Human cold and warm comfort prediction method based on deep learning

Abstract

The application is a human cold-heat comfort degree prediction method based on deep learning. First, a data set is obtained, each sample of the data set includes an environmental parameter sequence and an individual physiological parameter sequence, and the sample label includes two kinds of subjective cold-heat feeling and comfort degree. Then, the data set is preprocessed, and the preprocessed data is sliced to obtain a plurality of data slices. Then, a cold-heat feeling classification model is constructed to extract deep features reflecting human cold-heat feeling and cold-heat feeling prediction. A cold comfort classification model and a hot comfort classification model are constructed. Finally, the trained cold-heat feeling classification model, cold comfort classification model and hot comfort classification model are reorganized to obtain a cold-heat comfort degree prediction model. According to the cold-heat feeling prediction result of the cold-heat feeling classification model, the deep features are input into the cold comfort classification model or the hot comfort classification model for comfort degree prediction. The method follows the neural mechanism of human body temperature regulation, distinguishes the comfort degree under different feelings, respectively predicts the cold-heat feeling and the comfort degree, and is more in line with the subjective real feeling of individuals.

Description

Technical Field

[0001] This invention belongs to the field of thermal comfort prediction technology, specifically a method for predicting human thermal comfort based on deep learning. Background Technology

[0002] Thermal comfort prediction serves a dual purpose: on the one hand, it provides users with an optimal indoor environment; on the other hand, it improves the comfort of the built environment and the energy efficiency of buildings. The PMV model is currently one of the mainstream methods for thermal comfort prediction. The evaluation index used in the model to characterize human thermal response represents the average temperature sensation of most people in the same environment. As a general indicator of the thermal environment, it is difficult to predict individual temperature comfort in real time based on factors such as clothing thermal resistance and individual differences. While this model has reference value for assessing the thermal comfort of a population, it is difficult to accurately predict individual thermal comfort. Furthermore, due to subjectivity and individual differences, the quantification and prediction of temperature comfort remains challenging.

[0003] With the development of high-performance sensors and artificial intelligence, thermal comfort models built using data-driven and machine learning approaches are gradually becoming a future trend. Existing thermal comfort models primarily consider individual differences, such as age, gender, and clothing thermal resistance, acquiring human physiological indicators through smart wearable devices and then using traditional machine learning techniques to predict thermal comfort. However, the neural mechanisms of human thermoregulation indicate that temperature perception is mainly mediated by two types of sensory neurons, activated in warm and cold environments respectively. The degree and manner of activation differ in different environments, resulting in different characteristics of human physiological parameters under different sensations of heat and cold. Predicting thermal comfort under different sensations using the same thermal comfort model suffers from poor model specificity and low prediction accuracy.

[0004] To address this, this invention proposes a deep learning-based method for predicting human thermal comfort. It extracts deep features under different sensations of hot and cold, and uses data-driven modeling techniques to establish comfort prediction models for both sensations of hot and cold, thereby achieving the goal of real-time prediction of thermal sensations and comfort. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide a deep learning-based method for predicting human thermal comfort.

[0006] The present invention solves the aforementioned technical problem by adopting the following technical solution:

[0007] A deep learning-based method for predicting human thermal comfort, characterized by the following steps:

[0008] Step 1: Obtain the dataset. The dataset consists of several samples. Each sample includes a sequence of environmental parameters and a sequence of individual physiological parameters. The sample labels include two types: subjective feelings of hot and cold and comfort.

[0009] Step 2: Preprocess the dataset;

[0010] Step 3: Slice the preprocessed data to obtain several data slices, and divide them into training and test sets;

[0011] Step 4: Construct a cold and heat sensation classification model, which includes an encoder and a decoder to extract deep features reflecting human cold and heat sensation and to predict cold and heat sensation. Based on the cold and heat sensation prediction results, group all the deep features extracted by the cold and heat sensation classification model to obtain a cold comfort dataset and a hot comfort dataset.

[0012] Step 5: Construct a cold comfort classification model and a hot comfort classification model. Train the cold comfort classification model using the cold comfort dataset and the hot comfort classification model using the hot comfort dataset.

[0013] Step 6: Reorganize the trained hot and cold sensation classification model, cold comfort classification model, and hot comfort classification model to obtain a hot and cold comfort prediction model. Use the trained hot and cold comfort prediction model to predict human hot and cold comfort.

