A Neural Network-Based Analysis Method and System for Temperature Regulation Efficiency of Light-Green Combined Roofs

By constructing time-series input data on moisture and training a neural network, the problem of distinguishing preceding moisture processes in the analysis of the thermal regulation efficiency of light-green composite roofs was solved, and accurate analysis of the thermal response of light-green composite roofs was achieved.

CN123115707APending Publication Date: 2026-11-17TONGJI UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611624195.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing neural network-based methods for analyzing the temperature regulation efficiency of light-green roof combinations are unable to distinguish between different preceding moisture processes, resulting in insufficient accuracy in temperature regulation thermal response analysis.

Method used

By collecting measured data from the light-green roof combination, moisture time-series input data was constructed, the characteristics of the preceding moisture process were determined, and the average values ​​of environmental parameters were calculated. Analysis time points with differences within the measurement error range were selected to form training sample pairs. A fully connected layer was used to train a moisture history and thermal response mapping neural network to analyze the differences in canopy temperature, matrix midpoint temperature, and roof heat flow.

Benefits of technology

The neural network's ability to nonlinearly map moisture history and thermal response has been enhanced, the interference from differences in moisture and environmental conditions has been reduced, the impact of moisture changes on temperature regulation performance has been refined, and the accuracy of the analysis has been improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115707A_ABST
    Figure CN123115707A_ABST
Patent Text Reader

Abstract

This application discloses a method and system for analyzing the temperature regulation efficiency of light-green composite roofs based on neural networks, belonging to the field of neural network technology. The method includes: S1, constructing time-series input data on moisture content and determining the characteristics of preceding moisture processes and the average values ​​of environmental parameters; S2, selecting two analysis moments with similar current upper layer moisture content, current lower layer moisture content, and average environmental parameters, but with differences in the characteristics of preceding moisture processes, to form a training sample pair, and training a moisture history and thermal response mapping neural network; S3, obtaining calculated values ​​for canopy temperature, matrix midpoint temperature, and roof heat flux through the moisture history and thermal response mapping neural network; S4, performing thermal response difference analysis and recording the corresponding differences with historical moisture processes. This solves the problem that it is difficult to distinguish different preceding moisture processes based on the current matrix moisture content, leading to inaccurate thermal response analysis of light-green composite roofs.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of neural network technology, and in particular to a method and system for analyzing the temperature regulation efficiency of light-green roof combinations based on neural networks. Background Technology

[0002] With the development of building energy-saving technologies and roof space integration technologies, photovoltaic (PV) power generation combined with roof greening has gradually attracted attention. PV panels can alter the solar radiation received and the local thermal environment of green areas, while the green substrate and plants can influence roof temperature and heat transfer processes through heat storage, evaporation, and transpiration. Therefore, the temperature regulation performance of photovoltaic-green roofs is affected by various factors such as solar irradiance, air temperature, relative humidity, wind speed, substrate moisture content, and plant transpiration. These factors change over time, creating a complex nonlinear relationship between canopy temperature, substrate temperature, and roof heat flow. Existing research has begun to use data-driven models and neural networks to analyze and predict building energy consumption, building envelope heat flow, and related environmental responses.

[0003] For example, the invention patent with announcement number CN110046743B discloses a method and system for predicting the energy consumption of public buildings based on GAANN. The method involves collecting hourly energy consumption data of public buildings and its influencing factors, organizing and preprocessing the data; dividing the data into training and test sets, and selecting input variables using the correlation coefficient method; inputting test data, optimizing the relevant parameters of the artificial neural network model using a genetic algorithm, and then training the model using the training set data; predicting the energy consumption of public buildings by inputting the input variables for the predicted period; and finally evaluating the prediction effect on the test set using an error index, providing an allowable error range.

[0004] For example, the invention patent with announcement number CN109685252B discloses a building energy consumption prediction method based on recurrent neural networks and multi-task learning models, belonging to the field of integrated energy management technology. It addresses the technical problem of parallel prediction of multiple types of energy consumption while ensuring prediction accuracy and shortening model training time. This method first acquires building energy consumption data samples, processes missing data using "averaging data from multiple similar time points," then constructs multiple learning tasks based on energy consumption type, time step, and initial time. Next, the dataset for each learning task is normalized, and the similarity between the training sets of multiple tasks is measured using the Pearson correlation coefficient. After ensuring the similarity between multiple tasks, a neural network model is created and trained. Finally, a multi-task CIFGLSTM neural network model is used to predict comprehensive energy consumption.

[0005] However, for light-green composite roofs, the substrate moisture content is not only affected by rainfall and irrigation, but also changes with plant water absorption and surface evaporation. The same moisture content value may appear during the water loss process after irrigation and during the water loss process after a period of evaporation. At this time, the internal moisture distribution of the substrate, surface evaporation conditions, and available water for plants may differ, and the corresponding canopy temperature, substrate temperature, and roof heat flow response may also be different. Existing neural network-based building thermal environment or energy consumption analysis methods mainly rely on monitoring parameters at the sampling time or historical data within a certain time range to establish a mapping relationship between input and prediction results. When using the current substrate moisture content to characterize the moisture conditions of the green area, it is difficult to identify the moisture change process before the formation of the corresponding moisture state based solely on the current moisture content value. This can easily lead to a lack of effective differentiation between data with similar current moisture contents but different preceding moisture change processes during the analysis, thus affecting the analysis of differences in the actual temperature regulation efficiency of light-green composite roofs.

[0006] Therefore, in order to address the above problems, there is an urgent need for a method and system for analyzing the temperature regulation efficiency of light-green roof combinations based on neural networks. Summary of the Invention

[0007] This application provides a method and system for analyzing the temperature regulation efficiency of light-green composite roofs based on neural networks, which solves the problem that it is difficult to distinguish different preceding moisture processes based solely on the current substrate moisture content, resulting in insufficient accuracy in the analysis of the temperature regulation thermal response of light-green composite roofs.

[0008] This application provides a method for analyzing the temperature regulation efficiency of a light-green roof combination based on a neural network, including: S1, collecting measured data of the light-green roof combination, constructing moisture time-series input data based on the upper layer moisture content sequence and the lower layer moisture content sequence, determining the characteristics of the preceding moisture process, and calculating the average environmental parameters; S2, selecting two analysis times where the differences between the current upper layer moisture content, the current lower layer moisture content, and the average environmental parameters are within the corresponding measurement error range, and where the characteristics of the preceding moisture process differ, forming a training sample pair, inputting the moisture time-series input data, the characteristics of the preceding moisture process, and the average environmental parameters into a fully connected layer, and then... The moisture history and thermal response mapping neural network is trained using single-sample average error and sample-pair average error; S3, moisture time-series input data is constructed for the new analysis time, the characteristics of the previous moisture process and the average values ​​of environmental parameters are determined, and these are input into the moisture history and thermal response mapping neural network to obtain the calculated values ​​of canopy temperature, matrix midpoint temperature and roof heat flow; S4, thermal response difference analysis is performed using the current upper layer moisture content, current lower layer moisture content, average values ​​of environmental parameters, calculated values ​​of canopy temperature, calculated values ​​of matrix midpoint temperature and calculated values ​​of roof heat flow, and the results of thermal response difference analysis are recorded in correspondence with historical moisture process differences.

