Adaptive Snow Melting Method and System for Photovoltaic Snow Sheds

Abstract

This invention relates to the field of intelligent control technology and discloses an adaptive snow melting method and system for photovoltaic snow-proof sheds. The method includes: collecting multimodal real-time data of the photovoltaic panel surface; calling a digital twin physical model to calculate a first heating control parameter corresponding to the theoretical energy requirement for melting the current snow accumulation; calling a neural network model to predict a second heating control parameter; dynamically weighting and fusing the first and second heating control parameters using a weighting factor to generate a final heating control parameter; driving a heating unit to perform snow melting operations on the photovoltaic panel according to the final heating control parameter, and updating the weighting factor and / or the neural network model parameters based on real-time feedback of the photovoltaic panel surface temperature. The purpose of this invention is to overcome the cold start problem caused by the high dependence on historical data and the limited control accuracy and energy efficiency caused by the lack of an accurate physical model in the prior art.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to an adaptive snow melting method and system for photovoltaic snow shelters based on digital twins and multimodal perception. Background Technology

[0002] In cold regions, some sections of highways experience sudden snowfalls during winter. Snow shelters are typically installed to mitigate the impact of snow accumulation and black ice on driving and passenger safety. Photovoltaic panels are often installed on the outer surface of these shelters to provide some of the energy for heating the components. However, the power generation efficiency of these photovoltaic panels decreases significantly in winter due to snow cover. Therefore, developing efficient and intelligent automatic snow melting technology is crucial for ensuring the stable operation of these photovoltaic snow shelters.

[0003] In related technical fields, in order to achieve intelligent control, some solutions attempt to predict future snow accumulation conditions using historical data and adjust heating strategies accordingly. For example, Chinese patent application CN120434841A discloses a self-heating control method for photovoltaic snowproof sheds used in highway tunnels. Its core lies in dividing the photovoltaic panels into zones, collecting real-time snow thickness data, and obtaining an estimated snow thickness value and an estimated expected value based on historical data (data from multiple preset dates) and the target snow removal volume. Finally, the heating power is adjusted by comparing different expected values.

[0004] However, existing snowmelt control systems generally face a strong reliance on historical data. For a newly deployed system, without sufficient historical data accumulation (e.g., a complete winter), the accuracy of such predictive models is difficult to guarantee; this is the so-called "cold start" problem. Simultaneously, interannual climate change can also lead to a decrease in the representativeness of historical data, affecting the precision of control. Furthermore, existing control logic is mostly based on single macroscopic indicators such as snow depth, while rarely considering the physical properties of snow that decisively influence snowmelt energy consumption, such as density and phase (e.g., the energy required to melt 10 cm of loose, dry snow differs greatly from that required to melt 10 cm of wet, heavy snow). This control method, based on indirect indicators and fuzzy comparisons, has inherent limitations in achieving precise energy matching and maximizing energy efficiency, potentially leading to insufficient heating or energy waste.

[0005] Therefore, how to develop an adaptive snow melting method that does not rely on a large amount of historical data, can accurately identify the physical properties of snow accumulation, and achieve precise energy control is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The purpose of this invention is to overcome the cold start problem caused by the high dependence on historical data in the existing technology and the defects of limited control accuracy and energy efficiency due to the lack of accurate physical models, and to provide a brand-new adaptive snow melting method for photovoltaic snow shelters based on hybrid intelligence of digital twins and neural networks.

[0007] To address the problems existing in the prior art, the technical solution provided by the present invention includes:

[0008] The adaptive snow melting method for photovoltaic snow shelters includes:

[0009] Collect multimodal real-time data on the surface of the photovoltaic panel, wherein the multimodal real-time data includes at least snow thickness, photovoltaic panel surface temperature, and environmental parameters;

[0010] A digital twin physical model based on the laws of thermodynamics is invoked, and the first heating control parameter corresponding to the theoretical energy requirement for melting the current snow is calculated based on the multimodal real-time data.

[0011] A pre-trained neural network model is invoked to predict the second heating control parameter based on the multimodal real-time data;

[0012] The first heating control parameter and the second heating control parameter are dynamically weighted and fused using a weighting factor to generate the final heating control parameter.

[0013] Based on the final heating control parameters, the heating unit is driven to perform snow melting operation on the photovoltaic panel, and the weighting factor and / or the neural network model parameters are updated based on the real-time feedback of the surface temperature of the photovoltaic panel.

