Indoor thermal environment parameter reconstruction method and system based on multi-task neural network

Abstract

This invention belongs to the field of thermal environment parameter reconstruction technology, and particularly to a method and system for reconstructing indoor thermal environment parameters based on a multi-task neural network. It includes the following steps: S1: Establishing a numerical simulation model of the indoor thermal environment of urban residential buildings in a frigid climate zone; S2: Constructing neural network training samples; S3: Generating region labels based on the sample data source; S4: Constructing an indoor thermal environment reconstruction neural network based on Fourier position encoding and a multi-task output head; S5: Training and validating the model using an X-plane partitioning method; S6: Using the reconstructed indoor global thermal environment parameters for thermal comfort state assessment and regulation suggestion generation. This invention, by constructing indoor thermal environment numerical simulation samples, generating gas and solid region labels, and utilizing Fourier position encoding to enhance spatial coordinate expression capabilities, achieves global reconstruction of indoor thermal environment parameters while reducing the number of actual sensors deployed.

Description

Technical Field

[0001] This invention relates to the field of thermal environment parameter reconstruction technology, specifically to a method and system for reconstructing indoor thermal environment parameters based on a multi-task neural network. Background Technology

[0002] With the increasing demand for heating in urban residential buildings in frigid regions, indoor thermal environment monitoring and intelligent control of thermal comfort have become important aspects of building energy conservation and residential comfort control. Existing residential indoor environment monitoring technologies typically collect local environmental parameters by installing temperature sensors, humidity sensors, air quality sensors, or integrated environmental detection devices at fixed locations indoors. The collected data is then transmitted to indoor environmental control equipment such as air conditioners, underfloor heating, radiators, and fresh air systems, serving as the basis for controlling equipment start-up, power adjustment, or operating mode switching. These technical solutions are relatively simple in structure, easy to install and use, and have already been applied to some extent in ordinary residential environments.

[0003] However, in frigid regions, indoor temperature differences are significant in winter. Influenced by factors such as heat transfer from the external envelope, cold radiation from doors and windows, the arrangement of heating terminals, indoor airflow organization, and human activity, the indoor thermal environment typically exhibits significant spatial non-uniformity and dynamic changes. Existing technologies often characterize the overall indoor environment using monitoring results from a single or a small number of fixed measuring points, making it difficult to accurately reflect the physical field distribution of the thermal environment in different areas and the actual thermal comfort state of humans. Increasing the number of sensors to improve monitoring accuracy leads to increased equipment costs, installation and maintenance difficulties, and data processing complexity. Furthermore, existing control methods often lack the ability to reconstruct and correct errors across the entire domain from sparse monitoring data, easily causing localized overheating, overcooling, or equipment lag, thereby affecting indoor comfort and increasing operating energy consumption. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0006] The indoor thermal environment parameter reconstruction method based on multi-task neural networks includes the following steps:

[0007] S1: Establish a numerical simulation model of the indoor thermal environment of urban residential buildings in a frigid climate zone to obtain the distribution of solid temperature, gas temperature and gas concentration at different spatial locations inside the building;

[0008] S2: Extract cross-sectional data from the numerical simulation results of the indoor thermal environment and construct training samples for the neural network;

[0009] S3: Generate region labels based on the source of sample data;

[0010] S4: Construct an indoor thermal environment reconstruction neural network based on Fourier position encoding and multi-task output head. Using indoor spatial coordinates as input, the network achieves joint prediction of regional attributes and thermal environment parameters of the point to be predicted through coordinate normalization, Fourier position encoding, shared feature extraction and multi-task output structure.

[0011] S5: Model training and validation are performed by dividing the model into X-sections. Some X-sections are randomly selected as the validation set, and the remaining X-sections are used as the training set. At the same time, the tasks of region recognition, solid temperature prediction, gas temperature prediction, gas concentration prediction and weak heat source recognition are optimized.

[0012] S6: During the actual operation phase, a small number of optimized sensor nodes will be used to collect local indoor temperature or concentration data. Combined with a trained neural network model, the thermal environment parameters at locations where no sensors are deployed will be reconstructed. The reconstructed indoor global thermal environment parameters will be used for thermal comfort state assessment and control recommendations.

[0013] As a preferred embodiment of the indoor thermal environment parameter reconstruction method based on multi-task neural networks described in this invention, the specific method of S1 is as follows: based on the spatial dimensions, enclosure structure form, window location, indoor obstacle distribution, heating terminal location, and local heat source information of the target residence, an indoor air region, a solid enclosure structure region, an indoor furnishing and obstacle region, a heating terminal region, and a local heat source region are constructed; the enclosure structure includes walls, floors, ceilings, and windows; indoor solid components include beds, sofas, furniture, and heating terminals; the thermal conductivity, density, and specific heat capacity of various materials are assigned values ​​according to the material parameters set in the indoor thermal environment numerical simulation model; the model boundary conditions include outdoor low temperature boundary, enclosure structure heat transfer boundary, indoor air and solid surface coupling heat transfer boundary, heating terminal heat source boundary, and gas concentration source term boundary; wherein, the heating terminal or local heat source is described using a constant temperature boundary, an equivalent heat flux boundary, or a volume heat source form; when the heat source is in the form of a volume heat source, its equivalent heating power is expressed as:

[0014] (1)

[0015] in, q represents the equivalent heating power of the heat source. v Let V be the volumetric heat source intensity; V be the volume of the heat source region; when the heat source adopts the form of surface heat flux, its equivalent heating power is expressed as:

[0016] (2)

[0017] Where, q s denoted as , where A is the surface heat flux and A is the surface area of ​​the heat source.

[0018] As a preferred embodiment of the indoor thermal environment parameter reconstruction method based on multi-task neural network described in this invention, the specific method for data extraction in S2 is as follows: the indoor space is cut along the X direction using a YZ plane, with the X value remaining constant in each cut, while the Y and Z values ​​change with the spatial position; solid temperature, gas temperature, and gas concentration data are extracted respectively, and a sample file containing X, Y, Z, and corresponding physical quantities is formed.

[0019] As a preferred embodiment of the indoor thermal environment parameter reconstruction method based on multi-task neural network described in this invention, the specific method of S3 is as follows: for coordinate points with solid temperature data, they are marked as solid regions; for coordinate points with gas temperature or gas concentration data, they are marked as gas regions; if the same coordinate appears in both solid and gas regions, it is identified as a region conflict point and is eliminated or processed separately according to preset rules.

