WO2023024213A1 - 基于众包的温度预测方法、装置、设备和存储介质 - Google Patents

基于众包的温度预测方法、装置、设备和存储介质 Download PDF

Info

Publication number
WO2023024213A1
WO2023024213A1 PCT/CN2021/121177 CN2021121177W WO2023024213A1 WO 2023024213 A1 WO2023024213 A1 WO 2023024213A1 CN 2021121177 W CN2021121177 W CN 2021121177W WO 2023024213 A1 WO2023024213 A1 WO 2023024213A1
Authority
WO
WIPO (PCT)
Prior art keywords
vector
temperature prediction
temperature
data
training
Prior art date
Application number
PCT/CN2021/121177
Other languages
English (en)
French (fr)
Inventor
宋轩
陈达寅
史小丹
张浩然
Original Assignee
南方科技大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 南方科技大学 filed Critical 南方科技大学
Publication of WO2023024213A1 publication Critical patent/WO2023024213A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
    • G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F9/00—Arrangements for program control, e.g. control units
    • G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
    • G06F9/46—Multiprogramming arrangements
    • G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
    • G06F9/5061—Partitioning or combining of resources
    • G06F9/5072—Grid computing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/08—Thermal analysis or thermal optimisation

Definitions

  • the present invention relates to the technical field of temperature prediction, in particular to a temperature prediction method, device, equipment and storage medium based on crowdsourcing.
  • Smartphones are mobile devices that everyone is equipped with at present. If everyone's smart phones can be used to participate in the real-time monitoring of the temperature of the entire area, the required computing costs for the park will be greatly reduced.
  • the technical problem to be solved by the present invention is to provide a temperature prediction method, device, equipment and storage medium based on crowdsourcing, which can improve the accuracy and efficiency of temperature prediction.
  • the present invention provides a temperature prediction method based on crowdsourcing, including:
  • terminal data includes battery data, screen usage data and CPU usage data
  • the present invention also provides a temperature prediction device based on crowdsourcing, including:
  • the collection module is used to collect terminal data, and the terminal data includes battery data, screen usage data and CPU usage data;
  • a prediction module configured to input the terminal data into a preset optimal neural network model, and output a temperature prediction result, the temperature prediction result including a temperature prediction value and its reliability;
  • the first obtaining module is used to obtain the temperature prediction results of other terminals within the preset range, and obtain the temperature prediction value with the highest reliability as the final temperature according to its own temperature prediction results and the temperature prediction results of the other terminals. Predictive value.
  • the present invention also provides an electronic device, the electronic device comprising:
  • processors one or more processors
  • the one or more processors are made to implement the temperature prediction method based on crowdsourcing as provided in the first aspect.
  • the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the crowdsourcing-based temperature prediction method as provided in the first aspect is implemented.
  • the beneficial effect of the present invention is that: the neural network model considers a variety of influencing factors, which can ensure the accuracy of temperature prediction, and it will output the credibility of the data, which can be used for data screening, and has more extensiveness and anti-interference; By adopting crowdsourcing method for temperature prediction, the burden of data collection and calculation is evenly distributed to each terminal, making temperature monitoring accurate and efficient.
  • Fig. 1 is a flow chart of a temperature prediction method based on crowdsourcing provided by the present invention
  • Fig. 2 is a structural schematic diagram of a temperature prediction device based on crowdsourcing provided by the present invention
  • Fig. 3 is a schematic structural diagram of an electronic device provided by the present invention.
  • Embodiment 4 is a flow chart of a crowdsourcing-based temperature prediction method according to Embodiment 1 of the present invention.
  • FIG. 5 is a schematic diagram of a neural network model according to Embodiment 1 of the present invention.
  • FIG. 6 is a schematic diagram of an eight-head attention model according to Embodiment 1 of the present invention.
  • FIG. 7 is a schematic diagram of an attention model unit according to Embodiment 1 of the present invention.
  • first”, “second”, etc. may be used herein to describe various directions, actions, steps or elements, etc., but these directions, actions, steps or elements are not limited by these terms. These terms are only used to distinguish a first direction, action, step or element from another direction, action, step or element.
  • first information could be termed second information, and, similarly, second information could be termed first information, without departing from the scope of the present application. Both the first information and the second information are information, but they are not the same information.
  • the terms “first”, “second”, etc. should not be interpreted as indicating or implying relative importance or implying the number of indicated technical features. Thus, a feature defined as “first” and “second” may explicitly or implicitly include one or more of these features.
  • “plurality” means at least two, such as two, three, etc., unless otherwise specifically defined.
  • a temperature prediction method based on crowdsourcing including:
  • S101 collect terminal data, the terminal data includes battery data, screen usage data and CPU usage data;
  • S102 Input the terminal data into a preset optimal neural network model, and output a temperature prediction result, the temperature prediction result including a temperature prediction value and its reliability;
  • S103 Obtain temperature prediction results of other terminals within a preset range, and obtain a temperature prediction value with the highest reliability as a final temperature prediction value according to its own temperature prediction results and the temperature prediction results of the other terminals.
  • the neural network model of the present invention takes into account the different heating conditions of the terminal under different usage conditions (resting, listening to music, playing games, chatting, etc.), and at the same time some state information of the terminal itself (including CPU usage, screen usage time, etc.) are also different, and various influencing factors are used as input variables of the neural network model to ensure the accuracy of temperature prediction.
  • the neural network model of the present invention can output the credibility of data, can be used for data screening, has more universality and anti-interference, and can be applied to different terminals.
  • the method further includes:
  • the management center of the scenic spot can regulate the cooling resources according to the temperature of each area to ensure that the temperature of each area is suitable, so as to bring a more comfortable experience for tourists.
  • step S101 includes:
  • terminal data including battery data, screen usage data and CPU usage data
  • the battery data includes voltage, battery temperature sensor readings, battery temperature sensor readings when the screen was last activated, and battery temperature sensor readings when the screen was turned off last time
  • the screen usage data includes screen activation status, screen continuous activation time , the continuous screen-off time of the screen, the continuous screen-off time before the last activation of the screen, and the continuous use time before the last screen-off
  • the CPU usage data includes a CPU usage rate and an average CPU usage rate.
  • the method before the step S102, the method further includes:
  • the neural network model corresponding to this training is verified, and the mean absolute error corresponding to this training is obtained;
  • the average absolute error corresponding to the preset times of training is greater than the minimum average absolute error, and the neural network model corresponding to the minimum average absolute error is taken as the optimal neural network model.
  • the training ends, and the neural network model corresponding to the minimum MAE is selected as the optimal neural network model.
  • the neural network model includes a one-hot encoding module, an eight-head attention model, a hidden layer, an activation layer, a first linear layer, and a second linear layer, and the eight-head attention model Including eight attention model units and splicing modules and a third linear layer, the attention model unit includes four fourth linear layers and a self-attention module and a target attention module;
  • the one-hot encoding module is used to perform one-hot encoding on 11 variables in the input sample terminal data respectively to obtain 11 first vectors, and the first vectors are 12-dimensional vectors;
  • the one-hot encoding module is also used to perform one-hot encoding on the input temperature data to obtain a second vector, which is a 12-dimensional vector;
  • Three fourth linear layers in the four fourth linear layers are used to encode the 11 first vectors to obtain vector Q, vector K, and vector V, and the vector Q, vector K, and vector V are all is a 128-dimensional vector;
  • the remaining one of the four fourth linear layers is used to encode the second vector to obtain a target vector, and the target vector is a 128-dimensional vector;
  • the self-attention module is used to perform a self-attention operation on the vector Q, vector K and vector V to obtain 11 self-attention vectors, and the self-attention vector is a 128-dimensional vector;
  • the target attention module is used to calculate a third vector according to the 11 self-attention vectors and the target vector, and the third vector is a 128-dimensional vector;
  • the splicing module is used to splice the third vector output by the eight attention model units to obtain a fourth vector, which is a 1024-dimensional vector;
  • the third linear layer is used to encode the fourth vector to obtain a fifth vector, where the fifth vector is a 128-dimensional vector;
  • the hidden layer is used to encode the fifth vector to obtain a sixth vector, where the sixth vector is a 32-dimensional vector;
  • the activation layer is used to calculate the sixth vector according to a preset activation function to obtain a seventh vector, and the seventh vector is a 32-dimensional vector;
  • the first linear layer is used to encode the seventh vector to obtain a temperature prediction value
  • the second linear layer is used to encode the seventh vector to obtain the reliability of the temperature prediction value.
  • the attention mechanism pays more attention to the connection between different input variables and the correlation between input variables and labels. Therefore, the present invention uses both the multi-head self-attention mechanism and the target attention mechanism To strengthen the model, the average prediction error of the final model is smaller by using the attention mechanism to consider the correlation between input variables and the correlation between input variables and real air temperature.
  • the third vector is calculated according to the 11 self-attention vectors and the target vector, including:
  • the 11 self-attention vectors and the target vector respectively to obtain 11 dot products are in one-to-one correspondence with the 11 self-attention vectors;
  • the 11 weight vectors are added together to obtain a third vector.
  • the normalized dot product will tend to be stable during the training process, and a set of normalized dot products will be obtained after training. Later, when using the optimal neural network model for temperature prediction, only 11 variables of the terminal data are input, and after one-hot encoding, 11 self-attention vectors are obtained through three fourth linear layers and self-attention modules. In the force module, the 11 self-attention vectors are directly multiplied by this set of normalized dot products to obtain 11 weight vectors, and then the 11 weight vectors are added together to output the third vector.
  • the neural network model is trained according to the sample terminal data in the training set and the corresponding temperature data to obtain the neural network model corresponding to this training, and save the model parameters, include:
  • the loss function is , ⁇ is the first temperature prediction value, ⁇ is the reliability of the first temperature prediction value, and T is the corresponding temperature data.
  • the neural network model corresponding to this training is verified according to the sample terminal data in the verification set and its corresponding temperature data, and the mean absolute error corresponding to this training is obtained, including:
  • the present invention also provides a temperature prediction device based on crowdsourcing, including:
  • the collection module 201 is used to collect terminal data, and the terminal data includes battery data, screen usage data and CPU usage data;
  • a prediction module 202 configured to input the terminal data into a preset optimal neural network model, and output a temperature prediction result, the temperature prediction result including a temperature prediction value and its reliability;
  • the first obtaining module 203 is configured to obtain the temperature prediction results of other terminals within the preset range, and obtain the temperature prediction value with the highest reliability according to its own temperature prediction results and the temperature prediction results of the other terminals, as the final temperature forecast.
  • the present invention also provides an electronic device, the electronic device comprising:
