Deep learning-based indoor dynamic temperature field prediction method, system and device

Abstract

The application discloses a deep learning-based indoor dynamic temperature field prediction method, system and equipment, which takes a space-time deep convolution residual network as a prediction model, fully considers various complex factors influencing indoor temperature distribution of a building, adopts a data mechanism double-driven method, fully utilizes data driving advantages, constructs a deep convolution residual network model based on deep learning, reduces complexity of the model and requirements for prior knowledge while ensuring prediction accuracy, introduces mechanism information to improve interpretability and generalization ability of the data-driven model, constructs and applies boundary constraints with physical meanings to make the prediction result remain within a range allowed by physical reality, and improves reliability and accuracy of the prediction result. In addition, the method utilizes transfer learning to establish mapping association between a source task and a target task, realizes transfer by fine-tuning a pre-trained model, and improves generalization ability of the model.

Description

Technical Field

[0001] This invention belongs to the field of building room temperature prediction technology, specifically involving a method, system and device for predicting indoor dynamic temperature field based on deep learning. Background Technology

[0002] There is a strong coupling relationship between building energy consumption, environment, and people: building energy consumption is mainly generated by people's production and living activities. The high uncertainty of people's behavior affects the building environment, and the degree to which the indoor environment affects people's comfort varies from person to person. Furthermore, considering that changes in the indoor environment are a slow, dynamic process, the randomness of people's comfort requirements and actions brings uncertainty to building energy consumption. Statistics show that in developed countries, building energy consumption accounts for 20%-40% of total social energy consumption. Moreover, due to varying degrees of energy loss during production, transportation, conversion, and distribution, the energy demand of buildings will continue to grow with population growth and social development. Therefore, optimizing and controlling the demand side of buildings to achieve energy conservation has a certain amplifying effect, leading to better emission reduction and sustainable development. On the other hand, statistics show that a person spends almost 87% of their life in enclosed buildings, and the building environment also greatly affects people's comfort levels. The building environment involves thermal environment, humidity environment, gas distribution, and pollutant concentration levels, among which the study of the thermal environment is most noteworthy. This is because people's main comfort needs are for heating and cooling, and a significant portion of building energy is used for heating and cooling. The safe and stable operation of most equipment depends on a suitable operating temperature range. Therefore, rapid prediction and analysis of building indoor temperature is crucial for identifying the building's indoor thermal environment, regulating environmental comfort, optimizing the control of HVAC systems and heat source equipment, and supporting building operation scheduling decisions. Generally, the distribution of indoor temperature in buildings is assumed to be uniform. However, research shows that indoor temperature is not uniformly distributed. In fact, in rooms with HVAC systems, there are significant temperature differences in different locations. Therefore, effective prediction of the building's indoor temperature field is essential for optimizing equipment regulation, improving occupant comfort, and reducing building energy consumption.

[0003] The building's indoor thermal environment is a high-dimensional, nonlinear, and discontinuous dynamic physical process. It is influenced by numerous uncertainties, including outdoor weather, solar radiation, building layout, building envelope, human behavior, and equipment operation, making accurate description and analysis difficult. Currently, building thermal environment modeling methods mainly fall into two categories: mechanism-based methods and data-driven methods. Mechanism-based methods, based on prior knowledge and numerical modeling, are the mainstream approach for thermal environment modeling. These include nodal modeling, computational fluid dynamics-based modeling, the CRI index method, low-dimensional linearized ventilation models, and grey box models such as RC network models. Data-driven methods can simulate thermodynamics even without prior knowledge of the detailed characteristics of the target area, thus becoming a promising method for predicting indoor thermal environments. The most common data-driven method is linear system identification based on reduced-order models; other methods include machine learning models based on regression trees or neural networks. In summary, although many methods have been developed for predicting indoor building temperatures, limitations remain: First, mechanistic-based methods suffer from high model complexity and reliance on prior knowledge, resulting in high solution costs and low computational efficiency. While data-driven methods improve modeling speed, most are overly simplified and only applicable to single-node modeling. Furthermore, data-driven models, lacking physical interpretability, may produce results that deviate from physical laws, leading to overfitting. In conclusion, a comprehensive solution to the problem of building indoor thermal environment modeling still lacks a balance between computation time, computational complexity, prediction accuracy, input information requirements, and generalization ability. Summary of the Invention

[0004] To address the limitations of the existing technologies, this invention provides a deep learning-based method, system, and device for predicting indoor dynamic temperature fields. This method achieves an optimal balance between evaluation metrics such as prediction accuracy, computational complexity, prior information requirements, and generalization ability. It also provides visualization capabilities, thereby improving the overall performance of indoor temperature field prediction methods for buildings.

[0005] A deep learning-based method for predicting indoor dynamic temperature fields includes the following steps:

[0006] S1. Construct a dynamic temperature field deep neural prediction network, and input the large-scale dataset of the source task domain as the training set into the dynamic temperature field deep neural network for training to generate a pre-trained model.

[0007] S2. Acquire the perception information of the target task, which includes thermal disturbance information and basic physical information.

[0008] S3. Based on the spatial structure feature library and the basic physical information obtained in S2, perform integrated learning to generate the spatial structure features of the target task;

[0009] S4. Process the perception information of the target task obtained in S2 and construct the target task domain dataset.

[0010] S5. Fine-tune the pre-trained model obtained in S1 according to the characteristics of the target task domain dataset, and output the dynamic temperature field prediction results.

[0011] S6. Calculate the error between the predicted value of a node at a certain moment and its corresponding actual measurement value in the dynamic temperature field prediction result obtained in S5 as an indicator of accuracy, and construct a loss function to determine whether the error at this time is less than the expected value: if so, output the dynamic temperature field prediction result and collect the learned spatial structure features into the spatial structure feature library; otherwise, return to S3 and iteratively optimize the spatial structure features.

[0012] Furthermore, in S3, the spatial structure feature library is constructed through the following steps:

[0013] SA1. Obtain the sensing information of the source task, the sensing information including thermal disturbance information, temperature field node information and spatial structure information;

[0014] SA2. Perform data preprocessing and spatial grid partitioning on the acquired source task perception information, and construct a large-scale dataset of the source task domain.

[0015] SA3. Collect spatial structure information to form a spatial structure feature database.