[0020] As a preferred embodiment of the indoor thermal environment parameter reconstruction method based on a multi-task neural network described in this invention, the specific method of step S4 is as follows: Let the three-dimensional spatial coordinates of the point to be predicted be:

[0021] (3)

[0022] To eliminate the influence of different coordinate scales on the model training results, the input coordinates are first normalized:

[0023] (4)

[0024] In the formula, x is the original spatial coordinate. These are the normalized spatial coordinates. To train the mean of the spatial coordinates, To determine the standard deviation of spatial coordinates in the training set, Fourier position encoding is performed on the normalized coordinates. The input feature dimension after Fourier encoding is:

[0025] (5)

[0026] In the formula, B represents the number of Fourier frequencies, B=8, therefore the input dimension after model encoding is 51; the encoded spatial features are input into the shared feature extraction network to obtain the spatial shared features of the indoor thermal environment:

[0027] (6)

[0028] In the formula, This represents the spatial features after Fourier position encoding. Indicates a shared feature extraction network. Here, h represents the extracted shared features of the indoor thermal environment. This shared feature extraction network consists of multiple fully connected layers, normalization layers, SiLU activation functions, and Dropout layers. Based on the shared feature h, the model sets multiple task output heads. The region recognition output head is used to determine whether the input point belongs to a gas region or a solid region, and its output is expressed as follows:

[0029] (7)

[0030] In the formula, , This represents the probability that the input point belongs to the gas region. This represents the probability that the input point belongs to a solid region. The weight matrix for the output head of the region identification. The bias term for the region identification output head; the three regression output heads for solid temperature, gas temperature, and gas concentration are used to predict the corresponding thermal environment parameters, and their outputs are uniformly represented as:

[0031] (8)

[0032] In the formula, This represents the predicted value of the solid temperature. This represents the predicted gas temperature. This indicates the predicted gas concentration. First, the area identification output head determines the spatial attributes of the point to be predicted, and then selects the corresponding physical quantity as the effective output based on the area attributes.

[0033] As a preferred embodiment of the indoor thermal environment parameter reconstruction method based on multi-task neural networks described in this invention, the specific method of S5 is as follows: a masked mean square error loss is applied to the task, so that samples without corresponding physical quantities do not participate in the loss calculation. The masked loss is expressed as:

[0034] (9)

[0035] In the formula, For validity encoding, when the i-th sample has a corresponding physical quantity, When the i-th sample does not have a corresponding physical quantity, ; These are the model's predicted values. The values ​​are the true values; through this masking mechanism, solid region samples only participate in the calculation of solid temperature loss, and gas region samples only participate in the calculation of gas temperature and gas concentration loss; finally, the multi-task joint loss function is:

[0036] (10)

[0037] In the formula, For area identification loss, For solid temperature prediction loss, For predicting losses based on gas temperature, For predicting losses based on gas concentration, The coefficient 0.2 represents the weight of the weak heat source identification loss term.

[0038] After training, the system outputs the corresponding physical quantities based on the region recognition results. When the input point is identified as a solid region, the system only outputs the solid temperature, while the gas temperature and gas concentration are displayed as NaN. When the input point is identified as a gas region, the system outputs the gas temperature and gas concentration, while the solid temperature is displayed as NaN.

[0039] An indoor thermal environment parameter reconstruction system based on a multi-task neural network includes:

[0040] Indoor Space Basic Information Module: Used to acquire basic information on the geometric structure, enclosure structure, obstacles, heat sources and boundary conditions of residential indoor spaces, providing a spatial basis for sensor deployment and thermal environment reconstruction;

[0041] Optimize sensor deployment module: Based on the distribution characteristics of indoor thermal environment parameters, key sensor nodes are selected from candidate measurement points to reduce the number of sensors while ensuring the ability to characterize the overall indoor thermal environment.

[0042] Sparse temperature / concentration data acquisition module: Collects indoor local thermal environment data through temperature sensors and gas concentration sensors deployed in key locations, and transmits the data to the data processing terminal;

[0043] Data preprocessing and normalization module: performs unified processing on the collected data and the cross-sectional data exported from the indoor thermal environment numerical simulation model, including X / Y / Z header checking, duplicate data checking, coordinate normalization and unit unification, forming standardized samples that can be input into the neural network model;

[0044] The multi-task neural network reconstruction module takes spatial coordinates X, Y, Z and sparse observation data as input, and determines whether the input coordinates belong to a solid region or a gas region through region recognition and multi-task regression structure, and predicts the solid temperature, gas temperature and gas concentration at the corresponding location.

[0045] Thermal environment parameter output module: Outputs solid temperature, gas temperature, and gas concentration at different locations in the indoor space; when the coordinate is located in the solid region, the solid temperature, gas temperature, and gas concentration are output and displayed as NaN; when the coordinate is located in the gas region, the gas temperature and gas concentration are output, and the solid temperature is displayed as NaN.

[0046] Thermal comfort assessment and control suggestion module: Based on the reconstructed indoor thermal environment parameters, assess the thermal comfort status of the activity area and output control suggestions to indoor environmental control equipment such as heating, ventilation or air conditioning;

[0047] A virtual measurement point inference module for indoor thermal environment aimed at optimizing sensor deployment: This module is used to predict thermal environment parameters for candidate or random points in the indoor space based on a trained indoor thermal environment reconstruction model before actual sensor deployment, thereby assisting in determining the reasonable deployment location of sensors; without increasing the number of actual sensors, it can obtain the predicted results of solid temperature, gas temperature and gas concentration at different spatial points in advance, providing a basis for subsequent screening of key measurement points, reducing the number of sensors deployed and improving the ability to characterize the indoor thermal environment in its entirety.

[0048] As a preferred embodiment of the indoor thermal environment parameter reconstruction system based on a multi-task neural network according to the present invention, the indoor thermal environment virtual measurement point inference module includes:

[0049] Model weight loading unit: Used to load the weights of the trained indoor thermal environment reconstruction model, enabling the system to call the trained model to quickly predict virtual locations;

[0050] Virtual point input unit: used to input or generate the X, Y, Z coordinates of indoor spatial points to be predicted;

[0051] Model inference calculation unit: used to normalize the coordinates of the input points and call the multi-task neural network model to complete the inference of thermal environment parameters, and obtain the regional attributes and corresponding physical quantity prediction results at the point.

[0052] Region attribute discrimination unit: used to determine whether the input point belongs to a gas region or a solid region; when the model determines that the point belongs to a gas region, the point is used as a candidate placement location for gas temperature or gas concentration sensors; when the model determines that the point belongs to a solid region, the point is used as a candidate location for solid surface temperature monitoring or thermal state analysis near the heat source.

[0053] Thermal environment parameter output unit: used to output the solid temperature, gas temperature and gas concentration at the virtual point; when the point is located in the solid region, the system outputs the solid temperature, and the gas temperature and gas concentration are displayed as NaN; when the point is located in the gas region, the system outputs the gas temperature and gas concentration, and the solid temperature is displayed as NaN.

