate monitoring indicators; setting incremental learning trigger conditions, and generating incremental learning samples when the monitoring indicators meet the trigger conditions; incrementally training the temperature field prediction tensor based on the incremental learning samples to generate an optimized temperature field prediction tensor; performing a bidirectional target search based on the optimized temperature field prediction tensor to obtain target heating power configuration parameters; and applying the target heating power configuration parameters to the curing furnace for multi-region power adjustment to construct an energy efficiency optimization strategy for adaptive energy efficiency optimization control of the curing furnace.

[0013] In a possible implementation, the temperature field prediction tensor is incrementally trained based on the incremental learning samples to generate an optimized temperature field prediction tensor. The method includes: evaluating the contribution of the temperature field prediction tensor based on the incremental learning samples and calculating multiple weight importance evaluation values; applying weighted constraints to the temperature field prediction tensor based on the multiple weight importance evaluation values ​​to construct an incremental training dataset; and performing a finite number of iterations of fine-tuning based on the incremental training dataset to generate the optimized temperature field prediction tensor.

[0014] This application also provides an AI-based intelligent temperature field prediction system for curing oven energy efficiency optimization. The system includes: a data acquisition module for multi-dimensional real-time acquisition based on multiple spatial locations identified by the curing oven, obtaining a multi-dimensional time-series dataset for AI intelligent temperature field spatiotemporal distribution calculation, and obtaining a temperature field prediction tensor; a configuration parameter determination module for bidirectional target search based on the temperature field prediction tensor to determine target heating power configuration parameters; a prediction error sequence construction module for multi-region power adjustment of the curing oven according to the target heating power configuration parameters, continuously monitoring synchronous operation data for error calculation, and constructing a prediction error sequence; and an energy efficiency optimization control module for incremental learning by backtracking from the prediction error sequence to the temperature field prediction tensor, constructing an energy efficiency optimization strategy for adaptive energy efficiency optimization control of the curing oven.

[0015] This application proposes an AI-based intelligent temperature field prediction-based energy efficiency optimization method and system for curing ovens. This method involves multi-dimensional real-time data acquisition based on multiple spatial locations identified within the curing oven, obtaining a multi-dimensional time-series dataset for AI-based intelligent spatiotemporal distribution calculation of the temperature field, and acquiring a temperature field prediction tensor. It then determines target heating power configuration parameters, adjusts the power of the curing oven in multiple regions, continuously monitors synchronous operating data for error calculation, and constructs a prediction error sequence. Incremental learning is used to construct an energy efficiency optimization strategy for adaptive energy efficiency optimization control of the curing oven. This addresses the technical problems of temperature field distortion and prediction lag under dynamic operating conditions, and the dependence of heating power configuration on experience and the lack of global optimization in existing technologies. The method achieves improved accuracy in global temperature field uniformity control, heating power utilization efficiency, and adaptive control capabilities in the face of operating condition drift. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments of this disclosure will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a schematic diagram of the energy efficiency optimization method for curing ovens based on AI intelligent temperature field prediction, provided in an embodiment of this application.

[0018] Figure 2 A schematic diagram of the three-dimensional furnace chamber space model of the curing furnace in the curing furnace energy efficiency optimization method based on AI intelligent temperature field prediction provided in the embodiments of this application.

[0019] Figure 3 This is a schematic diagram of the curing oven spatial topology in the curing oven energy efficiency optimization method based on AI intelligent temperature field prediction provided in the embodiments of this application.

[0020] Figure 4 A schematic diagram of the energy efficiency optimization system for curing ovens based on AI intelligent temperature field prediction provided in this application embodiment.

[0021] Figure labeling: Data acquisition module 10, configuration parameter determination module 20, prediction error sequence construction module 30, energy efficiency optimization control module 40. Detailed Implementation

[0022] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0023] This application provides an embodiment of a curing oven energy efficiency optimization method based on AI intelligent temperature field prediction, such as... Figure 1 As shown, the method includes:

[0024] Step S100: Based on the multiple spatial locations identified by the curing oven, multidimensional real-time data acquisition is performed to obtain a multidimensional time-series dataset for AI intelligent temperature field spatiotemporal distribution calculation, and temperature field prediction tensor is obtained.

[0025] Step S100 further includes: performing three-dimensional analysis on the geometric features of the curing oven to identify the first density spatial location; performing three-dimensional analysis on the heating zone layout parameters of the curing oven to identify the second density spatial location; performing non-uniform distribution identification on the curing oven based on the first density spatial location and the second density spatial location to determine multiple spatial locations; performing multi-dimensional real-time acquisition based on the multiple spatial locations to obtain a multi-dimensional dataset for spatiotemporal alignment to generate a multi-dimensional time-series dataset; constructing a spatiotemporal graph attention network, synchronizing the multi-dimensional time-series dataset to the spatiotemporal graph attention network for AI intelligent temperature field spatiotemporal distribution calculation, and generating the temperature field prediction tensor.

[0026] Preferably, the physical dimensions of the curing furnace, such as its length, width, height, internal volume, wall thickness, and furnace geometry, are used as input parameters to establish a three-dimensional furnace space model in a three-dimensional coordinate system, such as... Figure 2 As shown, based on the geometric model, according to a relatively dense spatial sampling interval (i.e., high spatial resolution), the first set of spatial coordinate points is marked within the furnace volume to determine the first density spatial location, which is used to characterize areas where the temperature field within the furnace may change drastically, such as near the furnace wall, corners, or airflow confluences. Layout data such as the installation position, power level, radiation angle, and coverage of independent heating units such as electric heating tubes, infrared radiation plates, and hot air nozzles within the curing furnace are obtained, and a spatial distribution model of the heating elements is established in the same three-dimensional coordinate system. Based on this heating zone layout model, according to a relatively sparse spatial sampling interval (i.e., low spatial resolution), the second set of spatial coordinate points is marked within the furnace volume to determine the second density spatial location, which is used to characterize the center position of the area directly affected by the heating source and the boundary position between adjacent heating zones.

[0027] Preferably, the first batch of high-density sampling points and the second batch of low-density sampling points are spatially superimposed and comprehensively screened. Sampling points are retained or added at locations with large temperature gradients, such as near the furnace wall, heating elements, and area boundaries, and are marked with high density. Sampling points are removed or reduced at the furnace center or locations with relatively uniform temperature fields, and are marked with low density. This forms a non-uniformly distributed set of spatial coordinate points, determining multiple spatial locations whose distribution density is positively correlated with the rate of change of the temperature field. Then, at each determined spatial coordinate location, temperature sensors, such as thermocouples, are deployed, and multiple types of physical quantity data are collected simultaneously, including temperature values, collection timestamps, sensor numbers, and location coordinates. Data from different sensors are timestamped according to a unified time base and spatially matched according to spatial coordinates, ensuring that data at each moment corresponds to a specific three-dimensional spatial location, forming a multi-dimensional time-series dataset containing time, space, and physical quantity dimensions. A deep learning model based on a graph neural network architecture was constructed, defining a spatiotemporal graph attention network, including a spatial graph attention layer and a time-series attention layer. After inputting a multi-dimensional time-series dataset into this network, the model constructs a spatial topology graph of the curing furnace, using spatial locations as graph nodes and spatial distances or thermal conductivity relationships between locations as graph edges. Through an attention mechanism, the model automatically learns the temperature change characteristics of each node at different times and the mutual influence weights between nodes, calculating the predicted temperature value of each node at future times. The prediction results of all nodes are arranged in spatial coordinates to generate a temperature field prediction tensor, which comprehensively describes the temperature field distribution of the entire furnace at the present and multiple future times. An example of the curing furnace spatial topology diagram is shown below. Figure 3 As shown.

