Prediction system and method for vehicle-mounted intelligent air conditioner

Abstract

The invention discloses a prediction system for a vehicle-mounted intelligent air conditioner, and the system is characterized in that the system comprises a vehicle-mounted terminal which is used for building an input tensor needed by model reasoning through various collected data, extracting short-term historical time sequence features through LSTM, and making a final regression strategy model through XGBoost in combination with static variables, a strategy model reasoning result is sent to the vehicle-mounted air conditioner controller, whether a user receives the model reasoning result or not is judged, and a manual adjustment behavior is recorded and fed back to the cloud server; and the cloud server is used for obtaining an optimal strategy model and a strategy evaluation report for OTA optimization A / B test according to the multi-version air conditioner strategy model, the interactive feedback of the user and a simulation test result, and issuing parameters of the trained lightweight strategy model and a strategy table to the vehicle-mounted end in an OTA or API mode. The invention further discloses a prediction method of the vehicle-mounted intelligent air conditioner. The air conditioner setting expected by the user is accurately predicted, the comfort degree of the cabin is improved, and energy consumption management is optimized.

Description

Technical Field

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

[0002] At present, automotive air conditioner control systems mainly adopt the following two types of methods: 1. Rule-based temperature control strategies (such as PID control): Unable to adapt to user personalized needs, and the temperature control adjustment is not intelligent enough.

[0003] Prediction methods based on traditional machine learning (such as regression algorithms): Only use historical data for prediction, unable to accurately understand user habits, and lack the ability to integrate external environment and voice interaction.

[0004] The existing technologies have the following problems: 1. Insufficient personalized prediction ability: Traditional methods are only based on historical temperature control data, lacking data such as user driving styles and voice commands, and it is difficult to accurately predict the air conditioner settings expected by users.

[0005] Lack of multi-modal data fusion: Current methods cannot make full use of vehicle sensor data (indoor temperature, wind speed), weather APIs, user voices, etc., resulting in limited prediction accuracy.

[0006] Difficulty in adapting to dynamic changes: Situations such as sudden changes in environmental temperature, new users, and extreme weather will reduce the prediction accuracy, and existing methods are difficult to dynamically optimize air conditioner settings.

[0007] Therefore, in view of the above problems, a prediction system and method for an in-vehicle intelligent air conditioner are proposed. Summary of the Invention

[0008] The purpose of the present invention is to provide a prediction system and method for an in-vehicle intelligent air conditioner to overcome the existing defects, accurately predict the air conditioner settings expected by users, improve the comfort of the cockpit, and optimize energy consumption management.

[0009] The technical solution to achieve the above purpose is as follows: A prediction system for an in-vehicle intelligent air conditioner according to one aspect of the present invention includes: an in-vehicle terminal and a cloud server, The in-vehicle terminal is used to construct the input tensors required for model inference through various collected data, extract short-term historical time series features through LSTM, and combine static variables with XGBoost to make the final regression strategy model, send the results of the strategy model inference to the in-vehicle air conditioner controller, determine whether the user accepts the results of the model inference, record manual adjustment behaviors, and feedback them to the cloud server; The cloud server is used to collect the external environmental information of the current user's location and enhance the accuracy of the user's air-conditioning demand inference by using the context fusion method, so as to obtain the current environmental perception vector and multi-modal fusion features, which are used as part of the reinforcement learning state space; and construct a state-action-reward S-A-R model for air-conditioning control to optimize the fused state features and user behavior feedback, so as to obtain the air-conditioning policy model and the optimal policy model set for different user types; according to the multi-version air-conditioning policy model, the user's interaction feedback and the results of simulation tests, obtain the optimal policy model and the policy evaluation report for OTA optimization A / B testing; and send the parameters and policy table of the trained lightweight policy model to the vehicle-mounted terminal through OTA or API.