[0054] Point Result Recording Unit: Used to save information such as the coordinates, area identification results, area identification probability, solid temperature, gas temperature and gas concentration of each virtual point, forming a candidate point prediction result table;

[0055] Sensor Deployment Auxiliary Analysis Unit: Based on the prediction results of multiple virtual points, analyze the spatial differences in the indoor thermal environment, and select key points with large temperature gradients, significant concentration changes, proximity to areas of human activity, proximity to cold radiation areas of external windows, and proximity to heating terminals or heat sources as priority sensor deployment locations.

[0056] Compared with existing technologies, the advantages of this invention are as follows: This invention constructs numerical simulation samples of the indoor thermal environment, generates gas and solid region labels, utilizes Fourier position coding to enhance spatial coordinate representation capabilities, and employs a multi-task neural network to simultaneously achieve region attribute identification, solid temperature prediction, gas temperature prediction, and gas concentration prediction. While reducing the number of actual sensors deployed, it achieves full-domain reconstruction of indoor thermal environment parameters and provides data support for subsequent thermal comfort assessment and indoor environment control.

[0057] This invention combines indoor thermal environment numerical simulation samples, a multi-task neural network reconstruction model, and a virtual measurement point inference method to achieve full-domain prediction of indoor thermal environment parameters and selection of key measurement points while reducing the number of actual sensors deployed. Compared with traditional single-point or limited fixed-point monitoring methods, this invention can effectively identify indoor gas and solid regions and output gas temperature, gas concentration, and solid temperature respectively, avoiding invalid predictions in areas without physical significance. By using virtual measurement point prediction results to assist in sensor deployment, the number of sensors can be reduced by 40% to 70%, while maintaining the temperature reconstruction error of the main indoor activity areas within ±3.0℃ to ±5.0℃ and the gas concentration reconstruction error within approximately 5% to 10%. By identifying areas of localized overcooling, overheating, and abnormal concentration, the spatial integrity and control accuracy of indoor thermal environment status perception can be improved, providing more reliable control basis for heating, air conditioning, and fresh air equipment, thereby reducing ineffective equipment operation and local comfort deviations. Under ideal operating conditions, it can achieve an energy consumption reduction of approximately 10% to 20% and improve indoor thermal comfort and system operation stability in residential areas with severe cold climates. Attached Figure Description

[0058] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0059] Figure 1 This is an overall framework diagram of the present invention;

[0060] Figure 2 This is a structural diagram of the basic information module for indoor space in this invention;

[0061] Figure 3 This is a structural diagram of the neural network reconstruction module of the present invention;

[0062] Figure 4 This is a flowchart of the indoor thermal environment virtual measurement point inference module system for optimizing sensor deployment according to the present invention;

[0063] Figure 5 The above are the gas / solid region identification results (X=0.5) in Embodiment 2 of the present invention, where (a) is the simulated identification result of the gas / solid region and (b) is the predicted identification result of the gas / solid region.

[0064] Figure 6 The above is a gas temperature reconstruction cloud map (X=5.1) in Embodiment 2 of the present invention, wherein (a) is a gas temperature simulation cloud map and (b) is a gas temperature reconstruction cloud map;

[0065] Figure 7 The above is a gas concentration reconstruction cloud map (X=5.1) in Embodiment 2 of the present invention, wherein (a) is a gas concentration simulation cloud map and (b) is a gas concentration reconstruction cloud map;

[0066] Figure 8 The image shows the spatial distribution of prediction error in Embodiment 2 of the present invention (X=5.1), where (a) is the absolute error map of gas temperature and (b) is the absolute error map of gas concentration.

[0067] Figure 9 This is a map showing the random candidate points and recommended sensor locations in Embodiment 3 of the present invention. Detailed Implementation

[0068] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0069] Secondly, the present invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include three-dimensional spatial dimensions of length, width, and depth.

[0070] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0071] This invention achieves full-domain reconstruction of indoor thermal environment parameters and assisted optimization of sensor deployment schemes by constructing numerical simulation samples, training multi-task neural networks, predicting virtual measurement points, and screening key measurement points, while reducing the number of actual sensors.

[0072] Specifically, a method for reconstructing indoor thermal environment parameters based on a multi-task neural network includes the following steps:

[0073] S1: Establish a numerical simulation model of the indoor thermal environment of a residential building in a frigid climate zone. The target residential building has a room width of 6m, a depth of 10m, and a height of 3m. Based on the spatial dimensions, enclosure structure, window locations, indoor obstacle distribution, heating terminal locations, and localized heat sources of the target residential building, construct the indoor air region, solid enclosure structure region, indoor furnishings and obstacle region, heating terminal region, and localized heat source region. The enclosure structure includes walls, floors, ceilings, and windows, while indoor solid components include beds, sofas, furniture, and heating terminals. The thermal conductivity, density, and specific heat capacity of various materials are assigned values ​​according to the material parameters set in the numerical simulation model of the indoor thermal environment. The model boundary conditions mainly include the outdoor low temperature boundary, the heat transfer boundary of the enclosure structure, the coupled heat transfer boundary between indoor air and solid surfaces, the heating terminal heat source boundary, and the gas concentration source term boundary. Among them, the heating terminal or localized heat source can be described using a constant temperature boundary, an equivalent heat flux boundary, or a volumetric heat source. When the heat source is a volumetric heat source, its equivalent heating power can be expressed as:

[0074] (1)

[0075] in, q represents the equivalent heating power of the heat source, expressed in W; v Volumetric heat source intensity, unit: W / m 3 V represents the volume of the heat source region, in meters. 3 When the heat source adopts the form of surface heat flux, its equivalent heating power can be expressed as:

[0076] (2)

[0077] Where, q s Surface heat flux, in W / m 2 A represents the surface area of ​​the heat source, in m². 2 By solving the numerical simulation model of the indoor thermal environment, the distribution of solid temperature, gas temperature, and gas concentration at different locations within the room is obtained, providing sample data for subsequent virtual measurement point inference, sensor optimization and deployment, and thermal environment parameter reconstruction.

[0078] S2: Extract cross-sectional data from the numerical simulation results of the indoor thermal environment. The indoor space is divided into YZ planes along the X direction, with the X value remaining constant in each cross-section, while the Y and Z values ​​vary with spatial location. Solid temperature, gas temperature, and gas concentration data are extracted separately, forming sample files containing X, Y, Z, and their corresponding physical quantities. This data is used to construct neural network training samples, providing a data foundation for subsequent virtual measurement point prediction and sensor deployment auxiliary analysis.

