A laboratory heating ventilation load prediction and adaptive regulation method

Abstract

The present application relates to air conditioning technology field, specifically to a kind of laboratory heating and ventilation load prediction and self-adaptive control method.The present application is packaged by synchronizing infrared frame and power sampling time scale, aligns and generates heat diffusion evolution rate.Combined with power jump and temperature rise time, establish heterogeneous dynamics coupling model.Extract physical evolution parameter as mechanism operator injection prediction model hidden layer, reconstruct phase space and calculate air load increment in lag window.Based on heat exchange characteristics, reverse mapping is executed to generate feedforward instruction, which is issued and dynamically corrected weight when rate exceeds threshold, to complete closed-loop correction.The present application performs deep coupling of feedforward prediction compensation and feedback residual adjustment, and realizes precise control of laboratory air volume and cold and heat load.

Description

Technical Field

[0001] This invention relates to the field of air conditioning technology, specifically to a method for predicting and adaptively controlling laboratory HVAC loads. Background Technology

[0002] Laboratory HVAC systems are core facilities for maintaining precise constant temperature, pressure gradients, and controlled air circulation within a laboratory. This is especially crucial in optical laboratories equipped with high-power laser devices, where the ability to regulate the heat load on the air side directly impacts the stability of the optical path and the accuracy of experiments. Currently, traditional laboratory air conditioning solutions primarily rely on feedback from temperature and humidity sensors within the controlled environment to drive adjustments based on feedback errors. This approach works by monitoring air parameters in the controlled area in real time. When measured values ​​deviate from preset target values, it drives the actuators to adjust the refrigerant flow rate or air volume in the heat exchanger, thus balancing the energy supply to the air side with the indoor heat load.

[0003] However, in scenarios involving thermal disturbances induced by high-power heat sources, traditional air-side regulation logic faces the following physical limitations: air-side regulation time lag and thermal inertia mismatch. Air regulation relies on a passive response after the deviation signal becomes apparent, but there is a significant physical penetration delay in the conduction of heat from the excitation source through the equipment enclosure to the air medium. Due to the limitations of the air medium's heat capacity and the thermal resistance of the heat exchange interface, the compensation signal on the air side lags behind the energy injection moment. This thermal inertia prevents the system from performing advance counterbalancing before the heat load diffuses to the air field on a large scale, easily causing transient overshoot in the indoor temperature field; asynchronous sensing data and a disconnect between the excitation and response mechanisms. Existing systems only monitor scalar air parameters within the controlled environment, lacking synchronous alignment between the heat source energy pulse and the air thermal field response in terms of timing and mechanism. Because the energy pulse at the heat generation end and the air-side thermal field response are asynchronous in the sensing dimension, the regulation loop cannot extract the implicit characteristic evolution law between the excitation and response, making it difficult to accurately determine the dynamic nonlinear trend of the air-side load; the regulation process lacks thermodynamic constraints and noise discrimination. Air load forecasting logic often ignores energy balance criteria and unsteady heat conduction laws, making it prone to regulation deviations that violate heat transfer logic under sudden thermal shocks. Furthermore, the lack of identification and dynamic gain allocation mechanisms for background thermal noise such as ambient long-wave radiation fluctuations leads to frequent ineffective regulation actions by the actuators, making it difficult to guarantee the steady-state accuracy of air conditioning.

[0004] In summary, the urgent technical problem to be solved by existing technologies is to overcome the limitations of air-side regulation lag and error-driven regulation, and to achieve precise control of laboratory air volume and heating / cooling load.

[0005] Therefore, a method for predicting and adaptively controlling laboratory HVAC loads is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting and adaptively controlling laboratory HVAC loads. By constructing a source-end sensing flow based on infrared and power data, the method uses the hysteresis characteristics of electrothermal conversion to predict load increments and combines valve curves to generate feedforward thermal compensation commands. This overcomes the dependence of existing feedback logic on air sensor deviation signals, solves the problems of response time delay and disturbance misjudgment, and achieves precise control of laboratory air volume and heating / cooling loads.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for predicting and adaptively controlling HVAC loads in a laboratory includes:

[0009] Extract the frame header time stamp from the infrared thermal imaging data acquisition terminal of the optical laboratory, and acquire the power sampling sequence from the laser power monitoring terminal. Perform field concatenation on the image pixel matrix data of the frame header time stamp and the power sampling sequence to construct the source sensing data stream and generate the thermal diffusion evolution rate.

[0010] Extract the power pulse jump time of the high-power laser device and the heating start time of the heat source region from the source-end sensing data stream. Determine the electrothermal conversion hysteresis characteristics based on the power pulse jump time and the heating start time, and establish a heterogeneous dynamic coupling model.

[0011] The physical evolution parameters of the heterogeneous dynamic coupling model are extracted as prior constraint operators and injected into the hidden layer of the dynamic prediction model to reconstruct the characteristic phase space. Within the physical lag window determined by the electrothermal conversion lag characteristics, the air load increment at the end of the physical lag window is calculated.

[0012] Based on the nonlinear characteristics of valve heat transfer, an inverse mapping relationship is established to convert the air load increment into a terminal actuator response command and generate a feedforward thermal compensation command. When the thermal diffusion evolution rate exceeds the background disturbance threshold, the feedforward thermal compensation command is issued.

[0013] The HVAC equipment is driven to respond to the feedforward thermal compensation command and superimpose the feedback signal. The weights of the feedforward thermal compensation command are dynamically adjusted according to the laboratory temperature regression, thus completing the closed-loop correction.

[0014] Preferably, the specific implementation of constructing the source-end sensing data stream includes: synchronizing the local system clocks of the infrared thermal imaging data acquisition terminal and the laser power monitoring terminal to the same time reference via a timing protocol; real-time parsing of each frame of infrared thermal image, extracting the frame header time stamp and corresponding image pixel matrix data of the infrared thermal image, retrieving instantaneous power values ​​in the power sampling sequence of the laser power monitoring terminal whose deviation from the frame header time stamp is within a preset tolerance window, concatenating the image pixel matrix data and the corresponding instantaneous power values ​​at the field level and adding index markers to generate a source-end sensing data stream aligned with the time axis; calculating the rate of change of temperature gradient from the center of the heat source radially outward in the image pixel matrix data, and using the time derivative of the rate of change of temperature gradient within a continuous sampling period as the thermal diffusion evolution rate.

[0015] Preferably, the specific construction method of the source-end sensing data stream and the specific generation method of the heat diffusion evolution rate include: locking the power pulse jump moment of the real-time power sampling sequence, collecting the temperature change rate of the heat source region exceeding the preset threshold at the start of the temperature rise, calculating the time difference between the start of the temperature rise and the power pulse jump moment, and using it as the electrothermal conversion hysteresis feature; calculating the equivalent heat source excitation intensity using the instantaneous heat generation power of the high-power laser device, and using the equivalent heat source excitation intensity as the source term of the equation, introducing the structural thermal inertia constant of the high-power laser device and the air convection heat transfer coefficient, and establishing an unsteady-state heat conduction partial differential equation characterizing heat diffusion; performing spatiotemporal solution on the unsteady-state heat conduction partial differential equation, and constructing a heterogeneous dynamic coupling model characterizing the dynamic correlation between power input fluctuations and the indoor three-dimensional thermal field response trajectory.

[0016] Preferably, the specific steps of injecting the hidden layer of the dynamic prediction model include: the hidden layer of the dynamic prediction model is a layer containing long short-term memory neurons in the dynamic prediction model; the heat transfer time delay coefficient, thermal equilibrium time constant, and thermal diffusion step response slope extracted from the heterogeneous dynamic coupling model are defined as physical evolution parameters; the physical evolution parameters are mapped to a bias correction vector that matches the dimension of the hidden layer of the dynamic prediction model using a preset weight transformation matrix; the bias correction vector is superimposed on the weighted summation calculation unit of the neurons in the hidden layer of the dynamic prediction model; based on the thermal diffusion inertia law represented by the physical evolution parameters, the allocation of nonlinear mapping weights of the neurons is controlled to limit the parameter optimization of the dynamic prediction model.

