LSTM-based two-path photovoltaic temperature adaptive cascade prediction control method and system

Abstract

The application discloses a two-path photovoltaic temperature adaptive cascade prediction control method and system based on LSTM, relates to the technical field of photovoltaic temperature prediction process processing, and aims to solve the problem that there is a large lag between the process inside temperature and the heating outside temperature in a photovoltaic front furnace, which leads to the problem that the LSTM prediction result and the main and auxiliary temperature loop execution response are prone to mismatching. The method combines historical multi-source operation data, the current operation boundary and the loop bearing state, constructs a two-path thermal potential prediction model based on the LSTM neural network, identifies the two-path temperature change trend and the cascade link bearing offset, generates a two-path cooperative thermal state and a bearing synchronous control state, integrates a prediction adjustment target, synchronously performs cascade adaptation, obtains two-path thermal matching parameters, and finally updates the model parameters, the two-path thermal matching parameters and / or the loop bearing relationship according to the feedback deviation relationship, so as to improve the stability and adaptability of the photovoltaic high-temperature furnace temperature cascade control.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic temperature prediction process technology, and in particular to a dual-path photovoltaic temperature adaptive cascade predictive control method and system based on LSTM. Background Technology

[0002] In the front-end manufacturing of photovoltaic cells, processes such as diffusion, oxidation, PECVD, Poly PECVD, and annealing are typically completed in tube furnaces or plate furnaces under multi-temperature zones, high temperatures, and controlled atmospheres. Crystal growth furnaces also exhibit similar high heat capacity characteristics. In such equipment, there is multi-stage heat transfer between the heating rods, furnace walls, quartz tubes, graphite boats, or silicon wafers, resulting in a significant lag between the process point temperature and the heating element temperature. Simultaneously, each furnace tube or temperature zone is affected by factors such as wafer loading, quartz boat movement, process gas flow rate, plasma discharge, cooling water fluctuations, and grid voltage fluctuations, leading to different response speeds, overshoot, and steady-state deviations during the same heating or isothermal phases.

[0003] In existing technologies, to improve the predictability of temperature control in photovoltaic high-temperature furnaces, operational data such as process point temperature, heating element temperature, set temperature, heating power, solid-state relay duty cycle, process gas flow rate, furnace pressure, and batch load are typically collected. A Long Short-Term Memory (LSTM) network model is then used to predict temperature changes in subsequent control cycles. For cascade control systems, the main temperature loop uses the actual process temperature near the silicon wafer, graphite boat, or inside the furnace cavity as the primary controlled variable. Its output does not directly drive the heater but serves as the setpoint for the auxiliary temperature loop. The auxiliary temperature loop then outputs the solid-state relay duty cycle or power regulator output ratio based on the deviation between the temperature of the heating rod, heating cylinder, or furnace wall and this setpoint to control the heating power.

[0004] However, in the cascade control scenario of photovoltaic high-temperature furnaces, the temperature inside the process circuit and the temperature outside the heating circuit are not necessarily synchronized. The temperature of the heating rod or furnace wall can change within seconds, while the actual process temperature near the silicon wafer, quartz boat, or graphite boat often responds only after tens of seconds or even longer. When the wafer load changes, the quartz boat enters or leaves the furnace, the process gas is switched, or the plasma state changes, the temperature difference and the rate of temperature rise between the two will change further. If the heating output is adjusted directly based solely on the current deviation of the main circuit or a single prediction result, it is easy to cause the outside temperature to overshoot while the inside temperature still shows a lag, followed by a concentrated rise in the inside temperature, resulting in significant overshoot.

[0005] Furthermore, tubular diffusion furnaces, oxidation furnaces, annealing furnaces, and PECVD or Poly PECVD equipment typically have multiple furnace tubes and multiple temperature measurement points. For example, multiple temperature zones can be set on the same furnace tube, and a single device can integrate multi-tube, multi-point temperature control. Uneven thermal conductivity of the graphite boat, edge heat loss, inter-tube power distribution, and changes in the process atmosphere can cause different connection relationships between the main loop temperature and the auxiliary loop temperature in each temperature zone. If only the LSTM prediction results are used as the basis for adjusting the main temperature loop, without further distinguishing between dual-path prediction deviations, temperature difference trends, and heating channel assignment status, the auxiliary loop setpoint output by the main loop will be difficult to stably match the actual heating response.

[0006] Meanwhile, during the coordinated control of the main and auxiliary temperature loops, solid-state relays, power regulators, thyristor voltage regulators, and heating rods all have output ranges, single-cycle adjustment amplitudes, and thermal inertia constraints. If the main temperature loop directly generates the auxiliary loop setpoint based on the predicted temperature without adjusting for the actual temperature rise response, heating power changes, and feedback channel assignment of the auxiliary temperature loop, a mismatch can easily occur between the predicted results, the main temperature loop's adjustment target, and the auxiliary loop's actual output. This manifests as the main loop already requiring a temperature rise or overshoot suppression in advance, but the auxiliary loop's setpoint, output duty cycle, or heating power fails to adapt to the lag in the process point temperature change.

[0007] Therefore, existing technologies still suffer from the following technical problems when photovoltaic high-temperature furnaces have multiple temperature zones and intermediate heat capacities: mismatches easily occur between prediction results, main loop targets, and auxiliary loop execution, leading to heating lag, temperature overshoot, or steady-state temperature difference imbalance. More specifically, it is necessary to reduce the mismatch between the LSTM prediction bias of the process inner temperature and the heating outer temperature, and between the temperature difference change trend and the auxiliary loop execution response bias, to reduce the impact on the generation of the main temperature loop setpoint and its conversion to the output parameters of solid-state relays or power regulators, thereby improving the temperature stability of diffusion, oxidation, PECVD, Poly PECVD, annealing, or crystal growth processes. Summary of the Invention

[0008] Based on this, embodiments of the present invention provide a dual-path photovoltaic temperature adaptive cascade predictive control method and system based on LSTM. According to the internal process temperature, the external heating temperature, the operating boundary, and the feedback from the cascade loop, predictive modeling and adaptive control are performed, enabling the main temperature loop target to be accurately converted into the auxiliary temperature loop setpoint and heating execution parameters, thereby reducing the risks of heating lag, temperature overshoot, and dual-path temperature difference imbalance. The technical solution is as follows: Firstly, a dual-path photovoltaic temperature-adaptive cascade predictive control method based on LSTM is provided, which includes: Step 1: Obtain multi-source operating data generated during the historical process of the dual-temperature zone of the photovoltaic high-temperature furnace. Combine the current operating boundary state and loop connection state of the photovoltaic high-temperature furnace, reconstruct the input features, state embedding parameters and initial connection mapping relationship of the initial prediction model constructed using LSTM neural network to obtain the dual-path thermal potential prediction model, so as to establish the prediction connection relationship between the internal temperature of the process, the external temperature of the heating, and the cascade loop connection feedback.

[0009] Step 2: Obtain the dual-path thermal potential characterization data of the photovoltaic high-temperature furnace within the current control cycle, and call the dual-path thermal potential prediction model. After dual-path thermal potential identification and offset analysis, generate the dual-path collaborative thermal state and the synchronous control state.

[0010] Step 3: Based on the dual-path collaborative thermal state, the main temperature loop target is integrated, and cascade adaptation processing is performed simultaneously based on the received synchronization control state to obtain dual-path thermal matching parameters. These parameters are then used to control the lower-level execution parameters of the auxiliary temperature loop to cascade adjust the process temperature of the photovoltaic high-temperature furnace. The main temperature loop and the auxiliary temperature loop constitute a cascade matching relationship between the upper-level process temperature prediction target and the lower-level heating execution response.

[0011] Step 4: If, during the execution of the auxiliary temperature loop, the dual-path temperature feedback deviates from the cooperative state of the dual-path cooperative thermal state, feedback training samples are generated based on the deviated state, and the model parameters, dual-path thermal matching parameters, and / or loop connection relationships of the dual-path thermal potential prediction model are adapted and updated. Otherwise, the current dual-path thermal potential prediction model and dual-path thermal matching parameters are retained to participate in the next control cycle.

[0012] Secondly, a dual-path photovoltaic temperature adaptive cascade predictive control system based on LSTM is provided, which includes: The prediction model reconstruction module is used to acquire multi-source operating data generated during the historical process of the dual-temperature zone of the photovoltaic high-temperature furnace. Combined with the current operating boundary state and loop connection state of the photovoltaic high-temperature furnace, the module reconstructs the input features, state embedding parameters, and initial connection mapping relationship of the initial prediction model constructed using an LSTM neural network to obtain a dual-path thermal potential prediction model, thereby establishing the prediction connection relationship between the internal temperature of the process, the external temperature of the heating process, and the cascade loop connection feedback.