  • processors 301 one or more processors 301;
  • a storage device 302 configured to store one or more programs
  • the one or more processors 301 implement the temperature prediction method based on crowdsourcing as described above.
  • the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the crowdsourcing-based temperature prediction method as described above is realized.
  • Embodiment 1 of the present invention is: a temperature prediction method based on crowdsourcing, which can be applied to scenes such as scenic spots, and the method can be executed by tourists' smart phones.
  • the management center of the scenic spot divides the scenic spot into multiple areas in advance, and tourists can install the APP (application program) related to the scenic spot in their mobile phones, and realize the data interaction with the management center of the scenic spot through the APP.
  • APP application program
  • tourists can also view the scenic spot map, activity schedule, venue reservation, etc. through the APP.
  • bluetooth devices can be pre-installed in the venues of the scenic spot, and each bluetooth device has a different ID.
  • Each Bluetooth device will periodically send a Bluetooth broadcast packet.
  • a tourist holds a mobile device with the APP close to a certain Bluetooth device to a certain distance it will trigger the check-in task corresponding to the Bluetooth device to realize electronic check-in.
  • This punching method can get rid of the original paper punching method, save resources and increase fun.
  • the APP will regularly report the location information to the management center, and the management center can count the flow of people in each area. For venues with fewer people (for example, less than 10 people), the control of cooling resources may be reduced or abandoned; for venues with more people (for example, greater than or equal to 10 people), the area temperature information obtained by crowdsourcing will be used Make reasonable allocation of cooling resources.
  • Crowdsourcing refers to distributing data or calculations that need to be collected by oneself to volunteers.
  • each terminal installed with a corresponding APP in the park is allowed to jointly participate in the temperature prediction of the area where it is located.
  • the management center of the scenic spot can obtain the temperature of all passing areas in the scenic spot in real time.
  • the burden of calculation and data collection will be shared equally on the terminals of tourists, making temperature monitoring accurate and efficient.
  • a kind of temperature prediction method based on crowdsourcing includes the following steps:
  • S401 Obtain a temperature prediction model, that is, an optimal neural network model.
  • the process of building and training the neural network model is implemented in each terminal.
  • the construction and training of the neural network model can also be completed on the server side, and then the optimal neural network model is sent to each terminal.
  • a neural network model is constructed, and a preset number of sample terminal data and corresponding temperature data are obtained. The steps of building and getting samples are executed in no sequence.
  • the preset number of sample terminal data and the corresponding temperature data are divided according to a preset ratio to obtain a training set and a verification set. For example, 80% of the sample terminal data is used as a training set, and 20% of the sample terminal data is used as a validation set.
  • each set of sample terminal data includes screen activation status, voltage, CPU usage, battery temperature sensor readings, continuous activation time of the screen, continuous screen-off time, continuous screen-off time before the last activation of the screen, last screen-off time
  • the attention mechanism pays more attention to the connection between different input variables and the correlation between input variables and labels. Therefore, in this embodiment, both The multi-head self-attention mechanism and target attention mechanism are used to strengthen the model.
  • the average prediction error of the final model can be less than 1 Celsius.
  • the output of the neural network model in this embodiment is the Gaussian distribution function of the temperature prediction result
  • the mathematical expectation of the Gaussian distribution function represents the temperature prediction value with the highest probability
  • the standard deviation of the Gaussian distribution function represents the reliability of the temperature prediction result.
  • the larger the standard deviation the lower the reliability; the smaller the standard deviation, the higher the reliability.
  • the neural network model includes an eight-head attention model (Multi-Head-Attention), a hidden layer (Hidden Layer), an activation layer (ReLu), a first linear layer and a second linear Layer (Output Layer); as shown in Figure 6, the eight-head attention model includes eight attention model units and a splicing module and a third linear layer; as shown in Figure 7, the attention model unit includes four The fourth linear layer and a self-attention module (Self-Attention) and a target attention module (Taget-Attention); the neural network model also includes a one-hot encoding module.
  • Multi-Head-Attention Multi-Head-Attention
  • Hidden Layer hidden layer
  • ReLu activation layer
  • the attention model unit includes four The fourth linear layer and a self-attention module (Self-Attention) and a target attention module (Taget-
  • the input data is 11 variables in the sample terminal data and real-time temperature data
  • the output data is the temperature prediction value and its reliability.
  • 11 variables in a set of sample terminal data and real-time temperature data are input into the one-hot encoding module.
  • the one-hot encoding module performs one-hot encoding on the 11 variables in the input sample terminal data to obtain 11 first vectors, and the first vectors are 12-dimensional vectors; one-hot encoding is performed on the input temperature data to obtain the first Two vectors, the second vector is a 12-dimensional vector; then the 11 first vectors are sent to three fourth linear layers in each attention model unit, and the second vector is sent to each attention model unit respectively The remaining fourth linear layer of .
  • the fourth linear layer in the attention model unit is mainly used to encode a 12-dimensional vector into a 128-dimensional vector. Since the implementation of directly converting a 12-dimensional vector into a 128-dimensional vector is too complicated, in this embodiment, each fourth Each linear layer consists of two linear layers, one of which is used to encode a 12-dimensional vector into a 64-dimensional vector, and the other linear layer is used to encode a 64-dimensional vector into a 128-dimensional vector.
  • three of the four fourth linear layers encode the 11 first vectors and output a 128-dimensional vector, and the output results of the three fourth linear layers are respectively It is vector Q (Quary), vector K (Key) and vector V (Value), and then sent to the self-attention module (Self-Attention); the remaining fourth linear layer is used to encode the second vector, outputting 128
  • the target vector of dimension is given to the target attention module (Taget-Attention).
  • the self-attention module (Self-Attention) is used to perform a self-attention operation on the vector Q, vector K, and vector V to obtain 11 self-attention vectors, and the self-attention vectors are 128-dimensional vectors. Based on the principle of the self-attention mechanism, a new vector containing self-attention information can be obtained from the vector Q, vector K and vector V, that is, the vector contains the correlation information between different features.
  • the target attention module (Taget-Attention) is used to calculate and obtain a third vector according to the 11 self-attention vectors and the target vector, and the third vector is a 128-dimensional vector. That is, the vector containing the correlation information between different features obtained through the self-attention module will perform a target attention calculation with the real-time temperature data processed by the linear layer, and obtain a vector containing the correlation between each feature and the temperature label. A vector of degree information.
  • the internal implementation method of the target attention module includes: first, respectively performing dot multiplication of the 11 self-attention vectors and the target vector to obtain 11 dot products, and the 11 dot products and the 11 self-attention vectors Attention vectors correspond one-to-one. Then, normalize the 11 dot products, and save the normalized dot products, this step can realize the normalization through the Softmax function, of course, in other embodiments, other functions can also be used Implement the normalization operation. Next, the 11 self-attention vectors are multiplied by their corresponding normalized dot products to obtain 11 weight vectors. Finally, the 11 weight vectors are added together to obtain a third vector.
  • the normalized dot product will tend to be stable during the training process, and a set of normalized dot products will be obtained after training. Later, when using the optimal neural network model for temperature prediction, only 11 variables of the terminal data are input, and after one-hot encoding, 11 self-attention vectors are obtained through three fourth linear layers and self-attention modules. In the force module, the 11 self-attention vectors are directly multiplied by this set of normalized dot products to obtain 11 weight vectors, and then the 11 weight vectors are added together to output the third vector.
  • each of the eight attention model units will output a third vector, which is then sent to the splicing module respectively.
  • the splicing module splices the third vectors output by the eight attention model units to obtain a fourth vector, which is a 1024-dimensional vector; and then sends the fourth vector to the third linear layer.
  • the third linear layer encodes the fourth vector to obtain a fifth vector, where the fifth vector is a 128-dimensional vector.
  • the eight-head attention model (Multi-Head-Attention) sends the output fifth vector to the hidden layer (Hidden Layer).
  • the hidden layer encodes the fifth vector to obtain a sixth vector, and the sixth vector is a 32-dimensional vector; and then sends the sixth vector to the activation layer.
  • the number of activation layers may be two, and the hidden layer sends the sixth vector to the two activation layers respectively. Both activation layers are used to calculate the sixth vector according to a preset activation function to obtain a seventh vector, and the seventh vector is a 32-dimensional vector; then the seventh vectors output by the two activation layers are respectively Sent to the first linear layer and the second linear layer.
  • the activation function adopts ReLu.
  • the first linear layer encodes the seventh vector to obtain a mathematical expectation, that is, a temperature prediction value
  • the second linear layer encodes the seventh vector to obtain the standard deviation, that is, the reliability of the temperature prediction value.
  • the neural network model is constructed and the samples are obtained, then, the neural network model is trained according to the sample terminal data in the training set and the corresponding temperature data to obtain the neural network model corresponding to this training, and save the model Parameters: verify the neural network model corresponding to this training according to the sample terminal data in the verification set and the corresponding temperature data, and obtain the mean absolute error corresponding to this training.
  • the training set sequentially acquire a sample terminal data from the training set as the current training sample; input the current training sample and its corresponding temperature data into the latest neural network model for forward propagation, and obtain the current
  • the first output result corresponding to the training sample, the first output result includes the first temperature prediction value and its reliability; according to the preset loss function, calculate the first output result corresponding to the current training sample and the current A loss value of the temperature data corresponding to the training sample; according to the loss value, the neural network model is updated through a backpropagation algorithm.
  • the neural network model corresponding to this training is obtained, and the network parameters are saved.
  • the latest neural network model refers to the neural network model obtained after previous training, that is to say, the neural network model corresponding to the first training is based on the preset neural network model. obtained through training; the neural network model corresponding to the second training is obtained by training the training samples in the training set based on the neural network model corresponding to the first training; and so on.
  • the loss function used is , where ⁇ is the first temperature prediction value, ⁇ is the reliability of the first temperature prediction value, and T is the corresponding temperature data.
  • Traversing the verification set sequentially obtaining a sample terminal data from the verification set as a current verification sample; inputting the current verification sample and its corresponding temperature data into the neural network model corresponding to this training for forward propagation , to obtain the second output result corresponding to the current verification sample, the second output result including the second temperature prediction value and its reliability; calculate the second temperature prediction value corresponding to the current verification sample and the current verification
  • the absolute value of the difference of the temperature data corresponding to the sample is used to obtain the absolute value of the difference corresponding to the current verification sample.