[0028] Preferably, the spatiotemporal graph attention network consists of an input layer, a spatial graph attention encoding layer, a time series self-attention encoding layer, a feature fusion layer, and an output layer connected in series. The input layer receives a multi-dimensional time series dataset, including the length of the historical time window (ranging from 10 to 30 control periods), the total number of spatial nodes (i.e., the number of spatial locations, ranging from 50 to 200), and the input feature dimension of each node. The spatial graph attention encoding layer contains two graph attention sublayers, each with four attention heads, and each head outputs a 64-dimensional feature dimension, using Le... The akyReLU activation function is used with a negative slope of 0.2. The time-series self-attention encoding layer contains two Transformer encoder sublayers, each with four attention heads. The hidden layer of the feedforward network has a dimension of 256, and sinusoidal position encoding is used to mark the temporal positions. The feature fusion layer concatenates the features according to the node dimensions to obtain fused features, which are then mapped to the single-step temperature prediction value of each node through two fully connected networks (hidden layer dimensions of 128 and 64, respectively). The output layer arranges the predicted temperature values ​​of N nodes according to the original spatial coordinates to generate a temperature field prediction vector.

[0029] Preferably, the training dataset consists of historical operation records. Each sample contains input features and an output label. The input features include the measured temperature value of each node at time t, the temperature change rate (i.e., the temperature difference between the current time and the previous time divided by the sampling interval), and the current power value of the heating zone to which the node belongs. The output label is the measured temperature value of each node in the kth control cycle. The training set contains no less than 5000 valid samples, divided into training, validation, and test sets in an 8:1:1 ratio. The network weights are initialized using the Xavier uniform initialization method, the bias is initialized to 0, the AdamW optimizer is selected, and the initial learning rate is set to 1×10. -3 A cosine annealing decay strategy is adopted, with the temperature decaying to 10% of the initial value every 200 training rounds. The batch size is set to 32, and the maximum number of training rounds is 500. An early stopping strategy is adopted, with the validation set loss not decreasing for 20 consecutive rounds as the termination condition. The loss function is the mean square error between the predicted temperature value and the measured temperature value. Among them, the physical loss term is composed of two weighted parts: the residual loss of the heat conduction equation and the boundary condition loss. The residual loss of the heat conduction equation is calculated by substituting the predicted temperature field into the transient heat conduction equation, calculating the square of the absolute value of the difference between the left and right sides of the equation, and summing and averaging over all spatial nodes and the prediction period. The boundary condition loss is calculated for the furnace wall boundary nodes, by summing the squares of the differences between the predicted temperature and the set wall temperature boundary conditions, and dividing by the number of boundary nodes.

[0030] Furthermore, step S100 also includes: performing heat conduction analysis on the curing oven based on the multiple spatial locations to generate heat conduction reachable paths; connecting the multiple spatial locations as nodes and the heat conduction reachable paths as edges to construct a curing oven spatial topology graph; performing spatial attention analysis on the curing oven spatial topology graph to calculate multiple attention coefficients; weighting and aggregating multiple nodes according to the multiple attention coefficients to obtain spatial encoding features; performing temporal attention analysis on the curing oven spatial topology graph to calculate multiple local temporal features; performing self-attention long-range dependency analysis based on the multiple local temporal features to obtain temporal encoding features; fusing the spatial encoding features and the temporal encoding features to generate initial predicted temperature field data; applying physical loss constraints based on the initial predicted temperature field to obtain a physical loss term; and jointly training the physical loss term and the initial predicted temperature field to construct the spatiotemporal graph attention network.

[0031] Preferably, based on the physical mechanisms of heat transfer within the curing furnace, including thermal radiation, thermal convection, and thermal conduction, and based on the heat conduction equation and the thermal properties of the furnace medium such as thermal conductivity, specific heat capacity, and density, possible paths for heat transfer from any spatial location to another are calculated. All pairs of spatial locations with direct heat transfer relationships are identified and recorded as reachable heat conduction paths. Using these identified spatial locations as vertices of a graph and the reachable heat conduction paths as connecting edges, a graph data structure is established. Each edge is assigned a weight value, determined by the spatial distance between nodes and the heat conduction efficiency, thus forming a non-Euclidean structural graph characterizing the spatial thermal coupling relationships within the curing furnace—that is, the curing furnace spatial topology graph. Then, on the curing furnace spatial topology graph, a learnable attention weight function is used to quantify the correlation between each node and all its adjacent nodes. The input is the node's temperature and location characteristics, and the output is a scalar value characterizing the importance of adjacent nodes' influence on the current node's temperature. Furthermore, attention coefficients are calculated for each node's adjacent edges, and the sum of these coefficients is normalized to 1. The temperature feature vectors of each node's neighboring nodes are multiplied by their corresponding attention coefficients and then summed in a weighted manner. The sum is then concatenated or added to the node's own feature vector. This process is repeated for all nodes, and each node obtains an updated feature vector. This vector integrates the node's own temperature information and the weighted sum of the temperature contributions from its neighboring nodes. All the updated node feature vectors together constitute the spatial coding features.

[0032] Preferably, a fixed-length historical temperature sequence is extracted for each node in the time dimension, such as the measurement values ​​of the past K moments. Self-attention calculation is performed on the sequence data within the time window to extract the short-term fluctuation pattern and trend of temperature change over time for that node, and the corresponding local temporal feature vector is calculated for each node. The local temporal features of all nodes at multiple time steps are globally concatenated to construct a complete sequence matrix in the time dimension. A multi-head self-attention mechanism is applied to this matrix to calculate the correlation weight between any two moments in the sequence, thereby capturing the periodic changes of the temperature field over a long time span, the cumulative effect of thermal inertia, and the long-term impact of historical disturbances on the current state. After multi-layer attention stacking, the time-coded features are output, which contain the global temporal dependencies of each node within the complete time window.

[0033] Preferably, the spatial encoding feature vector and the temporal encoding feature vector of each node are concatenated and fused to form the fused feature of that node. The fused features of all nodes are input into a fully connected feedforward neural network. The network outputs the temperature prediction value of each node at a certain time in the future. The prediction values ​​of all nodes are arranged according to the original spatial coordinates to form the initial predicted temperature field data, that is, the preliminary temperature tensor without physical constraints. The initial predicted temperature field data is substituted into the Fourier partial differential equation of heat conduction to calculate the difference between the second derivative of the predicted field in space and the first derivative in time. This difference is used as the residual term. At the same time, the consistency between the predicted temperature field and the boundary conditions of the curing furnace (such as furnace wall heat flux and convective heat transfer coefficient) is checked, and the boundary deviation is calculated. The residual term and the boundary deviation together constitute the physical loss term, which is used to quantify the severity of the prediction result violating the physical laws. During model training, the data-driven prediction error loss, namely the mean square error between the initial predicted temperature field and the actual measured value, is weighted and summed with the physical loss term to form the total loss function. Finally, the network parameters are optimized through backpropagation, so that the model gradually reduces the deviation of the prediction results from the physical equation of heat conduction while fitting the training data. After multiple rounds of iterative training, a complete spatiotemporal graph attention network model with converged parameters is obtained, which can be directly used for real-time temperature field prediction.