[0014] Preferably, the dynamic weighted fusion satisfies the following formula:

[0015] ;

[0016] in, As a weighting factor, For the final heating control parameters, and These are the first heating control parameter and the second heating control parameter, respectively.

[0017] Preferably, the dynamic adjustment of the weighting factor includes:

[0018] Set α ≥ 0.95 for the cold start period;

[0019] During normal operation, α is set to any value in the range of 0.3-0.7 based on the neural network prediction confidence level;

[0020] When the prediction deviation of the neural network model is detected to exceed the first preset threshold, an abnormal calibration period is entered, and α is set to ≥ 0.8 until the prediction deviation is less than the second preset threshold.

[0021] Preferably, the pre-trained neural network model is a physical knowledge embedded neural network, and its loss function includes physical rule residual terms composed of thermodynamic equations.

[0022] Preferably, the step of calculating the theoretical energy requirement for melting the current snow cover further includes:

[0023] Based on the environmental parameters, the physical state of the current snow cover is classified, and the snow density is estimated based on the classification results.

[0024] The snow mass is calculated based on the estimated snow density and snow thickness, and the theoretical energy requirement for melting the current snow is calculated.

[0025] Preferably, the theoretical energy requirement for melting the current snow cover includes:

[0026] The heat required to heat snow from its current temperature to its melting point and cause it to undergo a complete phase change;

[0027] Environmental heat loss from photovoltaic panels during snow melting.

[0028] Preferably, the step of updating the neural network model parameters further includes:

[0029] Set a target snowmelt temperature curve;

[0030] The temperature error is obtained by continuously comparing the real-time feedback value of the photovoltaic panel surface temperature with the corresponding value on the target snow melting temperature curve;

[0031] Based on the temperature error, the parameters of the neural network model are updated using an online learning algorithm.

[0032] This invention also provides an adaptive snow melting system for photovoltaic snow shelters, comprising:

[0033] A multimodal sensing unit is configured to collect multimodal real-time data from the surface of a photovoltaic panel, wherein the multimodal real-time data includes at least snow thickness, photovoltaic panel surface temperature, and environmental parameters.

[0034] A data processing and decision-making unit is communicatively connected to the multimodal sensing unit, and the data processing and decision-making unit includes:

[0035] The physical model module is configured to call a digital twin physical model based on the laws of thermodynamics, and calculate the first heating control parameters corresponding to the theoretical energy requirement for melting the current snow based on the multimodal real-time data.

[0036] The neural network module is configured to call a pre-trained neural network model to predict the second heating control parameters based on the multimodal real-time data.

[0037] The fusion control module is configured to dynamically weight and fuse the first heating control parameter and the second heating control parameter using a weighting factor to generate the final heating control parameter.

[0038] The heating execution unit is communicatively connected to the data processing and decision-making unit and is configured to receive the final heating control parameters and perform heating and snow melting operations on the surface of the photovoltaic panel.

[0039] The feedback update unit is communicatively connected to the multimodal sensing unit and the data processing and decision-making unit, and is configured to update the weighting factors and / or the parameters of the neural network model based on the real-time feedback of the surface temperature of the photovoltaic panel.

[0040] Beneficial effects

[0041] 1. Combining high reliability and high precision: The thermodynamic model ensures that the system can make reliable decisions that conform to physical logic at all times, completely solving the "cold start" problem and providing a stable "safety net" for system decision-making. Meanwhile, the continuous learning and fine correction of the neural network maximizes the system's performance and energy efficiency, achieving control precision far exceeding that of a single model.

[0042] 2. Superior Adaptability: Whether facing the "cold start" challenge of a new deployment or dealing with complex field conditions caused by factors such as wind, obstruction, and equipment aging, the hybrid intelligent system of this invention can quickly adapt and continuously optimize to its best state. The dynamic weighted fusion mechanism enables the system to possess both high stability and strong adaptability to ever-changing field conditions.

[0043] 3. Significant Energy Saving Effect: This invention, through precise calculation and adaptive correction, can minimize energy waste. By real-time identification of the physical properties of snow (such as density and phase) and precise calculation of energy demand based on first principles of physics, optimal energy efficiency is achieved while ensuring snow melting effect.

