Clean energy bulk curing barn tobacco leaf curing thermal load analysis model establishment method

Abstract

The invention provides a method for establishing a tobacco leaf curing thermal load analysis model of a clean energy bulk curing barn, and belongs to the technical field of electric digital data processing. The volume density and the initial moisture content are determined by collecting temperature and power data of the bulk curing barn and combining the tobacco loading amount of the curing barn; a tobacco leaf dehydration and moisture dissipation physical process equation set is constructed as a constraint condition, and a thermodynamic constraint neural network model with tobacco leaf parameters, environment temperature and target temperature as input is established. According to the model, a multi-layer perceptron and a thermodynamic equation are fused, an adaptive temperature field attention mechanism is introduced, the number of attention heads is dynamically adjusted through the volume density, the initial water content and the target temperature, and accurate capture of mutual influence of temperature fields at different positions is achieved. Training to obtain a thermal load analysis model, and finally generating a thermal load prediction curve of different tobacco leaf types and tobacco loading amounts, thereby providing a scientific basis for energy conservation and consumption reduction of the bulk curing barn.

Description

Technical Field

[0001] The present invention belongs to the technical field of electronic digital data processing, and in particular relates to a method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn. Background Art

[0002] Tobacco curing is a critical step in tobacco processing. Traditional tobacco curing primarily relies on fossil fuels such as coal and diesel for heat, resulting in low energy efficiency and the generation of significant environmental pollutants. With increasing environmental protection requirements and the promotion of sustainable development concepts, the use of clean energy sources (such as electricity, natural gas, and heat pumps) in tobacco curing is becoming increasingly widespread. Modern intensive curing barns utilize a high-density arrangement to condition tobacco leaves, increasing the tobacco loading per unit area by over 40% compared to traditional curing barns, significantly improving space utilization efficiency.

[0003] However, existing clean, energy-intensive flue-curing barn heat load management relies primarily on empirical formulas or simple statistical models, lacking a deep understanding of the physical processes involved in tobacco leaf dehydration and moisture dissipation. These models typically treat the flue-curing barn as a simple thermal system, ignoring factors such as tobacco leaf packing density, initial moisture content, and dynamic changes during the curing process. This results in significant discrepancies between energy consumption predictions and actual operating conditions, and fails to meet the precise control requirements required under high-density tobacco packing conditions.

[0004] In the application of clean energy-intensive curing barns, problems such as the complexity of heat conduction paths and uneven changes in tobacco moisture content due to increased tobacco loading density are particularly prominent. Existing heat load prediction models are difficult to accurately describe the heat transfer and moisture migration processes under high-density tobacco loading conditions, resulting in energy waste and unstable baking quality. In other words, there is a technical problem in the existing technology of insufficient accuracy in heat load prediction for clean energy-intensive curing barns. Summary of the Invention

[0005] In view of this, the present invention provides a method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn, which can solve the technical problem of insufficient accuracy in heat load prediction for clean energy-intensive flue-curing barns in the prior art.

[0006] The present invention is implemented as follows: The present invention provides a method for establishing a heat load analysis model for tobacco leaf baking in a clean energy intensive flue-curing barn, including: collecting temperature data of the intensive flue-curing barn and operating power data of clean energy heating equipment; determining the volume density parameters of the intensive flue-curing barn and calculating the initial moisture content of tobacco leaves; constructing a set of physical process equations for tobacco leaf dehydration and moisture dissipation as physical constraints of a neural network; establishing a thermodynamic constraint neural network model with inputs of the initial moisture content of tobacco leaves, volume density parameters, ambient temperature, and target temperature, introducing an adaptive temperature field attention mechanism, and dynamically capturing the mutual influence of temperature fields at different positions in the flue-curing barn through a multi-head self-attention network; and using a measured data set to train a thermodynamic constraint neural network model as a heat load analysis model for tobacco leaf baking.

[0007] Among them, the volume density parameter refers to the ratio of the mass of tobacco leaves in unit volume to the total volume of the flue-curing room, which is used to calculate heat conduction efficiency and ventilation resistance.

[0008] Among them, the set of equations for the physical process of tobacco leaf dehydration and moisture dissipation includes the heat conduction equation, the moisture diffusion equation and the tissue structure change equation.

[0009] Among them, the heat conduction equation is used to describe the process of heat propagation in space in the flue-curing room. The input includes ambient temperature, heating equipment power distribution, and tobacco leaf distribution density, and the output is the temperature field distribution in the flue-curing room.

[0010] Among them, the moisture diffusion equation is used to describe the process of moisture migration outward from tobacco leaves. The input includes the initial moisture content of the tobacco leaves, the ambient air temperature and humidity, and the output is the rate of change of the moisture content of the tobacco leaves.

[0011] Among them, the tissue structure change equation is used to describe the changes in the cell structure of tobacco leaves during the drying process. The input includes the initial moisture content of the tobacco leaves, the temperature change rate, and the moisture content change rate. The output is the tobacco leaf tissue structure parameters.

[0012] Among them, the measured data set refers to a comprehensive data set collected through multiple flue-curing experiments, including the initial moisture content of tobacco leaves, volume density parameters, ambient temperature, target temperature, time series heat load value, and moisture content change rate, which is used to train the thermodynamic constraint neural network model.

[0013] Among them, the cross-validation method refers to a statistical method that divides the measured data set into a training set and a validation set, and tests the performance of the thermodynamic constraint neural network model under different segmentation methods to evaluate the generalization ability of the thermodynamic constraint neural network model.

[0014] Among them, the specific structure of the thermodynamic constraint neural network model is a hybrid architecture coupled with a multi-layer perceptron and thermodynamic equations, which includes an input layer, a hidden layer, a physical constraint layer and an output layer; among them, the number of neurons in the hidden layer is determined according to the volume density parameter, the initial moisture content of the tobacco leaves and the target temperature, and the physical constraint layer is embedded in the dehydration and moisture dissipation physical process equation group to ensure that the output of the thermodynamic constraint neural network model conforms to the physical laws.

[0015] Among them, the adaptive temperature field attention mechanism automatically adjusts the model's perception of the spatial relationship of the temperature field according to changes in smoke density, initial moisture content and target temperature. When the volume density is high, the number of attention heads is increased to capture more complex heat conduction paths. When the initial moisture content is low, the number of attention heads is reduced to adapt to a simpler dehydration process. When the target temperature is high, the number of attention heads is increased to handle stronger thermodynamic interactions.

