A prediction system and method for vehicle-mounted intelligent air conditioning

Abstract

The present invention discloses a prediction system for an on-board intelligent air conditioner, which is characterized by comprising: an on-board terminal, which is used to construct the input tensors required for model reasoning through various types of collected data, and to extract short-term historical time series features through LSTM and XGBoost combined with static variables to make a final regression strategy model, and to send the result of the strategy model reasoning to the on-board air conditioner controller, to judge whether the user accepts the result of the model reasoning, to record the manual adjustment behavior and to feed back to the cloud server; the cloud server is used to obtain the optimal strategy model and a strategy evaluation report for OTA optimization A / B testing based on multiple versions of the air conditioner strategy model, user interaction feedback and simulation test results, and to send the parameters of the trained lightweight strategy model and the strategy table to the on-board terminal via OTA or API. A prediction method for an on-board intelligent air conditioner is also disclosed. The present invention accurately predicts the air conditioner settings expected by the user, improves cabin comfort, and optimizes energy consumption management.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent cockpits and artificial intelligence temperature control systems, and specifically to a prediction system and method for vehicle-mounted intelligent air conditioning. Background Art

[0002] At present, automobile air conditioning control systems mainly adopt the following two methods:

[0003] 1. Rule-based temperature control strategies (such as PID control): They cannot adapt to user needs and the temperature control adjustment is not intelligent enough.

[0004] Prediction methods based on traditional machine learning (such as regression algorithms): They only use historical data for prediction, cannot accurately understand user habits, and lack the ability to integrate external environments and voice interactions.

[0005] The existing technology has the following problems:

[0006] 1. Insufficient personalized prediction capabilities: Traditional methods are based solely on historical temperature control data and lack data such as user driving style and voice commands, making it difficult to accurately predict the user's desired air conditioning settings.

[0007] Lack of multimodal data fusion: Current methods cannot fully utilize vehicle sensor data (in-vehicle temperature, wind speed), weather APIs, user voice and other data, resulting in limited prediction accuracy.

[0008] Difficult to adapt to dynamic changes: Sudden changes in ambient temperature, new users, extreme weather, etc. will reduce prediction accuracy, and existing methods are difficult to dynamically optimize air conditioning settings.

[0009] Therefore, to address the above problems, a prediction system and method for vehicle-mounted intelligent air conditioning are proposed. Summary of the Invention

[0010] The purpose of the present invention is to overcome the existing defects and provide a prediction system and method for vehicle-mounted intelligent air conditioning, which accurately predicts the air conditioning settings expected by the user, improves cabin comfort, and optimizes energy consumption management.

[0011] The technical solution to achieve the above purpose is:

[0012] A prediction system for an in-vehicle intelligent air conditioner according to one embodiment of the present invention includes: an in-vehicle terminal and a cloud server,

[0013] The vehicle-mounted terminal is used to construct the input tensors required for model inference through the collected data, and to extract short-term historical time series features through LSTM and XGBoost combined with static variables to form a final regression strategy model. The result of the strategy model inference is sent to the vehicle air conditioning controller to determine whether the user accepts the result of the model inference, record the manual adjustment behavior and feedback it to the cloud server;

[0014] The cloud server is used to collect external environmental information of the current user's location and use context fusion methods to enhance the accuracy of user air-conditioning demand reasoning, obtain the current environmental perception vector and multimodal fusion features as part of the reinforcement learning state space; and construct a state-action-reward SAR model for air-conditioning control to optimize the fusion state features and user behavior feedback, obtain the air-conditioning strategy model and the optimal strategy model set for different user types; based on multiple versions of air-conditioning strategy models, user interaction feedback and simulation test results, obtain the optimal strategy model and a strategy evaluation report for OTA optimization A / B testing; and send the parameters and strategy table of the trained lightweight strategy model to the vehicle-mounted terminal via OTA or API.

[0015] Preferably, the vehicle-mounted terminal specifically includes:

[0016] Data collection module, used to collect various types of data and unify the timestamps;

[0017] The data preprocessing module is used to clean and normalize the data, construct time series using sliding windows, and build the input tensors required for model inference;

[0018] The real-time model inference module is used to extract short-term historical time series features through LSTM (Long Short-Term Memory Network) and XGBoost (an integrated machine learning algorithm based on decision trees) combined with static variables to create a final regression strategy model, outputting air conditioning setting recommendations in real time;

[0019] The air conditioning control interface module is used to send the results of the policy model reasoning to the vehicle air conditioning controller, determine whether the user accepts the results of the model reasoning, record the manual adjustment behavior and feedback it to the cloud server and user behavior analysis module;

[0020] The user behavior analysis module is used to store the air conditioner usage records of the past 7 days, calculate the user's behavioral preferences during use, establish user profiles, and provide paranoid tuning strategies for the policy model;

[0021] The data upload and cloud communication module is used to periodically upload local behavior data to the cloud for policy model optimization, and receive the updated policy model parameters or policies from the cloud server.

[0022] The cloud server specifically includes:

[0023] The user interaction analysis module is used to process user voice commands, app commands, or historical interaction data through a large model to extract user preference keywords and semantic intents in air conditioning usage, and obtain user intent labels, keyword vectors, and usage preference embeddings;

[0024] The long-term preference modeling module aggregates users' historical air conditioning setting behaviors over multiple days or months. It uses clustering, collaborative filtering, or large-model embedding (the process of mapping high-dimensional data into a low-dimensional space) to generate preference vectors and user profile labels for inference.

[0025] The environmental fusion perception module is used to collect external environmental information of the current user's location and use context fusion methods to enhance the accuracy of the user's air conditioning demand reasoning. It obtains the current environmental perception vector and multimodal fusion features as part of the reinforcement learning state space;

[0026] A reinforcement learning training module is used to optimize the fused state features and user behavior feedback by building a state-action-reward SAR model for air conditioning control, thereby obtaining the air conditioning policy model and the optimal policy model set for different user types.

[0027] The strategy evaluation and optimization module is used to obtain the optimal strategy model and a strategy evaluation report for OTA (Over-the-Air) optimization A / B (A / B refers to two different versions or strategies) testing based on multiple versions of air conditioning strategy models, user interaction feedback, and simulation test results.

[0028] The strategy delivery interface module is used to deliver the parameters of the trained lightweight strategy model and the strategy table to the vehicle-mounted terminal via OTA or API.

[0029] Preferably, in the data acquisition module, various types of data include vehicle interior temperature, humidity, exterior ambient temperature, humidity sensor, vehicle speed, driving mode from CAN bus data, as well as calling third-party weather API weather, air quality, ultraviolet intensity, pollution index, current air conditioning setting value setting temperature, air volume, and air conditioning mode.

[0030] Preferably, the user interaction analysis module takes as input the voice recognition text uploaded by the vehicle-mounted terminal, APP interaction data, and user adjustment logs, and outputs the user intent label and keyword vector and usage preference embedding;

[0031] The long-term preference modeling module takes as input the user's intention label and keyword vector, the embedding of usage preferences, historical behavior data, and daily vehicle operation data. The module outputs the user's long-term air conditioning control preference model, the preference vector used for inference, and the user profile label.

[0032] In the environmental fusion perception module, external environmental information includes geographic location, real-time weather, time period, and traffic status; the input is a third-party weather API (application programming interface), geographic location information, traffic conditions, time period holiday labels, and the output is the current environmental perception vector and multimodal fusion features, which serve as part of the reinforcement learning state space.

[0033] Preferably, in the reinforcement learning training module, the status includes user preferences, environmental status, in-vehicle temperature and humidity, and driving mode; the action is the adjustment of air conditioning control parameters, including temperature setting, air volume setting, and setting mode; and the reward is based on user satisfaction, including the setting of energy consumption and comfort indicators;

[0034] User behavior feedback includes manual adjustment behavior, satisfaction score, and energy consumption.

[0035] Preferably, in the policy delivery interface module, the input is the policy model file, user identification, vehicle identification and delivery plan, and the output is the policy parameter configuration, synchronization status feedback and success / failure records.

[0036] A second method of predicting a vehicle-mounted intelligent air conditioner according to the present invention includes:

[0037] Step S1, collecting data information and performing pre-processing operations on the data;

[0038] Step S2: Analyze the user's voice command and, in combination with historical data, construct a personalized temperature control feature vector for fine-tuning the intelligent air conditioning strategy model or optimizing the strategy;

[0039] Step S3: Use LSTM to process time series data and predict the temperature change trend in the next 10 minutes. Use XGBoost to fuse multi-dimensional features, correct the LSTM prediction error, and calculate the optimal air conditioning setting value based on driving patterns and environmental factors.

[0040] Step S4: Optimizing the fused state features and user behavior feedback by constructing a state-action-reward SAR model for air conditioning control to obtain an air conditioning policy model and a set of optimal policy models for different user types;

[0041] Step S5: Obtaining an optimal policy model and a policy evaluation report for OTA optimization A / B testing based on the multi-version air conditioning policy model, user interaction feedback, and simulation test results;

[0042] Step S6: Send the trained lightweight strategy model parameters and strategy table to the vehicle-mounted terminal via OTA or API.

[0043] Preferably, in step S1, the data information includes vehicle data, environmental data, and user data; wherein the vehicle data includes vehicle interior temperature, vehicle interior humidity, air conditioning setting value, vehicle speed, engine speed, battery power, and driving mode; environmental data includes outdoor temperature, humidity, wind speed, precipitation, city, altitude, weather description, ultraviolet intensity, pollution index, time, and season; user data includes voice interaction logs, previous manual air conditioning setting records, driving habits, whether windows are open, and whether AUTO is turned on;

[0044] Pre-process the vehicle data, including:

[0045] Median filtering and mean filtering are used for denoising. Median filtering is used to suppress isolated outliers. The formula is:

[0046] ;

[0047] Where k is the half-width of the window, is the original sampling value;

[0048] Mean filtering is to smooth the fluctuation trend, the formula is:

[0049] ;

[0050] Where, is the value after median filtering;

[0051] In order to adapt to the input of the neural network, Min-Max normalization is adopted to map each variable to the interval [0,1] to maintain consistency in the physical sense. The formula is:

[0052] ;

[0053] Where, and are the minimum and maximum values ​​of the variable in the training set or sliding window respectively;

[0054] Outlier detection and removal: Eliminate erroneous and unreasonable data through a variety of rules, including:

[0055] Rule 1: Constrained by physical boundaries, exceeding them is considered an exception;

[0056] Rule 2: Sliding window difference method: if the difference between a value and the two previous and next values ​​is greater than a set threshold, it is determined to be a jump point and is eliminated;

[0057] Rule 3: Z-Score outlier detection, for normalized variables, that is:

[0058] ;

[0059] Where, is the overall mean, is the overall standard deviation. When the value of Zt is greater than a certain threshold, it is considered an outlier.

[0060] Time series windowing: Prepare historical window sequences for LSTM model input.

[0061] Preferably, in step S2, the user's voice commands, APP commands or historical interaction data are processed by a large model to extract the user's preferred keywords and semantic intentions in air conditioning usage, and obtain the user intention label and keyword vector and usage preference embedding;

[0062] Aggregate users' historical air conditioning setting behaviors over multiple days or months, and use clustering, collaborative filtering, or large-scale model embedding to generate a preference model. This model then generates preference vectors and user profile labels for inference.

[0063] The external environment information of the current user's location is collected and the context fusion method is used to enhance the accuracy of the user's air-conditioning demand reasoning, and the current environment perception vector and multimodal fusion features are obtained.

[0064] The beneficial effects of the present invention are as follows: the present invention proposes an intelligent automobile air-conditioning prediction system and method based on large models and machine learning, which improves the accuracy of air-conditioning prediction, ensures that temperature control meets user expectations, and reduces the number of manual adjustments; provides personalized temperature control services to adapt to different driving styles, environmental conditions, and historical habits; optimizes energy consumption management, reduces energy consumption while ensuring comfort, and extends cruising range; realizes end-cloud collaborative computing, and improves computing efficiency and intelligence through on-board real-time computing + long-term cloud optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a module diagram of a prediction system for a vehicle-mounted intelligent air conditioner according to the present invention;

[0066] Figure 2 The present invention is a flowchart of a prediction method for a vehicle-mounted intelligent air conditioner. DETAILED DESCRIPTION

[0067] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. In the description of the present invention, it should be noted that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0068] The present invention will be further described below with reference to the accompanying drawings.

[0069] A prediction system for an in-vehicle intelligent air conditioner, comprising: an in-vehicle terminal and a cloud server,

[0070] The vehicle-mounted device is used to construct the input tensors required for model inference using various collected data. It then uses LSTM to extract short-term historical time series features and XGBoost combined with static variables to create the final regression strategy model. The strategy model inference results are sent to the vehicle's air conditioning controller to determine whether the user accepts the model inference results. Manual adjustments are recorded and fed back to the cloud server.

[0071] The cloud server is used to collect external environmental information about the current user's location and use context fusion methods to enhance the accuracy of user air conditioning demand reasoning, obtaining the current environmental perception vector and multimodal fusion features as part of the reinforcement learning state space; and constructing a state-action-reward SAR model for air conditioning control to optimize the fused state features and user behavior feedback to obtain the air conditioning policy model and the optimal policy model set for different user types; based on multiple versions of air conditioning policy models, user interaction feedback, and simulation test results, the optimal policy model and a policy evaluation report for OTA optimization A / B testing are obtained; and the parameters and policy table of the trained lightweight policy model are sent to the vehicle end via OTA or API.

[0072] like Figure 1 As shown, the vehicle-mounted end specifically includes: data acquisition module 1, data preprocessing module 2, real-time model inference module 3, air conditioning control interface module 4, user behavior analysis module 5, data upload and cloud communication module 6;

[0073] like Figure 1 As shown, the cloud server specifically includes: a user interaction analysis module 7, a long-term preference modeling module 8, an environment fusion perception module 9, a reinforcement learning training module 10, a strategy evaluation and optimization module 11 and a strategy delivery interface module 12.

[0074] Vehicle terminal:

[0075] The vehicle-mounted end (ECU): completes the following three functions: collecting environmental and in-vehicle status data, including temperature, humidity, vehicle speed, driving mode, etc.; running a lightweight machine learning model (this invention uses LSTM+XGBoost) to predict real-time air conditioning settings; and analyzing user usage habits through historical data to adjust air conditioning parameter settings.

[0076] The data collection module 1 is used to collect various types of data and unify the timestamps.

[0077] In an embodiment, various types of data include but are not limited to vehicle interior temperature, humidity, exterior ambient temperature, humidity sensor, vehicle speed, driving mode from CAN bus data, as well as calling third-party weather API weather, air quality, ultraviolet intensity, pollution index, current air conditioning setting values, setting temperature, air volume, and air conditioning mode.

[0078] Data preprocessing module 2 is used to clean and normalize the data, construct time series using sliding windows, and build the input tensors required for model inference.

[0079] Real-time model inference module 3 is used to extract short-term historical time series features through LSTM and XGBoost combined with static variables to create the final regression strategy model. It outputs air conditioning setting recommendations in real time, for example, recommending a temperature of 23 degrees, air volume level 3, and foot-blowing mode.

[0080] The air conditioning control interface module 4 is used to send the results of the strategy model reasoning to the vehicle air conditioning controller, determine whether the user accepts the results of the model reasoning, record the manual adjustment behavior and feed it back to the cloud server and the user behavior analysis module 5.

[0081] The user behavior analysis module 5 is used to store the air conditioner usage records for the past 7 days, calculate the user's behavioral preferences during use, establish user profiles, and provide paranoid tuning strategies for the policy model.

[0082] The data upload and cloud communication module 6 is used to periodically upload local behavior data to the cloud for policy model optimization, and receive updated policy model parameters or policies from the cloud server.

[0083] Cloud Server:

[0084] The user interaction analysis module 7 is used to process user voice commands, APP commands or historical interaction data through large models (such as DeepSeek, Doubao, Tongyi Qianwen, etc.) to extract the user's preferred keywords and semantic intentions in air conditioning usage. The input is the voice recognition text uploaded by the vehicle end, as well as APP interaction data and user adjustment logs. The output is the user intention label (such as comfort-oriented, energy-saving-oriented, personalized temperature point) and the embedding of keyword vectors and usage preferences.

[0085] The long-term preference modeling module 8 is used to aggregate the user's historical air conditioning setting behavior over multiple days or months, aggregate the preference model through clustering, collaborative filtering or large model embedding, and generate a long-term behavior profile (such as preference for 24 degrees, low wind speed, and closed internal circulation). The input is the output of the user interaction analysis module, historical behavior data, and daily vehicle operation data. The output is the user's long-term air conditioning control preference model, the preference vector used for inference, and the user profile label (such as energy-saving, comfortable, and casual).

[0086] The environmental fusion perception module 9 is used to collect external environmental information of the current user's location and use context fusion methods to enhance the accuracy of user air conditioning demand reasoning. The input is a third-party weather API, geographic location information, traffic conditions, time period and holiday labels. The output is the current environmental perception vector (weather+location+time, etc.) and multimodal fusion features as part of the reinforcement learning state space.

[0087] The reinforcement learning training module 10 is used to optimize the fusion state features and user behavior feedback by constructing a state-action-reward SAR model for air conditioning control, thereby obtaining an air conditioning policy model and a set of optimal policy models for different user types.

[0088] In this embodiment, the status includes user preferences, environmental conditions, vehicle temperature and humidity, and driving mode; the action is the adjustment of air conditioning control parameters, including but not limited to temperature setting, air volume setting, and setting mode; the reward is based on user satisfaction, including but not limited to the setting of energy consumption and comfort indicators;

[0089] User behavior feedback includes manual adjustment behavior, satisfaction score, and energy consumption.

[0090] The strategy evaluation and optimization module 11 is used to verify the effects of different versions of strategies in a simulated environment or historical data offline, and to measure strategy performance indicators such as energy consumption reduction and manual adjustment rate reduction. The input is a multi-version air conditioning strategy model, user interaction feedback and simulation test results, and the output is the optimal strategy model for distribution and a strategy evaluation report for OTA optimization A / B testing.

[0091] The policy delivery interface module 12 is used to deliver the parameters and policy table of the trained lightweight policy model to the vehicle-mounted terminal via OTA or API, and supports model version control, push logs, and rollback mechanisms. The input is the policy model file, user ID, vehicle ID, and delivery plan (periodic delivery or trigger timing), and the output is the policy parameter configuration, synchronization status feedback, and success / failure records.

[0092] Preferably, in the policy delivery interface module, the input is the policy model file, user identification, vehicle identification and delivery plan, and the output is the policy parameter configuration, synchronization status feedback and success / failure records.

[0093] like Figure 2 As shown, a prediction method for a vehicle-mounted intelligent air conditioner includes:

[0094] Step S1: collect data information and perform pre-processing operations on the data.

[0095] In an embodiment, the data information includes vehicle data, environmental data, and user data; wherein the vehicle data includes but is not limited to vehicle temperature, vehicle humidity, air conditioning setting value, vehicle speed, engine speed, battery power, and driving mode; environmental data includes but is not limited to outdoor temperature, humidity, wind speed, precipitation, city, altitude, weather description, UV intensity, pollution index, time and season; user data includes but is not limited to voice interaction logs, previous manual air conditioning setting records, driving habits, whether the windows are open and whether AUTO is turned on;

[0096] Pre-process the vehicle data, including:

[0097] Median filtering and mean filtering are used for denoising. Median filtering is used to suppress isolated outliers. The formula is:

[0098] ;

[0099] Where k is the half-width of the window, is the original sampling value;

[0100] Mean filtering is to smooth the fluctuation trend, the formula is:

[0101] ;

[0102] Where, is the value after median filtering;

[0103] In order to adapt to the input of the neural network, Min-Max normalization is adopted to map each variable to the interval [0,1] to maintain consistency in the physical sense. The formula is:

[0104] ;

[0105] Where, and are the minimum and maximum values ​​of the variable in the training set or sliding window respectively;

[0106] Outlier detection and removal: Eliminate erroneous and unreasonable data through a variety of rules, including:

[0107] Rule 1: Constrained by physical boundaries, exceeding them is considered an exception;

[0108] Rule 2: Sliding window difference method: if the difference between a value and the two previous and next values ​​is greater than a set threshold, it is determined to be a jump point and is eliminated;

[0109] Rule 3: Z-Score outlier detection, for normalized variables, that is:

[0110] ;

[0111] Where, is the overall mean, is the overall standard deviation. When the value of Zt is greater than a certain threshold, it is considered an outlier.

[0112] Time series windowing: Prepare historical window sequences for LSTM model input.

[0113] Step S2: Analyze the user's voice commands, combine them with historical data, and construct a personalized temperature control feature vector for fine-tuning the intelligent air conditioning strategy model or optimizing the strategy.

[0114] In this embodiment, a large model processes user voice commands (such as "Raise the temperature by 2 degrees," "More air volume," "It's too hot"), app commands, or historical interaction data to extract user preference keywords and semantic intents for air conditioning usage, and obtain user intent labels, keyword vectors, and usage preference embeddings.

[0115] Aggregate users' historical air conditioning setting behaviors over multiple days or months, and use clustering, collaborative filtering, or large-scale model embedding to generate a preference model. This model then generates preference vectors and user profile labels for inference.

[0116] The external environment information of the current user's location is collected and the context fusion method is used to enhance the accuracy of the user's air-conditioning demand reasoning, and the current environment perception vector and multimodal fusion features are obtained.

[0117] The input of the large model is user command text and context input such as the current air conditioning settings; the output of the large model is intent such as temperature adjustment, air volume adjustment, air conditioning on / off, air conditioning mode switching, and slots structured control fields such as target_temp, fan_speed, mode, etc.

[0118] In step S3, the time series data is processed through LSTM to predict the temperature change trend in the next 10 minutes. Multi-dimensional features are integrated through XGBoost to correct the prediction error of LSTM. The optimal air conditioning setting value is calculated based on the driving mode and environmental factors.

[0119] LSTM time series model, with input features including historical temperature series and synchronized external temperature, wind speed, driving mode, etc.

[0120] The network structure is selected as Input [W x N] → LSTM(128) → Dropout → Dense(64) → Output: [T];

[0121] The loss function uses the mean square error between the MSE predicted value and the true value as the loss function, and the optimizer uses Adam;

[0122] Model output: predicted value of the vehicle interior temperature trend in the next 10 minutes.

[0123] XGBoost model: The correction target is the error in LSTM output due to unmodeled factors such as environmental changes, which requires the integration of external features for correction.

[0124] Input features: LSTM prediction results, used as initial prediction input; current vehicle interior conditions, such as temperature, humidity, wind speed, and air conditioning mode; current external environment characteristics, such as weather type, outside temperature, and light intensity; driving behavior characteristics, such as current vehicle speed, acceleration, and driving mode; and user profile characteristics, such as user profiles favoring cold or hot conditions and adjustment frequency.

[0125] Output features: predicted value error and final revised prediction.

[0126] In this embodiment, local adaptive adjustments on the terminal and long-term cloud-based model optimization are implemented. Local adaptive adjustments involve updating temperature control parameters locally and adjusting the prediction logic if the user adjusts the air conditioning settings multiple times within a short period of time. Triggering condition: two or more manual adjustments within 30 minutes. Adjustment method: LSTM predictions weight short-term adjustment trends (e.g., if the temperature has increased by 2°C in the past 10 minutes, future predictions will default to +1°C). The system prioritizes settings frequently used over the past seven days (e.g., if the temperature is frequently set to 24°C during nighttime driving, 24°C will be recommended by default). Long-term cloud-based optimization: User behavior data is regularly uploaded to the cloud and used to train a personalized model. Triggering condition: If the user adjusts the temperature to the same level multiple times within three days in the same environment (e.g., multiple changes from 22°C to 24°C). Adjustment method: An enhanced algorithm is used to update the personalized strategy, making the prediction more tailored to user habits. The LLM (Large Length Model) analyzes user voice preferences, such as implicit feedback like "a bit cold," and adjusts the air conditioning setting strategy.

[0127] In step S4, a state-action-reward SAR model for air conditioning control is constructed to optimize the fusion state features and user behavior feedback to obtain the air conditioning policy model and the optimal policy model set for different user types.

[0128] Step S5: Obtain an optimal policy model and a policy evaluation report for OTA optimization A / B testing based on the multi-version air conditioning policy model, user interaction feedback, and simulation test results.

[0129] Step S6: Send the trained lightweight strategy model parameters and strategy table to the vehicle-mounted terminal via OTA or API.

[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A prediction system for vehicle-mounted intelligent air conditioning, characterized in that: include: On-board terminal and cloud server, The vehicle-mounted terminal is used to construct the input tensors required for model inference through the collected data, and to extract short-term historical time series features through LSTM and XGBoost combined with static variables to form a final regression strategy model. The result of the strategy model inference is sent to the vehicle air conditioning controller to determine whether the user accepts the result of the model inference, record the manual adjustment behavior and feedback it to the cloud server; The cloud server is used to collect external environment information of the current user's location and use a context fusion method to enhance the accuracy of the user's air conditioning demand reasoning, thereby obtaining a current environment perception vector and a multimodal fusion feature as part of the reinforcement learning state space; A state-action-reward SAR model for air conditioning control was constructed to optimize the fusion of state features and user behavior feedback, resulting in an air conditioning policy model and a set of optimal policy models for different user types. Based on multiple versions of air conditioning policy models, user interaction feedback, and simulation test results, the optimal policy model and a policy evaluation report for OTA optimization A / B testing were obtained. And sending the parameters and strategy table of the trained lightweight strategy model to the vehicle-mounted terminal via OTA or API; The vehicle-mounted terminal specifically includes: Data collection module, used to collect various types of data and unify the timestamps; The data preprocessing module is used to clean and normalize the data, construct time series using sliding windows, and build the input tensors required for model inference; The real-time model inference module is used to extract short-term historical time series features through LSTM and combine XGBoost with static variables to create the final regression strategy model, outputting air conditioning setting recommendations in real time; The air conditioning control interface module is used to send the results of the policy model reasoning to the vehicle air conditioning controller, determine whether the user accepts the results of the model reasoning, record the manual adjustment behavior and feedback it to the cloud server and user behavior analysis module; The user behavior analysis module is used to store the air conditioner usage records of the past 7 days, calculate the user's behavioral preferences during use, establish user profiles, and provide paranoid tuning strategies for the policy model; A data upload and cloud communication module is used to periodically upload local behavior data to the cloud for policy model optimization and receive updated policy model parameters or policies from the cloud server; The cloud server specifically includes: The user interaction analysis module is used to process user voice commands, app commands, or historical interaction data through a large model to extract user preference keywords and semantic intents in air conditioning usage, and obtain user intent labels, keyword vectors, and usage preference embeddings; The long-term preference modeling module aggregates users' historical air conditioning setting behaviors over multiple days or months. It uses clustering, collaborative filtering, or large-scale model embedding to generate preference models for inference, including preference vectors and user profile labels. The environmental fusion perception module is used to collect external environmental information of the current user's location and use context fusion methods to enhance the accuracy of the user's air conditioning demand reasoning. It obtains the current environmental perception vector and multimodal fusion features as part of the reinforcement learning state space; A reinforcement learning training module is used to optimize the fused state features and user behavior feedback by building a state-action-reward SAR model for air conditioning control, thereby obtaining the air conditioning policy model and the optimal policy model set for different user types. The strategy evaluation and optimization module is used to obtain the optimal strategy model and a strategy evaluation report for OTA optimization A / B testing based on multiple versions of air conditioning strategy models, user interaction feedback, and simulation test results; The strategy delivery interface module is used to deliver the parameters of the trained lightweight strategy model and the strategy table to the vehicle-mounted terminal via OTA or API.

2. The prediction system for vehicle-mounted intelligent air conditioning according to claim 1, characterized in that: In the data acquisition module, various types of data include vehicle interior temperature, humidity, outdoor ambient temperature, humidity sensor, vehicle speed, driving mode from CAN bus data, as well as calling third-party weather API weather, air quality, ultraviolet intensity, pollution index, air conditioning current setting value setting temperature, air volume, air conditioning mode.

3. The prediction system for vehicle-mounted intelligent air conditioning according to claim 1, characterized in that: The user interaction analysis module takes as input the voice recognition text uploaded by the vehicle, APP interaction data, and user adjustment logs, and outputs the user intent label, keyword vector, and usage preference embedding. The long-term preference modeling module takes as input the user's intention label and keyword vector, the embedding of usage preferences, historical behavior data, and daily vehicle operation data. The module outputs the user's long-term air conditioning control preference model, the preference vector used for inference, and the user profile label. In the environmental fusion perception module, external environmental information includes geographic location, real-time weather, time period, and traffic status; The input is a third-party weather API, geographic location information, traffic conditions, and time period holiday labels. The output is the current environment perception vector and multimodal fusion features, which serve as part of the reinforcement learning state space.

4. The prediction system for vehicle-mounted intelligent air conditioning according to claim 1, characterized in that: In the reinforcement learning training module, the states include user preferences, environmental conditions, in-car temperature and humidity, and driving mode; the actions are the adjustment of air conditioning control parameters, including temperature setting, air volume setting, and setting mode; and the rewards are based on user satisfaction, including the setting of energy consumption and comfort indicators; User behavior feedback includes manual adjustment behavior, satisfaction score, and energy consumption.

5. The prediction system for vehicle-mounted intelligent air conditioning according to claim 1, characterized in that: In the policy delivery interface module, the input is the policy model file, user identification, vehicle identification and delivery plan, and the output is the policy parameter configuration, synchronization status feedback and success / failure records.

6. A prediction method for an in-vehicle intelligent air conditioner using the prediction system for an in-vehicle intelligent air conditioner according to claim 1, characterized in that: include: Step S1, collecting data information and performing pre-processing operations on the data; Step S2: Analyze the user's voice command and, in combination with historical data, construct a personalized temperature control feature vector for fine-tuning the intelligent air conditioning strategy model or optimizing the strategy; Step S3: Use LSTM to process time series data and predict the temperature change trend in the next 10 minutes. Use XGBoost to fuse multi-dimensional features, correct the LSTM prediction error, and calculate the optimal air conditioning setting value based on driving patterns and environmental factors. Step S4: Optimizing the fused state features and user behavior feedback by constructing a state-action-reward SAR model for air conditioning control to obtain an air conditioning policy model and a set of optimal policy models for different user types; Step S5: Obtaining an optimal policy model and a policy evaluation report for OTA optimization A / B testing based on the multi-version air conditioning policy model, user interaction feedback, and simulation test results; Step S6: Send the trained lightweight strategy model parameters and strategy table to the vehicle terminal via OTA or API; In step S1, the data information includes vehicle data, environmental data, and user data; wherein the vehicle data includes vehicle interior temperature, vehicle interior humidity, air conditioning setting value, vehicle speed, engine speed, battery power, and driving mode; environmental data includes outdoor temperature, humidity, wind speed, precipitation, city, altitude, weather description, UV intensity, pollution index, time, and season; user data includes voice interaction logs, previous manual air conditioning setting records, driving habits, whether windows are open, and whether AUTO is turned on; Pre-process the vehicle data, including: Median filtering and mean filtering are used for denoising. Median filtering is used to suppress isolated outliers. The formula is: ； Where, k is the half-width of the window, x t is the original sampling value; Mean filtering is to smooth the fluctuation trend, the formula is: ； Where, x j is the value after median filtering; In order to adapt to the input of the neural network, Min-Max normalization is adopted to map each variable to the interval [0,1] to maintain consistency in the physical sense. The formula is: ； Where, x min and x max are the minimum and maximum values ​​of the variable in the training set or sliding window respectively; Outlier detection and removal: Eliminate erroneous and unreasonable data through a variety of rules, including: Rule 1: Constrained by physical boundaries, exceeding them is considered an exception; Rule 2: Sliding window difference method: if the difference between a value and the two previous and next values ​​is greater than a set threshold, it is determined to be a jump point and is eliminated; Rule 3: Z-Score outlier detection, for normalized variables, that is: ； Where, μ is the overall mean, σ is the overall standard deviation. When the value of Zt is greater than a certain threshold, it is considered an outlier. Time series windowing: Prepare historical window sequences for LSTM model input.

7. The prediction method for a vehicle-mounted intelligent air conditioner according to claim 6, characterized in that: In step S2, the user's voice commands, APP commands, or historical interaction data are processed by the large model to extract the user's preferred keywords and semantic intentions in air conditioning usage, and obtain the user's intention label, keyword vector, and usage preference embedding; Aggregate users' historical air conditioning setting behaviors over multiple days or months, and use clustering, collaborative filtering, or large-scale model embedding to generate a preference model. This model then generates preference vectors and user profile labels for inference. The external environment information of the current user's location is collected and the context fusion method is used to enhance the accuracy of the user's air-conditioning demand reasoning, and the current environment perception vector and multimodal fusion features are obtained.

Images