Air conditioner expansion valve opening degree prediction method based on variational mode decomposition and deep learning

Abstract

The application discloses an air conditioner expansion valve opening degree prediction method based on variational mode decomposition and deep learning, and belongs to the technical field of air conditioning and intelligent control. The method is used for an air conditioner system provided with a condensing heat recovery branch, collects electronic expansion valve opening degrees and system operation parameters, and performs variational mode decomposition on the electronic expansion valve opening degree sequence to reduce the non-stationarity of the original signal; the decomposed mode components and the operation parameters are input into a prediction model composed of a gating cycle unit and an attention mechanism to obtain multi-step prediction results of the electronic expansion valve opening degree. The application can improve the prediction accuracy and stability of the electronic expansion valve opening degree, and can provide prediction basis for feedforward compensation, optimization control and model predictive control of the condensing heat recovery air conditioner system.

Description

Technical Field

[0001] This invention relates to the field of air conditioning and intelligent predictive control technology, specifically to a method for predicting the opening degree of an air conditioning expansion valve based on variational mode decomposition and deep learning. Background Technology

[0002] Condensing heat recovery air conditioning systems add a reheat branch to the traditional refrigeration cycle, utilizing the condensation heat from the compressor exhaust to reheat the air after it has been cooled and dehumidified by the evaporator. This replaces conventional electric heating devices, offering advantages such as reduced energy consumption and improved energy efficiency. These systems are widely applicable to environments requiring high precision temperature and humidity control, such as data centers, precision manufacturing facilities, and experimental environments.

[0003] In condensing heat recovery air conditioning systems, the electronic expansion valve is responsible for the dynamic distribution of refrigerant flow into the condensing heat recovery branch. Its opening changes directly affect the heat exchange capacity of the reheater and the accuracy of the outlet air temperature control. Existing electronic expansion valve opening control methods mostly employ deviation-based feedback control strategies, adjusting the valve opening according to the current temperature deviation. While this can meet certain steady-state regulation requirements, it struggles to reflect system trends in a timely manner under operating conditions with frequent changes and significant load disturbances.

[0004] With the widespread application of deep learning technology in time series forecasting, models such as Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, and Gated Recurrent Units (GRUs) have been used in fields such as energy load forecasting and industrial process modeling. Meanwhile, Variational Mode Decomposition (VMD) provides more structured input to deep learning models by decomposing complex non-stationary signals into multiple modal components with different frequency characteristics, helping to improve prediction accuracy and stability. However, research on multivariate, multi-step prediction of the opening of electronic expansion valves in condensing heat recovery air conditioning systems remains relatively limited.

[0005] Chinese patent application CN103925668A discloses a DC inverter constant temperature and humidity unit with condensing heat recovery and a heat and humidity control method. The patented technical solution uses traditional PID feedback control to adjust the opening of the auxiliary electronic expansion valve. The solution only relies on temperature deviation for lag adjustment, does not have the ability to predict the operating trend, and has insufficient control accuracy under multi-variable coupling conditions.

[0006] Therefore, it is necessary to propose an electronic expansion valve opening prediction method to solve the multi-step prediction problem of non-stationary, multivariable coupled time series data in condensing heat recovery air conditioning systems, and to provide a reliable basis for feedforward control, optimization control and model predictive control. Summary of the Invention

[0007] To address the problem of predicting the opening degree of electronic expansion valves in condensing heat recovery air conditioning systems in existing technologies, this invention provides a method for predicting the opening degree of expansion valves in condensing heat recovery air conditioning systems based on variational mode decomposition and deep learning.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting the opening degree of an electronic expansion valve in a condensing heat recovery air conditioning system is disclosed. This method is applied to an air conditioning system with a condensing heat recovery branch controlled by an electronic expansion valve. The air conditioning system includes a compressor, a condenser, a reheater, an evaporator, an electronic expansion valve, a thermostatic expansion valve, a fan, and sensors electrically connected to a controller. The compressor outlet is connected to both the reheater and the condenser. The reheater and condenser are connected to the evaporator via the electronic expansion valve and the thermostatic expansion valve, respectively. The evaporator is connected to the compressor. The method for predicting the opening degree of the electronic expansion valve includes the following steps: Step 1: Collect multivariate time series data during the actual operation of the air conditioning system. The data shall include at least the air inlet temperature, air inlet humidity, air outlet temperature, air outlet humidity, high pressure, low pressure, and electronic expansion valve opening of the air conditioning system.