[0016] The present invention combines traditional physical equations with deep learning technology by constructing a thermodynamic constraint neural network model to achieve an accurate description of the tobacco leaf baking process and a precise prediction of the heat load. The model takes the initial moisture content, bulk density parameters, ambient temperature and target temperature of the tobacco leaves as inputs, and embeds the dehydration and moisture dissipation physical process equations as constraints to ensure that the prediction results conform to the basic laws of thermodynamics. The adaptive temperature field attention mechanism introduced in the present invention can dynamically capture the mutual influence of the temperature fields at different locations in the baking room, and automatically adjust the model perception ability according to the changes in the tobacco loading density, initial moisture content and target temperature, effectively solving the problems of complex heat conduction paths and uneven moisture migration under high-density tobacco loading conditions. Through the multi-head self-attention network structure, this mechanism enables the model to optimize parameters for different baking room operating conditions, improve prediction accuracy, solve the technical problem of insufficient heat load prediction accuracy in clean energy-intensive baking rooms, provide more accurate energy consumption management and control strategies for the tobacco leaf baking process, and achieve the dual effects of improving baking quality stability and optimizing energy utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0019] like Figure 1 FIG. 1 is a flow chart of a method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn provided by the present invention. The method comprises the following steps:

[0020] S01. Collect real-time temperature data of temperature sensor distribution points in the intensive baking room and operating power data of clean energy heating equipment;

[0021] S02. Determine the bulk density parameter of the dense flue-curing barn according to the tobacco leaf loading amount and calculate the initial moisture content of the tobacco leaves;

[0022] S03, constructing a set of physical equations for tobacco leaf dehydration and moisture dissipation as physical constraints for the neural network;

[0023] S04. Determine the variation pattern of heat load at each stage of the baking process through time series analysis;

[0024] S05. Establishing a thermodynamic constraint neural network model whose inputs are initial moisture content of tobacco leaves, bulk density parameters, ambient temperature, and target temperature;

[0025] S06, setting the output of the thermodynamic constraint neural network model to the heat load value of the flue-curing barn and the rate of change of the moisture content of the tobacco leaves at different time nodes;

[0026] S07. Using the measured data set to train the thermodynamic constraint neural network model and perform parameter optimization;

[0027] S08. Using a cross-validation method to evaluate the prediction accuracy of the thermodynamic constraint neural network model and adjusting the structure of the thermodynamic constraint neural network model to obtain a tobacco leaf baking heat load analysis model;

[0028] S09. Optionally, the method further includes generating heat load prediction curves for different tobacco leaf types and tobacco loading amounts based on the tobacco leaf baking heat load analysis model.

[0029] Among them, the dense flue-curing room refers to a modern flue-curing room structure that adjusts tobacco leaves through a high-density arrangement. Compared with traditional flue-curing rooms, the tobacco loading capacity per unit area is increased by more than 40%.

[0030] Among them, the volume density parameter refers to the ratio of the mass of tobacco leaves in unit volume to the total volume of the flue-curing room, which is used to calculate heat conduction efficiency and ventilation resistance.

[0031] Among them, the dehydration and moisture dissipation physical process equations include heat conduction equation, water diffusion equation and tissue structure change equation;

[0032] The heat conduction equation is used to describe the heat propagation process in the flue-curing room. The input includes the ambient temperature, the power distribution of the heating equipment, and the density of the tobacco leaves. The output is the temperature field distribution in the flue-curing room.

[0033] The moisture diffusion equation is used to describe the process of moisture migration from tobacco leaves. The input includes the initial moisture content of the tobacco leaves, the ambient air temperature and humidity, and the output is the rate of change of the moisture content of the tobacco leaves.

[0034] The tissue structure change equation is used to describe the changes in the cell structure of tobacco leaves during the drying process. The input includes the initial moisture content of the tobacco leaves, the temperature change rate, and the moisture content change rate. The output is the tobacco leaf tissue structure parameters.

[0035] Among them, the measured data set refers to a comprehensive data set collected through multiple flue-curing experiments, including the initial moisture content of tobacco leaves, volume density parameters, ambient temperature, target temperature, time series heat load value, and moisture content change rate, which is used to train the thermodynamic constraint neural network model.

[0036] Among them, the cross-validation method refers to a statistical method of dividing the measured data set into a training set and a validation set, and testing the performance of the thermodynamic constraint neural network model under different segmentation methods to evaluate the generalization ability of the thermodynamic constraint neural network model.

[0037] Among them, the specific structure of the thermodynamic constraint neural network model is a hybrid architecture coupled with a multi-layer perceptron and thermodynamic equations, including an input layer, a hidden layer, a physical constraint layer and an output layer, wherein the number of neurons in the hidden layer is determined according to the volume density parameter, the initial moisture content of the tobacco leaves and the target temperature, and the physical constraint layer is embedded with a group of dehydration and moisture dissipation physical process equations to ensure that the output of the thermodynamic constraint neural network model conforms to the physical laws; the steps of establishing the training data set of the thermodynamic constraint neural network model specifically include collecting tobacco baking process data under different seasons, different varieties and different tobacco loading conditions from multiple dense baking rooms, recording the real-time temperature distribution, power consumption and moisture content changes of the entire baking process, dividing the data according to the time window and standardizing them, and filtering out abnormal values to form a training data set; the thermodynamic constraint neural network model is constructed by the following steps: The steps of training the thermodynamic constraint neural network model specifically include using a back-propagation algorithm with a physical loss function to train the thermodynamic constraint neural network model. The physical loss function combines the mean square error loss and the physical constraint loss to ensure that the thermodynamic constraint neural network model not only fits the measured data but also satisfies the laws of thermodynamics. The convergence process of the thermodynamic constraint neural network model is accelerated by an adaptive learning rate adjustment strategy, and the early stopping method is used to avoid overfitting problems. The thermodynamic constraint neural network model introduces an adaptive temperature field attention mechanism in the physical constraint layer. This mechanism dynamically captures the mutual influence of temperature fields at different positions in the flue-curing barn through a multi-head self-attention network, where the number of attention heads N is jointly determined by the volume density parameter (ρ), the initial moisture content of the tobacco leaves (M) and the target temperature (T), and the calculation formula is: Among them, ρ0, M0 and T0 are the reference volume density, reference moisture content and reference temperature respectively, N0 is the basic number of attention heads, α, β and γ are trainable weight exponent parameters, which control the influence of each physical quantity on the attention mechanism; the adaptive temperature field attention mechanism can automatically adjust the model's perception of the spatial relationship of the temperature field according to the changes in smoke density, initial moisture content and target temperature. When the volume density is high, the number of attention heads is increased to capture more complex heat conduction paths; when the initial moisture content is low, the number of attention heads is reduced to adapt to the simpler dehydration process; when the target temperature is high, the number of attention heads is increased to handle stronger thermodynamic interactions; at the same time, a receptive field of different scales is set for each attention head, and the size of the receptive field is inversely proportional to the combined function of these three key parameters, so that the model can achieve optimal heat load prediction accuracy under different baking room operating conditions.

[0038] The specific implementation of the above steps is described in detail below.

[0039] The specific implementation method of step S01 is to collect real-time temperature data by installing a high-precision temperature sensor network in a dense baking room, and at the same time connect the power monitoring system of the clean energy heating equipment to record the operating power data. The temperature sensor uses a PT100 platinum resistance temperature sensor with a measurement accuracy of ±0.1°C. It is arranged in a three-dimensional grid in the baking room, with a horizontal spacing of 1m and a vertical spacing of 0.5m to ensure coverage of the key thermodynamic areas in the baking room. The acquisition frequency is set to once every 60 seconds, and the data is summarized and transmitted to the data processing server through an industrial-grade data collector. The operating power data of the clean energy heating equipment is monitored in real time by a power analyzer, and the input power, output power and conversion efficiency data of the heating equipment are collected. The acquisition frequency is synchronized with the temperature data. The purpose of this step is to obtain time series data of the temperature field distribution and energy input in the baking room, providing basic data support for subsequent model construction.