  • the mean absolute error is the average value of the absolute value of the difference between the predicted value and the true value of each validation sample.
  • the minimum average absolute error is obtained in real time; when the average absolute error corresponding to the preset number of trainings after the minimum average absolute error is obtained is greater than the minimum average absolute error, then the minimum average absolute error corresponding to The neural network model is used as the optimal neural network model. That is, during the training process, when the MAE (mean absolute error) on the verification set does not decrease for 50 consecutive epochs, the model is taken as the final model.
  • terminal data includes battery data, screen usage data, and CPU usage data.
  • terminal data may be collected according to a preset cycle, for example, terminal data may be collected every 1 minute.
  • a set of terminal data includes 11 variables in total.
  • the battery data includes voltage, real-time battery temperature sensor readings, battery temperature sensor readings when the screen was last activated, and battery temperature sensor readings when the screen was turned off last time
  • the screen usage data includes screen activation status, screen duration Activation time, continuous screen-off time, continuous screen-off time before the last activation of the screen, and continuous use time before the last screen-off
  • the CPU usage data includes real-time CPU usage and average CPU usage.
  • the average CPU usage rate can be obtained by calculating the average value of the CPU usage rates collected in the latest 50 times.
  • S403 Input the terminal data into an optimal neural network model, and output a temperature prediction result, where the temperature prediction result includes a temperature prediction value and its reliability.
  • a temperature prediction is performed after the sequential terminal data is collected according to a preset period.
  • S404 Obtain temperature prediction results of other terminals within a preset range, and obtain a temperature prediction value with the highest reliability as a final temperature prediction value according to its own temperature prediction results and the temperature prediction results of the other terminals.
  • the temperature prediction results of the device can be exchanged between the terminal and the terminal through Bluetooth broadcasting.
  • the terminal will obtain the temperature prediction results of other terminals within a preset radius with itself as the center, and then select the temperature prediction value with the highest reliability among the temperature prediction results of other terminals and its own temperature prediction results.
  • the location information is the current location information of the terminal, which can be obtained through a positioning model (such as GPS).
  • the management center After receiving the final predicted temperature value and location information sent by each terminal, the management center can determine the final predicted temperature value of each area according to the area corresponding to each location information. Since when a terminal is at the edge of a certain area, the final temperature prediction value may be determined based on the temperature prediction results of terminals in other areas, therefore, there may be multiple different final temperature prediction values in the same area, and the management center may further base on These final temperature prediction values determine a temperature value as the temperature value of the area.
  • the accuracy of temperature measurement area can be arbitrarily reduced under the premise of GPS accuracy guarantee, which ensures real-time temperature monitoring in small areas.
  • the management center After the management center determines the temperature values of each area, it can reasonably control the cooling resources through the temperature distribution in different areas of the entire park, such as controlling small sprinklers to go to high-temperature outdoor areas, and adjusting the air-conditioning power at different positions in the room to achieve high efficiency and energy saving. comfort.
  • the temperature prediction model of this embodiment takes into account a variety of influencing factors and the correlation between the influencing factors, which can ensure the accuracy of temperature prediction, and the temperature prediction model will output the credibility of the data, which can be used for data screening, and more It has universality and anti-interference, and can be applied to different types of mobile phones.
  • temperature prediction is carried out in a crowdsourcing manner, and the burden of data collection and calculation is evenly shared on the terminals of tourists, so that temperature monitoring is accurate and efficient.
  • the management center of the scenic spot can adjust the cooling resources according to the temperature of each area to ensure that the temperature of each area is suitable, so as to bring a more comfortable experience for tourists.
  • the second embodiment of the present invention is: a temperature prediction device based on crowdsourcing, which can execute the temperature prediction method provided by the first embodiment of the present invention, and has corresponding functional modules and beneficial effects for executing the method.
  • the device can be implemented by software/or hardware, specifically including:
  • the collection module 201 is used to collect terminal data, and the terminal data includes battery data, screen usage data and CPU usage data;
  • a prediction module 202 configured to input the terminal data into a preset optimal neural network model, and output a temperature prediction result, the temperature prediction result including a temperature prediction value and its reliability;
  • the first obtaining module 203 is configured to obtain the temperature prediction results of other terminals within the preset range, and obtain the temperature prediction value with the highest reliability according to its own temperature prediction results and the temperature prediction results of the other terminals, as the final temperature forecast.
  • the temperature prediction device also includes:
  • a sending module configured to send the final predicted temperature value and location information to the management center, so that the management center can regulate the temperature according to the location information and the final predicted temperature value.
  • the collection module is specifically configured to collect terminal data according to a preset period, and the terminal data includes battery data, screen usage data, and CPU usage data;
  • the battery data includes voltage, battery temperature sensor readings, battery temperature sensor readings when the screen was last activated, and battery temperature sensor readings when the screen was turned off last time
  • the screen usage data includes screen activation status, screen continuous activation time , the continuous screen-off time of the screen, the continuous screen-off time before the last activation of the screen, and the continuous use time before the last screen-off
  • the CPU usage data includes a CPU usage rate and an average CPU usage rate.
  • the temperature prediction device also includes:
  • the second acquisition module is used to obtain a preset number of sample terminal data and corresponding temperature data, and divide the preset number of sample terminal data and corresponding temperature data according to a preset ratio to obtain a training set and validation set;
  • a training module configured to train the neural network model according to the sample terminal data in the training set and the corresponding temperature data, obtain the neural network model corresponding to this training, and save the model parameters;
  • the verification module is used to verify the neural network model corresponding to this training according to the sample terminal data and the corresponding temperature data in the verification set, and obtain the corresponding mean absolute error of this training;
  • the third acquisition module is used to obtain the minimum average absolute error in real time after each training
  • a determination module configured to use the neural network model corresponding to the minimum average absolute error as the optimal neural network model when the average absolute error corresponding to the preset times of training is greater than the minimum average absolute error after the minimum average absolute error is obtained. network model.
  • the training module includes:
  • the first traversal unit is configured to traverse the training set, and sequentially obtain a sample terminal data from the training set as the current training sample;
  • the first obtaining unit is configured to input the current training sample and its corresponding temperature data into the latest neural network model for forward propagation to obtain a first output result corresponding to the current training sample, the first output result including The first temperature prediction value and its reliability;
  • a first calculation unit configured to calculate a loss value of the first output result corresponding to the current training sample and the temperature data corresponding to the current training sample according to a preset loss function
  • An update unit configured to update the neural network model through a backpropagation algorithm according to the loss value
  • the second obtaining unit is used to obtain the neural network model corresponding to this training after traversing the training set, and save the network parameters;
  • the loss function is: ⁇ is the first temperature prediction value, ⁇ is the reliability of the first temperature prediction value, and T is the corresponding temperature data.
  • the verification module includes:
  • the second traversal unit is configured to traverse the verification set, and sequentially obtain a sample terminal data from the verification set as a current verification sample;
  • the third obtaining unit is configured to input the current verification sample and its corresponding temperature data into the neural network model corresponding to the current training for forward propagation, and obtain the second output result corresponding to the current verification sample, the said The second output result includes the second temperature prediction value and its reliability;
  • the second calculation unit is configured to calculate the absolute value of the difference between the second predicted temperature value corresponding to the current verification sample and the temperature data corresponding to the current verification sample, to obtain the absolute value of the difference corresponding to the current verification sample;
  • the third calculation unit is configured to calculate the average value of the absolute value of the difference corresponding to each sample terminal data in the verification set after traversing the verification set, so as to obtain the average absolute error corresponding to this training.
  • the third embodiment of the present invention is: an electronic device, the electronic device includes:
  • processors 301 one or more processors 301;
  • a storage device 302 configured to store one or more programs
  • the one or more programs are executed by the one or more processors 301, so that the one or more processors 301 implement the various processes in the above-mentioned embodiments of the crowdsourcing-based temperature prediction method, and The same technical effect can be achieved, so in order to avoid repetition, details will not be repeated here.
  • Embodiment 4 of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the various processes in the above-mentioned embodiment of the crowdsourcing-based temperature prediction method are implemented. , and can achieve the same technical effect, in order to avoid repetition, it will not be repeated here.
  • the present invention provides a crowdsourcing-based temperature prediction method, device, equipment, and storage medium.
  • the neural network model adopted takes into account a variety of influencing factors, and simultaneously uses a multi-head self-attention mechanism and target attention mechanism to strengthen the model, by using the attention mechanism to consider the correlation between the input variables and the correlation between the input variables and the real air temperature, so that the average prediction error of the final model is small, and the neural network model will also output data
  • the credibility, more extensive and anti-interference, can be applied to different terminals.
  • the burden of data collection and calculation is evenly distributed to each terminal, making temperature monitoring accurate and efficient; the management center can determine each area according to the final temperature prediction value and location information uploaded by each terminal The temperature is controlled, and the refrigeration resources are regulated to ensure that the temperature in each area is suitable, so as to bring a more comfortable experience for tourists.
  • the present invention can be realized by means of software and necessary general-purpose hardware, and of course it can also be realized by hardware, but in many cases the former is a better implementation mode .
  • the essence of the technical solution of the present invention or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as a floppy disk of a computer , read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), flash memory (FLASH), hard disk or optical disc, etc., including several instructions to make a computer device (it can be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments of the present invention.
  • a computer-readable storage medium such as a floppy disk of a computer , read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), flash memory (FLASH), hard disk or optical disc, etc.