[0013] The dual-path thermal potential identification module is used to acquire dual-path thermal potential characterization data of the photovoltaic high-temperature furnace within the current control cycle, and call the dual-path thermal potential prediction model. After dual-path thermal potential identification and offset analysis, it generates dual-path collaborative thermal state and synchronous control state.

[0014] The dual-path cascade adaptation module is used to integrate the main temperature loop target according to the dual-path collaborative thermal state, and simultaneously perform cascade adaptation processing according to the receiving synchronization control state to obtain dual-path thermal matching parameters, so as to control the lower-level execution parameters of the auxiliary temperature loop to cascade adjust the process temperature of the photovoltaic high-temperature furnace; the main temperature loop and the auxiliary temperature loop constitute a cascade matching relationship between the upper-level process temperature prediction target and the lower-level heating execution response.

[0015] The offset feedback update module is used to generate feedback training samples based on the offset state if the dual-path temperature feedback deviates from the cooperative state of the dual-path cooperative thermal state during the execution of the auxiliary temperature loop, and to adapt and update the model parameters, dual-path thermal matching parameters and / or loop connection relationship of the dual-path thermal potential prediction model. Otherwise, the current dual-path thermal potential prediction model and dual-path thermal matching parameters are retained for the next control cycle.

[0016] The beneficial effects achieved by the provided technical solution include at least the following: 1. This invention first improves the initial prediction model constructed using an LSTM neural network based on historical multi-source operating data and combined with the current operating boundary state and loop connection state of the photovoltaic high-temperature furnace, resulting in a dual-path thermal potential prediction model. This model input simultaneously reflects the relationship between the internal process temperature, the external heating temperature, and the cascade loop feedback. Subsequently, it acquires the dual-path thermal potential characterization data for the current control cycle and performs thermal potential identification and connection offset analysis through the dual-path thermal potential prediction model, generating a dual-path coordinated thermal state and a connection synchronization control state. This helps to distinguish the dual-path unidirectional bias in advance. Different temperature trends, such as deviation from single-path connection or temperature difference convergence, are identified, and the connection offset between the main and auxiliary loops is determined. Furthermore, the main loop target is integrated based on the dual-path coordinated thermal state, and cascade adaptation is performed in conjunction with the connection synchronization control state to obtain dual-path thermal matching parameters. The upper-level predicted adjustment target is converted into the setpoint correction amount and heating output correction amount that can be executed by the auxiliary temperature loop. Finally, based on the dual-path temperature feedback offset during the execution process, the model parameters, matching parameters, or loop connection relationships are adapted and updated, which helps to reduce the risks of heating lag, temperature overshoot, and steady-state temperature difference imbalance.

[0017] 2. This invention first acquires the cascade transmission data between the main temperature loop and the auxiliary temperature loop, and uses the temperature difference between the two loops, the change in heating power, the process load disturbance, and the change in temperature feedback between the two loops as the basis for judgment to determine whether the current photovoltaic high-temperature furnace is in a boundary-attached state or a boundary-detached state, thus avoiding abnormal operating conditions from directly participating in model training and causing prediction deviations. Subsequently, combining the temperature zone mapping relationship, the heating channel correspondence relationship, and the feedback sampling attribution relationship, it determines whether the heating channel of the main temperature loop, the auxiliary temperature loop, and the feedback sampling channel correspond to the same furnace tube temperature zone, and generates a mapping acceptance state or a mapping detachment state. When both the operating boundary state and the link attribution state are valid, the cascade transmission data within the same control cycle is subjected to change direction identification and amplitude correspondence processing to obtain the dual-path thermal potential correlation parameters and form an attribution acceptance sample. When a boundary detachment or mapping detachment occurs, a boundary deviation parameter is generated based on the data items that exceed the historical value range and a detachment constraint sample is formed. Finally, the two types of samples are used to update the input features, output labels, and initial acceptance mapping relationship of the initial prediction model, thereby obtaining a dual-path thermal potential prediction model that is more suitable for the current photovoltaic high-temperature furnace process state.

[0018] 3. This invention first calculates the temperature difference between the two paths based on the internal temperature of the process and the external temperature of the heating process, and then calculates the temperature rise rate difference between the two paths based on the temperature changes between the current control cycle and the previous control cycle. Subsequently, the temperature difference between the two paths and the temperature rise rate difference between the two paths are aligned within the same cycle to obtain the temperature difference trend parameters. These parameters are then combined with the boundary assignment state and the mapping acceptance state to form a dual-path thermal correlation matrix, which uniformly represents the temporal correspondence between the temperature deviation trend of the two paths, the boundary assignment result, and the cascade link acceptance result. Next, trend state values, boundary state values, and link state values ​​are extracted from the matrix of continuous control cycles and combined according to the same control cycle to form dual-path thermal characterization data. Then, the model outputs the predicted temperature deviation, predicted temperature difference change, and predicted link assignment identifier for both paths, generating dual-path same-direction deviation state, single-path acceptance deviation state, or temperature difference convergence state. Finally, the predicted link assignment identifier is compared with the link state identifier to determine the acceptance synchronous control state or the acceptance offset control state, which is beneficial for identifying whether the cascade acceptance between the main and auxiliary temperature loops is consistent.

[0019] 4. This invention first generates an upper-level predictive adjustment target for the main temperature loop based on the dual-path collaborative thermal state: when both paths deviate in the same direction, a dual-path collaborative correction target is generated by combining the predicted temperature deviation and the predicted temperature difference change; when a single path deviates, the path with the excessive deviation is used as the deviation target path and the other path as the reference path to generate a single-path correction target, avoiding premature overshoot of the outer heating temperature or continuous lag of the inner process temperature; when the temperature difference converges, a stable holding target is generated based on the predicted temperature difference change. Subsequently, only when the receiving control state is the receiving synchronous control state and the link affiliation state is the mapped receiving state, the upper-level predictive adjustment target is converted into dual-path thermal matching parameters. Finally, the auxiliary temperature loop setpoint, solid-state relay duty cycle, power regulator output ratio, heating current, or heating power are adjusted according to the deviation direction to match the lower-level heating execution response with the process temperature adjustment requirements.

[0020] 5. In the auxiliary temperature loop execution process, this invention first acquires the temperature feedback change of the first channel, the temperature feedback change of the second channel, and the actual temperature difference change. The difference between these and the expected feedback change corresponding to the dual-channel coordinated thermal state is calculated to obtain the feedback deviation, which helps determine whether the auxiliary execution result achieves the predicted adjustment intention. When both channels deviate in the same direction and both channels have feedback deviations, the main loop setting correction increment and the secondary loop output correction increment are generated respectively. When a single channel deviates, only the channel that caused the deviation is incremented. When the temperature difference convergence does not reach the expected level, the allocation ratio between the setpoint correction parameter and the heating output correction parameter is updated. Furthermore, during execution, the furnace tube identifier, the main temperature loop temperature zone identifier, the auxiliary temperature loop heating channel identifier, and the feedback sampling channel identifier are acquired. When these three do not correspond to the same furnace tube, the heating channel correspondence and feedback sampling attribution are updated according to the actual attribution of the feedback sampling channel to correct the loop connection relationship. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a dual-path photovoltaic temperature adaptive cascade predictive control method based on LSTM provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a dual-path photovoltaic temperature adaptive cascade predictive control system based on LSTM provided in an embodiment of the present invention; Figure 3 This is a dual-path temperature response trend diagram under irradiation load disturbance provided according to an embodiment of the present invention; Figure 4 This is a diagram showing the attribution and deviation of the operating boundary according to an embodiment of the present invention; Figure 5 This is a measured and predicted analysis chart of temperature and temperature difference provided according to an embodiment of the present invention; Figure 6 This is a correlation diagram of thermal control features provided according to an embodiment of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0024] In the embodiments of the present invention, "and / or" can mean both, or either one. To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the specific implementation methods, structures, features and effects of the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments.