[0079] S3: Generate region labels based on the sample data source. Coordinate points with solid temperature data are marked as solid regions; coordinate points with gas temperature or gas concentration data are marked as gas regions. If the same coordinate appears in both solid and gas regions, it is identified as a region conflict point and is either eliminated or processed separately according to preset rules. This method enables the model not only to predict temperature and concentration but also to determine whether the point to be predicted belongs to a solid or gas region.

[0080] S4: Constructing an indoor thermal environment reconstruction neural network based on Fourier position encoding and a multi-task output head. This model takes indoor spatial coordinates as input and, through coordinate normalization, Fourier position encoding, shared feature extraction, and a multi-task output structure, achieves joint prediction of regional attributes and thermal environment parameters at the point to be predicted. Let the three-dimensional spatial coordinates of the point to be predicted be:

[0081] (3)

[0082] To eliminate the influence of different coordinate scales on the model training results, the input coordinates are first normalized:

[0083] (4)

[0084] In the formula, x is the original spatial coordinate. These are the normalized spatial coordinates. To train the mean of the spatial coordinates, The standard deviation of the spatial coordinates is used for training. To enhance the model's ability to represent the spatial variations of indoor temperature and concentration fields, this invention performs Fourier position encoding on the normalized coordinates. The input feature dimensions after Fourier encoding are:

[0085] (5)

[0086] In the formula, B represents the number of Fourier frequencies. In this embodiment, B=8, therefore the input dimension after model encoding is 51. The encoded spatial features are input into the shared feature extraction network to obtain the shared spatial features of the indoor thermal environment:

[0087] (6)

[0088] In the formula, This represents the spatial features after Fourier position encoding. Indicates a shared feature extraction network. Here, represents the network parameters, and h represents the extracted shared features of the indoor thermal environment. This shared feature extraction network consists of multiple fully connected layers, normalized layers, SiLU activation functions, and Dropout layers. In this embodiment, the shared feature extraction network has 5 hidden layers, 256 hidden layer neurons, and a Dropout ratio of 0.05. Based on the shared feature h, the model sets multiple task output heads. Among them, the region recognition output head is used to determine whether the input point belongs to a gas region or a solid region, and its output can be expressed as:

[0089] (7)

[0090] In the formula, , This represents the probability that the input point belongs to the gas region. This represents the probability that the input point belongs to a solid region. The weight matrix for the output head of the region identification. The bias term for the region identification output head. The three regression output heads for solid temperature, gas temperature, and gas concentration are used to predict the corresponding thermal environment parameters, and their outputs can be uniformly expressed as:

[0091] (8)

[0092] In the above formula, This represents the predicted value of the solid temperature. This represents the predicted gas temperature. This represents the predicted gas concentration. The model first uses the region identification output head to determine the spatial attributes of the point to be predicted, and then selects the corresponding physical quantity as the effective output based on the region attributes. In this way, the model can simultaneously learn the spatial mapping relationship between gas regions, solid regions, and thermal environment parameters.

[0093] S5: Model training and validation are performed using an X-section partitioning method. Since the sample data of this invention originates from YZ sections at different X-coordinate positions, directly and randomly partitioning all spatial points might result in adjacent points within the same section appearing simultaneously in both the training and validation sets, thus preventing the validation results from fully reflecting the model's predictive ability for unknown sections. Therefore, this invention uses the X-section as the basic unit to partition the training and validation sets, randomly selecting a portion of the X-sections as the validation set and the remaining X-sections as the training set. In this embodiment, the validation set proportion is 20%.

[0094] During model training, the tasks of region identification, solid temperature prediction, gas temperature prediction, gas concentration prediction, and weak heat source identification are optimized simultaneously. Since solid regions lack gas temperature and gas concentration, and gas regions lack solid temperature, this invention employs a masked mean squared error loss for the regression task, ensuring that samples without corresponding physical quantities are not included in this loss calculation. The masked loss can be expressed as:

[0095] (9)

[0096] In the above formula, Encoding for validity. When the i-th sample has a corresponding physical quantity, When the i-th sample does not have a corresponding physical quantity, ; These are the model's predicted values. This represents the true value. Through this masking mechanism, solid region samples only participate in the calculation of solid temperature loss, and gas region samples only participate in the calculation of gas temperature and gas concentration loss. Finally, the multi-task joint loss function is:

[0097] (10)

[0098] In the above formula, For area identification loss, For solid temperature prediction loss, For predicting losses based on gas temperature, For predicting losses based on gas concentration, The loss term for weak heat source identification is represented by a coefficient of 0.2. Through this joint loss function, the model can simultaneously learn region identification capabilities and multi-physical quantity prediction capabilities.

[0099] After training, the model outputs corresponding physical quantities based on the region identification results. When the input point is identified as a solid region, the system only outputs the solid temperature, while the gas temperature and gas concentration are displayed as NaN; when the input point is identified as a gas region, the system outputs the gas temperature and gas concentration, while the solid temperature is displayed as NaN. This output method avoids outputting invalid parameters in regions that do not have corresponding physical meaning, ensuring that the prediction results match the actual properties of indoor solid and gas regions.

[0100] S6: During actual operation, a small number of optimized sensor nodes will be used to collect local indoor temperature or concentration data. These data will then be combined with a trained neural network model to reconstruct thermal environment parameters for locations where no sensors are deployed. The system can correct the reconstruction results based on the discrepancy between the measured sensor data and the model's predictions, thereby improving the prediction stability and reliability over long-term operation. Adaptive error correction methods can employ sliding window correction, distance-weighted error correction, or residual correction models.

[0101] S7: The reconstructed indoor thermal environment parameters are used for thermal comfort assessment and control recommendations. Based on differences in gas temperature, solid surface temperature, gas concentration, and spatial distribution within the activity area, the system determines whether there are areas of localized overcooling, overheating, or abnormal concentrations, and provides control recommendations to indoor environmental regulation equipment such as heating, air conditioning, or ventilation systems. When the indoor thermal environment deviates from the comfort range, the system outputs corresponding control recommendations; when the indoor thermal environment meets comfort requirements, the system can reduce equipment operating intensity to decrease energy consumption.