Abstract

本发明公开了一种基于众包的温度预测方法、装置、设备和存储介质，方法包括：采集终端数据，终端数据包括电池数据、屏幕使用数据和CPU使用数据；将所述终端数据输入预设的最优神经网络模型，输出温度预测结果，所述温度预测结果包括温度预测值及其可信度；获取预设范围内的其他终端的温度预测结果，并根据自身的温度预测结果和所述其他终端的温度预测结果，获取可信度最高的温度预测值，作为最终温度预测值。本发明可提高温度预测的准确性，且提高温度预测的效率。

Description

基于众包的温度预测方法、装置、设备和存储介质 技术领域
本发明涉及温度预测技术领域，尤其涉及一种基于众包的温度预测方法、装置、设备和存储介质。
 背景技术
夏天前往景区旅游时，通常会碰到需要在烈日下排队一两个小时的情况，用户体验比较差，并且容易导致中暑。一些园区可能会在每个场馆入口处投放降温设备（如空调）或是统一洒水降温，但是这种方式不能根据不同区域的温度高低进行针对性的制冷资源投递（比如在太阳直射区持续洒水，而阴影区则减少投入）。同理，在室内的制冷过程中，由于都是统一的制冷调控，经常会导致制冷效果不均匀的情况，有些区域可能温度较低（如靠近室内的区域），有些区域则温度较高（靠近门口或靠近一些热源的区域）。因此，如果能实时监控各个区域的温度就能够实现制冷资源更合理的配置。
智能手机是目前人人都配有的移动设备，如果能够利用每个人的智能手机来参与整个区域的温度实时监控，那么对于园区来说，所需的计算成本将会大大降低。
在公开号为CN110672231A的中国公开专利中公开了一种基于手机电池温度传感器的空气温度测量方法，其主要基于手机、人体和空气三者的热传导模型来计算，但以热传导温度预测模型考虑的相关因子较少，不确定因素较多，预测结果波动较大，无法保证预测的准确性。
 技术问题
本发明所要解决的技术问题是：提供一种基于众包的温度预测方法、装置、设备和存储介质，能够提高温度预测的准确性以及效率。
 技术解决方案
第一方面，本发明提供了一种基于众包的温度预测方法，包括：
采集终端数据，终端数据包括电池数据、屏幕使用数据和CPU使用数据；
将所述终端数据输入预设的最优神经网络模型，输出温度预测结果，所述温度预测结果包括温度预测值及其可信度；
获取预设范围内的其他终端的温度预测结果，并根据自身的温度预测结果和所述其他终端的温度预测结果，获取可信度最高的温度预测值，作为最终温度预测值。
第二方面，本发明还提供了一种基于众包的温度预测装置，包括：
采集模块，用于采集终端数据，终端数据包括电池数据、屏幕使用数据和CPU使用数据；
预测模块，用于将所述终端数据输入预设的最优神经网络模型，输出温度预测结果，所述温度预测结果包括温度预测值及其可信度；
第一获取模块，用于获取预设范围内的其他终端的温度预测结果，并根据自身的温度预测结果和所述其他终端的温度预测结果，获取可信度最高的温度预测值，作为最终温度预测值。
第三方面，本发明还提供了一种电子设备，所述电子设备包括：
一个或多个处理器；
存储装置，用于存储一个或多个程序；
当所述一个或多个程序被所述一个或多个处理器执行，使得所述一个或多个处理器实现如第一方面提供的基于众包的温度预测方法。
第四方面，本发明还提供了一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现如第一方面提供的基于众包的温度预测方法。
 有益效果
本发明的有益效果在于：神经网络模型考虑了多种影响因素，可保证温度预测的准确性，并且，其会输出数据的可信度，可用于数据筛选，更具有广泛性和抗干扰性；通过采用众包的方式进行温度预测，将数据采集负担和计算负担平摊到各个终端上，使得温度监控准确而高效。
 附图说明
图1为本发明提供的一种基于众包的温度预测方法的流程图；
图2为本发明提供的一种基于众包的温度预测装置的结构示意图；
图3为本发明提供的一种电子设备的结构示意图；
图4为本发明实施例一的一种基于众包的温度预测方法的流程图；
图5为本发明实施例一的神经网络模型的示意图；
图6为本发明实施例一的八头注意力模型的示意图；
图7为本发明实施例一的注意力模型单元的示意图。
 本发明的实施方式
下面结合附图和实施例对本发明作进一步的详细说明。可以理解的是，此处所描述的具体实施例仅仅用于解释本发明，而非对本发明的限定。另外还需要说明的是，为了便于描述，附图中仅示出了与本发明相关的部分而非全部结构。
在更加详细地讨论示例性实施例之前应当提到的是，一些示例性实施例被描述成作为流程图描绘的处理或方法。虽然流程图将各步骤描述成顺序的处理，但是其中的许多步骤可以被并行地、并发地或者同时实施。此外，各步骤的顺序可以被重新安排。当其操作完成时处理可以被终止，但是还可以具有未包括在附图中的附加步骤。处理可以对应于方法、函数、规程、子例程、子计算机程序等等。
此外，术语“第一”、“第二”等可在本文中用于描述各种方向、动作、步骤或元件等，但这些方向、动作、步骤或元件不受这些术语限制。这些术语仅用于将第一个方向、动作、步骤或元件与另一个方向、动作、步骤或元件区分。举例来说，在不脱离本申请的范围的情况下，可以将第一信息为第二信息，且类似地，可将第二信息称为第一信息。第一信息和第二信息两者都是信息，但其不是同一信息。术语“第一”、“第二”等而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。由此，限定有“第一”、“第二”的特征可以明示或者隐含地包括一个或者更多个该特征。在本发明的描述中，“多个”的含义是至少两个，例如两个，三个等，除非另有明确具体的限定。
如图1所示，一种基于众包的温度预测方法，包括：
S101：采集终端数据，终端数据包括电池数据、屏幕使用数据和CPU使用数据；
S102：将所述终端数据输入预设的最优神经网络模型，输出温度预测结果，所述温度预测结果包括温度预测值及其可信度；
S103：获取预设范围内的其他终端的温度预测结果，并根据自身的温度预测结果和所述其他终端的温度预测结果，获取可信度最高的温度预测值，作为最终温度预测值。
与以往利用热传导模型预测气温不同，本发明的神经网络模型考虑了终端在不同的使用情况下（静息，听音乐，玩游戏，聊天等）发热情况不同，同时终端本身的一些状态信息（包括CPU使用率，屏幕使用时长等）也有不同，将多种影响因素都作为神经网络模型的输入变量，以保证温度预测的准确性。并且，本发明的神经网络模型会输出数据的可信度，可用于数据筛选，更具有广泛性和抗干扰性，能够适用于不同的终端。
在一个可选的实施例中，所述步骤S103之后，所述方法还包括：
将所述最终温度预测值以及位置信息发送至管理中心，以使管理中心根据所述位置信息和最终温度预测值，进行温度调控。
由上述描述可知，景区的管理中心可以根据各个区域的温度，进行制冷资源的调控，保证各个区域温度适宜，从而为游客带来更加舒适的体验。
在一个可选的实施例中，所述步骤S101，包括：
根据预设的周期采集终端数据，所述终端数据包括电池数据、屏幕使用数据和CPU使用数据；
其中，所述电池数据包括电压、电池温度传感器读数、上一次激活屏幕时的电池温度传感器读数以及上一次熄屏时的电池温度传感器读数，所述屏幕使用数据包括屏幕激活状态、屏幕持续激活时间、屏幕持续熄屏时间、上一次激活屏幕前的持续熄屏时间以及上一次熄屏前持续使用时间，所述CPU使用数据包括CPU使用率以及平均CPU使用率。
通过设定合适的周期进行数据的采集和温度的预测，可在不影响终端本身的使用性能的情况下，保证温度预测的实时性。
在一个可选的实施例中，所述步骤S102之前，所述方法还包括：
构建神经网络模型；
获取预设数量的样本终端数据及其对应的温度数据，并按照预设的比例对所述预设数量的样本终端数据及其对应的温度数据进行划分，得到训练集和验证集；
根据所述训练集中的样本终端数据及其对应的温度数据对所述神经网络模型进行训练，得到本次训练对应的神经网络模型，并保存模型参数；
根据所述验证集中的样本终端数据及其对应的温度数据对本次训练对应的神经网络模型进行验证，得到本次训练对应的平均绝对误差；
每次训练后，实时获取最小平均绝对误差；
当获取到最小平均绝对误差之后连接预设次数的训练对应的平均绝对误差均大于所述最小平均绝对误差，则将所述最小平均绝对误差对应的神经网络模型作为最优神经网络模型。
通过对神经网络模型进行构建、训练、验证，当验证集上的MAE（平均绝对误差）连续50个epoch没有下降时，则结束训练，取最小MAE对应的神经网络模型作为最优神经网络模型。
在一个可选的实施例中，所述神经网络模型包括独热编码模块、八头注意力模型、隐含层、激活层、第一线性层和第二线性层，所述八头注意力模型包括八个注意力模型单元以及拼接模块和第三线性层，所述注意力模型单元包括四个第四线性层以及自注意力模块和目标注意力模块；
所述独热编码模块用于分别对输入的样本终端数据中的11个变量进行独热编码，得到11个第一向量，所述第一向量为12维向量；
所述独热编码模块还用于对输入的温度数据进行独热编码，得到第二向量，所述第二向量为12维向量；
所述四个第四线性层中的三个第四线性层用于对所述11个第一向量进行编码，得到向量Q、向量K和向量V，所述向量Q、向量K和向量V均为128维向量；
所述四个第四线性层中剩余的一个第四线性层用于对所述第二向量进行编码，得到目标向量，所述目标向量为128维向量；
所述自注意力模块用于对所述向量Q、向量K和向量V进行自注意力运算，得到11个自注意力向量，所述自注意力向量为128维向量；
所述目标注意力模块用于根据所述11个自注意力向量和目标向量，计算得到第三向量，所述第三向量为128维向量；
所述拼接模块用于对八个注意力模型单元输出的第三向量进行拼接，得到第四向量，所述第四向量为1024维向量；
所述第三线性层用于对所述第四向量进行编码，得到第五向量，所述第五向量为128维向量；
所述隐含层用于对所述第五向量进行编码，得到第六向量，所述第六向量为32维向量；
所述激活层用于根据预设的激活函数对所述第六向量进行计算，得到第七向量，所述第七向量为32维向量；
所述第一线性层用于对所述第七向量进行编码，得到温度预测值；
所述第二线性层用于对所述第七向量进行编码，得到所述温度预测值的可信度。
相较于传统的编码解码模型，注意力机制更加关注不同的输入变量之间的联系以及输入变量和标签之间的关联度，因此，本发明同时使用了多头自注意力机制和目标注意力机制来强化模型，通过使用注意力机制考虑输入变量之间的关联度以及输入变量与真实空气温度之间的关联度，使得最终模型的平均预测误差较小。
在一个可选的实施例中，所述根据所述11个自注意力向量和目标向量，计算得到第三向量，包括：
分别将所述11个自注意力向量与所述目标向量进行点乘，得到11个点积，所述11个点积与11个自注意力向量一一对应；
对所述11个点积进行归一化处理，并保存归一化后的点积；
分别将所述11个自注意力向量乘以其对应的归一化后的点积，得到11个权重向量；
将所述11个权重向量相加，得到第三向量。
