Building material stacking area protection prediction management method, system and equipment and storable medium

Abstract

The invention discloses a building material stacking area protection prediction management method, system and device and a storable medium, and belongs to the technical field of building construction big data processing, and the method comprises the following steps: weather forecast data acquisition, actual measurement data acquisition, multi-model collaborative prediction, climate abnormality analysis and correction, instruction generation linkage, and feedback evaluation fine tuning. According to the invention, based on the meteorological data of the hourly time and the real-time sensor data, the humidity change trend of the stacking area can be accurately predicted, the prediction result is corrected in combination with historical climate anomaly features, the adaptability and robustness of the model under seasonal change are enhanced, and the prediction accuracy is improved. And corresponding control instructions can be given according to relative humidity, temperature and illumination intensity states, moisture and water loss caused by exposure are prevented, the protection level of building materials on a construction site is improved, and the loss rate is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of construction big data processing, and in particular relates to a method, system, equipment and storable medium for predicting and managing protection of building material stacking areas. Background Art

[0002] As a lightweight, porous building material, autoclaved aerated wall insulation offers advantages such as thermal insulation, heat insulation, noise isolation, lightweight, earthquake resistance, fire resistance, easy construction, and environmental friendliness. However, its disadvantages are also significant. Due to its porous structure, it is highly absorbent and requires moisture-proofing during storage and transportation. At the same time, it is also important to note that high temperatures and exposure to sunlight can easily cause the already limited moisture in the building materials to escape, causing damage to the building materials. These disadvantages are common in many building materials. Therefore, how to monitor relative humidity, temperature, and light intensity in real time based on meteorological changes and the on-site environment, combine regional macroclimatic parameters with microclimatic parameters measured at the construction site, and integrate them into the BIM system to provide early warnings and coordinate construction processes, thereby improving the protection and prediction capabilities of building materials has become a key technical challenge in ensuring construction quality. Existing technologies often use single-sensor monitoring or static grading strategies, lacking dynamic prediction capabilities and failing to fully consider seasonal climate impacts, which can easily lead to misjudgments or control failures. From the perspective of transportation costs and transportation loss risks, the sales of construction materials have certain geographical restrictions, and the construction areas of regional construction units are generally also regional. This provides a feasibility basis for using historical meteorological data of specific construction areas to conduct building material protection forecasting and management. Summary of the Invention

[0003] In order to overcome the technical problems described in the above background technology, the present invention provides a protection prediction management method, system, equipment and storable medium for building material stacking areas. Based on meteorological data and real-time sensor data at the hour, it can more accurately predict the changing trends of relative humidity, temperature and light intensity in the stacking area, and correct the prediction results in combination with historical climate anomaly characteristics to enhance the adaptability and robustness of the model under seasonal changes. It can also give corresponding control instructions according to the relative humidity, temperature and light intensity status to prevent moisture and water loss due to exposure to the sun, thereby improving the protection level of building materials on the construction site and reducing the loss rate.

[0004] The technical solution of the present invention is: a method for predicting and managing protection of a building material stacking area, comprising the following steps: S1. Obtain weather forecast data and the past two years' historical weather data for the construction site's building material stacking area from the API interface of the meteorological department's open data platform. The weather data should include at least the temperature, relative humidity, and light intensity at each hour. S2. Real-time sensor data is acquired through a sensor network deployed in the building material stacking area. The real-time sensor data includes the temperature, humidity, and light intensity of the building material stacking area. The real-time sensor characterization data is obtained by averaging each parameter of the sensor network. S3: Normalize the weather forecast data for the next hour and the real-time sensor data for the current hour. If there are missing values, use linear interpolation to supplement them, and then splice them to form a multi-dimensional sequence input feature. Based on the multi-model collaborative mechanism, the relative humidity, temperature and light intensity of the next hour are predicted and processed to obtain the relative humidity prediction value. , temperature prediction value and light intensity prediction ; S4. Based on historical meteorological data, climate anomaly analysis was conducted, and the quantile intervals of humidity, temperature and light intensity were calculated by month to determine the historical normal intervals of humidity, temperature and light intensity. , ]、[ , ][ , ], and correct the relative humidity forecast value, temperature forecast value and light intensity forecast value according to the current month to obtain the corrected relative humidity forecast value , temperature prediction value and light intensity prediction ; S5. Corrected relative humidity prediction results based on the corrected , temperature prediction value and light intensity prediction ,adopting a multi-level response control strategy, the humidity range is divided into three risk levels, and the control instructions corresponding to each level are triggered; S6. The corrected relative humidity prediction value , temperature prediction value , light intensity prediction value , risk levels and control instructions are sent to the BIM system through the communication interface for construction management personnel to use as reference suggestions for moisture-proof management of building materials.

[0005] Furthermore, the sensor network includes a peripheral sensor group and an internal sensor group. The peripheral sensor group is arranged outside the building material stacking area to measure the ambient temperature, ambient humidity and ambient light intensity outside the stacking area. The internal sensor group is arranged inside the building material stacking area to measure the temperature, humidity and light intensity inside the stacking area.

[0006] Furthermore, in step S3, the multidimensional sequence input feature Including weather forecast temperature , Current measured temperature , Weather Forecast Relative Humidity , Current measured humidity , weather forecast light intensity and the current measured light intensity .