[0010] Preferably, the vehicle-mounted terminal specifically includes: The data collection module is used to collect various types of data and unify the timestamps; The data preprocessing module is used to clean and normalize the data, construct a time series by means of a sliding window, and construct the input tensor required for model inference; The real-time model inference module is used to extract short-term historical time series features through LSTM (Long Short-Term Memory Network) and combine with static variables by XGBoost (an integrated machine learning algorithm based on decision trees) to make the final regression policy model, and output the setting suggestions of the air conditioner in real time; The air-conditioning control interface module is used to send the results of policy model inference to the vehicle-mounted air-conditioning controller, judge whether the user accepts the results of model inference, record the manual adjustment behavior and feedback it to the cloud server and the user behavior analysis module; The user behavior analysis module is used to store the air-conditioning usage records of the past 7 days, calculate the behavior preferences of the user during the usage process, establish a user portrait, and provide a bias tuning strategy for the policy model; 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.

[0011] The cloud server specifically includes: The user interaction parsing module is used to process user voice commands, APP commands or historical interaction data through a large model, extract the preference keywords and semantic intentions of the user in terms of air-conditioning usage, and obtain the user intention label, keyword vector and embedding of usage preferences; The long-term preference modeling module is used to aggregate the historical air-conditioning setting behaviors of the user for multiple days or months, and aggregate the preference model through clustering, collaborative filtering or large model embedding (the process of mapping high-dimensional data to a low-dimensional space) for the preference vector and user portrait label used for inference; An environment fusion perception module, which is used to collect external environment information of the current location of the user and enhance the accuracy of inferring the user's air-conditioning demand by using the context fusion method, so as to obtain the current environment perception vector and multi-modal fusion features, which are used as part of the state space of reinforcement learning; A reinforcement learning training module, which is used to optimize the fused state features and user behavior feedback by constructing a state-action-reward S-A-R model for air-conditioning control, so as to obtain the policy model of the air-conditioning and the set of optimal policy models for different user types; A policy evaluation and optimization module, which is used to obtain the optimal policy model and a policy evaluation report for OTA (Over-the-Air) optimization A / B (A / B refers to two different versions or policies) testing according to the multi-version air-conditioning policy model, the interactive feedback of the user, and the results of simulation tests; A policy distribution interface module, which is used to distribute the parameters of the trained lightweight policy model and the policy table to the vehicle-mounted terminal through OTA or API.

[0012] Preferably, in the data collection module, various types of data include the temperature and humidity inside the vehicle, the temperature and humidity outside the vehicle, vehicle speed, driving mode from the can bus data, and the weather, air quality, ultraviolet intensity, pollution index, current setting value of the air-conditioning (set temperature, air volume, air-conditioning mode) called from the third-party weather api.

[0013] Preferably, in the user interaction parsing module, the input is the speech recognition text uploaded by the vehicle-mounted terminal, the APP interaction data, and the user adjustment log, and the output is the user intention label, the keyword vector, and the embedding of the usage preference; In the long-term preference modeling module, the input is the user intention label, the keyword vector, the embedding of the usage preference, the historical behavior data, and the daily operation data of the vehicle, and the output is the user's long-term air-conditioning control preference model, the preference vector for inference, and the user portrait label; In the environment fusion perception module, the external environment information includes geographical location, real-time weather, time period, and traffic status; the input is the third-party weather API (Application Programming Interface), geographical location information, traffic conditions, time period holiday label, and the output is the current environment perception vector and multi-modal fusion features, and they are used as part of the state space of reinforcement learning.

[0014] Preferably, in the reinforcement learning training module, the state includes user preference, environment state, temperature and humidity inside the vehicle, 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 the user's satisfaction, including the setting of energy consumption and comfort index; The user behavior feedback includes manual adjustment behavior, satisfaction score, and energy consumption.

[0015] Preferably, in the policy distribution interface module, the inputs are a policy model file, a user identification, a vehicle identification, and a distribution plan, and the outputs are policy parameter configurations, synchronization status feedbacks, and success / failure records.