其中，归一化后的点积在训练过程中会趋于稳定，训练完后会得到一组归一化后的点积。后续在运用最优神经网络模型进行温度预测时，只输入终端数据的11个变量，进行独热编码后通过三个第四线性层和自注意力模块得到11个自注意力向量，在目标注意力模块中直接将11个自注意力向量与这组归一化后的点积相乘，得到11个权重向量，再将11个权重向量相加，输出第三向量。
在一个可选的实施例中，所述根据所述训练集中的样本终端数据及其对应的温度数据对所述神经网络模型进行训练，得到本次训练对应的神经网络模型，并保存模型参数，包括：
遍历所述训练集，依次从所述训练集中获取一样本终端数据，作为当前训练样本；
将所述当前训练样本及其对应的温度数据输入最新的神经网络模型进行正向传播，得到所述当前训练样本对应的第一输出结果，所述第一输出结果包括第一温度预测值及其可信度；
根据预设的损失函数，计算所述当前训练样本对应的第一输出结果与所述当前训练样本对应的温度数据的损失值；
根据所述损失值，通过反向传播算法更新所述神经网络模型；
当遍历完所述训练集后，得到本次训练对应的神经网络模型，并保存网络参数；
其中，所述损失函数为
Figure dest_path_image001
，μ为第一温度预测值，σ为第一温度预测值的可信度，T为对应的温度数据。
在一个可选的实施例中，所述根据所述验证集中的样本终端数据及其对应的温度数据对本次训练对应的神经网络模型进行验证，得到本次训练对应的平均绝对误差，包括：
遍历所述验证集，依次从所述验证集中获取一样本终端数据，作为当前验证样本；
将所述当前验证样本及其对应的温度数据输入所述本次训练对应的神经网络模型进行正向传播，得到所述当前验证样本对应的第二输出结果，所述第二输出结果包括第二温度预测值及其可信度；
计算所述当前验证样本对应的第二温度预测值与所述当前验证样本对应的温度数据的差值绝对值，得到当前验证样本对应的差值绝对值；
当遍历完所述验证集后，计算所述验证集中各样本终端数据对应的差值绝对值的平均值，得到本次训练对应的平均绝对误差。
如图2所示，本发明还提供了一种基于众包的温度预测装置，包括：
采集模块201，用于采集终端数据，终端数据包括电池数据、屏幕使用数据和CPU使用数据；
预测模块202，用于将所述终端数据输入预设的最优神经网络模型，输出温度预测结果，所述温度预测结果包括温度预测值及其可信度；
第一获取模块203，用于获取预设范围内的其他终端的温度预测结果，并根据自身的温度预测结果和所述其他终端的温度预测结果，获取可信度最高的温度预测值，作为最终温度预测值。
如图3所示，本发明还提供了一种电子设备，所述电子设备包括：
一个或多个处理器301；
存储装置302，用于存储一个或多个程序；
当所述一个或多个程序被所述一个或多个处理器301执行，使得所述一个或多个处理器301实现如上所述的基于众包的温度预测方法。
本发明还提供了一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现如上所述的基于众包的温度预测方法。
实施例一
请参照图4-7，本发明的实施例一为：一种基于众包的温度预测方法，可适用于景区等场景，该方法可由游客的智能手机来执行。
景区的管理中心预先将景区划分为多个区域，游客的手机中可安装景区相关的APP（应用程序），通过该APP实现与景区的管理中心的数据交互。游客也可通过该APP查看景区地图、活动时间表、场馆预定等。
同时，可预先在景区的场馆中安装蓝牙设备，每个蓝牙设备具有不同的ID。每个蓝牙设备均会周期性地发送蓝牙广播包，当游客持有装有该APP的手机设备靠近某个蓝牙设备到一定距离时，会触发该蓝牙设备对应的打卡任务，实现电子打卡。这种打卡方式可以摆脱原有的纸质打卡方式，节约资源，增加趣味性。
进一步地，该APP会定时上报位置信息至管理中心，管理中心即可统计各区域的人流量。对于人员较少（例如少于10人）的场馆，可能会减少或者放弃制冷资源的调控；而对于人员较多（例如大于或等于10人）的场馆，将会根据众包获得的区域温度信息进行合理的制冷资源调配。
众包是指将原先需要自己采集的数据或者进行的计算分发给大众志愿者来进行。本实施例基于众包的方法，让处于园区内的装有相应APP的各个终端共同参与其所处区域的温度预测。通过这种方式，景区的管理中心可以实时地获得景区内所有有人经过的区域的温度，同时计算负担和数据采集负担将会平摊到游客的终端上，使得温度监控准确而高效。
如图4所示，本实施例提供的一种基于众包的温度预测方法包括如下步骤：
S401：获取温度预测模型，即最优神经网络模型。本实施例中，神经网络模型的构建和训练过程在各个终端中实现。在其他实施例中，也可以在服务端完成神经网络模型的构建和训练，再将最优神经网络模型发送给各个终端。
具体地，首先，构建神经网络模型，并获取预设数量的样本终端数据及其对应的温度数据。构建的步骤与获取样本的步骤不分先后执行。
获取到样本后，按照预设的比例对所述预设数量的样本终端数据及其对应的温度数据进行划分，得到训练集和验证集。例如，将80%的样本终端数据作为训练集，将20%的样本终端数据作为验证集。其中，每组样本终端数据包括屏幕激活状态、电压、CPU使用率、电池温度传感器读数、屏幕持续激活时间、屏幕持续熄屏时间、上一次激活屏幕前的持续熄屏时间、上一次熄屏前持续使用时间、上一次激活屏幕时的电池温度传感器读数、上一次熄屏时的电池温度传感器读数和平均CPU使用率共11个变量；样本终端数据对应的温度数据为采集该样本终端数据时的环境温度值。
对于神经网络模型的构建，相较于传统的编码解码模型，注意力机制更加关注不同的输入变量之间的联系以及输入变量和标签之间的关联度，因此，本实施例中，同时使用了多头自注意力机制和目标注意力机制来强化模型，通过使用注意力机制考虑输入变量之间的关联度以及输入变量与真实空气温度之间的关联度，使得最终模型的平均预测误差可小于1摄氏度。
本实施例的神经网络模型的输出为温度预测结果的高斯分布函数，该高斯分布函数的数学期望表示概率最高的温度预测值，该高斯分布函数的标准差表示该温度预测结果的可信度。其中，标准差越大，可信度越低；标准差越小，可信度越高。
具体地，如图5所示，所述神经网络模型包括八头注意力模型（Multi-Head-Attention）、隐含层（Hidden Layer）、激活层（ReLu）、第一线性层和第二线性层（Output Layer）；如图6所示，所述八头注意力模型包括八个注意力模型单元以及拼接模块和第三线性层；如图7所示，所述注意力模型单元包括四个第四线性层以及自注意力模块（Self-Attention）和目标注意力模块（Taget-Attention）；所述神经网络模型还包括独热编码模块。
对神经网络模型进行训练时，其输入的数据为样本终端数据中的11个变量以及实时的温度数据，其输出的数据为温度预测值及其可信度。
如图7所示，将一组样本终端数据中的11个变量以及实时的温度数据分别输入独热编码模块。独热编码模块分别对输入的样本终端数据中的11个变量进行独热编码，得到11个第一向量，所述第一向量为12维向量；对输入的温度数据进行独热编码，得到第二向量，所述第二向量为12维向量；然后将11个第一向量分别发送至各注意力模型单元中的三个第四线性层，将第二向量分别发送至各注意力模型单元中的剩余的一个第四线性层。
注意力模型单元中的第四线性层主要用于将12维向量编码为128维向量，由于直接将12维向量转换为128维向量的实现过于复杂，因此，本实施例中，每个第四线性层均由两个线性层构成，其中一个线性层用于将12维向量编码为64维向量，另一个线性层用于将64维向量编码为128维向量。
对于每个注意力模型单元，其四个第四线性层中的三个第四线性层均会对11个第一向量进行编码，输出一个128维向量，三个第四线性层的输出结果分别为向量Q（Quary）、向量K（Key）和向量V（Value），然后发送至自注意力模块（Self-Attention）；剩余的一个第四线性层用于对第二向量进行编码，输出128维的目标向量给目标注意力模块（Taget-Attention）。
自注意力模块（Self-Attention）用于对所述向量Q、向量K和向量V进行自注意力运算，得到11个自注意力向量，所述自注意力向量为128维向量。基于自注意力机制的原理，由向量Q、向量K和向量V可以得到一个新的包含自注意力信息的向量，即该向量包含了不同特征之间的关联度信息。
目标注意力模块（Taget-Attention）用于根据所述11个自注意力向量和目标向量，计算得到第三向量，所述第三向量为128维向量。即通过自注意力模块得到的包含了不同特征之间的关联度信息的向量，会与经过线性层处理的实时的温度数据进行一次目标注意力计算，得到一个包含了每个特征和温度标签关联度信息的向量。
具体地，目标注意力模块的内部实现方法包括：首先，分别将所述11个自注意力向量与所述目标向量进行点乘，得到11个点积，所述11个点积与11个自注意力向量一一对应。然后，对所述11个点积进行归一化处理，并保存归一化后的点积，本步骤可通过Softmax函数实现归一化处理，当然，在其他实施例中，也可采用其他函数实现归一化操作。接着，分别将所述11个自注意力向量乘以其对应的归一化后的点积，得到11个权重向量。最后，将所述11个权重向量相加，得到第三向量。
其中，归一化后的点积在训练过程中会趋于稳定，训练完后会得到一组归一化后的点积。后续在运用最优神经网络模型进行温度预测时，只输入终端数据的11个变量，进行独热编码后通过三个第四线性层和自注意力模块得到11个自注意力向量，在目标注意力模块中直接将11个自注意力向量与这组归一化后的点积相乘，得到11个权重向量，再将11个权重向量相加，输出第三向量。
如图6所示，八个注意力模型单元均会输出一个第三向量，然后分别发送至拼接模块。拼接模块对八个注意力模型单元输出的第三向量进行拼接，得到第四向量，所述第四向量为1024维向量；然后将第四向量发送至第三线性层。
第三线性层对所述第四向量进行编码，得到第五向量，所述第五向量为128维向量。
如图5所示，八头注意力模型（Multi-Head-Attention）将输出的第五向量发送至隐含层（Hidden Layer）。隐含层对所述第五向量进行编码，得到第六向量，所述第六向量为32维向量；然后将第六向量发送至激活层。
本实施例中，激活层的数量可为两个，隐含层分别将第六向量发送至两个激活层。两个激活层均用于根据预设的激活函数对所述第六向量进行计算，得到第七向量，所述第七向量为32维向量；然后两个激活层将其输出的第七向量分别发送至第一线性层和第二线性层。本实施例中，激活函数采用ReLu。
第一线性层对所述第七向量进行编码，得到数学期望，也即温度预测值；
第二线性层对所述第七向量进行编码，得到标准差，也即温度预测值的可信度。
构建完神经网络模型并获取到样本后，接着，根据所述训练集中的样本终端数据及其对应的温度数据对所述神经网络模型进行训练，得到本次训练对应的神经网络模型，并保存模型参数；根据所述验证集中的样本终端数据及其对应的温度数据对本次训练对应的神经网络模型进行验证，得到本次训练对应的平均绝对误差。
其中，对神经网络模型进行一次训练的过程如下：