[0102] This invention also provides an indoor thermal environment parameter reconstruction system based on a multi-task neural network, the overall block diagram of which is shown below. Figure 1As shown, the entire system mainly consists of an indoor space basic information module, an optimized sensor deployment module, a sparse temperature / concentration data acquisition module, a data preprocessing and normalization module, a multi-task neural network reconstruction module, a thermal environment parameter output module, and a thermal comfort assessment and control suggestion module. The indoor space basic information module is used to determine spatial objects such as indoor air zones, solid building envelopes, obstacles, and heat sources in residential buildings in frigid regions. The optimized sensor deployment module selects key measuring points and determines sensor node locations based on the indoor space structure and thermal environment parameter distribution characteristics. The sparse temperature / concentration data acquisition module acquires local indoor temperature and concentration data through a limited number of sensor nodes. The data preprocessing and normalization module performs header checks, duplicate data checks, invalid data ignoring, coordinate normalization, and data standardization on the acquired data and simulation sample data. Subsequently, the processed spatial coordinates and sparse observation data are input into the multi-task neural network reconstruction model to predict the indoor global thermal environment parameters. Finally, it outputs solid temperature, gas temperature, and gas concentration, which are further used for thermal comfort state assessment and indoor environment control suggestion generation.

[0103] The functions of each module in the system are described below:

[0104] Indoor space basic information module such as Figure 2 As shown, it is used to obtain basic information such as the geometry, enclosure structure, obstacles, heat sources and boundary conditions of the residential interior space, providing a spatial basis for sensor deployment and thermal environment reconstruction.

[0105] Optimize sensor deployment module: Based on the distribution characteristics of indoor thermal environment parameters, key sensor nodes are selected from candidate measurement points to reduce the number of sensors while ensuring the ability to characterize the overall indoor thermal environment.

[0106] Sparse temperature / concentration data acquisition module: Collects indoor local thermal environment data through temperature sensors and gas concentration sensors deployed at key locations, and transmits the data to the data processing terminal.

[0107] Data preprocessing and normalization module: Performs unified processing on the collected data and the cross-sectional data exported from the indoor thermal environment numerical simulation model, including X / Y / Z header checking, duplicate data checking, coordinate normalization and unit unification, forming standardized samples that can be input into the neural network model.

[0108] Multi-task neural network reconstruction module, such as Figure 3 As shown, using spatial coordinates X, Y, Z and sparse observation data as input, the system determines whether the input coordinates belong to a solid region or a gas region through region identification and multi-task regression structure, and predicts the solid temperature, gas temperature and gas concentration at the corresponding location.

[0109] Thermal environment parameter output module: Outputs solid temperature, gas temperature, and gas concentration at different locations within the indoor space. When the coordinates are located in the solid region, the solid temperature is output, while the gas temperature and gas concentration are displayed as NaN; when the coordinates are located in the gas region, the gas temperature and gas concentration are output, while the solid temperature is displayed as NaN.

[0110] Thermal comfort assessment and control suggestion module: Based on the reconstructed indoor thermal environment parameters, assess the thermal comfort status of the activity area and output control suggestions to indoor environmental conditioning equipment such as heating, ventilation or air conditioning.

[0111] Indoor thermal environment virtual measurement point inference module for optimizing sensor deployment

[0112] The system flow of the virtual measurement point inference module for indoor thermal environment aimed at optimizing sensor deployment is as follows: Figure 4 As shown, this module is used to predict thermal environment parameters for candidate or random points in an indoor space based on a trained indoor thermal environment reconstruction model before actual sensor deployment, thereby assisting in determining the appropriate sensor placement locations. This module mainly consists of a model weight loading unit, a virtual point input unit, a model inference calculation unit, a region attribute discrimination unit, a thermal environment parameter output unit, a point result recording unit, and a sensor deployment auxiliary analysis unit. Through this module, the predicted solid temperature, gas temperature, and gas concentration at different spatial points can be obtained in advance without increasing the number of actual sensors, providing a basis for subsequent selection of key measuring points, reduction of sensor deployment numbers, and improvement of the ability to characterize the indoor thermal environment across the entire area.

[0113] The functions of each unit are described below:

[0114] Model weight loading unit: Used to load the weights of the trained indoor thermal environment reconstruction model, enabling the system to call the trained model to quickly predict virtual locations.

[0115] Virtual point input unit: Used to input or generate the X, Y, Z coordinates of an indoor spatial point to be predicted. This point can be a single coordinate input manually, or a candidate measurement point or a random measurement point generated within the indoor space.

[0116] Model inference calculation unit: used to perform coordinate normalization processing on the input point and call the multi-task neural network model to complete the inference of thermal environment parameters, and obtain the regional attributes and corresponding physical quantity prediction results at the point.

[0117] Region attribute discrimination unit: Used to determine whether the input point belongs to a gas region or a solid region. When the model determines that the point belongs to a gas region, the point can be used as a candidate location for gas temperature or gas concentration sensors; when the model determines that the point belongs to a solid region, the point can be used as a candidate location for solid surface temperature monitoring or thermal state analysis near a heat source.

[0118] Thermal environment parameter output unit: Used to output the solid temperature, gas temperature, and gas concentration at the virtual location. When the location is in a solid region, the system outputs the solid temperature, while the gas temperature and concentration are displayed as NaN; when the location is in a gas region, the system outputs the gas temperature and concentration, while the solid temperature is displayed as NaN. This output method ensures that the predicted results are consistent with the physical properties of the indoor solid and gas regions.

[0119] The point result recording unit is used to store information such as the coordinates, region identification results, region identification probability, solid temperature, gas temperature, and gas concentration of each virtual point, forming a candidate point prediction result table. This result table can be used for subsequent analysis of temperature differences, concentration differences, and spatial distribution characteristics between different points.

[0120] Sensor Deployment Auxiliary Analysis Unit: Based on the prediction results of multiple virtual points, this unit analyzes the spatial differences in the indoor thermal environment and selects key points with large temperature gradients, significant concentration changes, proximity to areas of human activity, proximity to cold radiation areas from windows, and proximity to heating terminals or heat sources as priority sensor deployment locations. Specific selection criteria may include the predicted temperature change range, concentration change range, regional representativeness, spatial coverage, and measurement point redundancy.

[0121] In practical use, the trained indoor thermal environment reconstruction model is first loaded into the model weight loading unit; then, the coordinates of candidate indoor points are input or generated in the virtual point input unit; the model inference calculation unit predicts the thermal environment parameters for each point, and the regional attribute discrimination unit determines the spatial region to which it belongs; the thermal environment parameter output unit outputs the solid temperature, gas temperature, and gas concentration at each point; finally, the sensor deployment auxiliary analysis unit selects key measuring points that can represent the spatial differences in the indoor thermal environment based on the prediction results of different points. This method allows for the completion of virtual point selection and deployment scheme pre-evaluation before actual sensor installation, thereby reducing the number of sensors, lowering deployment costs, and improving the ability of a small number of sensors to characterize the overall indoor thermal environment.

[0122] Example:

[0123] Example 1: Model training instance based on numerical simulation cross-sectional data.