[0017] Preferably, the reconstructed feature phase space specifically includes: determining the time delay of the dynamic evolution process by calculating the autocorrelation function and average interaction information of the physical evolution parameter sequence, and calculating the minimum embedding dimension that satisfies the manifold reconstruction condition using a pseudo-nearest neighbor algorithm; the manifold reconstruction condition is determined by calculating the proportion of pseudo-nearest neighbor points generated by trajectory self-intersection during the process of increasing the embedding dimension, and using whether the proportion of pseudo-nearest neighbor points decreases to a preset convergence interval as the criterion for determining the manifold reconstruction condition; using the determined time delay and embedding dimension to perform time-delay coordinate translation and dimension-up transformation on the source-end sensing data stream to generate a multi-dimensional phase trajectory matrix characterizing the dynamic characteristics of load evolution, and using the multi-dimensional phase trajectory matrix as the input feature vector of the dynamic prediction model to obtain the implicit dynamic evolution characteristics.

[0018] Preferably, the calculation of the air load increment up to the end of the physical lag window specifically includes: inputting the multidimensional phase trajectory matrix into the dynamic prediction model, recursively updating the hidden layer state, and calculating the predicted value of the total indoor heat load from the current sampling time to the end of the preset physical lag time based on the electrothermal conversion lag characteristics; and determining the algebraic difference between the predicted value of the total indoor heat load and the real-time basic heat load value measured by the indoor temperature control device at the current time as the air load increment.

[0019] Preferably, the establishment of the reverse mapping relationship and the generation of feedforward thermal compensation instructions specifically include: establishing a heat transfer mapping model based on the heat exchanger energy balance criterion to describe the relationship between air-side heat transfer power and refrigerant-side mass flow rate, mean logarithmic temperature difference, and heat transfer efficiency; using the heat transfer mapping model to reverse calculate the refrigerant flow rate change demand value required to compensate for the increase in air load, and combining the linear flow characteristic curve of the HVAC equipment valve to convert the refrigerant flow rate change demand value into a response instruction of the HVAC equipment terminal actuator, and encapsulating the generated terminal actuator response instruction into a feedforward thermal compensation instruction.

[0020] Preferably, the specific steps for issuing the feedforward thermal compensation command include: under non-experimental conditions, acquiring a background thermal diffusion evolution rate sequence containing long-wave radiation fluctuations, calculating the mean and sample standard deviation of the background thermal diffusion evolution rate sequence, and using the sum of the mean and three times the sample standard deviation as the background disturbance threshold; compiling the generated feedforward thermal compensation command into an adjustment data packet conforming to the fieldbus communication protocol, and transmitting it to the terminal actuator driver of the HVAC equipment through a physical communication link; the terminal actuator driver parsing the adjustment data packet and generating analog voltage signals, analog current signals, and pulse modulation signals to drive the motor to move the physical valve core of the HVAC equipment valve to a preset position, thereby physically adjusting the refrigerant flow rate entering the laboratory heat exchange unit.

[0021] Preferably, the weighting of the dynamically corrected feedforward thermal compensation command specifically includes: using a temperature sensor array arranged in the laboratory to acquire the ambient temperature in real time, calculating the absolute value of the deviation of the ambient temperature from the preset target value, and differentiating the absolute value of the deviation over time to obtain the rate of decrease of the absolute value of the deviation; when the rate of decrease of the absolute value of the deviation is lower than the preset recovery threshold, increasing the weighting ratio of the feedforward thermal compensation command in the total adjustment output by a preset gradient step size until the ambient temperature returns to the preset allowable deviation range, and dynamically allocating the gain for the feedforward prediction compensation and feedback residual adjustment.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] 1. This invention synchronizes the acquisition time of infrared thermal imaging data frames from an optical laboratory with the real-time power timestamp of a high-power laser device, aligning and encapsulating energy consumption pulses with thermal field evolution. This constructs a source-end sensing data stream and generates a thermal diffusion evolution rate, achieving high-precision coupling of heterogeneous energy consumption characteristics and spatial thermal field evolution laws in the temporal dimension. This process eliminates the deviation in thermal field evolution trajectory determination caused by differences in sampling frequencies and asynchronous timing between multimodal sensors, providing a source-end sensing foundation with physical causal consistency for subsequent mechanism analysis.

[0024] 2. This invention extracts the physical evolution parameters of the heterogeneous dynamic coupling model and injects them as prior constraint operators into the hidden layer of the dynamic prediction model, reconstructing the feature phase space. This achieves the mechanistic constraint of the thermodynamic conduction law on the parameter optimization process of the deep learning model. This scheme utilizes the thermal diffusion inertia law, characterized by the heat transfer time delay coefficient and the thermal equilibrium time constant, to translate the activation reference point of neurons, thus limiting the physical solution space of the model prediction and correcting the non-physical prediction bias generated by the pure data-driven model under high-energy pulse jumps.

[0025] 3. This invention, by calculating the air load increment up to the end of the physical lag window, establishes a reverse mapping relationship based on the nonlinear characteristics of valve heat transfer, converting the air load increment into a response command for the terminal actuator and generating a feedforward thermal compensation command, thus achieving physical pre-countermeasures against instantaneous thermal shocks in the future. This mechanism performs reverse flow calculations through the heat exchanger energy balance criterion, enabling the heat exchange unit to pre-adjust the refrigerant circulation intensity before heat diffuses into the working area, compensating for the heat transfer lag effect caused by the thermal inertia of the heat source at the physical level.

[0026] 4. This invention completes closed-loop correction by executing superimposed feedback signals and dynamically adjusting the weights of feedforward thermal compensation commands based on laboratory temperature regression, establishing a dynamic gain allocation mechanism between feedforward predictive compensation and feedback residual adjustment. This logic adjusts the compensation ratio in real time according to the rate of decrease in ambient temperature deviation, ensuring that the temperature field returns to steady state at the optimal rate. Simultaneously, by balancing the ratio of feedforward and feedback components, it avoids system oscillations caused by a single adjustment logic, significantly improving the steady-state accuracy of high-power laboratory precision temperature control environments. Attached Figure Description

[0027] Figure 1 Here is a flowchart of a laboratory HVAC load prediction and adaptive control method provided in Embodiment 1 of the present invention;

[0028] Figure 2 This is a schematic diagram of the characteristic phase space reconstruction and load extrapolation logic in a laboratory HVAC load prediction and adaptive control method provided in Embodiment 1 of the present invention;

[0029] Figure 3 This is a schematic diagram of the hardware architecture of a laboratory HVAC load prediction and adaptive control system provided in Embodiment 2 of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figures 1 to 3 This invention provides a method for predicting and adaptively controlling laboratory HVAC loads, the technical solution of which is as follows:

[0032] A method for predicting and adaptively controlling HVAC loads in a laboratory includes:

[0033] Extract the frame header time stamp from the infrared thermal imaging data acquisition terminal of the optical laboratory, and acquire the power sampling sequence from the laser power monitoring terminal. Perform field concatenation on the image pixel matrix data of the frame header time stamp and the power sampling sequence to construct the source sensing data stream and generate the thermal diffusion evolution rate.

[0034] Extract the power pulse jump time of the high-power laser device and the heating start time of the heat source region from the source-end sensing data stream. Determine the electrothermal conversion hysteresis characteristics based on the power pulse jump time and the heating start time, and establish a heterogeneous dynamic coupling model.

[0035] The physical evolution parameters of the heterogeneous dynamic coupling model are extracted as prior constraint operators and injected into the hidden layer of the dynamic prediction model to reconstruct the characteristic phase space. Within the physical lag window determined by the electrothermal conversion lag characteristics, the air load increment at the end of the physical lag window is calculated.

[0036] Based on the nonlinear characteristics of valve heat transfer, an inverse mapping relationship is established to convert the air load increment into a terminal actuator response command and generate a feedforward thermal compensation command. When the thermal diffusion evolution rate exceeds the background disturbance threshold, the feedforward thermal compensation command is issued.

[0037] The HVAC equipment is driven to respond to the feedforward thermal compensation command and superimpose the feedback signal. The weights of the feedforward thermal compensation command are dynamically adjusted according to the laboratory temperature regression, thus completing the closed-loop correction.