[0016] A prediction method for an in-vehicle intelligent air conditioner according to the second aspect of the present invention includes: Step S1: Collect data information and perform preprocessing operations on the data; Step S2: Analyze the user's voice commands, combine historical data, and construct a personalized temperature control feature vector for fine-tuning the intelligent air conditioner policy model or optimizing the policy; Step S3: Process time series data through LSTM to predict the temperature change trend in the next 10 minutes, fuse multi-dimensional features through XGBoost to correct the prediction error of LSTM, and calculate the optimal air conditioner setting value in combination with the driving mode and environmental factors; Step S4: Optimize the fused state features and user behavior feedback by constructing a state-action-reward S-A-R model for air conditioner control to obtain an air conditioner policy model and a set of optimal policy models for different user types; Step S5: According to the multi-version air conditioner policy model, the user's interaction feedback, and the results of simulation tests, obtain the optimal policy model and a policy evaluation report for OTA optimization A / B testing; Step S6: Transmit the parameters of the trained lightweight policy model and the policy table to the in-vehicle terminal through OTA or API.

[0017] Preferably, in step S1, the data information includes vehicle-mounted data, environmental data, and user data; among them, the vehicle-mounted data includes the in-vehicle temperature, in-vehicle humidity, air conditioner setting value, vehicle speed, engine speed, battery power, and driving mode; the environmental data includes the outdoor temperature, humidity, wind speed, precipitation, city where located, altitude, weather description, ultraviolet intensity, pollution index, time, and season; the user data includes voice interaction logs, previous manual air conditioner setting records, driving habits, whether the window is opened, and whether AUTO is turned on; The preprocessing operations on the vehicle-mounted data include: Adopt median filtering and mean filtering methods for denoising. Median filtering is to suppress isolated outliers, and the formula is: ; In the formula, k is the half-width of the window, is the original sampling value; Mean filtering is to smooth the fluctuation trend, and the formula is: ; In the formula, is the value after median filtering; 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 physical meaning. The formula is: ; In the formula, and are the minimum and maximum values of the variable in the training set or the sliding window respectively; Outlier detection and removal, eliminating incorrect and unreasonable data through multiple rules, including: Rule 1: Constrained by physical boundaries, values beyond which are regarded as outliers; Rule 2: Sliding window difference method. If the difference between a value and the two adjacent values is greater than a set threshold, it is determined as a jump point and removed; Rule 3: Z-Score outlier detection. For the normalized variable, that is: ; In the formula, is the overall mean, is the overall standard deviation. When the value of Zt is greater than a certain threshold, it is regarded as an outlier; Time series windowing processing: Prepare the historical window sequence for the input of the LSTM model.

[0018] Preferably, in step S2, the large model processes the user voice command, APP command or historical interaction data to extract the preference keywords and semantic intentions of the user in air conditioner usage, obtaining the user intention label, as well as the keyword vector and the embedding of the usage preference; Aggregate the historical air conditioner setting behaviors of the user over multiple days or months, and aggregate a preference model through clustering, collaborative filtering or large model embedding for the preference vector for inference and the user portrait label; Collect the external environment information of the current user's location and use the context fusion method to enhance the accuracy of the user's air conditioner demand inference, obtaining the current environment perception vector and the multi-modal fusion feature.

[0019] The beneficial effects of the present invention are as follows: The present invention proposes an intelligent vehicle air conditioner prediction system and method based on a large model and machine learning, which improves the accuracy of air conditioner prediction, ensures that the temperature control meets the user's expectations, and reduces the number of manual adjustments; provides personalized temperature control services, adapting to different driving styles, environmental conditions, and historical habits; optimizes energy consumption management, reduces energy consumption on the premise of ensuring comfort, and extends the battery life; realizes end-cloud collaborative computing, improving the computing efficiency and intelligence through on-vehicle real-time computing + long-term cloud optimization. Brief Description of the Drawings

[0020] Figure 1 It is a module diagram of a prediction system for an in-vehicle intelligent air conditioner according to the present invention; Figure 2 It is a flowchart of a prediction method for an in-vehicle intelligent air conditioner according to the present invention. Specific embodiments

[0021] Next, the technical solution of the present invention will be clearly and completely described in conjunction with the accompanying drawings. In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying 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 a limitation to the present invention. In addition, the terms "first", "second", "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0022] Next, the present invention will be further described in conjunction with the accompanying drawings.