[0034] Step S200: Perform a bidirectional target search based on the temperature field prediction tensor to determine the target heating power configuration parameters.

[0035] Step S200 further includes: extracting prediction period data based on the temperature field prediction tensor; performing power accumulation analysis according to the prediction period data to set an energy consumption target; performing temperature accumulation analysis according to the prediction period data to set a temperature field uniformity target; locating the starting point based on the prediction period data to extract power and obtain initial parameters for heating power configuration; performing energy consumption priority search according to the energy consumption target based on the initial parameters for heating power configuration to obtain an energy consumption boundary solution; performing uniformity priority search according to the temperature field uniformity target based on the initial parameters for heating power configuration to obtain a uniformity boundary solution; constraining the energy consumption boundary solution and the uniformity boundary solution to generate multiple candidate power configuration parameters; traversing the multiple candidate power configuration parameters to perform energy consumption calculation to obtain multiple energy consumption values; traversing the multiple candidate power configuration parameters to perform temperature uniformity calculation to obtain multiple uniformity values, wherein the multiple energy consumption values, the multiple uniformity values, and the multiple candidate power configuration parameters have a corresponding relationship; and performing a comprehensive evaluation based on the multiple energy consumption values ​​and the multiple uniformity values ​​to determine the target heating power configuration parameters.

[0036] Preferably, the predicted temperature values ​​of each node within a specific future time interval (i.e., the prediction period) are extracted from the temperature field prediction tensor. The output power of all heating zones within this period is time-integrated and summed to calculate the total energy consumption. An upper limit for the curing oven's energy consumption is set as the energy consumption target. Simultaneously, the variance of the predicted temperature values ​​at each spatial location at each moment within this period is calculated. The change in the difference between the highest and lowest temperatures within the furnace over time is statistically analyzed, and a maximum allowable temperature difference threshold is set as the temperature field uniformity target. The actual operating power values ​​of each heating zone in the curing oven at the current moment are obtained, including the current voltage, current, and duty cycle of each heating zone. The actual operating power values ​​are used as the initial starting point for power optimization, forming an initial parameter vector for heating power configuration. The dimension of this vector is equal to the total number of heating zones. Then, starting from the initial parameters of the heating power configuration, a single-objective optimization search is performed in the power configuration parameter space along the direction of minimizing energy consumption. During the search, the uniformity of the temperature field is fixed as a constraint condition, that is, the temperature difference must not exceed the set threshold. Among all the power configuration sets that satisfy the uniformity constraint, the power configuration scheme that minimizes the total energy consumption is found, which is the energy consumption boundary solution.

[0037] Preferably, starting from the initial parameters of the heating power configuration, a single-objective optimization search is performed in the power configuration parameter space along the direction of optimal temperature field uniformity. During the search, the energy consumption objective is fixed as a constraint, i.e., the total energy consumption must not exceed a set upper limit. Among all power configurations that satisfy the energy consumption constraint, the power configuration scheme that minimizes the temperature difference in the furnace is found, which is the uniformity boundary solution. Using the energy consumption boundary solution and the uniformity boundary solution as the two endpoints of the Pareto front, linear interpolation or random sampling methods are used in the parameter space between them, such as using Latin hypercube sampling to generate multiple intermediate power configuration schemes. Each scheme is a combination vector of power values ​​of each heating zone. All schemes together constitute a candidate power configuration parameter set to cover the complete transition interval from pure energy consumption optimality to pure uniformity optimality. Specifically, let the total number of heating zones be R. The power value of each heating zone is divided into S equally spaced sub-intervals (S = 20) within its corresponding feasible interval with equal probability. Then, a value is randomly sampled from one of the sub-intervals for each heating zone, and these are combined to form a candidate scheme. The total number of candidate power configuration parameters repeatedly generated is M = 5 × R, i.e., M = 40 when R = 8, and M = 60 when R = 12. Simultaneously, one scheme is added at each of the two endpoints to ensure that the candidate set covers the complete Pareto front boundary, thus the total number of candidate schemes is M + 2. Linear interpolation is used as an auxiliary sampling method only when the number of heating zones R ≤ 4. When R > 4, only Latin hypercube sampling is used to avoid dimensionality explosion. After sampling, a feasibility check is performed on each candidate scheme, eliminating any scheme where the power value of any heating zone exceeds its physical rated power upper or lower limit. The final number of retained schemes is no less than 2 × R.

[0038] Next, for each power configuration scheme in the candidate power configuration parameter set, two calculations are performed. The first is to sum the power of each heating zone over time during the prediction period to obtain the energy consumption value of the scheme. The second is to calculate the variance and range of the temperature values ​​of all nodes in the temperature field prediction tensor under the power configuration to obtain the temperature uniformity value of the scheme. Each scheme corresponds to an energy consumption value and a uniformity value, forming a mapping relationship. The energy consumption value and uniformity value of each candidate scheme are normalized to eliminate the difference in their dimensions. Multi-objective decision-making is used to perform a weighted summation of the energy consumption index and uniformity index of each scheme or to calculate the Euclidean distance to the ideal point. Then, all candidate schemes are comprehensively ranked, and the candidate scheme with the highest comprehensive score is selected as the final target heating power configuration parameter output.

[0039] Furthermore, step S200 also includes: using the initial parameters of the heating power configuration as a starting point, calculating multiple local energy consumption gradient parameters according to the energy consumption target and the heating zone layout parameters; performing reverse updates according to the multiple local energy consumption gradient parameters, and setting a first power update direction vector; introducing real-time power values, and superimposing the first power update direction vector with the real-time power values ​​to calculate trial power configuration parameters; performing power constraint determination based on the trial power configuration parameters and the heating zone layout parameters; performing power amplitude analysis based on the multiple local energy consumption gradient parameters, setting power amplitude constraints, and based on the multiple local energy consumption gradient parameters... Local gradient parameters are analyzed for change, and power change rate constraints are set. Feasibility analysis is performed according to the power amplitude constraints and power change rate constraints to delineate the feasible region boundary. It is determined whether the trial power configuration parameters exceed the upper or lower power limit of the heating zone layout parameters. When the trial power configuration parameters exceed the upper or lower power limit of the heating zone layout parameters, the trial power configuration parameters are projected onto the feasible region boundary to obtain feasible power configuration parameters. The feasible power configuration parameters are used as real-time power configuration parameters, and projection correction is repeatedly performed until the energy consumption target reduction is lower than the convergence threshold to construct the energy consumption boundary solution.