[0038] Example 1:

[0039] This embodiment demonstrates the implementation method of the laboratory HVAC load prediction and adaptive control method provided by the present invention in the precision constant temperature control of a high-power optics laboratory. (See also...) Figures 1 to 2 The specific steps are as follows:

[0040] Furthermore, the specific implementation of constructing the source-end sensing data stream includes: synchronizing the local system clocks of the infrared thermal imaging data acquisition terminal and the laser power monitoring terminal to the same time reference via a timing protocol; real-time parsing of each frame of infrared thermal image, extracting the frame header time tag and corresponding image pixel matrix data of the infrared thermal image, retrieving instantaneous power values ​​in the power sampling sequence of the laser power monitoring terminal whose deviation from the frame header time tag is within a preset tolerance window, concatenating the image pixel matrix data and the corresponding instantaneous power values ​​at the field level and adding index markers to generate a source-end sensing data stream aligned with the time axis; calculating the rate of change of temperature gradient from the center of the heat source radially outward in the image pixel matrix data, and using the time derivative of the rate of change of temperature gradient within a continuous sampling period as the thermal diffusion evolution rate.

[0041] Specifically, the PTP precise time protocol is invoked to establish master-slave synchronization logic at the network layer. By periodically exchanging Sync messages, Follow_up messages, and delay request messages between the master clock node and the slave clock node, the transmission delay of the network link and the clock deviation between the two ends are calculated using hardware-level timestamps. The local system clocks of the infrared thermal imaging data acquisition terminal and the laser power monitoring terminal perform microsecond-level frequency compensation accordingly, aligning the clock phase to the same time reference, thus eliminating the risk of asynchronous data acquisition caused by crystal oscillator temperature drift in heterogeneous sensors.

[0042] Specifically, each frame of the raw bitstream output by the infrared thermal imager is analyzed in real time. A frame header time stamp in nanoseconds is extracted from the beginning of the data packet, and the image pixel matrix data, composed of M×N temperature scalars, is simultaneously separated. At the same time, the fixed-frequency power sampling sequence generated by the laser power monitoring terminal is retrieved from the cache. Using the frame header time stamp as a search index, records with an absolute deviation value within a preset tolerance window (e.g., ±0.5 milliseconds) are identified as instantaneous power values. The preset tolerance window is calculated as follows: half of the larger sampling period value between the infrared thermal imaging data acquisition terminal and the laser power monitoring terminal is taken as the upper limit of the allowable time deviation. The processor parses the matrix data and extracts two-dimensional pixel matrix data representing the current spatial thermal distribution panorama. Then, the processor uses the retrieved instantaneous power values ​​as an additional dimension. By broadcasting and expanding the instantaneous power values ​​to make their dimensions consistent with the infrared image pixel matrix data, a composite sensing tensor with 2 channels is constructed. A monotonically increasing sequence index is added, and the source sensing data stream containing the environmental thermal field and heat generation excitation is output.

[0043] Specifically, the region with the highest temperature value in the image pixel matrix data is identified as the heat source center. Based on the geometric topology of the laser device, the pixel value distribution within a preset physical spacing is extracted along multiple feature axes (including the main heat dissipation airflow direction and the major and minor axes of the structure). The temperature gradient between the heat source center point and the sampling points around each feature axis is calculated, and a weighted fusion operation is performed on the multi-path gradient values ​​to obtain the comprehensive temperature gradient change rate. The processor performs a first-order derivative of the comprehensive temperature gradient change rate with time over consecutive sampling periods to generate the thermal diffusion evolution rate.

[0044] As a preferred implementation, the parsing of the source-end sensing data stream further includes: performing a two-layer two-dimensional discrete wavelet transform on the acquired infrared thermal imaging data frame; decomposing the image into a low-frequency approximate component containing background isothermal information and horizontal, vertical, and diagonal high-frequency detail components containing transient temperature rise features; extracting the root of the square of the three high-frequency detail components as the pixel gradient magnitude, and using an eight-neighbor connected component labeling algorithm to group pixels with gradient magnitudes exceeding a preset mutation threshold into independent pixel clusters, locking the preset mutation threshold by a three-fold gain transform value of the background thermal distribution standard deviation; calculating the centroid coordinates of each pixel cluster, and locking the centroid of the pixel cluster with the largest area as the dynamic heat source center; correcting the initial energy reference of the thermal diffusion evolution rate by calculating the numerical difference between the dynamic heat source center and the low-frequency approximate component background thermal field, thus eliminating the interference of long-wave radiation from the laboratory environment on the diffusion rate calculation. This invention utilizes spatial frequency separation technology to improve the accuracy of heat source positioning and eliminates the interference of background thermal noise on diffusion rate identification.

[0045] This invention achieves microsecond-level clock phase alignment for heterogeneous sensors through the PTP timing protocol, eliminating time misalignment in the same-power sampling sequence of infrared image frames at the hardware level and ensuring the logical consistency of the source-end sensing data stream. By utilizing the radial evolution of the temperature gradient in thermal images and calculating the thermal diffusion evolution rate using time differentiation, the physical trend of heat diffusion into the laboratory space can be captured earlier and more accurately, improving the sensitivity of load sensing.

[0046] Furthermore, the specific implementation of establishing the heterogeneous dynamic coupling model includes: locking the power pulse jump moment of the real-time power sampling sequence, collecting the temperature change rate of the heat source region exceeding the preset threshold at the start of the temperature rise, calculating the time difference between the start of the temperature rise and the power pulse jump moment, and using it as the electrothermal conversion hysteresis feature; calculating the equivalent heat source excitation intensity using the instantaneous heat generation power of the high-power laser device, and using the equivalent heat source excitation intensity as the source term of the equation, introducing the structural thermal inertia constant of the high-power laser device and the air convection heat transfer coefficient, and establishing an unsteady-state heat conduction partial differential equation characterizing heat diffusion; performing spatiotemporal solution on the unsteady-state heat conduction partial differential equation, and constructing a heterogeneous dynamic coupling model characterizing the dynamic correlation between power input fluctuations and the indoor three-dimensional thermal field response trajectory.

[0047] Specifically, an edge detection algorithm is applied to monitor the real-time power sampling sequence with high-precision timestamps fed back from the laser power monitoring terminal, and the absolute value of the algebraic difference in power amplitude between adjacent sampling points is calculated. When the difference exceeds the step threshold of 5% of the rated power of the high-power laser device, a power input fluctuation is determined, and the timestamp corresponding to the abrupt change sampling point is locked as the power pulse jump moment. Simultaneously, a two-dimensional matrix composed of infrared pixels is retrieved, and the temperature values ​​of the pixels within the heat source region are subjected to real-time time-dimension differentiation. The first derivative, reflecting the temperature rise trend, is extracted as the temperature change rate of the heat source region. When the value of the first derivative first continuously exceeds a preset threshold (e.g., 0.01°C / s), the corresponding moment is locked as the heating start moment, indicating that heat has overcome the physical thermal resistance of the laser and begun to penetrate into the outer shell boundary. The algebraic time difference obtained by subtracting the power pulse jump moment from the heating start moment is calculated and defined as the electrothermal conversion hysteresis characteristic that quantifies the time delay of heat from the core heat generation penetrating into the device shell.

[0048] Specifically, Joule's law is used to perform energy mapping on the acquired instantaneous heat generation power to obtain the total heat generation. Then, a division operation is performed on the preset geometric heat dissipation area of ​​the high-power laser device's heat dissipation interface to derive the equivalent heat source excitation intensity per unit area. This equivalent heat source excitation intensity is loaded into the unsteady-state heat conduction partial differential equation as the equation's source term, i.e., the energy injection term maintaining thermal equilibrium. The structural thermal inertia constant, characterizing the heat storage resistance of the laser packaging material, and the air convection heat transfer coefficient, characterizing the heat dissipation capacity of the solid surface to the air, are introduced to establish the unsteady-state heat conduction partial differential equation. In establishing the unsteady-state heat conduction partial differential equation, the specific mapping method for the equation's source term is as follows: the processor divides the high-power laser device into a specific set of discretized grid nodes based on its physical coordinates in three-dimensional space, and injects the equivalent heat source excitation intensity as an energy increment into these node sets, characterizing the actual physical location of heat generation.