[0014] The thermal comfort prediction model includes an encoder and a decoder. The encoder is an encoder for the thermal sensation classification model, and the decoder includes a decoder for the thermal sensation classification model, a thermal comfort classification model, and a thermal comfort classification model. Based on the thermal sensation prediction results of the thermal sensation classification model, the deep features extracted by the encoder are input into the thermal comfort classification model or the thermal comfort classification model to predict comfort.

[0015] Furthermore, individual physiological parameters under different environmental conditions were obtained, and subjective evaluations of hot and cold sensations and comfort were conducted. Environmental parameters included ambient temperature and humidity, while individual physiological parameters included temperature and pulse rate at four sites: wrist, chest, forearm, and calf. In the subjective evaluation, hot and cold sensations were divided into seven levels: cold, cool, slightly cool, neutral, slightly warm, warm, and hot. Comfort was divided into seven levels: very uncomfortable, uncomfortable, slightly uncomfortable, neutral, slightly comfortable, comfortable, and very comfortable.

[0016] Furthermore, in step 2, preprocessing includes outlier removal, mapping of hot / cold sensations and comfort levels to labels according to levels, individual physiological parameter variations, normalization of human physiological parameters and individual physiological parameter variations, and smoothing and standardization of input data. Among these, hot / cold sensations are mapped to three categories of labels: cold, neutral, and hot, and comfort levels are mapped to two categories of labels: comfortable and uncomfortable. Individual physiological parameter variations include the variations in individual physiological parameters within a time period before the prediction time and the variations in individual physiological parameters within a time period after the prediction time.

[0017] Furthermore, the normalization calculation formula is as follows:

[0018]

[0019]

[0020] L norm =k·Weight 0.425 ·Height 0.725 ·G(5)

[0021] In the formula: For normalized human physiological parameters, L norm As the normalization factor, X represents the normalized change in an individual's physiological parameters. grad_past X grad_future To predict the changes in an individual's physiological parameters over a period of time before and after a given moment, k is the thermal resistance coefficient of the clothing, Weight is the individual's body weight, Height is the individual's height, and G is the sex coefficient.

[0022] Furthermore, the cold and heat sensation classification model, the cold comfort classification model, and the hot comfort classification model all include two bidirectional long short-term memory network layers, a long short-term memory network layer, and a fully connected layer. The bidirectional long short-term memory network layer and the long short-term memory network layer are followed by a Dropout layer and a batch normalization layer. The fully connected layer is followed by a ReLU layer and a Softmax classification layer.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. This invention follows the neural mechanism of human body temperature regulation, distinguishing between comfort levels under different sensations of cold and heat. It establishes separate classification models for cold comfort and hot comfort, and trains the two models using cold comfort datasets and hot comfort datasets respectively. Therefore, the parameters of the two models differ, making the models more targeted. Compared with traditional hot comfort models, it is closer to the real situation of human body temperature regulation mechanism, which is conducive to improving prediction accuracy and enhancing the interpretability and adaptability of the model.

[0025] 2. The prediction content of this invention includes two parts: hot and cold sensation and comfort. In fact, users pay more attention to comfort than hot and cold sensation. There is a comfortable state under both hot and cold sensation. Therefore, a hot and cold sensation classification model is constructed to classify hot and cold sensation according to environmental parameters and human physiological parameters. Then, the extracted deep features are input into the cold comfort classification model or the hot comfort classification model to predict comfort based on the hot and cold sensation classification results.

[0026] 3. To address individual differences, preprocessing individual physiological parameters can effectively reduce the impact of individual differences on prediction results, thereby improving prediction accuracy. Based on the real subjective feelings of the test users regarding hot and cold sensations, the hot and cold sensation classification model is trained to better reflect the users' actual feelings.

[0027] 4. In practical applications, real-time collection of environmental and individual physiological parameters can predict individual thermal comfort. The prediction results can be used as input parameters for controllers of air conditioning and other thermal control equipment to regulate the building environment, which is of great significance for improving the comfort of the building environment and energy efficiency. Attached Figure Description

[0028] Figure 1 This is an overall flowchart of the present invention;

[0029] Figure 2 This is a schematic diagram of the sliding window slicing principle of the present invention;

[0030] Figure 3 This is a structural diagram of the cold and heat sensation classification model of the present invention;

[0031] Figure 4 This is a structural diagram of the cold comfort classification model and the hot comfort classification model of the present invention. Detailed Implementation

[0032] Specific embodiments will be given below with reference to the accompanying drawings. These specific embodiments are only used to describe the technical solution of the present invention in detail, and are not intended to limit the scope of protection of this application.

[0033] This invention provides a deep learning-based method for predicting human thermal comfort, such as... Figure 1 As shown, the specific steps are as follows:

[0034] Step 1: Obtain the dataset. The dataset consists of several samples. Each sample includes a sequence of environmental parameters and a sequence of individual physiological parameters. The sample labels include two types of labels: subjective hot and cold sensation and comfort level, which reflect the individual's subjective feelings about the hot and cold environment.

[0035] This experiment, based on individual subjective evaluation of environmental changes, acquired individual physiological parameters under different environmental conditions and conducted subjective evaluations of hot and cold sensations and comfort. Environmental parameters included ambient temperature and humidity, while individual physiological parameters included temperature and pulse rate at four sites: wrist, chest, forearm, and calf. Temperature was sampled once per minute, and pulse rate was sampled once per second. To ensure consistent sampling frequency, the pulse rate was measured at the first second of each minute. Participants filled out a hot and cold sensation and comfort evaluation form every five minutes based on their subjective feelings. Hot and cold sensations were categorized into seven levels: cold, cool, slightly cool, neutral, slightly warm, warm, and hot, denoted as [-3,3]. Comfort levels were categorized into seven levels: very uncomfortable, uncomfortable, slightly uncomfortable, neutral, slightly comfortable, comfortable, and very comfortable, denoted as [-3,3].

[0036] Step 2: Preprocess the dataset;

[0037] (1) Remove outliers: If environmental parameters and individual physiological parameters change abruptly or exceed a reasonable range, the mean of the data before and after the outlier is used to replace the current value;

[0038] (2) The sensation of hot and cold is mapped to three categories of labels: cold, neutral and hot. -3 and -2 correspond to cold, -1, 0 and 1 correspond to neutral, and 2 and 3 correspond to hot. The comfort level is mapped to two categories of labels: comfortable and uncomfortable. [-3,0] corresponds to uncomfortable and [1,3] corresponds to comfortable.

[0039] (3) Calculating the changes in individual physiological parameters: The human body's sensation of hot and cold and its perception of comfort are the result of a slow heat exchange metabolic process, which is related not only to the current physiological state of the body, but also to the changes in physiological parameters over a period of time. Therefore, including the changes in individual physiological parameters over a period of time as part of the model input can improve the prediction accuracy; the formula for calculating the changes in individual physiological parameters is:

[0040] X grad_past =X(t)-X(t-△t1) (1)

[0041] X grad_future =X(t+Δt2)-X(t) (2)

[0042] Where X(t) represents the human physiological parameters at time t, and X(t-Δt1) and X(t+Δt2) represent the human physiological parameters at times t-Δt1 and t+Δt2, respectively. grad_past X represents the change in an individual's physiological parameters within Δt1 before time t. grad_future This represents the change in individual physiological parameters within Δt2 after time t.

[0043] (4) Normalize the human physiological parameters and individual physiological parameter variations based on the subject's specific parameters such as gender, height, and weight. The normalization calculation formula is as follows:

[0044]

[0045]

[0046] L norm =k·Weight 0.425 ·Height 0.725 ·G (5)

[0047] In the formula: For normalized human physiological parameters, L norm As the normalization factor, The normalized individual physiological parameter variation is represented by k, clothing thermal resistance coefficient, weight (kg), height (m), and gender coefficient.

[0048] (5) Smooth the input data and standardize it to 0-1:

[0049] The smoothing cost function is:

[0050]

[0051] In the formula: R(s,p) is the penalty coefficient, s(x) i ) is the fitting function, y i Let p be the i-th original data to be fitted, and p∈[0,1] be the smoothing coefficient. When p=0, the function to be fitted is the least squares fitting function; when p=1, the function to be fitted is the cubic spline interpolation function. The value of p is usually between [0,1] and [1,2]. Nearby, h is the average distance between data points;

[0052] Equation (6) means finding a fitting function s(x) among all functions that have continuous second derivatives. i To minimize the penalty coefficient, substitute the original data into the fitting function s(x). i In the process, the smoothed data is obtained;

[0053] The 0-1 normalized transition function is:

[0054]

[0055] In the formula: To standardize the data before and after, x min x is the minimum value of each set of data. max This represents the maximum value for each set of data.

[0056] Step 3: Divide the preprocessed dataset into training and test sets;

[0057] The preprocessed dataset is sliced ​​using a sliding window slicing method to obtain several two-dimensional data slices; the principle is as follows: Figure 2 As shown, in this embodiment, the time step is 25 and the sliding step (sliding distance) is 10. All data slices are divided into three groups according to the feeling of cold and hot, corresponding to cold, neutral and hot categories respectively. To solve the problem of uneven data distribution, data augmentation is performed by resampling the data slices with fewer data to balance the distribution of the three groups of data. Finally, the training set and test set are divided according to the group disordered split cross-validation method. This embodiment uses a five-fold cross-validation iterator with a batch size of 64.

[0058] Step 4: Construct a hot and cold sensation classification model to extract deep features reflecting human hot and cold sensations and to predict hot and cold sensations;

[0059] like Figure 3 As shown, the hot / cold sensation classification model includes an encoder and a decoder. The encoder employs two bidirectional long short-term memory (Bi-LSTM) network layers, each followed by a dropout layer and a batch normalization layer. The encoder extracts deep features reflecting human hot / cold sensations. These features serve as input to either the cold comfort classification model or the hot comfort classification model, and also as input to the decoder for hot / cold sensation classification. The decoder includes a long short-term memory (LSTM) network layer, a fully connected layer, a ReLU layer, and a softmax classification layer. To improve the model's convergence speed and prevent overfitting, a dropout layer and a batch normalization layer are added between the LSTM network layer and the fully connected layer. In this embodiment, the fully connected layer contains 100 neurons. The input data dimension of the hot / cold sensation classification model is (64, 25, 11), the output feature dimension is (64, 2), and the deep feature dimension is (64, 25, 64).

[0060] The preprocessed dataset is input into the cold and heat sensation classification model to extract deep features and predict cold and heat sensations. Based on the prediction results, the deep features of cold and heat sensations are grouped to form cold comfort datasets and hot comfort datasets for the next step of model training.

[0061] Step 5: Construct separate cold comfort and hot comfort classification models, and train both models using the cold comfort and hot comfort datasets. The two models are used to fit different human receptors to predict comfort levels under different sensations of cold and heat. The cold comfort classification model predicts the user's comfort level under cold sensations (cold comfort and cold discomfort); the hot comfort classification model predicts the user's comfort level under hot sensations (hot comfort and hot discomfort). Figure 4 As shown, the architectures of the cold comfort classification model and the hot comfort classification model are basically the same. The model parameters are obtained through training, including two bidirectional long short-term memory network layers, a long short-term memory network layer, a fully connected layer, a dropout layer and a batch normalization layer after the bidirectional long short-term memory network layer and the long short-term memory network layer, and a ReLU layer and a Softmax classification layer after the fully connected layer.

[0062] Step 6: Reconstruct and combine the trained hot and cold sensation classification model, cold comfort classification model, and hot comfort classification model to obtain a hot and cold comfort prediction model, which is used to predict human hot and cold comfort.

[0063] like Figure 1 As shown, the cold and heat comfort prediction model includes an encoder and a decoder. The encoder is the encoder of the cold and heat sensation classification model. The decoder consists of three parts: the decoder of the cold and heat sensation classification model, the cold comfort classification model, and the hot comfort classification model. Based on the cold and heat sensation prediction results of the cold and heat sensation classification model, the deep features extracted by the encoder are input into the cold comfort classification model or the hot comfort classification model for secondary prediction. The final prediction result of the cold and heat sensation classification model is cold comfort, hot comfort, cold discomfort, or hot discomfort. If the prediction result of the cold and heat sensation classification model decoder is neutral, it is directly regarded as comfortable.

[0064] Precision, recall, and accuracy were selected as three evaluation metrics to assess the performance of each model. For the hot / cold sensation classification model, hot / cold sensation was selected as positive examples, and neutral was selected as negative examples. For the cold comfort and hot comfort classification models, hot comfort and cold comfort were selected as positive examples, and hot discomfort and cold discomfort were selected as negative examples. T... p For a real example, F p F is a false positive example. n As a false counterexample, T n This is a true counterexample.

[0065] The formula for calculating accuracy is:

[0066]

[0067] The formula for calculating accuracy is:

[0068]

[0069] The formula for calculating recall rate is:

[0070]

[0071] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for predicting human thermal comfort based on deep learning, characterized in that, Includes the following steps: Step 1: Obtain the dataset. The dataset consists of several samples. Each sample includes a sequence of environmental parameters and a sequence of individual physiological parameters. The sample labels include two types: subjective feelings of hot and cold and comfort. Step 2: Preprocess the dataset; Step 3: Slice the preprocessed data to obtain several data slices, and divide them into training and test sets; Step 4: Construct a hot and cold sensation classification model, which includes an encoder and a decoder to extract deep features reflecting human hot and cold sensations and to predict hot and cold sensations. Based on the prediction results of cold and hot sensation, all deep features extracted by the cold and hot sensation classification model are grouped to obtain the cold comfort dataset and the hot comfort dataset. Step 5: Construct a cold comfort classification model and a hot comfort classification model. Train the cold comfort classification model using the cold comfort dataset and the hot comfort classification model using the hot comfort dataset. Step 6: Reorganize the trained hot and cold sensation classification model, cold comfort classification model, and hot comfort classification model to obtain a hot and cold comfort prediction model. Use the trained hot and cold comfort prediction model to predict human hot and cold comfort. The thermal comfort prediction model includes an encoder and a decoder. The encoder is an encoder for the thermal sensation classification model, and the decoder includes a decoder for the thermal sensation classification model, a thermal comfort classification model, and a thermal comfort classification model. Based on the thermal sensation prediction results of the thermal sensation classification model, the deep features extracted by the encoder are input into the thermal comfort classification model or the thermal comfort classification model to predict comfort.

2. The method for predicting human thermal comfort based on deep learning according to claim 1, characterized in that, In step 1, individual physiological parameters under different environmental parameters are obtained, and subjective evaluations of hot and cold sensations and comfort are conducted. Environmental parameters include ambient temperature and humidity, and individual physiological parameters include temperature and pulse rate of four parts: wrist, chest, forearm, and calf. In the subjective evaluation, hot and cold sensations are divided into seven levels: cold, cool, slightly cool, neutral, slightly warm, warm, and hot. Comfort is divided into seven levels: very uncomfortable, uncomfortable, slightly uncomfortable, neutral, slightly comfortable, comfortable, and very comfortable.

3. The method for predicting human thermal comfort based on deep learning according to claim 1 or 2, characterized in that, In step 2, preprocessing includes outlier removal, mapping of hot and cold sensations and comfort levels to labels according to levels, individual physiological parameter variations, normalization of human physiological parameters and individual physiological parameter variations, and smoothing and standardization of input data. Among them, the sensation of hot and cold is mapped to three categories of labels: cold, neutral, and hot, and the level of comfort is mapped to two categories of labels: comfortable and uncomfortable. The changes in individual physiological parameters include the changes in individual physiological parameters in the time period before the prediction time and the changes in individual physiological parameters in the time period after the prediction time.

4. The method for predicting human thermal comfort based on deep learning according to claim 3, characterized in that, The normalization calculation formula is: L norm =k·Weight 0.425 ·Height 0.725 ·G(5) In the formula: For normalized human physiological parameters, L norm As the normalization factor, X represents the normalized change in an individual's physiological parameters. grad_past X grad_future To predict the changes in an individual's physiological parameters over a period of time before and after a given moment, k is the thermal resistance coefficient of the clothing, Weight is the individual's body weight, Height is the individual's height, and G is the sex coefficient.

5. The method for predicting human thermal comfort based on deep learning according to claim 1, characterized in that, The cold and heat sensation classification model, the cold comfort classification model, and the hot comfort classification model all include two bidirectional long short-term memory network layers, a long short-term memory network layer, and a fully connected layer. The bidirectional long short-term memory network layer and the long short-term memory network layer are followed by a Dropout layer and a batch normalization layer. The fully connected layer is followed by a ReLU layer and a Softmax classification layer.

Images