[0007] Furthermore, in step S3, the multi-model collaboration mechanism includes the following steps: S301. At the beginning of the earthwork excavation phase of the construction project, collect historical meteorological data of the hourly weather forecast and the corresponding measured meteorological data of the building material stacking area at the previous moment to form a model training data set; S302: Based on the collected model training data set, the Autoformer model, LSTM model and XGBoost model are trained independently, and the input is a multi-dimensional sequence input feature. , the output is the relative humidity forecast value for the next hour , temperature prediction value and light intensity prediction , complete the pre-training of the Autoformer model, LSTM model and XGBoost model; S303: Deploy the pre-trained Autoformer model, LSTM model, and XGBoost model independently, and input multi-dimensional sequence input features. , and obtain three independent sets of prediction values ​​( 、 、 ), ( 、 、 )and( 、 、 ), corresponding to relative humidity, temperature and light intensity respectively; S304: Forward pass the three independent pre-trained Autoformer models, LSTM models, and XGBoost models once respectively, and start Monte Carlo Dropout to make multiple predictions during inference, so as to obtain K different prediction results for each prediction item, and obtain three prediction sets. 、 and , , based on the prediction set 、 and , respectively calculate the corresponding model prediction mean 、 and and the corresponding variance 、 and , , , , , , ; S305. The following uses relative humidity for processing. The temperature and light intensity are processed in a similar way to the relative humidity. The processing processes are independent of each other. The multidimensional sequence input features of the current time step are used. , through the fully connected layer Generate attention query vector , ; S306. Calculate the attention key value based on the model prediction mean and attention value , , ; S307, based on the attention mechanism and confidence adjustment, combined with the learnable variance penalty coefficient , and obtain the dynamic weights of the three models respectively , , in Represents the relative humidity fusion weight of the i-th model, i=1, 2, 3, This is a learnable parameter that is dynamically adjusted during the online adaptation process. It represents the variance penalty coefficient and controls the degree of influence of uncertainty on the weight. Note that the e in the current formula represents the natural constant of the exponential function, which is different from the e in step S701. The e in step S701 refers to the error. S308: Output the final relative humidity fusion prediction based on the prediction mean and dynamic weight of each model , , in, represents the fusion weight of the i-th model, Represents the predicted mean of each model, i takes values ​​of 1, 2, and 3 respectively; S309, at each time step, based on the final relative humidity fusion prediction results Compared with the actual observed humidity The error between the two is used to calculate the loss function , , Using this loss function to backpropagate, only update , without updating the model parameters, , in, for The learning rate, and in order to prevent Too large or too small will cause failure, set The value range is and between, and Based on experience Upper and lower limits of values; S310, the temperature and light intensity are also processed with reference to the relative humidity, and steps S303 to S309 are performed independently. In these steps, the relative humidity is replaced by temperature or light intensity, and the temperature fusion prediction is finally output. Fusion prediction with light intensity .

[0008] Furthermore, in step S4, climate anomaly analysis includes the following steps: S401, historical meteorological data acquisition and database construction phase, that is, to obtain the hourly meteorological observation historical data of the area where the building material stacking area is located in the past two years, including at least the temperature, relative humidity and light intensity at each hour, to form a historical meteorological database ; S402, the monthly grouping and statistical feature extraction stage, that is, statistically analyzing the quantile intervals of humidity, temperature, and light intensity by month, determining the historical normal ranges of humidity, temperature, and light intensity, and correcting the relative humidity prediction value, temperature prediction value, and light intensity prediction value based on the current month, and calculating the average humidity data for each month. , standard deviation , minimum value , maximum value , P10 quantile , P50 quantile , P75 quantile , P90 quantile ; S403, the climate anomaly model construction phase, that is, based on the quantile statistics results, determine the historical normal range of humidity, temperature and light intensity for each month [ , ]、[ , ][ , ], building a climate anomaly model for each month 、 and ; S404, climate anomaly correction factor calculation and prediction result correction stage, that is, according to the current month m, the corresponding climate anomaly model is extracted 、 and , the relative humidity prediction value output by the multi-model synergy mechanism , temperature prediction value and light intensity prediction , respectively, perform climate anomaly correction to obtain the corrected relative humidity forecast value , temperature prediction value , light intensity prediction value .

[0009] Furthermore, in step S5, the multi-level response control strategy includes the following steps: S501. According to expert experience, the storage conditions for building materials are 45%-60% relative humidity, the temperature should be below 45°C, and the light intensity should be below 1000 lux. Based on this, the deviation is set to divide the risk. The relative humidity deviation : , If the relative humidity exceeds the range of [45%, 60%], the greater the deviation, the higher the risk level. The temperature deviation ; , If the temperature exceeds 45°C, the greater the deviation, the higher the risk level; the light intensity deviation : , This means that if the light intensity exceeds 1000 lux, the greater the deviation, the higher the risk level; S502. Calculate the comprehensive risk score using the weighted risk score formula : , in, 、 and are the humidity risk weight, temperature risk weight, and light intensity risk weight, which are determined by expert experience and are generally 0.33 respectively, with a total of 1; S503, based on comprehensive risk score There are three risk levels: low risk ( ), medium risk ( ) and high risk ( ), low risk means all conditions are close to the standard value with no significant deviation, medium risk means some deviation, but not extreme, and high risk means significant deviation, and environmental conditions may cause damage to building materials; S504. Determine corresponding operation instructions based on the risk level. When the risk is low, adopt a no-response strategy; when the risk is medium, adopt a close monitoring strategy; when the risk is high, adopt a cover-up strategy. S505. Push the operation instructions to the BIM system through the communication interface, so that the construction management personnel can use them as reference suggestions for moisture-proof management of building materials.

[0010] Furthermore, the method further includes a feedback evaluation step, which includes the following steps: S701, real-time calculation of relative humidity prediction error , temperature prediction error and light intensity prediction error , , , , in 、 and They represent the relative humidity prediction error, temperature prediction error and light intensity prediction error at time t+1 respectively, 、 and Indicates the actual measured relative humidity, temperature and light intensity values. 、 and Indicates the revised predicted relative humidity value, predicted temperature value and predicted light intensity value; S702: Calculate the sliding mean square error for relative humidity, temperature, and light intensity. 、 and , and mean absolute percentage error 、 and , , , , , , , Where n represents the length of the sliding window based on a 24-hour environmental change cycle; S703: Safety threshold based on historical best performance set by experience 、 、 、 、 、 Determine whether the model's ability to predict the current environment is insufficient. If the sliding mean square error or mean absolute percentage error of any of the three items (relative humidity, temperature, or light intensity) is greater than the corresponding safety threshold, it indicates a persistently high error and insufficient predictive ability for the current environment. Fine-tune the multi-model coordination mechanism for this item. S704, calculate the KL divergence of the error distribution for relative humidity, temperature or light intensity respectively 、 and , , , , in 、 and are the probability distributions of the prediction errors of relative humidity, temperature and light intensity in the current window, 、 and are the probability distributions of historical stability period errors of relative humidity, temperature, and light intensity, respectively; S705: Preset drift alarm thresholds based on relative humidity, temperature, and light intensity 、 and ,judge Is it greater than 、 Is it greater than or Is it greater than If there is a case where is greater than , then the corresponding items in relative humidity, temperature, and light intensity in the current construction environment have significantly deviated from the environment during model training, and there is a risk of prediction failure. It is necessary to fine-tune the corresponding multi-model coordination mechanism for relative humidity, temperature, and light intensity. S706. Record the error changes before and after fine-tuning, the number of fine-tuning times, and the triggering reasons, generate a periodic report, and manually review it. If the manual review finds that a fine-tuning is incorrect, the parameters of the multi-model collaboration mechanism can be rolled back.