[0040] The specific implementation of step S02 is to calculate the bulk density parameter based on the tobacco leaf loading mass and the geometric dimensions of the flue-curing barn, and calculate the initial moisture content of the tobacco leaf using the difference between the initial mass of the tobacco leaf and the mass after drying. The bulk density parameter is calculated by dividing the total mass of the tobacco leaf loading by the effective volume of the flue-curing barn, and the unit is kg / m 3 The interior of a standard intensive curing room is 8000mm long, 2700mm wide and 3500mm high. The fresh tobacco loading capacity is 4500-6000kg, and the bulk density parameter is usually 55-70kg / m 3The initial moisture content of tobacco leaves is determined using the sample drying method. Multiple samples are randomly drawn from the packed tobacco leaves, their initial mass measured, and then dried at 105°C to constant weight. The initial moisture content is calculated based on the mass difference using the formula: Moisture content = (initial mass - dry mass) / initial mass × 100%. The initial moisture content of tobacco leaves is typically between 82% and 88%. This step determines the key parameters that affect heat conduction efficiency and ventilation resistance, providing boundary conditions for constructing the physical process equations.

[0041] The specific implementation of step S03 constructs a set of equations for the physical process of tobacco leaf dehydration based on heat transfer and mass transfer theory. The heat conduction equation uses a three-dimensional unsteady-state heat conduction equation to describe the heat transfer process within the flue-curing barn, taking into account the combined effects of air convection and radiation. The thermal diffusion coefficient in the equation is dynamically calculated based on the tobacco leaf volume density parameter. Boundary conditions include the heat source distribution at the heating equipment location and heat loss from the flue-curing barn walls. The moisture diffusion equation uses Fick's second law to describe the migration of moisture within tobacco leaf tissue. The diffusion coefficient exhibits a nonlinear relationship with temperature and moisture content, increasing significantly at high temperatures, with a critical temperature of 55°C. The structural change equation is based on a tobacco leaf cell contraction model, describing the relationship between the cell wall elastic modulus and moisture content. When the moisture content falls below a critical value of 30%, the cell structure undergoes irreversible changes, affecting the subsequent moisture migration path. This step establishes a mathematical description that conforms to physical laws, which serves as a constraint for the neural network model to ensure that the model's predictions conform to the fundamental laws of thermodynamics.

[0042] The specific implementation method of step S04 is to perform statistical analysis and clustering processing on the collected time series data to identify the heat load characteristics of different stages in the baking process. First, the temperature and power time series data are smoothed with a sliding window, with a window size of 30 minutes and a step length of 5 minutes to eliminate short-term fluctuation noise. Then, the dynamic time warping algorithm is used to align the time series of multiple baking experiments to compensate for the time differences in the baking process of different batches of tobacco leaves. Then, the wavelet transform is used to extract the multi-scale features of the time series to identify the heat load mutation points and stable intervals. Finally, the hierarchical clustering algorithm is applied to divide the baking process into four typical stages: color fixing period, yellowing period, dry tendon period and drying period, and the heat load change pattern and statistical characteristics of each stage are extracted. The role of this step is to discover the inherent law of the change of heat load over time in the baking process, and provide stage feature constraints for the neural network model.

[0043] The specific implementation of step S05 is to construct a hybrid neural network model that couples a multi-layer perceptron with thermodynamic equations. The input layer receives four key parameters: initial tobacco leaf moisture content, bulk density parameter, ambient temperature, and target temperature. These parameters are standardized and then input into the hidden layer. The hidden layer adopts a fully connected structure with 3 to 5 layers. The number of neurons in each layer is dynamically determined based on the bulk density parameter range and the initial moisture content interval. The calculation formula refers to the adaptive neuron configuration algorithm. The activation function uses a rectified linear unit function to overcome the gradient vanishing problem of the traditional sigmoid function in deep networks. The physical constraint layer embeds the heat conduction equation, the water diffusion equation, and the tissue structure change equation into the network structure. Through differentiable physical constraint operators, the neural network's prediction results are consistent with physical laws. The purpose of this step is to establish a neural network model framework with physical interpretation capabilities, integrating the advantages of data-driven and mechanism-driven models.

[0044] The specific implementation of step S06 involves designing the output structure of the neural network model to enable it to predict the heat load value of the flue-curing barn and the rate of change of tobacco moisture content at different time points. The output layer is divided into two parallel branches: one branch predicts the time series of the heat load of the flue-curing barn, and the other branch predicts the time series of the rate of change of tobacco moisture content. The time nodes are unevenly distributed, with denser time nodes at the initial stage of curing and near key transition points, with intervals of 10 to 30 minutes; and less dense time nodes at intervals of 30 to 60 minutes during the stable curing phase. The heat load value is output in kW, representing the real-time power required to maintain the target temperature of the flue-curing barn. The moisture content change rate is output in % / h, representing the rate of decrease in tobacco moisture content per unit time. This step clarifies the output targets of the neural network model, enabling dual prediction of both flue-curing barn energy consumption and tobacco drying progress.

[0045] The specific implementation method of step S07 is to use the measured data set to train the thermodynamic constraint neural network model and optimize the model parameters. The training adopts a small batch stochastic gradient descent algorithm with a batch size of 64, an initial learning rate of 0.001, a cosine annealing learning rate adjustment strategy, and 500 to 1000 training rounds. The loss function is designed to be the weighted sum of the mean square error loss and the physical constraint loss. The physical constraint loss includes three parts: thermal balance constraint, mass conservation constraint, and the second law of thermodynamics constraint. The initial value of the constraint weight coefficient is set to 0.1 and is dynamically adjusted as the training progresses. The parameter optimization adopts an adaptive moment estimation optimizer, with the momentum parameter set to 0.9, the second-order moment parameter set to 0.999, and the stability factor set to 10 -8 Regularization uses a combination of weight decay and dropout, with a weight decay coefficient of 0.0001 and a dropout rate of 0.2. This step adjusts the neural network parameters in a data-driven manner, enabling the model to fit the measured data while satisfying physical constraints.

[0046] The specific implementation of step S08 involves using a k-fold cross-validation method to evaluate the prediction accuracy of the thermodynamic constraint neural network model and adjusting the model structure based on the validation results. Cross-validation employs a 5-fold scheme, randomly dividing the measured dataset into five subsets. Four subsets are used as training sets each time, and the remaining subset is used as the validation set. Performance metrics are averaged after five cycles. Evaluation metrics include the root mean square error (RMS) of the heat load prediction, the mean absolute percentage error (MAPE), and the RMS error of the moisture content change rate prediction, with acceptable thresholds set at 0.5kW, 5%, and 0.2% / h, respectively. Based on the cross-validation results, the neural network structure is adjusted, including increasing or decreasing the number of hidden layers, adjusting the number of neurons, modifying the activation function type, and adjusting the physical constraint weights. A hyperparameter optimization strategy combining grid search and Bayesian optimization is used to search for the optimal model configuration within the preset parameter space. This step evaluates the model's generalization ability, avoids overfitting, and ensures good predictive performance on unseen data. Ultimately, training is completed to produce a tobacco flue-curing heat load analysis model.

[0047] Step S09 is an optional step, and its specific implementation method is to use the trained tobacco baking heat load analysis model to generate heat load prediction curves for different tobacco types and loading conditions. First, a typical tobacco variety library is established, which contains the characteristic parameters of major tobacco types such as flue-cured tobacco, sun-cured tobacco, and burley tobacco. The initial moisture content range and physical property parameters are set for each tobacco leaf. Then, a loading gradient sequence is designed to cover the range of 70% to 120% of the standard loading amount, corresponding to a volume density parameter of 16 to 40 kg / m 3 . Then, the ambient temperature conditions are set, including three typical meteorological conditions of spring, summer, and autumn, with an ambient temperature range of 5 to 35°C. Finally, for each combination of conditions, the conditions are input into the trained tobacco leaf baking heat load analysis model to generate a heat load prediction curve and moisture content change curve for the complete baking cycle, with a curve resolution of 10 minutes / point. The purpose of this step is to apply the trained model to actual production scenarios, provide a prediction basis for energy demand under different tobacco leaf types and loading conditions, and guide the rational configuration and optimized operation of clean energy equipment.