[0025] Embodiment 1 of this invention provides a dual-path photovoltaic temperature adaptive cascade predictive control method based on LSTM. This method can be applied to polycrystalline silicon deposition processes in diffusion furnaces, oxidation furnaces, annealing furnaces, tubular PECVD, and tubular LPCVD, or to equipment with intermediate heat capacity between the heating element temperature and the process point temperature, such as crystal growth furnaces. Typical carriers include AI-8648G, AI-86X9G, or AI-8x29 series multi-loop cascade temperature controllers. Specifically, in the polycrystalline silicon deposition process of TOPCon cells, this method can be applied to multi-temperature zone tubular LPCVD equipment, such as a tubular LPCVD equipment with nine temperature zones, to perform cascade predictive control on the inner process temperature, outer heating temperature, and heating output of each zone. This cascade method is generally not used for room temperature acid / alkali cleaning, nor is it used as a typical single-loop temperature control method for module lamination, EL testing, or string bonding equipment. Specifically, it includes: Step 1 The system acquires the dual-path temperature difference between the main temperature loop and the auxiliary temperature loop, the change in heating power, the process load disturbance, and the change in dual-path temperature feedback. The dual-path temperature difference is obtained by subtracting the temperature of the heating rod, heating cylinder, or outer wall detected by the auxiliary temperature loop from the process temperature near the silicon wafer, graphite boat, or inside the cavity detected by the main temperature loop, and taking the absolute value. This represents the degree of temperature deviation between the actual process point and the rapid heating response point. The change in heating power can be obtained from the change in the duty cycle of the solid-state relay in the current control cycle, the output ratio of the power regulator, or the change in actual heating power relative to the previous control cycle. The process load disturbance can be represented by one or more changes among the following: wafer loading quantity, quartz boat entry / exit status, process gas flow rate, furnace pressure, plasma discharge status, grid voltage fluctuations, changes in heating power in adjacent temperature zones, or changes in temperature feedback in adjacent temperature zones. The change in dual-path temperature feedback can be obtained from the changes in the current process temperature and the outer heating temperature after the auxiliary temperature loop execution relative to the previous control cycle. The dual-path temperature difference is in absolute form; if it is necessary to identify the direction of deviation, a signed difference can be recorded separately.

[0026] For a nine-zone tubular LPCVD equipment, the aforementioned data can be collected for each zone according to its actual zone numbering (furnace inlet side, middle section, and furnace tail side). When a change in the heating output of a certain zone causes a temperature change in an adjacent zone through furnace tube radiation or furnace body heat conduction, the change in heating power or temperature feedback in the adjacent zone can be included in the process load disturbance corresponding to the current zone. This allows the process load disturbance to simultaneously reflect process disturbances such as wafer loading and gas switching, as well as thermal coupling disturbances between multiple zones, without requiring additional furnace body detection hardware.

[0027] like Figure 3 The dual-path temperature response trend graph shown is illustrated under process load disturbances, including... Figure 3 (a) to Figure 3 (d) Figure 3 (a) The horizontal axis represents process time, and the vertical axis represents temperature, displaying the process point temperature of the main temperature loop, the heating outside temperature of the auxiliary temperature loop, and the process set temperature; during the heating stage, the heating outside temperature changes first, followed by the process point temperature, and the two gradually approach each other during the isothermal stage. Figure 3 (b) The horizontal axis represents the process time, and the vertical axis represents the temperature difference between the two paths. This shows the change in the absolute value of the difference between the process point temperature and the outside temperature of the heating. It can be seen that the temperature difference first expands and then converges after the loading, gas switching or power adjustment. Figure 3 (c) Displays the changes in heating power ratio and process load ratio.

[0028] Figure 3(d) The horizontal axis represents process time, and the vertical axis represents the normalized proportional gain, showing the duty cycle of the solid-state relay and the fluctuations in the mains voltage. The duty cycle exhibits a phased change. When the auxiliary temperature loop drives the heating rod through a solid-state relay, power regulator, or thyristor voltage regulator unit, the output proportional gain increases during the rapid heating phase, decreases as it approaches the setpoint of the secondary loop, and forms a short-cycle pulse regulation during the isothermal phase. Therefore, the duty cycle switching between rising and falling over multiple control cycles indicates that the secondary loop is rapidly correcting the temperature outside the heating circuit, but does not necessarily mean that the process point temperature has synchronously completed its response. This change illustrates that cascade control needs to incorporate predictions based on the dual-path temperature difference, heating power, process load, and temperature feedback changes.

[0029] Figure 3 (c) and Figure 3 The normalized ratio value in (d) can be obtained by the formula xnorm=(x-xmin) / (xmax-xmin), where x is the current value, xmin is the historical minimum value, and xmax is the historical maximum value. The normalized ratio value ranges from 0 to 1 and is used to eliminate the differences in the dimensions and numerical ranges of different variables such as heating power, process load, conduction duty cycle, and grid voltage fluctuation.

[0030] If the current control cycle is the first control cycle after equipment startup or process formula switching, and there is no data from the previous control cycle, then the initial value collected at startup time will be used as the reference value; alternatively, the average value under the same furnace tube, same process formula, and same batch load conditions in historical operating data can be used as the reference value to calculate the change in heating power, the disturbance of process load, and the change in dual-path temperature feedback.

[0031] The above data are compared with the historical value ranges of the corresponding parameters in the historical operating data. The historical value range is determined by the historical minimum and maximum values ​​under the same furnace tube temperature zone, the same process stage, or the same formula conditions. The width of the historical value range is the historical maximum value minus the historical minimum value. In a nine-zone tubular LPCVD equipment, the historical value range is established separately for each actual temperature zone, and the furnace mouth, middle, and furnace tail temperature zones are not shared by default. For example, the dual-path temperature difference, heating power change, and temperature feedback change corresponding to each temperature zone can be read separately under the same POLY process formula, the same loading amount, and the same heating stage to avoid the formation of an incompatible unified boundary due to different heat dissipation conditions and heat transfer distances in different temperature zones. When all the above data are within the corresponding historical value range, the current operating boundary state of the photovoltaic high-temperature furnace is determined as the boundary belonging state, indicating that the current operating condition is still within the reference range of the historical model; when any data exceeds the corresponding historical value range, the current operating boundary state of the photovoltaic high-temperature furnace is determined as the boundary deviating state.

[0032] like Figure 4The diagram showing the furnace temperature operating boundary assignment and deviation status includes... Figure 4 (a) and Figure 4 (b) Figure 4 (a) The horizontal axis represents the normalized furnace temperature cycle, used to uniformly convert different process formulas or different heating, isothermal, and cooling times to the same stage scale for comparison. For example, a complete heating-isothermal-cooling process is considered as a furnace temperature cycle unit. The vertical axis represents the boundary ratio value, which can be calculated by dividing the current value of the variable by the historical boundary upper limit, or by dividing the absolute value of the variable change by the allowable boundary change. When the boundary ratio value is not greater than 1, it indicates that the variable is within the historical boundary; when it is greater than 1, it indicates that the variable exceeds the historical boundary. Therefore, the vertical axis can display a range from 0 to 2 to observe the extent of exceeding the boundary.

[0033] Figure 4 The curve in (a) includes the normalized results of variables such as dual-path temperature difference, heating power, process load and temperature feedback change. The light red background area indicates that at least one variable exceeds the corresponding historical value range. Figure 4 (b) The horizontal axis also represents the normalized furnace temperature cycle, while the vertical axis represents the dual-path temperature difference, heating power, process load, temperature feedback change, heating status, and boundary status. Green indicates compliance with the boundary, while red indicates boundary violation. Locally, the red and green bars are relatively close together, indicating that disturbances such as quartz boat entry / exit, gas switching, plasma discharge, or voltage fluctuations may cause the relevant variables to repeatedly exceed and revert to the boundary within multiple adjacent sampling periods.

[0034] Figure 4 The boundary states in the data can also be used to distinguish between normal temperature rise and abnormal temperature deviations. For example, in the heating phase of the POLY process, increasing the heating power is a normal control action during process acceleration. However, if the corresponding temperature difference between the two paths, the temperature rise rate difference, or the temperature feedback change between adjacent temperature zones exceeds the historical boundaries of the same temperature zone and the same process stage, it is still identified as a boundary deviance state. Therefore, the compression of heating time is not achieved by removing boundary restrictions or raising the process temperature reference, but by redistributing the heating output in the initial heating phase and the phase close to the set temperature within the established temperature reference and historical safety boundaries.

[0035] After determining the operating boundary states, the temperature zone mapping relationship, heating channel correspondence, and feedback sampling attribution relationship between the main temperature loop and the auxiliary temperature loop are obtained. The temperature zone mapping relationship represents the correspondence between the process point monitored by the main temperature loop and the furnace tube temperature zone. The heating channel correspondence represents the correspondence between the auxiliary temperature loop, solid-state relay, or power regulator output channel and the heating rod zone. The feedback sampling attribution relationship represents the correspondence between the inner thermocouple, outer thermocouple, or other temperature sensor channel and the furnace tube temperature zone. For example, if the main temperature loop temperature zone identifier, the auxiliary temperature loop heating channel identifier, and the feedback sampling channel identifier all correspond to the same furnace tube temperature zone, it means that the three belong to the same cascade transmission link.