[0009] Step Two: Preprocess the collected data, including outlier handling, data time alignment, and input feature filtering, to ensure the stability of the subsequent training process. Outlier removal employs a boxplot method based on quartiles and interquartile ranges. Let a certain sequence of operating parameters be... Sort the sequence in ascending order to obtain the ordered sequence. ,in, In the ordered sequence, the first quartile Q 1 represents the value corresponding to the lower quartile position, the third quartile. Q 3 represents the value corresponding to the upper quartile position. Its calculation method can be expressed as follows: (1) (2) Then the interquartile range (IQR) is: (3) Lower bound for outlier detection L and upper limit U They are respectively: (4) (5) The input feature selection method uses the Pearson correlation coefficient method, retaining operating parameters whose absolute value of the correlation coefficient with the electronic expansion valve opening is greater than a preset threshold as model input features. The Pearson correlation coefficient calculation method is shown below: (6) In the formula, n The number of data points; x i , y i For the first i The actual value of each data point; , This is the average of the actual values.

[0010] Step 3: Perform variational mode decomposition on the electronic expansion valve opening time series to obtain multiple modal components with different center frequencies and residual terms. The electronic expansion valve opening signal typically contains both low-frequency trend changes and high-frequency fluctuations during system operation. Directly modeling the original signal is easily affected by non-stationarity. Variational mode decomposition decomposes the original opening sequence into multiple narrowband modal components, thereby reducing signal complexity and enhancing the subsequent model's ability to learn different frequency features. The variational problem employs the following constraints: (7) In the formula, f ( t ) represents the original time series signal. u k ( t ) indicates the first k One modal function, ω k This represents the corresponding center frequency. * indicates a convolution operation. δ ( t ) is the Dirac function, j The imaginary unit is used; the constraints ensure that the sum of all modal functions can reconstruct the original signal.

[0011] Step 4: Normalize the decomposed modal components and other preprocessed operating parameters, and divide them into training, validation, and test sets. Since variables such as electronic expansion valve opening, temperature, pressure, and humidity have different numerical ranges and dimensions, to avoid any one type of variable dominating the model training process due to excessively large numerical scales, the min-max normalization method is used to perform dimensionless processing on each variable. The calculation method is as follows: (8) In the formula For the normalized data, x, x min and x max These represent the data to be normalized, the maximum value of the data to be normalized, and the minimum value of the data to be normalized, respectively.

[0012] Step 5: Construct multivariate historical input samples and future predicted output samples based on a sliding time window. Specifically, the operating parameters and modal components at multiple consecutive historical time points are used as inputs, and the modal values ​​corresponding to the opening of the electronic expansion valve at multiple future time points are used as outputs, thereby establishing a multivariate time series multi-step prediction task. The expressions for the input and output samples are shown below: No. i Input Samples X i It can be represented as: (9) In the formula, M For the number of input variables, T For the length of the history window, H To predict the step size.

[0013] No. i Output samples Y i for: (10) Step Six: Input the input samples into a deep learning prediction model composed of a gated recurrent unit network (GRN) and an attention mechanism. The GRN extracts the temporal dependencies from the input samples to alleviate the gradient vanishing problem that traditional recurrent neural networks often encounter during long sequence modeling. The attention mechanism weights the importance of different historical time steps, thereby strengthening the representation of historical state information that contributes significantly to the prediction results. Finally, the prediction results for each modal component at multiple future time points are obtained.