[0048] The detailed structure of the thermodynamic constraint neural network model is as follows: the input layer contains 4 neurons, corresponding to the initial moisture content of tobacco leaves, volume density parameter, ambient temperature and target temperature respectively; the first hidden layer contains 64 neurons, using the rectified linear unit activation function; the second hidden layer contains 128 neurons, also using the rectified linear unit activation function; the third hidden layer contains 256 neurons, using the rectified linear unit activation function; after the third hidden layer, an adaptive temperature field attention mechanism is introduced, and the number of attention heads is determined by the volume density parameter, the initial moisture content of tobacco leaves and the target temperature. The basic number of attention heads is 8, and when the volume When the density parameter is high, the number of attention heads is increased to capture complex heat conduction paths. When the initial moisture content is low, the number of attention heads is reduced to adapt to the simplified dehydration process. When the target temperature is high, the number of attention heads is increased to handle strong thermodynamic interactions. The physical constraint layer converts the heat conduction equation, moisture diffusion equation, and tissue structure change equation into differentiable constraint operators and embeds them into the neural network. The output layer is divided into two parallel branches. The heat load prediction branch contains 48 neurons, corresponding to the heat load values ​​at 48 time nodes, and the moisture content change rate prediction branch also contains 48 neurons, corresponding to the moisture content change rate at the same time nodes. The detailed steps for establishing the training dataset include: first, collecting baking data for the spring, summer and autumn seasons from densely populated tobacco barns in 10 different regions, and collecting 5 to 8 batches of tobacco baking process data of different varieties and loading amounts in each barn; then preprocessing the collected raw data, including outlier detection and elimination, missing value interpolation, signal denoising and time alignment; then splitting the processed data into 10-minute time windows to form time series feature vectors; then standardizing the feature vectors to make the magnitudes of each feature consistent; then using the principal component analysis method to reduce the data dimension, retaining the principal components with an explained variance ratio of 95%; finally, dividing the processed dataset into training set, validation set and test set in a ratio of 7:2:1. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and early stopping strategy implementation, and the test set is used for final model performance evaluation.

[0049] Specifically, the present invention is based on a physical information-guided neural network architecture design. By combining thermodynamic principles with deep learning, this technology constructs a physically constrained prediction model. The core of this approach is to explicitly integrate physical processes such as heat conduction, moisture diffusion, and structural changes during tobacco leaf curing into the neural network structure, ensuring that the model's predictions adhere to the laws of thermodynamics.

[0050] To build the model, real-time data, such as temperature distribution and heating equipment power, was first collected experimentally within the dense flue-curing barn. This data was then combined with the tobacco leaf loading volume to calculate bulk density parameters and initial moisture content, establishing a comprehensive data foundation. Subsequently, a set of physical process equations, including heat conduction, moisture diffusion, and structural change equations, was constructed to serve as constraints for the neural network. These equations describe heat transfer within the flue-curing barn, moisture migration in the tobacco leaves, and changes in cellular structure, respectively, effectively representing the physical and chemical changes occurring throughout the flue-curing process.

[0051] The key innovation of this invention lies in the hybrid architecture that couples a multi-layer perceptron with thermodynamic equations. The physical constraint layer embeds the equations for the dehydration and moisture dissipation physics process, ensuring that predictions conform to physical laws. In particular, the introduction of an adaptive temperature field attention mechanism dynamically adjusts the number of attention heads, enabling the model to automatically adjust its perception of spatial relationships in the temperature field based on changes in smoke density, initial moisture content, and target temperature. When the volume density is high, the number of attention heads is increased to capture more complex heat conduction paths; when the initial moisture content is low, the number of attention heads is reduced to accommodate the simpler dehydration and moisture dissipation process; and when the target temperature is high, the number of attention heads is increased to handle stronger thermodynamic interactions.

[0052] The model training process utilizes a backpropagation algorithm with a physical loss function. This loss function combines mean squared error loss with physical constraint loss, enabling the model to fit measured data while also satisfying the laws of thermodynamics. Adaptive learning rate adjustment and early stopping accelerate model convergence while avoiding overfitting. Ultimately, an analytical model capable of accurately predicting thermal load changes under different tobacco leaf types and loading conditions was established.

[0053] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0054] The specific implementation method of step S01 is to collect real-time temperature data by installing a high-precision temperature sensor network in a dense baking room, and at the same time connect the power monitoring system of the clean energy heating equipment to record the operating power data. The temperature sensor uses a PT100 platinum resistance temperature sensor with a measurement accuracy of ±0.1°C. It is arranged in a three-dimensional grid in the baking room, with a horizontal spacing of 1m and a vertical spacing of 0.5m to ensure coverage of the key thermodynamic areas in the baking room. The acquisition frequency is set to once every 60 seconds, and is summarized and transmitted to the data processing server through an industrial-grade data collector. The operating power data of the clean energy heating equipment is monitored in real time by a power analyzer, and the input power, output power and conversion efficiency data of the heating equipment are collected. The acquisition frequency is synchronized with the temperature data. The purpose of this step is to obtain the time series data of the temperature field distribution and energy input in the baking room, and provide basic data support for subsequent model construction. During the sensor data acquisition process, the temperature matrix Ti，j，k (t) represents the temperature value at the time t and the position coordinate (i, j, k), where i, j, k represent the position index of the sensor in the three-dimensional space of the baking room; the heating equipment power matrix P m (t) represents the real-time power value of the mth heating device at time t. The power data statistical method is:

[0055] P m (t) = V m (t)·I m (t)·cosφ m (t);

[0056] Where V m (t) is the voltage value of the mth heating device at time t, in V; I m (t) is the current value of the mth heating device at time t, in A; cosφ m (t) is the dimensionless power factor of the mth heating device at time t. The power data must be collected with an accuracy of ±0.5% and a sampling frequency of no less than 1 Hz.

[0057] The specific implementation of step S02 is to calculate the bulk density parameter based on the tobacco leaf filling mass and the geometric dimensions of the flue-curing barn, and calculate the initial moisture content of the tobacco leaf using the difference between the initial mass and the mass after drying. The bulk density parameter calculation formula is:

[0058]

[0059] Where ρ is the bulk density parameter, unit is kg / m 3 ;M total is the total mass of tobacco leaves filled, in kg; V eff The effective volume of the baking room, in m 3 The initial moisture content of tobacco leaves was determined by the sample drying method, and the calculation formula is:

[0060]

[0061] Where, M is the initial moisture content of tobacco leaves, in %; M initial is the initial mass of tobacco leaves, in kg; M dry is the mass of the dried tobacco leaf, in kg. The initial moisture content of the tobacco leaf is typically between 82% and 88%. This step determines the key parameters that influence heat transfer efficiency and ventilation resistance, providing boundary conditions for constructing the physical process equations.