[0036] If the main temperature loop and auxiliary temperature loop corresponding to the current photovoltaic high-temperature furnace have the same furnace tube temperature zone identifier in the temperature zone mapping relationship and heating channel correspondence relationship, and the feedback sampling channel corresponding to the dual-path temperature feedback change belongs to the heating channel corresponding to the auxiliary temperature loop in the feedback sampling attribution relationship, then the link attribution status corresponding to the current photovoltaic high-temperature furnace is determined to be the mapping acceptance status, indicating that the process point temperature, the heating outer temperature, and the heating output are in the same acceptance link; otherwise, it is determined to be the mapping departure status.

[0037] When the operating boundary state is the boundary attribution state and the link attribution state is the mapping acceptance state, the direction of change and amplitude correspondence processing are performed on the dual-path temperature difference, heating power change, process load disturbance, and dual-path temperature feedback change within the same control cycle. Specifically, the current value of each data point is calculated by subtracting it from the corresponding value of the previous control cycle. If the difference is greater than a preset stability threshold, the corresponding direction of change is set to one; if the difference is less than a negative preset stability threshold, the corresponding direction of change is set to negative one; if the difference is within the preset stability threshold range, the corresponding direction of change is set to zero. Simultaneously, the absolute value of the difference is divided by the width of the corresponding historical value range to obtain the corresponding amplitude ratio. The direction of change of each data point is then multiplied by the corresponding amplitude ratio to obtain the individual correlation value for each data point. These individual correlation values ​​are then arranged and combined in the order of dual-path temperature difference, heating power change, process load disturbance, and dual-path temperature feedback change to obtain the dual-path thermal potential correlation parameters. This parameter is used to characterize the relationship between dual-path temperature difference changes, heating heat input changes, process load disturbances, and temperature feedback changes.

[0038] The above method first determines the direction of change of each data point relative to the previous control cycle by using the difference, and then normalizes the amplitude by using the width of the historical value range, which can eliminate the differences in the dimensions and numerical ranges of different parameters. Multiplying the change direction value by the amplitude ratio can simultaneously retain the change direction and change intensity, so that the dual-path thermal potential correlation parameters reflect not only the temperature difference change trend, but also the combined influence of heating power, process load disturbance, and feedback changes on the dual-path temperature state. It should be understood that the preset stability threshold is a positive threshold used to distinguish between effective changes and stable fluctuations, and the preset stability threshold range is the interval formed by the negative preset stability threshold to the preset stability threshold. When the difference of a data point relative to the previous control cycle is greater than the preset stability threshold, the change direction of the data point is determined to be increasing; when the difference is less than the negative preset stability threshold, the change direction of the data point is determined to be decreasing; when the difference is within the preset stability threshold range, the data point is determined to be in a stable change state. The preset stability threshold can be determined based on the average fluctuation of adjacent cycles in the historical operating phase or the standard deviation of fluctuation of adjacent cycles in the historical operating phase, and different data items can be set with corresponding preset stability thresholds.

[0039] After obtaining the dual-path thermal potential correlation parameters, the dual-path thermal potential correlation parameters, the corresponding cascade transmission link, and the dual-path temperature change status in the next control cycle are used as the assigned receiving samples. The cascade transmission link is jointly determined by the furnace tube identifier, the main temperature loop temperature zone identifier, the auxiliary temperature loop heating channel identifier, and the feedback sampling channel identifier. The dual-path temperature change status in the next control cycle includes the temperature change on the inner side of the process, the temperature change on the outer side of the heating, and the temperature difference change between the two paths. The assigned receiving samples are used to characterize the sample correspondence between the thermal potential status of the current control cycle, the cascade transmission link, and the temperature response of the next control cycle. For example, the current increase in temperature difference, increase in heating power, increase in batch load, and feedback lag are correlated with the result that the outer side of the heating rises first in the next cycle, the inner side of the process continues to follow, or the temperature difference increases.

[0040] When the operational boundary state is "boundary detachment" and / or the link affiliation state is "mapping detachment," boundary deviation parameters are obtained based on data items that exceed the historical value range. If a data item is greater than the historical maximum value, the current value is subtracted from the historical maximum value and then divided by the width of the historical value range to obtain the boundary deviation amount for that data item. If a data item is less than the historical minimum value, the historical minimum value is subtracted from the current value and then divided by the width of the historical value range to obtain the boundary deviation amount for that data item. The boundary deviation amounts corresponding to each data item that exceeds the historical value range are arranged in the order of dual-path temperature difference, heating power change, process load disturbance, and dual-path temperature feedback change to form boundary deviation parameters. Each element in the boundary deviation parameters corresponds one-to-one with the data item that generates the boundary deviation amount.

[0041] Then read the dual-path temperature change status of the next control cycle corresponding to the boundary deviation parameter, and combine the boundary deviation parameter and the dual-path temperature change status into a set of unconstrained samples. If there are multiple data items that exceed the historical value range in the same control cycle, the boundary deviation values ​​corresponding to each data item are unified into the same boundary deviation parameter, and together with the dual-path temperature change status of the next control cycle corresponding to the same control cycle, they form a set of unconstrained samples.

[0042] Finally, the input features, output labels, and initial connection mapping relationship of the initial prediction model constructed using an LSTM neural network are updated based on the assigned and unconstrained samples. The initial prediction model is a Long Short-Term Memory (LSTM) network model built based on machine learning, used to predict the dual-path temperature change state of the next control cycle based on the photovoltaic high-temperature furnace operation data within a continuous control cycle. It includes an input layer, an LSTM network temporal feature extraction layer, a fully connected layer, and an output layer. During training, the dual-path temperature difference, heating power change, process load disturbance, and dual-path temperature feedback change within the continuous control cycle are used as input features. The internal process temperature change, external heating temperature change, or dual-path temperature difference change direction of the next control cycle are used as output labels. The model parameters are updated through error backpropagation. During the update process, the dual-path thermal potential correlation parameters from the assigned samples are added to the input feature sequence, and the boundary deviation parameters from the unconstrained samples are added to the input feature sequence. The link identifier in the initial connection mapping relationship is updated based on the mapping connection state and mapping unconstraint state to obtain the dual-path thermal potential prediction model.

[0043] For the POLY process, training samples can be categorized into three stages: heating up, approaching the set temperature, and isothermal deposition. The actual temperature zone identifiers for each sample are retained, allowing the model to distinguish between the permissible rapid heating response before heating and the heat transfer that needs to be limited when approaching the process set temperature. In a nine-zone system, each zone can share the model structure, but the cascade transfer links and temperature zone identifiers within the receiving samples remain independent to preserve the different response characteristics of the furnace opening, middle, and tail zones.

[0044] like Figure 5 The chart showing the measured and predicted process temperature and temperature difference includes... Figure 5 (a) to Figure 5 (c) Figure 5 (a) The horizontal axis represents the continuous time after the start of the test process cycle, and the vertical axis represents the temperature. The black curve represents the measured temperature value of the process point, and the red curve represents the predicted temperature value of the process point. It can be used to observe the fitting situation in the heating, isothermal and disturbance recovery stages. Figure 5(b) The horizontal axis is also the continuous process time, and the vertical axis is the temperature difference between the two paths. The black curve represents the measured result of the absolute value of the difference between the process point temperature and the outside temperature of the heating, and the purple curve represents the temperature difference result predicted by the dual-path thermal potential prediction model. Figure 5 (c) The horizontal axis represents the error, and the vertical axis shows the mean absolute error of the process point temperature, the root mean square error of the process point temperature, the mean absolute error of the temperature difference, and the root mean square error of the temperature difference, respectively.

[0045] Step Two Based on the process internal temperature participating in the predictive regulation of the main temperature loop and the heating external temperature participating in the execution feedback of the auxiliary temperature loop, the temperature difference between the two loops is calculated; the temperature change of the two loops between the current control cycle and the previous control cycle is obtained; the difference of the process internal temperature between the two control cycles is divided by the control cycle duration to obtain the temperature rise rate of the main temperature loop; the temperature difference of the heating external temperature between the two control cycles is divided by the control cycle duration to obtain the temperature rise rate of the auxiliary temperature loop; and the absolute value of the difference between the two is determined as the temperature rise rate difference between the two loops.