[0023] A prediction system for an in-vehicle intelligent air conditioner includes: an in-vehicle terminal and a cloud server, The in-vehicle terminal is used to construct the input tensor required for model inference through various collected data, extract short-term historical time series features through LSTM, and combine static variables with XGBoost to make the final regression strategy model, send the result of the strategy model inference to the in-vehicle air conditioner controller, 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 the external environment information of the current user's location and enhance the accuracy of user air conditioner demand inference using the context fusion method to obtain the current environmental perception vector and multi-modal fusion features, which are used as part of the state space of reinforcement learning; and construct a state-action-reward S-A-R model for air conditioner control to optimize the fusion state features and user behavior feedback to obtain the air conditioner's strategy model and the optimal strategy model set for different user types; obtain the optimal strategy model and the strategy evaluation report for OTA optimization A / B testing according to the multi-version air conditioner strategy model, user interaction feedback, and simulation test results; and send the parameters and strategy table of the trained lightweight strategy model to the in-vehicle terminal through OTA or API.

[0024] As Figure 1 shown, the in-vehicle terminal specifically includes: a data acquisition module 1, a data preprocessing module 2, a real-time model inference module 3, an air conditioner control interface module 4, a user behavior analysis module 5, and a data upload and cloud communication module 6; AsFigure 1 As shown in the figure, the cloud server specifically includes: a user interaction parsing module 7, a long-term preference modeling module 8, an environment integration perception module 9, a reinforcement learning training module 10, a policy evaluation and optimization module 11, and a policy distribution interface module 12.

[0025] Vehicle side:

[0026] Vehicle side (ECU computing unit): The functions completed are as follows: collecting environmental and in-vehicle status data, including temperature, humidity, vehicle speed, driving mode, etc.; running a lightweight machine learning model (LSTM + XGBoost is used in the present invention) to predict real-time air conditioner settings; analyzing the user's usage habits through historical data and adjusting the air conditioner parameter settings.

[0027] A data acquisition module 1, used to collect various types of data and unify the timestamps.

[0028] In the embodiment, various types of data include but are not limited to vehicle interior temperature, humidity, exterior environmental temperature, humidity sensor, vehicle speed, driving mode from can bus data, and calling third-party weather api for weather, air quality, ultraviolet intensity, pollution index, current air conditioner setting value (set temperature, air volume, air conditioner mode).

[0029] A data preprocessing module 2, used to clean, normalize the data, construct a time series through a sliding window, and construct the input tensor required for model inference.

[0030] A real-time model inference module 3, used to extract short-term historical time series features through LSTM and combine with static variables by XGBoost to make a final regression policy model, and output real-time air conditioner setting suggestions, for example, recommend a temperature of 23 degrees, an air volume of 3 gears, and a foot-blowing mode.

[0031] An air conditioner control interface module 4, used to send the result of the policy model inference to the vehicle-mounted air conditioner controller, determine whether the user accepts the result of the model inference, record the manual adjustment behavior, and feedback it to the cloud server and the user behavior analysis module 5.

[0032] A user behavior analysis module 5, used to store the air conditioner usage records of the past 7 days, calculate the user's behavior preferences during use, establish a user profile, and provide a bias tuning strategy for the policy model.

[0033] A data upload and cloud communication module 6, 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.

[0034] Cloud server:

[0035] The user interaction parsing 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.), and extract preference keywords and semantic intentions of users in air conditioner use. The input is the speech recognition text uploaded by the in-vehicle terminal, APP interaction data, and user adjustment logs, and the output is user intention tags (such as comfort-oriented, energy-saving-oriented, personalized temperature points), as well as keyword vectors and embeddings of usage preferences.

[0036] The long-term preference modeling module 8 is used to aggregate the historical air conditioner setting behaviors of users for multiple days or months, and aggregate a preference model through clustering, collaborative filtering or large model embedding to generate a long-term behavior portrait (such as preferring 24 degrees, low wind speed, and turning off the internal circulation). The input is the output of the user interaction parsing module, historical behavior data, and vehicle daily operation data, and the output is the user's long-term air conditioner control preference model, the preference vector for inference, and user portrait tags (such as energy-saving type, comfort type, casual type).

