Temperature prediction method based on graph neural network

Abstract

The invention provides a temperature prediction method based on a graph neural network, and the method comprises the steps: obtaining weather data, inputting the weather data into a user behavior model, and predicting an air conditioner air supply mode; when the predicted air conditioner air supply mode is not consistent with the current air supply mode, the predicted air conditioner air supply mode is analyzed, and target parameters are obtained; the target parameters are input into a temperature distribution prediction model, and a temperature distribution feature prediction result in the room is obtained when the air conditioner air supply mode is predicted to be in a steady state; acquiring temperatures, acquired by a temperature sensor, of different positions in the current air conditioner air supply mode; and according to the temperature distribution characteristic prediction result and the temperature of the corresponding position in the current air conditioner air supply mode, the steady-state prediction temperature of the corresponding position in the predicted air conditioner air supply mode is obtained. By implementing the method, the influence of sudden change of the air supply mode on temperature prediction can be effectively dealt with, so that a user regulates and controls the air supply direction in advance according to the predicted temperature, and accurate regulation and control of the indoor temperature are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of temperature prediction, and specifically relates to a temperature prediction method based on graph neural network. Background Art

[0002] With the rapid development of smart buildings and Internet of Things technologies, intelligent control of air-conditioning systems has become a core research direction for improving energy efficiency and user comfort. Traditional air-conditioning control strategies are usually based on fixed temperature thresholds or simple feedback mechanisms (such as PID control). Such methods rely on local sensor data (such as a single temperature and humidity probe) and are difficult to cope with dynamic temperature distribution changes in complex spaces.

[0003] With the development of neural networks, machine learning models have been introduced into related technologies to predict temperature field distribution. However, existing studies mostly train models based on historical data to predict future temperatures, but do not fully consider sudden interference, such as the disturbance of the prediction results caused by sudden changes in air supply modes. Such interference will cause significant deviations between theoretical predictions and actual temperatures, affecting control accuracy. Therefore, it is urgent to propose a new method to accurately predict the temperature of each location in the room under different air conditioning modes, and achieve precise control of indoor temperature based on this. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a distributed temperature sensor and a temperature prediction method and device based on a graph neural network to meet the needs of achieving precise control of indoor temperature.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] The present invention provides a temperature prediction method based on a graph neural network, comprising: obtaining weather data, inputting the weather data into a user behavior model, and predicting an air-conditioning air supply mode; when the predicted air-conditioning air supply mode is inconsistent with the current air supply mode, parsing the predicted air-conditioning air supply mode to obtain target parameters; inputting the target parameters into a temperature distribution prediction model to obtain a prediction result of the temperature distribution characteristics in the room when the predicted air-conditioning air supply mode is in a steady state; obtaining the temperatures at different positions under the current air-conditioning air supply mode collected by a temperature sensor; and obtaining a steady-state predicted temperature at the corresponding position under the predicted air-conditioning air supply mode based on the temperature distribution characteristic prediction result and the temperature at the corresponding position under the current air-conditioning air supply mode.

[0007] Optionally, the temperature distribution prediction model is a graph attention network model, which includes an input layer, an attention layer and an output layer, wherein the attention layer includes a first attention layer and a second attention layer, and the target parameters are input into the temperature distribution prediction model to obtain a prediction result of the temperature distribution characteristics in the room when the air conditioning air supply mode is predicted to be in a steady state, including: constructing an original edge list according to the target parameters, wherein the original edge list includes node parameters and initial weights, and the node parameters include node position, material property parameters at the location, current temperature and target parameters in the air conditioning air supply mode; inputting the original edge list into the input layer for formatting; inputting the formatted original edge list into the first attention layer, calculating the first attention weights between single-hop nodes according to the original edge list, and obtaining the local heat exchange characteristics of each node; inputting the second edge list with the local heat exchange characteristics into the second attention layer, calculating the second attention weights between multi-hop nodes according to the second edge list, and obtaining the global thermal balance characteristics of each node; inputting the global thermal balance characteristics of each node into the output layer, and the output layer performs a fully connected layer mapping on the global thermal balance characteristics of each node to obtain the prediction result of the temperature distribution characteristics in the room when the air conditioning air supply mode is predicted to be in a steady state.

[0008] Optionally, the original edge list determination process includes: pre-deploying multiple temperature sensors, each temperature sensor serving as a node in the graph attention network; connecting two nodes that meet preset connection conditions to construct an edge, wherein the preset connection condition is that the distance value between the two nodes is less than a preset distance and / or the two nodes are in direct contact through a thermal conductive material; determining the initial weight of the edge based on the material property parameters of the location, the current temperature, and the target parameters in the air supply mode of the air conditioner; and constructing the original edge list based on the node parameters and the initial weight of the corresponding edge.

[0009] Optionally, the first attention layer includes: a first attention head, which extracts the distance value between single-hop nodes in the original edge list after formatting, determines the distance attention weight between the single-hop nodes according to the distance value, and obtains the first attention head output according to the distance attention weight between the single-hop nodes and the corresponding node parameters; a second attention head, which extracts the material attribute parameters of the nodes in the original edge list after formatting, determines the material attention weight between the single-hop nodes according to the material attribute parameters, and obtains the second attention head output according to the material attention weight between the single-hop nodes and the corresponding node parameters; a third attention head, which extracts the target parameters in the air conditioning supply mode of the nodes in the original edge list after formatting, determines the air supply direction attention weight between the single-hop nodes according to the target parameters in the air conditioning supply mode, and obtains the third attention head output according to the air supply direction attention weight between the single-hop nodes and the corresponding node parameters; a first splicing module, which is used to perform feature splicing on the first attention head output, the second attention head output and the third attention head output to obtain local heat exchange features.