遍历所述训练集，依次从所述训练集中获取一样本终端数据，作为当前训练样本；将所述当前训练样本及其对应的温度数据输入最新的神经网络模型进行正向传播，得到所述当前训练样本对应的第一输出结果，所述第一输出结果包括第一温度预测值及其可信度；根据预设的损失函数，计算所述当前训练样本对应的第一输出结果与所述当前训练样本对应的温度数据的损失值；根据所述损失值，通过反向传播算法更新所述神经网络模型。
当遍历完所述训练集后，得到本次训练对应的神经网络模型，并保存网络参数。
其中，最新的神经网络模型指的是经过之前各次训练后得到神经网络模型，也就是说，第一次训练对应的神经网络模型是基于预设的神经网络模型，通过训练集中的各训练样本训练得到的；第二次训练对应的神经网络模型是基于第一次训练对应的神经网络模型，通过训练集中的各训练样本训练得到的；以此类推。
本实施例中，采用的损失函数为
Figure dest_path_image002
，其中，μ为第一温度预测值，σ为第一温度预测值的可信度，T为对应的温度数据。
进一步地，对经过一次训练后的神经网络模型进行验证的过程如下：
遍历所述验证集，依次从所述验证集中获取一样本终端数据，作为当前验证样本；将所述当前验证样本及其对应的温度数据输入所述本次训练对应的神经网络模型进行正向传播，得到所述当前验证样本对应的第二输出结果，所述第二输出结果包括第二温度预测值及其可信度；计算所述当前验证样本对应的第二温度预测值与所述当前验证样本对应的温度数据的差值绝对值，得到当前验证样本对应的差值绝对值。
当遍历完所述验证集后，计算所述验证集中各样本终端数据对应的差值绝对值的平均值，得到本次训练对应的平均绝对误差。平均绝对误差即各验证样本的预测值与真实值之间的差值绝对值的平均值。
每次训练后，实时获取最小平均绝对误差；当获取到最小平均绝对误差之后连接预设次数的训练对应的平均绝对误差均大于所述最小平均绝对误差，则将所述最小平均绝对误差对应的神经网络模型作为最优神经网络模型。即在训练过程中，在验证集上的MAE（平均绝对误差）连续50个epoch没有下降时，则取该模型作为最终模型。
S402：采集终端数据，终端数据包括电池数据、屏幕使用数据和CPU使用数据。具体地，可按照预设的周期采集终端数据，如每1分钟采集一次终端数据。
本实施例中，一组终端数据总共包括11个变量。其中，所述电池数据包括电压、实时的电池温度传感器读数、上一次激活屏幕时的电池温度传感器读数以及上一次熄屏时的电池温度传感器读数，所述屏幕使用数据包括屏幕激活状态、屏幕持续激活时间、屏幕持续熄屏时间、上一次激活屏幕前的持续熄屏时间以及上一次熄屏前持续使用时间，所述CPU使用数据包括实时的CPU使用率以及平均CPU使用率。
其中，平均CPU使用率可通过计算最近50次采集的CPU使用率的平均值得到。
S403：将所述终端数据输入最优神经网络模型，输出温度预测结果，所述温度预测结果包括温度预测值及其可信度。
本实施例中，按照预设的周期采集依次终端数据后即进行一次温度预测。
S404：获取预设范围内的其他终端的温度预测结果，并根据自身的温度预测结果和所述其他终端的温度预测结果，获取可信度最高的温度预测值，作为最终温度预测值。
终端和终端之间可以通过蓝牙广播来交换本机的温度预测结果。终端会获取以自身为中心，预设半径范围内的其他终端的温度预测结果，然后在其他终端的温度预测结果和自身的温度预测结果中，选取可信度最高的温度预测值。
S405：将所述最终温度预测值以及位置信息发送至管理中心。其中，位置信息为终端当前的位置信息，可通过定位模型（如GPS）得到。
管理中心接收到各终端发送的最终温度预测值以及位置信息后，即可根据各位置信息对应的区域，确定各个区域的最终温度预测值。由于当终端处于某个区域的边缘时，可能会基于处于其他区域的终端的温度预测结果来确定最终温度预测值，因此，同一区域可能存在多个不同的最终温度预测值，管理中心可进一步根据这些最终温度预测值，确定出一个温度值，作为该区域的温度值。测温区域精度可以在GPS精度保证的前提下任意地缩小，保证了小区域的实时温度监测。
管理中心确定各个区域的温度值后，可通过整个园区不同区域的温度分布来合理调控制冷资源，例如控制小型洒水车前往高温室外地区，调控室内不同位置的空调功率达到高效节能的目的，提升游客舒适度。
本实施例的温度预测模型考虑了多种影响因素以及各影响因素之间的关联性，可保证温度预测的准确性，并且，温度预测模型会输出数据的可信度，可用于数据筛选，更具有广泛性和抗干扰性，能够适用于不同型号的手机之中。
本实施例通过采用众包的方式进行温度预测，将数据采集负担和计算负担平摊到游客的终端上，使得温度监控准确而高效。通过众包温度预测的方式，使得景区的管理中心可以根据各个区域的温度，进行制冷资源的调控，保证各个区域温度适宜，从而为游客带来更加舒适的体验。
实施例二
请参照图2，本发明的实施例二为：一种基于众包的温度预测装置，可执行本发明实施例一所提供的温度预测方法，具备执行方法相应的功能模块和有益效果。该装置可以由软件/或硬件实现，具体包括：
采集模块201，用于采集终端数据，终端数据包括电池数据、屏幕使用数据和CPU使用数据；
预测模块202，用于将所述终端数据输入预设的最优神经网络模型，输出温度预测结果，所述温度预测结果包括温度预测值及其可信度；
第一获取模块203，用于获取预设范围内的其他终端的温度预测结果，并根据自身的温度预测结果和所述其他终端的温度预测结果，获取可信度最高的温度预测值，作为最终温度预测值。
在一个可选的实施方式中，所述温度预测装置还包括：
发送模块，用于将所述最终温度预测值以及位置信息发送至管理中心，以使管理中心根据所述位置信息和最终温度预测值，进行温度调控。
在一个可选的实施方式中，所述采集模块具体用于根据预设的周期采集终端数据，所述终端数据包括电池数据、屏幕使用数据和CPU使用数据；
其中，所述电池数据包括电压、电池温度传感器读数、上一次激活屏幕时的电池温度传感器读数以及上一次熄屏时的电池温度传感器读数，所述屏幕使用数据包括屏幕激活状态、屏幕持续激活时间、屏幕持续熄屏时间、上一次激活屏幕前的持续熄屏时间以及上一次熄屏前持续使用时间，所述CPU使用数据包括CPU使用率以及平均CPU使用率。
在一个可选的实施方式中，所述温度预测装置还包括：
构建模块，用于构建神经网络模型；
第二获取模块，用于获取预设数量的样本终端数据及其对应的温度数据，并按照预设的比例对所述预设数量的样本终端数据及其对应的温度数据进行划分，得到训练集和验证集；
训练模块，用于根据所述训练集中的样本终端数据及其对应的温度数据对所述神经网络模型进行训练，得到本次训练对应的神经网络模型，并保存模型参数；
验证模块，用于根据所述验证集中的样本终端数据及其对应的温度数据对本次训练对应的神经网络模型进行验证，得到本次训练对应的平均绝对误差；
第三获取模块，用于每次训练后，实时获取最小平均绝对误差；
确定模块，用于当获取到最小平均绝对误差之后连接预设次数的训练对应的平均绝对误差均大于所述最小平均绝对误差，则将所述最小平均绝对误差对应的神经网络模型作为最优神经网络模型。
在一个可选的实施方式中，所述训练模块包括：
第一遍历单元，用于遍历所述训练集，依次从所述训练集中获取一样本终端数据，作为当前训练样本；
第一得到单元，用于将所述当前训练样本及其对应的温度数据输入最新的神经网络模型进行正向传播，得到所述当前训练样本对应的第一输出结果，所述第一输出结果包括第一温度预测值及其可信度；
第一计算单元，用于根据预设的损失函数，计算所述当前训练样本对应的第一输出结果与所述当前训练样本对应的温度数据的损失值；
更新单元，用于根据所述损失值，通过反向传播算法更新所述神经网络模型；
第二得到单元，用于当遍历完所述训练集后，得到本次训练对应的神经网络模型，并保存网络参数；
其中，所述损失函数为，μ为第一温度预测值，σ为第一温度预测值的可信度，T为对应的温度数据。
在一个可选的实施方式中，所述验证模块包括：
第二遍历单元，用于遍历所述验证集，依次从所述验证集中获取一样本终端数据，作为当前验证样本；
第三得到单元，用于将所述当前验证样本及其对应的温度数据输入所述本次训练对应的神经网络模型进行正向传播，得到所述当前验证样本对应的第二输出结果，所述第二输出结果包括第二温度预测值及其可信度；
第二计算单元，用于计算所述当前验证样本对应的第二温度预测值与所述当前验证样本对应的温度数据的差值绝对值，得到当前验证样本对应的差值绝对值；
第三计算单元，用于当遍历完所述验证集后，计算所述验证集中各样本终端数据对应的差值绝对值的平均值，得到本次训练对应的平均绝对误差。
实施例三
请参照图3，本发明的实施例三为：一种电子设备，所述电子设备包括：
一个或多个处理器301；
存储装置302，用于存储一个或多个程序；
当所述一个或多个程序被所述一个或多个处理器301执行，使得所述一个或多个处理器301实现如上所述的基于众包的温度预测方法实施例中的各个过程，且能达到相同的技术效果，为避免重复，这里不再赘述。
实施例四
本发明的实施例四提供了一种计算机可读存储介质，其上存储有计算机程序，所述计算机程序被处理器执行时实现如上所述的基于众包的温度预测方法实施例中的各个过程，且能达到相同的技术效果，为避免重复，这里不再赘述。
综上所述，本发明提供的一种基于众包的温度预测方法、装置、设备和存储介质，采用的神经网络模型考虑了多种影响因素，同时使用了多头自注意力机制和目标注意力机制来强化模型，通过使用注意力机制考虑输入变量之间的关联度以及输入变量与真实空气温度之间的关联度，使得最终模型的平均预测误差较小，并且，神经网络模型还会输出数据的可信度，更具有广泛性和抗干扰性，能够适用于不同的终端。通过采用众包的方式进行温度预测，将数据采集负担和计算负担平摊到各个终端上，使得温度监控准确而高效；管理中心可以根据各个终端上传的最终温度预测值和位置信息，确定各个区域的温度，并进行制冷资源的调控，保证各个区域温度适宜，从而为游客带来更加舒适的体验。
通过以上关于实施方式的描述，所属领域的技术人员可以清楚地了解到，本发明可借助软件及必需的通用硬件来实现，当然也可以通过硬件实现，但很多情况下前者是更佳的实施方式。基于这样的理解，本发明的技术方案本质上或者说对现有技术做出贡献的部分可以以软件产品的形式体现出来，该计算机软件产品可以存储在计算机可读存储介质中，如计算机的软盘、只读存储器（Read-Only Memory，ROM）、随机存取存储器（RandomAccess Memory，RAM）、闪存（FLASH）、硬盘或光盘等，包括若干指令用以使得一台计算机设备（可以是个人计算机，服务器，或者网络设备等）执行本发明各个实施例所述的方法。
值得注意的是，上述装置的实施例中，所包括的各个单元和模块只是按照功能逻辑进行划分的，但并不局限于上述的划分，只要能够实现相应的功能即可；另外，各功能单元的具体名称也只是为了便于相互区分，并不用于限制本发明的保护范围。
以上所述仅为本发明的实施例，并非因此限制本发明的专利范围，凡是利用本发明说明书及附图内容所作的等同变换，或直接或间接运用在相关的技术领域，均同理包括在本发明的专利保护范围内。