[0044] 4. Strong robustness and stability: The mechanism of mutual calibration between the two models provides an additional layer of fault tolerance for the system, making it more stable when facing sensor noise or extreme unseen weather patterns. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the adaptive snow melting method for photovoltaic snow shelters provided in a preferred embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram of the adaptive snow melting system for photovoltaic snow shelters provided in a preferred embodiment of the present invention. Detailed Implementation

[0047] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0048] Example 1

[0049] This embodiment provides an adaptive snow melting method for photovoltaic snow shelters based on digital twins and multimodal sensing. In a specific application scenario, the method is executed by a central controller deployed at the photovoltaic snow shelter site. The central controller is connected to a multimodal sensing unit (including but not limited to ultrasonic thickness sensors, patch temperature sensors, temperature and humidity sensors, anemometers, etc.) and one or more heating units (e.g., electrothermal films or heating wires built into the back of the photovoltaic panels).

[0050] like Figure 1 As shown, the adaptive snow melting method for photovoltaic snow shelters provided in this embodiment may specifically include the following steps:

[0051] S1. Collect multimodal real-time data of the photovoltaic panel surface, wherein the multimodal real-time data includes at least snow thickness, photovoltaic panel surface temperature and environmental parameters.

[0052] In this step, data is collected periodically (e.g., once per minute) by a multimodal sensing unit. Snow thickness can be obtained by ultrasonic or laser ranging sensors deployed at different locations on the photovoltaic panel; the surface temperature of the photovoltaic panel is measured by an array of thermocouples or thermistors closely attached to the back of the photovoltaic panel, providing crucial feedback for subsequent closed-loop control; environmental parameters are collected by a small weather station, including at least ambient temperature and humidity, and preferably also wind speed and solar radiation intensity. These data form the basis for subsequent dual-model intelligent decision-making.

[0053] S2. Call a digital twin physical model based on the laws of thermodynamics, and calculate the first heating control parameter corresponding to the theoretical energy requirement for melting the current snow based on the multimodal real-time data.

[0054] The digital twin physical model is pre-built offline based on the inherent physical properties of the photovoltaic panel, such as material properties (e.g., thermal conductivity of the glass cover and specific heat capacity of the backsheet), geometric dimensions, and parameters such as the rated power and spatial distribution of the built-in heating units. The core function of this model is to provide a stable and reliable theoretical reference value under any operating condition.

[0055] Specifically, this physical model can accurately simulate, based on fundamental thermodynamic laws (such as the law of conservation of energy and Fourier's law of heat conduction), the change of the temperature field on the surface of a photovoltaic panel over time when a specific power is input to the heating unit under given environmental conditions. More importantly, it can calculate the theoretical total energy (Q_total) required to melt a specific amount of snow. In one implementation, this total energy can be expressed as: Qtotal = Qmelt + Qloss.

[0056] Where Q_melt is the melting heat, which is the energy required to heat the snow from its current temperature to its melting point and complete the phase change; Q_loss is the environmental heat loss, which is the energy lost by the photovoltaic panel to the surrounding environment through convection, radiation and other means during the snow melting process.

[0057] The formula for calculating Q_melt is: Calculate the required heat. Where m is the total mass of snow, T_melt is the melting point of snow and ice (default 0°C), and c is the specific heat capacity of snow. This is the current snow temperature (which can be approximated as the surface temperature of the snowplow or the ambient temperature). It is the latent heat of fusion of ice (approximately 3.34 × 10^5 J / kg).

[0058] The calculation of Q_loss utilizes the simulation capabilities of a digital twin model. Real-time collected environmental parameters such as temperature and wind speed are input into the model, which then uses a built-in heat transfer model (e.g., combining Newton's law of cooling and the Stefan-Boltzmann law) to simulate and calculate heat loss during snowmelt, including heat dissipated from photovoltaic panels to the surrounding environment through convection and radiation. The specific calculation method can be reasonably designed by those skilled in the art based on existing technology and common knowledge.

[0059] Finally, based on the calculated total energy demand Q_total and the system's preset snow melting time target t (e.g., expecting to complete snow melting within 30 minutes), the first heating control parameter P1 can be determined, which can be a power value (unit: watts), calculated as P1 = Q_total / t.