[0014] In the t The calculation process of updating the gate, resetting the gate, and calculating the hidden state in the gated loop unit at each time step can be represented as: (11) (12) (13) (14) In the formula, z t and r t This indicates the update gate and the reset gate; σ Use the Sigmoid activation function;W z , W r , W h This is the weight matrix; h t－1 and h t These represent the hidden states at the previous and current time points, respectively. x t This represents the input feature vector at the current moment; This indicates the candidate is in a hidden state.

[0015] The computational process of the attention mechanism can be represented as follows: (15) (16) (17) In the formula, e t For the first t Attention score for each time step; W e This is the weight matrix; b e For bias; α t For the first t Time attention weights for each time step; c This is a context vector used to represent the summary of historical information that is most critical to the current prediction.

[0016] Step 7: Perform inverse normalization on the prediction results of each modal component, and then superimpose and reconstruct the inverse normalized results of each modal component to obtain the final multi-step prediction result of the electronic expansion valve opening.

[0017] Step 8: Output the final multi-step prediction results to the air conditioning controller for electronic expansion valve opening adjustment, feedforward compensation, or model predictive control.

[0018] This invention also provides a system for predicting the opening degree of an electronic expansion valve in a condensing heat recovery air conditioner, comprising a data acquisition module, a data preprocessing module, a decomposition module, a normalization and sample construction module, a prediction module, a reconstruction module, and a control interface module. The data acquisition module collects actual operating data such as temperature, humidity, pressure, and electronic expansion valve opening degree during the operation of the condensing heat recovery air conditioner. The data preprocessing module is signal-connected to the data acquisition module and is used to perform outlier removal and feature filtering on the collected operating data. The decomposition module receives the preprocessed data and performs variational mode decomposition on the electronic expansion valve opening degree time series. The normalization and sample construction module normalizes the decomposed modal components and air conditioner operating parameters, and completes dataset partitioning and prediction sample construction. The prediction module integrates a gated recurrent unit network and an attention mechanism to perform multi-step opening degree prediction for each modal component. The reconstruction module aggregates the predicted data of each modal component and completes sequence reconstruction, outputting the final predicted value of the electronic expansion valve opening degree. The control interface module is communicatively connected to the air conditioner controller and transmits the predicted opening degree value to the air conditioner controller to achieve closed-loop regulation of the electronic expansion valve opening degree.

[0019] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program for performing the above-described prediction method.

[0020] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described prediction method.

[0021] The beneficial effects of this invention include: (1) By performing variational mode decomposition on the electronic expansion valve opening sequence, the non-stationarity and complexity of the original signal are reduced, making it easier for the model to learn the dynamic change law of the electronic expansion valve opening at different frequencies; (2) By extracting multivariate time series dependent features through gated cyclic unit network and combining attention mechanism to strengthen the expression of key historical information, the accuracy and error stability of multi-step prediction are improved; (3) The prediction results can be used as a forward-looking basis for electronic expansion valve opening adjustment, and can be used for feedforward compensation or model predictive control, thereby improving the adjustment response capability and operation stability of condensing heat recovery air conditioner under changing operating conditions. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the structure of a condensing heat recovery air conditioning system. Figure 2 This is a framework diagram of an electronic expansion valve opening prediction model based on variational mode decomposition and deep learning. Figure 3 Heatmap of Pearson correlation coefficient; Figure 4 This is a schematic diagram of the variational mode decomposition results; Figure 5 This is a structural diagram of a gated loop unit; Figure 6 This is a network structure diagram of a gated loop unit; Figure 7 A diagram of the attention mechanism structure; Figure 8 A comparison chart of the predicted results from each model with the actual values; Figure 9 A comparison chart of evaluation metrics for different models; Figure 10 Histograms of relative errors for different models; Figure 11 This is a comparison chart of the predicted results and actual values ​​of each model in the ablation experiment; Figure 12 A comparison chart of ablation experiment evaluation indicators; Figure 13 This is a histogram of the relative error in the ablation experiment.