Claims (11)

  1. 一种基于众包的温度预测方法，其特征在于，包括：
    采集终端数据，终端数据包括电池数据、屏幕使用数据和CPU使用数据；
    将所述终端数据输入预设的最优神经网络模型，输出温度预测结果，所述温度预测结果包括温度预测值及其可信度；
    获取预设范围内的其他终端的温度预测结果，并根据自身的温度预测结果和所述其他终端的温度预测结果，获取可信度最高的温度预测值，作为最终温度预测值。
  2. 根据权利要求1所述的基于众包的温度预测方法，其特征在于，所述获取预设范围内的其他终端的温度预测结果，并根据自身的温度预测结果和所述其他终端的温度预测结果，获取可信度最高的温度预测值，作为最终温度预测值之后，所述方法还包括：
    将所述最终温度预测值以及位置信息发送至管理中心，以使管理中心根据所述位置信息和最终温度预测值，进行温度调控。
  3. 根据权利要求1所述的基于众包的温度预测方法，其特征在于，所述采集终端数据，终端数据包括电池数据、屏幕使用数据和CPU使用数据，包括：
    根据预设的周期采集终端数据，所述终端数据包括电池数据、屏幕使用数据和CPU使用数据；
    其中，所述电池数据包括电压、电池温度传感器读数、上一次激活屏幕时的电池温度传感器读数以及上一次熄屏时的电池温度传感器读数，所述屏幕使用数据包括屏幕激活状态、屏幕持续激活时间、屏幕持续熄屏时间、上一次激活屏幕前的持续熄屏时间以及上一次熄屏前持续使用时间，所述CPU使用数据包括CPU使用率以及平均CPU使用率。
  4. 根据权利要求1所述的基于众包的温度预测方法，其特征在于，所述将所述终端数据输入预设的最优神经网络模型，输出温度预测结果之前，所述方法还包括：
    构建神经网络模型；
    获取预设数量的样本终端数据及其对应的温度数据，并按照预设的比例对所述预设数量的样本终端数据及其对应的温度数据进行划分，得到训练集和验证集；
    根据所述训练集中的样本终端数据及其对应的温度数据对所述神经网络模型进行训练，得到本次训练对应的神经网络模型，并保存模型参数；
    根据所述验证集中的样本终端数据及其对应的温度数据对本次训练对应的神经网络模型进行验证，得到本次训练对应的平均绝对误差；
    每次训练后，实时获取最小平均绝对误差；
    当获取到最小平均绝对误差之后连接预设次数的训练对应的平均绝对误差均大于所述最小平均绝对误差，则将所述最小平均绝对误差对应的神经网络模型作为最优神经网络模型。
  5. 根据权利要求4所述的基于众包的温度预测方法，其特征在于，所述神经网络模型包括独热编码模块、八头注意力模型、隐含层、激活层、第一线性层和第二线性层，所述八头注意力模型包括八个注意力模型单元以及拼接模块和第三线性层，所述注意力模型单元包括四个第四线性层以及自注意力模块和目标注意力模块；
    所述独热编码模块用于分别对输入的样本终端数据中的11个变量进行独热编码，得到11个第一向量，所述第一向量为12维向量；
    所述独热编码模块还用于对输入的温度数据进行独热编码，得到第二向量，所述第二向量为12维向量；
    所述四个第四线性层中的三个第四线性层用于对所述11个第一向量进行编码，得到向量Q、向量K和向量V，所述向量Q、向量K和向量V均为128维向量；
    所述四个第四线性层中剩余的一个第四线性层用于对所述第二向量进行编码，得到目标向量，所述目标向量为128维向量；
    所述自注意力模块用于对所述向量Q、向量K和向量V进行自注意力运算，得到11个自注意力向量，所述自注意力向量为128维向量；
    所述目标注意力模块用于根据所述11个自注意力向量和目标向量，计算得到第三向量，所述第三向量为128维向量；
    所述拼接模块用于对八个注意力模型单元输出的第三向量进行拼接，得到第四向量，所述第四向量为1024维向量；
    所述第三线性层用于对所述第四向量进行编码，得到第五向量，所述第五向量为128维向量；
    所述隐含层用于对所述第五向量进行编码，得到第六向量，所述第六向量为32维向量；
    所述激活层用于根据预设的激活函数对所述第六向量进行计算，得到第七向量，所述第七向量为32维向量；
    所述第一线性层用于对所述第七向量进行编码，得到温度预测值；
    所述第二线性层用于对所述第七向量进行编码，得到所述温度预测值的可信度。
  6. 根据权利要求5所述的基于众包的温度预测方法，其特征在于，所述根据所述11个自注意力向量和目标向量，计算得到第三向量，包括：
    分别将所述11个自注意力向量与所述目标向量进行点乘，得到11个点积，所述11个点积与11个自注意力向量一一对应；
    对所述11个点积进行归一化处理，并保存归一化后的点积；
    分别将所述11个自注意力向量乘以其对应的归一化后的点积，得到11个权重向量；
    将所述11个权重向量相加，得到第三向量。
  7. 根据权利要求4所述的基于众包的温度预测方法，其特征在于，所述根据所述训练集中的样本终端数据及其对应的温度数据对所述神经网络模型进行训练，得到本次训练对应的神经网络模型，并保存模型参数，包括：
    遍历所述训练集，依次从所述训练集中获取一样本终端数据，作为当前训练样本；
    将所述当前训练样本及其对应的温度数据输入最新的神经网络模型进行正向传播，得到所述当前训练样本对应的第一输出结果，所述第一输出结果包括第一温度预测值及其可信度；
    根据预设的损失函数，计算所述当前训练样本对应的第一输出结果与所述当前训练样本对应的温度数据的损失值；
    根据所述损失值，通过反向传播算法更新所述神经网络模型；
    当遍历完所述训练集后，得到本次训练对应的神经网络模型，并保存网络参数；
    其中，所述损失函数为
    Figure dest_path_image001
    ，μ为第一温度预测值，σ为第一温度预测值的可信度，T为对应的温度数据。
  8. 根据权利要求4所述的基于众包的温度预测方法，其特征在于，所述根据所述验证集中的样本终端数据及其对应的温度数据对本次训练对应的神经网络模型进行验证，得到本次训练对应的平均绝对误差，包括：
    遍历所述验证集，依次从所述验证集中获取一样本终端数据，作为当前验证样本；
    将所述当前验证样本及其对应的温度数据输入所述本次训练对应的神经网络模型进行正向传播，得到所述当前验证样本对应的第二输出结果，所述第二输出结果包括第二温度预测值及其可信度；
    计算所述当前验证样本对应的第二温度预测值与所述当前验证样本对应的温度数据的差值绝对值，得到当前验证样本对应的差值绝对值；
    当遍历完所述验证集后，计算所述验证集中各样本终端数据对应的差值绝对值的平均值，得到本次训练对应的平均绝对误差。
  9. 一种基于众包的温度预测装置，其特征在于，包括：
    采集模块，用于采集终端数据，终端数据包括电池数据、屏幕使用数据和CPU使用数据；
    预测模块，用于将所述终端数据输入预设的最优神经网络模型，输出温度预测结果，所述温度预测结果包括温度预测值及其可信度；
    第一获取模块，用于获取预设范围内的其他终端的温度预测结果，并根据自身的温度预测结果和所述其他终端的温度预测结果，获取可信度最高的温度预测值，作为最终温度预测值。
  10. 一种电子设备，其特征在于，所述电子设备包括：
    一个或多个处理器；
    存储装置，用于存储一个或多个程序；
    当所述一个或多个程序被所述一个或多个处理器执行，使得所述一个或多个处理器实现如权利要求1-8中任一所述的基于众包的温度预测方法。
  11. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现如权利要求1-8任一项所述的基于众包的温度预测方法的步骤。
PCT/CN2021/121177 2021-08-23 2021-09-28 基于众包的温度预测方法、装置、设备和存储介质 WO2023024213A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110966275.0A CN113419868B (zh) 2021-08-23 2021-08-23 基于众包的温度预测方法、装置、设备和存储介质
CN202110966275.0 2021-08-23

