Cabin air conditioning prediction method based on channel-independent bidirectional interaction mamba

Abstract

The application relates to a cabin air conditioner prediction method based on channel-independent bidirectional interaction Mamba, and belongs to the technical field of intelligent power utilization. The method comprises the following steps: constructing a CIBG-Mamba-interaction circulation mechanism model, including a CIBG-Mamba and an interactive recursive mechanism; the CIBG-Mamba is obtained by integrating the Mamba model into a channel-independent recursive unit, and local modes and changes of different-dimensional time sequence data are captured; the interactive recursive mechanism is obtained by using a state transition function of a specific channel, information interaction between channels is realized, and complex dependency relationships between different dimensions are captured. The application can improve temperature prediction accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent power utilization, and relates to a cockpit air conditioner prediction method based on channel-independent bidirectional interaction Mamba. BACKGROUND

[0002] The cockpit air conditioner system is a key component of new energy vehicles and plays an important role in ensuring passenger comfort and optimizing energy utilization. These systems face multidimensional challenges such as temperature fluctuations, humidity changes, and passenger preferences. If they cannot be effectively predicted and controlled, it may lead to a decline in passenger experience, increased battery energy consumption, and even affect driving safety.

[0003] Accurate prediction of cockpit air conditioner temperature settings has a profound impact on the overall performance and user experience of new energy vehicles. Through accurate prediction, intelligent temperature regulation can be achieved, energy efficiency can be improved, battery life can be extended, passenger comfort can be enhanced, and decision-making can be provided for vehicle management systems, thereby promoting the sustainable development of the automotive industry.

[0004] Traditional prediction models have limitations in handling multi-dimensional time series data of cockpit air conditioners. Traditional cockpit air conditioner systems have weak capabilities in handling multi-dimensional nonlinear data and long-term time series dependencies, low prediction accuracy, poor adaptability to temperature mutations, fixed control strategies, and lack of dynamic adaptability, making it difficult to achieve personalized adjustment and optimal energy consumption balance. Therefore, a new prediction model is needed to improve the prediction accuracy of cockpit air conditioner temperature. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a cockpit air conditioner prediction method based on channel-independent bidirectional interaction Mamba, which effectively solves the shortcomings of existing technologies in handling complex multi-dimensional time series data and significantly improves prediction accuracy and model performance.

[0006] To achieve the above purpose, the present application provides the following technical scheme:

[0007] A cockpit air conditioner prediction method based on channel-independent bidirectional interaction Mamba, specifically comprising the following steps:

[0008] S1: Data acquisition and preprocessing: first, acquire real vehicle data from multiple vehicles, including environmental temperature, dew point, vehicle speed, and indoor temperature feature data; then, clean the acquired raw data;

[0009] S2: Construct a CIBG-Mamba-interactive recurrent mechanism model, including CIBG-Mamba and interactive recurrent mechanism (IRM); the CIBG-Mamba represents channel-independent bidirectional gated Mamba, which captures the local patterns and changes of different dimensional time series data by integrating the Mamba model into the channel-independent recurrent unit; the interactive recurrent mechanism (IRM) realizes the information interaction across channels through the state transition function of a specific channel, and captures the complex dependency relationship between different dimensions;

[0010] S3: Input the cleaned data into the CIBG-Mamba-interactive recurrent mechanism model to predict the user driving setting temperature at time T.

[0011] Further, in step S1, the obtained raw data is cleaned, and the specific steps include: a. dividing a large data set into blocks for subsequent processing; b. processing each data block, and filling the null data appearing according to the time stamp with the data in the upper and lower time windows; c. splicing and arranging the filled data blocks to obtain a complete data set after cleaning.

[0012] Further, in step S2, the structure of the CIBG-Mamba is to integrate the recurrent unit and the Mamba in a channel-independent and bidirectional manner, specifically including:

[0013] First, channel-independent embedding is used to extract channel-independent features of multi-dimensional time series data while preserving the unique properties of each time series variable; the process starts with an input tensor U in , where B is the batch size, L is the sequence length, and D is the number of channels; channel-independent embedding independently extracts local time features of each channel by applying a slicing transformation through the following formula:

[0014] U E = Linear(Unfold(U in ))

[0015] where Unfold(·) represents performing a patch operation to transform the input tensor U in into U p ∈(B,S,D,P), where each channel is subjected to S patches, and each patch has a size of P; the function further applies a channel-specific learnable linear transformation to output; then, in order to obtain the independent mode of each channel, U E is converted into U T ;