[0016] Furthermore, step SA2 includes the following steps:

[0017] SA201: The various thermal disturbance information and temperature field node information of the source task obtained in SA1 are preprocessed, and then feature selection based on correlation analysis is performed to obtain the indoor heat source characteristics. Outdoor thermal disturbance characteristics Other characteristics of disturbance variables Event disturbance characteristics and temperature field nodal features Y t Where t = n-1-r,...,n-1, t represents time t, r represents the length of the time series, and n represents the current time;

[0018] SA202. The spatial structure information of the source task obtained in SA1 is processed by spatial grid division. A temperature sensor is deployed in each grid of the source task.

[0019] SA203: Integrate and process the data obtained from SA201 and SA202 to construct a large-scale dataset of the source task domain. The first four items represent input features, and the last item represents label features;

[0020] Xn-1-r→n-1 is the heat source feature and the heat source spatial position distribution tensor L t integrated and processed heat source spatio-temporal tensor

[0021] [X n-1-r ,X n-r ,...,X t ,...,X n-1 , where denotes the heat source spatio-temporal tensor at time t.

[0022] Further, in step S1, the deep neural prediction network for dynamic temperature field includes:

[0023] a deep convolutional residual network for training the heat source spatio-temporal tensor, a feature extraction structure for training other thermal disturbance factors, and a predictor for outputting prediction results, wherein the output ends of the deep convolutional residual network and the feature extraction structure are both connected with the input end of the predictor;

[0024] The deep convolutional residual network includes a proximity module, a trend module and a periodic module, which are respectively used to capture the proximity, trend and periodicity of time features, and the time sequences of proximity, trend and periodicity are respectively defined as and where d c , d e , d p are sampling durations, 1, e, p are corresponding time sampling intervals, 0<1<e<p<n; these three parts share a similar network structure, and the network structure is composed of a plurality of three-dimensional convolutional layers and residual units; the three-dimensional convolution of the l-th layer is defined as:

[0025] Y (l) =α(X (l) *W (l) +b (l) )

[0026] where, X (l) and Y (l) are the input and output of the l-th layer network, W (l) and b (l) are the weight and bias parameters in the l-th layer, * represents the three-dimensional convolution operation, and α(·) is the activation function operation;

[0027] The residual unit of the l-th layer is defined as:

[0028] X (l+1) =X (l) +F(X (l) ,θ (l) ), l=1,...,L;

[0029] Among them, X (l) and X (l+1) are the input and output of the l-th residual unit; F(·) is the operation that enables nonlinearization, normalization and feature extraction of the layer, and L is the total number of residual units;

[0030] The outputs of proximity, trend, and periodicity are fused through feature matrix matching and defined as follows:

[0031]

[0032] Among them, X Fu It is the output of the fusion; M c M e M p These are matrices that are respectively matched with proximity, trend and periodic time features, and are learnable parameters in the model training process; Represents matrix dot product;

[0033] Finally, the output of feature matrix matching is passed through a convolutional long short-term memory network module to achieve temporal multi-step prediction. The final output of this part is represented as X. D ;

[0034] The feature extraction structure includes outdoor thermal disturbance features, other disturbance features, and event disturbance features. These features are merged along the feature axis and then extracted by the feature extraction unit. A fully connected neural network layer is then connected, and the result undergoes non-linear processing using an activation function. Finally, the result is reshaped to match X. D Output X of the same shape O ;

[0035] The outputs of the deep convolutional network module and the feature extraction module are fused and nonlinearized to obtain the predicted value of the temperature field. Then, an optimization strategy based on a penalty factor is introduced to impose physically meaningful boundary constraints on the model, so that the variables are kept within the range allowed by physical reality.

[0036] The loss function of the predictor is a loss function with a regularization penalty term.

[0037] Furthermore, step S3 includes the following steps:

[0038] First, distance-based clustering is performed on the basic physical information and spatial structure feature library of the target task. The distance between the physical information of the target task and the known spatial features in the spatial feature library is obtained through clustering. Weights are assigned according to the clustering distance, and then ensemble learning is performed. Finally, grid search of the spatial location of the heat source is performed to generate the initial spatial structure features of the target task. The spatial structure features are a set of basic physical information of the room in the target task domain and spatial location distribution information of the heat source.

[0039] Furthermore, step S4 includes the following steps:

[0040] S401. Perform data preprocessing on the thermal disturbance information obtained in S2 to obtain heat source characteristics. Outdoor thermal disturbance characteristics Other characteristics of disturbance variables Event disturbance characteristics Where t = n-1-r,...,n-1;

[0041] S402. Perform spatial mesh re-division processing on the spatial structure features generated in S3 so that the input of the target task and the source task maintain the same shape.

[0042] Based on the re-divided grid and according to the spatial distribution characteristics of heat sources, assuming there are M types of heat sources indoors, the spatial distribution of heat sources at time t is represented as follows: in Represent the grid coordinates of the m-th type of heat source; process the spatial distribution of the heat sources into a spatial tensor form, represented as: in It is a four-dimensional tensor representing the spatial location of the m-th type of heat source at time t, where the first dimension represents time, and is defined as:

[0043]

[0044] in, It represents the spatial occupancy status of the m-th type of heat source in the (w,d,h) grid at time t. 1 indicates that it occupies the grid at this time, and 0 otherwise.

[0045] S403. Integrate and process the data obtained from S401 and S402 to construct a large-scale dataset for the target task domain. XT n-1-r→n-1 It is to identify the characteristics of the heat source and the spatial distribution tensor of heat source LT t The integrated heat source spacetime tensor [XT] n-1-r ,XT n-r ,...,XT t ,...,XT n-1 ],in Let t represent the spacetime tensor of the heat source at time t.

[0046] Furthermore, in step S5, the method for fine-tuning the pre-trained model is as follows:

[0047] a) When the target task domain dataset exhibits characteristics of large data volume and data distribution similar to that of the source task domain: based on the pre-trained model, train the front input layer or the rear output layer that does not conform to the shape of the current target task, while retaining other layers;

[0048] b) When the target task domain dataset has the characteristics of large data volume but different data distribution from the source task domain: based on the constructed model architecture, use the target task domain dataset to train the entire model, that is, relearn a new model.