[0011] A building material stacking area protection prediction and management system, characterized by comprising: The meteorological data collection module is used to obtain weather forecast data and historical meteorological data from the past two years for the area where the building material stacking area is located through the API interface of the meteorological department's open data platform. The meteorological data includes at least the temperature, relative humidity, and light intensity at each hour. The historical meteorological data can be used to build a historical database by recording hourly meteorological data from the past two years, providing data support for subsequent climate anomaly analysis. A measured data acquisition module is used to obtain real-time ambient temperature, humidity, and light intensity data inside and outside the building material stacking area through a peripheral sensor group installed outside the building material stacking area and an internal sensor group installed inside the building material stacking area. The peripheral sensor group is used to measure the ambient temperature, humidity, and light intensity, and the internal sensor group is used to measure the internal temperature, humidity, and light intensity. The real-time collected data is uploaded to a remote server via wireless means, and the mean value of each parameter is calculated to obtain real-time sensor characterization data; Multi-source data preprocessing module: This module is used to normalize the weather forecast data for the next hour and the real-time sensor data for the current hour. If there are missing values, linear interpolation is used to supplement them to generate multidimensional sequence input features. The input features include the forecast temperature, the current measured temperature, the forecast relative humidity, the current measured humidity, the forecast light intensity, and the current measured light intensity, which are used for subsequent model prediction processing. The collaborative prediction processing module is used to independently build models using three pre-trained models: Autoformer, LSTM, and XGBoost, to predict relative humidity, temperature, and light intensity. The Autoformer is trained based on multidimensional sequence features and a trend-season decomposition mechanism, the LSTM model is trained based on a gated recursive unit structure, and the XGBoost model is trained based on the expansion of time series features into tabular features. The prediction results of the three models are then fused through a dynamic attention mechanism and a variance penalty coefficient to obtain the final predicted values ​​of relative humidity, temperature, and light intensity. The climate analysis and correction module is used to perform climate anomaly analysis based on historical meteorological data, calculate the quantile intervals of relative humidity, temperature, and light intensity by month, determine the historical normal intervals for each month, and correct the forecast results of relative humidity, temperature, and light intensity based on the current month to obtain the corrected relative humidity, temperature, and light intensity forecast values; The risk level determination module is used to adopt a multi-level response control strategy based on the corrected relative humidity, temperature, and light intensity prediction results, divide the humidity range into three risk levels, and trigger the control instructions corresponding to each level to be sent to the BIM system; The feedback evaluation and fine-tuning module is used to analyze the error between each prediction result and the actual measurement result, and judge the health status of the model based on error statistics and drift detection (KL divergence). If the error or drift exceeds the threshold, the fine-tuning mechanism is triggered, and only the weight calculation logic or variance penalty coefficient of the fusion layer is updated to keep the model adaptable to field changes. At the same time, the feedback evaluation module records the fine-tuning process and generates periodic operation evaluation reports for manual review.

[0012] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned method for predicting and managing protection of a building material stacking area when executing the computer program.

[0013] 10. A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the above-mentioned method for predicting and managing protection of a building material stacking area.

[0014] The present invention has the following beneficial effects due to the adoption of the above technology.

[0015] 1. The present invention improves the effectiveness of moisture-proof management in building material stacking areas by introducing multi-source data fusion, climate anomaly correction, collaborative prediction, and sending linkage information to BIM. It achieves short-term (next hour) prediction of relative humidity, temperature, and light intensity with high prediction accuracy, effectively improving the timeliness, accuracy, and adaptability of moisture-proof and sun-proof responses, reducing the loss rate of building materials, and improving construction quality.

[0016] 2. This invention uses multi-model collaboration (Autoformer+LSTM+XGBoost) to enhance the robustness of prediction, integrating the advantages of neural networks and tree models, and taking into account both long-term trend capture and short-term response.

[0017] 3. Based on the humidity meteorological data of the past two years, the present invention statistically analyzes the normal ranges (P10-P90) of relative humidity, temperature and light intensity by month, establishes climate anomaly models respectively, and dynamically adjusts the prediction results of relative humidity, temperature and light intensity according to seasonal characteristics to resist regional seasonal climate deviations.

[0018] 4. The present invention adopts a weighted fusion risk assessment method based on the corrected relative humidity, temperature and light intensity predicted values, and can predict the moisture and sun protection risks at the next hour according to the risk assessment score, which can effectively improve the storage and protection capabilities of building materials in the BIM system. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1It is a flow chart of the building material protection prediction and management method in the present invention.

[0020] Figure 2 It is a flowchart of the multi-model collaboration mechanism in the present invention.

[0021] Figure 3 It is a flow chart of climate anomaly analysis in the present invention.

[0022] Figure 4 It is a flow chart of the multi-stage response control strategy in the present invention.

[0023] Figure 5 It is a flowchart of the feedback evaluation fine-tuning in the present invention.

[0024] Figure 6 It is a schematic diagram of the module structure of the building material protection prediction and management system in the present invention.

[0025] In the figure: 100, image data acquisition module, 200, measured data acquisition module, 300, multi-source data processing module, 400, collaborative prediction processing module, 500, climate analysis and correction module, 600, risk level determination module, 700, feedback evaluation and fine-tuning module. DETAILED DESCRIPTION

[0026] Example 1: Figure 1 As shown, the present invention provides a method for predicting and managing protection of a building material stacking area, comprising the following steps: S1. Obtain weather forecast data and the past two years' historical weather data for the construction site's building material stacking area from the API interface of the meteorological department's open data platform. The weather data should include at least the temperature, relative humidity, and light intensity at each hour. S2. Real-time sensor data is acquired through a sensor network deployed in the building material stacking area. The real-time sensor data includes the temperature, humidity, and light intensity of the building material stacking area. The real-time sensor characterization data is obtained by averaging each parameter of the sensor network. S3: Normalize the weather forecast data for the next hour and the real-time sensor data for the current hour. If there are missing values, use linear interpolation to supplement them, and then splice them to form a multi-dimensional sequence input feature. , multi-dimensional sequence input features Including weather forecast temperature , Current measured temperature , Weather Forecast Relative Humidity , Current measured humidity , weather forecast light intensity and the current measured light intensity Based on the multi-model collaborative mechanism, the relative humidity, temperature and light intensity of the next hour are predicted and processed to obtain the relative humidity prediction value. , temperature prediction value and light intensity prediction ; S4. Based on historical meteorological data, climate anomaly analysis was conducted, and the quantile intervals of humidity, temperature and light intensity were calculated by month to determine the historical normal intervals of humidity, temperature and light intensity. , ]、[ , ][ , ], and correct the relative humidity forecast value, temperature forecast value and light intensity forecast value according to the current month to obtain the corrected relative humidity forecast value , temperature prediction value and light intensity prediction ; S5. Corrected relative humidity prediction results based on the corrected , temperature prediction value and light intensity prediction ,adopting a multi-level response control strategy, the humidity range is divided into three risk levels, and the control instructions corresponding to each level are triggered; S6. The corrected relative humidity prediction value , temperature prediction value , light intensity prediction value , risk levels and control instructions are sent to the BIM system through the webhook interface for construction management personnel to use as reference suggestions for moisture-proof management of building materials.