[0016] Second, the CIBG-Mamba dynamically processes U T, update gate Mamba, reset gate Mamba and candidate activation Mamba are updated to extract hidden features; this helps to better preserve and utilize information over a longer period of time, improving the model's ability to learn fine temporal representations and handle long-term dependencies. In particular, the present application positively and negatively input U T into the CIBG-Mamba to conduct comprehensive feature learning; the CIBG-Mamba unit operates on each input in both directions at time step T, as follows:

[0017] Mamba is a selective state space model with a dynamic selection mechanism; this innovative feature enables Mamba to adjust parameters according to input data, greatly improving the ability to model complex time series data and provide accurate predictions. In traditional state space models (SSM), the system's data name is represented by a first-order differential:

[0018] h' t =Ah t +Bx t

[0019] y t =Ch t

[0020] Where x t is the input sequence at time t, h t is the hidden state, h' t is the first-order differential representation of the hidden state, y t is the output; matrices A, B and C are learnable parameters, and the model maintains a linear time-varying structure; to handle continuous data, the model uses time step Δ to discretize the sequence, as follows:

[0021]

[0022] y t =Ch t

[0023] Where, I represents the identity matrix; this discretization method allows efficient computation using linear recurrence methods. Mamba further extends this model by taking A, B and C as input functions, as follows:

[0024] h t =A(x t )h t-1 +B(x t )x t

[0025] y t =C(x t )ht

[0026] where A(·), B(·) and C(·) are state transition functions of input x t , and linear projection is the operation tool used in this case.

[0027] Further, in step S2, the network structure of Mamba is:

[0028] (1) Update gate Mamba: update gate U A specialized Mamba module is used to determine how much of the previous hidden state H t-1 should be retained in the new state in the following equation:

[0029]

[0030] where σ(·) is the sigmoid function, Mamba z (·) is the update gate Mamba operation, is the input, is the update gate, and arrows → and ← represent the forward and backward processes, respectively;

[0031] (2) Reset gate Mamba: reset gate R R is calculated by its specific Mamba module to adjust the influence of the past state on the candidate state, as shown in the following equation:

[0032]

[0033] where Mamba r is the reset gate Mamba operation;

[0034] (3) Candidate activation Mamba: This component combines the current input with the adjusted previous state in the following equation:

[0035]

[0036] where, is the candidate activation, tanh(·) represents the hyperbolic tangent function, and Mamba h is the candidate activation Mamba operation, and ⊙ represents element-wise multiplication;

[0037] (4) Hidden state update: The final state is the composition of the old state and the new candidate activation state in the forward and backward directions, where the hidden state update in the forward and backward directions is shown in the following equation:

[0038]

[0039] where H tis the hidden layer output.

[0040] Further, in step S2, the interactive recurrent mechanism includes forward interaction and backward interaction, by adding forward and backward hidden layer outputs and form a recursive interaction of the overall information flow, as follows:

[0041] Forward interaction: combine the forward hidden state of the previous time step with the current backward hidden state H t to obtain an interactive forward hidden state, as shown in the following formula:

[0042]

[0043] where Interactive(·,·) represents linear projection of two inputs to overfit the dimension of the data stream;

[0044] Backward interaction: combine the backward hidden state of the next time step with the current forward hidden state to create an interactive backward hidden state in the following formula:

[0045]

[0046] The above interaction state is used to calculate the update gate reset gate U R and candidate activation to enhance the time processing capability of the model; after the interactive loop mechanism, we can obtain the information of the Mamba enhanced independent recurrent unit in the bidirectional interaction channel, and the forward and backward hidden states imply rich temporal features, so we can obtain the final output H out of the interactive loop mechanism unit after processing by the following formula:

[0047]

[0048] where drop represents dropout, convert represents conversion function, and norm represents layer normalization;

[0049] The interactive loop mechanism combines bidirectional information to reconstruct time information and create a time sequence; then, the sequence is mapped and converted to the final output U out , as shown in the following formula:

[0050] U out = Projection(H out )

[0051] where Projection represents linear mapping.