[0046] The temperature difference between the two channels and the temperature rise rate difference between the two channels are aligned in time according to the same control cycle. The temperature difference between the two channels is used as the temperature deviation, and the temperature rise rate difference between the two channels is used as the deviation change. The two channels temperature difference trend parameter is obtained by combining them in the order of the control cycle. This parameter is used to simultaneously represent the degree of temperature deviation and the rate of change of the deviation. For example, in the t-th control cycle, if the absolute value of the difference between the temperature zone of the main temperature loop and the feedback temperature zone of the auxiliary temperature loop is the temperature difference between the two channels, and this temperature difference continues to increase relative to the (t-1)-th control cycle, while the temperature rise rate difference between the two channels also increases, then the corresponding temperature difference trend parameter indicates that the temperature deviation between the two channels is aggravated. If the temperature difference between the two channels decreases and the temperature rise rate difference between the two channels decreases, then the corresponding temperature difference trend parameter indicates that the temperature deviation between the two channels is converged.

[0047] The dual-path temperature difference trend parameters, boundary attribution status, and mapping acceptance status are combined using state encoding to obtain a dual-path thermal correlation matrix. Specifically, each control cycle is treated as a row of the matrix, and the trend state values ​​corresponding to the dual-path temperature difference trend parameters, the boundary state values ​​corresponding to the boundary attribution status, and the link state values ​​corresponding to the mapping acceptance status are treated as columns of the matrix, filled into the corresponding matrix cells in the order of the control cycles. The boundary attribution status and boundary detachment status can be denoted as 1 and 0, respectively; the mapping acceptance status and mapping detachment status can be denoted as 1 and 0, respectively; the dual-path temperature difference trend parameters can be determined based on the comprehensive change direction of the dual-path temperature difference and the dual-path temperature rise rate difference within the same control cycle. When the comprehensive change direction is increasing, decreasing, or stable, the trend state value is encoded as 1, -1, and 0, respectively. The rows of the dual-path thermal correlation matrix correspond to continuous control cycles, and the columns correspond to the trend state values, boundary state values, and link state values.

[0048] The trend state values ​​corresponding to the temperature difference trend parameters of the two paths, the boundary state values ​​corresponding to the boundary attribution state, and the link state values ​​corresponding to the mapping acceptance state are extracted from the dual-path thermal correlation matrix within a continuous preset number of control cycles. Following the control cycle order, the trend state values, boundary state values, and link state values ​​within the same control cycle are grouped into a set of thermal state vectors. A continuous preset number of these thermal state vector sets constitute the dual-path thermal characterization data. The dual-path thermal characterization data is input into the dual-path thermal potential prediction model. After time-series feature extraction, state encoding parsing, and prediction mapping processing, the model outputs the predicted temperature deviation on the inner side of the process, the predicted temperature deviation on the outer side of the heating process, the predicted temperature difference change, and the predicted link attribution identifier.

[0049] The preset quantity here is usually determined based on the heat transfer lag time of the photovoltaic high-temperature furnace, the control cycle length, and the input window length of the dual-path thermal potential prediction model. For example, when the temperature of the outer heating element changes within a few seconds, but the process point temperature requires multiple control cycles to form a stable feedback, the preset quantity can be set to an integer covering this lag period; in tubular PECVD or diffusion furnaces, it can also be determined based on the response time corresponding to the heat capacity of the graphite boat, quartz boat, and furnace tube.

[0050] When the absolute values ​​of both the first and second predicted temperature deviations are greater than the preset deviation thresholds, and their deviation directions are the same, the current dual-path coordinated thermal state is recorded as a dual-path same-direction deviation state. When the absolute value of either predicted temperature deviation is greater than the preset deviation threshold, and the other is not greater than the preset deviation threshold, it is recorded as a single-path acceptance deviation state. When the predicted temperature difference change is less than the preset temperature difference convergence threshold, it is recorded as a temperature difference convergence state. The preset deviation threshold can be determined based on the average temperature fluctuation during a historical stable process stage, the thermocouple acquisition accuracy, or the allowable process temperature difference. The preset temperature difference increase threshold and the preset temperature difference convergence threshold can be determined based on the steady-state temperature difference range of different formulations.

[0051] The predicted link attribution identifier is compared with the current link attribution status. If the current link attribution status is a mapped acceptance status, and the predicted link attribution identifier matches the link status identifier corresponding to the current mapped acceptance status, then the acceptance control status of the current photovoltaic high-temperature furnace in the current control cycle is determined as the acceptance synchronization control status. If the current link attribution status is a mapped decoupling status, or the predicted link attribution identifier does not match the link status identifier corresponding to the current mapped acceptance status, then it is determined as the acceptance offset control status. The acceptance synchronization control status indicates that the main temperature loop process point, the auxiliary temperature loop heating outer measuring point, and the solid-state relay or power regulator output channel are in the same effective cascade link.

[0052] Step 3 In practice, the operator or the host process system sets the target process temperature Tset for the main temperature loop. Based on the deviation between the internal process temperature and Tset, and the dual-path coordinated thermal state, the main temperature loop generates an auxiliary temperature loop setpoint SV2. The auxiliary temperature loop then outputs a solid-state relay duty cycle of 0 to 100% or a power regulator control signal of 4 to 20 mA, based on the deviation between the external heating temperature and SV2, to drive the heating rod. The main temperature loop does not directly drive the solid-state relay; instead, it determines the intermediate target temperature that the external heating temperature should reach. The auxiliary temperature loop is responsible for quickly following this target, thus primarily confining external disturbances within the external fast-track loop.

[0053] When using the AI-8648G high-performance intelligent multi-channel temperature controller, its cascade control channel can execute the auxiliary temperature loop setpoint SV2 and the lower-level heating output. Self-tuning and PID regulation with minimal overshoot maintain stable external heating temperature. The LSTM model is used for upper-level continuous prediction and SV2 correction, without replacing the lower-level PID execution within the temperature controller. Therefore, the output handling process during the heating phase can be optimized while retaining the original heater, thermocouple, solid-state relay, or power regulator.

[0054] In the POLY process, the target process temperature Tset continues to use the original process formula setting value. The reduction in heating time is mainly achieved by improving the effective heating utilization rate in the early stage of heating, and by reducing SV2 or heating output in advance when it is predicted that the residual heat in the furnace wall, quartz boat and furnace carrier will continue to be transferred to the process point, rather than by increasing Tset to form a larger temperature margin.

[0055] When the dual-path coordinated thermal state is a state of dual-path deviation in the same direction, if both the process inner temperature and the heating outer temperature are lower than their respective targets, then SV2 is increased and the heating output is increased; if both are higher than the targets, then SV2 is decreased and the heating output is reduced to synchronously correct the deviation of the two paths. When the state is a state of single-path deviation, if the process inner temperature is still lower than Tset while the heating outer temperature is close to or higher than SV2, then the heating output is limited and the process waits for the inner temperature to catch up to avoid subsequent concentrated overshoot; if the heating outer temperature fails to follow SV2, then the secondary loop PID or heating output is adjusted first. A stable target is generated when the temperature difference converges.

[0056] The upper-level predictive regulation target is converted into dual-path thermal matching parameters only when the receiving control state is the receiving synchronous control state and the link affiliation state is the mapping receiving state. If the above conditions are not met, the parameter conversion is paused and the correspondence between the furnace tube temperature zone, heating output channel, and feedback sampling channel is checked. Specifically, the predicted deviation amount inside the process, the predicted deviation amount outside the heating, and the predicted temperature difference change amount are first normalized to obtain the main loop target correction ratio, the secondary loop setting correction ratio, and the heating output correction ratio. Then, the total regulation intensity is formed by combining the deviation direction and allocated to the SV2 correction amount and the heating output correction amount according to the receiving relationship between the slow response of the main loop and the fast response of the secondary loop. The dual-path thermal matching parameters represent the regulation intensity or allocation ratio of the cascade setting and execution output, and are not directly equivalent to the temperature value.

[0057] When the upper-level predicted adjustment target is a dual-path collaborative correction target, the main circuit target correction amount is used to adjust the auxiliary temperature circuit setpoint SV2, and the heating output correction amount is used to adjust the solid-state relay duty cycle, power regulator output ratio, thyristor conduction angle, heating current, or heating power. When the temperature is lower than the target, the adjustment is made in the direction of increasing heating; when the temperature is higher than the target, the adjustment is made in the direction of decreasing heating. The adjustment amount can be obtained by multiplying the current output parameter by the corresponding matching parameter ratio, and is constrained by the equipment's allowable upper and lower limits and the maximum change amplitude per cycle.