[0062] The specific implementation of step S03 is to construct a set of equations for the physical process of tobacco leaf dehydration and moisture dissipation based on heat transfer and mass transfer theory. The heat conduction equation uses a three-dimensional unsteady-state heat conduction equation to describe the heat transfer process in the flue-curing barn, and its mathematical expression is:

[0063]

[0064] Where, ρ a is the air density in kg / m 3 ;c p is the specific heat capacity of air, in J / (kg·K); T is the temperature field, in K; t is the time, in s; is the gradient operator; k is the thermal conductivity, in W / (m·K); q is the heat source term, which represents the heat input by the heating equipment, in W / m 3 ; h is the convection heat transfer coefficient, unit is W / (m 2 ·K); T s is the surface temperature of the tobacco leaf, in K. The thermal conductivity coefficient k is related to the bulk density parameter ρ, and the relationship is:

[0065] k=k0·(1-α·ρ);

[0066] Where k0 is the reference thermal conductivity, ranging from 0.026 to 0.030 W / (m·K); α is the correction coefficient, ranging from 0.01 to 0.02 m 3 / kg. The water diffusion equation uses Fick's second law to describe the migration process of water in tobacco leaf tissue. Its mathematical expression is:

[0067]

[0068] Where C is the moisture concentration in tobacco leaves, unit is kg / m 3 ; D is the water diffusion coefficient, unit is m 2 / s; S is the moisture source term, representing the evaporation effect, and the unit is kg / (m 3 ·s). The relationship between the water diffusion coefficient D, temperature T and moisture content M is:

[0069]

[0070] Where D0 is the reference diffusion coefficient, which is 10 -6 ~10 -5 m 2 / s;E a is the diffusion activation energy, ranging from 20 to 30 kJ / mol; R is the gas constant, ranging from 8.314 J / (mol·K); a and b are empirical coefficients, ranging from approximately 0.05 to 0.08 and 0.1 to 0.2, respectively. The tissue structure change equation describes the changes in the cell structure of tobacco leaves during the drying process, and its mathematical expression is:

[0071]

[0072] Where S is the organizational structure parameter, dimensionless, and its initial value is 1; k s is the structural change rate constant, in s -1 ; f(M) is the moisture content influence function; g(T) is the temperature influence function. The moisture content influence function and temperature influence function are expressed as:

[0073]

[0074] Where M c is the critical moisture content, which is 30%; T c is the critical temperature, set at 55°C; β and γ are adjustment coefficients, ranging from approximately 0.1 to 0.2 and 0.05 to 0.1, respectively. This step establishes a mathematical description that conforms to physical laws, serving as constraints for the neural network model and ensuring that the model's predictions conform to the fundamental laws of thermodynamics.

[0075] The specific implementation of step S04 is to perform statistical analysis and clustering on the collected time series data to identify the heat load characteristics of different stages in the baking process. First, the temperature and power time series data are smoothed by sliding windows with a window size of 30 minutes and a step size of 5 minutes. The mathematical expression for the smoothing process is:

[0076]

[0077] Where, is the smoothed data at time t; X(i) is the original data at time i; w is the window size, in minutes. Then the dynamic time warping algorithm is used to align the time series of multiple baking experiments, and its objective function is:

[0078]

[0079] Where DTW(A, B) is the dynamic time warping distance between sequence A and sequence B; w is the alignment path; d(w k ) is the distance metric of the kth point on the path; K is the path length. Then, wavelet transform is used to extract the multi-scale features of the time series. The wavelet transform calculation formula is:

[0080]

[0081] Where W f (a, b) are the wavelet coefficients of the signal f(t); a is the scale parameter; b is the translation parameter; ψ is the wavelet basis function; ψ * is the complex conjugate of the wavelet basis function. The heat load mutation point detection adopts the cumulative sum control chart method, and its statistics are:

[0082] S i=max(0, S i-1 +(X i -μ)-kσ);

[0083] Where S i is the cumulative sum statistic; X i is the heat load value at time i; μ is the mean heat load; σ is the standard deviation of the heat load; and k is a sensitivity parameter ranging from 0.3 to 0.5. Finally, a hierarchical clustering algorithm is used to divide the baking process into four typical stages, using the dynamic time warping distance as the cluster distance metric. This step aims to discover the inherent patterns of heat load variation over time during the baking process, providing stage-specific feature constraints for the neural network model.

[0084] The specific implementation of step S05 is to construct a hybrid architecture neural network model that couples a multi-layer perceptron with thermodynamic equations. The input layer receives the initial moisture content M0 of the tobacco leaf, the volume density parameter ρ, the ambient temperature T env , target temperature T target The four key parameters are input into the hidden layer after being standardized. The standardization formula is:

[0085]

[0086] Where, is the i-th input parameter after standardization; x i is the original input parameter value; μ i is the mean of parameter i; σ i is the standard deviation of parameter i. The hidden layer adopts a fully connected structure with 3 to 5 layers. The number of neurons in each layer is dynamically determined according to the range of volume density parameter value and initial moisture content interval. The calculation formula is:

[0087]

[0088] Where N l is the number of neurons in the lth hidden layer; N0 is the number of basic neurons, which is 32; ρ0 is the reference volume density, which is 25 kg / m 3 ;M 00 is the reference moisture content, which is 85%; α and β are weight indices, which are 0.5 and 0.3 respectively; l is the hidden layer index, which starts from 1. The forward propagation calculation formula of the hidden layer is:

[0089] h l =σ(W l ·h l-1 +b l );

[0090] Where h lis the output vector of the lth hidden layer; σ is the activation function, using the rectified linear unit function ReLU(x)=max(0,x); W l is the weight matrix of the lth layer; b l is the bias vector of the lth layer; h l-1 is the output vector of the previous layer. When l = 1, h0 is the normalized input vector. The physical constraint layer embeds the heat conduction equation, water diffusion equation, and tissue structure change equation into the network structure and implements them through differentiable physical constraint operators. The mathematical expression is:

[0091]

[0092] Where, is the total physical constraint loss; is the heat conduction constraint loss; To constrain losses for water diffusion; is the structural change constraint loss; λ1, λ2, and λ3 are weight coefficients, which are approximately 0.5, 0.3, and 0.2 respectively. The calculation formula for the heat conduction constraint loss is:

[0093]

[0094] Where N is the number of sampling points; T i is the temperature prediction value of the i-th sampling point; Q i is the heat source term of the i-th sampling point; T s，i is the tobacco leaf surface temperature at the i-th sampling point. Similarly, the moisture diffusion constraint loss and the structural change constraint loss are calculated using the sum of squared residuals of the corresponding physical equations. This step establishes a neural network model framework with physical interpretation capabilities, integrating the advantages of data-driven and mechanism-driven approaches.

[0095] The specific implementation of step S06 is to design the output structure of the neural network model so that it can predict the heat load value of the flue-curing barn and the rate of change of the moisture content of the tobacco leaves at different time nodes. The output layer is divided into two parallel branches, one branch predicts the time series of the heat load of the flue-curing barn, and the other branch predicts the time series of the rate of change of the moisture content of the tobacco leaves. The time nodes are non-uniformly distributed, with denser time nodes set at the initial stage of baking and near the key transition points, with a time interval of 10 to 30 minutes; and sparser time nodes set at the stable stage of baking, with a time interval of 30 to 60 minutes. The calculation formula of the heat load prediction branch is:

[0096] P(t j )=W P ·h L +b P ;

[0097] Where, P(t j) is the time node t j The heat load forecast value is in kW; P is the heat load output layer weight matrix; h L is the output vector of the last hidden layer; b P is the thermal load output layer bias vector. The calculation formula for the moisture content change rate prediction branch is:

[0098]

[0099] Where, is the time node t j The predicted value of the water content change rate, in % / h; W M is the weight matrix of the output layer of water content change rate; b M The function of this step is to clarify the output target of the neural network model and realize the dual prediction of energy consumption of the flue-curing barn and the progress of tobacco leaf drying.