[0049] c) When the target task domain dataset has the characteristics of small data volume and different data distribution from the source task domain: based on the pre-trained model, train the front input layer or the rear output layer that does not conform to the shape of the current target task, while retaining other layers;

[0050] d) When the target task domain dataset has a small data volume but a data distribution similar to that of the source task domain: directly use the pre-trained model without any training process.

[0051] A deep learning-based indoor dynamic temperature field prediction system includes:

[0052] The information sensing module is used to collect thermal disturbance information of the source task and the target task, temperature field node information of the source task, spatial structure information of the source task, and basic physical information of the target task through a ubiquitous sensing sensor network.

[0053] The pre-training module is used to clean and process the source task thermal disturbance information and temperature field node information collected by the information perception module through the data preprocessing module, and process the source task spatial structure information through the spatial grid partitioning sub-module. Then, the processed data is used to generate the source task dataset through the source task domain large-scale training set construction sub-module. The dataset is input into the dynamic temperature field prediction network module and then the pre-trained model is generated through the pre-trained model generation sub-module.

[0054] The spatial structure feature library module is used to collect and store various architectural interior spatial structure features and form a spatial structure feature library;

[0055] The dynamic temperature field prediction module is used to construct a deep neural prediction network for dynamic temperature fields.

[0056] The transfer learning module is used to establish a mapping relationship between the source task and the target task, realize the model transfer from the source task to the target task, and obtain the temperature field prediction results.

[0057] Furthermore, the step transfer learning module includes:

[0058] The target task spatial structure feature generation submodule is used to perform grid search on the basic physical information of the target task and integrate learning with the spatial feature library to output the initial target task spatial structure features.

[0059] The spatial grid re-division submodule is used to re-divide the initial target task spatial structure features into a spatial grid.

[0060] The data preprocessing submodule preprocesses the thermal disturbance information of the target task obtained by the information perception module.

[0061] The target task domain training set construction submodule is used to generate the target task dataset by combining the processed spatial structure features and thermal perturbation information.

[0062] The pre-trained model fine-tuning submodule is used to fine-tune the pre-trained model based on the distribution characteristics of the target task dataset;

[0063] The dynamic temperature field generation submodule is used to output the dynamic temperature field prediction tensor results based on the fine-tuned pre-trained model.

[0064] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described indoor dynamic temperature field prediction method.

[0065] Compared with the prior art, the present invention has at least the following beneficial technical effects:

[0066] This method employs a dual-driven approach, combining data and mechanism. On one hand, it fully leverages the advantages of data-driven approaches to construct a deep convolutional residual network model based on deep learning, thereby reducing model complexity and the need for prior knowledge while ensuring prediction accuracy. On the other hand, it introduces mechanistic information to enhance the interpretability and generalization ability of the data-driven model. Simultaneously, by constructing and applying physically meaningful boundary constraints, it keeps the prediction results within the range allowed by physical reality, thereby improving the reliability and accuracy of the prediction results.

[0067] This method utilizes transfer learning to establish a mapping relationship between the source task and the target task, and achieves transfer by fine-tuning the pre-trained model, thereby improving the model's generalization ability. Furthermore, when predicting the target task, a spatial structure feature library is constructed, and ensemble learning is used to generate spatial structure features for the target task. These features are then iteratively optimized using the temperature field prediction error as a loss function. When the accuracy reaches the expected level, the optimized spatial structure features are added to the feature library for expansion. This allows the method to further improve generalization ability and reduce the requirement for prior knowledge.

[0068] This method uses a spatiotemporal deep convolutional residual network as the prediction model, fully considering various complex factors affecting the indoor temperature distribution of buildings, including spatial dependence, temporal dependence, thermal disturbance effects, and event-related factors. Spatial dependence refers to the fact that temperature changes in each region of the indoor temperature field are influenced by thermal interactions with its neighboring regions and other distant regions. Temporal dependence refers to the fact that due to the dynamic changes in indoor thermal processes, the temperature distribution over a certain period is affected by continuous time periods, including short-term and long-term. Thermal disturbance effects refer to the influence of various indoor and outdoor thermal disturbance factors, manifested in the form of heat convection, heat conduction, and heat radiation. Event-related factors refer to the fact that certain human-driven events and other variables may also affect the indoor temperature distribution. While fully considering the above influencing factors, highly relevant, necessary, and concise features are constructed through efficient feature engineering as input to the deep learning network, achieving dynamic, accurate, and rapid prediction of the indoor temperature field.

[0069] Furthermore, this method visualizes the dynamic temperature field prediction results using a three-dimensional cloud map, making the results more visually appealing and easier to analyze and apply.

[0070] In summary, this method achieves an optimal balance between evaluation metrics such as prediction accuracy, computational complexity, prior information requirements, and generalization ability. It also enables visualization, improving the overall performance of building indoor temperature field prediction methods and facilitating further analysis and application. Attached Figure Description

[0071] Figure 1 A flowchart of a deep learning-based method for predicting dynamic indoor temperature fields in buildings;

[0072] Figure 2 This is a schematic diagram of a deep learning-based building indoor dynamic temperature field prediction system.

[0073] Figure 3 A schematic diagram of the dynamic temperature field deep neural prediction network structure;

[0074] Figure 4 This is a visualized 3D cloud map of the indoor dynamic temperature field of a room in an office building in Xi'an, predicted based on the above scheme. Detailed Implementation

[0075] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0076] Example 1

[0077] Reference Figure 1 As shown, an embodiment of the present invention provides a method for predicting dynamic indoor temperature fields in buildings based on deep learning, specifically including the following steps:

[0078] S1. Obtain the sensing information of the source task, including thermal disturbance information, temperature field node information and spatial structure information;

[0079] In this context, the source task refers to a regression or classification task in a known domain during transfer learning. In this embodiment, it refers to predicting the temperature field of a known building.

[0080] Thermal disturbance information refers to various thermal disturbance factors that affect indoor temperature distribution. In thermodynamics, these factors influence changes in the indoor temperature field through heat convection, heat conduction, and heat radiation. Since this solution is essentially a time-series prediction problem, the information collected in this embodiment is time-series, mainly including various indoor heat source information, outdoor thermal disturbance information, other disturbance variables, and event disturbance information. Indoor heat source information includes the heat consumption of various heat-generating devices, per capita heat dissipation, and furniture surface temperature, etc. Outdoor thermal disturbance information includes outdoor temperature and solar radiation, etc. Other disturbance variables include indoor and outdoor humidity, indoor carbon dioxide, and indoor fine particulate matter concentration, and event disturbance information includes event states such as turning lights on and off, people entering and exiting, and turning air conditioners on and off.