[0023] In the diagram: 1. Compressor; 2. Outlet air temperature sensor; 3. Reheater; 4. Evaporator; 5. Electronic expansion valve; 6. Thermal expansion valve; 7. Condenser; 8. Fan; 9. Inlet air temperature sensor; 10. Air conditioning duct. Detailed Implementation

[0024] To clearly illustrate the objectives and technical solutions of this invention, the following detailed description, in conjunction with specific embodiments and accompanying drawings, further clarifies the invention. It should be understood that the specific embodiments described herein are merely illustrative and do not constitute a limitation on the scope of protection of this invention.

[0025] Example 1 like Figure 1 As shown, the condensing heat recovery air conditioning system in this embodiment adds a condensing heat recovery branch to the traditional vapor compression refrigeration cycle. The compressor 1 outlet is connected to the reheater 3 and the condenser 7 respectively. The reheater 3 branch is connected to the evaporator 4 via an electronic expansion valve 5, and the condenser 7 branch is connected to the evaporator 4 via a thermostatic expansion valve 6. The two refrigerant paths converge before the evaporator 4 and return to the compressor 1, forming a closed loop. The fan 8 drives the treated air to flow sequentially through the evaporator 4 and the reheater 3 to achieve cooling, dehumidification, and reheating processes. The electronic expansion valve 5 is used to regulate the refrigerant flow distribution ratio into the reheater 3 branch.

[0026] In the system described in this embodiment, changes in the opening degree of the electronic expansion valve 5 directly affect the refrigerant mass flow rate into the reheater branch 3. Changes in the refrigerant mass flow rate in the reheater branch, in turn, affect the heat release capacity of the reheater and further influence the outlet air temperature regulation process of the air conditioning system. Therefore, the opening degree of the electronic expansion valve 5 is not only related to operating parameters such as system temperature, pressure, and humidity, but also has a direct impact on the air supply status of the air conditioning system. Based on the above coupling relationship, a multivariate time series modeling method is used to predict the opening degree of the electronic expansion valve, providing a forward-looking reference for subsequent adjustment of the electronic expansion valve 5.

[0027] The flowchart of the electronic expansion valve opening prediction method proposed in this embodiment is as follows: Figure 2 As shown, the specific steps include the following: Step 1, Data Collection. From Figure 1 Multivariate time series data were collected during the actual operation of the air conditioning system shown. The sampling interval was 2 seconds. The collected variables included the actual operating data of the air conditioning system, such as inlet air temperature, inlet air humidity, outlet air temperature, outlet air humidity, high pressure, low pressure, and electronic expansion valve opening. The total sample size was approximately 100,000 records.

[0028] Step two, data preprocessing. For outliers identified during data collection, a box plot method based on quartiles and interquartile ranges was first used to identify abnormal samples to ensure data quality. Subsequently, time alignment was performed on each variable. To reduce feature redundancy and improve training efficiency, Pearson correlation analysis was further employed, retaining input variables whose correlation with the opening degree of the electronic expansion valve 5 met the preset threshold conditions. The results are as follows: Figure 3 As shown, in this embodiment, the preset threshold is 0.3.

[0029] Step 3, decomposition of the opening signal. Variational mode decomposition is performed on the opening time series of the electronic expansion valve 5, decomposing it into two modal components with different center frequencies and one residual component. Partial data after decomposition is shown below. Figure 4 As shown in the diagram, the low-frequency mode characterizes the opening trend of the electronic expansion valve 5, the high-frequency mode characterizes rapid fluctuations, and the residual term characterizes the remaining components not captured by the primary mode. This step decouples the different frequency components in the original electronic expansion valve 5 opening signal, thereby reducing the non-stationarity of the original sequence and improving the ability of subsequent deep learning models to model dynamic features.

