Air conditioning device

Abstract

The invention provides an air conditioning device. The air conditioning device comprises a data acquisition module and a control module, the data acquisition module is configured to collect environment parameters, operation parameters and / or user behavior parameters of the air conditioning device; the control module is configured to establish a prediction model based on a Transform architecture according to a historical sample; predicting predicted set temperatures of a plurality of future target moments based on the established prediction model; moreover, when the user set temperature at the future target moment deviates from the predicted set temperature, the user set temperature is corrected to be the predicted set temperature; the historical samples are generated and stored in a preset database based on the environment parameters, the operation parameters and / or the user behavior parameters associated with the corresponding timestamps. According to the invention, high-precision, stable and self-adaptive future set temperature prediction can be realized, and the intelligent degree of the air conditioning device is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of air conditioning, and in particular to an air conditioning device. Background Art

[0002] Intelligent control of residential air conditioners has always been one of the main research topics in the relevant field. Manufacturers hope that their products can achieve automatic control based on user habits, such as automatically turning on the air conditioner before the user returns home to make the air-conditioned environment more suitable, or automatically turning off the air conditioner after the user leaves home to avoid wasting energy.

[0003] There are related technical solutions disclosed in the prior art to solve this problem, such as the technical solution disclosed in the Chinese patent application (CN107504656A): "A self-learning control method for an air-conditioning system, comprising an information collection module and an automatic module, the self-learning control method comprising: step S10: the information collection module collects data of a user using a remote control or an APP to turn on or off the air-conditioning system, and sends the collected data to the automatic control module; step S11: the automatic control module analyzes and processes the collected data by self-learning the user's on-off habits, and identifies the on-off time period of the air-conditioning system; step S12: the automatic control module processes the on-off time period of the air-conditioning system respectively, and determines the on-off time of the air-conditioning system, wherein determining the on-time of the air-conditioning system comprises: recording the on-time in all the on-time periods respectively, and calculating the self-learning on-time of the nth hour time period; determining the off-time of the air-conditioning system comprises: recording the off-time in all the off-time periods respectively, and calculating the self-learning off-time of the mth hour time period; step S13: the automatic control module automatically controls the on-off time of the air-conditioning system according to the determined on-off time of the air-conditioning system."

[0004] The self-learning control method provided in the comparative document generates self-learning on / off time based on the average number of on / off times within 10 days. But in fact, the user's operation control of the air-conditioning system, especially the control in a certain scene, is a discrete time series with an indefinite time interval. In addition to on / off control, it also includes mode control (cooling / heating), function control (sleep / voice / human sensing / auxiliary heating), air supply control (wind direction / wind speed), etc. The prediction based on traditional data statistics has a large deviation and poor intelligence.

[0005] The above information disclosed in the background technology is only used to increase the understanding of the background technology of the present application, and therefore, it may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0006] In view of the problems pointed out in the background technology, a first aspect of the present application provides an air conditioning device.

[0007] Specifically, the air conditioning device includes a data acquisition module and a control module; the data acquisition module is configured to collect environmental parameters, operating parameters, and / or user behavior parameters of the air conditioning device; the control module is configured to establish a prediction model based on the Transformer architecture according to historical samples; and, based on the established prediction model, predict the predicted set temperatures at several future target moments; and, when the user set temperature at the future target moment deviates from the predicted set temperature, the user set temperature is corrected to the predicted set temperature.

[0008] In one or more embodiments of the present application, the historical samples are generated based on the environmental parameters, operating parameters and / or user behavior parameters associated with corresponding timestamps and stored in a predetermined database.

[0009] For air conditioning devices, due to various events, such as extreme weather (e.g., rapid temperature changes) or force majeure factors, the patterns in the observation sequence may change significantly over time, that is, a number of abnormal points that cannot be eliminated and observation sequences corresponding to the change points will be generated and input into the prediction model, resulting in each vector being referred to in the dot product processing without making full use of similar context information, resulting in local context confusion of the attention sublayer, bringing potential optimization problems. To solve this problem, a causal convolution layer is configured in the prediction model; the prediction model uses the output sequence of the causal convolution layer to calculate the query matrix and the key matrix, and uses the query matrix and the key matrix to calculate the attention weights of the prediction model to process the dependencies between features at different positions in the observation sequence; the observation sequence is generated based on one or more parameters in the historical samples and serves as the input sequence of the prediction model.

[0010] Since the environmental parameters, operating parameters and user behavior parameters collected by the data acquisition module are fine-grained, long-term series with strong long-term dependencies, in order to reduce the difficulty of modeling from a hardware perspective, in one or more embodiments of the present application, a sparse mode is configured in the prediction model; the prediction model can be configured to selectively calculate the attention weights in the sparse mode to enhance the attention weights of the features at some positions of the observation sequence and reduce the attention weights of the features at another part of the observation sequence; the observation sequence is generated based on one or more parameters in the historical samples and serves as the input sequence of the prediction model.

[0011] In one or more embodiments of the present application, the sparse pattern is formed based on the LogSparse SelfAttention strategy; the prediction model includes multiple sub-attention layers, each of which is configured with the sparse pattern; the prediction model stacks multiple sub-attention layers configured with the sparse pattern.

[0012] In one or more embodiments of the present application, the operating parameters include the actual operating state of the air conditioning device; the actual operating state includes user-set parameters.

[0013] In one or more embodiments of the present application, in order to eliminate invalid data, the air conditioning device also includes: a preprocessing module, which is configured to enter a waiting state when the data acquisition module collects any one or more user-set parameters, and determine whether the set parameters collected within a set time period remain unchanged; when the set parameters collected within the set time period remain unchanged, it is inferred that the collected set parameters are valid user-set parameters, and the valid user-set parameters are associated with the corresponding timestamps and stored in a predetermined database as part of the historical samples.