[0040] Preferably, using the initial parameters of the current heating power configuration of each heating zone as the starting vector, the partial derivative of the total energy consumption function is solved in the power configuration parameter space. Combined with the heating zone layout parameters, including the spatial location, radiation coverage, and thermal coupling coefficient of each heating zone, the change in total energy consumption when the power value of each heating zone changes by a unit is calculated, serving as the local energy consumption gradient parameter for that heating zone. This process is repeated for all heating zones to obtain multiple local energy consumption gradient parameters. The negative value of each local energy consumption gradient parameter is taken, causing the power update direction to point towards the direction of decreasing total energy consumption. The negative gradient values ​​of all heating zones are arranged in order of heating zone number, forming a square with a dimension equal to the total number of heating zones. The first power update direction vector represents the direction in which the power of each heating zone should be adjusted in the current iteration step to reduce the total energy consumption. The actual output power value of each heating zone of the curing oven at the current moment is obtained to form a real-time power vector. The first power update direction vector is multiplied by a preset step size coefficient and then added element by element to the real-time power vector to obtain a new power value combination, which is the trial power configuration parameter. Then, the power value of each heating zone in the trial power configuration parameter is compared item by item with the rated power range (including the lower limit and the upper limit of power) of each heating zone defined in the heating zone layout parameter of the curing oven to determine whether each power value falls within its corresponding allowable range.

[0041] Preferably, the maximum adjustable power output value of each heating zone is statistically analyzed, and upper and lower limits of the power amplitude for each heating zone are set as power amplitude constraints. Simultaneously, the maximum allowable change in power for each heating zone within a unit time step is specified. Based on the heater's thermal inertia response time and temperature control accuracy requirements, a power change rate constraint is set, meaning the power difference between two consecutive control cycles must not exceed this maximum allowable change. The power amplitude constraint and the power change rate constraint are treated as a system of simultaneous inequality equations. The set of all power value combinations that simultaneously satisfy both constraints is solved in the power configuration parameter space. The boundary of this set is the feasible region boundary, and any power configuration parameter located within this boundary can be physically executed by the curing oven's heating system.

[0042] Preferably, each element in the trial power configuration parameters is compared one by one with the upper and lower power limits of the corresponding heating zone. If an element is greater than the upper power limit, its value is modified to be equal to the upper power limit; if an element is less than the lower power limit, its value is modified to be equal to the lower power limit. After performing correction operations on all elements that exceed the boundaries, a power configuration parameter is obtained in which all elements fall within the allowable range and satisfy the power change rate constraint, which is recorded as the feasible power configuration parameter. Finally, the feasible power configuration parameter obtained in this iteration is used as the starting real-time power value for the next iteration. The gradient is recalculated, trial parameters are generated, constraints are determined, and projection correction is performed to form a cyclic iterative process. After each iteration, the difference between the current total energy consumption and the total energy consumption of the previous round is calculated. When the difference is less than the preset convergence threshold, the iteration is terminated, and the feasible power configuration parameter obtained in the last iteration is output as the energy consumption boundary solution.

[0043] Step S300: Adjust the power of the curing oven in multiple zones according to the target heating power configuration parameters, continuously monitor the synchronous operation data to calculate the error, and construct a predicted error sequence.

[0044] Step S300 further includes: parsing the target heating power configuration parameters according to the heating zone layout parameters to generate a time-series power command, wherein the time-series power command corresponds to the heating zone layout parameters; executing the time-series power command to synchronously adjust the heating zone power; continuously collecting measured temperature parameters at multiple spatial locations to construct multiple temperature field measurement matrices, wherein the multiple temperature field measurement matrices contain multiple control times; matching the temperature field prediction tensor according to the multiple control times to extract multiple predicted temperature values; performing spatiotemporal point-by-point comparison between the multiple predicted temperature values ​​and the multiple temperature field measurement matrices to calculate multiple temperature prediction residuals; aggregating the multiple temperature prediction residuals to construct the multiple residual vectors arranged in time sequence to construct the prediction error sequence.

[0045] Preferably, the target heating power configuration parameters are matched and associated with the heating zone layout parameters such as the physical location, number, and control channel address of each heating zone. Each power value is converted into the control signal format of the corresponding heating zone, including voltage amplitude, pulse width modulation duty cycle, or current setpoint. According to the communication protocol and data frame structure of the curing oven control system, the control signals of each heating zone are packaged into a sequence of instructions arranged in chronological order. Each instruction explicitly specifies the target heating zone number, execution timestamp, and target power value, so that the instructions and heating zones form a one-to-one mapping relationship. The time-series power instructions are sent to the curing oven control system, so that each heating zone synchronously outputs and adjusts according to the power value specified in the instruction. During the power adjustment process, the measurement values ​​of temperature sensors at multiple spatial locations are continuously read, and the temperature readings of each sensor at multiple control times are recorded. The temperature measurement values ​​of all spatial locations within the same control time are arranged in a matrix form according to spatial coordinates. A temperature field measurement matrix is ​​generated for each control time, and multiple temperature field measurement matrices are generated for multiple control times. The multiple temperature field measurement matrices contain multiple control times.

[0046] Preferably, using the timestamp of each control moment as an index, the predicted temperature field data slice corresponding to that moment is searched in the temperature field prediction tensor. For the temperature prediction value at each spatial location in the temperature field prediction tensor, the spatial coordinates of that location are matched with the sensor position coordinates in the measured matrix to extract the predicted temperature values ​​of all spatial locations at the same control moment, forming a predicted temperature matrix with the same spatial dimension as the measured temperature field matrix. Then, the predicted temperature matrix at each control moment is matched one-to-one with the measured temperature field matrix at the same moment in spatial coordinates. Subtraction is performed on each spatial location, that is, the predicted temperature value at that location is subtracted from the measured temperature value to obtain the prediction residual value at that location at that moment. Subtraction is performed on all control moments and all spatial locations to obtain the temperature prediction residual corresponding to each spatial location at each control moment. Then, the temperature prediction residuals of all spatial locations at each control time are arranged in a fixed spatial order to form a residual vector. The dimension of this vector is equal to the total number of spatial locations. The residual vectors corresponding to all control times are arranged vertically in chronological order to form a two-dimensional array. The rows of this array correspond to the control times, and the columns correspond to the spatial locations. This array is the prediction error sequence, which is used to completely record the deviation distribution between the predicted temperature field and the measured temperature field in both time and space dimensions.

[0047] Furthermore, step S300 also includes constructing a multi-scale sliding window, which includes a short-term sliding window and a long-term sliding window; performing statistical analysis on the prediction error sequence through the short-term sliding window to calculate the residual mean data; performing statistical analysis on the prediction error sequence through the long-term sliding window to calculate the residual trend slope; generating a performance drift signal when the residual mean data exceeds a preset noise threshold and the residual trend slope remains positive; and generating a sudden fault alarm signal when the residual mean data exceeds a preset impact threshold.