[0010] Optionally, the second attention layer includes: a fourth attention head, identifying a heat conduction path that meets preset conditions in a second edge list with local heat exchange characteristics, determining an end node of the heat conduction path, determining a heat conduction attention weight of a starting node of the heat conduction path to an end node according to a preset first attention mechanism, and obtaining a fourth attention head output according to the heat conduction attention weight and node parameters of the end node; a fifth attention head, identifying a thermal barrier area in a second edge list with local heat exchange characteristics, determining a thermal barrier attention weight between a thermal barrier area node and surrounding nodes according to a preset second attention mechanism, and obtaining a fifth attention head output according to the thermal barrier attention weight and node parameters of the surrounding nodes, wherein the surrounding nodes are nodes directly adjacent to the barrier area nodes; and a second splicing module, used to perform feature splicing on the fourth attention head output and the fifth attention head output to obtain a global thermal balance feature.

[0011] Optionally, based on the temperature distribution characteristic prediction results and the temperature of the corresponding position under the current air-conditioning and air-supply mode, a steady-state predicted temperature of the corresponding position under the predicted air-conditioning and air-supply mode is obtained, including: inputting the temperature distribution characteristic prediction results and the temperature of the corresponding position under the current air-conditioning and air-supply mode into a pre-trained temperature prediction model to obtain the steady-state predicted temperature of the corresponding position under the predicted air-conditioning and air-supply mode, wherein the pre-trained temperature prediction model is trained based on different current temperatures, the temperature distribution characteristic prediction results in the room when the air-conditioning and air-supply mode is in a steady state, and the actual temperature value when the air-conditioning and air-supply mode is predicted to be in a steady state as input.

[0012] Optionally, the temperature sensor is a distributed temperature sensor, including a main control module, multiple fixed temperature sensor modules and at least one movable temperature sensor module, the movable temperature sensor module is embedded with a temperature sensing unit, a positioning unit, an intelligent perception unit and an intelligent decision-making unit; obtaining the temperatures of different positions under the current air-conditioning air supply mode collected by the temperature sensor, including: obtaining the spatial information of temperature collection and the preset number of nodes; sending the spatial information of temperature collection and the preset number of nodes to the distributed temperature sensor, so that the distributed temperature sensor main control module allocates the target position for the required temperature collection to the movable temperature sensor module; the movable sensor module plans the movement path according to the intelligent perception unit and the intelligent decision-making unit until it reaches the target position; the movable sensor module collects the temperature of the target position.

[0013] Optionally, the distributed temperature sensor assigns a target position for the required temperature collection to the movable temperature sensor module, including: the distributed temperature sensor main control module obtains the position of each fixed temperature sensor module; calculates the vacant position according to the position of the fixed temperature sensor module, the preset number of nodes and the spatial information of temperature collection, and uses the vacant position as the target position.

[0014] Optionally, based on the position of the fixed temperature sensor module, the preset number of nodes and the spatial information of temperature collection, the vacant position is calculated and the vacant position is used as the target position, including: determining the minimum coverage radius of each node based on the spatial information and the preset number of nodes; obtaining the position of the fixed temperature sensor module, and determining the coverage range represented by the minimum coverage radius of the fixed temperature sensor with the position of the fixed temperature sensor module as the center; determining the vacant area based on the coverage range represented by the minimum coverage radius of the fixed temperature sensor and the spatial information; determining the vacant position based on the vacant area and the coverage range represented by the minimum coverage radius.

[0015] An embodiment of the present invention provides a temperature prediction method based on a graph neural network. According to weather data, the change of the user's air conditioning air supply mode is predicted. When the air supply mode changes, the temperatures at different positions under the current air conditioning air supply mode collected by the temperature sensor and the predicted air conditioning air supply mode are in a steady state. The temperature distribution characteristics in the room are predicted to obtain a steady-state predicted temperature at the corresponding position under the predicted air conditioning air supply mode. This method can effectively deal with the impact of sudden changes in the air supply mode on temperature prediction, and can give a predicted temperature in a steady state after the air supply mode changes, so that the user can understand the actual impact of the change in the air supply mode on the indoor temperature, and can enable the user to adjust the air supply direction according to the predicted temperature in advance to achieve precise control of the indoor temperature.