[0037] The environment fusion perception module 9 is used to collect the external environment information of the current user location and enhance the accuracy of user air conditioner demand inference using the context fusion method. The input is the third-party weather API, geographical location information, traffic conditions, and time period holiday tags, and the output is the current environment perception vector (weather + location + time, etc.) and multi-modal fusion features, which are used as part of the reinforcement learning state space.

[0038] The reinforcement learning training module 10 is used to optimize the fused state features and user behavior feedback by constructing a state-action-reward S-A-R model for air conditioner control, and obtain the air conditioner's policy model and the optimal policy model set for different user types.

[0039] In the embodiment, the state includes user preferences, environmental status, in-vehicle temperature and humidity, and driving mode. The action is the adjustment of air conditioner control parameters, including but not limited to temperature setting, air volume setting, and setting mode. The reward is based on the user's satisfaction, including but not limited to the setting of energy consumption and comfort indicators; User behavior feedback includes manual adjustment behavior, satisfaction score, and energy consumption.

[0040] The policy evaluation and optimization module 11 is used to offline verify the effects of different versions of policies in the simulation environment or historical data. The performance metrics for measuring the policy are energy consumption reduction and the decrease in the manual adjustment rate. The input is the multi-version air conditioner policy model, the user's interaction feedback, and the results of the simulation test, and the output is the optimal policy model for distribution and the policy evaluation report for OTA optimization A / B testing.

[0041] The policy distribution interface module 12 is used to distribute the parameters of the trained lightweight policy model and the policy table to the in-vehicle terminal through OTA or API, and support model version control, push logs, and rollback mechanisms. The input is the policy model file, user identification, vehicle identification, and distribution plan (periodic distribution or trigger timing), and the output is the policy parameter configuration, synchronization status feedback, and success / failure records.

[0042] Preferably, in the policy distribution interface module, the input is the policy model file, user identification, vehicle identification, and distribution plan, and the output is the policy parameter configuration, synchronization status feedback, and success / failure records.

[0043] As Figure 2 shown, a prediction method for an in-vehicle intelligent air conditioner includes: Step S1, collect data information and perform preprocessing operations on the data.

[0044] In the embodiment, the data information includes in-vehicle data, environmental data, and user data; among them, the in-vehicle data includes but is not limited to the in-vehicle temperature, in-vehicle humidity, air conditioner set value, vehicle speed, engine speed, battery power, and driving mode; the environmental data includes but is not limited to the outdoor temperature, humidity, wind speed, precipitation, city where located, altitude, weather description, ultraviolet intensity, pollution index, time, and season; the user data includes but is not limited to voice interaction logs, previous manual air conditioner setting records, driving habits, whether the window is opened, and whether AUTO is turned on. The preprocessing operations on the in-vehicle data include: Adopt median filtering and mean filtering methods for denoising. Median filtering is to suppress isolated outliers, and the formula is: ; In the formula, k is the half-width of the window, is the original sampling value; Mean filtering is to smooth the fluctuation trend, and the formula is: ; In the formula, is the value after median filtering; To adapt to the input of the neural network, adopt Min-Max normalization processing to map each variable to the [0,1] interval and maintain consistency in physical meaning. The formula is: ; In the formula, and are the minimum and maximum values of the variable in the training set or sliding window respectively; Outlier detection and removal, eliminate incorrect and unreasonable data through multiple rules, including: Rule 1: Constrained by physical boundaries, values beyond which are regarded as anomalies; Rule 2: Sliding window difference method. If the difference between a value and the two adjacent values is greater than a set threshold, it is determined as a jump point and removed; Rule 3: Z-Score outlier detection. For the normalized variable, i.e., ; In the formula, is the overall mean, is the overall standard deviation. When the value of Zt is greater than a certain threshold, it is regarded as an outlier; Time series windowing processing: Prepare the historical window sequence for the input of the LSTM model.

[0045] Step S2, Analyze the user's voice command, combine with historical data, construct a personalized temperature control feature vector for fine-tuning the intelligent air conditioner strategy model or optimizing the strategy.