[0058] When the upper-level predictive adjustment target is a single-path receiving correction target, if the deviation occurs at the internal process temperature while the external heating temperature has already reached the intermediate target, then SV2 is corrected first, and the duty cycle is limited to continue increasing, in order to wait for heat to be transferred from the heating rod, furnace wall, and quartz tube to the silicon wafer or graphite boat. If the deviation occurs at the external heating temperature, then the secondary loop PID output or power regulator output ratio is corrected first, so that the external heating temperature quickly follows SV2. This can distinguish between insufficient execution of the fast loop and lag in heat transfer of the slow loop, reducing false heating and over-adjustment.

[0059] When the upper-level predicted adjustment target is to maintain stability, the auxiliary temperature loop maintains the current SV2, solid-state relay duty cycle, power regulator output ratio, heating current or heating power, while simultaneously monitoring the internal process temperature, external heating temperature, and actual temperature difference changes in real time. If any prediction deviation exceeds the preset deviation threshold again, or if the actual temperature difference changes from convergence to increase, the stability maintenance target is stopped, and a new dual-path collaborative correction target or a single-path correction target is generated.

[0060] It's important to understand that the aforementioned lower-level execution parameters do not require the solid-state relays, power regulators, and thyristor voltage regulators to operate simultaneously. Instead, the appropriate output method is selected based on the actual configuration of the photovoltaic high-temperature furnace. When a solid-state relay is used in the auxiliary temperature circuit, the dual-path thermal matching parameters are used to calculate the duty cycle. When a power regulator or thyristor voltage regulator is used, it can be used to calculate the output ratio, conduction angle, heating current, or heating power. For multi-tube, multi-temperature zone equipment, cascade links can be formed according to the furnace tubes and temperature measurement points. For example, a six-tube, nine-point equipment can form multiple independent main and auxiliary circuits that are constrained by the total power.

[0061] Step Four During the execution of the auxiliary temperature loop according to the dual-path thermal matching parameters, the changes in the internal process temperature and the external heating temperature relative to the previous control cycle, as well as the actual temperature difference between the two paths, are acquired. Based on the dual-path coordinated thermal state, the corresponding expected feedback change is determined. The difference between the actual feedback change and the corresponding expected feedback change is calculated to obtain the feedback deviation. The larger the absolute value of the feedback deviation, the higher the degree of inconsistency between the current heating execution response and the predicted result.

[0062] When the feedback deviation exceeds a preset feedback deviation threshold, the dual-path thermal characterization data of the current control cycle, the predicted temperature change output by the dual-path thermal potential prediction model, the predicted link attribution identifier, the actual feedback change, and the feedback deviation constitute the feedback training sample. The dual-path thermal characterization data serves as the model input sample, the actual feedback change serves as the temperature prediction correction label, the actual link attribution status serves as the link attribution correction label, and the feedback deviation serves as the basis for sample correction. Based on the feedback training sample, the LSTM time-series feature extraction layer parameters, state embedding parameters, and output layer parameters in the dual-path thermal potential prediction model are updated. Specifically, the LSTM time-series feature extraction layer parameters are used to correct the temporal relationship between temperature difference changes, heating power changes, and process load disturbances within the continuous control cycle; the state embedding parameters are used to correct the influence of boundary attribution status and loop connection status on the prediction results; and the output layer parameters are used to correct the predicted temperature change and the predicted link attribution identifier.

[0063] When the dual-path coordinated thermal state is a state of unidirectional deviation and both paths have feedback deviations, the feedback deviation corresponding to the temperature inside the process is multiplied by a preset parameter conversion coefficient to obtain the main loop setting correction increment; the feedback deviation corresponding to the temperature outside the heating is multiplied by the preset parameter conversion coefficient to obtain the secondary loop output correction increment. The parameter conversion coefficient is used to represent the adjustment amount of SV2, solid-state relay duty cycle, or power regulator output ratio corresponding to a unit temperature deviation, and can be determined based on the historical response of the heating rod and furnace body. The two types of parameter increments are respectively superimposed on the original dual-path thermal matching parameters to obtain the updated matching parameters used in the next control cycle.

[0064] Through the above processing, on the one hand, the dual-path thermal potential prediction model is updated using feedback training samples, so that the subsequent prediction results are closer to the actual temperature response; on the other hand, the dual-path thermal matching parameters are updated using parameter increments, so that the auxiliary temperature loop setpoint and heating output intensity in the next control cycle match the actual feedback deviation.

[0065] It's important to understand that the feedback deviation represents the difference between the actual and expected changes in the internal and external temperatures of the process after the auxiliary temperature loop is executed. A larger feedback deviation indicates a greater mismatch between the current SV2 or heating output and the predicted adjustment target. Therefore, it's necessary to increase or decrease the adjustment along the corresponding deviation direction based on the original matching parameters, rather than regenerating completely unrelated control parameters. This preserves the effective adjustment direction established in the previous control cycle while allowing for incremental correction based on the actual degree of inadequacy, reducing sudden output changes.

[0066] When the dual-path coordinated thermal state is in a single-path receiving deviation state, the parameter increment is only added to the dual-path thermal matching parameters corresponding to the path that has deviated, while the reference path maintains its original matching parameters or only makes minor adjustments to limit the amplitude, to avoid continuously increasing the output when the external heating temperature has reached the target. When the dual-path coordinated thermal state is in a temperature difference convergence state, the temperature difference deviation is obtained by subtracting the expected temperature difference change from the actual temperature difference change, and the distribution ratio between the SV2 correction and the heating output correction is updated according to the temperature difference deviation, so that the fast response of the secondary circuit and the slow reception of the primary circuit are coordinated.

[0067] In the first embodiment described above, a method for updating dual-path thermal matching parameters based on feedback deviation is presented. However, the generation of feedback deviation is not necessarily caused solely by insufficient heating execution response; it may also be due to inconsistencies in the attribution relationships between the main temperature loop process point, the auxiliary temperature loop heating channel, and the feedback sampling channel. For example, the predicted adjustment target formed by the main temperature loop corresponds to the first zone temperature of furnace tube A, while the actual solid-state relay channel or feedback sampling channel executed by the auxiliary temperature loop corresponds to furnace tube B. In this case, even if the heating output has been adjusted, the collected feedback changes still cannot accurately reflect the temperature response of furnace tube A. Therefore, it is necessary to determine whether the loop attribution relationships are consistent before adding parameter increments.

[0068] Example 2 During the execution of the auxiliary temperature loop, the furnace tube identifier, main temperature loop temperature zone identifier, auxiliary temperature loop heating channel identifier, and feedback sampling channel identifier corresponding to the current photovoltaic high-temperature furnace are obtained, and the furnace tube and temperature zone corresponding to each identifier are determined respectively. When the main temperature loop process point, auxiliary temperature loop heating channel, and feedback sampling channel all correspond to the same furnace tube temperature zone, the loop connection relationship is determined to be consistent; when the three do not correspond to the same furnace tube temperature zone, the loop connection relationship is determined to be offset.

[0069] Furnace tube markings are used to distinguish different tube furnaces or different process chambers; main temperature loop temperature zone markings are used to indicate the target process point near the silicon wafer, graphite boat, or inside the chamber; auxiliary temperature loop heating channel markings are used to indicate the output channels of solid-state relays, power regulators, or thyristors that actually drive the corresponding heating rod zones; feedback sampling channel markings are used to indicate thermocouple channels that collect the temperature inside the process or the temperature outside the heating process.

[0070] After determining that there is a misalignment in the loop connection relationship, the actual furnace tube temperature zone to which the feedback sampling channel identifier belongs is read. The actual assignment can be obtained from the temperature controller channel configuration table, thermocouple installation calibration record, heating rod wiring record, most recent manual verification record, or historical response consistency record; when there is inconsistency among the above records, the latest calibrated and active record shall prevail. If the actual assignment can be determined, the auxiliary temperature loop heating channel identifier and the furnace tube temperature zone are re-established to correspond, and the feedback sampling assignment relationship is updated synchronously; if it cannot be determined, the current link status is recorded as a mapping disconnection state, and the update of dual-path thermal matching parameters based on the feedback channel is suspended.