[0052] The beneficial effects of the present application are that the present application constructs a channel independent bidirectional gated Mamba model (CIBG-Mamba-interactive loop mechanism) of interactive recursive mechanism, which captures global patterns and local changes through CIBG-Mamba, and integrates cross-channel and multi-scale features using interactive recursive mechanism, so as to improve the temperature prediction accuracy. The model of the present application can effectively solve the shortcomings of the prior art in processing complex multi-dimensional time series data, and significantly improve the prediction accuracy and model performance. This is embodied in the following aspects:

[0053] (1) Channel independent bidirectional gated Mamba (CIBG-Mamba) structure: In view of the limitations of Mamba model in the application of multi-dimensional time series data, the present application proposes a CIBG-Mamba structure. The structure integrates the Mamba model into a channel independent recursive unit, and each unit is responsible for processing a specific dimensional time series data. Through this design, the model can better learn and represent fine-grained time sequence features, and enhance the ability to capture complex time patterns.

[0054] (2) Interactive recursive mechanism (interactive loop mechanism): The present application designs an interactive recursive mechanism, which skillfully integrates the historical hidden state of forward information and the immediate hidden state of backward information (and vice versa) through a specific channel state transition function. This design enables the model to realize cross-channel information interaction and effectively capture complex dependency relationships between different dimensions.

[0055] (3) CIBG-Mamba-interactive loop mechanism model: Based on the CIBG-Mamba structure and the interactive recursive mechanism, the present application constructs a complete CIBG-Mamba-interactive loop mechanism model. The model realizes high-precision prediction of multi-dimensional time series data by integrating the above two key components.

[0056] Other advantages, objects, and features of the present application will be set forth in part in the following specification, and in part will become apparent to those skilled in the art from the examination of the following specification, or can be learned from practice of the present application. The objects and other advantages of the present application can be realized and attained by the below description. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to make the purposes, technical solutions and advantages of the present application clearer, the preferred detailed description of the present application will be combined with the drawings as follows, wherein:

[0058] Figure 1 The structure block diagram of the CIBG-Mamba-interactive loop mechanism model of the present application is shown in the figure;

[0059] Figure 2 A schematic diagram for classifying model input data;

[0060] Figure 3 A model architecture diagram for Mamba;

[0061] Figure 4 All predicted contents for a vehicle;

[0062] Figure 5 A comparison test effect diagram of the model of the present application and various existing models. DETAILED DESCRIPTION

[0063] The implementation of the present application will be described in detail through specific examples. Those skilled in the art can easily understand other advantages and effects of the present application from the disclosure. The present application can also be implemented or applied in different specific embodiments, and various modifications or changes can be made based on different views and applications without departing from the spirit of the present application. It should be noted that the diagrams provided in the following examples only illustrate the basic concept of the present application in a schematic manner, and the following examples and features in the examples can be combined with each other without conflict.

[0064] Referring to Figures 1-5 , the present application provides a channel-independent bidirectional interaction Mamba cabin air conditioner prediction method, and a CIBG-Mamba-interactive cycle mechanism model (as shown in Figure 1 ) is constructed. The model fully utilizes the capturing ability of CIBG-Mamba for global patterns and local changes in multi-dimensional time series data, and realizes effective integration of cross-channel and multi-scale features through an interactive recursive mechanism, significantly improving the utilization rate and prediction accuracy of multi-dimensional time series data. The model performs well in long sequence data modeling and fine-grained time representation, effectively solving the key limitations of existing Mamba models in multi-dimensional time series prediction.

[0065] The channel-independent bidirectional interaction Mamba cabin air conditioner prediction method provided by the present application specifically includes the following steps:

[0066] S1: Data acquisition and preprocessing: first, real vehicle data is acquired from multiple vehicles, mainly including environmental temperature, dew point, vehicle speed, indoor temperature and other feature data. The acquired raw data is cleaned, and the specific steps include: a. block division is performed on a large number of data sets to facilitate subsequent processing; b. each data block is processed, and the null data appearing according to the time stamp is filled in the upper and lower time window data; c. the filled data blocks are spliced and arranged to obtain a complete data set after cleaning.

[0067] S2: Data input and feature construction: Input the cleaned data into the CIBG-Mamba-interaction cycle mechanism model. Figure 2 As shown in the figure, input feature data can be divided into three categories: a. External features (blue boxes): such as ambient temperature and dew point; b. Internal features (cyan boxes): such as internal temperature; c. User-set features (red boxes): such as the user-set target temperature. The constructed input data format is (B, L, D), where B represents the batch size, L represents the input sequence length, and D represents the feature dimension.