[0048] Preferably, a multi-scale sliding window is defined on the time series data, including a short-term sliding window and a long-term sliding window. The short-term sliding window covers the prediction error data of the most recent N consecutive control moments, where N is a small value, such as 10 to 30 moments, and is used to capture the instantaneous fluctuation characteristics of the prediction error in the recent period. The long-term sliding window covers the prediction error data of the most recent M consecutive control moments, where M is a large value, such as 100 to 300 moments, and is used to capture the cumulative change trend of the prediction error over a longer period. Both windows slide forward over time, always maintaining coverage of the latest data. Extract all residual values ​​currently covered by the short-term sliding window from the prediction error sequence, sum them, and divide by the total number of data points within the window to obtain a scalar value. This value represents the direction and magnitude of the average deviation between the predicted temperature and the measured temperature in the recent period and is denoted as the residual mean. Extract all residual values ​​currently covered by the long-term sliding window from the prediction error sequence. Using time order as the independent variable and the residual value as the dependent variable, perform linear regression fitting on the data points within the window using the least squares method. Calculate the slope of the fitted line to represent the monotonic rate and direction of change of the prediction error in the long-term period. A positive slope indicates that the prediction deviation continues to increase over time, while a negative slope indicates that the prediction deviation continues to decrease over time.

[0049] Preferably, the absolute value of the residual mean data is compared with a preset noise threshold, where the preset noise threshold represents the acceptable deviation range of the prediction model under normal operating conditions. Simultaneously, it is determined whether the slope of the residual trend is positive for multiple consecutive long-term sliding windows, i.e., the slope remains greater than zero. When both conditions are met, it is determined that the prediction model's ability to represent the current state of the curing furnace has degraded, and a performance drift signal is output as a start flag for triggering incremental learning operations. Then, the absolute value of the residual mean data is compared with a preset impact threshold, where the preset impact threshold is a pre-set maximum allowable error limit. Its value is greater than the noise threshold, representing the limit deviation range the system can withstand. When the absolute value of the residual mean data exceeds this impact threshold, it is determined that an unexpected and severe thermal anomaly has occurred in the curing furnace, such as heater burnout, fan stoppage, or accidental opening of the furnace door. A sudden fault alarm signal is then output to trigger emergency protection shutdown or manual intervention.

[0050] Step S400: Incremental learning is performed by backtracking to the temperature field prediction tensor according to the prediction error sequence to construct an energy efficiency optimization strategy for adaptive energy efficiency optimization control of the curing oven.

[0051] Step S400 further includes: performing multi-scale statistical monitoring on the prediction error sequence, identifying temperature field prediction performance degradation patterns, and generating monitoring indicators; setting incremental learning trigger conditions, and generating incremental learning samples when the monitoring indicators meet the incremental learning trigger conditions; performing incremental training on the temperature field prediction tensor based on the incremental learning samples to generate an optimized temperature field prediction tensor; performing bidirectional target search based on the optimized temperature field prediction tensor to obtain target heating power configuration parameters; applying the target heating power configuration parameters to the curing oven for multi-region power adjustment, and constructing an energy efficiency optimization strategy for adaptive energy efficiency optimization control of the curing oven.

[0052] Preferably, the prediction error sequence is input into multiple statistical monitoring units at different time scales. Each unit calculates the mean, variance, autocorrelation coefficient, and skewness of the residuals. The statistics at each scale are compared to determine whether the distribution of the prediction error has changed from an initial zero-mean white noise distribution to a distribution with a significant non-zero mean or trend. Based on the judgment results, monitoring indicators such as the mean drift of the error, the rate of change of the error variance, and the trend strength coefficient are output to quantify the current performance deviation of the prediction model. A set of logical judgment rules is set as the incremental learning trigger condition, which includes the threshold intervals and combined logical relationships corresponding to each component of the monitoring indicators. For example, if the mean drift of the error is greater than the set threshold and the trend strength coefficient is higher than the set value, the trigger condition is determined to be met. The monitoring indicators calculated in real time are compared with the trigger condition item by item. When all logical judgments return true values, the input feature data in the latest time period is extracted from the prediction error sequence, including the measured temperature sequence, power configuration sequence, timestamp, and the corresponding measured temperature true value, to generate incremental learning samples.

[0053] Preferably, the incremental learning triggering conditions include: error mean drift, the absolute value of the residual mean data (calculated by a short-term sliding window with a window length of 15 control cycles) exceeding a preset noise threshold, which is obtained by taking 3 times the standard deviation of the predicted residual data of the curing oven running continuously for 24 hours under stable conditions, such as 0.8℃; a continuously deteriorating trend, the residual trend slope (calculated by a long-term sliding window with a window length of 120 control cycles, using least squares linear fitting) being positive for 3 consecutive long-term windows and the slope value being greater than 0.02℃ / cycle; and accumulated error energy, the root mean square error of the predicted error sequence in the long-term window exceeding a set threshold of 1.5℃, which is determined by 1.2 times the maximum value of RMSE in historical normal operation data.

[0054] Preferably, the incremental learning samples are used as new training data and input into the spatiotemporal graph attention network model. Mini-batch gradient descent is used to update the network parameters in a new round. During training, the network simultaneously fits the distribution characteristics of the original historical data and the new operating conditions reflected in the incremental samples. After the parameter update, the updated network model is used to recalculate the temperature field prediction for the future period of the curing furnace, generating a temperature field prediction tensor with optimized prediction accuracy, denoted as the temperature field prediction optimization tensor. Using the temperature field prediction optimization tensor as input data, the energy consumption priority search and uniformity priority search are re-executed to obtain the energy consumption boundary solution and uniformity boundary solution, respectively. Then, within the constraints of the two, a set of candidate power configuration parameters is generated and comprehensively evaluated. The power configuration scheme with the highest score is selected as the new target heating power configuration parameter. Finally, the target heating power configuration parameters are sent to the curing oven control system, parsed into time-series power commands for each heating zone, and executed to complete synchronous power adjustment across multiple zones. The parameter calculation process and adjustment execution process used for this adjustment are recorded as complete control strategy rules, including the prediction optimization tensor, target power configuration parameters, and their corresponding expected energy consumption and uniformity indices, forming an energy efficiency optimization strategy. This strategy continuously applies to the curing oven operation, enabling the control parameters to automatically update with changes in operating conditions, achieving adaptive optimization control of the curing oven's energy efficiency. The gradient calculation in gradient descent and projection correction uses an automatic differentiation method. Specifically, with total energy consumption as the objective function and temperature field uniformity as the constraint function, a spatiotemporal graph attention network is used as the forward propagation model. After inputting the current power configuration parameters, the corresponding predicted temperature field is output, thereby calculating the total energy consumption and temperature field uniformity. Then, the automatic differentiation engine built into the deep learning framework is used to calculate the partial derivative of the objective function with respect to the power of each heating zone, which is the local gradient parameter of energy consumption. For the gradient calculation in the uniformity-first search, the temperature field uniformity index is used as the objective function, and the uniformity local gradient parameter is also calculated by automatic differentiation. In this process, the calculation is repeated once in each iteration step, and the network parameters are frozen during the calculation (i.e., no network parameter update is performed). Only the power configuration parameter is used as a trainable variable for gradient backpropagation.

[0055] Furthermore, step S400 also includes: evaluating the contribution of the temperature field prediction tensor based on the incremental learning samples, and calculating multiple weight importance evaluation values; applying weighted constraints to the temperature field prediction tensor based on the multiple weight importance evaluation values ​​to construct an incremental training dataset; and performing a finite number of iterations of fine-tuning based on the incremental training dataset to generate the optimized temperature field prediction tensor.