[0027] Among them, the sensor network in step S2 mainly includes a peripheral sensor group and an internal sensor group, wherein the peripheral sensor group is arranged outside the building material stacking area, and is used to measure the ambient temperature, ambient humidity and ambient light intensity outside the stacking area; the internal sensor group is arranged inside the building material stacking area, and is used to measure the temperature, humidity and light intensity inside the stacking area; wherein the peripheral sensor group is used to sense the atmospheric environmental conditions of the stacking area, and the internal sensor group is used to sense the microenvironmental state of the inner wall surface of the stacking area and the stacking core area, and the two work together to constitute a multi-source data input for better humidity prediction; and the number and layout of the peripheral sensor group and the internal sensor group can be flexibly set according to factors such as the stacking area area, material stacking density, sensitivity to environmental changes, etc. The type and model of the sensor are not limited to specific brands or models, and other sensor components with equivalent functions can also be used.

[0028] The peripheral sensor group has the same internal components as the internal sensor group, both of which include a single-chip microcomputer controller with a wireless transmission module, a temperature sensor, a humidity sensor and a light sensor. The single-chip microcomputer controller is electrically connected to the temperature sensor, the humidity sensor and the light sensor in turn. Specifically, it is arranged on a shell, and the shell is fixed on a base to form a sensor assembly that is easy to move. When deploying, care should be taken not to allow the internal sensor group to be covered by the piles of building materials in the stacking area, and it should be placed as much as possible in the center of the stacking area. The peripheral sensor group should be placed at the south edge of the stacking area of ​​building materials, and should not be covered by the piles of building materials in the stacking area.

[0029] like Figure 2 As shown, in step S3, the multi-model collaboration mechanism includes the following steps: S301. At the beginning of the earthwork excavation phase of the construction project, collect historical meteorological data of the hourly weather forecast and the corresponding measured meteorological data of the building material stacking area at the previous moment to form a model training data set; S302: Based on the collected model training data set, the Autoformer model, LSTM model and XGBoost model are trained independently, and the input is a multi-dimensional sequence input feature. , the output is the relative humidity forecast value for the next hour , temperature prediction value and light intensity prediction , complete the pre-training of the Autoformer model, LSTM model and XGBoost model; S303: Deploy the pre-trained Autoformer model, LSTM model, and XGBoost model independently, and input multi-dimensional sequence input features. , and obtain three independent sets of prediction values ​​( 、 、 ), ( 、 、 )and( 、 、 ), corresponding to relative humidity, temperature and light intensity respectively; S304: Forward pass the three independent pre-trained Autoformer models, LSTM models, and XGBoost models once respectively, and start Monte Carlo Dropout to make multiple predictions during inference, so as to obtain K different prediction results for each prediction item, and obtain three prediction sets. 、 and , , based on the prediction set 、 and , respectively calculate the corresponding model prediction mean 、 and and the corresponding variance 、 and , , , , , , ; S305. The following uses relative humidity for processing. The temperature and light intensity are processed in a similar way to the relative humidity. The processing processes are independent of each other. The multidimensional sequence input features of the current time step are used. , through the fully connected layer Generate attention query vector , ; S306. Calculate the attention key value based on the model prediction mean and attention value , , ; S307, based on the attention mechanism and confidence adjustment, combined with the learnable variance penalty coefficient , and obtain the dynamic weights of the three models respectively , , in Represents the relative humidity fusion weight of the i-th model, i=1, 2, 3, It is a learnable parameter that is dynamically adjusted during the online adaptation process. It represents the variance penalty coefficient and controls the impact of uncertainty on the weight. S308: Output the final relative humidity fusion prediction based on the prediction mean and dynamic weight of each model , , in, represents the fusion weight of the i-th model, Represents the predicted mean of each model, i takes values ​​of 1, 2, and 3 respectively; S309, at each time step, based on the final relative humidity fusion prediction results Compared with the actual observed humidity The error between the two is used to calculate the loss function , , Using this loss function to backpropagate, only update , without updating the model parameters, , in, for The learning rate, and in order to prevent Too large or too small will cause failure, set The value range is and between, and Based on experience Upper and lower limits of values; S310, the temperature and light intensity are also processed with reference to the relative humidity, and steps S303 to S309 are performed independently. In these steps, the relative humidity is replaced by temperature or light intensity, and the temperature fusion prediction is finally output. Fusion prediction with light intensity .

[0030] In step S301, the adaptation characteristics of the three pre-training models, Autoformer, LSTM, and XGBoost, are shown in Table 1.

[0031] Table 1 Model comparison and adaptation analysis table

[0032] In step S302, since XGBoost is an ensemble learning model based on gradient boosted trees (GBDT) rather than a neural network and does not support Dropout, Bootstrap resampling (a bagging-like mechanism) is used instead to achieve the "multiple sub-model sampling" effect. XGBoost's Bootstrap resampling alternative operation is specifically as follows: repeatedly sampling subsets of the input data, calling XGBoost prediction on each subset separately, and obtaining a set of prediction samples. If it is a single test sample (not a batch), the input cannot be resampled, and multiple weak models are used (multiple XGBoost models are trained with different random seeds), or a certain degree of uncertainty is introduced by setting different ntree_limit, colsample_bytree, and subsample.