[0046] In the embodiment, the large model processes the user's voice commands (such as "raise the temperature by 2 degrees", "increase the air volume", "too hot"), APP commands or historical interaction data, extracts the preference keywords and semantic intentions of the user in air conditioner usage, and obtains the user intention label, as well as the keyword vector and the embedding of usage preferences; Aggregate the historical air conditioner setting behaviors of the user for multiple days or months, and aggregate a preference model through clustering, collaborative filtering or large model embedding for the preference vector for inference and the user portrait label; Collect the external environmental information of the current user's location and use the context fusion method to enhance the accuracy of inferring the user's air conditioner demand, and obtain the current environment perception vector and multi-modal fusion features.

[0047] The input of the large model is the user command text and context input such as the current air conditioner settings; the output of the large model is the intent such as temperature adjustment, air volume adjustment, air conditioner on / off, air conditioner mode switching, and the slots structured control fields such as target_temp, fan_speed, mode, etc.

[0048] Step S3, Process the time series data through LSTM, predict the temperature change trend in the next 10 minutes, fuse multi-dimensional features through XGBoost to correct the prediction error of LSTM, and calculate the optimal air conditioner setting value in combination with the driving mode and environmental factors.

[0049] The LSTM time series model, the input features are the historical temperature sequence + synchronous external temperature, wind speed, driving mode, etc.; Network structure selection: Input [W x N] → LSTM(128) → Dropout → Dense(64) → Output: [T]; The loss function adopts the mean squared error between the predicted value and the true value as the loss function, and the optimizer uses Adam; Model output: Predicted value of the in-vehicle temperature trend in the next 10 minutes.

[0050] XGBoost model: The correction target is the error in the LSTM output. Due to unmodeled factors such as environmental changes, external features need to be fused for correction.

[0051] Input features: The LSTM prediction result, as the initial prediction input; the current in-vehicle state such as temperature, humidity, wind speed, air-conditioning mode, etc.; the current external environmental features such as weather type, external temperature, light intensity, etc.; driving behavior features such as the current vehicle speed, acceleration, driving mode, etc.; user portrait features such as cold / hot user portraits and adjustment frequencies; Output features: Prediction value error and the final corrected prediction.

[0052] In the embodiment, terminal local adaptive adjustment and cloud model long-term optimization are adopted: Local adaptive adjustment: If the user adjusts the air-conditioning settings multiple times in a short period, the local temperature control parameters are updated, and the prediction logic is adjusted: Trigger condition: Manual adjustment more than 2 times within 30 minutes. Adjustment method: The LSTM prediction weight is increased to enhance the short-term adjustment trend (for example, if the temperature has been raised by 2°C in the past 10 minutes, then +1°C is default in future predictions), and the settings frequently used by the user in the recent 7 days are preferentially recommended (for example, if the setting is often 24°C during night driving, then 24°C is default recommended); Cloud long-term optimization: The user behavior data will be regularly uploaded to the cloud and used to train the personalized model. Trigger condition: The user adjusts to the same temperature multiple times within 3 days in the same environment (for example, adjusts from 22°C to 24°C multiple times). Adjustment method: The reinforcement algorithm is used to update the personalized strategy to make the prediction more in line with the user's habits. The LLM (large model) analyzes the user's voice preferences, such as implicit feedback like "a bit cold", and adjusts the air-conditioning setting strategy.

[0053] Step S4, optimize the fused state features and user behavior feedback by constructing the state-action-reward S-A-R model for air-conditioning control to obtain the air-conditioning policy model and the set of optimal policy models for different user types.

[0054] Step S5, based on the multi-version air-conditioning policy model, the user's interaction feedback, and the results of the simulation test, obtain the optimal policy model and the policy evaluation report for OTA optimization A / B testing.

[0055] Step S6: Transmit the parameters of the trained lightweight policy model and the policy table to the vehicle side through OTA or API.