[0060] In some preferred embodiments, to more accurately calculate the melting heat Q_melt, this method also includes the identification of the physical state of the snow and density estimation, the specific steps of which are as follows:

[0061] S21. Based on the environmental parameters, classify the physical state of the current snow cover and estimate the snow density based on the classification results. The central controller runs a built-in snow physical characteristic identification model. This model, based on the ambient temperature and humidity data collected in S1, classifies the current snow cover into states such as dry snow, wet snow, or ice / freezing rain. For example, a rule-based classifier can be set to classify the snow cover as "dry snow" when the ambient temperature is below -5°C and assign it a low density value, such as 100 kg / m³, from the built-in physical parameter library. 3 When the ambient temperature is between -2°C and 1°C and the humidity is greater than 90%, it is classified as "wet snow" and assigned a higher density value, such as 300 kg / m³. 3 When the ambient temperature remains below 0°C and the snow is hard, it is classified as "ice" and given a higher density value, such as 900 kg / m³. 3 .

[0062] S22. Calculate the snow mass based on the estimated snow density and snow thickness, and calculate the theoretical energy requirement for melting the current snow.

[0063] Based on the snow thickness measured by the sensor in S1, the effective area of ​​the photovoltaic panel, and the snow density estimated in S21, the total mass m of snow covering the photovoltaic panel is calculated. Then, this snow mass m is substituted into the above calculation formula for Qmelt to obtain a more accurate melting heat requirement.

[0064] S3. Call a pre-trained neural network model to predict the second heating control parameters based on the multimodal real-time data.

[0065] The role of this neural network model is to supplement the physical model, and to learn and compensate for complex nonlinear influencing factors (such as local wind field disturbances, snow accumulation unevenness, equipment aging, etc.) that are difficult to be accurately described by ideal physical models through data-driven methods.

[0066] Considering the difficulty in obtaining a large number of labeled training samples in actual deployments, this embodiment preferably employs a pre-training method optimized for few-shot learning. In a preferred embodiment, the neural network model is a Physics-Informed Neural Network (PINN). Its core idea is to add a physical rule residual term composed of thermodynamic equations (such as the heat conduction equation) to the loss function of the neural network, in addition to the conventional prediction error term (i.e., the difference between the model's predicted value and the true value). This is equivalent to forcing the output of the neural network to obey fundamental physical laws during training, thereby greatly reducing the dependence on training data, solving the "cold start" problem, and significantly improving the model's generalization ability. The loss function L can be expressed as: L = L_data + λ * L_physics, where L_data is the data prediction error, L_physics is the physical rule residual term, and λ is the error weight coefficient.

[0067] The model receives the same multimodal real-time data as the physical model as input and outputs a second heating control parameter P2. P2 is a modified or empirical heating power value predicted by the model based on the complex and nonlinear relationships (such as the effects of local wind field disturbances and snow accumulation unevenness) that it learns from the data and cannot be fully described by the ideal physical model.

[0068] S4. Dynamically weight and fuse the first heating control parameter and the second heating control parameter using a weighting factor to generate the final heating control parameter.

[0069] This step is designed to organically combine the high reliability of the physical model with the high adaptability of the neural network model. By dynamically adjusting the weights, the system can prioritize the more reliable model output under different operating stages and conditions, thereby achieving a balance between stability and accuracy.

[0070] In some preferred embodiments, the dynamic weighted fusion can be a linear weighted fusion, with the formula: P_final = α * P1 + (1-α) * P2. Wherein, P_final is the final heating control parameter output to the heating unit, and α is a dynamically changing weighting factor, ranging from [0, 1].

[0071] The dynamic adjustment strategy for the weighting factor α is designed to cover the entire lifecycle of the system, as follows:

[0072] Cold Start Period: During the initial deployment of the system or after a long period of shutdown and restart, the neural network model's predictions may be unreliable because it has not yet undergone sufficient online learning. At this time, the system automatically enters a cold start period, setting the weight factor α to a high value close to 1, such as 0.95, 0.98, or 1.0. This ensures that the final decision P_final is almost entirely determined by the reliable physical model P1, guaranteeing the stability and security of the system in the initial startup phase.

[0073] Normal Operation Phase: As the system continues to run and learn online, the neural network model is fine-tuned by constantly absorbing real-time data, and its prediction accuracy and confidence gradually improve. At this point, the system enters the normal operation phase. The value of the weight factor α is in the range of 0.3 to 0.7, dynamically adjusted according to the confidence level of the neural network prediction. For example, when the model outputs a high confidence score, it means that the model is very confident in its prediction, and α can be set to a lower value, such as 0.35, making more use of the neural network's experience; when the confidence level is low, α is set to a higher value, such as 0.65, relying more on the theoretical calculations of the physical model.

[0074] Abnormal Calibration Period: During operation, the deviation between P1 and P2 is continuously monitored. If this deviation exceeds a preset first threshold (e.g., the absolute value of the deviation is greater than 20% of P1), it indicates a significant discrepancy between the outputs of the two models. This may mean that the neural network has encountered an unprecedented extreme condition, and its predictions may no longer be reliable. At this time, the system automatically enters the abnormal calibration period, temporarily setting α to a high value not less than 0.8 (e.g., 0.8 or 0.9), forcing the system to revert to a more stable physical model-dominated mode, achieving a "safe rollback". This state will remain until the prediction deviation between the two models recovers to less than a more stringent second preset threshold (e.g., the deviation is less than 10% of P1), at which point the system returns to normal operation.

[0075] S5. Based on the final heating control parameters, drive the heating unit to perform snow melting operation on the photovoltaic panel, and update the weighting factor and / or the neural network model parameters based on the real-time feedback of the surface temperature of the photovoltaic panel.

[0076] This step establishes a complete closed-loop feedback control system. The central controller sends the calculated final heating control parameter P_final (e.g., a specific power value) to the drive circuit of the heating unit to drive it to operate.

[0077] Simultaneously, the system utilizes real-time feedback of the plate surface temperature for online learning and optimization. In a preferred embodiment, the update step of the neural network model parameters specifically includes:

[0078] S51. Set a target snow melting temperature curve. This curve defines the ideal heating and snow melting process. For example, the system expects to linearly raise the surface temperature from -5°C to +2°C within 15 minutes and maintain it for 5 minutes thereafter. This curve can be preset or dynamically adjusted according to actual needs and environmental conditions.

[0079] S52. Continuously compare the real-time feedback value of the photovoltaic panel surface temperature with the corresponding value on the target snow melting temperature curve to obtain the temperature error.

[0080] S53. Based on the temperature error, update the parameters (such as weights and biases) of the neural network model using an online learning algorithm. For example, a gradient descent-based fine-tuning algorithm or a Kalman filter algorithm can be used. This process enables the neural network to continuously learn from the actual control effect, adapt to the unique characteristics of the current environment, and thus continuously reduce the error of its subsequent predictions. Optionally, the temperature error can also be used to optimize the dynamic adjustment strategy of the weight factor α (such as the confidence-α mapping relationship during normal operation).

[0081] In another embodiment, to achieve more refined control, the photovoltaic panel surface is physically or logically divided into multiple independent control zones, for example, into a top zone, a middle zone, and a bottom zone based on snow distribution patterns. Each zone is equipped with an independent sensor group (for measuring the snow thickness and surface temperature of that zone) and an independently controllable heating unit. In this case, the adaptive snow melting method described in Embodiment 1 (i.e., steps S1 to S5) will be executed independently and in parallel for each zone. The central controller maintains an independent model calculation and decision-making process for each zone, outputting the optimal final heating control parameters for that zone. In this way, the system can achieve extremely fine-grained zone control. For example, when only the top zone has a thick snow layer, the system will only drive the heating unit in the top zone to operate at higher power, while the middle and bottom zones may operate at lower power or not at all, thereby maximizing energy efficiency while ensuring snow melting effect.

[0082] Example 2

[0083] like Figure 2 As shown, the photovoltaic snow-proof shed adaptive snow melting system provided in this embodiment specifically includes:

[0084] A multimodal sensing unit is configured to collect multimodal real-time data from the surface of a photovoltaic panel, wherein the multimodal real-time data includes at least snow thickness, photovoltaic panel surface temperature, and environmental parameters.

[0085] A data processing and decision-making unit is communicatively connected to the multimodal sensing unit, and the data processing and decision-making unit includes:

[0086] The physical model module is configured to call a digital twin physical model based on the laws of thermodynamics, and calculate the first heating control parameters corresponding to the theoretical energy requirement for melting the current snow based on the multimodal real-time data.

[0087] The neural network module is configured to call a pre-trained neural network model to predict the second heating control parameters based on the multimodal real-time data.

[0088] The fusion control module is configured to dynamically weight and fuse the first heating control parameter and the second heating control parameter using a weighting factor to generate the final heating control parameter.