[0014] To help the prediction model obtain more stable and accurate predictions, the control module is also configured with a correction model independent of the prediction model, and the correction model is configured to generate a set temperature correction value; the control module is configured to associate the set temperature correction value with a corresponding timestamp and store it in a predetermined database as part of the historical sample.

[0015] In one or more embodiments of the present application, the air conditioning device further includes: a training module, wherein the training module is configured to input a plurality of the observation sequences to train the prediction model to adjust the hyperparameters of the prediction model until the training model converges.

[0016] In one or more embodiments of the present application, the air conditioning device further includes: a testing module, which tests the prediction model trained by the training module to evaluate whether the prediction result of the prediction model meets the set accuracy condition.

[0017] In one or more embodiments of the present application, the air conditioning device also includes: an evaluation module, which is configured to correct the user set temperature to the predicted set temperature when the user set temperature at the future target moment deviates from the predicted set temperature; and collect whether the intervention set temperature output by the user is received within the set evaluation period; and when the intervention set temperature is received, update the observation sequence and configure the training module to train the prediction model again.

[0018] The second aspect of the present application provides an air conditioning device, including a data acquisition module and a control module; the data acquisition module is configured to collect environmental parameters, operating parameters, and / or user behavior parameters of the air conditioning device; the control module is configured to establish a prediction model based on the Transformer architecture according to historical samples; and, based on the established prediction model, predict the predicted set temperature of the target future moment corresponding to the restart of the air conditioning device; and, when the air conditioning device is restarted, the set temperature stored in standby mode is corrected to the predicted set temperature; the historical samples are generated based on the environmental parameters, operating parameters and / or user behavior parameters associated with the corresponding timestamps and stored in a predetermined database.

[0019] This application can realize high-precision adaptive and intelligent generation of predicted set temperatures.

[0020] After reading the specific embodiments of the present invention in conjunction with the accompanying drawings, other features and advantages of the present invention will become more clear. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0022] Figure 1 A schematic block diagram of the structure of an air conditioning device provided in one or more embodiments of the present application;

[0023] Figure 2 A schematic block diagram of the structure of a data acquisition module in an air conditioning device provided in one or more embodiments of the present application;

[0024] Figure 3 A schematic block diagram of the structure of a data acquisition module in an air conditioning device provided in one or more embodiments of the present application;

[0025] Figure 4 A schematic block diagram of the structure of a data acquisition module in an air conditioning device provided in one or more embodiments of the present application;

[0026] Figure 5 A schematic block diagram of the structure of a data acquisition module in an air conditioning device provided in one or more embodiments of the present application;

[0027] Figure 6 A flow chart of a control module in an air conditioning device provided in one or more embodiments of the present application;

[0028] Figure 7 A schematic diagram of the structure of a causal convolutional layer in an air conditioning device provided in one or more embodiments of the present application;

[0029] Figure 8 A schematic diagram of the structure of a causal convolutional layer in an air conditioning device provided in one or more embodiments of the present application;

[0030] Fig. 9 A point product curve corresponding to a causal convolutional layer in an air conditioning device provided in one or more embodiments of the present application;

[0031] Fig.10 A schematic diagram of the principle of a sparse mode in an air conditioning device provided in one or more embodiments of the present application;

[0032] Fig.11 A schematic block diagram of the structure of an air conditioning device provided in one or more embodiments of the present application;

[0033] Fig.12 A flow chart of a pre-processing module in an air conditioning device provided in one or more embodiments of the present application;

[0034] Fig.13 A schematic block diagram of the structure of a control module of an air conditioning device provided in one or more embodiments of the present application;

[0035] Fig.14 A schematic block diagram of the structure of an air conditioning device provided in one or more embodiments of the present application;

[0036] Fig.15 A schematic block diagram of the structure of an air conditioning device provided in one or more embodiments of the present application;

[0037] Fig.16 A schematic block diagram of the structure of an air conditioning device provided in one or more embodiments of the present application;

[0038] Fig.17 A flow chart of a control module in an air conditioning device provided in one or more embodiments of the present application. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0040] In the description of the present application, it should be understood that the terms "center", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0041] The terms "first", "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, unless otherwise specified, "plurality" means two or more.

[0042] In the description of this application, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0043] In the present invention, unless otherwise clearly stipulated and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly below and obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0044] The disclosure below provides many different embodiments or examples to realize different structures of the present invention. In order to simplify the disclosure of the present invention, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the present invention. In addition, the present invention can repeat reference numbers and / or reference letters in different examples, and this repetition is for the purpose of simplicity and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed. In addition, the present invention provides various specific examples of processes and materials, but those of ordinary skill in the art can be aware of the application of other processes and / or the use of other materials.

[0045] One or more embodiments of the present invention provide an air conditioning device. In the present application, the air conditioning device performs a refrigeration cycle of the air conditioning device by using a compressor, a condenser, an expansion valve and an evaporator. The refrigeration cycle includes a series of processes involving compression, condensation, expansion and evaporation to cool or heat an indoor space.

[0046] The low-temperature and low-pressure refrigerant enters the compressor, which compresses it into high-temperature and high-pressure refrigerant gas and discharges the compressed refrigerant gas. The discharged refrigerant gas flows into the condenser. The condenser condenses the compressed refrigerant into a liquid phase, and the heat is released to the surrounding environment through the condensation process.

[0047] The expansion valve expands the high-temperature and high-pressure liquid refrigerant condensed in the condenser into a low-pressure liquid refrigerant. The evaporator evaporates the refrigerant expanded in the expansion valve and returns the low-temperature and low-pressure refrigerant gas to the compressor. The evaporator can achieve a cooling effect by utilizing the latent heat of evaporation of the refrigerant to exchange heat with the material to be cooled. In the entire cycle, the air conditioning device can adjust the temperature of the indoor space.

[0048] The outdoor unit of the air conditioning device refers to a portion of a refrigeration cycle including a compressor and an outdoor heat exchanger, the indoor unit of the air conditioning device includes an indoor heat exchanger, and the expansion valve may be provided in the indoor unit or the outdoor unit.

[0049] The indoor heat exchanger and the outdoor heat exchanger function as a condenser or an evaporator. When the indoor heat exchanger functions as a condenser, the air conditioning device functions as a heater in a heating mode, and when the indoor heat exchanger functions as an evaporator, the air conditioning device functions as a cooler in a cooling mode.

[0050] Figure 1 The structure schematic diagram of the air conditioning device provided in this embodiment is as follows; Figure 1 The air conditioning device shown can automatically predict and configure future set temperatures to create a smarter, more efficient and personalized air conditioning system for better user experience and energy efficiency.

[0051] like Figure 1As shown, the air conditioning device includes a data acquisition module 100 and a control module 200; wherein the data acquisition module 100 is configured to collect environmental parameters, operating parameters, and / or user behavior parameters of the air conditioning device. The control module 200 is configured to establish a prediction model based on the Transformer architecture according to historical samples; and, based on the established prediction model, predict the predicted set temperature at several future target moments; and, when the user set temperature at the future target moment deviates from the predicted set temperature, correct the user set temperature to the predicted set temperature. Wherein, the historical samples are generated based on the environmental parameters, operating parameters and / or user behavior parameters associated with the corresponding timestamps and stored in a predetermined database.

[0052] Figure 2 This is a schematic block diagram of the structure of a data acquisition module 100 provided in one or more embodiments of the present invention. In this embodiment, the data acquisition module 100 is configured to collect environmental parameters of an air conditioning device.

[0053] In one or more embodiments of the present application, the environmental parameters include indoor environmental parameters and outdoor environmental parameters.

[0054] In one or more embodiments of the present application, the indoor environmental parameters include: indoor environmental temperature, indoor environmental humidity and indoor light intensity.

[0055] In one or more embodiments of the present application, the indoor environmental parameters also include: air quality parameters, such as the concentrations of carbon dioxide, formaldehyde, and volatile organic compounds (VOCs).

[0056] In one or more embodiments of the present application, indoor environmental parameters may also include the number of people indoors and odor levels, etc.

[0057] In one or more embodiments of the present application, outdoor environmental parameters include: outdoor environmental temperature, outdoor environmental humidity and weather parameters; wherein weather parameters include: precipitation level, wind speed, wind direction and solar radiation intensity, etc.

[0058] The indoor ambient temperature and the outdoor ambient temperature can be collected by a temperature sensor, and the indoor ambient humidity and the outdoor ambient humidity can be collected by a humidity sensor.

[0059] Indoor light intensity can be collected by light meters, spectroradiometers, light sensors, or cameras and image processing. Indoor light intensity is used to measure and evaluate the intensity and distribution of light in indoor environments. For example, light intensity can be measured by photosensors, which work based on the principle of photoelectric effect. For example, the resistance value of a photoresistor changes with the change of light intensity, that is, when light shines on a photosensitive material, it can excite the electrons in the photosensitive material, thereby changing the resistance value. In addition to photoresistors, photosensitive sensors can also be implemented based on photodiodes, phototransistors, photosensitive capacitors, etc.

[0060] The concentrations of carbon dioxide, formaldehyde, volatile organic compounds (VOCs), odor levels, etc. can all be collected by corresponding sensors, and the noise level can be collected by a decibel meter.

[0061] Weather parameters can be obtained from the server.

[0062] FIG3 is a schematic block diagram of the structure of a data acquisition module 100 provided in one or more embodiments of the present invention. In this embodiment, the data acquisition module 100 is configured to collect operating parameters of an air conditioning device.

[0063] In one or more embodiments of the present application, the operating parameters include the actual operating state of the air conditioning device; the actual operating state includes but is not limited to: one or more of set temperature, set wind speed, set working mode, timing setting, set humidity, set energy efficiency mode, fault state, filter state, and energy consumption state.

[0064] Figure 4 This is a schematic block diagram of the structure of the data acquisition module 100 provided in one or more embodiments of the present invention. In this real-time mode, the data acquisition module 100 is configured to collect user behavior parameters of the air conditioning device.

[0065] The user behavior parameters of the air conditioning device refer to information such as the user's related behaviors, habits, and preferences when using the air conditioning device.

[0066] In one or more embodiments of the present invention, user behavior parameters include: setting temperature preferences, setting wind direction preferences, setting wind speed preferences, adjusting frequency preferences, setting mode preferences, setting timing preferences, indoor and outdoor temperature difference preferences, time period preferences, preferences for collaboration with other devices, calibration preferences, and one or more of feedback preferences.

[0067] Among them, the setting temperature preference represents the temperature preference set by the user in different time periods, such as daytime, nighttime, weekdays and weekends; the setting wind direction preference represents the wind direction preference set in different time periods; the setting wind speed preference represents the wind speed preference set in different time periods; the adjustment frequency preference represents the frequency of adjustment of the set temperature by the user; the setting mode preference represents the user's preference for different working modes (cooling, heating, fresh air, etc.), and whether to use the energy-saving mode; the setting timing preference represents whether the user uses the timing function to set the air conditioner to automatically turn on or off at a specific time; the indoor and outdoor temperature difference preference represents the user's preference for the difference between indoor and outdoor temperatures, such as the power on and off action or temperature adjustment action corresponding to the indoor and outdoor temperature difference; the time period preference represents the user's preference for using the air conditioner in different time periods, such as daytime, nighttime and rest time; the preference for coordination with other devices represents whether the user uses the air conditioner when using other smart devices, such as using multiple devices in the smart home system at the same time; the calibration preference represents whether the user has a preference to adjust the set temperature according to the actual comfort level, and the degree of calibration; the feedback preference represents whether the user has provided feedback on the experience of using the air conditioning device, such as the control effect is not ideal in some scenarios.

[0068] Figure 5 This is a schematic diagram of the structure of the data acquisition module 100 provided in one or more embodiments of the present invention. In this embodiment, the data acquisition module 100 is configured to simultaneously collect environmental parameters, operating parameters and user behavior parameters of the air conditioning device.

[0069] In one or more embodiments of the present application, each environmental parameter, operating parameter, and user behavior parameter collected by the data collection module 100 includes a corresponding timestamp. For user behavior parameters, the timestamp is the time when the user behavior parameters are generated, for example, the corresponding preference calculation method is configured on the server, and the time when the corresponding preference is received from the server is used as the timestamp of the user behavior parameter.

[0070] The environmental parameters, operating parameters and / or user behavior parameters associated with the corresponding timestamps are generated into historical samples and stored in a predetermined database.

[0071] The control module 200 is configured to establish a prediction model based on the Transformer architecture according to historical samples; and, based on the established prediction model, predict the predicted set temperatures at several future target moments; and, when the user set temperature at the future target moment deviates from the predetermined set temperature, correct the user set temperature to the predicted set temperature; Figure 6 shown.

[0072] The Transformer architecture is a deep learning architecture for processing sequence data. Its core idea is to map the input sequence into a set of key-value pairs and perform weighted aggregation of these key-value pairs through the attention mechanism to generate an output sequence. The Transformer consists of two main components: an encoder and a decoder.

[0073] The encoder maps the input sequence into a set of key-value pairs. Each input is embedded as a vector and then processed by multiple layers of sub-attention layers and feedforward neural networks. The sub-attention mechanism allows the encoder to automatically adjust the weights based on other positions in the input sequence when processing each input, thereby capturing global context information.

[0074] The decoder also contains sub-attention layers and feed-forward neural networks. When generating sequence data, the decoder can not only refer to the encoded representation of the input sequence, but also perform attention calculations on the generated partial sequence, thereby capturing the information of the previously generated part when generating the next element.

[0075] The following is an introduction to the prediction model based on the Transformer architecture established according to historical samples in one or more embodiments of the present application:

[0076] Mathematically, the target variable, or observation (such as set temperature), can be represented by a set:

[0077]

[0078] The above formula includes a time series consisting of observations from time step 1 to time step t0, N represents the number of observations in the set, i represents the index of the observation in the time series, which can be an integer from 1 to N, t0 represents the maximum value of the time step, and the time step in the observation sequence starts from 1 and gradually increases until t0; is defined as:

[0079]

[0080] The time step is used to represent the position of each observation sequence in the time dimension. The time step can correspond to different time intervals, such as one second, one hour, etc.

[0081] Assuming that the time step of the future target moment is τ, the external sequence can also be represented by a set:

[0082]

[0083] That is, it includes external sequences from time step 1 to time step t0+τ, N represents the number of external sequences in the set, i represents the index of the external sequence, which can be an integer from 1 to N, t0+τ represents the maximum value of the time step, and the time step in the external sequence starts from 1 and gradually increases until t0+τ.

[0084] The prediction goal is transformed into modeling and prediction of the following conditional distribution:

[0085]

[0086] represents the future observation sequence, from time step t0+1 to t0+τ;

[0087] represents the past observation sequence, from time step 1 to t0;

[0088] represents the external sequence, from time step 1 to t0+τ;

[0089] Φ represents the parameter set of the model, i.e. the learnable parameters

[0090] Represents a known past observation sequence external sequence and parameter set Φ, the future observation sequence The conditional probability density function of ;

[0091] Represents the continuous multiplication of the probability density function in the time steps from t0+1 to t0+τ.

[0092] Based on the above model, we can further simplify it into a learning prediction model

[0093] p(z t |z i，1：t-1 , x 1：t ; Φ)

[0094] Among them, z t represents the observed value at time step t; z i，1：t-1 represents the past observation sequence from time step 1 to time step t-1, x 1：t represents the external sequence of the past and present, from time step 1 to t, Φ represents the parameter set of the model, that is, the learnable parameters; p(z t |z i，1：t-1 , x 1：t ; Φ); represents the known past observation sequence z i，1：t-1 and the external sequence x 1：t , and the observation value z at time step t for the parameter set Φ tThe conditional probability density function of .

[0095] To make full use of the observations and covariates, we concatenate them to obtain the following augmented matrix

[0096]

[0097]

[0098] y t represents a vector, z t-1 represents the observed value of vector z at time step t-1, x t represents the observed value of vector x at time step t, ο represents the element-level operation, represents the d+1-dimensional real number space; Y t Represented by vector y1,…,y t The matrix is ​​composed of t×(d+1), y1,…,y t represents the observed value of vector y at different time steps, T represents transposition, represents the real number space with t rows and d+1 columns.

[0099] In the above formula, y t The definition includes z t-1 (observation) and x t The connection or combination of (covariate) element levels further produces augmented vectors containing observations and covariate information. In the observation sequence analysis, these augmented vectors are stacked together by time step to obtain the augmented matrix Y t . Augmented matrix Y t Each row in corresponds to a time step, which contains the observations and covariate information of that time step, thereby providing a more comprehensive data representation. The augmented matrix can help capture the relationship between the observations and other related information, as well as their dynamic changes between different time steps.

[0100] Further, we need to generate a suitable model z t ~f(Y t ) to predict a given distribution. In one or more embodiments of the present application, a prediction model is established based on the Transformer architecture. The Transformer architecture uses a multi-head attention mechanism to instantiate f(Y t ), because sub-attention enables the prediction model under the Transformer architecture to capture long-term or short-term dependencies, and different attention heads learn to focus on different aspects of the observation sequence.

[0101] In the sub-attention layer, a multi-head attention sub-layer simultaneously t Transformed into H different query matrices key matrix and value matrix where h = 1, ..., H. After these linear projections, the attention calculation vector output sequence is:

[0102]

[0103] The attention calculation vector output sequence is calculated by applying the attention mechanism to obtain a series of context vectors, which reflect the relationship between different positions in the input sequence and different positions in the generated target sequence. These context vectors are used in the decoder to generate each element of the target sequence and finally generate the prediction result.

[0104] In one or more embodiments of the present application, 15 features are selected from environmental parameters, operating parameters, and user behavior parameters, and multiple observation sequences arranged in time series are generated based on timestamps. That is, the observation sequence is generated based on one or more parameters in the historical samples and used as the input sequence of the prediction model. An example of the observation sequence can be:

[0105] [′iu34Status01′, ′iu46TempCv′, ′iu55Wind′, ′I1′, ′12′, ′I3′, ′14′, ′I5′, ′season′, ′inDl′, ′outDl′, ′temp_range′, ′temp_rec′, wind_rec′, mode_rec′]

[0106] Among them, iu34Status01 represents the actual operating status, iu46TempCv represents the temperature setting correction value; for example, it can be 0-4°C, iu55Wind represents the indoor fan speed, I1 represents the outdoor ambient temperature, I2 represents the outdoor ambient humidity, I3 represents the outdoor wind speed, I4 represents the outdoor wind direction, and I5 represents the weather type; inDl represents the indoor comfort level; outDl represents the outdoor comfort level; temp_range represents the indoor and outdoor temperature difference; temp_rec represents the user temperature preference; wind_rec represents the user wind speed preference; mode_rec represents the user mode preference.

[0107] Among them, Fig.13As shown, the set temperature correction value is generated by a correction model independent of the prediction model, and the correction model is configured in the control module; for example, the correction model can fine-tune or correct the set temperature according to factors that may affect the actual temperature of the room, such as weather changes and direct sunlight; or fine-tune or correct the set temperature according to the heat generated by the activities of people in the room; or fine-tune or correct the set temperature according to the difference between the temperatures at different locations in the room; or fine-tune or correct the set temperature according to the different perceptions of the user to the comfortable temperature. The correction model can adopt a variety of models disclosed in the prior art (such as the algorithm disclosed in the background technology), which is not the focus of protection of this application and will not be repeated here. The control module 200 is configured to associate the set temperature correction value with the corresponding timestamp and store it in a predetermined database as a component of the historical sample.

[0108] Since the set temperature correction value itself comes from the correction model, it is input into the neural network as a feature of the sample parameter. The prediction result of the correction model can be combined with the actual correction value to provide more comprehensive information for the neural network, helping the neural network to better understand the performance and deviation of the correction model. The set temperature correction value and weather changes, direct sunlight, heat generated by people's activities in the room, differences in temperature at different locations in the room, and users' different perceptions of comfortable temperature can enrich the input features of the neural network, improve the model's prediction ability, and calibrate the deviation of the correction model. The advantages of the correction model and the neural network are combined to obtain more stable and accurate predictions. The neural network can make real-time adjustments based on the patterns and trends of historical correction values ​​to show periodic changes.

[0109] In one or more embodiments of the present application, the user temperature preference, the user wind speed preference and the user mode preference can be obtained by analyzing the user's behavior pattern based on historical data. For example, the user's set temperature data, set wind speed data and set mode data in a set time window in the past are collected, and the trend of the set temperature is analyzed using a mathematical model to obtain the user temperature preference, the user wind speed preference and / or the user mode preference. For example, a user may prefer a lower temperature in a specific time period, a lower wind speed in a specific time period, or a certain mode (such as silent mode or sleep mode, etc.) in a specific time period.

[0110] In one or more embodiments of the present application, indoor comfort and outdoor comfort can be represented by a temperature humidity index (also called a discomfort index) or a temperature-humidity index, which is based on temperature and relative humidity and is used to describe the comfort of the human body in a thermal environment.

[0111] In one or more embodiments of the present application, the indoor comfort and outdoor comfort can be obtained by the following formula:

[0112] DI=T-(0.55-0.55f)(T-58)

[0113] Where T represents the real-time temperature (in degrees Fahrenheit) and f represents the real-time humidity.

[0114] In one or more embodiments of the present application, indoor comfort and outdoor comfort can also be represented by a perceived temperature index (also called a heat index), which is also based on temperature and humidity and is used to measure the actual temperature felt by the human body. The perceived temperature index further describes the index of the actual temperature felt by the human body in a cold wind and low temperature environment, and is based on a combination of air temperature and wind speed, further reflecting the effect of wind speed on the heat dissipation of the human body.

[0115] In one or more embodiments of the present application, indoor comfort and outdoor comfort can also be represented by a PMV index or a PPD index, which comprehensively considers temperature, humidity, air flow speed, clothing, metabolic rate, etc.

[0116] The above 15 features are encoded into a feature dimension of (15, 512) as the input sequence of the prediction model. This can be achieved by using methods such as embedding layers and fully connected layers.

[0117] The prediction is further implemented by the encoder and decoder in the Transformer architecture. The prediction process is described in the above introduction to the Transformer model.

[0118] like Fig.11 As shown, in one or more embodiments of the present application, the air conditioning device further includes a pre-processing module 300, and the pre-processing module 300 is used to screen and process the historical samples.

[0119] In one or more embodiments of the present application, the preprocessing module 300 may also be configured with a preprocessing algorithm, which is used to remove the user's invalid operation data interference, remove the user's unstable data, ensure the validity of the data, and thus implement the screening of the collected indoor environment parameters.

[0120] In one or more embodiments of the present application, the preprocessing algorithm includes: screening effective indoor unit setting parameters; and merging outdoor environment data.

[0121] In one or more embodiments of the present application, screening the valid indoor unit setting parameters includes: acquiring the indoor unit setting parameters of the same indoor unit based on the same device code and machine number.

[0122] A device code is a unique identifier used to distinguish different devices. It can be a string of numbers, letters, or a combination of numbers and letters. The device code can be used to manage licenses, authorize devices to access specific services or features, and track and manage devices.

[0123] The machine number is a unique number that identifies the machine. It is a number assigned by the manufacturer to track, identify and manage different machines in the production process. The machine number can be used for maintenance records, warranty tracking, identifying problems in production, etc.

[0124] like Fig.12 As shown, in one or more embodiments of the present application, the preprocessing module is configured to enter a waiting state when the data acquisition module 100 acquires any one or more user-set parameters, and determines whether the collected setting parameters remain unchanged within a set time period; when the collected setting parameters remain unchanged within the set time period, it is inferred that the collected setting parameters are valid user setting parameters, and the valid user setting parameters are configured to be associated with the corresponding timestamps and stored in a predetermined database as historical samples.

[0125] In one or more embodiments of the present application, the set duration may be 10 minutes.

[0126] In one or more embodiments of the present application, merging outdoor environment data includes: acquiring outdoor environment parameters; merging the outdoor environment data when the operation time and the collection time of the outdoor environment parameters do not exceed the set time.

[0127] In one or more embodiments of the present application, the set duration may be 30 minutes.

[0128] In one or more embodiments of the present application, the preprocessing algorithm processes the screened indoor environmental parameters including: missing value processing, outlier processing, and / or normalization processing, etc.

[0129] In one or more embodiments of the present application, missing value processing adopts interpolation. Interpolation is to estimate missing values ​​through known data points, such as linear interpolation, polynomial interpolation, spline interpolation, etc. Exemplarily, spline interpolation can fill in missing values ​​between data points by fitting a smooth curve or surface, which divides the entire data range into multiple small intervals, and then fits a polynomial or other smooth function in each interval. The connection points between the small intervals are nodes, and the nodes can be selected by rules such as uniform distribution and quantiles. Spline interpolation uses piecewise polynomials, and a low-degree polynomial is fitted in each interval, optionally, for example, a quadratic or cubic polynomial to ensure smoothness in the local range. These polynomials are continuous at the nodes, and the derivatives at the nodes are also continuous to maintain overall smoothness.

[0130] In one or more embodiments of the present application, missing value processing can also be filled using mean, median, mode, fixed value, or value of similar samples.

[0131] In one or more embodiments of the present application, missing value processing can also be implemented by observation sequence interpolation, regression methods, machine learning models or multiple interpolation methods.

[0132] In one or more embodiments of the present application, outlier processing includes two steps: identifying outliers and deleting outliers; or includes two steps: identifying outliers and correcting outliers; or includes both of the above. Among them, outliers can be identified by statistical methods, such as using visualization tools (such as box plots, scatter plots, etc.) to identify outliers; or outlier detection algorithms, such as clustering-based methods, can be used to automatically identify and process outliers. Correcting outliers can replace them with other suitable values, such as using statistical indicators such as mean values ​​or medians to fill them.

[0133] In one or more embodiments of the present application, the normalization process may adopt minimum-maximum normalization, standardization, decimal scaling standardization or norm normalization; wherein minimum-maximum normalization refers to linearly mapping the eigenvalues ​​between a specified minimum and maximum value; standardization refers to making the eigenvalues ​​have a distribution with a mean of 0 and a standard deviation of 1 by subtracting the mean and dividing by the standard deviation; decimal scaling standardization refers to dividing the eigenvalues ​​by the cardinality so that the absolute value of the eigenvalues ​​is in the range of [0.1, 1); norm normalization is to scale the eigenvector of each sample so that the norm of the vector is equal to 1. Norm normalization can ensure that the position of each sample in the multidimensional space is the same distance from the origin. The normalized sample data can eliminate the dimensional differences between the features, so that the influence of each feature on the model is more balanced.

[0134] In one or some embodiments of the present application, for air conditioning devices, due to various events, such as extreme weather (e.g., rapid temperature changes) or force majeure, the pattern in the observation sequence may change significantly over time, that is, a number of abnormal points that cannot be eliminated by the preprocessing module 300 and the observation sequences corresponding to the change points will be generated and input into the Transformer model, but in the sub-attention layer of the Transformer, the similarity between the query matrix and the key matrix is ​​calculated based on their point-wise. When the observation sequence corresponding to the abnormal points and change points is generated, the point refers to each vector, and similar contextual information is not fully utilized. At this time, the local context will confuse the attention sub-layer, that is, the observed value is an abnormal value, and the sub-attention score is only a reflection of the correlation between single time points, which is different from the original intention of the observation sequence prediction and brings potential optimization problems.

[0135] In particular, if a convolution operation with a kernel of 1 and a stride of 1 is used when calculating Query, Key, and Value, since the calculation of the attention weight is completed by the dot product operation between the query matrix (Query) and the key matrix (Key), the dot product operation after the convolution operation may cause the relationship between features at different positions to become unclear.

[0136] In one or more embodiments of the present application, a causal convolution layer is configured in the prediction model; the prediction model uses the output sequence of the causal convolution layer to calculate the query matrix and the key matrix, and uses the query matrix and the key matrix to calculate the attention weight of the prediction model to process the dependency between features at different positions in the observation sequence.

[0137] Specifically, in the causal convolutional layer, such as Figure 8 and Fig. 9 As shown in the figure, the query matrix, key matrix and value are calculated through the convolution operation with a convolution kernel of k and a step size of 1, so that the query matrix and key matrix at different positions can consider the relationship between each other and their differences in position. This position awareness can help the model better understand the correlation between different positions, thereby improving the performance of the model, especially when processing observation sequences. The causal convolution layer can ensure that the current position does not access future information. By using the causal convolution layer, the generated query matrix and key matrix can more clearly understand the remaining time points relative to the current position or the so-called context. Therefore, their similarity is calculated through their local context information, which improves the prediction accuracy and overcomes the interference of outliers. For the trend of frequent extreme weather, the model can be more flexible to adapt to extreme situations and further improve the intelligent performance of air conditioning devices.

[0138] For air conditioning devices, the environmental parameters, operating parameters and user behavior parameters collected by the data acquisition module 100 are fine-grained, long-term series with strong long-term dependencies. In order to meet higher accuracy requirements, the collection time intervals of temperature, humidity parameters, etc. are small (i.e., fine-grained). At the same time, there are long-term dependencies between data points (such as preference), which also means that the observation value of the current time step is affected by multiple time steps in the past; when calculating the sub-attention score, it will result in O(L2) memory usage, which will further increase the difficulty of modeling from a hardware perspective.

[0139] To address this problem, a sparse mode is configured in the prediction model; the prediction model can be configured to selectively calculate attention weights in the sparse mode to enhance the attention weights of features at some positions of the observation sequence and reduce the attention weights of features at another part of the observation sequence. Specifically, the sparse mode calculates the similarity between the query matrix and the key matrix, such as using a dot product operation or other methods, and outputs a similarity matrix. Each element in the similarity matrix represents the similarity score between the query matrix and the key matrix. Create an attention weight matrix of the same size as the similarity matrix and initialize all elements to zero. Specify some positions in the sparse mode, and apply the similarity score to the attention weight for the specified position in the sparse mode. Exponentiate the similarity score at the specified position in the sparse mode. Normalize the indexed attention weights to ensure that their sum is 1 or close to 1. By exponentializing the similarity score and applying it to the attention weight, the specified position in the sparse mode will receive an enhanced weight when calculating the attention, so that the information that the model considers important can be better captured. This approach can enhance the model's attention at specific positions while maintaining computational efficiency. For the observation sequence, the attention weights of the features at some positions of the observation sequence can be enhanced and the attention weights of the features at other positions of the observation sequence can be reduced.

[0140] The specified positions in the sparse pattern can be randomly selected or selected based on feature importance. For example, the user's preference can be used as the specific position of the sparse pattern.

[0141] like Fig.10 As shown in Figure 1, the sparse pattern can be formed based on the LogSparse Self Attention strategy. The prediction model includes multiple sub-attention layers, each of which is configured with a sparse pattern. That is, in each sub-attention layer, the above steps are performed, and finally multiple sub-attention layers are stacked to enhance the expression ability of the model and significantly reduce the memory cost.

[0142] Based on the Transformer prediction model provided by the present invention, long-term dependencies can be captured. Compared with recent RNN-based methods, matrix decomposition methods and classical statistical methods, the proposed convolutional sub-attention further improves the performance of Transformer and reaches the most advanced level in different environments. In addition, under the same memory budget, the sparse attention model can obtain better results on data with long-term dependencies.

[0143] The control module 200 can also be configured to configure the first sparse mode strategy, the second sparse mode strategy and the third sparse mode strategy in the prediction model based on the Transformer architecture established according to historical samples. The first sparse mode strategy, the second sparse mode strategy and the third sparse mode strategy all adopt the LogSparse Self Attention strategy. The specified position in the first sparse mode strategy is the user preference, the specified position in the second sparse mode strategy is the environmental parameters of the closest several time points, and the specified position in the third sparse mode strategy is the energy price parameter. The control module 200 can choose to adopt the first sparse mode strategy, the second sparse mode strategy and the third sparse mode strategy according to user needs. For example, during normal use, the control module 200 executes a first sparse mode strategy to increase the influence of user preferences on the prediction; when predicting the set temperature for a short period of time in the future, the control module 200 executes a second sparse mode strategy to increase the influence of environmental parameters at the most recent time points on the prediction; when paying more attention to energy prices, the control module 200 executes a third sparse mode strategy to increase the influence of energy price parameters on the prediction, thereby enabling the prediction model to better capture past trends and patterns and achieve optimal configuration so that the model can perform well in predicting future set temperatures.

[0144] like Fig.14 As shown, in some embodiments of the present application, the air conditioning device further includes a training module 400. The training module 400 is configured to input a plurality of observation sequences to train the prediction model to adjust the hyperparameters of the prediction model until the training model converges. More specifically, the training module 400 uses the observation sequence to perform parameter tuning on the prediction model, and selects appropriate hyperparameters according to the loss function; for example, the number of layers of the encoder and the decoder, the number of attention heads, etc., until the prediction model reaches a converged state.

[0145] like Fig.15 As shown, in some embodiments of the present application, the air conditioning device further includes a testing module 500. The testing module 500 tests the prediction model trained by the training module 400 to evaluate whether the prediction result of the prediction model meets the set accuracy condition. The evaluation index includes the root mean square error (RMSE) or the mean absolute error (MAE), etc.

[0146] like Fig.16 As shown, in some embodiments of the present application, the air conditioning device further includes an evaluation module 600. The evaluation module 600 is configured to correct the user set temperature to the predicted set temperature when the user set temperature at the future target moment deviates from the predicted set temperature; and collect whether the intervention set temperature output by the user is received within the set evaluation period; and when the intervention set temperature is received, update the observation sequence and configure the training module 400 to train the prediction model again. When the intervention set temperature output by the user is received, it indicates that the user's subjective feeling is not good during the evaluation stage, and it is necessary to modify the observation sequence and conduct the experiment again.

[0147] like Fig.17 As shown, in one or more embodiments of the present application, an air conditioning device is also provided. The air conditioning device includes a data acquisition module 100 and a control module 200, wherein the data acquisition module 100 is configured to collect environmental parameters, operating parameters, and / or user behavior parameters of the air conditioning device; the control module 200 is configured to establish a prediction model based on the Transformer architecture according to historical samples; and, based on the established prediction model, predict the predicted set temperature at the target future moment corresponding to the restart of the air conditioning device; and, when the air conditioning device is restarted, the predicted set temperature is used as the set temperature of the air conditioning device; the historical samples are generated based on the environmental parameters, operating parameters and / or user behavior parameters associated with the corresponding timestamps and stored in a predetermined database.

[0148] During the use of the air conditioning device, there is a situation where the indoor unit or the controller has a standby function, that is, the air conditioning device is usually set to retain the previous set temperature when it is turned off. That is, the set temperature is stored in the non-volatile memory of the air conditioning device. When the user leaves the room or goes to sleep, he often chooses to turn off the air conditioning equipment, and when the air conditioning device is started again, the set temperature stored in the standby state is usually very inconsistent with the actual situation. In the present application, when the air conditioning device is started again, the control module 200 corrects the set temperature stored in the standby state to the predicted set temperature, which can meet the needs of the actual current situation, and can respond to temperature changes and user needs more quickly instead of relying solely on the previous static settings. For example, when just getting up in the morning, it can provide the user with a comfortable environment.

[0149] In the description of the above embodiments, specific features, structures, materials or characteristics may be combined in a suitable manner in any one or more embodiments or examples.

[0150] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. An air conditioning device, comprising: A data collection module, wherein the data collection module is configured to collect environmental parameters, operating parameters, and / or user behavior parameters of the air conditioning device; Characterized in that the air conditioning device also includes: A control module, wherein the control module is configured to establish a prediction model based on a Transformer architecture according to historical samples; and predict a predicted set temperature at a plurality of future target moments based on the established prediction model; and when the user set temperature at the future target moment deviates from the predicted set temperature, correct the user set temperature to the predicted set temperature; The historical samples are generated based on the environmental parameters, operating parameters and / or user behavior parameters associated with corresponding timestamps and stored in a predetermined database.

2. The air conditioning device according to claim 1, characterized in that: A causal convolution layer is configured in the prediction model; the prediction model uses the output sequence of the causal convolution layer to calculate a query matrix and a key matrix, and uses the query matrix and the key matrix to calculate the attention weight of the prediction model to process the dependency between features at different positions in the observation sequence; The observation sequence is generated based on one or more parameters in the historical samples and serves as an input sequence of the prediction model.

3. The air conditioning device according to claim 1, characterized in that: A sparse mode is configured in the prediction model; the prediction model can be configured to selectively calculate attention weights in the sparse mode to enhance the attention weights of features at a portion of the observation sequence and reduce the attention weights of features at another portion of the observation sequence; The observation sequence is generated based on one or more parameters in the historical samples and serves as an input sequence of the prediction model.

4. The air conditioning device according to claim 3, characterized in that: The sparse pattern is formed based on the LogSparse Self Attention strategy; the prediction model includes a plurality of sub-attention layers, each of which is configured with the sparse pattern; The prediction model stack is configured with multiple sub-attention layers of the sparse pattern.

5. The air conditioning device according to claim 2 or 3, characterized in that: The operating parameters include the actual operating status of the air conditioning device; The actual operating state includes user-set parameters; The air conditioning device further comprises: A preprocessing module, wherein the preprocessing module is configured to enter a waiting state when the data acquisition module acquires any one or more user-set parameters, and to determine whether the set parameters acquired within a set time period remain unchanged; when the set parameters acquired within the set time period remain unchanged, it is inferred that the acquired set parameters are valid user-set parameters, and the valid user-set parameters are associated with corresponding timestamps and stored in a predetermined database as part of historical samples.

6. The air conditioning device according to claim 2 or 3, characterized in that: The control module is also configured with a correction model independent of the prediction model, and the correction model is configured to generate a set temperature correction value; The control module is configured to associate the set temperature correction value with the corresponding time stamp and store it in a predetermined database as a component of a historical sample.

7. The air conditioning device according to claim 2 or 3, characterized in that: The air conditioning device further comprises: A training module is configured to input a plurality of the observation sequences to train the prediction model to adjust the hyperparameters of the prediction model until the training model converges.

8. The air conditioning device according to claim 7, characterized in that: The air conditioning device further comprises: A testing module is used to test the prediction model trained by the training module to evaluate whether the prediction result of the prediction model meets the set accuracy condition.

9. The air conditioning device according to claim 8, characterized in that: The air conditioning device further comprises: An evaluation module, wherein the evaluation module is configured to correct the user set temperature to the predicted set temperature when the user set temperature at the future target moment deviates from the predicted set temperature; and collect whether the intervention set temperature output by the user is received within a set evaluation period; and when the intervention set temperature is received, update the observation sequence and configure the training module to train the prediction model again.

10. An air conditioning device, comprising: A data collection module, wherein the data collection module is configured to collect environmental parameters, operating parameters, and / or user behavior parameters of the air conditioning device; Characterized in that the air conditioning device also includes: A control module, wherein the control module is configured to establish a prediction model based on a Transformer architecture according to historical samples; Furthermore, based on the established prediction model, a predicted set temperature at a target future time corresponding to when the air conditioning device is restarted is predicted; and when the air conditioning device is restarted, the set temperature stored during standby is corrected to the predicted set temperature; The historical samples are generated based on the environmental parameters, operating parameters and / or user behavior parameters associated with corresponding timestamps and stored in a predetermined database.

Images