[0033] like Figure 3As shown, in step S4, in order to effectively resist the impact of meteorological seasonal changes on the humidity prediction accuracy, improve the adaptability and stability of humidity prediction in different months and different climatic conditions, and realize a prediction correction mechanism with climate anomaly perception capability, climate anomaly analysis includes the following steps: S401, historical meteorological data acquisition and database construction phase, that is, to obtain the hourly meteorological observation historical data of the area where the building material stacking area is located in the past two years, including at least the temperature, relative humidity and light intensity at each hour, to form a historical meteorological database ; S402, the monthly grouping and statistical feature extraction stage, that is, statistically analyzing the quantile intervals of humidity, temperature, and light intensity by month, determining the historical normal ranges of humidity, temperature, and light intensity, and correcting the relative humidity prediction value, temperature prediction value, and light intensity prediction value based on the current month, and calculating the average humidity data for each month. , standard deviation , minimum value , maximum value , P10 quantile , P50 quantile , P75 quantile , P90 quantile ; S403, the climate anomaly model construction phase, that is, based on the quantile statistics results, determine the historical normal range of humidity, temperature and light intensity for each month [ , ]、[ , ][ , ], building a climate anomaly model for each month 、 and ; S404, climate anomaly correction factor calculation and prediction result correction stage, that is, according to the current month m, the corresponding climate anomaly model is extracted 、 and , the relative humidity prediction value output by the multi-model synergy mechanism , temperature prediction value and light intensity prediction , respectively make climate anomaly corrections: , , , in, is the relative humidity climate anomaly correction factor. If the fusion prediction humidity value In the historical normal humidity range [ , ], then set Equal to 1, if the fusion predicted humidity value Outside the historical normal range and greater than , then set Greater than 1, if the fusion prediction humidity value Outside the historical normal range and less than , then set Less than 1 or equal to 1, relative humidity correction factor The specific value of is determined by the linear amplification and piecewise function adjustment strategy according to the degree of exceeding the normal range, so as to achieve the effect of resisting seasonal deviation; is the temperature climate anomaly correction factor. If the predicted humidity value is integrated In the historical normal humidity range [ , ], then set Equal to 1, if the fusion predicted humidity value Outside the historical normal range and greater than , then set Greater than 1, if the fusion prediction humidity value Outside the historical normal range and less than , then set Less than 1 or equal to 1, temperature correction factor The specific value of is determined by the linear amplification and piecewise function adjustment strategy according to the degree of exceeding the normal range, so as to achieve the effect of resisting seasonal deviation; is the correction factor for the climate anomaly of light intensity. If the predicted humidity value is integrated In the historical normal humidity range [ , ], then set Equal to 1, if the fusion predicted humidity value Outside the historical normal range and greater than , then set Greater than 1, if the fusion prediction humidity value Outside the historical normal range and less than , then set Less than 1 or equal to 1, light intensity correction factor The specific value of is determined by the linear amplification and piecewise function adjustment strategy according to the degree of exceeding the normal range, so as to achieve the effect of resisting seasonal deviation.

[0034] In step S402, the pth quantile It refers to the value at the pth percentile position after a set of sample data is sorted from small to large. P10 (10th percentile) is the number at the 10th percentile position from small to large, P50 (median, 50%) is the number in the middle, P75 (75th percentile position) is the number in the upper quartile, and P90 (90th percentile position) is the number close to the maximum value. Linear interpolation is used to calculate the quantile, which is the position of the pth quantile: , Where p is the quantile percentage, N is the number of historical humidity observations, and k is the theoretical index position; If k is an integer, then , if k is not an integer (assuming k=i+f, where i is the integer part and f is the fractional part), then ,in Represents historical humidity observation data for a certain month After sorting A value in .

[0035] like Figure 4 As shown, in step S5, a multi-level response control strategy is used to dynamically divide risk levels based on the humidity prediction value, temperature prediction value, light intensity prediction value, climate anomaly correction result, and on-site environmental conditions, and generate corresponding control instructions to implement protective management of the building material stacking area. Specifically, the strategy includes the following steps: S501. According to the recommended relative humidity for common building materials (as shown in Table 2), the relative humidity should be between 45% and 60%. At the same time, according to expert experience, the temperature should be below 45°C and the light intensity should be below 1000 lux. The deviation is set to divide the risk. : , If the relative humidity exceeds the range of [45%, 60%], the greater the deviation, the higher the risk level. The temperature deviation ; , If the temperature exceeds 45°C, the greater the deviation, the higher the risk level; the light intensity deviation : , This means that if the light intensity exceeds 1000 lux, the greater the deviation, the higher the risk level;

[0036] Table 2 Recommended relative humidity for storage of common building materials

[0037] S502. Calculate the comprehensive risk score using the weighted risk score formula : , in, 、 and are the humidity risk weight, temperature risk weight, and light intensity risk weight, which are determined by expert experience and are generally 0.33 respectively, with a total of 1; S503, based on comprehensive risk score There are three risk levels: low risk ( ), medium risk ( ) and high risk ( ), low risk means all conditions are close to the standard value, with no significant deviation; medium risk means some deviation, but not extreme; high risk means significant deviation, and environmental conditions may cause damage to building materials, such as based on the revised relative humidity prediction value , temperature prediction value and light intensity prediction are 65%, 47°C and 1100 lux respectively, then 0.33, 0.044, is 0.1, then =0.1574, which is a low-risk situation; S504. Determine corresponding operational control instructions based on the risk level, for example, adopt a no-response strategy for low risk, a close-watch strategy for medium risk, and a cover-up strategy for high risk; S505. Push control instructions to the BIM system through the communication interface to provide construction management personnel with reference suggestions for moisture-proof management of building materials.

[0038] Among them, such as Figure 5 As shown, in order to improve the accuracy of humidity prediction, a feedback evaluation step needs to be performed after step S6 is executed. The feedback evaluation includes the following steps: S701, real-time calculation of relative humidity prediction error , temperature prediction error and light intensity prediction error , , , , in 、 and They represent the relative humidity prediction error, temperature prediction error and light intensity prediction error at time t+1 respectively, 、 and Indicates the actual measured relative humidity, temperature and light intensity values. 、 and Indicates the revised predicted relative humidity value, predicted temperature value and predicted light intensity value; S702: Calculate the sliding mean square error for relative humidity, temperature, and light intensity. 、 and , and mean absolute percentage error 、 and , , , , , , , Where n represents the length of the sliding window based on a 24-hour environmental change cycle; S703: Safety threshold based on historical best performance set by experience 、 、 、 、 、 Determine whether the model's ability to predict the current environment is insufficient. If the sliding mean square error or mean absolute percentage error of any of the three items (relative humidity, temperature, or light intensity) is greater than the corresponding safety threshold, it indicates a persistently high error and insufficient predictive ability for the current environment. Fine-tune the multi-model coordination mechanism for this item. S704, calculate the KL divergence of the error distribution for relative humidity, temperature or light intensity respectively 、 and , , , , in 、 and are the probability distributions of the prediction errors of relative humidity, temperature and light intensity in the current window, 、 and are the probability distributions of historical stability period errors of relative humidity, temperature, and light intensity, respectively; S705: Preset drift alarm thresholds based on relative humidity, temperature, and light intensity 、 and ,judge Is it greater than 、 Is it greater than or Is it greater than If there is a case where is greater than , then the corresponding items in relative humidity, temperature, and light intensity in the current construction environment have significantly deviated from the environment during model training, and there is a risk of prediction failure. It is necessary to fine-tune the corresponding multi-model coordination mechanism for relative humidity, temperature, and light intensity. S706. Record the error changes before and after fine-tuning, the number of fine-tuning times, and the triggering reasons, generate a periodic report, and manually review it. If the manual review finds that a fine-tuning is incorrect, the parameters of the multi-model collaboration mechanism can be rolled back.