[0049] Specifically, the construction and solution process of the unsteady-state heat conduction partial differential equation is embodied in a multi-channel iterative calculation logic for three-dimensional spatial grid state data. Specifically, the boundary conditions of the equation are set to the third type of convective heat transfer boundary conditions, meaning that for the interface node between the equipment casing and the air, the heat dissipation flux is determined by the difference between the node temperature and the ambient background average temperature, as well as the air convective heat transfer coefficient. The initial conditions of the equation are set to the initial thermal field at time zero, meaning that at the initial point of calculation startup, the temperature values ​​of all grid nodes are initialized to the real-time background temperature measured by a laboratory temperature sensor array. The specific steps for constructing the unsteady-state heat conduction partial differential equation include: first, starting the energy balance calculation process for the discretized grid nodes, receiving three parallel input data streams to determine the temperature change gradient of the current node. The first path is the internal conduction data stream: it calls the neighborhood difference operator to read the temperature difference between the current node and its neighboring nodes in 3D space, and performs weighted aggregation based on the thermal conductivity of the laser encapsulation material to generate a value characterizing the heat diffusion flux within the medium. The second path is the source-end excitation data stream: it directly reads the energy value converted from the equivalent heat source excitation intensity and injects it as an external gain signal into the superposition operation to characterize the energy input introduced by laser heating. The third path is the boundary exchange data stream: it calculates the deviation between the current node temperature and the ambient background temperature, and uses the air convection heat transfer coefficient to perform linear scaling on this deviation value to generate a value characterizing surface heat dissipation loss. Finally, the superposition result of the above three data streams is divided by the thermal inertia parameter determined by the material density and specific heat capacity to obtain the rate of temperature change over time.

[0050] Specifically, the heterogeneous dynamic coupling model achieves data flow through three cascaded layers: a source-end energy sensing layer, a physical dynamic evolution layer, and a state-space mapping layer. The source-end energy sensing layer receives real-time power sampling sequences as input and outputs the equivalent heat source excitation intensity through electrothermal conversion logic. The physical dynamic evolution layer receives the excitation intensity as the source term of the equation and performs spatiotemporal solution of the unsteady-state heat conduction partial differential equation under the constraint of the electrothermal conversion hysteresis characteristics, outputting the indoor three-dimensional thermal field response trajectory characterizing the heat load evolution law. The state-space mapping layer receives the trajectory matrix and constructs a heterogeneous dynamic coupling model characterizing the heat transfer logic from energy excitation to load manifestation by establishing a multi-dimensional state function mapping between the power input fluctuation and the three-dimensional thermal field response trajectory.

[0051] This invention establishes a digital physical model of the response from electrical energy input to the three-dimensional thermal field by quantifying the hysteresis characteristics of electrothermal conversion and introducing unsteady-state partial differential equations of heat conduction. This model utilizes the finite difference method to perform spatiotemporal solutions, quantifying the thermal inertia barrier introduced by the laser packaging structure. This enables the system to identify the physical attenuation and time delay of the heat load along the propagation path, providing a solid physical causal background for subsequent predictions.

[0052] Further, the specific steps of injecting the hidden layer of the dynamic prediction model include: the hidden layer of the dynamic prediction model is a layer containing long short-term memory neurons in the dynamic prediction model; the heat transfer time delay coefficient, thermal equilibrium time constant, and thermal diffusion step response slope extracted from the heterogeneous dynamic coupling model are defined as physical evolution parameters; the physical evolution parameters are mapped to a bias correction vector that matches the dimension of the hidden layer of the dynamic prediction model using a preset weight transformation matrix; the bias correction vector is superimposed on the weighted summation calculation unit of the neurons in the hidden layer of the dynamic prediction model; based on the thermal diffusion inertia law represented by the physical evolution parameters, the allocation of nonlinear mapping weights of the neurons is controlled to limit the parameter optimization of the dynamic prediction model.

[0053] Specifically, in the state transition matrix of the heterogeneous dynamic coupling model, the heat transfer time delay coefficient reflecting the depth of heat conduction delay, the thermal equilibrium time constant reflecting the system's convergence rate to steady state, and the thermal diffusion step response slope reflecting the sensitivity to step temperature rise are extracted using parameter identification techniques. These three sets of characteristic values ​​reflecting thermodynamic evolution are then encapsulated into a one-dimensional sequence of physical evolution parameters. The heat transfer time delay coefficient is a scalar describing the time-domain displacement between the energy pulse and the manifestation of the boundary temperature rise, obtained by calculating the peak time of the coupled model's impulse response function. The thermal equilibrium time constant describes the time it takes for the system to reach equilibrium exponentially, calculated from the time required for the unsteady-state heat conduction partial differential equation to reach 63.2% of its steady-state value under step excitation. The thermal diffusion step response slope describes the rate of change of physical intensity in the initial stage of temperature rise, obtained by performing a first-order time derivative on the temperature rise trajectory output by the coupled model.

[0054] Specifically, a pre-defined weight transformation matrix of dimension M×H is constructed, where M is the feature dimension of the physical evolution parameter sequence, and H is the number of hidden layer neurons in the long short-term memory network layer of the dynamic prediction model. The pre-defined weight transformation matrix is ​​calibrated in the following way: First, the response sensitivity of the predicted air load value relative to the heat transfer time lag coefficient, the thermal equilibrium time constant, and the thermal diffusion step response slope is calculated using historical experimental data; then, the sensitivity values ​​of each parameter are converted to absolute values ​​and divided by the sum of the absolute values ​​of the sensitivity of all parameters to complete the normalization calculation, thereby obtaining the basic coefficients reflecting the strength of the physical constraint; finally, each basic coefficient is used as a mapping weight and spatially tiled according to the dimension of the hidden layer neurons of the dynamic prediction model to generate the weight transformation matrix.

[0055] Specifically, the dynamic prediction model consists of a cascaded input layer, a hidden layer containing long short-term memory neurons, and a fully connected regression layer. The input layer receives the multidimensional phase trajectory matrix and performs tensor slicing, outputting a formatted feature stream to the hidden layer. The bias correction vector is directly injected into the weighted summation unit of the hidden layer neurons; specifically, a multidimensional regression layer is constructed. The weight transformation matrix, where This represents the number of neurons in the hidden layer. Multiplying the three-element physical evolution parameter vectors yields a vector with dimension [dimensionality missing]. The bias correction vector is used to modify the cell state update formula during the computation of long short-term memory neurons. Subsequently, within the hidden layer, the product of the input vector and the input weight matrix, and the product of the previous hidden state vector and the circular weight matrix are summed. This bias correction vector is then used as an additional physical additive offset for superposition.

[0056] Specifically, the superposition action occurs before the linear summation result is output to the nonlinear activation function (such as the hyperbolic tangent activation function Tanh). Utilizing the heat diffusion inertia laws characterized by unidirectional heat transfer and spatiotemporal lag, as represented by physical evolution parameters, the activation reference point of the hidden layer neurons is directly shifted. Through the threshold shift of the bias correction vector, the nonlinear mapping weight allocation of the hidden layer neurons during training iterations is forcibly intervened. This ensures that the dynamic prediction model is constrained within a physical solution space that conforms to the physical logic of heat transfer during parameter optimization, eliminating the physically uninterpretable bias caused by overfitting in traditional deep learning models.

[0057] This invention utilizes the heat transfer time delay coefficient and thermal equilibrium time constant to generate a bias correction vector and injects it into the hidden layer of a long short-term memory network, achieving hard intervention under physical constraints. This mechanism, by shifting the activation reference point of neurons, forces the neural network to follow the inertial law of heat diffusion when performing nonlinear weight optimization, thus solving the problem of physically meaningless biases that deep learning models may produce when facing drastic thermal fluctuations, significantly enhancing the model's generalization ability and interpretability.

[0058] Further, the reconstructed feature phase space specifically includes: determining the time delay of the dynamic evolution process by calculating the autocorrelation function and average interaction information of the physical evolution parameter sequence, and calculating the minimum embedding dimension that satisfies the manifold reconstruction condition using the pseudo-nearest neighbor algorithm; the manifold reconstruction condition is determined by calculating the proportion of pseudo-nearest neighbor points generated by trajectory self-intersection during the process of increasing the embedding dimension, and using whether the proportion of pseudo-nearest neighbor points decreases to a preset convergence interval as the criterion for determining the manifold reconstruction condition; using the determined time delay and embedding dimension to perform time-delay coordinate translation and dimension-up transformation on the source-end sensing data stream to generate a multi-dimensional phase trajectory matrix characterizing the dynamic characteristics of load evolution, and using the multi-dimensional phase trajectory matrix as the input feature vector of the dynamic prediction model to obtain the implicit dynamic evolution characteristics.

[0059] Specifically, the autocorrelation function is invoked to perform linear similarity analysis on the physical evolution parameter sequence. By calculating the covariance of the original signal and its own displacement signal at different time offsets and performing standardization, the time span for the autocorrelation coefficient to decrease to the initial value 1 / e is determined. Simultaneously, the average interaction information is calculated, and the statistical correlation between the sequence and its lagged replica is calculated based on the joint probability density and marginal probability density distributions. The offset step corresponding to the first local minimum point of the interaction information curve is retrieved, and the offset step is determined as the time delay characterizing the dynamic evolution process that minimizes the uncertainty of system information. The pseudo-nearest neighbor algorithm is used to increase the embedding dimension. The nearest neighbor of each state point is found in the d-dimensional reconstruction space and the Euclidean distance is calculated. Then, the reconstructed manifold is simulated to be raised to the d+1-dimensional space and the incremental projection distance of the nearest neighbor in the newly added coordinate axis direction is recalculated. If the ratio of the incremental projection distance to the original Euclidean distance exceeds a preset threshold, the current nearest neighbor is determined to be a pseudo-nearest neighbor caused by trajectory self-intersection due to the incomplete unfolding of the manifold. The criteria for determining the manifold reconstruction condition are whether the proportion of statistically generated false nearest neighbors decreases monotonically and converges to a preset convergence interval. The preset convergence interval is determined by calculating the second derivative of the proportion of false nearest neighbors. When the increase in dimension causes the rate of change of the proportion to become stable and the absolute value is less than 5%, it is considered to have entered the convergence interval, thereby locking the minimum embedding dimension and ensuring that the complex nonlinear load evolution dynamics characteristics inside the controlled environment can be fully unfolded.

[0060] Specifically, a time-delay coordinate translation is performed on the source-end sensing data stream using a determined time delay and a minimum embedding dimension. The original one-dimensional time series is then sampled at equal intervals according to integer multiples of the time delay, completing the dimensionality upgrade from a one-dimensional scalar signal to a multi-dimensional spatial state vector. By sliding the sampling window on the time axis, the state vectors generated at each time point, exhibiting dynamic evolution relationships, are arranged row by row, outputting a multi-dimensional phase trajectory matrix characterizing the co-evolution of indoor heat generation and dissipation. The number of rows corresponds to the number of sampling times, and the number of columns is determined by the minimum embedding dimension. Each row vector in the matrix represents the system evolution state point after the time-delay coordinate translation at that time point. This reconstructs the one-dimensional time series data into a spatial trajectory tensor reflecting the co-evolution of indoor heat generation and dissipation, serving as the input features of the dynamic prediction model at each time step. The input layer receives the multidimensional phase trajectory matrix and performs tensor slicing and formatting, outputting a time-series slice stream that meets the requirements of recursive operations to the Long Short-Term Memory (LSTM) network layer. The LSM network layer processes the input feature vector using its internal forget gate, input gate, and output gate mechanisms, captures long-term temporal dependencies through recursive updates of the hidden layer states, and outputs a latent dynamic evolution feature scalar that characterizes the diffusion nature of the indoor temperature field to the fully connected regression layer, thereby calculating the predicted total heat load value for future moments under controlled conditions.

[0061] As a preferred implementation, the injection of the hidden layer of the dynamic prediction model specifically includes: constructing physical feature residual connection paths in the long short-term memory network layer; constructing the extracted heat transfer time delay coefficient and thermal diffusion step response slope as the original physical feature vector, and performing linear projection using a preset dimension mapping matrix to map out a physical feature residual vector with the same dimension as the state vector of the hidden layer of the dynamic prediction model; performing element-wise summation operation between the physical feature residual vector and the state vector output by the hidden layer; and using the hyperbolic tangent activation function to perform mapping range restriction on the activation values ​​of neurons, thereby compressing the output trajectory of neurons into the value range determined by the physical heat balance equation, and maintaining the feature expression strength of physical prior information in the long-term recursive deduction process. This invention maintains the transmission efficiency of physical constraints in deep networks through residual connections, reducing the deviation of the predicted trajectory under long-term delay conditions.

[0062] This invention utilizes autocorrelation functions, average mutual information, and the pseudo-nearest neighbor algorithm to realize the topological expansion of complex nonlinear dynamic systems. By finding the minimum embedding dimension that satisfies the manifold reconstruction condition, one-dimensional low-dimensional features are mapped into a multi-dimensional phase trajectory matrix containing implicit dynamic evolution features. This fully preserves the co-evolutionary information of heat generation, heat storage, and diffusion within the laboratory, providing more complete input features for predictive models to capture high-order nonlinear load fluctuations.

[0063] Furthermore, the calculation of the air load increment up to the end of the physical lag window specifically includes: inputting the multidimensional phase trajectory matrix into the dynamic prediction model, recursively updating the hidden layer state, and calculating the predicted value of the total indoor heat load from the current sampling time to the end of the preset physical lag time based on the electrothermal conversion lag characteristics; and determining the algebraic difference between the predicted value of the total indoor heat load and the real-time basic heat load value measured by the indoor temperature control device at the current time as the air load increment.

[0064] Specifically, the multidimensional phase trajectory matrix is ​​loaded into a prediction model consisting of an input layer, a hidden layer containing long short-term memory neurons, and a fully connected regression layer cascaded together. The input layer receives the multidimensional phase trajectory matrix as input data stream, and through tensor dimension transformation and temporal slicing processing, maps the high-dimensional phase space features into a temporal tensor sequence that conforms to the model input and outputs it to the hidden layer. The hidden layer receives the temporal tensor sequence and uses the gating weight matrix inside the neurons to perform recursive updates of the hidden layer state. Specifically, at the current sampling time, the forget gate weight matrix is ​​used to selectively filter the cell state at the previous sampling time, and the input gate weight matrix is ​​used to perform a nonlinear transformation on the input features at the current sampling time to update the cell state. The recursive update process performs continuous iterative deduction with a preset time step increment, gradually deriving from the current sampling time to the preset physical lag time endpoint. The preset physical lag time is obtained by repeatedly collecting the electrothermal conversion lag features under the operation of a high-power laser device and taking its maximum statistical value as the preset value. In each recursive step, the implicit state vector is used to capture the dynamic law of thermal inertia contained in the multidimensional phase trajectory matrix, and the output vector at the current moment is used as the input parameter of the implicit layer at the next moment. This realizes the temporal recursive transmission of thermal field evolution characteristics within the physical lag window, thereby deriving the high-order implicit feature vector reflecting the heat load distribution at the preset physical lag time endpoint and outputting it to the fully connected regression layer.

[0065] Specifically, the fully connected regression layer receives the hidden layer state vector output by the hidden layer at the preset physical lag time endpoint. The fully connected regression layer maps the high-dimensional hidden features into a scalar value reflecting the total indoor heat generation intensity, i.e., the predicted total indoor heat load, by performing a weighted summation operation on the linear weight matrix. Simultaneously, the real-time baseline heat load value is acquired through the sensor array of the indoor temperature control system. This real-time baseline heat load value is obtained by calculating the algebraic product of the circulating air volume, refrigerant specific heat capacity, and the temperature difference between the inlet and outlet of the heat exchanger, and is used to characterize the background baseline power consumption of the laboratory after excluding the thermal shock effect of the high-power laser device. An algebraic difference calculation is performed, specifically by subtracting the real-time baseline heat load value from the predicted total indoor heat load value, and the resulting algebraic difference is determined as the air load increment caused by the operation of the high-power laser device and which needs to be compensated by HVAC equipment in the future.

[0066] This invention utilizes the hidden layer recursive update mechanism of Long Short-Term Memory (LSTM) networks to achieve trend-based forward prediction within a physical lag window. By calculating the algebraic difference between the predicted total load and the real-time base load, it accurately isolates the transient air load increments generated by the operation of high-power equipment, enabling HVAC adjustments to precisely locate the controlled variable and avoid erroneous adjustments caused by background heat load fluctuations.

[0067] Furthermore, the establishment of the reverse mapping relationship and the generation of feedforward thermal compensation instructions specifically include: establishing a heat transfer mapping model based on the heat exchanger energy balance criterion to describe the relationship between air-side heat transfer power and refrigerant-side mass flow rate, mean logarithmic temperature difference, and heat transfer efficiency; using the heat transfer mapping model to reverse calculate the refrigerant flow rate change demand value required to compensate for the increase in air load; and combining the linear flow characteristic curve of the HVAC equipment valve to convert the refrigerant flow rate change demand value into a response instruction for the HVAC equipment terminal actuator; and encapsulating the generated terminal actuator response instruction into a feedforward thermal compensation instruction.