[0056] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A prediction system for an in-vehicle intelligent air conditioner, characterized in that, Including: The in-vehicle terminal and the cloud server, The in-vehicle terminal is used to construct the input tensor required for model inference through various collected data, extract short-term historical time series features through LSTM, and combine XGBoost with static variables to make the final regression strategy model. Then, it sends the result of the strategy model inference to the vehicle air conditioner controller, judges whether the user accepts the result of the model inference, records the manual adjustment behavior, and feedbacks it to the cloud server; The cloud server is used to collect the external environment information of the current user's location and enhance the accuracy of user air conditioner demand inference by using the context fusion method, so as to obtain the current environmental perception vector and multi-modal fusion features, which are used as part of the state space of reinforcement learning; And construct a state-action-reward S-A-R model for air conditioner control to optimize the fusion state features and user behavior feedback, so as to obtain the air conditioner's strategy model and the optimal strategy model set of different user types; According to the multi-version air conditioner strategy model, the user's interaction feedback, and the results of simulation tests, obtain the optimal strategy model and the strategy evaluation report for OTA optimization A / B testing; And send the parameters and strategy table of the trained lightweight strategy model to the in-vehicle terminal through OTA or API.

2. The prediction system of an in-vehicle intelligent air conditioner according to claim 1, wherein, The in-vehicle terminal specifically includes: The data collection module is used to collect various data and unify the timestamps; The data preprocessing module is used to clean and normalize the data, construct time series through sliding windows, and construct the input tensor 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 make the final regression strategy model, and output the setting suggestions for the air conditioner in real time; The air conditioner control interface module is used to send the result of the strategy model inference to the vehicle air conditioner controller, judge whether the user accepts the result of the model inference, record the manual adjustment behavior, and feedback it to the cloud server and the 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 behavior preferences during use, establish a user portrait, and provide a bias tuning strategy for the strategy model; The data upload and cloud communication module is used to periodically upload local behavior data to the cloud for strategy model optimization, and receive the updated strategy model parameters or strategies from the cloud server.

3. A prediction system for an in-vehicle intelligent air conditioner according to claim 1, characterized in that, The cloud server specifically includes: The user interaction parsing module is used to process user voice commands, APP commands, or historical interaction data through a large model, extract preference keywords and semantic intentions of the user in air conditioner use, and obtain user intention tags, keyword vectors, and embeddings of usage preferences; The long-term preference modeling module is used to aggregate the historical air conditioner setting behaviors of the user for multiple days or months, and aggregate a preference model through clustering, collaborative filtering, or large model embedding for the preference vector and user portrait tags used in inference; The environmental fusion perception module is used to collect the external environmental information of the current location of the user and enhance the accuracy of the user's air-conditioning demand inference by using the context fusion method, so as to obtain the current environmental perception vector and multi-modal fusion features, which are used as part of the reinforcement learning state space; The reinforcement learning training module is used to optimize the fused state features and user behavior feedback by constructing a state-action-reward S-A-R model for air-conditioning control, so as to obtain the policy model of the air-conditioning and the set of optimal policy models for different user types; The policy evaluation and optimization module is used to obtain the optimal policy model and the policy evaluation report for OTA optimization A / B testing according to the multi-version air-conditioning policy model, the user's interaction feedback and the results of simulation tests; The policy distribution interface module is used to distribute the parameters of the trained lightweight policy model and the policy table to the vehicle end through OTA or API; 4. The prediction system of an in-vehicle intelligent air conditioner according to claim 2, characterized in that, In the data collection module, various types of data include the temperature and humidity inside the vehicle, the temperature and humidity outside the vehicle, the vehicle speed, the driving mode from the can bus data, and the weather, air quality, ultraviolet intensity, pollution index, and the current setting values of the air-conditioning, including the set temperature, air volume, and air-conditioning mode, by calling the third-party weather api; 5. The prediction system of an in-vehicle intelligent air conditioner according to claim 3, characterized in that, In the user interaction parsing module, the input is the speech recognition text uploaded by the vehicle end, the APP interaction data, and the user adjustment log, and the output is the user intention label, the keyword vector, and the embedding of the usage preference; In the long-term preference modeling module, the input is the user intention label, the keyword vector, the embedding of the usage preference, the historical behavior data, and the daily operation data of the vehicle, and the output is the user's long-term air-conditioning control preference model, the preference vector for reasoning, and the user portrait label; In the environmental fusion perception module, the external environmental information includes geographical location, real-time weather, time period, and traffic status; The input is the third-party weather API, geographical location information, traffic conditions, and time period holiday labels, and the output is the current environmental perception vector and multi-modal fusion features, which are used as part of the reinforcement learning state space; 6. The prediction system of an in-vehicle intelligent air conditioner according to claim 3, characterized in that, In the reinforcement learning training module, the state includes user preference, environmental state, temperature and humidity inside the vehicle, 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 the user's satisfaction, including the setting of energy consumption and comfort index; The user behavior feedback includes manual adjustment behavior, satisfaction score, and energy consumption; 7. The prediction system of an in-vehicle intelligent air conditioner according to claim 3, characterized in that, In the policy distribution interface module, the input is the policy model file, the user identification vehicle identification, and the distribution plan, and the output is the policy parameter configuration, the synchronization status feedback, and the record of success / failure; 8. A prediction method for an in-vehicle intelligent air conditioner, characterized in that, Including: Step S1, collect data information and perform preprocessing operations on the data; Step S2, analyze the user's voice command, combine historical data, and construct a personalized temperature control feature vector for fine-tuning the intelligent air-conditioning policy model or optimizing the policy; Step S3, process the time series data through LSTM to predict the temperature change trend in the next 10 minutes. Then, fuse multi-dimensional features through XGBoost to correct the prediction error of LSTM, and calculate the optimal air conditioner setting value in combination with the driving mode and environmental factors; Step S4, optimize the fused state features and user behavior feedback by constructing a state-action-reward S-A-R model for air conditioner control to obtain the air conditioner's policy model and the optimal policy model set for different user types; Step S5, based on the multi-version air conditioner policy model, the user's interaction feedback, and the results of simulation tests, obtain the optimal policy model and the policy evaluation report for OTA optimization A / B testing; Step S6, send the parameters of the trained lightweight policy model and the policy table to the in-vehicle terminal through OTA or API.