[0056] Preferably, the input feature data from the incremental learning samples is input into the current spatiotemporal graph attention network model, and forward propagation is performed to obtain the predicted value of the network output layer. Then, backpropagation is performed to solve the partial derivative of the loss function with respect to each trainable weight parameter in the network model. At the same time, the activation frequency and parameter change magnitude of each weight parameter on the current batch of incremental samples are recorded. Then, the absolute value of the partial derivative, activation frequency and change magnitude are weighted and calculated. The weighting coefficients of the three indicators are 0.5, 0.3 and 0.2 respectively, and the sum of the three coefficients is 1, so as to balance the contribution of gradient sensitivity, usage frequency and change magnitude to importance assessment. After all weights are calculated, they are normalized so that their values ​​are in the range of [0, 1]. Finally, the normalized values ​​are used as the scalar values ​​corresponding to each weight parameter, that is, the weight importance assessment value, which represents the degree of contribution of the weight parameter to the prediction result of the current incremental learning samples. Then, the importance evaluation value of each weight parameter is used as the elastic weight constraint coefficient for that parameter. When constructing the loss function for incremental training, an additional regularization term is added to each weight parameter. This term is the importance evaluation value multiplied by the square of the difference between the parameter value and the parameter value at the end of the historical training. This regularization term imposes a stronger constraint on weight parameters with higher importance, limiting their large changes during incremental training, and a weaker constraint on weight parameters with lower importance. The constrained loss function is then combined with the incremental learning samples to form the incremental training dataset. This dataset contains input features, target ground truth values, and a vector of weight constraint coefficients for each sample. The incremental training dataset is then input into the spatiotemporal graph attention network model for parameter updates. In each iteration, forward propagation is performed to calculate the predicted value, the total loss including the regularization term is calculated, backpropagation is performed to calculate the gradient, and the weight parameter values ​​are adjusted according to the optimizer update rules. The number of iterations is preset to a small value, such as 5 to 20 times. After completion, training is stopped without full retraining. The fine-tuned network model is used for temperature field prediction calculation, and a new temperature field prediction tensor is output. This tensor is denoted as the temperature field prediction optimization tensor.

[0057] In the above text, refer to Figure 1 This paper describes in detail a curing oven energy efficiency optimization method based on AI intelligent temperature field prediction according to embodiments of the present invention. Next, reference will be made to... Figure 4 This invention describes an AI-based intelligent temperature field prediction system for optimizing the energy efficiency of a curing oven according to an embodiment of the present invention.

[0058] The energy efficiency optimization system for curing ovens based on AI-powered intelligent temperature field prediction, according to embodiments of the present invention, addresses the technical problems in existing technologies, such as temperature field distortion and prediction lag under dynamic operating conditions, and the dependence of heating power configuration on experience and lack of global optimization. It achieves the technical effects of improving the accuracy of global temperature field uniformity control, heating power utilization efficiency, and adaptive control capability in the face of operating condition drift. Figure 4 As shown, the curing oven energy efficiency optimization system based on AI intelligent temperature field prediction includes: a data acquisition module 10, a configuration parameter determination module 20, a prediction error sequence construction module 30, and an energy efficiency optimization control module 40.

[0059] The data acquisition module 10 is used to perform multi-dimensional real-time acquisition based on multiple spatial locations identified by the curing oven, obtain a multi-dimensional time-series dataset, perform AI intelligent temperature field spatiotemporal distribution calculation, and obtain a temperature field prediction tensor; the configuration parameter determination module 20 is used to perform bidirectional target search based on the temperature field prediction tensor to determine the target heating power configuration parameters; the prediction error sequence construction module 30 is used to adjust the power of the curing oven in multiple regions according to the target heating power configuration parameters, continuously monitor synchronous operation data to perform error calculation, and construct a prediction error sequence; the energy efficiency optimization control module 40 is used to backtrack to the temperature field prediction tensor according to the prediction error sequence to perform incremental learning, construct an energy efficiency optimization strategy, and perform adaptive energy efficiency optimization control of the curing oven.

[0060] The specific configuration of the data acquisition module 10 will be described in detail below. The data acquisition module 10 further includes: incorporating the geometric features of the curing oven for three-dimensional analysis to identify the first density spatial position; locating the heating zone layout parameters of the curing oven for three-dimensional analysis to identify the second density spatial position; non-uniformly distributing the curing oven based on the first and second density spatial positions to determine multiple spatial positions; performing multi-dimensional real-time acquisition based on the multiple spatial positions to obtain a multi-dimensional dataset, performing spatiotemporal alignment to generate a multi-dimensional time-series dataset; constructing a spatiotemporal graph attention network, synchronizing the multi-dimensional time-series dataset to the spatiotemporal graph attention network for AI intelligent temperature field spatiotemporal distribution calculation, and generating the temperature field prediction tensor.

[0061] The specific configuration of the data acquisition module 10 will be described in detail below. The data acquisition module 10 further includes: performing heat conduction analysis on the curing oven based on the multiple spatial locations to generate heat conduction reachable paths; connecting the multiple spatial locations as nodes and the heat conduction reachable paths as edges to construct a curing oven spatial topology graph; performing spatial attention analysis on the curing oven spatial topology graph to calculate multiple attention coefficients; weighting and aggregating multiple nodes according to the multiple attention coefficients to obtain spatial coding features; performing temporal attention analysis on the curing oven spatial topology graph to calculate multiple local temporal features; performing self-attention long-range dependency analysis based on the multiple local temporal features to obtain temporal coding features; fusing the spatial coding features and the temporal coding features to generate initial predicted temperature field data; applying physical loss constraints based on the initial predicted temperature field to obtain a physical loss term; and jointly training the physical loss term and the initial predicted temperature field to construct the spatiotemporal graph attention network.

[0062] The specific configuration of the configuration parameter determination module 20 will be described in detail below. The configuration parameter determination module 20 further includes: extracting prediction period data based on the temperature field prediction tensor; performing power accumulation analysis according to the prediction period data to set an energy consumption target; performing temperature accumulation analysis according to the prediction period data to set a temperature field uniformity target; locating the starting point based on the prediction period data and extracting power to obtain initial parameters for heating power configuration; performing energy consumption priority search according to the energy consumption target based on the initial parameters for heating power configuration to obtain an energy consumption boundary solution; performing uniformity priority search according to the temperature field uniformity target based on the initial parameters for heating power configuration to obtain a uniformity boundary solution; constraining the energy consumption boundary solution and the uniformity boundary solution to generate multiple candidate power configuration parameters; traversing the multiple candidate power configuration parameters to perform energy consumption calculation to obtain multiple energy consumption values; traversing the multiple candidate power configuration parameters to perform temperature uniformity calculation to obtain multiple uniformity values, wherein the multiple energy consumption values, the multiple uniformity values, and the multiple candidate power configuration parameters have a corresponding relationship; and performing a comprehensive evaluation based on the multiple energy consumption values ​​and the multiple uniformity values ​​to determine the target heating power configuration parameters.