[0100] The specific implementation of step S07 is to use the measured data set to train the thermodynamic constraint neural network model and optimize the model parameters. The training adopts the small batch stochastic gradient descent algorithm with a batch size of 64, an initial learning rate of 0.001, and a cosine annealing learning rate adjustment strategy. Its mathematical expression is:

[0101]

[0102] Where η t is the learning rate of the tth iteration; η min is the minimum learning rate, which is set to 10 -5 η max is the maximum learning rate, which is set to 10 -3 ; T is the total number of iterations, ranging from 500 to 1000. The loss function is designed as the weighted sum of the mean square error loss and the physical constraint loss, and its mathematical expression is:

[0103]

[0104] Where, is the total loss function; is the mean square error loss; is the physical constraint loss; ω is the weight coefficient, with an initial value of 0.1, which is dynamically adjusted as the training progresses. The formula for calculating the mean square error loss is:

[0105]

[0106] Where, P true (t i ) is the time node t i The actual heat load value; is the time node t i The actual moisture content change rate is λ, which is the balance coefficient and is set to 0.5. The adaptive moment estimation optimizer is used for parameter optimization, and its parameter update rule is:

[0107] m t =β1·m t-1 +(1-β1)·g t ;

[0108]

[0109] Where m t and v t are the first-order moment estimate and the second-order moment estimate respectively; g t is the current gradient; β1 and β2 are the momentum decay rate and the second-order moment decay rate, which are 0.9 and 0.999 respectively; and are the first-order moment estimate and the second-order moment estimate after bias correction respectively; θ t is the model parameter after the tth iteration; ∈ is the stability factor, which is 10 -8 Regularization adopts a strategy combining weight decay and dropout, and the weight decay correction term is:

[0110] Δθ decay =-λ decay θ;

[0111] Where λ decay is the weight decay coefficient, set to 0.0001. The dropout method randomly sets the outputs of some neurons in the hidden layer to zero during training, with a dropout rate of 0.2. This step adjusts the neural network parameters in a data-driven manner, ensuring that the model fits the measured data while meeting physical constraints.

[0112] The specific implementation of step S08 is to use the k-fold cross-validation method to evaluate the prediction accuracy of the thermodynamic constraint neural network model, and adjust the model structure according to the validation results to obtain the tobacco baking heat load analysis model. The cross-validation adopts a 5-fold scheme, randomly dividing the measured data set into 5 subsets, using 4 subsets as training sets each time and the remaining 1 subset as validation set. After 5 cycles, the average performance index is obtained. The evaluation indicators include the root mean square error RMSE of the heat load prediction P , mean absolute percentage error MAPE P and the root mean square error RMSE of the water content change rate prediction M , the calculation formulas are:

[0113]

[0114] Where N is the number of evaluation time nodes. The acceptable thresholds are set as RMSE P ≤0.5kW, MAPE P ≤5% and rMSE M ≤0.2% / h. Adjust the neural network structure based on the cross-validation results, including increasing or decreasing the number of hidden layers, adjusting the number of neurons, modifying the activation function type, adjusting the physical constraint weights, etc. A hyperparameter optimization strategy combining grid search and Bayesian optimization is used to search for the optimal model configuration in the preset parameter space. The parameter space of the grid search is defined as:

[0115] S={(n layers , n base , α, β, γ, λ1, λ2, λ3, ω)∣n layers ∈{3, 4, 5}, n base ∈

[0116] {16, 32, 64}, α∈{0.3, 0.5, 0.7}, β∈{0.2, 0.3, 0.4}, γ∈{0.1, 0.2, 0.3}, λ1∈{0.4, 0.5, 0.6}, λ2∈{0.2, 0.3, 0.4}, λ3∈{0.1, 0.2, 0.3}, ω∈{0.05, 0.1, 0.2}};

[0117] Where n layers is the number of hidden layers; n base is the number of basic neurons; α, β, and γ are the exponential parameters in the formula for calculating the number of neurons; λ1, λ2, and λ3 are the weight coefficients of each part of the physical constraint loss; and ω is the weight coefficient of the physical constraint loss in the total loss function. For the optimal model structure, an adaptive temperature field attention mechanism is further introduced, and the formula for calculating the number of attention heads is:

[0118]

[0119] Where N heads is the number of attention heads; N base is the number of basic attention heads, which is 8; ρ0, M 00 and T0 are reference bulk density, reference moisture content and reference temperature, respectively, and their values are 25 kg / m 3 , 85% and 60℃; α, β and γ are trainable weight index parameters, which respectively represent the influence of volume density, initial moisture content and target temperature on the attention mechanism, with initial values of 0.5, -0.3 and 0.2 respectively. The calculation formula of the attention mechanism is:

[0120]

[0121] Where Q, K and V are query matrix, key matrix and value matrix respectively, which are generated by linear transformation of hidden layer features; d k is the dimension of the key vector. The calculation formula of the multi-head attention mechanism is:

[0122] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O ;

[0123]

[0124] Where h is the number of attention heads; is the parameter matrix of the i-th attention head; W O The output projection matrix is used to evaluate the generalization ability of the model, avoid overfitting, ensure that the model still has good prediction performance on unseen data, and form the final tobacco leaf baking heat load analysis model.

[0125] The specific implementation of step S09 is to use the trained tobacco leaf baking heat load analysis model to generate heat load prediction curves for different tobacco leaf types and tobacco loading conditions. First, a typical tobacco leaf variety library is established, which contains the characteristic parameters of the main tobacco leaf types such as flue-cured tobacco. The characteristic parameters of each tobacco leaf include the initial moisture content range M 0，min ~M 0，max and physical characteristic parameter vector The mathematical expression of the physical characteristic parameter vector is:

[0126]

[0127] In the formula, φ1 is the specific heat capacity coefficient, the unit is J / (kg·K), and the value range is 1500~2000; φ2 is the thermal conductivity coefficient, the unit is W / (m·K), and the value range is 0.2~0.5; φ3 is the water diffusion coefficient correction factor, dimensionless, and the value range is 0.8~1.2; φ4 is the tissue density coefficient, dimensionless, and the value range is 0.9~1.1; φ5 is the evaporation enthalpy correction factor, dimensionless, and the value range is 0.95~1.05. Then, a smoke loading gradient sequence is designed to cover the range of 70%~120% of the standard smoke loading, and the corresponding volume density parameter ρ ranges from 16~40kg / m 3 , the mathematical expression of the gradient sequence is:

[0128]

[0129] Where, ρ i is the value of the i-th volume density parameter; ρ min is the minimum bulk density parameter, which is 16 kg / m 3ρ max is the maximum volume density parameter, which is 40kg / m 3 ; n is the gradient sequence length, ranging from 10 to 15. Then set the ambient temperature condition T env , including three typical meteorological conditions of spring, summer and autumn, with an ambient temperature range of 5 to 35°C. target , set to a typical baking temperature of 68°C. Finally, for each combination of conditions, the parameter quaternary (M0, ρ, T env , T target ) is input into the trained neural network model to generate the heat load prediction curve P(t) and moisture content change curve M(t) for the complete baking cycle, with a curve resolution of 10 minutes / point. The calculation formula for the moisture content change curve is:

[0130]

[0131] Where, M(t j ) is the time node t j The predicted value of moisture content; is the time node t i Predicted value of the rate of change of moisture content; Δt i is the time step, in h. The cumulative heat energy consumption calculation formula is:

[0132]

[0133] Where, E(t j ) is the time node t j The cumulative heat energy consumption is in kWh; P(t i ) is the time node t i The heat load prediction value is in kW. The thermal efficiency calculation formula is:

[0134]

[0135] Where, η(t j ) is the time node t j Thermal efficiency, dimensionless; m is the tobacco packing mass, unit is kg; L v is the latent heat of vaporization of water, which is 2257 kJ / kg; M0 is the initial water content, in %; M(t j ) is the time node t j The predicted value of moisture content is expressed in %. For the prediction results under different combination conditions, three-dimensional visualization technology is used to display the relationship between heat load, moisture content change rate, time, and volume density parameters. The heat load surface equation is:

[0136]

[0137] Where P is the heat load value, unit is kW; t is the time, unit is h; ρ is the volume density parameter, unit is kg / m 3 ;a ij is the fitting coefficient; n t and n ρ The polynomial orders of the time and volume density parameters are respectively, and are usually 3 to 5. The purpose of this step is to apply the trained model to actual production scenarios, provide a prediction basis for energy demand under different tobacco leaf types and loading conditions, and guide the rational configuration and optimized operation of clean energy equipment.

[0138] The detailed structure of the thermodynamic constraint neural network model is as follows: the input layer contains 4 neurons, corresponding to the initial moisture content M0 of tobacco leaves, the volume density parameter ρ, the ambient temperature T env and target temperature T target The first hidden layer contains 64 neurons and uses the rectified linear unit activation function; the second hidden layer contains 128 neurons and also uses the rectified linear unit activation function; the third hidden layer contains 256 neurons and uses the rectified linear unit activation function; the adaptive temperature field attention mechanism is introduced after the third hidden layer, and the number of attention heads N heads The volume density parameter ρ, the initial moisture content of tobacco leaves M0 and the target temperature T target The three parameters are jointly determined and the calculation formula is:

[0139]

[0140] Where N base is the number of basic attention heads, which is 8; ρ0 is the reference volume density, which is 25 kg / m 3 ;M 00 is the reference moisture content, which is 85%; T0 is the reference temperature, which is 60°C; α, β, and γ are trainable weight index parameters that control the influence of each physical quantity on the attention mechanism, with initial values of 0.5, -0.3, and 0.2, respectively. The attention calculation formula is:

[0141]

[0142] Where a ij is the attention weight of position i to position j; e ij is the attention score between position i and position j; W Q 、W K and W V are the query transformation matrix, key transformation matrix and value transformation matrix respectively; h i and h j are the hidden state vectors at position i and position j respectively; d is the dimension of the hidden state vector; ci is the context vector at position i; n is the sequence length. The physical constraint layer transforms the heat conduction equation, moisture diffusion equation, and tissue structure change equation into differentiable constraint operators and embeds them into the neural network. The output layer is divided into two parallel branches. The heat load prediction branch contains 48 neurons, corresponding to the heat load values at 48 time points, and the moisture content change rate prediction branch also contains 48 neurons, corresponding to the moisture content change rate at the same time points.

[0143] The detailed steps of establishing the training data set include: first, collecting baking data in spring, summer and autumn from densely populated tobacco barns in 10 different regions. Each barn collects 5 to 8 batches of tobacco baking process data of different varieties and loading amounts. Specifically, the interval for collecting data during each batch of baking process is 1 minute, and the collection duration is the entire baking cycle, which is usually 5 to 7 days. The collection parameters include: the temperature T at different locations in the baking barn; i，j，k (t), real-time power P of heating equipment m (t), real-time moisture content M of tobacco leaf samples s (t), humidity in the drying room H(t), wind speed of the ventilation system v(t), etc. Then, the collected raw data is preprocessed, including outlier detection and elimination, missing value interpolation, signal denoising and time alignment; outlier detection uses a method based on the interquartile range, and the mathematical expression is:

[0144] [Q1-k·IQR, Q3+k·IQR];

[0145] Where Q1 and Q3 are the first and third quartiles of the data, respectively; IQR = Q3 - Q1 is the interquartile range; k is the expansion coefficient, which is set to 1.5. Missing values are interpolated using cubic spline interpolation, signal denoising using Savitzky-Golay filter, and time alignment using dynamic time warping. The processed data is then segmented into 10-minute time windows to form a time series feature vector. The mathematical expression is:

[0146]

[0147] Where, is the characteristic vector at time t; M0 is the initial moisture content; ρ is the bulk density parameter; T env (t) is the ambient temperature; T target (t) is the target temperature; P(t) is the heat load; is the rate of change of moisture content. Then the eigenvectors are normalized to make the magnitudes of each feature consistent. The normalization formula is:

[0148]

[0149] Where, is the normalized feature vector; is the mean vector of each feature; is the standard deviation vector of each feature. Principal component analysis was then used to reduce the data dimension, retaining the principal components that explained 95% of the variance. Finally, the processed dataset was divided into training, validation, and test sets in a ratio of 7:2:1. The training set was used for model parameter learning, the validation set for hyperparameter tuning and early stopping, and the test set for final model performance evaluation.

[0150] In this embodiment, the principles of physics and deep learning technology are combined to achieve accurate prediction of the heat load of the flue-curing room. In particular, in step S08, a cross-validation method is introduced to evaluate the prediction accuracy of the model, and a better tobacco leaf baking heat load analysis model is obtained by adjusting the model structure; in step S09, the trained model is used to generate heat load prediction curves under different tobacco leaf types and loading conditions, providing a scientific basis for energy demand prediction for actual production. The core of the model is the introduction of an adaptive temperature field attention mechanism, which can dynamically adjust the number of attention heads according to the volume density parameters, initial moisture content and target temperature, so as to more accurately capture complex heat conduction paths and thermodynamic interactions. The complete implementation method ensures that the model prediction results are consistent with both actual data and the basic laws of thermodynamics through strict mathematical formulas and physical constraints, providing strong support for the application of clean energy in the field of tobacco leaf baking.

[0151] In order to better understand and implement the present invention, the following provides a specific application scenario of the present invention, Example 2: Researchers conducted a practical application of the heat load analysis model for tobacco leaf baking in a clean energy intensive flue-curing barn in a tobacco area. This example is for a standard intensive flue-curing barn with an effective volume of 107m 3 Data is collected by installing a network of high-precision temperature sensors. These sensors use PT100 platinum resistance temperature sensors with a measurement accuracy of ±0.1°C. Within the baking room, 86 temperature sensors are arranged in a three-dimensional grid, with horizontal spacing of 1 meter and vertical spacing of 0.5 meters. The clean energy heating equipment utilizes an electric heating system with a rated power of 60kW, capable of real-time adjustment of output power based on control commands.

[0152] In this example, tobacco leaves of K326 variety were selected for the baking experiment. The length of the baking room was 8000 mm, the width was 2700 mm, and the height was 3500 mm. The tobacco loading capacity was 5500 kg. The calculated volume density parameter ρ was 29.9 kg / m 3 The initial moisture content of the tobacco leaves, measured by the sample drying method, was 85.7%. The average ambient temperature during the experiment was 28°C, and the target baking temperature was set at 68°C. The actual baking process lasted for 128 hours, with data collected every 60 seconds.