9. A prediction method for an in-vehicle intelligent air conditioner according to claim 8, characterized in that, In the above-mentioned step S1, the data information includes vehicle data, environmental data, and user data; among them, the vehicle data includes the in-vehicle temperature, in-vehicle humidity, air conditioner setting value, vehicle speed, engine speed, battery power, and driving mode; the environmental data includes the outdoor temperature, humidity, wind speed, precipitation, city where the vehicle is located, altitude, weather description, ultraviolet intensity, pollution index, time, and season; the user data includes voice interaction logs, previous manual air conditioner setting records, driving habits, whether the window is opened, and whether AUTO is turned on; Perform preprocessing operations on the vehicle data, including: Adopt median filtering and mean filtering methods for denoising. Median filtering is to suppress isolated outliers, and the formula is: ； where k is the half-width of the window, is the original sampled value; Mean filtering is to smooth the fluctuation trend, and the formula is: ； Wherein, is the value after median filtering; For adapting to the input of the neural network, perform Min-Max normalization processing to map each variable to the [0,1] interval and maintain the consistency in physical meaning. The formula is: ； wherein, and are the minimum and maximum values of the variable in the training set or the sliding window, respectively; Detect and remove outliers, and eliminate incorrect and unreasonable data through multiple rules, including: Rule 1: Constrain through physical boundaries, and values beyond the boundaries are regarded as outliers; Rule 2: Sliding window difference method. If the difference between a value and the two adjacent values is greater than a set threshold, it is determined as a jump point and removed; Rule 3: Z-Score outlier detection. For the normalized variables, that is: ； Wherein, is the overall mean, is the overall standard deviation. When the value of Zt is greater than a certain threshold, it is regarded as an outlier; Time series windowing processing: Prepare the historical window sequence for the input of the LSTM model.

10. A prediction method for an in-vehicle intelligent air conditioner according to claim 8, characterized in that, In the above-mentioned step S2, process the user's voice commands, APP commands, or historical interaction data through a large model to extract the preference keywords and semantic intentions of the user in air conditioner usage, and obtain the user intention label, keyword vector, and embedding of usage preferences; Aggregate the historical air conditioner setting behaviors of the user over multiple days or months, and aggregate the preference model for inference, preference vector, and user portrait label through clustering, collaborative filtering, or large model embedding; Collect the external environmental information of the current location of the user and use the context fusion method to enhance the accuracy of inferring the user's air conditioner demand, and obtain the current environmental perception vector and multi-modal fusion features.

Images