[0063] The specific configuration of the configuration parameter determination module 20 will be described in detail below. The configuration parameter determination module 20 further includes: using the initial heating power configuration parameters as a starting point, calculating multiple local energy consumption gradient parameters according to the energy consumption target and heating zone layout parameters; performing reverse updates according to the multiple local energy consumption gradient parameters, setting a first power update direction vector; introducing real-time power values, and superimposing the first power update direction vector with the real-time power values ​​to calculate trial power configuration parameters; performing power constraint determination based on the trial power configuration parameters and heating zone layout parameters; performing power amplitude analysis based on the multiple local energy consumption gradient parameters, setting power amplitude constraints, and based on the multiple local energy consumption gradient parameters... The system performs a change analysis on the local gradient parameters and sets a power change rate constraint. A feasibility analysis is then conducted based on the power amplitude constraint and the power change rate constraint to delineate the feasible region boundary. It is determined whether the trial power configuration parameters exceed the upper or lower power limit of the heating zone layout parameters. If the trial power configuration parameters exceed the upper or lower power limit of the heating zone layout parameters, the trial power configuration parameters are projected onto the feasible region boundary to obtain feasible power configuration parameters. These feasible power configuration parameters are used as real-time power configuration parameters, and projection correction is repeatedly performed until the energy consumption target decrease is lower than the convergence threshold, thus constructing the energy consumption boundary solution.

[0064] The specific configuration of the prediction error sequence construction module 30 will be described in detail below. The prediction error sequence construction module 30 further includes: parsing the target heating power configuration parameters according to the heating zone layout parameters to generate a time-series power command, wherein the time-series power command corresponds to the heating zone layout parameters; executing the time-series power command to synchronously adjust the heating zone power; continuously collecting measured temperature parameters at multiple spatial locations to construct multiple temperature field measurement matrices, wherein the multiple temperature field measurement matrices contain multiple control times; traversing the temperature field prediction tensor according to the multiple control times for matching to extract multiple predicted temperature values; performing spatiotemporal point-by-point comparison between the multiple predicted temperature values ​​and the multiple temperature field measurement matrices to calculate multiple temperature prediction residuals; aggregating the multiple temperature prediction residuals to construct the multiple residual vectors, arranging them in time sequence to construct the prediction error sequence.

[0065] The specific configuration of the prediction error sequence construction module 30 will be described in detail below. The prediction error sequence construction module 30 further includes: constructing a multi-scale sliding window, which includes a short-term sliding window and a long-term sliding window; performing statistical analysis on the prediction error sequence through the short-term sliding window to calculate the residual mean data; performing statistical analysis on the prediction error sequence through the long-term sliding window to calculate the residual trend slope; generating a performance drift signal when the residual mean data exceeds a preset noise threshold and the residual trend slope remains positive; and generating a sudden fault alarm signal when the residual mean data exceeds a preset impact threshold.

[0066] The specific configuration of the energy efficiency optimization control module 40 will be described in detail below. The energy efficiency optimization control module 40 further includes: performing multi-scale statistical monitoring on the prediction error sequence, identifying temperature field prediction performance degradation modes, and generating monitoring indicators; setting incremental learning trigger conditions, and generating incremental learning samples when the monitoring indicators meet the incremental learning trigger conditions; incrementally training the temperature field prediction tensor based on the incremental learning samples to generate an optimized temperature field prediction tensor; performing a bidirectional target search based on the optimized temperature field prediction tensor to obtain target heating power configuration parameters; applying the target heating power configuration parameters to the curing oven for multi-region power adjustment, and constructing an energy efficiency optimization strategy for adaptive energy efficiency optimization control of the curing oven.

[0067] The specific configuration of the energy efficiency optimization control module 40 will be described in detail below. The energy efficiency optimization control module 40 further includes: evaluating the contribution of the temperature field prediction tensor based on the incremental learning samples, and calculating multiple weighted importance evaluation values; applying weighted constraints to the temperature field prediction tensor based on the multiple weighted importance evaluation values ​​to construct an incremental training dataset; and performing a finite number of iterations of fine-tuning based on the incremental training dataset to generate the optimized temperature field prediction tensor.

[0068] The curing oven energy efficiency optimization system based on AI intelligent temperature field prediction provided in this embodiment of the invention can execute the curing oven energy efficiency optimization method based on AI intelligent temperature field prediction provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for optimizing the energy efficiency of a curing oven based on AI-powered intelligent temperature field prediction, characterized in that, The method includes: Based on the curing oven, multiple spatial locations are identified and multidimensional real-time data are collected to obtain a multidimensional time-series dataset. AI intelligent temperature field spatiotemporal distribution calculation is then performed to obtain the temperature field prediction tensor. A bidirectional target search is performed based on the temperature field prediction tensor to determine the target heating power configuration parameters; Based on the target heating power configuration parameters, the power of the curing oven is adjusted in multiple zones, and the synchronous operation data is continuously monitored to calculate the error and construct a predicted error sequence. Incremental learning is performed by backtracking the predicted error sequence to the temperature field prediction tensor to construct an energy efficiency optimization strategy for adaptive energy efficiency optimization control of the curing oven.

2. The method for optimizing the energy efficiency of a curing oven based on AI intelligent temperature field prediction as described in claim 1, characterized in that, Based on the curing oven's identification of multiple spatial locations, multidimensional real-time data acquisition is performed to obtain a multidimensional time-series dataset. This dataset is then used for AI-powered intelligent calculation of the spatiotemporal distribution of the temperature field, yielding a temperature field prediction tensor. The method includes: The geometric features of the curing oven are used for three-dimensional analysis to identify the spatial location of the first density. The layout parameters of the heating zone of the positioning curing oven are analyzed in three dimensions to identify the spatial location of the second density. Based on the first density spatial location and the second density spatial location, the curing oven is marked with non-uniform distribution to determine multiple spatial locations; Multidimensional real-time data acquisition is performed based on the multiple spatial locations to obtain a multidimensional dataset, which is then spatiotemporally aligned to generate a multidimensional time-series dataset. A spatiotemporal graph attention network is constructed, and the multidimensional time series dataset is synchronized to the spatiotemporal graph attention network to perform AI intelligent temperature field spatiotemporal distribution calculation, generating the temperature field prediction tensor.

3. The method for optimizing the energy efficiency of a curing oven based on AI intelligent temperature field prediction as described in claim 2, characterized in that, The construction process of spatiotemporal graph attention networks includes the following methods: Based on the multiple spatial locations, a heat conduction analysis of the curing oven is performed to generate heat conduction accessible paths; By connecting the multiple spatial locations as nodes and the heat conduction reachable paths as edges, a spatial topology map of the curing oven is constructed. Spatial attention analysis is performed based on the spatial topology diagram of the curing oven, and multiple attention coefficients are calculated. The spatial coding features are obtained by weighting and aggregating multiple nodes according to the multiple attention coefficients. Based on the spatial topology of the curing oven, temporal attention analysis is performed to calculate multiple local temporal features; Self-attention long-range dependency analysis is performed based on the multiple local temporal features to obtain temporal coding features; The spatial coding features are fused with the temporal coding features to generate initial predicted temperature field data; Based on the initial predicted temperature field, physical loss constraints are applied to obtain the physical loss term; The physical loss term and the initial predicted temperature field are jointly trained to construct the spatiotemporal graph attention network.