[0071] After re-correspondence, the updated heating channel correspondence and feedback sampling attribution are written into the current photovoltaic high-temperature furnace's corresponding connection mapping, forming an updated loop connection relationship. The updated loop connection relationship includes at least the furnace tube identifier, main temperature loop temperature zone identifier, auxiliary temperature loop heating channel identifier, feedback sampling channel identifier, update time, and update source marker, used for link attribution determination in the next control cycle. This process avoids inconsistencies between the SV2 output of the main temperature loop, the heating output of the auxiliary temperature loop, and the feedback sampling data source, which could lead to the heating quantity being sent to the wrong temperature zone.

[0072] Example 3 First, determine if the feedback deviation is greater than a preset feedback deviation threshold, then determine if there is a shift in the loop connection relationship. When the feedback deviation is greater than the preset feedback deviation threshold and the loop connection relationship is consistent, incremental updates are performed on the dual-path thermal matching parameters according to Example 1. When the feedback deviation is not greater than the preset feedback deviation threshold but the loop connection relationship is shifted, only the loop connection relationship is updated according to Example 2. When the feedback deviation is greater than the preset feedback deviation threshold and the loop connection relationship is also shifted, the loop connection relationship is updated first, then the feedback deviation is recalculated based on the updated heating channel correspondence, and the dual-path thermal matching parameters are updated based on the recalculated feedback deviation. This ensures that the parameter increment is superimposed on the correct heating channel, avoiding amplification of control deviation due to incorrect link attribution.

[0073] The embodiments of the present invention provide, as follows Figure 1The flowchart shown illustrates a dual-path photovoltaic temperature adaptive cascade predictive control method based on LSTM. Through the aforementioned processing, the matching between process-internal temperature prediction, main loop target generation, and secondary loop heating execution feedback can be improved, reducing the risks of heating lag, temperature overshoot, and steady-state temperature difference imbalance. This invention provides, as shown in the embodiments,... Figure 2 The diagram shows the structure of a dual-path photovoltaic temperature adaptive cascade predictive control system based on LSTM. This system is used to execute the above control method and achieves coordinated prediction and adaptive cascade adjustment of the photovoltaic high-temperature furnace process temperature through predictive model reconstruction, dual-path thermal potential identification, dual-path cascade adaptation, and offset feedback update.

[0074] It should be added that, such as Figure 6 The correlation diagram of the thermal control characteristics of the high-temperature furnace shown has the following variables on both the horizontal and vertical axes: process point temperature, external heating temperature, temperature difference between the two circuits, temperature feedback change, solid-state relay duty cycle, heating power, process load, process atmosphere flow rate, and grid voltage fluctuation. The number in each cell represents the Pearson correlation coefficient between the two variables; warm colors indicate a positive correlation, cool colors indicate a negative correlation, and darker colors indicate a stronger correlation. The correlation diagram is used to illustrate the statistical association between variables and does not directly represent a causal control relationship. Figure 6 It can be seen that there is a strong correlation between the process point temperature and the external heating temperature, heating power and conduction duty cycle. The temperature difference between the two paths may also show a positive or negative correlation with the process load or atmosphere change. This indicates that the above variables can be used as candidate variables for thermal potential prediction and cascade control analysis.

[0075] The above combination Figures 3 to 6 The following provides an illustrative description of the internal temperature, external temperature, disturbance response, boundary assignment, prediction error, and correlation of thermal control variables in a photovoltaic high-temperature furnace. The curve shapes, data values, state colors, normalization ratios, and error indices in the accompanying figures are only used to illustrate the processing relationships between multi-source operational data acquisition, dual-path thermal potential characterization, boundary state judgment, prediction result verification, and cascade control variable correlation analysis. They do not limit specific data sources, sampling periods, number of variables, normalization methods, or threshold values. In actual implementation, the data in the figures can be replaced or adjusted based on the number of furnace tubes, temperature zone locations, sensor accuracy, heating actuator type, and historical process range of diffusion, oxidation, PECVD, Poly PECVD, annealing, or crystal growth equipment. Any simple modifications, equivalent changes, or alterations made by those skilled in the art without departing from the scope of this invention should be considered within the protection scope of this invention.

Claims

1. A dual-path photovoltaic temperature adaptive cascade predictive control method based on LSTM, characterized in that, The method includes: Step 1: Obtain multi-source operating data of the dual-temperature zone of the photovoltaic high-temperature furnace during the historical process. Combine the current operating boundary state and loop connection state of the photovoltaic high-temperature furnace, reconstruct the input features, state embedding parameters and initial connection mapping relationship of the initial prediction model constructed using LSTM neural network to obtain the dual-path thermal potential prediction model, so as to establish the prediction connection relationship between the internal temperature of the process, the external temperature of the heating and the cascade loop connection feedback. Step 2: Obtain the dual-path thermal potential characterization data of the photovoltaic high-temperature furnace within the current control cycle, and call the dual-path thermal potential prediction model. After dual-path thermal potential identification and offset analysis, generate the dual-path collaborative thermal state and the synchronous control state. Step 3: Based on the dual-path collaborative thermal state, the main temperature loop target is integrated, and cascade adaptation processing is performed simultaneously based on the received synchronous control state to obtain dual-path thermal matching parameters, so as to control the lower-level execution parameters of the auxiliary temperature loop to cascade adjust the process temperature of the photovoltaic high-temperature furnace. The main temperature loop and the auxiliary temperature loop constitute a cascaded matching relationship between the upper-level process temperature prediction and adjustment target and the lower-level heating execution and response. Step 4: If, during the execution of the auxiliary temperature loop, the dual-path temperature feedback deviates from the cooperative state of the dual-path cooperative thermal state, then feedback training samples are generated based on the deviated state, and the model parameters, dual-path thermal matching parameters, and / or loop connection relationships of the dual-path thermal potential prediction model are adapted and updated. Otherwise, the current dual-path thermal potential prediction model and dual-path thermal matching parameters are retained to participate in the next control cycle.

2. The LSTM-based dual-path photovoltaic temperature adaptive cascade predictive control method as described in claim 1, characterized in that, Step one includes: Acquire cascaded transfer data between the main temperature loop and the auxiliary temperature loop. The cascaded transfer data includes the dual-path temperature difference, heating power change, process load disturbance, and dual-path temperature feedback change. When all of the above data are within the corresponding historical value range, the current operating boundary state of the photovoltaic high-temperature furnace is determined as the boundary belonging state; otherwise, it is determined as the boundary detachment state. Obtain the temperature zone mapping relationship, heating channel correspondence, and feedback sampling attribution relationship between the main temperature loop and the auxiliary temperature loop; If the main temperature circuit and auxiliary temperature circuit corresponding to the current photovoltaic high-temperature furnace have the same furnace tube temperature zone identifier in the temperature zone mapping relationship and heating channel correspondence relationship, and the feedback sampling channel corresponding to the dual-path temperature feedback change belongs to the heating channel corresponding to the auxiliary temperature circuit in the feedback sampling attribution relationship, then the link attribution status corresponding to the current photovoltaic high-temperature furnace is determined to be the mapping acceptance status; otherwise, it is determined to be the mapping disengagement status.

3. The LSTM-based dual-path photovoltaic temperature adaptive cascade predictive control method as described in claim 2, characterized in that, Step one also includes: When the operating boundary state is the boundary attribution state and the link attribution state is the mapping acceptance state, the dual-path temperature difference, heating power change, process load disturbance and dual-path temperature feedback change within the same control cycle are identified by change direction and processed by amplitude correspondence to obtain dual-path thermal potential correlation parameters. The dual-path thermal potential correlation parameters, the corresponding cascade transmission links, and the dual-path temperature change status in the next control cycle are used as the assigned receiving samples. When the running boundary state is the boundary detachment state, and / or the link affiliation state is the mapping detachment state, the boundary deviation parameter is obtained based on the data item that exceeds the historical value range, and the boundary deviation parameter and the corresponding dual-path temperature change state are used as the detachment constraint sample. The input features, output labels, and initial connection mapping relationships of the initial prediction model constructed using the LSTM neural network are updated based on the assigned and unconstrained samples to obtain the dual-path thermal potential prediction model.