[0089] The heating execution unit is communicatively connected to the data processing and decision-making unit and is configured to receive the final heating control parameters and perform heating and snow melting operations on the surface of the photovoltaic panel.

[0090] The feedback update unit is communicatively connected to the multimodal sensing unit and the data processing and decision-making unit, and is configured to update the weighting factors and / or the parameters of the neural network model based on the real-time feedback of the surface temperature of the photovoltaic panel.

[0091] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive snow melting method for photovoltaic snow shelters, characterized in that, include: Collect multimodal real-time data on the surface of the photovoltaic panel, wherein the multimodal real-time data includes at least snow thickness, photovoltaic panel surface temperature, and environmental parameters; A digital twin physical model based on the laws of thermodynamics is invoked, and the first heating control parameter corresponding to the theoretical energy requirement for melting the current snow is calculated based on the multimodal real-time data. A pre-trained neural network model is invoked to predict the second heating control parameter based on the multimodal real-time data; The first heating control parameter and the second heating control parameter are dynamically weighted and fused using a weighting factor to generate the final heating control parameter. Based on the final heating control parameters, the heating unit is driven to perform snow melting on the photovoltaic panel, and the weighting factors and / or the neural network model parameters are updated based on real-time feedback of the photovoltaic panel surface temperature; The dynamic adjustment of the weighting factors includes: Set α ≥ 0.95 for the cold start period; During normal operation, α is set to any value in the range of 0.3-0.7 based on the neural network prediction confidence level; When the prediction deviation of the neural network model is detected to exceed a first preset threshold, an abnormal calibration period is entered, and α is set to ≥ 0.8 until the prediction deviation is less than a second preset threshold; The pre-trained neural network model is a physical knowledge embedded neural network, and its loss function includes physical rule residuals composed of thermodynamic equations.

2. The adaptive snow melting method for photovoltaic snow shelters according to claim 1, characterized in that, The dynamic weighted fusion satisfies the following formula: ; in, As a weighting factor, For the final heating control parameters, and These are the first heating control parameter and the second heating control parameter, respectively.

3. The adaptive snow melting method for photovoltaic snow-proof sheds according to claim 1, characterized in that, The steps for calculating the theoretical energy requirement to melt the current snow cover further include: Based on the environmental parameters, the physical state of the current snow cover is classified, and the snow density is estimated based on the classification results. The snow mass is calculated based on the estimated snow density and snow thickness, and the theoretical energy requirement for melting the current snow is calculated.

4. The adaptive snow melting method for photovoltaic snow-proof sheds according to claim 3, characterized in that, The theoretical energy requirement for melting the current snow cover includes: The heat required to heat snow from its current temperature to its melting point and cause it to undergo a complete phase change; Environmental heat loss from photovoltaic panels during snow melting.

5. The adaptive snow melting method for photovoltaic snow shelters according to claim 1, characterized in that, The update step of the neural network model parameters further includes: Set a target snowmelt temperature curve; The temperature error is obtained by continuously comparing the real-time feedback value of the photovoltaic panel surface temperature with the corresponding value on the target snow melting temperature curve; Based on the temperature error, the parameters of the neural network model are updated using an online learning algorithm.

6. A photovoltaic snow-proof shed / cave adaptive snow melting system, characterized in that, include: A multimodal sensing unit is configured to collect multimodal real-time data from the surface of a photovoltaic panel, wherein the multimodal real-time data includes at least snow thickness, photovoltaic panel surface temperature, and environmental parameters. A data processing and decision-making unit is communicatively connected to the multimodal sensing unit, and the data processing and decision-making unit includes: The physical model module is configured to call a digital twin physical model based on the laws of thermodynamics, and calculate the first heating control parameters corresponding to the theoretical energy requirement for melting the current snow based on the multimodal real-time data. The neural network module is configured to call a pre-trained neural network model to predict the second heating control parameters based on the multimodal real-time data. The fusion control module is configured to dynamically weight and fuse the first heating control parameter and the second heating control parameter using a weighting factor to generate the final heating control parameter. The heating execution unit is communicatively connected to the data processing and decision-making unit and is configured to receive the final heating control parameters and perform heating and snow melting operations on the surface of the photovoltaic panel. The feedback update unit is communicatively connected to the multimodal sensing unit and the data processing and decision-making unit, and is configured to update the weighting factors and / or the parameters of the neural network model based on the real-time feedback of the surface temperature of the photovoltaic panel.

Images