[0124] This study focuses on the indoor space of a residential building in a town in a frigid climate zone. The room is 6m wide, 10m deep, and 3m high. Based on the geometric dimensions of the indoor space, the external envelope, the location of windows, heating terminals, indoor obstacles, and local heat sources, a numerical simulation model of the indoor thermal environment is established. By solving this model, the distribution of solid temperature, gas temperature, and gas concentration at different locations within the indoor space is obtained.

[0125] The simulation results are divided along the X-axis into YZ planes to export solid temperature cross-sectional data, gas temperature cross-sectional data, and gas concentration cross-sectional data, respectively. Each sample file includes at least X, Y, and Z three-dimensional spatial coordinates and corresponding physical quantity values. Coordinate points containing solid temperature data are marked as solid regions; coordinate points containing gas temperature or gas concentration data are marked as gas regions. If the same coordinate appears in both solid and gas regions, it is treated as a conflict point and either removed or processed separately.

[0126] During model training, the X, Y, and Z coordinates are first normalized and Fourier positional encoded. The encoded spatial features are then input into a shared feature extraction network, and multiple task output heads are used to perform region attribute recognition, solid temperature prediction, gas temperature prediction, and gas concentration prediction, respectively. During training, the training and validation sets are divided according to the X-section to avoid adjacent points within the same section appearing in both sets simultaneously, thereby improving the representativeness of the validation results for unknown sections.

[0127] In this embodiment, the number of training rounds can be set to 500, the validation set ratio to 20%, and the learning rate to 1×10⁻⁶. -3 The model has 5 hidden layers, 256 hidden layer neurons, a Dropout ratio of 0.05, and 8 Fourier frequencies. The sample data structure and model training parameters for this embodiment are shown in Tables 1 and 2, respectively.

[0128] Table 1: Sample Data Structure Table for Example Implementation

[0129]

[0130] Supplementary Table 2: Model Training Parameter Table

[0131]

[0132] Example 2: Result of indoor thermal environment parameter reconstruction.

[0133] After model training is complete, the X-section, which was not used in training, is selected as the validation section. The X, Y, and Z coordinates of each spatial point on this section are input into the trained multi-task neural network model. The model first outputs the probability that each spatial point belongs to a gas region or a solid region, and then outputs the corresponding physical quantity based on the region identification result. When a spatial point is identified as a solid region, the solid temperature is output, while the gas temperature and gas concentration are recorded as invalid values; when a spatial point is identified as a gas region, the gas temperature and gas concentration are output, while the solid temperature is recorded as an invalid value.

[0134] By comparing the model predictions with the numerical simulation results, the region identification accuracy, solid temperature prediction error, gas temperature prediction error, and gas concentration prediction error can be obtained. The validation results can be evaluated using mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (CDO). In Table 3, MAE represents the average level of the absolute deviation between the model predictions and the numerical simulation values; a smaller MAE value indicates a smaller overall prediction deviation. RMSE represents the square root of the average squared prediction errors; compared to MAE, RMSE is more sensitive to larger local errors and can therefore be used to reflect the error amplification of the model in areas of sudden changes or complex boundaries. Table 3 shows that the region identification accuracy reaches 88.14%, indicating that the model can effectively distinguish between indoor gas and solid regions. The R² values ​​for solid temperature, gas temperature, and gas concentration prediction tasks are 0.9756, 0.9532, and 0.9471, respectively, all close to 1, indicating a high consistency between the model predictions and the numerical simulation results. The MAE and RMSE for solid temperature prediction were 1.6716 and 2.5984, respectively; for gas temperature prediction, they were 3.3029 and 9.4975, respectively; and for gas concentration prediction, they were 0.0196 and 0.0631, respectively. This indicates that the model can achieve joint reconstruction of indoor solid temperature, gas temperature, and gas concentration while completing regional attribute identification. Figure 5 It can be seen that the predicted region is largely consistent with the actual region in terms of main spatial structure. The model can identify most gas and solid regions indoors, with only a few misclassifications at local boundaries, narrow slits, or gas-solid interfaces. This result demonstrates that the region identification output head can provide an effective basis for spatial attribute judgment in subsequent temperature and concentration tasks, avoiding the output of inconsistent thermal environment parameters in physically meaningless regions. Figure 6 It can be seen that the reconstructed gas temperature cloud map maintains the same overall spatial distribution trend as the numerical simulation cloud map. The model can reconstruct the main spatial variation patterns of indoor gas temperature and reflect local high-temperature zones, low-temperature zones, and temperature gradient change areas. Some deviations exist near complex boundaries and in areas with drastic temperature changes, but the overall reconstruction results meet the needs of comprehensive indoor thermal environment prediction and thermal comfort evaluation. Figure 7 It can be seen that the gas concentration reconstruction results can reflect the main spatial distribution characteristics of the concentration field, and the predicted and simulated cloud maps show good consistency in the main concentration change areas. Due to the potentially stronger non-uniform distribution of gas concentration near local source terms, boundary regions, and spatially obstructed areas, some errors exist in local locations, but the overall prediction results can still be used to identify areas with relatively high or significantly changing indoor concentrations. Figure 8 It can be seen that the absolute errors in gas temperature and gas concentration are mainly concentrated at local boundaries, gas-solid interfaces, near obstacles, and in areas with large changes in physical quantity gradients, while the errors are relatively small in most indoor areas. This error distribution indicates that the model has a relatively stable reconstruction capability over a large spatial area, and the high local errors are mainly related to complex geometric boundaries, sample sparsity, and gas-solid region switching. Combined with the quantitative evaluation results in Table 3, it can be proven that the multi-task neural network model can accurately complete the regional identification and multi-physical quantity reconstruction of indoor thermal environment parameters.

[0135] Table 3: Validation set prediction results

[0136]

[0137] Example 3: Virtual Measurement Point Inference and Sensor Deployment Assistance:

[0138] Before the actual sensor installation, multiple candidate measurement points are first generated in the indoor space of the target residence. These candidate measurement points can be generated using a uniform grid, or they can be generated more densely in areas with high foot traffic, near windows, near heating terminals, near heat sources, and near ventilation paths. The X, Y, and Z coordinates of each candidate measurement point are input into a trained multi-task neural network model to obtain the predicted region category, region identification probability, solid temperature, gas temperature, and gas concentration for each candidate point.