[0030] Step 4: Normalization and Data Partitioning. The data is normalized using the min-max normalization method. The normalized data is then divided into training, validation, and test sets; in this embodiment, the training, validation, and test sets are divided in an 8:1:1 ratio.

[0031] Step 5, Sample Construction. Using continuous historical time windows as model input, and combining the filtered operating parameters with the decomposed modal components, multivariate time series samples are constructed. The corresponding modal values ​​at multiple future sampling times are used as the model output. In this embodiment, the historical window length is set to 150 sampling times, and the prediction step size is set to 10 sampling times.

[0032] Step six, construct the prediction model. For example... Figures 5 to 7 As shown, a deep learning model consisting of a gated recurrent unit network (GRN) and an attention mechanism is established. The GRN encodes the input time series samples and extracts time-dependent features to enhance the model's ability to express historical time-series information. The attention mechanism scores and assigns weights to the hidden states at different historical time steps, thereby highlighting key historical state information that contributes significantly to the prediction results. Finally, the prediction results of each modal component at multiple future time steps are output through a fully connected layer. This approach enables effective modeling of the multi-step change trend of the electronic expansion valve opening.

[0033] Step 7, Model Training. Input the constructed training set samples into the prediction model for training. The model training uses the Adam optimization algorithm, with mean squared error as the loss function, a learning rate of 0.0005, a maximum number of training epochs of 100, a batch size of 128, a gated recurrent unit hidden unit of 64, an attention hidden dimension of 64, and a Dropout rate of 0.2.

[0034] Step 8: Result Reconstruction and Output. The prediction results of each modal component are inversely normalized, and then the inversely normalized results of each modal component are superimposed and reconstructed to obtain the final multi-step prediction result of the opening degree of the electronic expansion valve 5.

[0035] In this embodiment, to verify the effectiveness of the method of the present invention, the proposed prediction model is compared with RNN, LSTM, GRU and Temporal Convolutional Network (TCN) models.

[0036] Select the root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R-squared, R0). 2 This is used as an evaluation index for the prediction results. The expression is as follows: (18) (19) (20) (twenty one) In the formula, n The number of data points; y i For the first i The actual value of each data point; For the first i Predicted values ​​for each data point. This is the average of the actual values.

[0037] from Figure 8 The comparison between the actual and predicted values ​​shows that the prediction curve of the prediction model described in this application generally follows the same trend as the actual value, and can effectively track the dynamic change process of the opening degree of the electronic expansion valve 5, indicating that the method has good prediction accuracy. Furthermore, from... Figure 9 Based on the evaluation metrics shown, the prediction model described in this application has relatively low RMSE, MAE, and MAPE on the test set, and R... 2 The error rates remained high across all prediction step sizes. As the prediction step size increased, all model error metrics showed an upward trend, while R0 remained high. 2 The corresponding decrease is related to error accumulation and increased prediction uncertainty during multi-step prediction. However, the degree of performance degradation varies among different models. Taking R... 2 For example, the prediction model described in this application decreased from 0.9892 in the first step to 0.9767 in the tenth step, a decrease of 1.26%. In comparison, the R-values ​​of the GRU, LSTM, RNN, and TCN models are significantly lower. 2 The decreases were 3.0%, 5.0%, 5.1%, and 7.3%, respectively. This demonstrates that, even with multi-step prediction, the prediction model described in this application can still maintain good predictive performance.

[0038] from Figure 10 Looking at the relative error distributions of different models, the prediction errors of each model generally exhibit an approximately symmetrical distribution, but the concentration of the distribution varies. The prediction model described in this application has a more concentrated error distribution, with approximately 98.2% of the relative prediction errors falling within the ±2% range. Its relative prediction error concentration is improved by 8.7%, 10.9%, 11.5%, and 14.6% compared to GRU, RNN, LSTM, and TCN, respectively. This indicates that the proposed model has better stability in error fluctuation control. The obtained multi-step prediction results of the electronic expansion valve opening can be further used as a feedforward reference or as a candidate control sequence input to the controller in model predictive control, thereby providing a basis for the outlet air temperature regulation of the condensing heat recovery air conditioning system.