[0009] Another embodiment of this application provides a neural network-based system for analyzing the temperature regulation efficiency of a light-green roof combination, comprising: a data and historical feature acquisition unit, used to collect measured data of the light-green roof combination, construct moisture time-series input data based on the upper layer moisture content sequence and the lower layer moisture content sequence, determine the characteristics of the preceding moisture process, and calculate the average environmental parameter value; and a sample construction and network training unit, used to select two analysis times where the differences between the current upper layer moisture content, the current lower layer moisture content, and the average environmental parameter value are within the corresponding measurement error range, and where the characteristics of the preceding moisture process differ, to form a training sample pair, and input the moisture time-series input data, the characteristics of the preceding moisture process, and the average environmental parameter value into a fully connected layer. The system employs a moisture history and thermal response mapping neural network trained based on the single-sample average error and sample-pair average error. A roof thermal response calculation unit is used to construct time-series input data for new analysis moments, determine the characteristics of previous moisture processes and the average values ​​of environmental parameters, and input these into the moisture history and thermal response mapping neural network to obtain calculated values ​​for canopy temperature, matrix midpoint temperature, and roof heat flow. A temperature difference analysis unit is used to perform thermal response difference analysis using the current upper layer moisture content, current lower layer moisture content, average environmental parameters, calculated canopy temperature, calculated matrix midpoint temperature, and calculated roof heat flow, and records the results of the thermal response difference analysis in relation to historical moisture process differences.

[0010] The beneficial effects achieved by the technical solutions provided in this application include at least the following: (1) In this invention, moisture time-series input data is extracted by one-dimensional time-series convolution to extract moisture time-series features. The moisture time-series features, along with the features of preceding moisture processes and the average values ​​of environmental parameters, are used together for neural network modeling. Furthermore, the network parameters are trained by combining the single-sample average error and the sample-pair average error. This enables the moisture history and thermal response mapping neural network to learn the nonlinear mapping relationship between different preceding moisture processes and canopy temperature, matrix midpoint temperature, and roof heat flow. This enhances the neural network's ability to characterize differences in moisture history and corresponding thermal response differences, effectively solving the problem in the prior art where the neural network mainly maps thermal response based on the current moisture content, making it difficult to identify the influence of different preceding moisture processes.

[0011] (2) In this invention, by selecting two analysis times with different characteristics of the preceding water process from the analysis times when the difference between the current upper layer water content, the current lower layer water content and the average value of environmental parameters is within the corresponding measurement error range, the training process can highlight the difference in thermal response formed by the preceding water process when the current water conditions and external environmental conditions are similar, thereby achieving the effect of reducing the interference of the difference between the current water conditions and environmental conditions on the analysis of water process, and effectively solving the problem in the prior art that samples with similar current water content but different preceding water processes are easily treated as similar water conditions.

[0012] (3) In this invention, water time series input data is constructed by upper layer water content sequence and lower layer water content sequence, and the preceding water process characteristics are formed based on the upper layer historical net change, lower layer historical net change, upper layer water increase detection mark, lower layer water increase detection mark, lower layer response detection mark and upper and lower layer first detection interval input items. This enables the analysis data to reflect the changes in water content at different depths with the sampling time and the response relationship between upper and lower layers, thereby realizing the effect of characterizing the preceding water change process from the dual-depth water time series, effectively solving the problem that the water content at a single moment in the prior art is difficult to reflect the preceding water change process of the greening matrix.

[0013] (4) This invention obtains the calculated values ​​of canopy temperature, matrix midpoint temperature and roof heat flow through a moisture history and thermal response mapping neural network. It also performs thermal response difference analysis on two analysis moments with similar current moisture and environmental conditions, and records the differences in canopy temperature, matrix midpoint temperature and roof heat flow with the differences in historical moisture processes. This achieves the effect of establishing the correspondence between the differences in previous moisture processes and the differences in thermal response of the light-green combination roof. It effectively solves the problem in the prior art that the roof thermal response analysis results lack a correspondence with the previous moisture processes and it is difficult to refine the analysis of the influence of moisture history on temperature regulation performance.

[0014] It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the method for analyzing the temperature regulation efficiency of light-green roof combinations based on neural networks, provided in this application. Figure 2 This is a structural diagram of the neural network-based light-green roof temperature regulation efficiency analysis system provided in this application; Figure 3 This is a schematic diagram of the layout for collecting measured data of the light-green combination roof provided in this application; Figure 4 This is a schematic diagram constructed based on the time-series input data of moisture and the characteristics of preceding moisture processes provided in this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0018] Please see Figures 1-4 This invention provides a technical solution: a method for analyzing the temperature regulation efficiency of a light-green roof combination based on neural networks, including S1, collecting measured data of the light-green roof combination, constructing moisture time-series input data based on the upper layer moisture content sequence and the lower layer moisture content sequence, determining the characteristics of the preceding moisture process, and calculating the average value of environmental parameters; S2, selecting two analysis times where the differences between the current upper layer moisture content, the current lower layer moisture content, and the average value of environmental parameters are within the corresponding measurement error range, and where the characteristics of the preceding moisture process differ, forming a training sample pair, inputting the moisture time-series input data, the characteristics of the preceding moisture process, and the average value of environmental parameters into a fully connected layer, and then... A moisture history and thermal response mapping neural network is trained based on the single-sample average error and sample-pair average error; S3, moisture time-series input data is constructed for the new analysis time, the characteristics of the previous moisture process and the average value of environmental parameters are determined, and these are input into the moisture history and thermal response mapping neural network to obtain the calculated values ​​of canopy temperature, matrix midpoint temperature and roof heat flow; S4, thermal response difference analysis is performed using the current upper layer moisture content, current lower layer moisture content, average environmental parameters, calculated canopy temperature, calculated matrix midpoint temperature and calculated roof heat flow, and the results of the thermal response difference analysis are recorded in correspondence with the differences in historical moisture processes.

[0019] Specifically, the measured data collected for the photovoltaic-green roof combination included: collecting data on the upper layer moisture content, lower layer moisture content, canopy temperature, midpoint temperature of the substrate, roof heat flux, air temperature, relative humidity, wind speed, and shortwave irradiance of the green surface at the same sampling interval for the area below the photovoltaic panels and adjacent green areas, and recording the sampling time; specifically, corresponding measuring points were arranged in the green area below the photovoltaic panels and adjacent green areas respectively. The upper layer moisture content was collected at the upper part of the green substrate, the lower layer moisture content was collected at the lower part of the green substrate, and the canopy temperature was collected at the location corresponding to the plant canopy. The temperature at the midpoint of the substrate was collected at the midpoint of the substrate thickness, while the roof heat flow was collected above the waterproof layer beneath the substrate. Air temperature, relative humidity, and wind speed were collected above the corresponding green areas, and shortwave irradiance on the green surface was collected at the corresponding green surface location. This ensures a consistent spatial correspondence between the upper and lower layer moisture content, as well as the corresponding thermal response and environmental parameters within the same area. Data collected from the green area beneath the photovoltaic panels and adjacent green areas were arranged according to their respective sampling times, and measurements from two areas were not placed in the same field. The data are mixed to form a single sequence. When analyzing any region, the measured data obtained within that region are used. When comparing two regions, data with the same sampling time are used for comparison. The sampling interval is determined based on the effective update cycle of each acquisition parameter and the time scale of the matrix moisture change process. Multiple sampling times are included within the historical time window. Each acquisition channel uses the same acquisition clock for sampling and recording the sampling time. Parameters corresponding to the same sampling are used as data at the same sampling time, and data from different sampling times are not spliced ​​together. When an individual acquisition parameter does not obtain a valid measurement value at the corresponding sampling time, a missing value marker is written at the corresponding position. Data containing missing values ​​is not used to construct the moisture time series input data for the corresponding analysis time. Measurement values ​​from adjacent sampling times are not used to replace missing values, thus avoiding changes to the actual time relationship corresponding to the upper layer change sequence, the lower layer change sequence, and the first detection interval input items for the upper and lower layers. The upper layer moisture content and the lower layer moisture content are arranged according to the sampling time to form upper layer moisture content sequences and lower layer moisture content sequences. Measured values ​​of canopy temperature, matrix midpoint temperature, and roof heat flow are recorded. Figure 3 The diagram shows the data collection layout for the photovoltaic-green roof in this embodiment. It illustrates the collection locations of the green area below the photovoltaic panel and adjacent green areas, and shows the corresponding collection locations for the upper layer moisture content, lower layer moisture content, canopy temperature, midpoint temperature of the substrate, roof heat flow, air temperature, relative humidity, wind speed, and shortwave irradiance of the green surface. The upper layer moisture content and lower layer moisture content correspond to different depths of the green substrate, the midpoint temperature of the substrate corresponds to the midpoint of the green substrate thickness, and the roof heat flow corresponds to the location below the green substrate, so that the moisture data, thermal response data, and environmental parameters of the same area correspond to each other according to the sampling time.

[0020] In this implementation plan, by synchronously collecting data on the upper and lower layers of moisture content, canopy temperature, midpoint temperature of the matrix, roof heat flux, and environmental parameters in the area below the photovoltaic panel and adjacent green areas, and by unifying the sampling time and missing measurement handling rules, the consistency of the upper and lower layer moisture content sequences with the corresponding thermal response data and environmental parameters in terms of spatial location and time dimension can be ensured. This provides a reliable data foundation for the subsequent construction of moisture time series input data, determination of the characteristics of preceding moisture processes, and conduct thermal response difference analysis.

[0021] Specifically, constructing the moisture time-series input data based on the upper and lower layer moisture content sequences includes: taking the analysis time as the end of the historical time window, reading the upper and lower layer moisture contents at n sampling times backward, where n represents the number of sampling times participating in the moisture time-series analysis within the historical time window, and n is a positive integer greater than one, determined according to the length of the historical time window and the sampling interval, and using the same value at different analysis times to ensure that the upper and lower layer moisture contents read at each analysis time have the same number of sampling points, so as to form an equal-length time-series input matrix subsequently; determining the earliest sampling time among the n sampling times... The historical time window is defined as the starting point. The upper and lower layer water contents at the end of the historical time window are respectively determined as the current upper and lower layer water contents. The time span from the starting point of the historical time window to the analysis time is determined as the historical time window length. Since the sampling times are arranged according to the same sampling interval, the same n value corresponds to the same historical time window length. The upper layer water content minus the lower layer water content at the same sampling time forms a sequence of upper and lower layer water content differences, which is used to characterize the water content difference between different depths of the greening matrix at the same sampling time. The water content difference at the same depth at two adjacent sampling times is calculated to form... The data is divided into upper-layer and lower-layer change sequences. The upper-layer change sequence reflects the direction and magnitude of change in upper-layer water content between adjacent sampling times, while the lower-layer change sequence reflects the direction and magnitude of change in lower-layer water content between adjacent sampling times. Zero values ​​are written at the beginning of both sequences to ensure they have the same data length and sampling time correspondence with the upper-layer and lower-layer water content sequences, respectively. The water content at the end of the historical time window is subtracted from the water content at the beginning of the historical time window. The water content of the upper and lower layers at different depths is used to obtain the historical net change in the upper and lower layers, which reflects the overall change in water content at the two depths from the start of the historical time window to the analysis time. The water content sequences of the upper and lower layers, the water content sequences of the lower layers, the water content sequence of the upper layer, the water content sequence of the lower layer, and the water content difference sequence between the upper and lower layers are used as water time series input data. Each of the five sequences contains n data positions arranged according to the sampling time, so that the water time series input data formed at different analysis times maintain the same data length and arrangement, providing a data basis for the subsequent formation of an equal-length time series input matrix according to the sampling time.

[0022] In this implementation scheme, a historical time window is constructed by unifying n sampling times, and upper layer water content sequence, lower layer water content sequence, upper layer change sequence, lower layer change sequence, and upper and lower layer water content difference sequence are formed. This ensures that the water time series input data at different analysis times have a consistent data length and sampling time correspondence. The upper layer historical net change and the lower layer historical net change are combined to characterize the water content change process at different depths within the historical time window, thereby providing a unified data foundation for the subsequent formation of an equal-length time series input matrix and the extraction of features of the preceding water process.

[0023] Specifically, determining the characteristics of the preceding moisture process and calculating the average environmental parameters includes: when there is a positive value in the upper layer change sequence that is greater than the sum of the upper layer moisture content measurement absolute error limits of the two preceding and following upper layer change measurements, the sampling time corresponding to the earliest upper layer change that meets the condition is determined as the upper layer first detection time, and the upper layer moisture increase detection flag is written as 1; otherwise, it is written as 0. In this embodiment, the upper layer moisture content measurement absolute error limit is taken from the maximum absolute measurement error under the corresponding working conditions in the metrological performance data and calibration records of the probe used for upper layer moisture content acquisition, using the same unit of measurement as the upper layer moisture content, and not replacing the upper layer moisture content measurement absolute error limit with the fluctuation range of historical sampling data; At the sampling times preceding and following a certain upper-layer change, the upper-layer moisture content measurement absolute error limits applicable to the two sampling times are read and summed. This sum is used as the criterion for determining whether the corresponding upper-layer change exceeds the influence of measurement error, thus preventing small positive changes caused by the probe's allowable measurement error from being identified as an increase in moisture. When a positive value exists in the bottom-layer change sequence that is greater than the sum of the upper-layer moisture content measurement absolute error limits of the two preceding and following bottom-layer changes, the sampling time corresponding to the earliest occurrence of the condition-compliant bottom-layer change is determined as the first detection time of the bottom layer, and the bottom-layer moisture increase detection flag is written as 1; otherwise, it is written as 0. The upper-layer moisture content measurement absolute error limit is determined according to the upper-layer... The upper limit of the absolute error for moisture content measurement is determined based on the metrological performance data and calibration records of the probe used for bottom layer moisture content acquisition, and uses the same unit of measurement as the bottom layer moisture content. This ensures that both the upper and lower layer change sequences use the corresponding depth moisture content acquisition accuracy as the error boundary for moisture increase detection. When both the upper and lower layer moisture increase detection flags are one, and the first detection time of the lower layer is not earlier than the first detection time of the upper layer, the lower layer response detection flag is written as one, and the difference between the first detection time of the lower layer and the first detection time of the upper layer is written into the upper and lower layer first detection interval input field. Since both the first detection time of the upper layer and the first detection time of the lower layer are taken from the sampling time written under the same sampling interval... The upper and lower layer first detection interval input item directly maintains the actual time interval between the corresponding sampling times; when the upper layer moisture increase detection flag is zero, the lower layer moisture increase detection flag is zero, or the lower layer first detection time is earlier than the upper layer first detection time, the lower layer response detection flag is written to zero, and the sum of the historical time window length and a sampling interval is written into the upper and lower layer first detection interval input item, so that the analysis time that has not formed a corresponding lower layer response has a unified input value that is different from the effective first detection interval within the historical time window; the upper layer historical net change, the lower layer historical net change, the upper layer moisture increase detection flag, the lower layer moisture increase detection flag, the lower layer response detection flag, and the upper and lower layer first detection interval input item are used as features of the preceding moisture process;The average values ​​of air temperature, relative humidity, wind speed, and shortwave irradiance on green surfaces are calculated as the average environmental parameters. Specifically, each environmental parameter average is calculated based on the same historical time window corresponding to the current analysis time. Air temperature, relative humidity, wind speed, and shortwave irradiance on green surfaces are read from each sampling time from the beginning to the end of the historical time window. Only sampling locations with valid measurements are included in the arithmetic mean calculation of the corresponding environmental parameter. Sampling locations with missing markers are not included in the calculation of the corresponding environmental parameter average, nor are measurements from adjacent sampling times used to replace missing measurements. When a valid measurement value for a certain environmental parameter cannot be obtained within a historical time window, the environmental parameter average for the corresponding analysis time is not generated, and the corresponding analysis time is not used for constructing subsequent training sample pairs. This ensures that the average air temperature, relative humidity, wind speed, and shortwave irradiance on green surfaces have a consistent historical time range and sampling time basis with the moisture time-series input data. For example; Figure 4 The diagram shows the moisture time-series input data and the construction of preceding moisture process characteristics in this embodiment. The upper layer moisture content sequence and the lower layer moisture content sequence are arranged according to the sampling time, and the starting point of the historical time window, the analysis time, the current upper layer moisture content, the current lower layer moisture content, the first detection time of the upper layer, the first detection time of the lower layer, and the interval between the first detections of the upper and lower layers are marked. The upper layer change sequence and the lower layer change sequence are shown below the corresponding sampling time to represent the changes in moisture content at different depths within the historical time window with the sampling time and the sequential detection relationship of the increase in moisture in the upper and lower layers.

[0024] In this implementation scheme, by combining the upper limit of the absolute error of the upper layer moisture content measurement and the upper limit of the absolute error of the lower layer moisture content measurement, the first detection time of the upper layer and the first detection time of the lower layer are identified, and based on this, upper layer moisture increase detection flag, lower layer moisture increase detection flag, lower layer response detection flag, and upper and lower layer first detection interval input items are formed, which can reduce the interference of measurement error on moisture change identification. Furthermore, by combining the historical net change of the upper layer, the historical net change of the lower layer, and the average value of environmental parameters within the same historical time window, the preceding moisture process characteristics are formed, which can provide a data foundation with a unified time range for subsequent training samples to train the neural network for mapping the structure and moisture history with thermal response.

[0025] Specifically, selecting two analysis timestamps to form a training sample pair, where the differences between the current upper layer moisture content, the current lower layer moisture content, and the average environmental parameters are within the corresponding measurement error range and the characteristics of the preceding moisture processes differ, includes: selecting two analysis timestamps from different analysis timestamps. In this embodiment, to avoid interference from differences in spatial location and sampling conditions on the training sample pair, selection is prioritized from analysis timestamps corresponding to the same measurement point; when measurement points are set below the photovoltaic panel and in adjacent green areas, cross-regional measurement is only performed under the condition that the roof structure, greening matrix composition, matrix thickness, and measurement point arrangement are consistent. Measurement points were selected, and the same sampling interval, the same n value, and the same historical time window length were ensured for the two analysis times. This ensured that the upper layer moisture content sequence, the lower layer moisture content sequence, and the resulting moisture time series input data corresponding to the two analysis times had consistent data length and sampling time arrangement, thereby avoiding misinterpreting differences in measurement point location, roof structure, and sampling grid as differences in the preceding moisture process. Upper layer moisture content, lower layer moisture content, air temperature, relative humidity, wind speed, and shortwave irradiance of the green surface were all collected at the same sampling interval and written into the sampling time, providing a data basis for the aforementioned comparability control.For two analysis times that meet the above comparability conditions, calculate the absolute value of the current upper layer moisture content difference, the absolute value of the current lower layer moisture content difference, and the absolute value of the difference between the average values ​​of each environmental parameter. The average values ​​of each environmental parameter are calculated based on the air temperature, relative humidity, wind speed, and shortwave irradiance of the green surface within the corresponding historical time window. For the error range of the average environmental parameter values, sum the upper limits of the absolute errors of each environmental parameter measurement involved in the average value calculation within the historical time window, divide by the corresponding number of samplings, and use this as the upper limit of the error for the corresponding environmental parameter average value. When the same upper limit of absolute error is used for the same environmental parameter within the historical time window, the upper limit of the error for the corresponding environmental parameter average value is consistent with the upper limit of the absolute error for a single environmental parameter measurement. Therefore, the sum of the upper limits of the errors for the average environmental parameters at the two analysis times is used as the comparison range for the absolute values ​​of the difference between the average values ​​of each environmental parameter. The absolute values ​​of the current upper layer moisture content difference and the current lower layer moisture content difference are respectively not greater than the sum of the upper limits of the absolute errors of the two corresponding moisture content measurements, and the absolute values ​​of the difference between the average values ​​of each environmental parameter are respectively not greater than the sum of the upper limits of the errors of the two corresponding environmental parameter average values. The system assumes that the current water content and environmental conditions at two analysis times are within a comparable range of measurement error. Based on this, a training sample pair is formed if any of the following conditions exist between the two analysis times: the upper layer water increase detection flag has different values; the upper layer water increase detection flag is always one and the lower layer response detection flag has different values; both the upper layer water increase detection flag and the lower layer response detection flag are one, and the absolute value of the difference between the input terms of the first detection interval between the upper and lower layers is greater than or equal to one sampling interval; both upper layer water increase detection flags are zero, and the absolute value of the difference in the historical net change in the upper layer is greater than the sum of the upper limit of the absolute error of the four upper layer water content measurements at the beginning and end of the historical time window at both analysis times; both upper layer water increase detection flags are zero, and the absolute value of the difference in the historical net change in the lower layer is greater than the sum of the upper limit of the absolute error of the four lower layer water content measurements at the beginning and end of the historical time window at both analysis times. This ensures that the training sample pair can retain the differences in the characteristics of the preceding water process under similar current water content and environmental conditions, providing a comparable sample basis for subsequent use of the training samples to train the water history and thermal response mapping neural network.

[0026] In this implementation scheme, by limiting the measurement point conditions, sampling intervals and historical time window lengths to consistent at two analysis times, and by using the absolute value of the current upper layer water content difference, the absolute value of the current lower layer water content difference and the absolute value of the average difference of various environmental parameters to screen out incomparable samples, and then forming training sample pairs based on the differences in the characteristics of preceding water processes, the interference of spatial location, sampling conditions and environmental fluctuations on sample differences is reduced, thereby providing comparable training data for the water history and thermal response mapping neural network to learn the thermal response differences corresponding to different preceding water processes.

[0027] Specifically, inputting the time-series water content data, preceding water process characteristics, and average environmental parameters into the fully connected layer includes: arranging the time-series water content data according to sampling times to form an equal-length time-series input matrix; specifically, using the same historical time window length and sampling interval during the training and application of the same water history and thermal response mapping neural network, ensuring that the time-series water content data corresponding to each analysis time contains the same number of sampling times, and constructing an equal-length time-series input matrix according to a fixed channel order: upper layer water content sequence, lower layer water content sequence, upper layer change sequence, lower layer change sequence, and upper and lower layer water content difference sequence. The data within each channel are arranged according to sampling times. The sampling times are arranged from earliest to latest, ensuring a one-to-one correspondence between different training samples in terms of channel position and sampling time position. When missing markers exist within the historical time window, the missing positions are not filled with measurements from adjacent sampling times, nor are the historical time window truncated or altered. The corresponding analysis time is excluded from the construction of the equal-length time-series input matrix to avoid artificially altering the original relationship between the upper-layer change sequence, the lower-layer change sequence, and the water content difference sequence between the upper and lower layers. The equal-length time-series input matrix is ​​input into a one-dimensional time-series convolution. The one-dimensional time-series convolution slides along the sampling time direction and extracts the water content change relationship between adjacent sampling times on the five input channels. This embodiment uses a single... One-dimensional temporal convolution is performed, with a stride of one sampling position. The convolution boundaries use zero-padding to maintain a consistent number of sampling positions before and after convolution. The convolution output undergoes modified linear unit activation processing. The number of convolution output channels is read from the network structure configuration of the moisture history and thermal response mapping neural network, determined before training begins and remaining constant throughout a complete training cycle, serving as the dimension of the subsequent moisture temporal feature vector. After one-dimensional temporal convolution, an arithmetic mean is calculated along the sampling time direction, i.e., the arithmetic mean of the output values ​​of each convolution output channel at each sampling position within the historical time window is calculated. Each convolution output channel forms a component in the moisture temporal feature vector. The system obtains a water temporal feature vector, thus forming a fixed-dimensional temporal representation while retaining water change information within historical time windows. The water temporal feature vector, previous water process features, and average environmental parameters are concatenated in a fixed field order and input into a fully connected layer. The previous water process features and average environmental parameters are not entered into the one-dimensional temporal convolution, but are used together with the water temporal feature vector as input to the fully connected layer after the water temporal feature vector is formed. The fully connected layer maps the normalized values ​​of canopy temperature, matrix midpoint temperature, and roof heat flux to ensure that a consistent network data flow is used during the training phase and at new analysis times.

[0028] In this implementation scheme, by constructing an equal-length time-series input matrix based on the upper layer moisture content sequence, the lower layer moisture content sequence, the upper layer change sequence, the lower layer change sequence, and the upper and lower layer moisture content difference sequence at a unified sampling time, and using one-dimensional time-series convolution to extract the moisture time-series feature vector, it is possible to maintain the consistency of the input structure at different analysis times and characterize the moisture change relationship within the historical time window. Furthermore, by concatenating the moisture time-series feature vector, the previous moisture process features, and the average values ​​of environmental parameters and inputting them into the fully connected layer, it is possible to provide a unified input basis with time-series information for the calculation of the canopy temperature normalization calculation value, the matrix midpoint temperature normalization calculation value, and the roof heat flux normalization calculation value.

[0029] Specifically, the moisture history and thermal response mapping neural network trained based on the single-sample average error and sample-pair average error includes: calculating the mean and standard deviation of canopy temperature, matrix midpoint temperature, and roof heat flow in the training samples. The training samples consist of each analysis time point within the training sample pairs. The mean and standard deviation are calculated only from the training samples used in the network training and are fixed and saved after the moisture history and thermal response mapping neural network completes training. They are reused when calculating at new analysis times without recalculating using data from new analysis times, thus maintaining consistency in the standardization criteria between the training and calculation phases. For any of the canopy temperature, matrix midpoint temperature, and roof heat flow, when the calculated standard deviation is zero... During the stabilization and standardization process, the stable standard deviation is replaced with a stable value of the standard deviation that is greater than zero. The stable standard deviation remains unchanged during the same network training and subsequent new analysis time calculations. The standardized measured values ​​of canopy temperature, matrix midpoint temperature, and roof heat flow are obtained by subtracting their respective means from the measured values ​​of canopy temperature, matrix midpoint temperature, and roof heat flow, and then dividing by their respective standard deviations. The moisture history and thermal response mapping neural network consists of one-dimensional temporal convolution and fully connected layers. It takes the upper layer moisture content sequence, the lower layer moisture content sequence, the upper layer change sequence, the lower layer change sequence, and the upper and lower layer moisture content difference sequence as five input channels to form an equal-length temporal input matrix. The one-dimensional temporal convolution is performed along the sampling time. Moisture temporal features are extracted along the direction of sampling time, and the arithmetic mean is calculated to obtain a moisture temporal feature vector. This feature vector is then concatenated with previous moisture process features and average environmental parameters before being input into a fully connected layer. The outputs are standardized calculated values ​​for canopy temperature, matrix midpoint temperature, and roof heat flux, establishing a correspondence between the network input and the historical moisture changes, environmental conditions, and thermal response output of the green roof system. The squared differences between the standardized calculated values ​​and the standardized measured values ​​of canopy temperature, matrix midpoint temperature, and roof heat flux are calculated separately. The sum of these three squared differences is used to calculate the arithmetic mean over all training samples, yielding the single-sample average error, which is used to constrain the thermal response calculations for each training sample. The deviation between the results and the measured results: For the training sample pairs, the standardized calculated difference and standardized measured difference of the canopy temperature, matrix midpoint temperature and roof heat flow are obtained in the order of subtracting the early sample from the late sample. The standardized calculated difference is obtained by subtracting the corresponding standardized calculated values ​​of the two analysis times in the training sample pair, and the standardized measured difference is obtained by subtracting the corresponding standardized measured values. The square difference between the corresponding standardized calculated difference and the standardized measured difference is calculated. The sum of the three square differences is used to calculate the arithmetic mean of all training sample pairs to obtain the average error of the sample pair. This error is used to constrain the difference in thermal response between two analysis times with similar current water content and environmental conditions but different characteristics of the preceding water process.The training loss is formed by weighted summation of the single-sample average error and the sample-pair average error. To avoid introducing additional uncertainty due to weight changes during training, this embodiment uses fixed positive weights for both error terms, which remain unchanged throughout a complete training cycle. The same weights can be used to ensure that both single-sample thermal response fitting and training sample-pair difference fitting participate in network parameter updates. The gradient of the training loss with respect to the network parameters is calculated using the backpropagation algorithm, and the one-dimensional temporal convolutional parameters and fully connected layer parameters are updated based on the gradient. This process of forward computation, training loss calculation, gradient calculation, and network parameter updates is repeated. The training termination condition is determined by mapping the moisture history and thermal response neural network. The training configuration is read from the network, and the training termination conditions include reaching the upper limit of the training epochs or the training loss meeting the loss convergence condition, which are determined before the start of a complete training cycle. Forward computation, training loss calculation, gradient calculation, and network parameter updates are repeatedly executed. When the training termination conditions are met, network parameter updates are stopped, and a moisture history and thermal response mapping neural network is constructed. After training, the one-dimensional temporal convolution parameters, fully connected layer parameters, and the mean and standard deviation of canopy temperature, matrix midpoint temperature, and roof heat flow in the training samples are fixed for subsequent calculations of canopy temperature, matrix midpoint temperature, and roof heat flow at new analysis times.

[0030] In this implementation scheme, the standardization of canopy temperature, matrix midpoint temperature, and roof heat flow is unified by using the mean and standard deviation of training samples. The training loss is formed by combining the single-sample average error and the sample-pair average error. This enables the moisture history and thermal response mapping neural network to both constrain the thermal response calculation deviation of individual training samples and learn the thermal response differences corresponding to the differences in the characteristics of preceding moisture processes. This improves the stability of the training of one-dimensional temporal convolution parameters and fully connected layer parameters, providing a model basis for obtaining the calculated values ​​of canopy temperature, matrix midpoint temperature, and roof heat flow.

[0031] Specifically, for the new analysis time, moisture time-series input data is constructed, the characteristics of the preceding moisture process and the average values ​​of environmental parameters are determined, and these are input into the moisture history and thermal response mapping neural network to obtain the calculated values ​​of canopy temperature, matrix midpoint temperature, and roof heat flow. This includes: constructing moisture time-series input data for the new analysis time, and then calculating the characteristics of the preceding moisture process and the average values ​​of environmental parameters; specifically, using the new analysis time as the end of the historical time window, the corresponding upper and lower layer moisture contents are read according to the historical time window length and sampling interval used in the training phase, and processed according to the same method as in the training phase. The formula generates upper layer moisture content sequence, lower layer moisture content sequence, upper layer change sequence, lower layer change sequence, and upper and lower layer moisture content difference sequence, which serve as the time series input data for moisture. Simultaneously, based on the input items of upper layer historical net change, lower layer historical net change, upper layer moisture increase detection indicator, lower layer moisture increase detection indicator, lower layer response detection indicator, and the first detection interval between upper and lower layers, the characteristics of the preceding moisture process are determined. Furthermore, the average values ​​of environmental parameters are calculated based on air temperature, relative humidity, wind speed, and shortwave irradiance of green surfaces within the same historical time window, ensuring that the data fields, arrangement order, and calculation methods used at the new analysis time are consistent with... The training phase maintains consistency. Moisture time-series input data, preceding moisture process features, and average environmental parameters are input into the moisture history and thermal response mapping neural network. The moisture time-series input data is arranged according to sampling time to form an equal-length time-series input matrix. After one-dimensional time-series convolution, the arithmetic mean is calculated along the sampling time direction to obtain the moisture time-series feature vector. Then, the moisture time-series feature vector, preceding moisture process features, and average environmental parameters are concatenated according to the same field order as in the training phase and input into the fully connected layer to obtain the standardized calculated values ​​of canopy temperature, matrix midpoint temperature, and roof heat flux. This ensures that the network data stream used at the new analysis time is consistent with that used during the training phase. The standardized calculated values ​​of canopy temperature, matrix midpoint temperature, and roof heat flow are multiplied by the standard deviation of the corresponding parameters in the training sample, and then the mean of the corresponding parameters in the training sample is added. The mean and standard deviation are the corresponding statistics used when calculating the standardized measured values ​​of canopy temperature, matrix midpoint temperature, and roof heat flow during the training phase, and are kept at the same values ​​when the moisture history and thermal response mapping neural network is used for calculation at the new analysis time, thus obtaining the calculated values ​​of canopy temperature, matrix midpoint temperature, and roof heat flow.

[0032] In this implementation plan, by constructing moisture time-series input data for new analysis moments using the same processing method as the training phase, determining the characteristics of preceding moisture processes and the average values ​​of environmental parameters, and inputting them into the moisture history and thermal response mapping neural network, it is possible to ensure that the data fields, arrangement order, and network data flow are consistent between the new analysis moments and the training phase. Furthermore, by utilizing the mean and standard deviation of the corresponding parameters in the training samples to recover the calculated values ​​of canopy temperature, matrix midpoint temperature, and roof heat flow, it is possible to provide a unified set of calculation results for subsequent thermal response difference analysis.

[0033] Specifically, thermal response difference analysis is performed using the current upper layer moisture content, current lower layer moisture content, average environmental parameters, calculated canopy temperature, calculated matrix midpoint temperature, and calculated roof heat flow. The results of the thermal response difference analysis are then recorded in relation to historical moisture process differences. This includes selecting two analysis times from the data to be analyzed. In this embodiment, the data to be analyzed consists of data generated by the light-green composite roof at multiple analysis times. Each analysis time corresponds to a sampling time and includes the current upper layer moisture content, current lower layer moisture content, characteristics of previous moisture processes, average environmental parameters, and calculated canopy temperature, calculated matrix midpoint temperature, and calculated matrix midpoint temperature obtained using a moisture history and thermal response mapping neural network. For calculating roof heat flux, the time range of the data to be analyzed is limited to the sampling period where complete historical time window data can be obtained. The spatial range uses the same measuring point or measuring points with the same roof structure and sampling conditions to ensure the comparability of data at different analysis times. After completing the construction of moisture time series input data, determination of previous moisture process characteristics, calculation of average environmental parameters, and thermal response calculation in the aforementioned manner, the new analysis time can also be written as an analysis time into the data to be analyzed. The new analysis time reads the upper layer moisture content, lower layer moisture content, air temperature, relative humidity, wind speed, and shortwave irradiance of the green surface within the corresponding historical time window, and further obtains the corresponding previous moisture process characteristics and thermal response calculation results.For two selected analysis times, if the absolute values ​​of the current upper layer moisture content difference and the current lower layer moisture content difference are not greater than the sum of the upper limits of the absolute errors of the two corresponding moisture content measurements, and the absolute values ​​of the average differences of each environmental parameter are not greater than the sum of the upper limits of the absolute errors of the two corresponding environmental parameter measurements, then the two analysis times are retained. If the values ​​of the upper layer moisture increase detection flags are different at the two analysis times, then the difference in upper layer moisture increase detection is recorded. If the upper layer moisture increase detection flags are both one and the values ​​of the lower layer response detection flags are different, then the difference in lower layer response detection is recorded. If both the upper layer moisture increase detection flags and the lower layer response detection flags are one, and the absolute value of the difference between the input items for the first detection interval of the upper and lower layers is greater than or equal to one sampling interval, then the difference in the first detection interval of the upper and lower layers is recorded. If the upper layer moisture increase detection flags are both zero and the absolute value of the difference in the historical net change of the upper layer is greater than the sum of the upper layer moisture content absolute errors of the four measurements at the beginning and end of the historical time window of the two analysis times, then the difference in the historical net change of the upper layer is recorded. The net historical change difference of the bottom layer is recorded when the sum of the upper limit of the absolute error of the four bottom layer moisture content measurements at the beginning and end of the time window is reached. The above five types of differences are regarded as historical moisture process differences. The aforementioned differences are used to characterize the observational differences of the moisture change process before the two analysis times, and are not directly used to identify the specific water replenishment source or moisture migration mechanism. Thermal response difference analysis is performed on the two retained analysis times. The early time and late time are determined according to the sampling time sequence. The difference between the calculated values ​​of the canopy temperature, the calculated values ​​of the matrix midpoint temperature, and the calculated values ​​of the roof heat flow between the late time and the early time are calculated to obtain the canopy temperature difference, matrix midpoint temperature difference, and roof heat flow difference. The combination of the sampling times of the two analysis times is used as the correlation basis. The canopy temperature difference, matrix midpoint temperature difference, roof heat flow difference, and the corresponding historical moisture process differences are written into the same structured record. This allows each group of thermal response differences to be traced back to the corresponding historical moisture process differences through the sampling times of the two analysis times, thereby forming a queryable correspondence. This provides a data basis for subsequent analysis of the thermal response differences of the light and green combination roof corresponding to different preceding moisture processes.

[0034] In this implementation plan, two analysis points are selected where the current upper layer moisture content, the current lower layer moisture content, and the average values ​​of various environmental parameters are within a comparable range. Differences in the detection of increased upper layer moisture, differences in the detection of lower layer response, differences in the first detection interval between upper and lower layers, differences in the historical net change of upper layer, and differences in the historical net change of lower layer are identified. These differences are then combined with the corresponding records of the canopy temperature difference, the midpoint temperature difference of the matrix, and the roof heat flow difference. This establishes a traceable correspondence between historical moisture process differences and thermal response differences, providing data support for analyzing the temperature regulation response of the light and green roof combination under different preceding moisture processes.

[0035] Specifically, this embodiment provides a neural network-based system for analyzing the temperature regulation efficiency of light-green roof composites, applied to a neural network-based method for analyzing the temperature regulation efficiency of light-green roof composites. The system includes: a data and historical feature acquisition unit, used to collect measured data from the light-green roof composites; constructing moisture time-series input data based on the upper and lower layer moisture content sequences; determining the characteristics of the preceding moisture process; and calculating the average environmental parameters. Specifically, upper and lower layer moisture content sequences are formed according to the sampling time, and further, upper layer change sequence, lower layer change sequence, and upper and lower layer moisture content difference sequence are formed. The characteristics of the preceding moisture process are determined based on the input items: upper layer historical net change, lower layer historical net change, upper layer moisture increase detection indicator, lower layer moisture increase detection indicator, lower layer response detection indicator, and upper and lower layer first detection interval. The sample construction and network training unit selects two analysis timestamps where the differences between the current upper layer moisture content, the current lower layer moisture content, and the average environmental parameters are within the corresponding measurement error range, and where the characteristics of the preceding moisture process differ. These two timestamps form training sample pairs. The moisture time-series input data is processed through one-dimensional temporal convolution to obtain a moisture time-series feature vector, which is then concatenated with the preceding moisture process features and the average environmental parameters before being input into a fully connected layer. The moisture history and thermal response mapping neural network is trained based on the single-sample average error and the sample pair average error. The roof thermal response calculation unit constructs moisture time-series input data for a new analysis time, determines the preceding moisture process characteristics and the average environmental parameters, and inputs them into the moisture history and thermal response mapping neural network to obtain the calculated values ​​of canopy temperature, matrix midpoint temperature, and roof heat flux. The temperature regulation difference analysis unit selects two analysis timestamps where the current upper layer moisture content, the current lower layer moisture content, and the average environmental parameters are comparable, determines the historical moisture process differences, calculates the canopy temperature difference, matrix midpoint temperature difference, and roof heat flux difference, and records the thermal response difference analysis results in correspondence with the historical moisture process differences.

[0036] In this implementation scheme, the data and historical feature acquisition unit, sample construction and network training unit, roof thermal response calculation unit and temperature regulation difference analysis unit work together to process the moisture time series input data, previous moisture process characteristics and average environmental parameters. This enables the moisture history and thermal response mapping neural network to complete the training and thermal response calculation at the new analysis time. It also records the canopy temperature difference, matrix midpoint temperature difference and roof heat flow difference with the differences in historical moisture processes, thereby realizing the integrated analysis of moisture history and thermal response differences of the light and green roof combination.

[0037] The terms "first," "second," etc., in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the term "comprising," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or modules, but may optionally include steps or modules not listed, or may optionally include other step units inherent to these processes, methods, apparatuses, products, or devices.

[0038] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0039] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0040] The methods and related apparatuses provided in this application are described with reference to the method flowcharts and / or structural diagrams provided in this application. Specifically, each block of the method flowchart and / or structural diagram, as well as combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to create a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A schematic diagram of one or more processes and / or structures. Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1A schematic diagram of one or more processes and / or structures. Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 A process or multiple processes and / or structures illustrate the steps of the functions specified in one or more boxes.

[0041] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A method for analyzing the temperature regulation efficiency of light-green roof combinations based on neural networks, characterized in that, include: S1. Collect measured data of the light and green roof combination, construct moisture time series input data based on the upper layer moisture content sequence and the lower layer moisture content sequence, determine the characteristics of the preceding moisture process and calculate the average value of environmental parameters; S2, select two analysis times where the differences between the current upper layer water content, the current lower layer water content and the average value of environmental parameters are within the corresponding measurement error range and the characteristics of the preceding water process are different to form a training sample pair. Input the water time series input data, the characteristics of the preceding water process and the average value of environmental parameters into the fully connected layer, and train the water history and thermal response mapping neural network based on the single sample average error and the sample pair average error. S3: Construct moisture time series input data for the new analysis time, determine the characteristics of the previous moisture process and the average value of environmental parameters, and input them into the moisture history and thermal response mapping neural network to obtain the calculated values ​​of canopy temperature, matrix midpoint temperature and roof heat flow. S4. Perform thermal response difference analysis using the current upper layer moisture content, current lower layer moisture content, average environmental parameters, calculated canopy temperature, calculated matrix midpoint temperature, and calculated roof heat flow, and record the thermal response difference analysis results in correspondence with historical moisture process differences.

2. The method for analyzing the temperature regulation efficiency of a light-green roof based on a neural network according to claim 1, characterized in that, The measured data collected from the green roof combination include: For the area below the photovoltaic panels and adjacent green areas, the moisture content of the upper layer, the moisture content of the lower layer, the canopy temperature, the midpoint temperature of the matrix, the roof heat flow, the air temperature, the relative humidity, the wind speed, and the shortwave irradiance of the green surface were collected at the same sampling interval and recorded at the sampling time. The moisture content of the upper layer and the moisture content of the lower layer were arranged according to the sampling time to form the upper layer moisture content sequence and the lower layer moisture content sequence, and the measured values ​​of the canopy temperature, the midpoint temperature of the matrix, and the roof heat flow were recorded.

3. The method for analyzing the temperature regulation efficiency of a light-green roof combination based on a neural network according to claim 2, characterized in that, The construction of moisture time-series input data based on the upper and lower layer moisture content sequences includes: Using the analysis time as the end of the historical time window, the upper and lower layer water contents of n sampling times are read backwards. The earliest sampling time is determined as the start of the historical time window. The upper and lower layer water contents at the end of the historical time window are determined as the current upper and lower layer water contents, respectively. The time span from the start of the historical time window to the analysis time is determined as the length of the historical time window. The upper layer water contents minus the lower layer water contents at the same sampling time form a sequence of water content difference between the upper and lower layers. The water content difference at the same depth at two adjacent sampling times is calculated to form a sequence of upper layer change and a sequence of lower layer change. Zero values ​​are written at the beginning of the upper layer change and lower layer change sequences. The net historical change in the upper layer and the net historical change in the lower layer are obtained by subtracting the water content at the depth corresponding to the start of the historical time window from the water content at the end of the historical time window. The upper layer water content sequence, the lower layer water content sequence, the upper layer change sequence, the lower layer change sequence, and the sequence of water content difference between the upper and lower layers are used as moisture time series input data.

4. The method for analyzing the temperature regulation efficiency of a light-green roof combination based on a neural network according to claim 3, characterized in that, The process of determining the characteristics of preceding moisture processes and calculating the average values ​​of environmental parameters includes: When a positive value exists in the upper layer change sequence that is greater than the sum of the upper layer moisture content measurement absolute errors of the two preceding and following upper layer change measurements, the sampling time corresponding to the earliest occurrence of the upper layer change that meets the condition is determined as the first detection time of the upper layer, and the upper layer moisture increase detection flag is written as 1; otherwise, it is written as 0. When a positive value exists in the lower layer change sequence that is greater than the sum of the upper layer moisture content measurement absolute errors of the two preceding and following lower layer change measurements, the sampling time corresponding to the earliest occurrence of the lower layer change that meets the condition is determined as the first detection time of the lower layer, and the lower layer moisture increase detection flag is written as 1; otherwise, it is written as 0. When both the upper and lower layer moisture increase detection flags are one, and the first detection time of the lower layer is not earlier than the first detection time of the upper layer, the lower layer response detection flag is written as one, and the difference between the first detection time of the lower layer and the first detection time of the upper layer is written into the upper and lower layer first detection interval input field. When the upper layer moisture increase detection flag is zero, the lower layer moisture increase detection flag is zero, or the lower layer first detection time is earlier than the upper layer first detection time, the lower layer response detection flag is written as zero, and the sum of the historical time window length and a sampling interval is written into the upper and lower layer first detection interval input field. The upper layer historical net change, the lower layer historical net change, the upper layer moisture increase detection flag, the lower layer moisture increase detection flag, the lower layer response detection flag, and the upper and lower layer first detection interval input field are used as the characteristics of the preceding moisture process. The average air temperature, the average relative humidity, the average wind speed, and the average shortwave irradiance of the green surface are calculated as the average environmental parameters.

5. The method for analyzing the temperature regulation efficiency of a light-green roof combination based on a neural network according to claim 4, characterized in that, The training sample pair consists of two analysis times where the differences between the current upper layer water content, the current lower layer water content, and the average environmental parameters are within the corresponding measurement error range, and where the characteristics of the preceding water process differ. Two analysis times are selected from different analysis times. The absolute values ​​of the current upper layer moisture content difference, the current lower layer moisture content difference, and the absolute values ​​of the average differences of each environmental parameter are calculated. A training sample pair is formed when the absolute values ​​of the current upper layer moisture content difference and the current lower layer moisture content difference are not greater than the sum of the upper limits of the absolute errors of the two corresponding moisture content measurements, and the absolute values ​​of the average differences of each environmental parameter are not greater than the sum of the upper limits of the absolute errors of the two corresponding environmental parameter measurements. If either of the following conditions exists between the two analysis times, the values ​​of the upper layer moisture increase detection markers are different; or the upper layer moisture increase detection markers are both one and the lower layer moisture content difference is the same. The values ​​of the layer response detection flags are different; the upper layer moisture increase detection flag and the lower layer response detection flag are both one, and the absolute value of the difference between the input items of the first detection interval between the upper and lower layers is greater than or equal to one sampling interval; the upper layer moisture increase detection flags are all zero, and the absolute value of the difference in the historical net change of the upper layer is greater than the sum of the upper limit of the absolute error of the four upper layer moisture content measurements at the beginning and end of the historical time window of the two analysis time points; the upper layer moisture increase detection flags are all zero, and the absolute value of the difference in the historical net change of the lower layer is greater than the sum of the upper limit of the absolute error of the four lower layer moisture content measurements at the beginning and end of the historical time window of the two analysis time points.

6. The method for analyzing the temperature regulation efficiency of a light-green roof combination based on a neural network according to claim 5, characterized in that, The step of inputting time-series water input data, characteristics of preceding water processes, and average environmental parameters into the fully connected layer includes: The moisture time-series input data is arranged according to the sampling time to form an equal-length time-series input matrix. After one-dimensional time-series convolution, the arithmetic mean is calculated along the sampling time direction to obtain the moisture time-series feature vector. The moisture time-series feature vector, the previous moisture process features and the average environmental parameters are concatenated and input into the fully connected layer to obtain the canopy temperature normalized calculation value, the matrix midpoint temperature normalized calculation value and the roof heat flux normalized calculation value.

7. The method for analyzing the temperature regulation efficiency of a light-green roof combination based on a neural network according to claim 6, characterized in that, The neural network for mapping moisture history and thermal response, trained based on the average error of a single sample and the average error of a sample pair, includes: The mean and standard deviation of canopy temperature, midpoint matrix temperature, and roof heat flux in the training samples were calculated. The measured values ​​were then standardized based on their respective mean and standard deviation to obtain standardized measured values ​​of canopy temperature, midpoint matrix temperature, and roof heat flux. The squared differences between the standardized calculated values ​​and the standardized measured values ​​of canopy temperature, midpoint matrix temperature, and roof heat flux were calculated separately. The sum of the three squared differences was then used to calculate the arithmetic mean of all training samples to obtain the single-sample mean error. For each training sample pair, the standardized calculated difference and standardized measured difference of canopy temperature, matrix midpoint temperature, and roof heat flow are obtained by subtracting the early sample from the late sample. The squared difference between the corresponding standardized calculated difference and standardized measured difference is calculated. The sum of the three squared differences is used to calculate the arithmetic mean of all training sample pairs to obtain the average error of the sample pair. The average error of the single sample and the average error of the sample pair are weighted and summed to form the training loss. The gradient of the training loss with respect to the network parameters is calculated by the backpropagation algorithm, and the one-dimensional temporal convolution parameters and fully connected layer parameters are updated to construct a neural network for mapping moisture history and thermal response.

8. The method for analyzing the temperature regulation efficiency of a light-green roof combination based on a neural network according to claim 7, characterized in that, The process involves constructing time-series input data for the new analysis time, determining the characteristics of the preceding moisture process and the average values ​​of environmental parameters, and inputting these into a moisture history and thermal response mapping neural network to obtain calculated values ​​for canopy temperature, matrix midpoint temperature, and roof heat flow, including: For the new analysis time, moisture time-series input data is constructed, and the characteristics of the preceding moisture process and the average values ​​of environmental parameters are calculated. The moisture time-series input data, the characteristics of the preceding moisture process, and the average values ​​of environmental parameters are input into the moisture history and thermal response mapping neural network to obtain the standardized calculated values ​​of canopy temperature, matrix midpoint temperature, and roof heat flow. The standardized calculated values ​​of canopy temperature, matrix midpoint temperature, and roof heat flow are multiplied by the standard deviation of the corresponding parameters in the calculation training sample, and then the mean of the corresponding parameters is added to obtain the calculated values ​​of canopy temperature, matrix midpoint temperature, and roof heat flow.

9. The method for analyzing the temperature regulation efficiency of a light-green roof based on a neural network according to claim 8, characterized in that, The thermal response difference analysis is performed using the current upper layer moisture content, the current lower layer moisture content, the average environmental parameters, the calculated canopy temperature, the calculated midpoint temperature of the matrix, and the calculated roof heat flow. The results of the thermal response difference analysis are then recorded in correspondence with historical moisture process differences. Two analysis times are selected from the data to be analyzed. When the absolute values ​​of the current upper layer moisture content difference and the current lower layer moisture content difference are not greater than the sum of the upper limits of the absolute errors of the two corresponding moisture content measurements, and the absolute values ​​of the average differences of each environmental parameter are not greater than the sum of the upper limits of the absolute errors of the two corresponding environmental parameter measurements, the two analysis times are retained. When the values ​​of the upper layer moisture increase detection indicator differ at two analysis times, record the difference in upper layer moisture increase detection; when both upper layer moisture increase detection indicators are one and the values ​​of the lower layer response detection indicator differ, record the difference in lower layer response detection; when both the upper layer moisture increase detection indicator and the lower layer response detection indicator are one, and the absolute value of the difference between the input items for the first detection interval of the upper and lower layers is greater than or equal to one sampling interval, record the difference in the first detection interval of the upper and lower layers; when both upper layer moisture increase detection indicators are zero and the absolute value of the difference in the historical net change of the upper layer is greater than the sum of the upper limit of the absolute error of the four upper layer moisture content measurements at the beginning and end of the historical time window at two analysis times, record the difference in the historical net change of the upper layer; when both upper layer moisture increase detection indicators are zero and the absolute value of the difference in the historical net change of the lower layer is greater than the sum of the upper limit of the absolute error of the four lower layer moisture content measurements at the beginning and end of the historical time window at two analysis times, record the difference in the historical net change of the lower layer; these five types of differences are considered as historical moisture process differences. Thermal response difference analysis was performed on the two retained analysis times: the early and late times were determined according to the sampling time sequence, and the differences between the calculated values ​​of canopy temperature, matrix midpoint temperature and roof heat flow were calculated between the late and early times to obtain the canopy temperature difference, matrix midpoint temperature difference and roof heat flow difference. The canopy temperature difference, matrix midpoint temperature difference and roof heat flow difference were then recorded in correspondence with the historical moisture process differences according to the sampling times of the two analysis times.

10. A neural network-based system for analyzing the temperature regulation efficiency of light-green roof combinations, characterized in that, include: The data and historical feature acquisition unit is used to collect measured data of the light and green roof combination, construct moisture time series input data based on the upper layer moisture content sequence and the lower layer moisture content sequence, determine the characteristics of the preceding moisture process, and calculate the average value of environmental parameters. The sample construction and network training unit is used to select two analysis times that are within the corresponding measurement error range and have differences in the current upper layer water content, current lower layer water content and average environmental parameters, and have differences in the characteristics of the preceding water process to form a training sample pair. The water time series input data, the characteristics of the preceding water process and the average environmental parameters are input into the fully connected layer, and the water history and thermal response mapping neural network is trained based on the single sample average error and the sample pair average error. The roof thermal response calculation unit is used to construct moisture time series input data for a new analysis time, determine the characteristics of the previous moisture process and the average value of environmental parameters, and input them into the moisture history and thermal response mapping neural network to obtain the calculated values ​​of canopy temperature, matrix midpoint temperature and roof heat flow. The temperature difference analysis unit is used to perform thermal response difference analysis using the current upper layer moisture content, current lower layer moisture content, average environmental parameters, calculated canopy temperature, calculated matrix midpoint temperature, and calculated roof heat flow, and records the thermal response difference analysis results in correspondence with historical moisture process differences.

Citation Information

Patent Citations

  • A Building Energy Consumption Prediction Method Based on Recurrent Neural Networks and Multi-Task Learning Models

    CN109685252B

  • A method and system for predicting energy consumption in public buildings based on GA-ANN

    CN110046743B