[0016] Other advantages, objectives and features of the present invention will be described in the following description and will be apparent to those skilled in the art to some extent, or those skilled in the art may be taught from the practice of the present invention. The objectives and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:

[0018] Figure 1 This is a specific example flow chart of a temperature prediction method based on graph neural network in the present invention;

[0019] Figure 2 Schematic diagram of the structure of the graph attention network model in the present invention;

[0020] Figure 3 It is a schematic diagram of the module of the distributed temperature sensor in the present invention. DETAILED DESCRIPTION

[0021] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, it can also be the internal connection of two components, it can be a wireless connection, or it can be a wired connection. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0023] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0024] The embodiment of the present invention provides a temperature prediction method based on a graph neural network. Figure 1 As shown, including:

[0025] S101, obtaining weather data, inputting the weather data into a user behavior model, and predicting an air supply mode of an air conditioner;

[0026] S102, when the predicted air-conditioning air supply mode is inconsistent with the current air supply mode, analyzing the predicted air-conditioning air supply mode to obtain target parameters;

[0027] S103, inputting the target parameters into the temperature distribution prediction model to obtain a prediction result of the temperature distribution characteristics in the room when the air conditioning air supply mode is predicted to be in a steady state;

[0028] S104, obtaining the temperatures at different locations under the current air-conditioning air supply mode collected by the temperature sensor;

[0029] S105, obtaining a steady-state predicted temperature of the corresponding position under the predicted air-conditioning air-supply mode according to the temperature distribution feature prediction result and the temperature of the corresponding position under the current air-conditioning air-supply mode.

[0030] Exemplarily, weather data includes temperature, humidity, wind speed, sunshine intensity, weather conditions, etc., and weather conditions include cloudy, sunny, rainy, etc. Weather data can be obtained from the public data interface of the meteorological department, a meteorological website, or a local meteorological monitoring station. By inputting weather data into the user behavior model, the user's air conditioning air supply mode selection behavior under the corresponding weather can be predicted. The user behavior model can be a decision tree model, a logistic regression model, or a neural network model, and its training process is: first, the user's selection history of air conditioning air supply mode under different weather conditions is collected from the monitoring module installed on the air conditioning equipment, or from the smart home system. The air conditioning air supply mode may include cooling, heating, ventilation, automatic, etc. Taking the user behavior model as a neural network model as an example, the air supply mode selected by the user and the corresponding weather data are input into the neural network model to be trained. The neural network model learns the mapping relationship between weather data and air conditioning air supply mode by adjusting weights and biases. The specific training process is consistent with the existing neural network training process, which will not be repeated here.

[0031] When the predicted air-conditioning air supply mode is inconsistent with the current air-conditioning air supply mode, it means that the air supply mode has changed. In this case, the predicted air-conditioning air supply mode is analyzed to determine the target parameters in the predicted air-conditioning air supply mode. The target parameters may include the air supply direction and temperature under the air-conditioning air supply mode. The target parameters are input into the temperature distribution prediction model to obtain the prediction results of the temperature distribution characteristics in the room when the predicted air-conditioning air supply mode is in a steady state. It should be explained that the temperature distribution characteristic prediction results characterize the characteristics of the temperature distribution under the air-conditioning air supply mode, including temperature gradient characteristics, isotherm characteristics, etc.

[0032] The temperature distribution prediction model can be a graph attention model. When obtaining data for training the graph attention network, the temperature distribution data after reaching steady state under different air-conditioning air supply modes can be collected. The graph attention network can learn the mapping relationship between different air supply modes and steady-state temperature distribution from these data, and then predict the corresponding steady-state temperature distribution when the target parameters (such as air supply direction, temperature, etc.) are given. Specifically, the temperature in the steady-state scene under any air-conditioning air supply mode is obtained by using temperature sensors distributed at different positions in the air-conditioned room, which is used as the graph structure node of the input graph attention network, and the edge represents the heat transfer relationship between different positions. The heat exchange between the area close to the air outlet of the air conditioner and the area far from the air outlet can be represented by the edge. In the graph attention network, a weight is assigned to each edge through the attention mechanism. This weight reflects the degree of influence of one node on another node. For example, the influence of nodes close to the air outlet on nodes far from the air outlet can be represented by a higher attention weight, while nodes with less influence compared to other areas are represented by lower weights. In this way, the graph attention network can learn the temperature distribution of different air-conditioning air supply modes in a steady state, and by learning the mapping relationship between existing different air supply modes and steady-state temperature distribution, it can predict the steady-state temperature distribution characteristics under new and unknown air supply modes.

[0033] After obtaining the predicted results of the temperature distribution characteristics in the room when the predicted air conditioning air supply mode is in a steady state, it is also necessary to rely on the temperatures at different locations under the current air conditioning air supply mode to predict the actual steady-state temperature after changing the air supply mode. Specifically, the graph attention network gives the temperature distribution characteristics in a steady state, but in the actual temperature prediction, it is also necessary to consider the impact of the current air conditioning air supply mode on the actual steady-state temperature. Therefore, in this embodiment, the current temperature and the predicted steady-state temperature distribution characteristics can be used as input features, and input into a pre-trained temperature prediction model to predict the temperature at the corresponding location when the predicted air conditioning air supply mode is in a steady state. The pre-trained temperature prediction model can be a neural network model, such as a convolutional model. For different air supply modes, the temperature prediction model mainly learns the following relationships:

[0034] T=σ 2 (W[T 1 ,T 2 ]+b);

[0035] Where T is the output of the temperature prediction model, σ 2 is a nonlinear activation function, T 1 ,T 2 They represent the temperature predicted by the model and the temperature before prediction, respectively, and W and b represent the learnable parameters.