Publications (1)

Publication Number Publication Date
WO2023024213A1 true WO2023024213A1 (zh) 2023-03-02

Family

ID=77719175

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/121177 WO2023024213A1 (zh) 2021-08-23 2021-09-28 基于众包的温度预测方法、装置、设备和存储介质

Country Status (2)

Country Link
CN (1) CN113419868B (zh)
WO (1) WO2023024213A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116306032A (zh) * 2023-05-17 2023-06-23 宁德时代新能源科技股份有限公司 温度预测方法、装置、设备以及存储介质
CN116470190A (zh) * 2023-06-19 2023-07-21 江西五十铃汽车有限公司 动力电池包加热方法、系统、计算机及可读存储介质
CN117309195A (zh) * 2023-11-30 2023-12-29 中控技术股份有限公司 测温仪表的检测方法、装置及电子设备

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113419868B (zh) * 2021-08-23 2021-11-16 南方科技大学 基于众包的温度预测方法、装置、设备和存储介质
CN114139805A (zh) * 2021-11-29 2022-03-04 杭州逗酷软件科技有限公司 温度预测及模型训练方法和相关装置

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107610087A (zh) * 2017-05-15 2018-01-19 华南理工大学 一种基于深度学习的舌苔自动分割方法
CN112766337A (zh) * 2021-01-11 2021-05-07 中国科学院计算技术研究所 用于预测众包数据的正确标签的方法及系统
CN112834067A (zh) * 2020-12-31 2021-05-25 南方科技大学 移动终端获取环境温度的方法、装置、移动终端及介质
CN112910690A (zh) * 2021-01-18 2021-06-04 武汉烽火技术服务有限公司 基于神经网络模型的网络流量预测方法、装置及设备
US20210201147A1 (en) * 2018-11-28 2021-07-01 Tencent Technology (Shenzhen) Company Limited Model training method, machine translation method, computer device, and storage medium
CN113419868A (zh) * 2021-08-23 2021-09-21 南方科技大学 基于众包的温度预测方法、装置、设备和存储介质

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101424747A (zh) * 2008-11-28 2009-05-06 飞图科技（北京）有限公司 一种手机的对局部区域内的天气状况预测的方法
KR20190102391A (ko) * 2018-02-26 2019-09-04 포항공과대학교 산학협력단 인공 신경망을 이용한 개인의 만족도 예측을 통해 건물의 열 에너지 시스템을 제어하는 방법 및 장치
CN108507700A (zh) * 2018-03-20 2018-09-07 中国农业大学 一种猪舍多点温度预测方法及系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107610087A (zh) * 2017-05-15 2018-01-19 华南理工大学 一种基于深度学习的舌苔自动分割方法
US20210201147A1 (en) * 2018-11-28 2021-07-01 Tencent Technology (Shenzhen) Company Limited Model training method, machine translation method, computer device, and storage medium
CN112834067A (zh) * 2020-12-31 2021-05-25 南方科技大学 移动终端获取环境温度的方法、装置、移动终端及介质
CN112766337A (zh) * 2021-01-11 2021-05-07 中国科学院计算技术研究所 用于预测众包数据的正确标签的方法及系统
CN112910690A (zh) * 2021-01-18 2021-06-04 武汉烽火技术服务有限公司 基于神经网络模型的网络流量预测方法、装置及设备
CN113419868A (zh) * 2021-08-23 2021-09-21 南方科技大学 基于众包的温度预测方法、装置、设备和存储介质

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116306032A (zh) * 2023-05-17 2023-06-23 宁德时代新能源科技股份有限公司 温度预测方法、装置、设备以及存储介质
CN116306032B (zh) * 2023-05-17 2023-08-25 宁德时代新能源科技股份有限公司 温度预测方法、装置、设备以及存储介质
CN116470190A (zh) * 2023-06-19 2023-07-21 江西五十铃汽车有限公司 动力电池包加热方法、系统、计算机及可读存储介质
CN116470190B (zh) * 2023-06-19 2023-09-22 江西五十铃汽车有限公司 动力电池包加热方法、系统、计算机及可读存储介质
CN117309195A (zh) * 2023-11-30 2023-12-29 中控技术股份有限公司 测温仪表的检测方法、装置及电子设备
CN117309195B (zh) * 2023-11-30 2024-04-12 中控技术股份有限公司 测温仪表的检测方法、装置及电子设备

Also Published As

Publication number Publication date
CN113419868B (zh) 2021-11-16
CN113419868A (zh) 2021-09-21

Similar Documents

Publication Publication Date Title
WO2023024213A1 (zh) 基于众包的温度预测方法、装置、设备和存储介质
CN110285532B (zh) 基于人工智能的机房空调控制方法、装置及系统
Chen et al. Deep reinforcement learning for computation offloading in mobile edge computing environment
Kumrai et al. Multiobjective optimization in cloud brokering systems for connected Internet of Things
Yang et al. Ai-powered green cloud and data center
CN110739696B (zh) 智能配网环境中需求侧资源与可再生能源的集成调度方法
CN110866592B (zh) 模型训练方法、装置、能效预测方法、装置和存储介质
KR102128363B1 (ko) 에너지 거래 관리 시스템 및 그 방법
CN113348330B (zh) 空调系统的管理方法、控制方法及存储介质、控制平台
Arouj et al. Towards energy-aware federated learning on battery-powered clients
US11790289B2 (en) Systems and methods for managing dynamic transportation networks using simulated future scenarios
Kiljander et al. Intelligent consumer flexibility management with neural network-based planning and control
Cai et al. Dynacomm: Accelerating distributed CNN training between edges and clouds through dynamic communication scheduling
Sun et al. Task offloading method of internet of vehicles based on cloud-edge computing
CN117134380A (zh) 一种基于云边协同分布式储能分层级优化运行方法及系统
CN109413746B (zh) 一种混合能源供能的通信系统中最优化能量分配方法
CN115051374A (zh) 电采暖设备参与电力调峰的控制方法、装置和存储介质
KR102383321B1 (ko) 자가 발전원을 갖는 네트워크 내에서의 에너지 관리 방법, 이를 수행하기 위한 장치 및 시스템
US11687689B2 (en) Capacity forecasting
Yang et al. A noval bidding strategy of electric vehicles participation in ancillary service market
Liu et al. A Proactive Stable Scheme for Vehicular Collaborative Edge Computing
Yang et al. Non-cooperative game approach for task offloading in edge clouds
CN113259472A (zh) 一种面向视频分析任务的边缘节点资源分配方法
Simoiu et al. Markov Chain model for energy community member reaction to recommendations: a multi-agent case study
Ke et al. Service caching decision‐making policy for mobile edge computing using deep reinforcement learning

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21954733

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE