Air conditioner terminal intelligent control method based on user behavior preference

Abstract

The invention provides an air conditioner terminal intelligent control method based on user behavior preferences. The method comprises the steps that 1, an Internet of Things monitoring platform is built; 2, extracting user and behavior monitoring data, and performing preprocessing; step 3, obtaining the probability of occurrence of the target behavior aw under the condition that the non-target behavior bj occurs, namely, the behavior confidence of the next state; and 4, predicting air conditioner use recommendation by using a neural collaborative filtering method according to the environmental parameter set E, the user set U, the non-target environmental control behavior set B, the target environmental control behavior set A and the behavior confidence sequence. When the method is used in real time in the indoor environment, the user characteristics and the environment characteristics are continuously used as input, and intelligent regulation and control of target behaviors are carried out.

Description

Technical Field

[0001] The present invention belongs to the technical field of air-conditioning terminal control, and in particular relates to an air-conditioning terminal intelligent control method based on user behavior preference. Background Art

[0002] With the increase in the total area and the proportion of large-scale buildings, as well as the growth in energy demand, the energy consumption per unit area of ​​public buildings has increased year by year, which is the focus of building energy conservation. As the object of building operation services, the user's usage behavior is an important factor affecting the building's energy consumption level, just like the building's meteorology, enclosure structure, and equipment system. Due to different user usage behaviors, the building's energy consumption level varies by several times or even dozens of times.

[0003] With the development of artificial intelligence, smart homes are becoming more and more popular. However, the current intelligent control still needs to be realized through user instructions. Although some public buildings use IoT to monitor the energy system, they lack intelligent control based on user needs. Therefore, deepening the understanding of user behavior to achieve intelligent regulation is of great significance to improving building energy efficiency. Users' use of air conditioners has certain preferences, habits, and uncertainties, and it is necessary to achieve intelligent control through reinforcement learning of energy usage rules. Summary of the invention

[0004] The purpose of the present invention is to provide an intelligent control method for air conditioner terminals based on user behavior preferences, which continuously uses user characteristics and environmental characteristics as inputs to perform intelligent regulation of target behavior when used in real time in an indoor environment. The technical solution adopted is as follows:

[0005] An air-conditioning terminal intelligent control method based on user behavior preference comprises the following steps:

[0006] Step 1: Build an IoT monitoring platform;

[0007] Step 2: Extract user and behavior monitoring data and perform preprocessing;

[0008] Step 3: Obtain the probability of the target behavior aw occurring when the non-target behavior bj occurs, that is, the behavior confidence of the next state;

[0009] Step 4: The environmental parameter set E, the user set U, the non-target environmental control behavior set B, the target environmental control behavior set A and the behavior confidence sequence are used to predict the air conditioning usage recommendation using the neural collaborative filtering method.

[0010] Preferably, step 4 specifically includes the following steps:

[0011] Step 41, calculate the eigenvectors of the user behavior multi-feature matrix P and the environment parameter multi-feature matrix E:

[0012] Through the one-hot encoding method, the environmental parameters and user IDs are encoded respectively to generate the user feature matrix P and the environmental parameter matrix E. Then, based on the number of occurrences of different behaviors, a multi-feature matrix that integrates user behaviors and environmental parameters is constructed.

[0013] Step 42: Based on the feature vector, a neural collaborative filtering algorithm is used to establish a learning model of the behavior;

[0014] Step 43: predict the non-target behavior and the target behavior, and add the confidence of the non-target behavior to the target behavior in the prediction process of the target behavior, and generate a behavior recommendation based on the prediction result of the target behavior.

[0015] Preferably, step 3 specifically includes the following steps:

[0016] Step 31: Obtain an environmental parameter set E, a user set U, a non-target environmental control behavior set B, and a target environmental control behavior set A, and establish association rules between non-target behaviors bj and target behaviors aw;

[0017] Step 32: Obtain a confidence sequence based on the association rule;

[0018] Step 33: discretize the confidence sequence to obtain a confidence state sequence;

[0019] Step 34: Based on the confidence state sequence, calculate the transition probability between confidence states, and based on the transition probability, calculate the behavior confidence of the next state.

[0020] Compared with the prior art, the advantages of the present invention are:

[0021] The present invention uses a spade-based correlation analysis algorithm to divide human behavior data according to time periods, and predicts the confidence of the next time period by calculating the confidence of the occurrence of target environmental control behavior under the condition that non-target environmental control behavior occurred in the past time period.

[0022] In the multi-behavior neural collaborative filtering recommendation model, the environmental control behaviors of different users under certain environmental parameters are considered, a feature matrix is ​​constructed, and a composite function is used to represent the complex nonlinear relationship between user behavior and environmental preferences.

[0023] The model is solved using the behavior classification learning strategy: non-target environmental control behaviors are learned and predicted independently, and for target environmental control behaviors, the impact of non-target environmental control behaviors is considered and joint learning is performed. When used in real time in indoor environments, user characteristics and environmental characteristics are continuously used as input to perform intelligent regulation of target behaviors. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1This is a flow chart of the intelligent control method of air-conditioning terminal based on user behavior preference.

[0025] Figure 2 Recommend algorithm logic diagram for neural network. DETAILED DESCRIPTION

[0026] The following is a more detailed description of the air-conditioning terminal intelligent control method based on user behavior preference of the present invention in conjunction with a schematic diagram, which shows a preferred embodiment of the present invention. It should be understood that those skilled in the art can modify the present invention described herein and still achieve the beneficial effects of the present invention. Therefore, the following description should be understood as being widely known to those skilled in the art and not as a limitation of the present invention.

[0027] like Figure 1 As shown, this method extracts user behavior data obtained through monitoring based on the Internet of Things platform, and uses data mining methods to analyze the correlation characteristics between air-conditioning use and window opening, sun shading and other equipment under different indoor environmental parameters. Based on the correlation characteristics between indoor and outdoor environmental parameters and user behavior, the method combines neural network and matrix decomposition learning to automatically match user habits and perform intelligent regulation of air-conditioning equipment.

[0028] During use, this method integrates intelligent sensors to monitor user behavior and related equipment status to achieve multi-source information fusion; establishes a neural collaborative filtering model to predict behavior in real time, and incorporates association rules as domain information to achieve the purpose of intelligently regulating energy consumption behavior.

[0029] 1. Build an IoT monitoring platform to monitor personnel activities, the use of sunshade facilities, the operating status of the lighting system, the adjustment of air conditioning, and the opening and closing status of windows.

[0030] The sensors use a unified standard protocol to ensure compatibility and data consistency between different devices and systems, while monitoring indoor and outdoor environmental parameters such as temperature, humidity and carbon dioxide concentration.

[0031] All collected data will be centrally stored and processed for subsequent analysis.

[0032] 2. Extract user and behavior monitoring data, and pre-process various types of data by cleaning, noise reduction, compression, etc.

[0033] Data cleaning includes two parts: missing value processing and outlier processing.

[0034] The missing value processing method is: using the multiple imputation method, multiple regression predictions are performed through the multiple imputation method of chained equations (MICE), and then randomly extracting the predicted values ​​to fill the missing variables;

[0035] The outlier processing method is: use the Raida criterion to determine whether the variable is abnormal, then delete the outlier and fill it with the missing value processing method as described above;

[0036] Data compression adopts segment compression.

[0037] Environmental data compression requires smoothing the data curve through data denoising, and then performing segmented compression through a top-down (TD) global algorithm. First, the Bartlett window function is used to roll and process the numerical environmental parameters, and then segmentation is performed after denoising. The TD algorithm is used to identify key points of the denoised curve and divide it into segments. Cluster analysis and classification are then performed on the segment feature values, such as slope and average value. Different environmental parameter categories are obtained based on the segment classification results as the environmental parameter set E.

[0038] 3. Identification of user multi-behavior correlation features.

[0039] By using association sequence rule mining, the mathematical association features between the use of air-conditioning equipment (target behavior) and lighting, shading, and windows (non-target behavior) are mined.

[0040] In this embodiment, a time series correlation analysis algorithm is used to obtain the confidence of the non-target environmental control behavior to the target environmental control behavior (that is, the probability of the target behavior occurring under the condition that a non-target behavior occurs), and a confidence state sequence is constructed, which specifically includes the following steps:

[0041] 1) Each set of data in the total data set Du represents the behavior of user u when the environmental parameter state is e, which is obtained by the Internet of Things monitoring platform.

[0042] For example, the infrared rangefinder monitors the height change of the sunshade curtain and feeds it back to the platform, where it is combined with the environmental parameters detected by the temperature and humidity sensors and the time when the corresponding behavior occurs into a set of data.

[0043] Based on the total data set Du, a user set, an environmental parameter set, a non-target environmental control behavior set, and a target environmental control behavior set are obtained.

[0044] U represents the user set, E represents the environmental parameter set, B represents the non-target environmental control behavior set, A represents the target environmental control behavior set, and Du represents the total data set.

[0045] The number of people in the room and the time spent in the room are used as the basis for user classification;

[0046] The user data u mainly includes different combinations of the number of people in the room and the time periods in which the users are in the room, for example, combinations of the number of people in the room being 1 and the time periods in the room being 8-9, 9-10, 10-11, and so on.

[0047] The air conditioner switching behavior is taken as the target environmental control behavior aw, and the shading, ventilation, and lighting behaviors are taken as non-target environmental control behaviors bj.

[0048] Sunshade behavior refers to the height of the sunshade curtains, and there are four types according to the height range: no sunshade, low, medium, and full sunshade.

[0049] Ventilation behavior refers to the opening and closing of windows.

[0050] Lighting behavior refers to the turning on and off of indoor lights.

[0051] The relationship between the non-target environmental control behavior bj and the target environmental control behavior aw is represented as the association rule R: bj→aw, and its confidence is represented as m (bj→aw) , which is used to represent the probability of the target behavior aw occurring when the non-target behavior bj occurs.

[0052] Hereinafter, non-target environmental control behavior is referred to as “non-target behavior”; target environmental control behavior is referred to as “target behavior”.

[0053] 2) The confidence sequence Mjw of the association rule bj→aw is:

[0054] Mjw={m (bj→aw)1 ,m (bj→aw)2 ,…,m (bj→aw)t}

[0055] t represents the time period, m (bj→aw)t Indicates that the association rule bj→aw is in the data subset It is the confidence of the data set divided by time period in the data subset divided by time period. This patent uses the spade association rule mining algorithm to mine confidence.

[0056] 3) Map the continuous confidence values ​​in the confidence sequence Mjw into discrete state values, and the set of discrete state values ​​constitutes the confidence state sequence Sjw.

[0057] In the interval [0,1], n subintervals of equal width are divided according to the hyperparameter n, and each subinterval corresponds to a state value 1, 2, ..., n.

[0058] According to this method, the confidence state sequence Sjw can be expressed as a set containing t elements, Sjw = {s (bj→aw)1 ,s (bj→aw)2 ,…s (bj→aw)t}.

[0059]

[0060] s x (x∈1, 2, …, n) represents the confidence state corresponding to the subinterval divided by the hyperparameter.

[0061] For example, s1=1, the corresponding interval is [0,1 / n]; s2=2, the corresponding interval is [1 / n,2 / n]; ………s n =n, the corresponding interval is [n-1 / n,n].

[0062] And s (bj→aw)t For the association rule bj→aw in the data subset It is the confidence status of the data subset divided by time period in the data set).

[0063] 4) Using the confidence state sequence Sjw, according to the existing technology, calculate the transition probability Q of each confidence state of the association rule bj→aw to any other confidence state xy =Q (sx→sy)j , forming the state transfer matrix Q jw :

[0064]

[0065] According to the confidence state sequence obtained in step 3) of process 3, the final state is s (bj→aw)t , then the confidence state vector is represented by one-hot encoding The confidence state vector for the next time period is predicted to be:

[0066] Finally, let Number be The subscript corresponding to the maximum value in the time period t is s. (bj→aw)t+1 = Number, then we can get the confidence m of the next state (bj→aw)t+1 , then the confidence level is the confidence level of the next predicted time period t+1.

[0067] m (bj→aw)t+1 =s (bj→aw)t+1 The corresponding median confidence interval is as follows:

[0068]

[0069] x∈1,2,…,n.

[0070] x represents The subscript corresponding to s in ; s represents a number that is not 0.

[0071] 4. Based on the environmental parameter set E, user set U, behavior data set (set A and set B) and user behavior confidence state transition matrix, the neural collaborative filtering method is used to predict air conditioning usage recommendations, and the preferences of different users for indoor environment are obtained based on the predicted behavior probability. The following steps are included:

[0072] 1) Through the one-hot encoding method, the environment parameters and user IDs are encoded to generate the user feature matrix P and the environment parameter matrix E. and Represent the unique hot encoding vectors of user u and environment parameter e respectively, and the matrix is ​​represented as follows:

[0073] User feature matrix P

[0074]

[0075] Environmental parameter matrix E

[0076]

[0077] The user feature matrix P is generated based on the ID codes of the users in the user set U.

[0078] The environment parameter matrix E is generated based on the environment parameters in the environment parameter set E.

[0079] Among them, M is the number of user feature categories, N is the number of environmental parameter categories, and J is the feature vector dimension

[0080] The user feature categories include: different combinations of the number of people in the room and the time periods in which the users are in the room, for example: combinations of the number of people in the room being 1 and the time periods in the room being 8-9, 9-10, 10-11, and the like.

[0081] Environmental parameter categories include: temperature and humidity data categories with different data characteristics obtained by segmenting continuous temperature and humidity data, such as a temperature category with an average value of 25 degrees Celsius and a slope greater than 0.

[0082] Then, based on the number of occurrences of different behaviors, a multi-feature matrix that integrates user behavior and environmental parameter characteristics is constructed.

[0083] That is, after increasing the number of occurrences of K kinds of behaviors on the basis of matrices P and E, the matrix sizes are expanded to M×(J+K) and N×(J+K).

[0084] Among them, the “K kinds of behaviors” include 4 types: air conditioning switch behavior, ventilation, lighting and shading behavior.

[0085] The functions of the behavioral data sets (set A and set B) are as follows:

[0086] The number of behavior occurrences is used as a behavior feature and combined with user features and environmental parameter features to obtain an augmented matrix

[0087] use and Represent the unique hot encoding vectors of user u and environment parameter e respectively, and the matrix is ​​represented as follows:

[0088] User feature matrix P

[0089]

[0090] Environmental parameter matrix E

[0091]

[0092] In the Matrix represents the number of occurrences of the t-th behavior under the u-th user condition;

[0093] In the Matrix It represents the number of occurrences of the t-th behavior under the e-th type of environmental parameter conditions;

[0094] The augmented matrix is:

[0095] User feature matrix P

[0096]

[0097] Environmental parameter matrix E

[0098]

[0099] The corresponding eigenvectors, that is, the u-th row of matrix P and the e-th row of matrix E, can be obtained using the following formula:

[0100]

[0101]

[0102] 2) Use the neural collaborative filtering algorithm to build a behavioral learning model.

[0103]

[0104] in: Represents the predicted probability of the target behavior aw or the non-target behavior bj.

[0105] r represents the research object. When the research object is non-target behavior, r = j. When the research object is target behavior, r = i.

[0106] ue indicates that the predicted probability is obtained based on the user feature matrix and the environment parameter matrix;

[0107] is a function of the feature vectors of the user and the environment;

[0108] θ represents the neural network function;

[0109] σ is the sigmoid function that converts the output into probability.

[0110]

[0111] The neural matrix factorization method (NeuMF) is adopted as follows;

[0112]

[0113] in:

[0114] h T represents the weight vector;

[0115] p u ⊙e e It is obtained by adding weights to each item on the result of matrix decomposition. ⊙ is the Hadamard product, which is an element-by-element multiplication operation of a matrix.

[0116] z L The result of the last hidden layer obtained through neural network learning is as follows:

[0117]

[0118] …

[0119]

[0120] Where λ represents the number of hidden layers in the multilayer perceptron, Wx and bx represent the weight and bias term of the xth layer respectively, and the activation function uses the linear rectification function (ReLU) by default.

[0121] 3) Predict non-target behaviors and target behaviors, and add the confidence of non-target behaviors to target behaviors in the prediction process of target behaviors, and generate behavior recommendations based on the prediction results of target behaviors.

[0122] Association rules are added to the target behavior, and then the neural collaborative filtering algorithm is used for prediction.

[0123] First, the set of user-environment states in the target behavior aw, that is, the total data set Du, is represented as an interaction matrix Zw. The size of Zw is M×N, where M is the number of user categories and N is the number of environment parameter categories.

[0124] This matrix is ​​used to verify the predicted values ​​and is also the basis for optimizing the loss function.

[0125]

[0126] For the target behavior aw, the confidence level is input to consider the non-target environmental control behavior and its relevance.

[0127]

[0128] Among them, m jw is the confidence m obtained in step 4) of process 3 (bj→aw)t+1 ,

[0129] nw=jw;

[0130]

[0131] is the bias term of the environmental parameter e on the target behavior aw; it is the target optimization value of the loss function, optimizing the bias term and function To reduce the loss function value;

[0132] is the target behavior function representation;

[0133] σ is the sigmoid function that converts the output into probability.

[0134] is the predicted probability of non-target behavior bj.

[0135] The probability of non-target behavior bj The prediction steps are as follows:

[0136] a. Represent the user-environment state when non-target behavior bj occurs as a matrix Yj, the size of Yj is M×N, M is the user feature category, and N is the environment parameter category.

[0137] The method of constructing Yj is the same as that of Zw. As shown in formula (10):

[0138] b. Use the neural collaborative filtering algorithm to predict the probability of non-target behavior bj

[0139]

[0140] is the bias term of the environmental parameter e on the non-target behavior bj;

[0141] function Functions used with target behaviors The same is formula (7).

[0142] c. The same as the loss function of the target behavior below, the negative logarithmic loss function is adopted as formula (13) and the stochastic gradient descent method is used to independently optimize the loss function of each non-target environmental control behavior function.

[0143] 4) Verify the trained model.

[0144] The likelihood function for the target behavior aw is as follows:

[0145]

[0146] in represents the entries that are 1 in the behavior matrix Zw, represents the entries that are 0 in the behavior matrix Zw;

[0147] By applying a negative logarithmic transformation to the above probability function, the corresponding loss function L is derived.

[0148] The loss function L is exactly the objective function in the optimization process.

[0149]

[0150] in represents the entries that are 1 in the behavior matrix Zw, represents the entries that are 0 in the behavior matrix Zw;

[0151] If L < the set value, the model established according to formula (7) is applicable, otherwise the model is adjusted.

[0152] Build a recommendation list based on the prediction results;

[0153] Enter the environmental parameter e and user classification u in formula 11 and judge If the recommended condition is met, the target behavior is adopted

[0154]

[0155] Predict the likelihood of the target behavior based on the previous step It can reflect the preference of user u for environmental parameter e:

[0156]

[0157] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any technician in the relevant technical field, without departing from the scope of the technical solution of the present invention, makes any form of equivalent replacement or modification to the technical solution and technical content disclosed in the present invention, which does not depart from the content of the technical solution of the present invention and still falls within the protection scope of the present invention.

Claims

1. An intelligent control method for air-conditioning terminals based on user behavior preferences, characterized in that: The following steps are involved: Step 1: Build an IoT monitoring platform; Step 2: Extract user and behavior monitoring data and perform preprocessing; Step 3: Obtain the probability of the target behavior aw occurring when the non-target behavior bj occurs, that is, the behavior confidence of the next state; Step 4: The environmental parameter set E, the user set U, the non-target environmental control behavior set B, the target environmental control behavior set A and the behavior confidence sequence are used to predict the air conditioning usage recommendation using the neural collaborative filtering method.

2. The air-conditioning terminal intelligent control method based on user behavior preference according to claim 1 is characterized in that: Step 4 specifically includes the following steps: Step 41, calculate the eigenvectors of the user behavior multi-feature matrix P and the environment parameter multi-feature matrix E: Through the one-hot encoding method, the environmental parameters and user IDs are encoded respectively to generate the user feature matrix P and the environmental parameter matrix E. Then, based on the number of occurrences of different behaviors, a multi-feature matrix that integrates user behaviors and environmental parameters is constructed. Step 42: Based on the feature vector, a neural collaborative filtering algorithm is used to establish a learning model of the behavior; Step 43: predict the non-target behavior and the target behavior, and add the confidence of the non-target behavior to the target behavior in the prediction process of the target behavior, and generate a behavior recommendation based on the prediction result of the target behavior.

3. The air-conditioning terminal intelligent control method based on user behavior preference according to claim 1 is characterized in that: Step 3 specifically includes the following steps: Step 31: Obtain an environmental parameter set E, a user set U, a non-target environmental control behavior set B, and a target environmental control behavior set A, and establish association rules between non-target behaviors bj and target behaviors aw; Step 32: Obtain a confidence sequence based on the association rule; Step 33: discretize the confidence sequence to obtain a confidence state sequence; Step 34: Based on the confidence state sequence, calculate the transition probability between confidence states, and based on the transition probability, calculate the behavior confidence of the next state.

Images