[0139] Based on the temperature differences, concentration differences, spatial coverage, regional representativeness, and measurement point redundancy among candidate points, key points that can represent the spatial distribution characteristics of the indoor thermal environment are selected. Points with large temperature gradients, significant concentration changes, proximity to areas of human activity, proximity to cold radiation areas from windows, and proximity to heating terminals or heat sources are prioritized for sensor deployment. This approach reduces the number of actual sensors while improving the ability of a small number of sensors to characterize the overall indoor thermal environment. Table 4 lists the recommended point numbers, spatial coordinates, region, simulated actual value, model predicted value, absolute error, relative error, and heat source prediction probability. X, Y, and Z represent the coordinates of the recommended point in the indoor three-dimensional space; "region" indicates whether the point belongs to a solid or gaseous region; absolute error represents the absolute value of the difference between the model predicted value and the numerical simulation actual value; relative error represents the proportion of absolute error to the simulated actual value, used to evaluate prediction bias under different magnitudes of data; and the heat source prediction probability characterizes the likelihood that the point is affected by a local heat source or is close to a heat source region. Table 4 shows that F01-F05 are recommended solid area points, mainly used for surface temperature monitoring near solid enclosure structures, indoor obstacles, or heat-generating terminals. The prediction error of these points is generally small, with absolute errors of 0.3962, 0.8179, and 0.3156 for F01, F02, and F04, respectively, indicating good predictive stability of the model for solid area temperatures. F06-F10 are recommended gas area points, mainly used to characterize the temperature or concentration changes in indoor air areas. Among them, F09 and F10 have higher simulated values, and the predicted probability of the heat source for F10 is 0.5489, significantly higher than other points, indicating that this point may be located in an area with strong heat source influence or significant thermal environment changes, making it suitable as a priority location for gas temperature or concentration sensor deployment. As shown in Table 4, the selected key points include both solid and gas area points, which can meet the thermal environment characterization needs of different physical areas indoors. Figure 9 This is a map showing random candidate points and recommended sensor locations. Light gray dots represent random candidate points, triangles represent recommended solid regions, and circles represent recommended gas regions. Figure 9 It can be seen that this embodiment generated a total of 80 random candidate points, and after screening, 12 recommended area points and 68 unrecommended area points were obtained. The recommended points are distributed in a dispersed manner in the three-dimensional space of the room, covering different X, Y, and Z directions. They include points near the building envelope, cold radiation areas of external windows, heating terminals or heat sources, as well as representative points in areas of human activity and indoor air. This layout avoids excessive concentration of sensor points, reduces measurement point redundancy, and enhances the ability of a small number of sensors to characterize the spatial distribution of the indoor thermal environment. (See Table 4 and...) Figure 9It can be seen that the key point screening method can select a small number of representative sensor deployment locations from a large number of candidate points. While reducing the actual number of sensors, it retains the main spatial variation characteristics of the indoor thermal environment, providing a reliable data foundation for subsequent full-domain reconstruction of indoor thermal environment parameters, thermal comfort state assessment, and heating or ventilation control.

[0140] Table 4: Differences between prediction results and simulation results for random fixed points

[0141]

[0142] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for reconstructing indoor thermal environment parameters based on a multi-task neural network, characterized in that, Includes the following steps: S1: Establish a numerical simulation model of the indoor thermal environment of urban residential buildings in a frigid climate zone to obtain the distribution of solid temperature, gas temperature and gas concentration at different spatial locations inside the building; S2: Extract cross-sectional data from the numerical simulation results of the indoor thermal environment and construct training samples for the neural network; S3: Generate region labels based on the source of sample data; S4: Construct an indoor thermal environment reconstruction neural network based on Fourier position encoding and multi-task output head. Using indoor spatial coordinates as input, the network achieves joint prediction of regional attributes and thermal environment parameters of the point to be predicted through coordinate normalization, Fourier position encoding, shared feature extraction and multi-task output structure. S5: Model training and validation are performed by dividing the model into X-sections. Some X-sections are randomly selected as the validation set, and the remaining X-sections are used as the training set. At the same time, the tasks of region recognition, solid temperature prediction, gas temperature prediction, gas concentration prediction and weak heat source recognition are optimized. S6: During the actual operation phase, a small number of optimized sensor nodes will be used to collect local indoor temperature or concentration data. Combined with a trained neural network model, the thermal environment parameters at locations where no sensors are deployed will be reconstructed. The reconstructed indoor global thermal environment parameters will be used for thermal comfort state assessment and control recommendations.

2. The indoor thermal environment parameter reconstruction method based on a multi-task neural network according to claim 1, characterized in that, The specific method of S1 is as follows: based on the spatial dimensions, enclosure structure, window location, indoor obstacle distribution, heating terminal location, and local heat source information of the target residence, an indoor air area, a solid enclosure structure area, an indoor furnishing and obstacle area, a heating terminal area, and a local heat source area are constructed. The enclosure structure includes walls, floors, ceilings, and windows; the indoor solid components include beds, sofas, furniture, and heating terminals. The thermal conductivity, density, and specific heat capacity of various materials are assigned values ​​according to the material parameters set in the indoor thermal environment numerical simulation model. The model boundary conditions include outdoor low temperature boundary, enclosure structure heat transfer boundary, indoor air-solid surface coupling heat transfer boundary, heating terminal heat source boundary, and gas concentration source term boundary. Among them, the heating terminal or local heat source is described using a constant temperature boundary, an equivalent heat flux boundary, or a volume heat source form. When the heat source is in the form of a volume heat source, its equivalent heating power is expressed as: (1) in, q represents the equivalent heating power of the heat source; v Let V be the volumetric heat source intensity; V be the volume of the heat source region; when the heat source adopts the form of surface heat flux, its equivalent heating power is expressed as: (2) Where, q s denoted as , where A is the surface heat flux and A is the surface area of ​​the heat source.

3. The method for reconstructing indoor thermal environment parameters based on a multi-task neural network according to claim 1, characterized in that, The specific method for data extraction in S2 is as follows: the indoor space is cut along the X direction using a YZ plane, with the X value remaining constant in each cut, while the Y and Z values ​​change with the spatial position; solid temperature, gas temperature, and gas concentration data are extracted respectively, and a sample file containing X, Y, Z, and the corresponding physical quantities is formed.

4. The method for reconstructing indoor thermal environment parameters based on a multi-task neural network according to claim 1, characterized in that, The specific method of S3 is as follows: for coordinate points with solid temperature data, they are marked as solid regions; for coordinate points with gas temperature or gas concentration data, they are marked as gas regions; if the same coordinate appears in both solid and gas regions, it is identified as a region conflict point and is eliminated or processed separately according to preset rules.