[0081] Temperature field node information refers to the establishment of a temperature sensor network on a known building, that is, in each A temperature sensor is deployed on a cubic grid, which serves as a node in the temperature field, and the actual indoor temperature field is obtained through measurement. In the diagram, S represents the source task domain to which the cube belongs, r represents the resolution at which the cube can represent the room space division, and w, d, and h represent the length, width, and height, respectively. This indicates that the length, width, and height dimensions of the source task domain are respectively... The resolution.

[0082] Spatial structure information includes basic physical information of rooms in a known building and spatial distribution information of heat sources. Basic physical information of rooms includes room envelope type, room size, type and number of heat sources, etc. Spatial distribution information of heat sources (i.e., layout of indoor furniture, heating equipment, HVAC and lighting, etc.).

[0083] S2. Perform data preprocessing and spatial grid partitioning on the perceptual information of the source task obtained in S1, and construct a large-scale dataset of the source task domain. This step specifically includes the following steps:

[0084] S201. The various thermal disturbance information and temperature field node information of the source task obtained in S1 are preprocessed. In this embodiment, the preprocessing work mainly includes exploratory data analysis, cleaning of abnormal and missing data, data standardization, and then feature selection based on correlation analysis, etc., to finally obtain the indoor heat source characteristics. Outdoor thermal disturbance characteristics Other characteristics of disturbance variables Event disturbance characteristics Temperature field nodal characteristics Y t Where t = n-1-r,...,n-1, t represents time t, r represents the length of the time series, and n represents the current time;

[0085] S202. Perform spatial grid partitioning on the spatial structure information of the source task obtained in S1.

[0086] Specifically, based on the room's length, width, and height dimensions, it is set to a resolution greater than 0 and smaller than the room's spatial dimensions. The area is divided into W×D×H grid regions, where W×D×H are scalars representing the number of grids. For example, a room with dimensions of 5 meters (length), 4 meters (width), and 3 meters (height) can be divided into grid regions according to resolution. The system is divided into W×D×H=5×4×3=60 regional grids. (w,d,h) represents the region located in row w, column d, and layer h. In the source task, each grid is equipped with a temperature sensor.

[0087] Based on the predefined grid area and the spatial distribution information of heat sources, assuming there are M types of heat sources indoors, the spatial distribution of heat sources at time t can be represented as follows: in Let (w1, d1, h1) represent the grid coordinates of the m-th heat source, and (w1, d1, h1) represent the spatial coordinates of the first grid cell occupied by the m-th heat source. x ,d x ,h x () represents the spatial coordinates of the x-th grid it occupies; further, the spatial distribution of the heat source is processed into a spatial tensor form, represented as: t represents time t, r represents the length of the time series, and n represents the current time. It is a four-dimensional tensor representing the spatial location of the m-th type of heat source at time t. The first dimension represents time, and the other three dimensions represent space. Specifically, it is defined as:

[0088]

[0089] in, It represents the spatial occupancy status of the m-th type of heat source in the (w,d,h) grid at time t. 1 indicates that it occupies the grid at this time, and 0 otherwise.

[0090] S203. Integrate and process the data obtained from S201 and S202 to construct a large-scale dataset of the source task domain. The first four items represent input features, and the last item represents label features.

[0091] X n-1-r→n-1 It is to identify the characteristics of the heat source and the spatial distribution tensor L of the heat source t The integrated heat source spacetime tensor [X] n-1-r ,X n-r ,...,X t ,...,X n-1 ],in The spacetime tensor of the heat source at time t is defined as:

[0092]

[0093]

[0094]

[0095]

[0096] in, Let m be the spatial thermal characteristic distribution tensor of the m-th heat source at time t. This represents its distribution within the (w,d,h) grid; Thermal characteristic data representing the m-th type of heat source at time t In this embodiment, it is in the form of temperature or heat dissipation; (L t ) m This represents the spatial location distribution of the m-th type of heat source at time t; The thermal characteristics of the grid that is not occupied by the heat source at time t can be represented by average room temperature or zero heat loss in this embodiment.

[0097] S3. Collect spatial structure information to form a spatial structure feature database;

[0098] S4. Construct a dynamic temperature field deep neural prediction network, and use the large-scale dataset of the source task domain obtained in S2 as the training set to train the dynamic temperature field deep neural network to generate a pre-trained model.

[0099] In this embodiment, the structure of the dynamic temperature field deep neural prediction network is shown in the attached figure. Figure 3 As shown, this network is a three-dimensional dynamic spatiotemporal prediction model that predicts the temperature field distribution by inputting highly relevant, necessary, and concise features. It mainly consists of three parts: a deep convolutional residual network for training the spatiotemporal tensor of the heat source (…). Figure 3 (Top left part), feature extraction structure used for training other thermal disturbance factors ( Figure 3 The upper right part) and the predictor used to output the prediction results ( Figure 3 (Lower right part) The outputs of the deep convolutional residual network and the feature extraction structure are both connected to the input of the predictor.

[0100] Reference Figure 3In the upper left part, the deep convolutional residual network structure consists of three modules: a proximity module, a trend module and a periodic module, which are respectively used to capture the proximity, trend and periodicity of temporal features. The difference between these three parts is that they are respectively input to time series representing short-term, near-term and long-term histories, which is achieved in this embodiment by dividing the time axis into three corresponding segments. The time series for proximity, trend and periodicity can be respectively defined as and where d c , d e , d p are the sampling durations, 1, e, p are the corresponding time sampling intervals, 0<1<e<p<n. These three parts share a similar network structure, which is composed of several 3D convolutional layers and residual units. The 3D convolution of the l-th layer is defined as:

[0101] Y (l) =α(X (l) *W (l) +b (l) )

[0102] where X (l) and Y (l) are the input and output of the l-th layer of the network, W (l) and b (l) are the weight and bias parameters in the l-th layer, * represents the 3D convolution operation, and α(·) is the activation function operation.