[0068] S3: Model prediction: The CIBG-Mamba-interactive recurrent mechanism model of the present invention mainly consists of two key components: a. CIBG-Mamba: By integrating the Mamba model into the channel-independent recurrent unit, it captures the local patterns and changes of time series data in different dimensions; b. Interactive recurrent mechanism (IRM): Through the state transfer function of a specific channel, it realizes cross-channel information interaction and captures the complex dependencies between different dimensions. The specific structure of the model is as follows Figure 1 The prediction procedure is shown in Table 1.

[0069] Table 1 Prediction algorithm process

[0070]

[0071]

[0072] S4: Prediction result output: Through the CIBG-Mamba-interaction loop mechanism model, the user's driving setting temperature at the next time T can be predicted. The output format is (B, T, D), where B represents the batch size, T represents the number of predicted future time steps, and D represents the predicted feature dimension. This algorithm mainly predicts the temperature set by the user.

[0073] 1) CIBG-Mamba: Channel Independent Bidirectional Gating Mamba

[0074] CIBG-Mamba enhances multidimensional time series data modeling by integrating recurrent units with Mamba in a channel-independent and bidirectional manner. This integration improves information flow control by using a method designed specifically for multidimensional time series prediction that is more efficient than traditional gating mechanisms.

[0075] First, channel-independent embedding is used to extract channel-independent features of multidimensional time series data while preserving the unique attributes of each time series variable. This process starts with an input tensor U of shape (B, L, D) in Starting from, where B is the batch size, L is the sequence length, and D is the number of channels. CIE applies the slice-wise transformation to extract local temporal features for each channel independently by the following formula:

[0076] U E =Linear(Unfold(U in ))

[0077] where Unfold(·) denotes performing patch operation to transform the input tensor U in into U p ∈(B,S,D,P), where each channel is passed through S patches individually, and each patch has a size of P. The function further applies a channel-specific learnable linear transformation to output. Then, in order to obtain the independent mode for each channel, the invention will transform into U T .

[0078] Secondly, CIBG-Mamba is proposed for dynamic processing of U T , which extracts hidden features by calculating update-gated Mamba, reset-gated Mamba and candidate-activated Mamba. This helps to better preserve and utilize information over a longer period of time, improving the model's ability to learn fine temporal representations and handle long-term dependencies. In particular, we forward and backward input U T into CIBG-Mamba for comprehensive feature learning. The CIBG-Mamba unit operates on each input in both directions at time step T, as follows:

[0079] Mamba is a selective state space model with a dynamic selection mechanism. This innovative feature enables Mamba to adjust parameters according to input data, greatly improving the ability to model complex time series data and provide accurate predictions. In traditional state space models (SSM), the system's data name is represented by a first-order differential:

[0080] h' t =Ah t +Bx t

[0081] y t =Ch t

[0082] where x t is the input sequence at time t, h t is the hidden state, h' t is the first-order differential representation of the hidden state, y t is the output; matrices A, B and C are learnable parameters, and the model maintains a linear time-varying structure; to handle continuous data, the model uses time step Δ to discretize the sequence, as follows:

[0083]

[0084] y t = Ch t

[0085] where, This discretization method allows efficient computation using linear recurrence methods. Mamba goes further on this model by taking A, B and C as input functions as follows:

[0086] h t = A(x t )h t-1 + B(x t )x t

[0087] y t = C(x t )h t

[0088] where A(·), B(·) and C(·) are state transition functions of input x t , linear projections are the operational tools used in this case. The network structure of Mamba is shown in Figure 3 .

[0089] (1) Update gate Mamba: The update gate U is computed using a dedicated Mamba module to determine how much of the previous hidden state H t-1 should be preserved in the new state y :

[0090]

[0091]

[0092] where σ(·) is the sigmoid function, arrows → and ← represent the forward and backward processes, respectively;

[0093] (2) Reset gate Mamba: The reset gate U R is computed through its specific Mamba module to regulate the influence of past states on the candidate state y :

[0094]

[0095] (3) Candidate activation Mamba: This component merges the current input x with the regulated previous state y :

[0096]

[0097] where tanh(·) denotes the hyperbolic tangent function, and ⊙ represents element-wise multiplication;

[0098] (4) Hidden state update: The final state is a composition of the old state and the new candidate activation state in the forward and backward directions, where the hidden state update in the forward and backward directions is shown in the following formula:

[0099]

[0100] 2) Interactive recurrent mechanism

[0101] The interactive recurrent mechanism includes forward interaction and backward interaction, which adds forward and backward hidden layer outputs and to form a recursive interaction of the overall information flow, as follows:

[0102] Forward interaction: combine the forward hidden state of the previous time step with the current backward hidden state to obtain an interactive forward hidden state, as shown in the following formula:

[0103]

[0104] where Interactive(·,·) represents linear projection of two inputs to overfit the dimension of the data stream;

[0105] Backward interaction: merge the backward hidden state of the next time step with the current forward hidden state to create an interactive backward hidden state in the following formula:

[0106]

[0107] The above interactive state is used to calculate the update gate reset gate U R and candidate activation to enhance the time processing capability of the model; after the interactive recurrent mechanism, the present application can obtain the information of the Mamba enhanced independent recurrent unit in the bidirectional interaction channel, and the forward and backward hidden states thereof imply rich temporal features, so that the present application can obtain the final output H out of the interactive recurrent mechanism unit after processing according to the following formula.

[0108]

[0109] where drop represents dropout, convert represents a conversion function, and norm represents layer normalization.

[0110] The interactive recurrent mechanism combines bidirectional information to reconstruct time information and create a time sequence; then, the sequence is mapped and converted into the final output U outAs shown in the following formula:

[0111] U out = Projection(H out )

[0112] Experimental results analysis:

[0113] For the final results of temperature setting, the experiment was conducted for 30 times, and the final MSE and MAE mean value and standard deviation results were compared with the existing most advanced algorithm, as shown in Table 2.

[0114] Table 2 Comparison of performance of different models

[0115]

[0116] In Table 2, the italicized indicates the best effect, and the black bold indicates the second best effect. It can be seen that the model proposed by us has excellent performance in prediction accuracy and prediction accuracy, among which the prediction accuracy is the highest in all models, the mean value of MSE is 0.904, and the mean value of MAE is 0.641, the prediction accuracy ranks second in all models, the prediction standard deviation of MSE is 0.057, and the prediction standard deviation of MAE is 0.038.

[0117] The experiment visualizes all the predicted contents of a car, as shown in Figure 4 Compared with other models, the predicted value of the model of the application is closer to the true value, indicating that its performance is better than that of other models. In addition, the application also magnifies the fluctuation points in the graph, which are caused by the rapid changes of the user and the environment. From the magnified subgraph, it can be seen that the prediction curve of the model of the application has a large slope, which can produce faster and more accurate adaptive results. This phenomenon further proves that the model proposed by us has fine time representation ability.

[0118] In order to prove the more fine time representation ability of the model proposed by us, this time we carried out comparative test on more models, as shown in Figure 5 It can be seen that at the last tens of data points of time fluctuation, the algorithm proposed by the application obtains the most accurate prediction result, and obtains the fastest convergence ability.

[0119] Therefore, it can be seen that the CIBG-Mamba-interactive cycle mechanism model has the following obvious advantages:

[0120] (1) High prediction accuracy: by skillfully combining the CIBG-Mamba structure with the interactive recursive mechanism, the model can fully utilize the rich information contained in the multi-dimensional time series data, model and predict from the global and local angles, and significantly improve the prediction accuracy of the model.

[0121] (2) Strong feature capturing ability: The CIBG-Mamba structure endows the model with the ability to capture fine-grained time series features, and each channel-independent recurrent unit can focus on processing time series data in a specific dimension. At the same time, the interactive recurrent mechanism realizes the information interaction across dimensions, enabling the model to capture complex dependency relationships between different dimensions.

[0122] (3) Ability to process complex data: The CIBG-Mamba-interactive recurrent mechanism model is designed to process complex time series data with long sequences and high dimensions. By introducing channel-independent recurrent units and an interactive recurrent mechanism, the model can effectively deal with the high-dimensional characteristics and long-term dependence of data, and has broad application prospects in fields such as climate prediction and financial analysis that involve complex time series.

[0123] In summary, the CIBG-Mamba-interactive recurrent mechanism model of the present application has significant innovations and advantages in terms of prediction accuracy, feature capturing ability, and processing complex data, providing a new solution and method for multi-dimensional time series prediction tasks, and has important theoretical significance and practical value.

[0124] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and all should be included in the scope of the claims of the present application.