4. The LSTM-based dual-path photovoltaic temperature adaptive cascade predictive control method as described in claim 3, characterized in that, Step two includes: The temperature difference between the two paths is calculated based on the process temperature of the first path and the heating temperature of the second path. The temperature rise rate difference between the two paths is calculated based on the temperature change of the two paths in the current control cycle and the previous control cycle. The first process temperature and the second heating temperature correspond to the temperature inside the silicon wafer, graphite boat or process cavity that the main temperature loop participates in predictive regulation, and the temperature outside the heating rod, furnace wall or heating cavity that the auxiliary temperature loop participates in execution feedback. The temperature difference between the two paths and the temperature rise rate difference between the two paths are aligned with the same period to obtain the temperature difference trend parameter for characterizing the temperature change trend of the two paths. The parameter is then combined with the boundary attribution state and the mapping acceptance state to obtain the thermal correlation matrix of the two paths. Extract the trend state value, boundary state value, and link state value corresponding to the dual-path thermal correlation matrix within the continuous control cycle, which are the trend state values ​​corresponding to the dual-path temperature difference trend parameters, the boundary state values ​​corresponding to the boundary attribution state, and the link state values ​​corresponding to the mapping acceptance state. The trend state values, boundary state values, and link state values ​​within the same control cycle are combined to obtain dual-path thermal characterization data, which is then input into the dual-path thermal potential prediction model for dual-path thermal potential identification and offset analysis.

5. The LSTM-based dual-path photovoltaic temperature adaptive cascade predictive control method as described in claim 4, characterized in that, The dual-path thermal potential identification and bearing offset analysis include: Obtain the predicted temperature deviation of the first path, the predicted temperature deviation of the second path, the predicted temperature difference change, and the predicted link affiliation identifier; When the absolute values ​​of the first predicted temperature deviation and the second predicted temperature deviation are both greater than the preset deviation threshold, and the two deviation directions are the same, a dual-path same-direction deviation state is generated. When the absolute value of the predicted temperature deviation of any one path is greater than the preset deviation threshold, and the absolute value of the predicted temperature deviation of the other path is not greater than the preset deviation threshold, a single-path acceptance deviation state is generated. When the predicted temperature difference change is less than the preset temperature difference convergence threshold, a temperature difference convergence state is generated. The predicted link affiliation identifier is compared with the link status identifier corresponding to any of the above states. If they match, the current photovoltaic high-temperature furnace is determined to be in the current control cycle as the receiving synchronization control state; otherwise, it is determined to be in the receiving offset control state.

6. The LSTM-based dual-path photovoltaic temperature adaptive cascade predictive control method as described in claim 5, characterized in that, Step three includes: When the dual-path collaborative thermal state is a dual-path same-direction deviation state, the predicted temperature deviation of the first path, the predicted temperature deviation of the second path, and the predicted temperature difference change are used together as the main temperature loop target integration basis to generate a dual-path collaborative correction target for synchronously correcting the deviation of the internal temperature of the process and the external temperature of the heating. When the dual-path collaborative thermal state is a single-path acceptance deviation state, one path that meets the preset deviation conditions is taken as the deviation target path, and the other path is taken as the reference path, generating a single-path acceptance correction target to correct the deviation target path and limit the continued expansion of the temperature difference between the two paths. When the dual-path collaborative thermal state is a temperature difference convergence state, a stable maintenance target is generated based on the predicted temperature difference change to maintain the temperature difference convergence trend of the dual paths. The dual-path collaborative correction target, single-path receiving correction target, or stable maintenance target are determined as the upper-level predictive adjustment target of the main temperature loop within the current control cycle.

7. The LSTM-based dual-path photovoltaic temperature adaptive cascade predictive control method as described in claim 6, characterized in that, Step three also includes: Only when the acceptance control state is the acceptance synchronization control state and the link affiliation state is the mapping acceptance state, the upper-layer prediction adjustment target is assigned parameters to the first predicted temperature deviation, the second predicted temperature deviation and the predicted temperature difference change to obtain dual-path thermal matching parameters. When the upper-level predictive adjustment target is a dual-path collaborative correction target, the dual-path thermal matching parameters are converted into the set value correction amount and heating output correction amount of the auxiliary temperature loop. Based on the deviation direction, one or more of the following are simultaneously increased or decreased: solid-state relay duty cycle, power regulator output ratio, thyristor conduction angle, heating current, and heating power, so that the lower-level execution parameters are adjusted towards the target temperature correction direction. When the upper-level predictive adjustment target is a single-path acceptance correction target, the dual-path thermal matching parameters are preferentially used to correct the auxiliary temperature loop set value or heating output, and the parameter increment corresponding to the other temperature feedback is limited, so as to avoid the outer heating temperature overshooting first or the inner process temperature continuously lagging behind. When the upper-level predicted adjustment target is to maintain stability, the auxiliary temperature circuit maintains the current set value, the solid-state relay duty cycle, the power regulator output ratio, the heating current, or the heating power.

8. The LSTM-based dual-path photovoltaic temperature adaptive cascade predictive control method as described in claim 7, characterized in that, Step four includes: During the execution of the auxiliary temperature loop, the change in temperature feedback from the first channel, the change in temperature feedback from the second channel, and the actual change in temperature difference are obtained. Based on the dual-path coordinated thermal state, the corresponding expected feedback change is determined, and the difference between the expected feedback change and the actual temperature difference change is calculated to obtain the feedback deviation. The expected feedback change represents the target change result that the process inner temperature and the heating outer temperature should achieve after the auxiliary temperature loop is executed according to the current dual-path thermal matching parameters. When the dual-path collaborative thermal state is a dual-path same-direction deviation state and both paths have feedback deviation, the first parameter increment and the second parameter increment are generated according to the temperature feedback change of the first path and the temperature feedback change of the second path, and are superimposed on the auxiliary temperature loop set value correction parameter and the heating output correction parameter. When the dual-path collaborative hot state is a single-path acceptance deviation state, the parameter increment is only added to the dual-path hot state matching parameter corresponding to the path that caused the acceptance deviation. When the dual-path collaborative thermal state is in the temperature difference convergence state, the allocation ratio between the auxiliary temperature loop setpoint correction parameter and the heating output correction parameter is updated based on the difference between the actual temperature difference change and the expected temperature difference change.

9. The LSTM-based dual-path photovoltaic temperature adaptive cascade predictive control method as described in claim 7, characterized in that, Step four also includes: During the execution of the auxiliary temperature loop, the furnace tube identifier, main temperature loop temperature zone identifier, auxiliary temperature loop heating channel identifier, and feedback sampling channel identifier corresponding to the current photovoltaic high-temperature furnace are obtained; When the main temperature loop temperature zone identifier, auxiliary temperature loop heating channel identifier, and feedback sampling channel identifier do not correspond to the same furnace tube identifier, read the furnace tube identifier to which the feedback sampling channel identifier actually belongs. At the same time, the correspondence between the auxiliary temperature circuit heating channel identifier and the actual furnace tube identifier is re-established, and the attribution relationship between the feedback sampling channel identifier and the furnace tube identifier is updated. The updated heating channel correspondence and feedback sampling attribution relationship are then written into the initial acceptance mapping relationship corresponding to the current photovoltaic high-temperature furnace to form the updated loop acceptance relationship, which is used for link attribution judgment in the next control cycle.

10. A dual-path photovoltaic temperature adaptive cascade predictive control system based on LSTM, characterized in that, The system includes: The prediction model reconstruction module is used to acquire multi-source operating data formed during the historical process of the dual-temperature zone of the photovoltaic high-temperature furnace. Combined with the current operating boundary state and loop connection state of the photovoltaic high-temperature furnace, the input features, state embedding parameters and initial connection mapping relationship of the initial prediction model constructed by LSTM neural network are reconstructed to obtain the dual-path thermal potential prediction model, so as to establish the prediction connection relationship between the internal temperature of the process, the external temperature of the heating and the cascade loop connection feedback. The dual-path thermal potential identification module is used to acquire dual-path thermal potential characterization data of the photovoltaic high-temperature furnace in the current control cycle, and call the dual-path thermal potential prediction model. After dual-path thermal potential identification and offset analysis, it generates dual-path collaborative thermal state and synchronous control state. The dual-path cascade adaptation module is used to integrate the main temperature loop target according to the dual-path collaborative thermal state, and simultaneously perform cascade adaptation processing according to the receiving synchronous control state to obtain dual-path thermal matching parameters, so as to control the lower-level execution parameters of the auxiliary temperature loop to cascade adjust the process temperature of the photovoltaic high-temperature furnace. The main temperature loop and the auxiliary temperature loop constitute a cascaded matching relationship between the upper-level process temperature prediction and adjustment target and the lower-level heating execution and response. The offset feedback update module is used to generate feedback training samples based on the offset state if the dual-path temperature feedback deviates from the cooperative state of the dual-path cooperative thermal state during the execution of the auxiliary temperature loop, and to adapt and update the model parameters, dual-path thermal matching parameters and / or loop connection relationship of the dual-path thermal potential prediction model. Otherwise, the current dual-path thermal potential prediction model and dual-path thermal matching parameters are retained for the next control cycle.

Images