[0103] The residual unit of the l-th layer is defined as:

[0104] X (l+1) =X (l) +F(X (l) ,θ (l) ), l=1,...,L

[0105] where X (l) and X (l+1) are the input and output of the l-th residual unit; F(·) in this embodiment may be "ReLu+BatchNorm+Conv3D" or other structures capable of realizing layer nonlinearization, normalization and feature extraction operations, and L is the total number of layers of the residual units.

[0106] The outputs of the three parts of proximity, trend and periodicity are fused through feature matrix matching, which is defined as:

[0107]

[0108] where X Fu is the fused output; M c , M e , Mp These are matrices that are respectively matched with proximity, trend and periodic time features, and are learnable parameters in the model training process; This represents the matrix dot product.

[0109] Finally, the output of feature matrix matching is passed through a convolutional long short-term memory (ConvLstm) network module to achieve temporal multi-step prediction. The final output of this part can be represented as X. D ConvLstm is a deep learning model that performs convolutional operations on a Long Short-Term Memory (LSTM) network. For details, please refer to the paper "Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting".

[0110] The feature extraction structure used for training other thermal disturbance factors mainly includes three types of input features: outdoor thermal disturbance features, other disturbance features, and event disturbance features. These three types of features are merged along the feature axis and then extracted by the feature extraction unit. In this embodiment, the feature extraction unit can consist of a fully connected neural network layer, an activation function, and a Dropout operation, or it can consist of other types of feature extraction structures. Then, a fully connected neural network layer is connected, and the result is non-linearly processed by the activation function. Finally, the result is reshaped to be similar to X. D Output X of the same shape O .

[0111] The outputs of the deep convolutional network module and the feature extraction module are fused and then nonlinearized to obtain the predicted value of the temperature field. In this embodiment, it is represented as: Sigmoid() is a commonly used non-linear activation function. This represents matrix addition. Then, a penalty factor-based optimization strategy is introduced to impose physically meaningful boundary constraints on the model, keeping the variables within physically permissible limits. This property is achieved by applying inequality constraints to each time step t using a penalty function:

[0112]

[0113]

[0114] in, These are the upper and lower bounds of the constraints, and the slack variables. and Indicates the degree of violation of constraints.

[0115] Making a predictor work requires defining a loss function and setting an optimizer. Deep learning open-source frameworks like PyTorch and TensorFlow allow direct access to and configuration of various optimizers, such as SGD, Adam, and RMSProp. This example defines a loss function with a regularization penalty term:

[0116]

[0117] Where l(θ) represents the loss function, and θ represents the parameters that need to be learned during model training; Smooth L1 (·) represents the Huber loss function, which is a combination of mean absolute error and mean square error; the second term on the right-hand side of the equation represents the penalty factor, where Q ineq It is the regularization coefficient, obtained by adding weighted slack variables. Penalize violations of inequality constraints in the output boundary conditions.

[0118] S5. Acquire the perception information of the target task, which includes thermal disturbance information and basic physical information.

[0119] The target task refers to regression or classification tasks in a new domain during transfer learning. In this embodiment, it refers to predicting the temperature field of an unknown building. Thermal disturbance information refers to various thermal disturbance factors that affect the indoor temperature distribution. In thermodynamics, these factors affect changes in the indoor temperature field in the form of heat convection, heat conduction, and heat radiation. Since this solution is essentially a time-series prediction problem, the information collected in this embodiment is time-series, mainly including various indoor heat source information, outdoor thermal disturbance information, other disturbance variables, and event disturbance information. Indoor heat source information includes the heat consumption of various heat-generating devices, per capita heat dissipation, and furniture surface temperature. Outdoor thermal disturbance information includes outdoor temperature and solar radiation. Other disturbance variables include indoor and outdoor humidity, indoor carbon dioxide, and indoor fine particulate matter concentration. Event disturbance information includes event states such as turning lights on and off, people entering and exiting, and turning air conditioners on and off.

[0120] Basic physical information includes room envelope type, room size, type and number of heat sources, etc.

[0121] S6. Based on the basic physical information obtained in S5 and the spatial structure feature library, perform integrated learning to generate the spatial structure features of the target task.

[0122] In this embodiment, distance-based clustering, such as minimax clustering, nearest neighbor clustering, and K-means clustering, is first performed on the basic physical information and spatial structure feature library of the target task. Clustering yields the distances between the target task's physical information and known spatial features in the feature library. Weights are assigned based on these clustering distances, followed by ensemble learning. Finally, a grid search for the spatial locations of heat sources is performed to generate initialized spatial structure features for the target task. These features are a collection of basic physical information of the rooms in the target task domain and information on the spatial distribution of heat sources. The basic physical information of the rooms in the target task domain includes the room envelope type, room size, and type and quantity of heat sources, serving as the input for this ensemble learning step. The spatial distribution information of heat sources includes the layout information of indoor furniture, heating equipment, HVAC, and lighting, serving as the output for this ensemble learning step.

[0123] S7. Process the perceptual information of the target task obtained in S5 and construct the target task domain dataset.

[0124] S701. Perform data preprocessing on the thermal disturbance information obtained in S5. In this embodiment, the preprocessing mainly includes exploratory data analysis, cleaning of abnormal and missing data, data standardization, feature selection, etc., and finally obtain the heat source characteristics. Outdoor thermal disturbance characteristics Other characteristics of disturbance variables Event disturbance characteristics Where t = n-1-r,...,n-1;

[0125] S702. Perform spatial mesh re-division processing on the spatial structure features generated in S6;

[0126] To utilize the pre-trained model, the input to the target task needs to maintain the same shape as the source task. Therefore, the spatial structure of the target task needs to be re-divided according to the shape of the source task's spatial grid. In this embodiment, it is assumed that the rooms in the source task are arranged at an appropriate resolution. If the target room is divided into W×D×H grid areas, then the room in the target task should also be divided into W×D×H grid areas. This grid redivision can be mathematically described as follows:

[0127]

[0128]

[0129]

[0130] in, It is the grid resolution of the target task, (w T ,d T ,h T ) and (wS ,d S ,h S These represent the length, width, and height dimensions of the room for the target task and the room for the source task, respectively. Therefore, any target task room of any size can be transformed into input that the pre-trained model can accept through this process.

[0131] Based on the re-divided mesh and the spatial structure features learned by S6, assuming there are M types of heat sources indoors, the spatial distribution of the heat sources at time t can be represented as follows: in Let represent the grid coordinates of the m-th type of heat source. Further, the spatial distribution of the heat sources is processed into a spatial tensor form, represented as: in It is a four-dimensional tensor representing the spatial location of the m-th type of heat source at time t, with the first dimension representing time. Specifically, it is defined as:

[0132]

[0133] in, It represents the spatial occupancy status of the m-th type of heat source in the (w,d,h) grid at time t. 1 indicates that it occupies the grid at this time, and 0 otherwise.

[0134] S703. Integrate and process the data obtained from S701 and S702 to construct a large-scale dataset for the target task domain. These four items all represent input features.

[0135] XT n-1-r→n-1 It is to identify the characteristics of the heat source and the spatial distribution tensor of heat source LT t The integrated heat source spacetime tensor [XT] n-1-r ,XT n-r ,...,XT t ,...,XT n-1 ],in The spacetime tensor of the heat source at time t is defined as:

[0136]

[0137]

[0138]

[0139]

[0140] in, Let m be the spatial thermal characteristic distribution tensor of the m-th heat source at time t. This represents its distribution within the (w,d,h) grid; Thermal characteristic data representing the m-th type of heat source at time t In this embodiment, it is in the form of temperature or heat loss;

[0141] (LT t ) m This represents the spatial location distribution of the m-th type of heat source at time t; The thermal characteristics of the grid that is not occupied by the heat source at time t can be represented by average room temperature or zero heat loss in this embodiment.

[0142] S8. Fine-tune the pre-trained model obtained in S4 according to the characteristics of the target task domain dataset, and use the fine-tuned pre-trained model to output the dynamic temperature field prediction results.

[0143] Among them, the characteristics of the target task domain dataset and the fine-tuning strategies for pre-training are divided into the following categories depending on the situation:

[0144] (1) The target task domain dataset exhibits characteristics of large data volume and data distribution similar to that of the source task domain. The fine-tuning strategy applicable to this scenario is: based on the pre-trained model, train some front input layers or rear output layers that do not conform to the shape of the current target task, while retaining other layers. The layers that need to be trained depend on the target task and the model architecture.

[0145] (2) The target task domain dataset exhibits a large data volume, but its data distribution differs from that of the source task domain. The appropriate fine-tuning strategy for this scenario is to train the entire model using the target task domain dataset, based on the constructed model architecture; that is, to relearn a new model.

[0146] (3) The target task domain dataset has the characteristics of small data volume and different data distribution from the source task domain. In this scenario, the appropriate strategy is to train some front input layers or rear output layers that do not conform to the shape of the current target task based on the pre-trained model, while retaining other layers. At the same time, attention should be paid to the balance between the number of trained layers and the number of retained layers, as well as the number of training rounds. Too much or too deep training may lead to overfitting, which is more difficult than fine-tuning in case (1).

[0147] (4) The target task domain dataset has the characteristics of small data volume, but the data distribution is similar to that of the source task domain. The fine-tuning strategy applicable to this scenario is to directly use the pre-trained model without any training process.

[0148] S9. Calculate the error between the predicted value of a node at a certain moment and its corresponding actual measured value in the dynamic temperature field prediction results obtained in S8. As a metric for measuring accuracy, a loss function is constructed to determine whether the current error is less than the expected value, i.e., e≤ζ. If yes, proceed to the next step and return to S3 to collect the learned spatial structure features into the spatial structure feature library; otherwise, return to S6 to iteratively optimize the spatial structure features, i.e., optimize and adjust the cluster-based weight allocation and the grid search for the spatial location of the heat source in the direction of loss reduction.

[0149] S10. Visualize the dynamic temperature field prediction results obtained in S8.

[0150] In this embodiment, the Axes3D object is imported using Python's mpl_toolkits.mplot3d to represent the thermal field in the form of a three-dimensional cloud plot, thereby realizing the visualization of the dynamic temperature field.

[0151] Example 2

[0152] Reference Figure 2 This invention provides a deep learning-based system for predicting dynamic indoor temperature fields in buildings, comprising an information perception module, a pre-training module, a building indoor spatial structure feature library module (hereinafter referred to as the spatial structure feature library), a dynamic temperature field deep learning network prediction module (hereinafter referred to as the dynamic temperature field prediction module), a transfer learning module, and a visualization module.

[0153] The information perception module collects thermal disturbance information of the source task and the target task, temperature field node information of the source task, spatial structure information of the source task, and basic physical information of the target task through a ubiquitous sensing sensor network.

[0154] The pre-training module is used to clean and process the source task thermal disturbance information and temperature field node information collected by the information perception module through the data preprocessing module, and to process the source task spatial structure information through the spatial grid division sub-module. Then, the processed data is used to generate the source task dataset through the source task domain large-scale training set construction sub-module. The dataset is input into the dynamic temperature field prediction network module and then the pre-trained model is generated through the pre-trained model generation sub-module.

[0155] The spatial structure feature library module is used to collect and store various architectural interior spatial structural features and form a spatial structure feature library;

[0156] The dynamic temperature field prediction module is used to construct a deep neural prediction network for dynamic temperature fields;

[0157] The transfer learning module is used to establish a mapping relationship between the source task and the target task, realizing model transfer from the source task to the target task. Specifically, firstly, the basic physical information of the target task (including room envelope type, room size, heat source type and quantity, etc.) obtained by the information perception module is subjected to grid search and integrated learning with a spatial feature library. The target task spatial structure feature generation submodule outputs the initial target task spatial structure features, and these features are processed by the spatial grid re-partitioning submodule. At the same time, the thermal disturbance information of the target task obtained by the information perception module is preprocessed by the data preprocessing submodule, and then... The processed spatial structure features and thermal disturbance information are used to generate the target task dataset through the target task domain training set construction submodule. Based on the distribution characteristics of the target task dataset, the pre-trained model is fine-tuned through the pre-trained model fine-tuning submodule. Finally, the dynamic temperature field generation submodule outputs the dynamic temperature field prediction tensor result. The error between the value of a certain node in the prediction result and the measured value of that node is calculated as the accuracy index and a loss function is established. If the accuracy does not reach the preset value, the spatial structure features of the target task are continuously iterated and optimized until the accuracy requirement is met. The spatial structure features learned at this time are then input into the spatial structure feature library to expand the library.

[0158] The visualization module is used to visualize the obtained dynamic temperature field prediction tensor in the form of a three-dimensional cloud map.

[0159] like Figure 4 This is a visualized 3D cloud map of the indoor dynamic temperature field of a room in an office building in Xi'an, predicted based on the above scheme. The map shows the dynamic changes of the temperature field between two adjacent time steps, which can vividly illustrate how the indoor temperature is distributed, so as to carry out further analysis and building optimization control based on the temperature field.

[0160] To further evaluate and compare the performance of different methods on indoor temperature field prediction tasks across different datasets, we established several evaluation metrics and constructed three datasets for result validation through experiments in different rooms, as shown in Table 1. These metrics considered prediction accuracy, computation time, prior knowledge, and generalization ability. In Table 1, the third column represents the input information required to predict indoor temperature, i.e., prior knowledge. The fourth column indicates whether these methods can be transferred to these datasets, which measures the generalization ability of each method. RMSE and resolution are used to measure prediction accuracy. As shown in Table 1, this method provides more spatial structure information than the data-driven method in terms of required prior knowledge. This mechanistic information is beneficial for improving the interpretability of the data-driven model. Furthermore, the heat consumption of the heating equipment required by this method is easier to obtain than the temperature information of the heating equipment required by the data-driven model. Compared with the CFD model, this method does not require complex boundary conditions and prior knowledge. Regarding generalization ability, machine learning methods other than ARIMA and ANN have significant differences in data distribution from the source task domain, making transfer difficult on LR and BS sets. ANN, however, can fine-tune the pre-trained model based on the neural network structure, thus achieving transfer. CFD models are unique, requiring remodeling for each new scenario before running the simulation model, making transfer impossible. This method can achieve transfer through fine-tuning the pre-trained model. Although machine learning methods perform best in terms of RMSE, their resolution is too high, meaning they can only predict the temperature of a single node in the temperature field to represent the entire indoor temperature distribution. While CFD models have the best resolution and lower RMSE error, their average computation time is the longest and increases with increasing resolution. This method achieves a comprehensive prediction accuracy second only to CFD, considering both RMSE and resolution, while its average computation time, even including fine-tuning, is significantly shorter than CFD. In summary, Table 1 shows that this method achieves the optimal trade-off between prediction accuracy, computation time, required prior knowledge, and generalization performance compared to other methods.

[0161] Table 1

[0162]

[0163] Example 3

[0164] Embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, for performing the aforementioned deep learning-based method for predicting dynamic indoor temperature fields in buildings. The system modules can be functionally divided according to the above method; for example, each function can be divided into separate functional modules, or two functions can be integrated into one processing module. The integrated module can be implemented in hardware or as a software functional module. It should be noted that the module division in this embodiment is illustrative and represents only one logical functional division; other division methods may be used in actual implementation.

[0165] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks, SSDs).

[0166] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed invention. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0167] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely exemplary descriptions of the invention as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A deep learning-based method for predicting indoor dynamic temperature fields, characterized in that, Includes the following steps: S1. Construct a dynamic temperature field deep neural prediction network, and input the large-scale dataset of the source task domain as the training set into the dynamic temperature field deep neural network for training to generate a pre-trained model. S2. Acquire the perception information of the target task, which includes thermal disturbance information and basic physical information. S3. Based on the spatial structure feature library and the basic physical information obtained in S2, perform integrated learning to generate the spatial structure features of the target task; S4. Process the perception information of the target task obtained in S2 and construct the target task domain dataset. S5. Fine-tune the pre-trained model obtained in S1 according to the characteristics of the target task domain dataset, and output the dynamic temperature field prediction results. S6. Calculate the error between the predicted value of a node at a certain moment and its corresponding actual measurement value in the dynamic temperature field prediction result obtained in S5 as an indicator of accuracy, and construct a loss function to determine whether the error at this time is less than the expected value: if so, output the dynamic temperature field prediction result and collect the learned spatial structure features into the spatial structure feature library; otherwise, return to S3 and iteratively optimize the spatial structure features. In S3, the spatial structure feature library is constructed through the following steps: SA1. Obtain the sensing information of the source task, the sensing information including thermal disturbance information, temperature field node information and spatial structure information; SA2. Perform data preprocessing and spatial grid partitioning on the acquired source task perception information, and construct a large-scale dataset of the source task domain. SA3. Collect spatial structure information to form a spatial structure feature database; The SA2 includes the following steps: SA201: The various thermal disturbance information and temperature field node information of the source task obtained in SA1 are preprocessed, and then feature selection based on correlation analysis is performed to obtain the indoor heat source characteristics. Outdoor thermal disturbance characteristics Other characteristics of disturbance variables Event disturbance characteristics and temperature field nodal characteristics ,in t represents time t, r represents the length of the time series, and n represents the current time. SA202. The spatial structure information of the source task obtained in SA1 is processed into a spatial grid. A temperature sensor is deployed in each grid of the source task. SA203: Integrate and process the data obtained from SA201 and SA202 to construct a large-scale dataset of the source task domain. The first four items represent input features, and the last item represents label features; It is to identify the characteristics of indoor heat sources and the spatial distribution tensor of heat source Integrated heat source spacetime tensor ,in Let t represent the spacetime tensor of the heat source at time t.