[0039] Example 2 To verify the function of each component module, further ablation experiments were conducted. The results are as follows: Figures 11 to 13 As shown. From Figure 11 The comparison results between the predicted and actual values ​​show that as the model structure is gradually improved, the degree of fit between the predicted curve and the actual value gradually increases. Among them, the prediction result of the method described in this application is closest to the actual value and can better reflect the dynamic change trend of the opening of the electronic expansion valve 5, indicating that the method has good prediction accuracy.

[0040] From a statistical and error distribution perspective, using the GRU model as the baseline model, adding an attention mechanism reduces the average RMSE, MAE, and MAPE over ten prediction steps by 4.6%, 7.1%, and 7.0%, respectively, and the average R... 2 The improvement was 0.35%; after further superimposing variational mode decomposition, the prediction error distribution range was further reduced, mainly concentrated in the ±2% range. The average RMSE, MAE, and MAPE of the ten prediction steps decreased by 34.8%, 34.7%, and 34.7% respectively compared with the baseline model, and the average R... 2 The improvement was 2.4%. The results show that variational mode decomposition achieves frequency feature decomposition at the input level, GRU establishes time series dependencies, and the attention mechanism strengthens the expression of key historical information. The combination of these three factors helps to improve the accuracy and stability of multi-step prediction of the opening of the electronic expansion valve.

[0041] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the opening degree of an electronic expansion valve in a condensing heat recovery air conditioner, characterized in that, This invention relates to an air conditioning system with a condensing heat recovery branch. The air conditioning system includes a compressor, condenser, reheater, evaporator, electronic expansion valve, thermostatic expansion valve, fan, and sensors electrically connected to a controller. The compressor outlet is connected to both the reheater and condenser. The reheater and condenser are connected to the evaporator via the electronic expansion valve and thermostatic expansion valve, respectively. The evaporator is connected to the compressor. The electronic expansion valve opening prediction method includes the following steps: S1: Synchronously collect and record multivariate time series data during the actual operation of the air conditioning system. The multivariate time series data includes at least the opening degree of the electronic expansion valve and parameters such as temperature, pressure and humidity that characterize the operating status of the system. S2: Preprocess the multivariate time series data, including outlier removal from box plots based on quartiles and interquartile ranges, time alignment of each operating parameter according to the same sampling period, and input feature filtering based on Pearson correlation coefficient, to obtain preprocessed data for prediction modeling. S3: Perform variational mode decomposition on the electronic expansion valve opening time series to obtain multiple modal components and residual components with different center frequencies; S4: Normalize each modal component and other preprocessed operating parameters, and divide the training set, validation set and test set according to a preset ratio; S5: Construct multivariate historical input samples and future multi-step prediction output samples based on the sliding time window mechanism. The input samples consist of the operating parameters and modal components of continuous historical moments, and the output samples are the modal values ​​corresponding to the opening of the electronic expansion valve at multiple future moments. S6: Input the input sample into a deep learning prediction model composed of a gated recurrent unit network and an attention mechanism. Extract time-dependent features through the gated recurrent unit and strengthen the information representation of key historical moments through the attention mechanism to obtain the prediction results of each modality component at multiple future moments. S7: Perform inverse normalization on the prediction results of each modal component, and then superimpose and reconstruct the inverse normalized results of each modal component to obtain the final multi-step prediction result of the electronic expansion valve opening. S8: Output the final multi-step prediction result to the air conditioning controller as a reference input for electronic expansion valve opening adjustment, feedforward compensation, or model predictive control.