Claims

1. A cabin air conditioning prediction method based on channel-independent bidirectional interaction Mamba, characterized in that: The method specifically comprises the following steps: S1: Data acquisition and preprocessing: First, real vehicle data is obtained from multiple vehicles, including ambient temperature, dew point, vehicle speed, and interior temperature; then the acquired raw data is cleaned; S2: Construct a CIBG-Mamba-interactive recurrent mechanism model, including CIBG-Mamba and an interactive recursive mechanism. CIBG-Mamba stands for channel-independent bidirectional gated Mamba, which captures local patterns and changes in time series data of different dimensions by integrating the Mamba model into channel-independent recursive units. The interactive recursive mechanism uses the state transfer function of a specific channel to achieve cross-channel information interaction and capture the complex dependencies between different dimensions. S3: Input the cleaned data into the CIBG-Mamba-interaction loop mechanism model to predict the user's driving setting temperature at the next time T; In step S2, the structure of the CIBG-Mamba is to integrate the cyclic unit with Mamba in a channel-independent and bidirectional manner, specifically including: First, channel-independent embedding is used to extract channel-independent features of multidimensional time series data while preserving the unique attributes of each time series variable; this process starts from a shape of The input tensor Start, where B is the batch size, L is the sequence length, D is the number of channels; the local temporal features of each channel are extracted independently by applying the slice-wise transformation: in, Indicates the execution of a patch operation, which takes the input tensor Transformed into , where each channel is passed through a separate S patch, each of size P; in order to obtain an independent pattern for each channel, Convert to ; Secondly, CIBG-Mamba dynamic processing is used , extract hidden features by calculating update gate Mamba, reset gate Mamba and candidate activation Mamba; forward and reverse input into CIBG-Mamba for comprehensive feature learning; the CIBG-Mamba unit operates on each input in two directions at time steps T as follows: In the traditional state-space model, the data name of the system is expressed as a first-order differential: in, is the input sequence at time t, is hidden state, is the first-order differential representation of the hidden state, is the output; matrices A, B, and C are learnable parameters, and the model maintains a linear time-varying structure; in order to process continuous data, the model uses a time step Δ to discretize the sequence, as shown in the following formula: in, , , Represents the identity matrix; Mamba takes this model a step further and uses A, B, and C as input functions, as shown below: in, 、 and is input The state transfer function, linear projection is the operation tool used in this case; In step S2, the network structure of Mamba is: (1) Update gate Mamba: Update gate Use a specialized Mamba module to determine how much of the previously hidden state should be retained in the new state using the following formula : in, is the sigmoid function, It is the Mamba operation of the update door, is the input, is the update gate, and the arrows → and ← represent the forward and backward processes respectively; (2) Reset gate Mamba: reset gate Calculated through its specific Mamba module to adjust the influence of past states on candidate states, as shown in the following formula: in, It is the Mamba operation that resets the door; (3) Candidate Activation Mamba: Merge the current input with the previous state adjusted as follows: in, is a candidate activation, represents the hyperbolic tangent function, is the Mamba operation of the candidate activation gate, Represents element-wise multiplication; (4) Hidden state update: The final state is a composite of the old state and the new candidate activation state in the forward and reverse directions, where the hidden state updates in the forward and reverse directions are shown in the following formulas: in, is the hidden layer output; In step S2, the interactive recursive mechanism includes forward interaction and backward interaction, by adding forward and backward hidden layer outputs and , forming a recursive interaction of the overall information flow, as follows: Forward interaction: forward hidden state of the previous time step With the current backward hidden state Combined, we get an interactive forward hidden state, as shown in the following formula: in, Indicates linear projection of the two inputs to fit the dimension of the data stream; Reverse interaction: reverse the hidden state of the next time step With the current forward hidden state Merge to create the interactive inverse hidden state in the following formula: The above interaction state is used to calculate the update gates in both directions , Reset Gate and candidate activation , thereby enhancing the model's time processing capability; after the interactive loop mechanism, the information of the Mamba enhanced independent loop unit in the bidirectional interactive channel is obtained, and its forward and backward hidden states imply rich temporal features. Therefore, the final output of the interactive loop mechanism unit after processing by the following formula is obtained ; in, represents dropout, convert represents the conversion function, norm Representation layer normalization; The interactive loop mechanism combines bidirectional information to reconstruct temporal information and create a time series; this series is then mapped and transformed into the final output , as shown in the following formula: in, Represents a linear map.

2. The cabin air conditioning prediction method based on channel-independent bidirectional interactive Mamba according to claim 1 is characterized in that: In step S1, the acquired raw data is cleaned. The specific steps include: a. dividing a large number of data sets into blocks; b. processing each data block and filling the upper and lower time window data of the null value data according to the timestamp; c. splicing and arranging the filled data blocks to obtain a complete cleaned data set.

Images