[0068] Specifically, a heat exchange mapping model is constructed based on the heat exchanger energy balance criterion, consisting of a power input layer, a thermodynamic feature processing layer, and a flow calculation layer cascaded together. The power input layer receives the air load increment determined in the preceding steps and defines the air load increment as the air-side heat exchange power, which is an energy scalar reflecting the intensity of heat released by the air-side fluid of the heat exchanger per unit time. The thermodynamic feature processing layer retrieves the real-time temperature measurements of the inlet and outlet of the air-side and refrigerant-side of the heat exchange unit of the HVAC equipment and calculates the average logarithmic temperature difference based on the logarithmic distribution law of the fluid temperature difference at both ends. The average logarithmic temperature difference represents the physical characteristic describing the intensity of the logarithmic average driving force of the heat exchange process, and is obtained by calculating the difference between the hot and cold ends of the heat exchanger and the quotient of the natural logarithm of the ratio of the temperature difference. The heat exchange mapping model introduces the heat transfer efficiency, which characterizes the inherent structural heat transfer capability of the heat exchange equipment, to achieve parameterized encapsulation of the physical performance of the heat exchange unit.

[0069] Specifically, during the reverse analysis of the refrigerant-side mass flow rate, the flow calculation layer configures the previously predicted air load increment as the target total heat exchange power demand on the air side of the heat exchanger, and establishes the physical equivalence between this power demand and the energy released in the refrigerant circulation loop, following the heat exchange energy balance criterion. The processor acquires the temperature characteristic parameters of the inlet and outlet of the air and refrigerant sides of the heat exchanger in real time, and analyzes the mean logarithmic temperature difference, which characterizes the heat exchange drive intensity throughout the process, based on the nonlinear heat transfer law of the fluid. That is, it takes the natural logarithm of the ratio of the difference between the fluid temperature difference at the inlet and outlet of the heat exchanger to this temperature difference. On this basis, the system constructs a multi-physical quantity coupled mapping model including air-side heat exchange power, refrigerant mass flow rate, refrigerant isobaric specific heat capacity, and mean logarithmic temperature difference. By configuring the air load increment as the divisor and configuring the algebraic product of heat transfer efficiency, mean logarithmic temperature difference, and refrigerant isobaric specific heat capacity as the divisor, the mapping operation is performed, thereby reversely analyzing the refrigerant-side fluid mass flow rate adjustment deviation required to maintain heat exchange balance.

[0070] Specifically, the linear flow characteristic curve of the HVAC equipment valve is retrieved. This curve stores the proportional mapping logic between the valve core's physical displacement and the refrigerant mass flow rate. Using the refrigerant flow rate change demand value, the linear flow characteristic curve is retrieved to pinpoint the required stroke percentage or rotation angle of the HVAC equipment's terminal actuator, which is then defined as the HVAC equipment's terminal actuator response command. This terminal actuator response command manifests as specific level signal characteristics driving the proportional control valve or stepper motor. Subsequently, the terminal actuator response command is encapsulated with the equipment address code and function bit execution data to generate a feedforward thermal compensation command conforming to the industrial communication protocol.

[0071] Specifically, a fluid transport delay compensation factor is introduced during the generation of the feedforward thermal compensation command. This compensation factor is calibrated based on the physical length of the laboratory refrigerant pipeline, the rated fluid velocity, and the response slope of the convective heat transfer on the heat exchanger surface. The specific method is as follows: the ratio of the physical length of the laboratory refrigerant pipeline to the rated fluid velocity is calculated to determine the fluid transport time delay; the thermodynamic response time required for the heat exchanger to reach a preset response threshold is calculated using the convective heat transfer response slope; and the compensation factor is obtained by summing the fluid transport time delay and the thermodynamic response time. This compensation factor is used to advance the phase of the terminal actuator's response command issuance time. This advance processing ensures that the refrigerant flow change caused by the valve's physical action can penetrate the physical displacement delay generated by fluid transport, ensuring that the time when the refrigerant's cooling capacity reaches the heat exchange interface and the time when the heat pulse generated by the heat source reaches the working area remain temporally coupled, thereby reducing the impact of the physical transport delay.

[0072] As a preferred implementation, establishing the reverse mapping relationship further includes: acquiring the pressure difference values ​​of the inlet and outlet pressure sensors of the heat exchanger in real time and calculating the instantaneous flow resistance coefficient; calculating the deviation ratio of the instantaneous flow resistance coefficient relative to the pre-stored standard rated flow resistance; loading the deviation ratio as an independent variable into a preset efficiency decay function, and using exponential mapping logic to reduce the heat transfer efficiency value in the heat transfer mapping model; using the corrected heat transfer efficiency to perform reverse calculation, calculating the quotient of the air load increment relative to the product of heat transfer efficiency, average logarithmic temperature difference, and refrigerant isobaric specific heat capacity, deriving the refrigerant flow rate change demand value corresponding to the air load increment, realizing real-time compensation for nonlinear fluctuations in heat transfer efficiency caused by pipe fouling or pump frequency conversion, and ensuring that the generated terminal actuator response command accurately offsets the predicted heat load. This invention uses real-time flow resistance feedback to correct the reverse calculation model, enhancing the accuracy of the system's heat balance adjustment under fluctuating pipe operating conditions.

[0073] This invention utilizes a heat exchange mapping model based on the heat exchanger energy balance criterion to achieve accurate inverse calculation from "heat energy demand" to "refrigerant flow demand." By combining the nonlinear flow characteristic curves of HVAC valves to generate execution commands, it ensures seamless integration of the regulation logic at the thermodynamic and mechanical execution levels, resolving the problem of mismatch between power regulation demand and the physical action of the actuator.

[0074] Furthermore, the specific steps for issuing the feedforward thermal compensation command include: under non-experimental conditions, acquiring a background thermal diffusion evolution rate sequence containing long-wave radiation fluctuations, calculating the mean and sample standard deviation of the background thermal diffusion evolution rate sequence, and using the sum of the mean and three times the sample standard deviation as the background disturbance threshold; compiling the generated feedforward thermal compensation command into an adjustment data packet conforming to the fieldbus communication protocol, and transmitting it to the terminal actuator driver of the HVAC equipment through a physical communication link; the terminal actuator driver parses the adjustment data packet and generates analog voltage signals, analog current signals, and pulse modulation signals to drive the motor to move the physical valve core of the HVAC equipment valve to a preset position, thereby physically adjusting the refrigerant flow rate entering the laboratory heat exchange unit.

[0075] Specifically, during the system calibration phase, the time evolution variance of the thermal diffusion rate is calculated by observing the infrared thermal field sequence under conditions without high-power equipment operation for an extended period; the time evolution variance is then... The upper limit of the confidence interval is defined as the basic background noise level, and the sum of the mean of the background thermal diffusion evolution rate sequence and three times the sample standard deviation is determined as the background disturbance threshold. This background disturbance threshold isolates interference signals caused by long-wave radiation fluctuations in the environment through statistical probability distribution characteristics. The feedforward thermal compensation instruction is compiled into low-level binary machine code conforming to the fieldbus communication protocol. The binary machine code is encapsulated according to an adjustment data packet with a specified topology for address, function, data, and parity fields. The generated adjustment data packet exhibits a level characteristic sequence with timing logic and is transmitted to the terminal actuator driver via a physical communication link such as shielded twisted-pair cable or industrial Ethernet cable.

[0076] Specifically, the terminal actuator driver receives the adjustment data packet and performs electrical signal calculation and power mapping; the driver parses the opening command value in the data packet and generates an internal digital-to-analog converter circuit. V analog voltage signal, Analog current signals or pulse-modulated signals are used. The analog voltage signals and analog current signals achieve a linear mapping between the valve opening degree of the actuator, while the pulse-modulated signals adjust the motor speed and torque by changing the duty cycle. These various level signals work together to power the valve unit of the HVAC equipment, converting electrical signals into physical torque that drives the mechanical mechanism to translate or rotate. This torque drives the physical valve core to perform mechanical movement relative to the valve seat, moving the valve core to a preset position. The preset position determines the effective flow cross-sectional area of ​​the valve's internal flow channel, and as the valve core opening degree changes, the flow rate of refrigerant entering the laboratory heat exchange unit is physically adjusted.