[0039] Example 2: In order to implement a method for predicting and managing moisture-proof building materials according to Example 1, based on Example 1, Figure 6 As shown, the present invention also provides a building material moisture-proof prediction and management system, including a meteorological data acquisition module 100, a measured data acquisition module 200, a multi-source data processing module 300, a collaborative prediction and processing module 400, a climate analysis and correction module 500, a risk level determination module 600, and a feedback evaluation and fine-tuning module 700. These modules cooperate with each other to realize environmental information perception, relative humidity prediction, temperature prediction, light intensity prediction and corresponding response control of the building material stacking area, as follows: Meteorological data acquisition module 100, which obtains weather forecast data and historical meteorological data of the past two years in the area where the building material stacking area is located through the open data platform API interface of the meteorological department. The meteorological data includes at least the temperature, relative humidity and light intensity at each hour. The historical meteorological data can be constructed through the hourly meteorological records of the past two years. , providing data support for subsequent climate anomaly analysis; The measured data acquisition module 200 is provided outside the building material stacking area and is used to measure the ambient temperature. , ambient humidity and ambient light intensity and at least one peripheral sensor group arranged inside the building material stacking area for measuring the internal temperature. , internal humidity and internal light intensity The internal sensor group builds a sensor Internet of Things for the building material stacking area, and obtains real-time sensor data through this sensor Internet of Things. The real-time sensor data includes the temperature, humidity and light intensity inside and outside the building material stacking area, that is, the external temperature , internal temperature , ambient humidity , internal humidity , ambient light intensity , internal light intensity ,These sensor sensors in the sensor Internet of Things adopt wireless ,collection method, and upload them to the remote server according to the set ,cycle and average the values ​​of each parameter of the sensor network to obtain ,real-time sensor representation data, which serves as the ,real-time detection data of temperature, humidity and light intensity in the ,building material stacking area; Multi-source data processing module 300: used to normalize the weather forecast data for the next hour and the real-time sensor data for the current hour respectively. If there are missing values, linear interpolation is used to supplement them to generate multi-dimensional sequence input features. , multi-dimensional sequence input features Including weather forecast temperature , Current measured temperature , Weather Forecast Relative Humidity , Current measured humidity , weather forecast light intensity and the current measured light intensity , and used for subsequent model prediction processing; The collaborative prediction processing module 400 independently models three pre-trained models, namely Autoformer, LSTM, and XGBoost, to realize three sets of prediction models for relative humidity, temperature, and light intensity. The Autoformer is trained based on multidimensional sequence features and trend-season decomposition mechanism, and adopts mean square error (MSE) as the loss function; the LSTM model is trained based on a gated recursive unit structure, using historical relative humidity, temperature, and light intensity sequences as input, and can output the predicted relative humidity, temperature, and light intensity at the next hour; the XGBoost model is based on the time series features expanded into the form of tabular features, and is trained using gradient boosting tree regression, with the training goal also to minimize the mean square error; for each parameter of relative humidity, temperature, and light intensity, the three pre-trained models of Autoformer, LSTM, and XGBoost are fused through a dynamic attention mechanism and a variance penalty coefficient to obtain the final relative humidity, temperature, and light intensity prediction results respectively; The climate analysis and correction module 500 is used to perform climate anomaly analysis based on historical meteorological data, calculate the quantile intervals of relative humidity, temperature, and light intensity by month, determine the historical normal intervals of relative humidity, temperature, and light intensity, and correct the predicted results of relative humidity, temperature, and light intensity based on the current month to obtain corrected relative humidity, temperature, and light intensity predicted values; The risk level determination module 600 is configured to use a multi-level response control strategy based on the corrected relative humidity, temperature, and light intensity prediction results to divide the humidity range into three risk levels and trigger control instructions corresponding to each level to be sent to the BIM system; Feedback evaluation and fine-tuning module 700: This module performs error analysis between each prediction result and the actual measurement result, and judges the health status of the model based on error statistics and drift detection (KL divergence). If the error or drift exceeds the threshold, the fine-tuning mechanism is triggered, and only the weight calculation logic or variance penalty coefficient of the fusion layer is updated to keep the model adaptable to field changes. At the same time, the feedback evaluation module records the fine-tuning process and generates periodic operation evaluation reports for manual review.

[0040] Example 3: Based on Example 1, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of a method for predicting and managing protection of a building material stacking area as described in Example 1 are implemented. The processor may be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a field programmable gate array (FPGA), a neural network acceleration chip (NPU), or a combination of the above processors.

[0041] Example 4: Based on Example 1, the present invention also provides a computer-readable storage medium. The computer-readable storage medium can be a non-volatile storage medium, including but not limited to: optical disks (CD-ROM, DVD-ROM), tapes, disks, flash memory chips, U disks, SD cards, mechanical hard disks, solid-state disks (SSDs) or other electrically writable and erasable storage devices. The computer-readable storage medium stores a computer program. When the computer program is executed by the processor, the steps of the construction material stacking area protection prediction management method as in Example 1 are implemented. The computer program can exist in the form of an installation package, a script file, a container image, a virtual machine image, a cloud service configuration file, etc., and is adaptable to various computing platforms in this field.