5. The method for reconstructing indoor thermal environment parameters based on a multi-task neural network according to claim 1, characterized in that, The specific method of S4 is as follows: Let the three-dimensional spatial coordinates of the point to be predicted be: (3) To eliminate the influence of different coordinate scales on the model training results, the input coordinates are first normalized: (4) In the formula, x is the original spatial coordinate. These are the normalized spatial coordinates. To train the mean of the spatial coordinates, To determine the standard deviation of spatial coordinates in the training set, Fourier position encoding is performed on the normalized coordinates. The input feature dimension after Fourier encoding is: (5) In the formula, B represents the number of Fourier frequencies, B=8, therefore the input dimension after model encoding is 51; the encoded spatial features are input into the shared feature extraction network to obtain the spatial shared features of the indoor thermal environment: (6) In the formula, Represents the spatial features after Fourier position encoding. Indicates a shared feature extraction network. Here, h represents the extracted shared features of the indoor thermal environment. This shared feature extraction network consists of multiple fully connected layers, normalization layers, SiLU activation functions, and Dropout layers. Based on the shared feature h, the model sets multiple task output heads. The region recognition output head is used to determine whether the input point belongs to a gas region or a solid region, and its output is expressed as follows: (7) In the formula, , This represents the probability that the input point belongs to the gas region. This represents the probability that the input point belongs to a solid region. The weight matrix for the output head of the region identification. The bias term for the region identification output head; the three regression output heads for solid temperature, gas temperature, and gas concentration are used to predict the corresponding thermal environment parameters, and their outputs are uniformly represented as: (8) In the formula, This represents the predicted value of the solid temperature. This represents the predicted gas temperature. This indicates the predicted gas concentration. First, the spatial attributes of the point to be predicted are determined by the region identification output head, and then the corresponding physical quantity is selected as the effective output based on the region attributes.

6. The method for reconstructing indoor thermal environment parameters based on a multi-task neural network according to claim 1, characterized in that, The specific method of S5 is to apply a masked mean squared error loss to the task, so that samples without corresponding physical quantities are not included in the loss calculation. The masked loss is expressed as: (9) In the formula, For validity encoding, when the i-th sample has a corresponding physical quantity, ; When the i-th sample does not have a corresponding physical quantity ; These are the model's predicted values. The values ​​are the true values; through this masking mechanism, solid region samples only participate in the calculation of solid temperature loss, and gas region samples only participate in the calculation of gas temperature and gas concentration loss; finally, the multi-task joint loss function is: (10) In the formula, For area identification loss, For solid temperature prediction loss, For predicting losses based on gas temperature, For predicting losses based on gas concentration, The coefficient 0.2 represents the weight of the weak heat source identification loss term. After training, the system outputs the corresponding physical quantities based on the region recognition results. When the input point is identified as a solid region, the system only outputs the solid temperature, while the gas temperature and gas concentration are displayed as NaN. When the input point is identified as a gas region, the system outputs the gas temperature and gas concentration, while the solid temperature is displayed as NaN.

7. An indoor thermal environment parameter reconstruction system based on a multi-task neural network, used to implement the indoor thermal environment parameter reconstruction method based on a multi-task neural network as described in any one of claims 1-6, characterized in that, include: Indoor Space Basic Information Module: Used to acquire basic information on the geometric structure, enclosure structure, obstacles, heat sources and boundary conditions of residential indoor spaces, providing a spatial basis for sensor deployment and thermal environment reconstruction; Optimize sensor deployment module: Based on the distribution characteristics of indoor thermal environment parameters, key sensor nodes are selected from candidate measurement points to reduce the number of sensors while ensuring the ability to characterize the overall indoor thermal environment. Sparse temperature / concentration data acquisition module: Collects indoor local thermal environment data through temperature sensors and gas concentration sensors deployed in key locations, and transmits the data to the data processing terminal; Data preprocessing and normalization module: performs unified processing on the collected data and the cross-sectional data exported from the indoor thermal environment numerical simulation model, including X / Y / Z header checking, duplicate data checking, coordinate normalization and unit unification, forming standardized samples that can be input into the neural network model; The multi-task neural network reconstruction module takes spatial coordinates X, Y, Z and sparse observation data as input, and determines whether the input coordinates belong to a solid region or a gas region through region recognition and multi-task regression structure, and predicts the solid temperature, gas temperature and gas concentration at the corresponding location. Thermal environment parameter output module: Outputs solid temperature, gas temperature, and gas concentration at different locations in the indoor space; when the coordinate is located in the solid region, the solid temperature, gas temperature, and gas concentration are output and displayed as NaN; when the coordinate is located in the gas region, the gas temperature and gas concentration are output, and the solid temperature is displayed as NaN. Thermal comfort assessment and control suggestion module: Based on the reconstructed indoor thermal environment parameters, assess the thermal comfort status of the activity area and output control suggestions to indoor environmental control equipment such as heating, ventilation or air conditioning; A virtual measurement point inference module for indoor thermal environment aimed at optimizing sensor deployment: This module is used to predict thermal environment parameters for candidate or random points in the indoor space based on a trained indoor thermal environment reconstruction model before actual sensor deployment, thereby assisting in determining the reasonable deployment location of sensors; without increasing the number of actual sensors, it can obtain the predicted results of solid temperature, gas temperature and gas concentration at different spatial points in advance, providing a basis for subsequent screening of key measurement points, reducing the number of sensors deployed and improving the ability to characterize the indoor thermal environment in its entirety.

8. The indoor thermal environment parameter reconstruction system based on a multi-task neural network according to claim 7, characterized in that, The indoor thermal environment virtual measurement point inference module includes: Model weight loading unit: Used to load the weights of the trained indoor thermal environment reconstruction model, enabling the system to call the trained model to quickly predict virtual locations; Virtual point input unit: used to input or generate the X, Y, Z coordinates of indoor spatial points to be predicted; Model inference calculation unit: used to normalize the coordinates of the input points and call the multi-task neural network model to complete the inference of thermal environment parameters, and obtain the regional attributes and corresponding physical quantity prediction results at the point. Region attribute discrimination unit: used to determine whether the input point belongs to a gas region or a solid region; when the model determines that the point belongs to a gas region, the point is used as a candidate placement location for gas temperature or gas concentration sensors; when the model determines that the point belongs to a solid region, the point is used as a candidate location for solid surface temperature monitoring or thermal state analysis near the heat source. Thermal environment parameter output unit: used to output the solid temperature, gas temperature and gas concentration at the virtual point; when the point is located in the solid region, the system outputs the solid temperature, and the gas temperature and gas concentration are displayed as NaN; when the point is located in the gas region, the system outputs the gas temperature and gas concentration, and the solid temperature is displayed as NaN. Point Result Recording Unit: Used to save information such as the coordinates, area identification results, area identification probability, solid temperature, gas temperature and gas concentration of each virtual point, forming a candidate point prediction result table; Sensor Deployment Auxiliary Analysis Unit: Based on the prediction results of multiple virtual points, analyze the spatial differences in the indoor thermal environment, and select key points with large temperature gradients, significant concentration changes, proximity to areas of human activity, proximity to cold radiation areas of external windows, and proximity to heating terminals or heat sources as priority sensor deployment locations.

Images