2. The method for predicting indoor dynamic temperature fields based on deep learning according to claim 1, characterized in that, In step S1, the dynamic temperature field deep neural prediction network includes: The deep convolutional residual network is used to train the spatiotemporal tensor of the heat source, the feature extraction structure is used to train other thermal disturbance factors, and the predictor is used to output the prediction results. The output of the deep convolutional residual network and the feature extraction structure are both connected to the input of the predictor. The deep convolutional residual network includes a proximity module, a trend module and a periodic module, which are respectively used to capture the proximity, trend and periodicity of temporal features, and the time series of proximity, trend and periodicity are respectively defined as , and , where , , are the sampling durations, 1 , e, p is the corresponding time sampling interval, 0<1<e<p<n; these three parts share a similar network structure, and the network structure is composed of several three-dimensional convolutional layers and residual units; the three-dimensional convolution of the l th layer is defined as: in, and For the first l The input and output of the layer network, and For the first l Weights and bias parameters in the layer, Represents three-dimensional convolution operation. It is an activation function operation; No. The residual element of a layer is defined as: ； in, and It is the first Inputs and outputs of layer residual units; To enable layer nonlinearity, normalization, and feature extraction operations, L It is the total number of residual units; The outputs of proximity, trend, and periodicity are fused through feature matrix matching and defined as follows: ； in, It is the output of fusion; , , These are matrices that are respectively matched with proximity, trend and periodic time features, and are learnable parameters in the model training process; Represents matrix dot product; Finally, the output of feature matrix matching is passed through a convolutional long short-term memory network module to achieve temporal multi-step prediction. The final output of this part is represented as follows: ; The feature extraction structure includes outdoor thermal disturbance features, other disturbance features, and event disturbance features. These features are merged along the feature axis and then extracted by the feature extraction unit. A fully connected neural network layer is then connected, and the result undergoes non-linear processing using an activation function. Finally, the result is reshaped to resemble... same shape output ; The outputs of the deep convolutional network module and the feature extraction module are fused and nonlinearized to obtain the predicted value of the temperature field. Then, an optimization strategy based on a penalty factor is introduced to impose physically meaningful boundary constraints on the model, so that the variables are kept within the range allowed by physical reality. The loss function of the predictor is a loss function with a regularization penalty term.

3. The method for predicting indoor dynamic temperature fields based on deep learning according to claim 1, characterized in that, S3 includes the following steps: First, distance-based clustering is performed on the basic physical information and spatial structure feature library of the target task. The distance between the physical information of the target task and the known spatial features in the spatial feature library is obtained through clustering. Weights are assigned according to the clustering distance, and then ensemble learning is performed. Finally, grid search of the spatial location of the heat source is performed to generate the initial spatial structure features of the target task. The spatial structure features are a set of basic physical information of the room in the target task domain and spatial location distribution information of the heat source.

4. The method for predicting indoor dynamic temperature fields based on deep learning according to claim 1, characterized in that, S4 includes the following steps: S401. Perform data preprocessing on the thermal disturbance information obtained in S2 to obtain heat source characteristics. Outdoor thermal disturbance characteristics Other characteristics of disturbance variables Event disturbance characteristics ,in ; S402. Perform spatial mesh re-division processing on the spatial structure features generated in S3 so that the input of the target task and the source task maintain the same shape. Based on the re-divided grid and according to the spatial distribution characteristics of heat sources, assuming there are M types of heat sources indoors, the spatial distribution of heat sources at time t is expressed as follows: ,in Represent the grid coordinates of the m-th type of heat source; process the spatial distribution of heat sources into a spatial tensor form, represented as: ,in It is a four-dimensional tensor representing the spatial location of the m-th type of heat source at time t, where the first dimension represents time, and is defined as: in, Is the m-th type of heat source at time t? The space occupancy status of the grid, 1 indicates that it occupies the grid at this time, otherwise it is 0; S403. Integrate and process the data obtained from S401 and S402 to construct a large-scale dataset for the target task domain. , It is to identify the characteristics of the heat source and the spatial distribution tensor of heat source Integrated heat source spacetime tensor ,in Let t represent the spacetime tensor of the heat source at time t.

5. The method for predicting indoor dynamic temperature fields based on deep learning according to claim 1, characterized in that, In S5, the method for fine-tuning the pre-trained model is as follows: a) When the target task domain dataset exhibits characteristics of large data volume and data distribution similar to that of the source task domain: based on the pre-trained model, train the front input layer or the rear output layer that does not conform to the shape of the current target task, while retaining other layers; b) When the target task domain dataset has the characteristics of large data volume but different data distribution from the source task domain: based on the constructed model architecture, use the target task domain dataset to train the entire model, that is, relearn a new model. c) When the target task domain dataset has the characteristics of small data volume and different data distribution from the source task domain: based on the pre-trained model, train the front input layer or the rear output layer that does not conform to the shape of the current target task, while retaining other layers; d) When the target task domain dataset has a small data volume but a data distribution similar to that of the source task domain: directly use the pre-trained model without any training process.

6. A deep learning-based indoor dynamic temperature field prediction system, used to implement the method of claim 1, characterized in that, include: The information sensing module is used to collect thermal disturbance information of the source task and the target task, temperature field node information of the source task, spatial structure information of the source task, and basic physical information of the target task through a ubiquitous sensing sensor network. The pre-training module is used to clean and process the source task thermal disturbance information and temperature field node information collected by the information perception module through the data preprocessing module, and process the source task spatial structure information through the spatial grid partitioning sub-module. Then, the processed data is used to generate the source task dataset through the source task domain large-scale training set construction sub-module. The dataset is input into the dynamic temperature field prediction network module and then the pre-trained model is generated through the pre-trained model generation sub-module. The spatial structure feature library module is used to collect and store various architectural interior spatial structure features and form a spatial structure feature library; The dynamic temperature field prediction module is used to construct a deep neural prediction network for dynamic temperature fields. The transfer learning module is used to establish a mapping relationship between the source task and the target task, realize the model transfer from the source task to the target task, and obtain the temperature field prediction results.

7. The deep learning-based indoor dynamic temperature field prediction system according to claim 6, characterized in that, The transfer learning module includes: The target task spatial structure feature generation submodule is used to perform grid search on the basic physical information of the target task and integrate learning with the spatial feature library to output the initial target task spatial structure features. The spatial grid re-division submodule is used to re-divide the initial target task spatial structure features into a spatial grid. The data preprocessing submodule preprocesses the thermal disturbance information of the target task obtained by the information perception module. The target task domain training set construction submodule is used to generate the target task dataset by combining the processed spatial structure features and thermal perturbation information. The pre-trained model fine-tuning submodule is used to fine-tune the pre-trained model based on the distribution characteristics of the target task dataset; The dynamic temperature field generation submodule is used to output the dynamic temperature field prediction tensor results based on the fine-tuned pre-trained model.

8. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the indoor dynamic temperature field prediction method as described in any one of claims 1-5.

Images