[0036] The training samples of the temperature prediction model can be the steady-state temperature distribution characteristics predicted by the first air-conditioning air supply mode and the temperature in the second air-conditioning air supply mode. The sample label is the actual temperature value collected by the sensor when the first air-conditioning air supply mode is in a steady state. The specific training process is consistent with the supervised learning process in the prior art and will not be repeated here. It should be explained that the judgment of the steady state can be that in this air-conditioning air supply mode, the temperature of the sensors placed at various positions does not change within a preset time. The preset time can be half an hour or ten minutes. This embodiment does not limit this, and those skilled in the art can determine it as needed.

[0037] An embodiment of the present invention provides a temperature prediction method based on a graph neural network. According to weather data, the change of the user's air conditioning air supply mode is predicted. When the air supply mode changes, the temperatures at different positions under the current air conditioning air supply mode collected by the temperature sensor and the predicted air conditioning air supply mode are in a steady state. The temperature distribution characteristics in the room are predicted to obtain a steady-state predicted temperature at the corresponding position under the predicted air conditioning air supply mode. This method can effectively deal with the impact of sudden changes in the air supply mode on temperature prediction, and can give a predicted temperature in a steady state after the air supply mode changes, so that the user can understand the actual impact of the change in the air supply mode on the indoor temperature, and can enable the user to adjust the air supply direction according to the predicted temperature in advance to achieve precise control of the indoor temperature.

[0038] As an optional implementation, the temperature distribution prediction model is a graph attention network model, such as Figure 2 As shown in the figure, the attention model includes an input layer, an attention layer and an output layer, wherein the attention layer includes a first attention layer and a second attention layer. The target parameters are input into the temperature distribution prediction model to obtain the prediction results of the temperature distribution characteristics in the room when the air-conditioning air supply mode is predicted to be in a steady state, including:

[0039] Construct an original edge list according to the target parameters, wherein the original edge list includes node parameters and initial weights, and the node parameters include node position, material property parameters at the position, current temperature, and target parameters in the air supply mode of the air conditioner;

[0040] Input the original edge list into the input layer for formatting;

[0041] The formatted original edge list is input into the first attention layer, and the first attention weights between single-hop nodes are calculated according to the original edge list to obtain the local heat exchange features of each node;

[0042] The second edge list with local heat exchange features is input into the second attention layer, and the second attention weights between multi-hop nodes are calculated according to the second edge list to obtain the global heat balance features of each node;

[0043] The global thermal balance characteristics of each node are input into the output layer, and the output layer performs a fully connected layer mapping on the global thermal balance characteristics of each node to obtain the predicted results of the temperature distribution characteristics in the room when the air-conditioning air supply mode is predicted to be in a steady state.

[0044] For example, Figure 2 As shown, the graph attention model includes an input layer, an attention layer and an output layer, wherein the attention layer includes a first attention layer and a second attention layer, wherein the first attention layer is used to learn local heat exchange, and the second attention layer is used to model global thermal balance. As the input of the entire model, the original edge list determination process includes: pre-deployment of multiple temperature sensors, each temperature sensor as a node in the graph attention network; connecting two nodes that meet the preset connection conditions to build edges, wherein the preset connection conditions are that the distance value between the two nodes is less than the preset distance and / or that the two nodes are in direct contact through heat-conducting materials; determining the initial weight of the edge according to the material property parameters of the location, the current temperature and the target parameters in the air-conditioning supply mode; and building the original edge list according to the node parameters and the initial weight of the corresponding edge.

[0045] Specifically, the node position in the node parameters is the temperature sensor position, and the material property parameters at the position can be input by the user according to actual conditions, or the temperature sensor at the position can be equipped with an image acquisition module and an image recognition module. The image acquisition module acquires the temperature material image, and the image recognition module identifies the material in the image, and retrieves the pre-stored material property parameters of the material according to the identified material. The material property parameters may include thermal conductivity, specific heat capacity, thermal conductivity, emissivity, absorptivity, and the like.

[0046] When the distance between the two temperature sensors is less than the preset distance and / or the two temperature sensors are in direct contact with each other through the heat-conducting material, it can be considered that the two temperature sensors are connected to form an edge. The preset distance can be 0.5 meters. This embodiment does not limit the preset distance and can be determined according to actual needs. The initial weight of the edge can be determined based on the material property parameters of the location, the current temperature, and the target parameters in the air conditioning supply mode. The details are as follows:

[0047]

[0048] Among them, ω i,j is the initial weight of the edge formed by node i and node j, λ ij represents the equivalent thermal conductivity of the material between two nodes, |T i -T j | represents the absolute value of the temperature difference between nodes i and j, Δx ij represents the distance between node i and node j, P represents the heat flux density, θ ijrepresents the wind direction from node i to node θ air Indicates the wind direction of the air conditioner.

[0049] The input layer receives the original edge list and formats it to convert it into a data format that can be applied and recognized by the model. The formatted original edge list is input into the first attention layer, and the first attention weight between single-hop nodes is calculated based on the original edge list to obtain the local heat exchange characteristics of each node.

[0050] The first attention layer includes: a first attention head, which extracts the distance value between the single-hop nodes in the original edge list after formatting, determines the distance attention weight between the single-hop nodes according to the distance value, and obtains the first attention head output according to the distance attention weight between the single-hop nodes and the corresponding node parameters; a second attention head, which extracts the material attribute parameters of the nodes in the original edge list after formatting, determines the material attention weight between the single-hop nodes according to the material attribute parameters, and obtains the second attention head output according to the material attention weight between the single-hop nodes and the corresponding node parameters; a third attention head, which extracts the target parameters in the air-conditioning air supply mode of the nodes in the original edge list after formatting, determines the air supply direction attention weight between the single-hop nodes according to the target parameters in the air-conditioning air supply mode, and obtains the third attention head output according to the air supply direction attention weight between the single-hop nodes and the corresponding node parameters; a first splicing module, which is used to perform feature splicing on the first attention head output, the second attention head output and the third attention head output to obtain local heat exchange features.

[0051] Specifically, the first attention head takes the Euclidean distance or other distance measurement between nodes as part of the input features, and uses a learnable first parameter matrix. The first parameter matrix determines its optimal parameter value through an optimization algorithm during the model training process, and then maps the distance information to a distance attention weight. For example, a linear transformation is used to convert the distance value into a score, and then the score is converted into a normalized distance attention weight through a softmax function, so that nodes with closer distances have higher weights and nodes with farther distances have lower weights. According to the calculated distance attention weights, the features of neighboring nodes are weighted and summed to obtain the output of the first attention head. In this way, the features of neighboring nodes closer to the current node will have a greater impact on the representation of the current node, thereby capturing the impact of physical distance on heat transfer.

[0052] The second attention head transforms and learns the material property features through a learnable second parameter matrix. The parameter matrix can automatically learn the degree of influence of different material properties on heat transfer. For example, nodes corresponding to materials with high thermal conductivity are given higher weights to indicate their greater importance in heat transfer. By comparing the material property features of the current node and neighboring nodes, the similarity or difference between them is calculated and converted into material attention weights. For example, the dot product or other similarity metric functions are used to calculate the similarity between the material property features of the current node and the neighboring nodes. The higher the similarity, the larger the corresponding material attention weight, which means that the heat transfer between the two nodes is more affected by the material properties, thereby highlighting the role of material property differences in heat transfer. The second attention head is obtained based on the material attention weights between single-hop nodes and the feature vectors corresponding to the node parameters.

[0053] The third attention head uses an asymmetric attention calculation method to capture the asymmetry of heat transfer caused by the air supply direction. For each node, the attention weights of different air supply directions from the current node to the neighbor node and from the neighbor node to the current node are calculated according to the relative position relationship between the air supply direction and the node. For example, if a neighbor node is located downstream of the air supply direction of the air conditioner, the heat transfer from the current node to the neighbor node may be promoted by the air supply, while the heat transfer from the neighbor node to the current node may be hindered to a certain extent. By learning the attention weights in different directions, this asymmetric heat transfer phenomenon can be effectively captured. Furthermore, the third parameter matrix and the fourth parameter matrix can be used to calculate the forward and reverse attention weights, respectively. The two different parameter matrices are pre-trained, and the asymmetric heat transfer process is better modeled by the positive and negative attention weights. According to the attention weights of the air supply direction between single-hop nodes and the eigenvectors corresponding to the node parameters, the output of the third attention head is obtained.

[0054] Finally, the first concatenation module is used to concatenate the first attention head output, the second attention head output, and the third attention head output to obtain a local hot exchange feature. A second edge list is constructed based on the local hot exchange feature. Specifically, the local hot exchange feature data can be added to the original edge list.

[0055] The second edge list with local heat exchange characteristics is input into the second attention layer, and the second attention layer includes: a fourth attention head, identifying a heat conduction path that meets a preset condition in the second edge list with local heat exchange characteristics, determining an end node of the heat conduction path, and determining a heat conduction attention weight of a starting node of the heat conduction path to an end node according to a preset first attention mechanism, and obtaining a fourth attention head output according to the heat conduction attention weight and the node parameters of the end node; a fifth attention head, identifying a thermal barrier region in the second edge list with local heat exchange characteristics, determining a thermal barrier attention weight between a thermal barrier region node and surrounding nodes according to a preset second attention mechanism, and obtaining a fifth attention head output according to the thermal barrier attention weight and the node parameters of the surrounding nodes, wherein the surrounding nodes are nodes directly adjacent to the barrier region nodes; and a second splicing module, used to perform feature splicing on the output of the fourth attention head and the output of the fifth attention head to obtain a global thermal balance feature.

[0056] Specifically, the local heat exchange feature already contains preliminary heat transfer relationship information between nodes. The fourth attention head will calculate the attention weight of the edge information in the second edge list according to the first attention mechanism. The first attention mechanism can refer to the thermal state parameters such as the temperature and heat capacity of the node. If the temperature of the starting node is high and the heat capacity is large, and the temperature of the end node is low, then the heat transfer potential of the starting node to the end node is large, and a higher thermal conductivity attention weight will be obtained. For example, in a system composed of different temperature zones, the node in the high temperature zone is used as the starting node, and the thermal conductivity attention weight of the end node in the low temperature zone will increase due to the temperature difference and heat capacity factors. Finally, these thermal conductivity attention weights are weighted and combined with the node parameters of the end node to obtain the output of the fourth attention head.

[0057] At the same time, the fifth attention head identifies the nodes representing the thermal barrier area based on the node parameters in the second edge list. Since the material properties of these nodes are significantly different from those of other areas, they can be marked and identified by analyzing the material property parameters. The second attention mechanism can be to use historical heat exchange data to analyze the heat transfer between the nodes in the thermal barrier area and the surrounding nodes. If it is found in historical observations that the amount of heat exchange between the nodes in the thermal barrier area and some surrounding nodes is small, it means that the thermal barrier effect between them is better, and a lower thermal barrier attention weight will be given. Finally, according to the obtained attention weights, the features of the surrounding nodes are adjusted to highlight the local temperature differences caused by the thermal barrier area. Specifically, the features of the surrounding nodes can be adjusted by assuming that the surrounding node is i, and the feature vector of the initial node parameters of the node is x i , the adjustment method can be to calculate the new feature vector of the node through the following formula: i '=x i+β j ΔT, β j Represents the thermal barrier area node and its thermal barrier attention weight, and ΔT represents the temperature difference between the surrounding nodes and the thermal barrier area. The adjusted features of all surrounding nodes are summarized to obtain the output of the fifth attention head. Finally, the output of the fourth attention head and the output of the fifth attention head are sent to the second splicing module for splicing to obtain the global thermal balance feature.

[0058] Finally, the global thermal balance characteristics of each node are input into the output layer, and the output layer performs a fully connected layer mapping on the global thermal balance characteristics of each node to obtain the predicted results of the temperature distribution characteristics in the room when the air-conditioning supply mode is predicted to be in a steady state.

[0059] The loss function value of the training process of the above-mentioned attention layer (including the first attention layer and the second attention layer) is determined based on the difference between the output temperature distribution feature prediction result and the distribution feature result obtained by the temperature actually collected by the temperature sensor. Specifically, it can be a mean square error loss function, whose input is the output temperature distribution feature prediction result and the distribution feature result obtained by the temperature sensor.

[0060] The embodiment of the present invention provides a temperature prediction method based on a graph neural network. By constructing multiple attention heads, the influence of node distance, material properties and air conditioning air supply mode on heat transfer is explored, local heat exchange features are extracted, and based on the local heat exchange features, key heat conduction paths are extracted to accurately grasp the heat transfer in the key heat conduction paths, thereby improving the modeling accuracy of the heat transfer process and enabling the model to learn the role of the heat barrier area, thereby more accurately predicting the temperature changes around the heat barrier area. It can be seen that the first attention layer extracts local heat exchange features from multiple basic dimensions, and the second attention layer further focuses on specific heat exchange phenomena on this basis, forming a hierarchical feature extraction structure, which allows the model to gradually and deeply learn the laws of heat exchange and improve the learning and expression capabilities of the model.

[0061] As an optional implementation, the temperature sensor is a distributed temperature sensor, such as Figure 3 As shown, it includes a main control module, multiple fixed temperature sensor modules and at least one movable temperature sensor module, and the movable temperature sensor module is embedded with a temperature sensing unit, a positioning unit, an intelligent perception unit and an intelligent decision-making unit; obtaining the temperature of different positions under the current air-conditioning air supply mode collected by the temperature sensor, including:

[0062] Acquire the spatial information of temperature collection and the preset number of nodes; send the spatial information of temperature collection and the preset number of nodes to the distributed temperature sensor, so that the distributed temperature sensor main control module can allocate the target position of the required temperature collection to the movable temperature sensor module; the movable sensor module plans the movement path according to the intelligent perception unit and the intelligent decision-making unit until it reaches the target position; the movable sensor module collects the temperature at the target position.

[0063] Exemplarily, the main control module serves as the core control unit of the distributed temperature sensor, responsible for coordinating and managing the operation of the entire system. The main control module receives instructions and information from the outside, and controls and schedules the fixed temperature sensor module and the movable temperature sensor module. For example, it can set the working parameters of the sensor (such as sampling frequency, data transmission interval, etc.), receive and process the data collected by the sensor, etc. The fixed temperature sensor module is installed at a predetermined fixed position to continuously monitor the temperature changes in a specific area. The movable temperature sensor module can be a temperature sensor module carrying a mobile device, and the movable temperature sensor module is embedded with a temperature sensing unit, a positioning unit, an intelligent perception unit and an intelligent decision unit. The temperature sensing unit uses a high-precision temperature sensor element, such as a thermocouple, a thermistor, etc.; the positioning unit can use indoor positioning technology (such as Bluetooth positioning, Wi-Fi positioning, ultra-wideband positioning, etc.) to determine the real-time position of the movable temperature sensor module; the intelligent perception unit can perceive the information of the surrounding environment, such as the location of obstacles, the terrain of the environment, etc. It can realize the environmental perception function by installing devices such as cameras, laser radars, ultrasonic sensors, etc.; the intelligent decision unit performs path planning and decision-making based on the environmental information obtained by the intelligent perception unit and the instructions of the main control module. Specifically, an algorithm, such as the Dijkstra algorithm, can be used to calculate the optimal path from the current location to the target location, and dynamically adjust the path according to real-time environmental changes.

[0064] The user can manually input the spatial information of temperature collection and the preset number of nodes through the system's operating interface. The spatial information may include the scope and shape of the space, whether there are special areas, such as obstacle areas, key monitoring areas, etc. Or the system can support the import of electronic map data (such as CAD drawings, GIS maps, etc.), and obtain the spatial information of temperature collection by parsing and processing the map data. Then, the main control module allocates the target position for the required temperature collection to the movable temperature sensor module. The specific allocation method can be to divide the space evenly into several areas according to the spatial information of temperature collection and the preset number of nodes, each area corresponds to a target position, and then these target positions are allocated to the movable temperature sensor module. After the movable temperature sensor module moves to the corresponding position, it collects the ambient temperature.

[0065] As an optional implementation, the distributed temperature sensor assigns a target position for collecting the required temperature to the movable temperature sensor module, including: the distributed temperature sensor main control module obtains the position of each fixed temperature sensor module; according to the position of the fixed temperature sensor module, the preset number of nodes and the spatial information of temperature collection, the vacant position is calculated, and the vacant position is used as the target position.

[0066] Specifically, according to the position of the fixed temperature sensor module, the preset number of nodes and the spatial information of temperature collection, the vacant points are calculated and the vacant points are used as the target positions, including: determining the minimum coverage radius of each node according to the spatial information and the preset number of nodes; obtaining the position of the fixed temperature sensor module, and determining the coverage range represented by the minimum coverage radius of the fixed temperature sensor with the position of the fixed temperature sensor module as the center; determining the vacant area according to the coverage range represented by the minimum coverage radius of the fixed temperature sensor and the spatial information; determining the vacant point according to the coverage range represented by the vacant area and the minimum coverage radius.

[0067] Exemplarily, according to the spatial information and the preset number of nodes, the method of determining the minimum coverage radius of each node can be to determine the spatial area of ​​the temperature to be detected according to the spatial information, divide the spatial area by the preset number of nodes to obtain the average area, solve the radius of the minimum circle covering the average area in a circle manner, and use the radius as the minimum coverage radius. Then, the position of the fixed temperature sensor module is obtained, and the minimum coverage range of each fixed temperature sensor module is determined with the position of the fixed temperature sensor module as the center of the circle and the minimum coverage radius as the radius. By comparing the coverage range of all fixed temperature sensors with the temperature area to be detected determined by the entire spatial information, an uncovered area can be obtained, and the uncovered area is defined as a vacant area. Therefore, it is necessary to determine a specific vacant point in the vacant area, and send the point information to the movable temperature sensor module so that it moves to the corresponding position for temperature detection. The method of determining a specific vacant point in the vacant area can be that when there are multiple vacant areas, the vacant areas are sorted according to the area, and the larger the area of ​​the vacant area, the higher the priority of determining the vacant point. In areas where the area of ​​the vacant region is larger than the minimum coverage range, the first attention layer can be used to extract local heat exchange features from multiple basic dimensions according to the grid method with twice the minimum coverage radius. The second attention layer further focuses on specific heat exchange phenomena on this basis, forming a hierarchical feature extraction structure. This structure allows the model to gradually and deeply learn the laws of heat exchange, from macroscopic multi-dimensional information to microscopic specific heat exchange mechanisms, improving the learning and expression capabilities of the model. Divide it as the length and width of the grid to obtain each grid, and take the center of each grid as a vacant point. For areas where the area of ​​the vacant region is smaller than the minimum coverage range, when the number of nodes meets the preset number, such areas are not processed. When the number of nodes is less than the preset number, a position is randomly selected in such areas as a vacant point.

[0068] An embodiment of the present invention provides a temperature prediction method based on a graph neural network. Distributed sensors are added to achieve unified control of sensors. At the same time, a movable sensor module is introduced to flexibly fill in vacant positions, greatly improving data acquisition efficiency.

[0069] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A temperature prediction method based on graph neural network, characterized in that: include: Obtain weather data, input the weather data into the user behavior model, and predict the air conditioning supply mode; When the predicted air-conditioning air supply mode is inconsistent with the current air supply mode, the predicted air-conditioning air supply mode is analyzed to obtain the target parameters; The target parameters are input into the temperature distribution prediction model to obtain the prediction results of the temperature distribution characteristics in the room when the air-conditioning air supply mode is predicted to be in a steady state. The temperature distribution prediction network is a graph attention network. Get the temperature at different locations under the current air-conditioning air supply mode collected by the temperature sensor; According to the temperature distribution characteristic prediction result and the temperature of the corresponding position under the current air conditioning air supply mode, the steady-state predicted temperature of the corresponding position under the predicted air conditioning air supply mode is obtained.

2. A temperature prediction method based on graph neural network according to claim 1, characterized in that: The temperature distribution prediction model is a graph attention network model, which includes an input layer, an attention layer, and an output layer. The attention layer includes a first attention layer and a second attention layer. The target parameters are input into the temperature distribution prediction model to obtain the prediction results of the temperature distribution characteristics in the room when the air-conditioning air supply mode is predicted to be in a steady state, including: Construct an original edge list according to the target parameters, wherein the original edge list includes node parameters and initial weights, and the node parameters include node position, material property parameters at the position, current temperature, and target parameters in the air supply mode of the air conditioner; Input the original edge list into the input layer for formatting; The formatted original edge list is input into the first attention layer, and the first attention weights between single-hop nodes are calculated according to the original edge list to obtain the local heat exchange features of each node; The second edge list with local heat exchange features is input into the second attention layer, and the second attention weights between multi-hop nodes are calculated according to the second edge list to obtain the global heat balance features of each node; The global thermal balance characteristics of each node are input into the output layer, and the output layer performs a fully connected layer mapping on the global thermal balance characteristics of each node to obtain the predicted results of the temperature distribution characteristics in the room when the air-conditioning air supply mode is predicted to be in a steady state.

3. A temperature prediction method based on graph neural network according to claim 2, characterized in that: The original edge list determination process includes: Pre-deploy multiple temperature sensors, each of which serves as a node in the graph attention network; Connecting two nodes that meet a preset connection condition to construct an edge, wherein the preset connection condition is that the distance between the two nodes is less than a preset distance and / or the two nodes are in direct contact with each other through a heat conductive material; Determine the initial weight of the edge based on the material property parameters of the location, the current temperature, and the target parameters in the air conditioning mode; Construct the original edge list based on the node parameters and the initial weights of the corresponding edges.

4. The temperature prediction method based on graph neural network according to claim 2 is characterized in that: The first attention layer includes: The first attention head extracts the distance value between the single-hop nodes in the formatted original edge list, determines the distance attention weight between the single-hop nodes according to the distance value, and obtains the first attention head output according to the distance attention weight between the single-hop nodes and the corresponding node parameters; The second attention head extracts the material attribute parameters of the nodes in the formatted original edge list, determines the material attention weights between single-hop nodes according to the material attribute parameters, and obtains the output of the second attention head according to the material attention weights between single-hop nodes and the corresponding node parameters; The third attention head extracts the target parameter of the air conditioning air supply mode of the node from the formatted original edge list, determines the attention weight of the air supply direction between the single-hop nodes according to the target parameter in the air conditioning air supply mode, and obtains the output of the third attention head according to the attention weight of the air supply direction between the single-hop nodes and the corresponding node parameters; The first splicing module is used to perform feature splicing on the output of the first attention head, the output of the second attention head and the output of the third attention head to obtain a local heat exchange feature.

5. The temperature prediction method based on graph neural network according to claim 4 is characterized in that: The second attention layer includes: The fourth attention head identifies a heat conduction path that meets a preset condition in the second edge list with a local heat exchange feature, determines an end node of the heat conduction path, determines a heat conduction attention weight of the start node of the heat conduction path to the end node according to a preset first attention mechanism, and obtains a fourth attention head output according to the heat conduction attention weight and a node parameter of the end node; The fifth attention head identifies the heat blocking region in the second edge list with the local heat exchange feature, determines the heat blocking attention weight between the heat blocking region node and the surrounding nodes according to the preset second attention mechanism, and obtains the fifth attention head output according to the heat blocking attention weight and the node parameters of the surrounding nodes, wherein the surrounding nodes are nodes directly adjacent to the blocking region node; The second splicing module is used to perform feature splicing on the output of the fourth attention head and the output of the fifth attention head to obtain a global thermal balance feature.

6. A temperature prediction method based on graph neural network according to any one of claims 1 to 5, characterized in that: According to the prediction result of the temperature distribution characteristics and the temperature of the corresponding position under the current air conditioning air supply mode, the steady-state predicted temperature of the corresponding position under the predicted air conditioning air supply mode is obtained, including: The temperature distribution feature prediction results and the temperature of the corresponding position under the current air-conditioning air supply mode are input into a pre-trained temperature prediction model to obtain the steady-state predicted temperature of the corresponding position under the predicted air-conditioning air supply mode, wherein the pre-trained temperature prediction model is trained based on different current temperatures, the air-conditioning air supply mode is in a steady state, the temperature distribution feature prediction results in the room, and the actual temperature value when the air-conditioning air supply mode is predicted to be in a steady state as input.

7. The temperature prediction method based on graph neural network according to claim 1 is characterized in that: The temperature sensor is a distributed temperature sensor, including a main control module, multiple fixed temperature sensor modules and at least one movable temperature sensor module, wherein the movable temperature sensor module is embedded with a temperature sensing unit, a positioning unit, an intelligent sensing unit and an intelligent decision-making unit; Get the temperature of different locations in the current air conditioning air supply mode collected by the temperature sensor, including: Get the spatial information of temperature collection and the number of preset nodes; The spatial information of temperature collection and the preset number of nodes are sent to the distributed temperature sensor, so that the distributed temperature sensor main control module can allocate the target position of the required temperature collection to the movable temperature sensor module; The movable sensor module plans the motion path according to the intelligent sensing unit and the intelligent decision-making unit until it reaches the target position; The movable sensor module collects the temperature at the target location.

8. The temperature prediction method based on graph neural network according to claim 7 is characterized in that: The distributed temperature sensor allocates the target location of the required temperature collection to the movable temperature sensor module, including: The distributed temperature sensor main control module obtains the position of each fixed temperature sensor module; According to the position of the fixed temperature sensor module, the preset number of nodes and the spatial information of temperature collection, the vacant position is calculated and the vacant position is used as the target position.

9. A temperature prediction method based on graph neural network according to claim 8, characterized in that: According to the position of the fixed temperature sensor module, the number of preset nodes and the spatial information of temperature collection, the vacant position is calculated and the vacant position is used as the target position, including: Determine the minimum coverage radius of each node based on the spatial information and the preset number of nodes; Obtain the position of the fixed temperature sensor module, and determine the coverage range represented by the minimum coverage radius of the fixed temperature sensor with the position of the fixed temperature sensor module as the center; Determine the vacant area according to the coverage range represented by the minimum coverage radius of the fixed temperature sensor and the spatial information; The vacancy location is determined based on the coverage range represented by the vacancy area and the minimum coverage radius.

Images