[0077] This invention sets a background disturbance threshold by calculating the upper limit of the confidence interval of the background thermal diffusion rate, effectively filtering measurement noise caused by long-wave radiation fluctuations within the laboratory. It employs an industrial-grade bus protocol to encapsulate adjustment data packets and directly drives the valve actuator to respond to analog electrical or pulse signals, ensuring that the predicted thermal compensation amount is faithfully converted into physical adjustment of refrigerant flow, thus enhancing the system's anti-interference performance under actual operating conditions.

[0078] Furthermore, the weighting of the dynamically corrected feedforward thermal compensation command specifically includes: acquiring the ambient temperature in real time using a temperature sensor array deployed in the laboratory, calculating the absolute value of the deviation of the ambient temperature from a preset target value, and differentiating the absolute value of the deviation over time to obtain the rate of decrease of the absolute value of the deviation; when the rate of decrease of the absolute value of the deviation is lower than a preset recovery threshold, increasing the weighting ratio of the feedforward thermal compensation command in the total adjustment output by a preset gradient step size until the ambient temperature returns to a preset allowable deviation range, and dynamically allocating the gain for feedforward prediction compensation and feedback residual adjustment.

[0079] Specifically, feedback values ​​from an array of temperature sensors distributed across a spatial grid node in the laboratory are retrieved in parallel. The ambient temperature is obtained by performing a spatial weighted average operation on the multi-channel sampled data. The absolute value of the algebraic difference between the ambient temperature and a preset target value is calculated to obtain the absolute value of the deviation, which is mathematically represented as a scalar value reflecting the distance of the measured value from the ideal equilibrium state. A first-order time derivative operation is performed on the continuously sampled absolute value sequence of deviations, i.e., the derivative is calculated as a function of time, and the derivative value is extracted as the rate of decrease of the absolute value of the deviation. This rate of decrease is represented by the slope of the tangent line to the deviation curve on the time axis, characterizing the dynamic response performance of the controlled temperature field regressing to the steady-state target value.

[0080] Specifically, when the rate of decrease of the absolute value of the deviation is lower than a preset recovery threshold, it is determined that the feedforward predicted power is insufficient to cover the real-time thermal shock. The preset recovery threshold is obtained by performing a reciprocal operation on the thermal response time constant of the laboratory environment in open-loop condition, representing the minimum temperature rise suppression slope that the system should possess under standard load. A weight correction mechanism is triggered, increasing the weighted allocation ratio of the feedforward thermal compensation command in the total adjustment allocation according to a fixed preset gradient step size (e.g., 0.02). The preset gradient step size is derived from simulation iteration based on the physical minimum step resolution of the HVAC actuator and the system stability criterion. While increasing the feedforward allocation ratio, the weight of the feedback adjustment component based on the measured temperature difference is simultaneously reduced according to the principle of proportional conservation (i.e., the sum of the feedforward weight and the feedback weight is 1), thereby realizing the dynamic gain allocation between feedforward predictive compensation and feedback residual adjustment.

[0081] Specifically, the dynamic gain allocation process performs closed-loop iterations as the sampling period progresses, adjusting the overall regulation output to the HVAC actuator using continuously updated weight allocation coefficients until the ambient temperature returns to a preset allowable deviation range (e.g., ...). When the overall measurement uncertainty of the temperature sensor array is summed with the temperature deviation value corresponding to the physical minimum step resolution of the HVAC actuator, the sum is used as the threshold of the allowable deviation range.

[0082] This invention establishes a dynamic weight allocation mechanism based on the absolute value of temperature deviation and the rate of decrease, enabling autonomous correction of the adjustment gain. When the predicted feedforward compensation is insufficient to suppress the regression of the actual deviation, the system autonomously increases the feedforward ratio and decreases the feedback weight by adjusting the gradient step size, effectively solving the prediction deviation problem caused by system aging and environmental drift. While ensuring rapid temperature control response, it minimizes system oscillation.

[0083] Example 2:

[0084] This second embodiment details the specific application of a laboratory HVAC load prediction and adaptive control system in the precision temperature control of a high-power optics laboratory. (See also...) Figure 3 The system hardware architecture includes an infrared thermal imager deployed on the top of the laboratory, a high-frequency power sampling module connected in series with the laser power supply line, multiple air temperature sensors distributed in the indoor work area, and HVAC terminal actuators connected to the central processor. The specific application method is as follows:

[0085] Environmental benchmark calibration and nonlinear characteristic modeling were performed. During the background period when the laser equipment was not activated, time-series data of temperature fluctuations in the laboratory's natural environment were collected. By analyzing the spatial evolution of the temperature field, the background thermal diffusion evolution rate, reflecting the spatial distribution density of heat, was calculated. The maximum value of the background thermal diffusion evolution rate over a 24-hour period was statistically analyzed, multiplied by a preset safety factor to obtain the calculation result, and a background disturbance threshold to avoid environmental noise interference was determined. Next, a nonlinear functional mapping relationship between valve opening and heat exchange was established through pre-testing. Based on this, the terminal valve characteristic curve was established as the benchmark data stream for subsequent reverse calculations and stored in local data storage.

[0086] When the laser starts up or switches operating conditions, it enters a real-time prediction process. A high-frequency power sampling module monitors abrupt changes in the laser's power input and applies an edge detection algorithm to pinpoint the moment of the power pulse jump. The real-time infrared pixel matrix is ​​analyzed to synchronously acquire thermal field characteristics, identifying the pixel with the highest temperature as the center of the heat source. The radially outward temperature gradient rate of change is calculated, and by calculating the time derivative of this gradient rate over a continuous sampling period, the real-time thermal diffusion evolution rate is determined. The construction of this baseline data stream completes the synchronous alignment of heterogeneous energy consumption characteristics with the spatial thermal evolution trajectory.

[0087] Heat transfer delay quantization and multimodal feature reconstruction are performed. A preset radius region around the center of the heat source is locked, and the average temperature change curve is monitored in real time. When the first derivative of the average temperature change curve exceeds a preset threshold, the heating start time is locked. The algebraic time difference between the heating start time and the power pulse jump time is calculated and defined as the electrothermal conversion hysteresis feature. The physical evolution parameter sequence reflecting the heat transfer delay depth and convergence rate is extracted, mapped into a bias correction vector using a preset weight transformation matrix, and injected into the hidden layer neuron computing unit of the dynamic prediction model, realizing the hard constraint intervention of physical thermodynamic laws on the deep learning training process.

[0088] The pseudo-nearest neighbor algorithm is used to statistically determine the proportion of pseudo-nearest neighbor points generated by trajectory self-intersections, and the minimum embedding dimension is locked when the proportion converges. A dimensionality-up transformation is performed on the source-end sensing data stream using a determined time delay and embedding dimension to generate a multi-dimensional phase trajectory matrix, recovering the implicit dynamic evolution characteristics of the controlled environment. This matrix is ​​input into the prediction model, and using the recursive update mechanism of the hidden layer state, the predicted value of the total indoor heat load at the end of the physical lag time is calculated in advance within the physical lag window. The algebraic difference obtained by subtracting the real-time baseline heat load value from the predicted total heat load value is used to determine the air load increment.

[0089] Based on the heat exchanger energy balance criterion, a reverse calculation is performed, and the required refrigerant flow rate for compensating for changes in air load is derived through a heat transfer mapping model. By retrieving the linear flow characteristic curve, a response command for the terminal actuator is generated and encapsulated into a protocol-compliant feedforward heat compensation command. The command is issued and dynamic weight correction is performed. The driver parses the data packet and generates analog voltage, current, or pulse modulation signals to drive the motor, moving the physical valve core to the preset position required by the command, thus physically regulating the refrigerant flow rate in the heat exchange unit.

[0090] The system retrieves feedback values ​​from the temperature sensor array in parallel and calculates the absolute value of the ambient temperature deviation and its rate of decrease. If the regression rate is lower than a preset threshold, the weight of the feedforward thermal compensation command is increased according to a fixed gradient step size, achieving dynamic gain allocation between feedforward prediction and feedback adjustment. This process continues until the ambient temperature returns to the preset allowable deviation range, completing the closed-loop correction. The system achieves trend capture and precise closed-loop control of the laboratory load.