Claims

1. A method for predicting and managing protection of a building material stacking area, characterized by: The steps include: S1. Obtain weather forecast data and the past two years' historical weather data for the construction site's building material stacking area from the API interface of the meteorological department's open data platform. The weather data should include at least the temperature, relative humidity, and light intensity at each hour. S2. Real-time sensor data is acquired through a sensor network deployed in the building material stacking area. The real-time sensor data includes the temperature, humidity, and light intensity of the building material stacking area. The real-time sensor characterization data is obtained by averaging each parameter of the sensor network. S3: Normalize the weather forecast data for the next hour and the real-time sensor data for the current hour. If there are missing values, use linear interpolation to supplement them, and then splice them to form a multi-dimensional sequence input feature. Based on the multi-model collaborative mechanism, the relative humidity, temperature and light intensity of the next hour are predicted and processed to obtain the relative humidity prediction value. , temperature prediction value and light intensity prediction ; S4. Based on historical meteorological data, climate anomaly analysis was conducted, and the quantile intervals of humidity, temperature and light intensity were calculated by month to determine the historical normal intervals of humidity, temperature and light intensity. , ]、[ , ][ , ], and correct the relative humidity forecast value, temperature forecast value and light intensity forecast value according to the current month to obtain the corrected relative humidity forecast value , temperature prediction value and light intensity prediction ; S5. Corrected relative humidity prediction results based on the corrected , temperature prediction value and light intensity prediction ,adopting a multi-level response control strategy, the humidity range is divided into three risk levels, and the control instructions corresponding to each level are triggered; S6. The corrected relative humidity prediction value , temperature prediction value , light intensity prediction value , risk levels and control instructions are sent to the BIM system through the communication interface for construction management personnel to use as reference suggestions for moisture-proof management of building materials.

2. A method for predicting and managing moisture-proofing of building materials according to claim 1, characterized in that: The sensor network includes a peripheral sensor group and an internal sensor group. The peripheral sensor group is set outside the building material stacking area and is used to measure the ambient temperature, ambient humidity and ambient light intensity outside the stacking area. The internal sensor group is set inside the building material stacking area and is used to measure the temperature, humidity and light intensity inside the stacking area.

3. A method for predicting and managing moisture-proofing of building materials according to claim 2, characterized in that: In step S3, the multidimensional sequence input features Including weather forecast temperature , Current measured temperature , Weather Forecast Relative Humidity , Current measured humidity , Weather forecast light intensity and the current measured light intensity .

4. A method and system for predicting and managing moisture-proofing of building materials according to claim 3, characterized in that: In step S3, the multi-model collaboration mechanism includes the following steps: S301. At the beginning of the earthwork excavation phase of the construction project, collect historical meteorological data of the hourly weather forecast and the corresponding measured meteorological data of the building material stacking area at the previous moment to form a model training data set; S302: Based on the collected model training data set, the Autoformer model, LSTM model and XGBoost model are trained independently, and the input is the multi-dimensional sequence input feature. , the output is the relative humidity forecast value for the next hour , temperature prediction value and light intensity prediction , complete the pre-training of the Autoformer model, LSTM model and XGBoost model; S303: Deploy the pre-trained Autoformer model, LSTM model, and XGBoost model independently, and input multi-dimensional sequence input features. , and obtain three independent sets of prediction values ​​( 、 、 ), ( 、 、 )and( 、 、 ), corresponding to relative humidity, temperature and light intensity respectively; S304: Forward pass the three independent pre-trained Autoformer models, LSTM models, and XGBoost models once respectively, and start Monte Carlo Dropout to make multiple predictions during inference, so as to obtain K different prediction results for each prediction item, and obtain three prediction sets. 、 and , , based on the prediction set 、 and , respectively calculate the corresponding model prediction mean 、 and and the corresponding variance 、 and , ， , ， , , ； S305. The following uses relative humidity for processing. The temperature and light intensity are processed in a similar way to the relative humidity. The processing processes are independent of each other. The multidimensional sequence input features of the current time step are used. , through the fully connected layer Generate attention query vector , ; S306. Calculate the attention key value based on the model prediction mean and attention value , ， ； S307, based on the attention mechanism and confidence adjustment, combined with the learnable variance penalty coefficient , and obtain the dynamic weights of the three models respectively , ， in Represents the relative humidity fusion weight of the i-th model, i=1, 2, 3, It is a learnable parameter that is dynamically adjusted during the online adaptation process. It represents the variance penalty coefficient and controls the impact of uncertainty on the weight. S308: Output the final relative humidity fusion prediction based on the prediction mean and dynamic weight of each model , ， in, represents the fusion weight of the i-th model, Represents the predicted mean of each model, i takes values ​​of 1, 2, and 3 respectively; S309, at each time step, based on the final relative humidity fusion prediction results Compared with the actual observed humidity The error between the two is used to calculate the loss function , ， Using this loss function to backpropagate, only update , without updating the model parameters, ， in, for The learning rate, and in order to prevent Too large or too small will cause failure, set The value range is and between, and Based on experience Upper and lower limits of values; S310, the temperature and light intensity are also processed with reference to the relative humidity, and steps S303 to S309 are performed independently. In these steps, the relative humidity is replaced by temperature or light intensity, and the temperature fusion prediction is finally output. Fusion prediction with light intensity .

5. A method for predicting and managing moisture-proofing of building materials according to claim 4, characterized in that: In step S4, climate anomaly analysis includes the following steps: S401, historical meteorological data acquisition and database construction phase, that is, to obtain the hourly meteorological observation historical data of the area where the building material stacking area is located in the past two years, including at least the temperature, relative humidity and light intensity at each hour, to form a historical meteorological database ; S402, the monthly grouping and statistical feature extraction stage, that is, statistically analyzing the quantile intervals of humidity, temperature, and light intensity by month, determining the historical normal ranges of humidity, temperature, and light intensity, and correcting the relative humidity prediction value, temperature prediction value, and light intensity prediction value based on the current month, and calculating the average humidity data for each month. , standard deviation , minimum value , maximum value , P10 quantile , P50 quantile , P75 quantile , P90 quantile ; S403, the climate anomaly model construction phase, that is, based on the quantile statistics results, determine the historical normal range of humidity, temperature and light intensity for each month [ , ]、[ , ][ , ], building a climate anomaly model for each month 、 and ; S404, climate anomaly correction factor calculation and prediction result correction stage, that is, according to the current month m, the corresponding climate anomaly model is extracted 、 and , the relative humidity prediction value output by the multi-model synergy mechanism , temperature prediction value and light intensity prediction , respectively, perform climate anomaly correction to obtain the corrected relative humidity forecast value , temperature prediction value , light intensity prediction value .

6. A method for predicting and managing moisture-proofing of building materials according to claim 5, characterized in that: In step S5, the multi-level response control strategy includes the following steps: S501. According to expert experience, the storage conditions for building materials are relative humidity of 45%-60%, temperature should be below 45℃, and light intensity should be below 1000lux. Based on this, the deviation is set to divide the risk. The relative humidity deviation : ， If the relative humidity exceeds the range of [45%, 60%], the greater the deviation, the higher the risk level. The temperature deviation ; ， If the temperature exceeds 45°C, the greater the deviation, the higher the risk level; the light intensity deviation : ， This means that if the light intensity exceeds 1000 lux, the greater the deviation, the higher the risk level; S502. Calculate the comprehensive risk score using the weighted risk score formula : ， in, 、 and are the humidity risk weight, temperature risk weight, and light intensity risk weight, which are determined by expert experience and are generally 0.33 respectively, with a total of 1; S503, based on comprehensive risk score There are three risk levels: low risk ( ), medium risk ( ) and high risk ( ), low risk means all conditions are close to the standard value with no significant deviation, medium risk means some deviation, but not extreme, and high risk means significant deviation, and environmental conditions may cause damage to building materials; S504. Determine corresponding operation instructions based on the risk level. When the risk is low, adopt a no-response strategy; when the risk is medium, adopt a close monitoring strategy; when the risk is high, adopt a cover-up strategy. S505. Push the operation instructions to the BIM system through the communication interface, so that the construction management personnel can use them as reference suggestions for moisture-proof management of building materials.

7. A method for predicting and managing moisture-proofing of building materials according to claim 6, characterized in that: It also includes the steps of feedback evaluation, which includes the following steps: S701, real-time calculation of relative humidity prediction error , temperature prediction error and light intensity prediction error , ， ， ， in 、 and They represent the relative humidity prediction error, temperature prediction error and light intensity prediction error at time t+1 respectively, 、 and Indicates the actual measured relative humidity, temperature and light intensity values. 、 and Indicates the revised predicted relative humidity value, predicted temperature value and predicted light intensity value; S702: Calculate the sliding mean square error for relative humidity, temperature, and light intensity. 、 and , and mean absolute percentage error 、 and , ， , , , , ， Where n represents the length of the sliding window based on a 24-hour environmental change cycle; S703: Safety threshold based on historical best performance set by experience 、 、 、 、 、 Determine whether the model's ability to predict the current environment is insufficient. If the sliding mean square error or mean absolute percentage error of any of the three items (relative humidity, temperature, or light intensity) is greater than the corresponding safety threshold, it indicates a persistently high error and insufficient predictive ability for the current environment. Fine-tune the multi-model coordination mechanism for this item. S704, calculate the KL divergence of the error distribution for relative humidity, temperature or light intensity respectively 、 and , ， ， ， in 、 and are the probability distributions of the prediction errors of relative humidity, temperature and light intensity in the current window, 、 and are the probability distributions of historical stability period errors of relative humidity, temperature, and light intensity, respectively; S705: Preset drift alarm thresholds based on relative humidity, temperature, and light intensity 、 and ,judge Is it greater than 、 Is it greater than or Is it greater than If there is a case where is greater than , then the corresponding items in relative humidity, temperature, and light intensity in the current construction environment have significantly deviated from the environment during model training, and there is a risk of prediction failure. It is necessary to fine-tune the corresponding multi-model coordination mechanism for relative humidity, temperature, and light intensity. S706. Record the error changes before and after fine-tuning, the number of fine-tuning times, and the triggering reasons, generate a periodic report, and manually review it. If the manual review finds that a fine-tuning is incorrect, the parameters of the multi-model collaboration mechanism can be rolled back.

8. A building material stacking area protection prediction and management system, characterized by: include: The meteorological data collection module is used to obtain weather forecast data and historical meteorological data from the past two years for the area where the building material stacking area is located through the API interface of the meteorological department's open data platform. The meteorological data includes at least the temperature, relative humidity, and light intensity at each hour. The historical meteorological data can be used to build a historical database by recording hourly meteorological data from the past two years, providing data support for subsequent climate anomaly analysis. A measured data acquisition module is used to obtain real-time ambient temperature, humidity, and light intensity data inside and outside the building material stacking area through a peripheral sensor group installed outside the building material stacking area and an internal sensor group installed inside the building material stacking area. The peripheral sensor group is used to measure the ambient temperature, humidity, and light intensity, and the internal sensor group is used to measure the internal temperature, humidity, and light intensity. The real-time collected data is uploaded to a remote server via wireless means, and the mean value of each parameter is calculated to obtain real-time sensor characterization data; Multi-source data preprocessing module: This module is used to normalize the weather forecast data for the next hour and the real-time sensor data for the current hour. If there are missing values, linear interpolation is used to supplement them to generate multidimensional sequence input features. The input features include the forecast temperature, the current measured temperature, the forecast relative humidity, the current measured humidity, the forecast light intensity, and the current measured light intensity, which are used for subsequent model prediction processing. The collaborative prediction processing module is used to independently build models using three pre-trained models: Autoformer, LSTM, and XGBoost, to predict relative humidity, temperature, and light intensity. The Autoformer is trained based on multidimensional sequence features and a trend-season decomposition mechanism, the LSTM model is trained based on a gated recursive unit structure, and the XGBoost model is trained based on the expansion of time series features into tabular features. The prediction results of the three models are then fused through a dynamic attention mechanism and a variance penalty coefficient to obtain the final predicted values ​​of relative humidity, temperature, and light intensity. The climate analysis and correction module is used to perform climate anomaly analysis based on historical meteorological data, calculate the quantile intervals of relative humidity, temperature, and light intensity by month, determine the historical normal intervals for each month, and correct the forecast results of relative humidity, temperature, and light intensity based on the current month to obtain the corrected relative humidity, temperature, and light intensity forecast values; The risk level determination module is used to adopt a multi-level response control strategy based on the corrected relative humidity, temperature, and light intensity prediction results, divide the humidity range into three risk levels, and trigger the control instructions corresponding to each level to be sent to the BIM system; The feedback evaluation and fine-tuning module is used to analyze the error between each prediction result and the actual measurement result, and judge the health status of the model based on error statistics and drift detection (KL divergence). If the error or drift exceeds the threshold, the fine-tuning mechanism is triggered, and only the weight calculation logic or variance penalty coefficient of the fusion layer is updated to keep the model adaptable to field changes. At the same time, the feedback evaluation module records the fine-tuning process and generates periodic operation evaluation reports for manual review.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method for predicting and managing protection of a building material stacking area as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting and managing protection of a building material stacking area as claimed in any one of claims 1 to 7 are implemented.

Images