[0153] During the model construction phase, the dehydration and moisture dissipation physical process equations were first established as the physical constraints of the neural network. The thermal conductivity coefficient k was calculated based on the bulk density parameter ρ and was 0.024 W / (m·K). The water diffusion coefficient D varied with temperature and moisture content within a range of 2.5×10 -7 ~8.3×10 -6 m 2 / s. The researchers conducted a statistical analysis of the collected time series data and identified four typical stages in the baking process. The characteristic parameters of each stage are shown in Table 1:

[0154] Table 1 Characteristic parameters of four typical stages of tobacco leaf curing

[0155]

[0156] The researchers constructed a specific thermodynamically constrained neural network model for this tobacco variety and loading conditions. The model employs a five-layer structure, with four neurons in the input layer and three hidden layers with 64, 128, and 256 neurons, respectively. An adaptive temperature field attention mechanism was introduced, resulting in a calculated number of attention heads of 12. The model ultimately outputs predicted heat load values and moisture content change rates for 48 time points. The model was trained using a measured dataset, with the training parameter settings shown in Table 2.

[0157] Table 2 Thermodynamic constraint neural network model training parameter settings

[0158] Parameter name Parameter value Parameter name Parameter value Batch size 64 Initial learning rate 0.001 Training rounds 800 Weight decay coefficient 0.0001 discard rate 0.2 Physics Constraint Weights 0.15 Heat transfer constraint coefficient 0.5 Moisture diffusion constraint coefficient 0.3 Structural change constraint coefficient 0.2 Early stopping patience value 30

[0159] Through cross-validation to evaluate model performance, the root mean square error of heat load prediction was 0.42kW, the mean absolute percentage error was 4.3%, and the root mean square error of moisture content change rate prediction was 0.15% / h, meeting the preset accuracy requirements. The heat load prediction results for different smoke loading conditions are shown in Table 3:

[0160] Table 3 Prediction of heat load of tobacco leaf baking under different tobacco loading conditions

[0161]

[0162] The researchers further tested the model's predictive performance under varying initial moisture content conditions. The results demonstrated that the model accurately captures the impact of changes in initial moisture content on heat load. When the initial moisture content changes from 83% to 88%, total energy consumption increases by approximately 12.5%, and heat load values at each stage also increase accordingly. The model also calculates thermal efficiency under these conditions, providing a basis for the configuration and optimization of clean energy equipment.

[0163] Traditional tobacco leaf curing heat load analysis relies mainly on empirical estimation or simple statistical regression models, which are difficult to accurately reflect the complex nonlinear relationship between tobacco leaf characteristics and heat load, and the prediction accuracy is usually within the range of ±10% to 15%. The present invention, by integrating physical mechanisms and deep learning technology, constructs a thermodynamically constrained neural network model with physical interpretation capabilities, and the prediction accuracy is improved to within ±5%. Compared with traditional methods, the main advancements of the present invention are reflected in three aspects: first, the dehydration and moisture dissipation physical process equations are introduced as constraints of the neural network to ensure that the model prediction results conform to the basic laws of thermodynamics; second, an adaptive temperature field attention mechanism is developed, which can automatically adjust the perception of the spatial relationship of the temperature field according to the volume density parameter, initial moisture content and target temperature; finally, accurate heat load prediction is achieved for different tobacco leaf types and loading conditions, providing a scientific basis for energy planning and optimization control of clean energy-intensive curing barns.

[0164] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 4, 5 and 6 below.

[0165] Table 4 Variable Explanation Table (Part 1)

[0166]

[0167]

[0168] Table 5 Variable Explanation Table (Part 2)

[0169]

[0170] Table 6 Variable Explanation Table (Part 3)

[0171]

[0172]

[0173] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn, characterized in that: include: Collect temperature data of dense flue-curing rooms and operating power data of clean energy heating equipment; determine the volume density parameters of dense flue-curing rooms and calculate the initial moisture content of tobacco leaves; construct a set of equations for the physical process of tobacco leaf dehydration and moisture dissipation as physical constraints of the neural network; establish a thermodynamic constraint neural network model with inputs of initial moisture content of tobacco leaves, volume density parameters, ambient temperature, and target temperature, introduce an adaptive temperature field attention mechanism, and dynamically capture the mutual influence of temperature fields at different locations in the flue-curing room through a multi-head self-attention network; use the measured data set to train the thermodynamic constraint neural network model as a tobacco leaf baking heat load analysis model.

2. The method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn according to claim 1, characterized in that: The volume density parameter refers to the ratio of the mass of tobacco leaves filled in unit volume to the total volume of the flue-curing room, which is used to calculate heat conduction efficiency and ventilation resistance.

3. The method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn according to claim 2, characterized in that: The equations for the physical process of tobacco leaf dehydration and moisture dissipation include the heat conduction equation, the moisture diffusion equation, and the tissue structure change equation.

4. The method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn according to claim 3, characterized in that: The heat conduction equation is used to describe the process of heat propagation in space in the flue-curing room. The input includes ambient temperature, heating equipment power distribution, and tobacco leaf distribution density, and the output is the temperature field distribution in the flue-curing room.

5. The method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn according to claim 4, characterized in that: The moisture diffusion equation is used to describe the process of moisture migration from tobacco leaves. The input includes the initial moisture content of the tobacco leaves, the ambient air temperature and humidity, and the output is the rate of change of the moisture content of the tobacco leaves.

6. The method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn according to claim 5, characterized in that: The tissue structure change equation is used to describe the changes in the cell structure of tobacco leaves during the drying process. The input includes the initial moisture content of the tobacco leaves, the temperature change rate, and the moisture content change rate. The output is the tobacco leaf tissue structure parameters.

7. The method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn according to claim 6, characterized in that: The measured data set refers to a comprehensive data set collected through multiple flue-curing barn experiments, including the initial moisture content of tobacco leaves, bulk density parameters, ambient temperature, target temperature, time series heat load values, and moisture content change rate, which is used to train the thermodynamic constraint neural network model.

8. The method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn according to claim 7, characterized in that: The cross-validation method refers to a statistical method that divides the measured data set into a training set and a validation set, and tests the performance of the thermodynamic constraint neural network model under different segmentation methods to evaluate the generalization ability of the thermodynamic constraint neural network model.

9. The method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn according to claim 8, characterized in that: The specific structure of the thermodynamic constraint neural network model is a hybrid architecture that couples a multi-layer perceptron with thermodynamic equations, including an input layer, a hidden layer, a physical constraint layer, and an output layer. The number of neurons in the hidden layer is determined according to the volume density parameter, the initial moisture content of the tobacco leaves, and the target temperature. The physical constraint layer is embedded in the dehydration and moisture dissipation physical process equations to ensure that the output of the thermodynamic constraint neural network model conforms to physical laws.

10. The method for establishing a heat load analysis model for tobacco leaf baking in a clean energy-intensive flue-curing barn according to claim 9, characterized in that: The adaptive temperature field attention mechanism automatically adjusts the model's ability to perceive the spatial relationship of the temperature field according to changes in smoke density, initial moisture content and target temperature. When the volume density is high, the number of attention heads is increased to capture more complex heat conduction paths. When the initial moisture content is low, the number of attention heads is reduced to adapt to a simpler dehydration process. When the target temperature is high, the number of attention heads is increased to handle stronger thermodynamic interactions.

Images