[0091] This invention constructs a sensing data stream with a physical mechanism background by deeply aligning infrared thermal field evolution with electrical power pulses, thus solving the problem of heat load identification deviation caused by asynchronous data from heterogeneous sensors in traditional HVAC regulation. By injecting physical mechanism operators into the dynamic prediction model, the deep learning model achieves optimization under the constraints of thermodynamic laws, eliminating the risk of prediction failure under extreme conditions inherent in purely data-driven models. The dynamic weighted allocation of feedforward compensation and feedback correction ensures that the system maintains both trend-based proactive regulation and high-precision closed-loop stability under high-power thermal shock scenarios, effectively improving the precision regulation level of laboratory ambient temperature.

[0092] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting and adaptively controlling laboratory HVAC loads, characterized in that, include: Extract the frame header time stamp from the infrared thermal imaging data acquisition terminal of the optical laboratory, and acquire the power sampling sequence from the laser power monitoring terminal. Perform field concatenation on the image pixel matrix data of the frame header time stamp and the power sampling sequence to construct the source sensing data stream and generate the thermal diffusion evolution rate. The specific methods for constructing the source-end sensing data stream and the specific methods for generating the thermal diffusion evolution rate include: The local system clocks of the infrared thermal imaging data acquisition terminal and the laser power monitoring terminal are synchronized to the same time base through a timing protocol. Each frame of infrared thermal image is analyzed in real time, and the frame header time tag and corresponding image pixel matrix data are extracted. The instantaneous power value with a deviation from the frame header time tag within a preset tolerance window is retrieved from the power sampling sequence of the laser power monitoring terminal. The image pixel matrix data and the corresponding instantaneous power value are concatenated at the field level and indexed to generate a source-end sensing data stream aligned with the time axis. The rate of change of temperature gradient from the center of the heat source radially outward in the image pixel matrix data is calculated, and the time derivative of the rate of change of temperature gradient within a continuous sampling period is taken as the thermal diffusion evolution rate. Extract the power pulse jump time of the high-power laser device and the temperature rise start time of the heat source region from the source-end sensing data stream. Determine the electrothermal conversion hysteresis characteristics based on the power pulse jump time and the temperature rise start time, and establish a heterogeneous dynamic coupling model. The specific implementation of establishing the heterogeneous dynamic coupling model includes: The power pulse jump moment of the real-time power sampling sequence is locked, and the temperature rise start moment when the temperature change rate of the heat source region exceeds a preset threshold is collected. The time difference between the temperature rise start moment and the power pulse jump moment is calculated and used as the electrothermal conversion hysteresis feature. The equivalent heat source excitation intensity is calculated using the instantaneous heat generation power of the high-power laser device, and the equivalent heat source excitation intensity is used as the source term of the equation. The structural thermal inertia constant of the high-power laser device and the air convection heat transfer coefficient are introduced to establish an unsteady heat conduction partial differential equation characterizing heat diffusion. The unsteady heat conduction partial differential equation is solved in time and space to construct a heterogeneous dynamic coupling model characterizing the dynamic correlation between power input fluctuations and the response trajectory of the indoor three-dimensional thermal field. The physical evolution parameters of the heterogeneous dynamic coupling model are extracted as prior constraint operators and injected into the hidden layer of the dynamic prediction model to reconstruct the characteristic phase space. Within the physical lag window determined by the electrothermal conversion lag characteristics, the air load increment at the end of the physical lag window is calculated. The specific steps of injecting the parameters into the hidden layer of the dynamic prediction model include: The hidden layer of the dynamic prediction model is a layer containing long short-term memory neurons. The heat transfer time delay coefficient, thermal equilibrium time constant, and thermal diffusion step response slope extracted from the heterogeneous dynamic coupling model are defined as physical evolution parameters. The physical evolution parameters are mapped to a bias correction vector that matches the dimension of the hidden layer of the dynamic prediction model using a preset weight transformation matrix. The bias correction vector is superimposed on the weighted summation calculation unit of the hidden layer neurons of the dynamic prediction model. Based on the thermal diffusion inertia law represented by the physical evolution parameters, the allocation of nonlinear mapping weights of neurons is controlled to limit the parameter optimization of the dynamic prediction model. Based on the nonlinear characteristics of valve heat transfer, an inverse mapping relationship is established to convert the air load increment into a terminal actuator response command and generate a feedforward thermal compensation command. When the thermal diffusion evolution rate exceeds the background disturbance threshold, the feedforward thermal compensation command is issued. The HVAC equipment is driven to respond to the feedforward thermal compensation command and superimpose the feedback signal. The weights of the feedforward thermal compensation command are dynamically adjusted according to the laboratory temperature regression, thus completing the closed-loop correction.

2. The laboratory HVAC load prediction and adaptive control method according to claim 1, characterized in that, The reconstructed feature phase space specifically includes: By calculating the autocorrelation function and average interaction information of the physical evolution parameter sequence, the time delay of the dynamic evolution process is determined, and the minimum embedding dimension that satisfies the manifold reconstruction condition is calculated using the pseudo-nearest neighbor algorithm. The manifold reconstruction condition is determined by calculating the proportion of pseudo-nearest neighbor points generated by trajectory self-intersection during the process of increasing the embedding dimension, and whether the proportion of pseudo-nearest neighbor points decreases to a preset convergence interval as the criterion for determining the manifold reconstruction condition. The determined time delay and embedding dimension are used to perform time-delay coordinate translation and dimension-up transformation on the source-end sensing data stream to generate a multi-dimensional phase trajectory matrix characterizing the dynamic characteristics of load evolution. The multi-dimensional phase trajectory matrix is ​​used as the input feature vector of the dynamic prediction model to obtain the implicit dynamic evolution characteristics.

3. The laboratory HVAC load prediction and adaptive control method according to claim 2, characterized in that, The air load increment calculated up to the end of the physical lag window specifically includes: The multidimensional phase trajectory matrix is ​​input into the dynamic prediction model. Through recursive updates of the hidden layer state, the predicted value of the total indoor heat load from the current sampling time to the preset physical lag time end point is calculated based on the electrothermal conversion lag characteristics. The algebraic difference between the predicted value of the total indoor heat load and the real-time basic heat load value measured by the indoor temperature control device at the current time is determined as the air load increment.

4. The laboratory HVAC load prediction and adaptive control method according to claim 1, characterized in that, The establishment of the reverse mapping relationship and the generation of feedforward thermal compensation instructions specifically include: Based on the heat exchanger energy balance criterion, a heat transfer mapping model is established to describe the relationship between air-side heat transfer power and refrigerant-side mass flow rate, mean logarithmic temperature difference, and heat transfer efficiency. The heat transfer mapping model is used to calculate the refrigerant flow rate change requirement value required to compensate for the air load increment. Combined with the linear flow characteristic curve of the HVAC equipment valve, the refrigerant flow rate change requirement value is converted into a response command of the HVAC equipment terminal actuator. The generated terminal actuator response command is then encapsulated as a feedforward heat compensation command.

5. The laboratory HVAC load prediction and adaptive control method according to claim 1, characterized in that, The specific steps for issuing the feedforward thermal compensation command include: Under non-experimental conditions, a background heat diffusion evolution rate sequence containing long-wave radiation fluctuations is acquired. The mean and standard deviation of the background heat diffusion evolution rate sequence are calculated, and the sum of the mean and three times the standard deviation is used as the background disturbance threshold. The generated feedforward thermal compensation command is compiled into an adjustment data packet following the fieldbus communication protocol and transmitted to the terminal actuator driver of the HVAC equipment through a physical communication link. The terminal actuator driver parses the adjustment data packet and generates analog voltage signals, analog current signals, and pulse modulation signals to drive the motor to move the physical valve core of the HVAC equipment valve to a preset position, thereby physically adjusting the refrigerant flow rate entering the laboratory heat exchange unit.

6. The laboratory HVAC load prediction and adaptive control method according to claim 1, characterized in that, The weights involved in the dynamically corrected feedforward thermal compensation command specifically include: The ambient temperature is acquired in real time using a temperature sensor array deployed in the laboratory. The absolute value of the deviation of the ambient temperature from the preset target value is calculated, and the derivative of the absolute value of the deviation over time is obtained to obtain the rate of decrease of the absolute value of the deviation. When the rate of decrease of the absolute value of the deviation is lower than the preset recovery threshold, the weighted allocation ratio of the feedforward thermal compensation command in the total regulation output is increased by a preset gradient step size until the ambient temperature returns to the preset allowable deviation range. Dynamic gain allocation is performed on the feedforward prediction compensation and feedback residual adjustment.

Images