4. The method for optimizing the energy efficiency of a curing oven based on AI intelligent temperature field prediction as described in claim 2, characterized in that, Based on the temperature field prediction tensor, a bidirectional target search is performed to determine the target heating power configuration parameters. The method includes: Based on the temperature field prediction tensor, data for the prediction period is extracted. Power accumulation analysis is performed according to the data for the prediction period to set an energy consumption target. Temperature accumulation analysis is performed according to the data for the prediction period to set a temperature field uniformity target. Based on the predicted time period data, the starting point is located and power is extracted to obtain the initial parameters for heating power configuration. Based on the initial parameters configured with heating power, an energy consumption priority search is performed according to the energy consumption target to obtain the energy consumption boundary solution; Based on the initial parameters configured with the heating power, a uniformity-first search is performed according to the temperature field uniformity objective to obtain the uniformity boundary solution; Constraints are applied based on the energy consumption boundary solution and the uniformity boundary solution to generate multiple candidate power configuration parameters; Energy consumption is calculated by iterating through the multiple candidate power configuration parameters to obtain multiple energy consumption values. Temperature uniformity is calculated by iterating through the multiple candidate power configuration parameters to obtain multiple uniformity values. There is a corresponding relationship between the multiple energy consumption values, the multiple uniformity values ​​and the multiple candidate power configuration parameters. The target heating power configuration parameters are determined by comprehensively evaluating the multiple energy consumption values ​​and the multiple uniformity values.

5. The method for optimizing the energy efficiency of a curing oven based on AI intelligent temperature field prediction as described in claim 4, characterized in that, Based on the initial parameters of the heating power configuration, an energy-first search is performed according to the energy consumption target to obtain the energy consumption boundary solution. The method includes: Using the initial parameters of the heating power configuration as a starting point, and combining the energy consumption target with the heating zone layout parameters, multiple local energy consumption gradient parameters are obtained. The power update direction vector is set by performing reverse updates based on the multiple local energy consumption gradient parameters. Introducing real-time power values, and superimposing the first power update direction vector with the real-time power values, the trial power configuration parameters are calculated; Based on the aforementioned trial power configuration parameters, power constraints are determined according to the heating zone layout parameters. Power amplitude analysis is performed based on the multiple local energy consumption gradient parameters, and power amplitude constraints are set. Change analysis is performed based on the multiple local energy consumption gradient parameters, and power change rate constraints are set. A feasibility analysis is performed based on the power amplitude constraint and the power change rate constraint to delineate the boundary of the feasible region. Determine whether the trial power configuration parameters exceed the upper or lower power limit of the heating zone layout parameters. When the trial power configuration parameters exceed the upper or lower power limit of the heating zone layout parameters, project the trial power configuration parameters to the feasible domain boundary to obtain feasible power configuration parameters. The feasible power configuration parameters are used as real-time power configuration parameters, and projection correction is repeatedly performed until the energy consumption target decrease is lower than the convergence threshold, thus constructing the energy consumption boundary solution.

6. The method for optimizing the energy efficiency of a curing oven based on AI intelligent temperature field prediction as described in claim 1, characterized in that, The curing oven is adjusted in multiple zones according to the target heating power configuration parameters. Synchronous operation data is continuously monitored to calculate errors and construct a predicted error sequence. The method includes: Based on the target heating power configuration parameters, the heating zone layout parameters are parsed to generate a time-series power command, and the time-series power command has a corresponding relationship with the heating zone layout parameters. The time-series power command is executed to synchronously adjust the power of the heating zone, and the measured temperature parameters at the multiple spatial locations are continuously collected to construct multiple temperature field measurement matrices, which contain multiple control moments. The temperature field prediction tensor is matched according to the multiple control times to extract multiple predicted temperature values. Based on the spatiotemporal point-by-point comparison between the multiple predicted temperature values ​​and the multiple measured temperature field matrices, multiple temperature prediction residuals are calculated. The multiple temperature prediction residuals are aggregated, and the multiple residual vectors are arranged in time sequence to construct the prediction error sequence.

7. The method for optimizing the energy efficiency of a curing oven based on AI intelligent temperature field prediction as described in claim 6, characterized in that, After constructing the prediction error sequence, the methods include: Construct a multi-scale sliding window, which includes a short-term sliding window and a long-term sliding window; The prediction error sequence is statistically analyzed using the short-term sliding window to calculate the residual mean data. The prediction error sequence is statistically analyzed using the long-term sliding window to calculate the residual trend slope. When the residual mean data exceeds the preset noise threshold and the residual trend slope remains positive, a performance drift signal is generated. When the mean residual data exceeds the preset impact threshold, a sudden fault alarm signal is generated.

8. The method for optimizing the energy efficiency of a curing oven based on AI intelligent temperature field prediction as described in claim 1, characterized in that, Incremental learning is performed by backtracking from the predicted error sequence to the temperature field prediction tensor to construct an energy efficiency optimization strategy for adaptive energy efficiency optimization control of the curing oven. The method includes: Multi-scale statistical monitoring is performed on the prediction error sequence to identify temperature field prediction performance degradation patterns and generate monitoring indicators. Set incremental learning trigger conditions; when the monitoring indicator meets the incremental learning trigger conditions, an incremental learning sample is generated. Based on the incremental learning samples, the temperature field prediction tensor is incrementally trained to generate an optimized temperature field prediction tensor. Based on the temperature field prediction optimization tensor, a bidirectional target search is performed to obtain the target heating power configuration parameters; The target heating power configuration parameters are applied to the curing oven for multi-zone power adjustment, and an energy efficiency optimization strategy is constructed to perform adaptive energy efficiency optimization control of the curing oven.

9. The method for optimizing the energy efficiency of a curing oven based on AI intelligent temperature field prediction as described in claim 8, characterized in that, Incremental training is performed on the temperature field prediction tensor based on the incremental learning samples to generate an optimized temperature field prediction tensor. The method includes: The contribution of the temperature field prediction tensor is evaluated based on the incremental learning samples, and multiple weighted importance evaluation values ​​are calculated. The temperature field prediction tensor is weighted and constrained based on the multiple weight importance evaluation values ​​to construct an incremental training dataset; Based on the incremental training dataset, a finite number of iterations of fine-tuning are performed to generate the optimized tensor for temperature field prediction.

10. An energy efficiency optimization system for curing ovens based on AI-powered intelligent temperature field prediction, characterized in that: The system is used to implement the curing oven energy efficiency optimization method based on AI intelligent temperature field prediction as described in any one of claims 1 to 9, and the system includes: The data acquisition module is used to perform multi-dimensional real-time acquisition based on multiple spatial locations identified by the curing oven, obtain multi-dimensional time-series datasets, perform AI intelligent temperature field spatiotemporal distribution calculations, and obtain temperature field prediction tensors. The configuration parameter determination module is used to perform a bidirectional target search based on the temperature field prediction tensor to determine the target heating power configuration parameters. The prediction error sequence construction module is used to adjust the power of the curing oven in multiple regions according to the target heating power configuration parameters, continuously monitor synchronous operation data to calculate errors, and construct a prediction error sequence. The energy efficiency optimization control module is used to perform incremental learning by backtracking to the temperature field prediction tensor according to the prediction error sequence, and to construct an energy efficiency optimization strategy to perform adaptive energy efficiency optimization control of the curing oven.

Images