2. The prediction method according to claim 1, characterized in that, Outlier removal in step S2 employs a box plot method based on quartiles and interquartile ranges; let a certain operating parameter sequence be... Sort the sequence in ascending order to obtain the ordered sequence. ,in, ; In the ordered sequence, the first quartile Q 1 represents the value corresponding to the lower quartile position, the third quartile. Q 3 represents the value corresponding to the upper quartile position, and its calculation method can be expressed as follows: (1) (2) Then the interquartile range (IQR) is: (3) Lower bound for outlier detection L and upper limit U They are respectively: (4) (5) When sample value x i < L or x i > U When this happens, the sample is identified as an outlier and removed. The input feature selection in step S2 employs a correlation analysis method based on the Pearson correlation coefficient, retaining operating parameters whose absolute value of the correlation coefficient with the electronic expansion valve opening is greater than a preset threshold as model input features; the calculation method of the Pearson correlation coefficient is as follows: (6) In the formula, n The number of data points; x i , y i For the first i The actual value of each data point; , This is the average of the actual values.

3. The prediction method according to claim 1, characterized in that, The variational mode decomposition in step S3 is used to decompose the original electronic expansion valve opening sequence into at least two modal components and one residual term, so as to separate high-frequency fluctuation information from low-frequency trend information and reduce the non-stationarity and complexity of the original opening signal; the variational mode decomposition adopts the following constraint method: (7) In the formula, f ( t ) represents the original time series signal. u k ( t ) indicates the first k One modal function, ω k The corresponding center frequency is represented by *, where * indicates convolution operation. δ ( t ) is the Dirac function, j The imaginary unit is used; the constraints ensure that the sum of all modal functions can reconstruct the original signal.

4. The prediction method according to claim 1, characterized in that, In step S4, the max-min normalization method is used to perform dimensionless processing on each modal component and operating parameter to reduce the impact of differences in the dimensions of different variables on the stability of model training; the normalization calculation method is as follows: (8) In the formula For the normalized data, x , x min and x max These represent the data to be normalized, the maximum value of the data to be normalized, and the minimum value of the data to be normalized, respectively.

5. The prediction method according to claim 1, characterized in that, In step S5, a sliding time window is used to construct supervised learning samples. The multivariate sequence of multiple consecutive historical moments is used as the input sample, and the modal values ​​corresponding to the opening of the electronic expansion valve at multiple future moments are used as the output sample, thereby establishing a multivariate time series multi-step prediction task. The expressions for the input and output samples are as follows: No. i Input Samples X i It can be represented as: (9) In the formula, M For the number of input variables, T For the length of the history window, H To predict the step size, No. i Output samples Y i for: (10)。 6. The prediction method according to claim 1, characterized in that, The gated recurrent unit network in step S6 is used to extract time-dependent features from the input samples, and the attention mechanism is used to weight the hidden states at each historical time step to enhance the ability to express key historical information. The deep learning prediction model adopts a direct multi-step output method to output the predicted opening values ​​of the electronic expansion valve at multiple future time steps. t The calculation process of updating the gate, resetting the gate, and calculating the hidden state in the gated loop unit at each time step can be represented as: (11) (12) (13) (14) In the formula, z t and r t This indicates the update gate and the reset gate; σ Use the Sigmoid activation function; W z , W r , W h This is the weight matrix; h t－1 and h t These represent the hidden states at the previous and current time points, respectively. x t This represents the input feature vector at the current moment; This indicates the candidate is in a hidden state.

7. The prediction method according to claim 6, characterized in that, The attention mechanism includes: scoring the hidden states output by the gated recurrent unit network at each time step, calculating the attention weights corresponding to each time step; weighting and summing the hidden states according to the attention weights to obtain a context vector; and then inputting the context vector into a fully connected layer to output the prediction results of each modal component. The calculation process of the attention mechanism can be expressed as follows: (15) (16) (17) In the formula, e t For the first t Attention score for each time step; W e This is the weight matrix; b e For bias; α t For the first t Time attention weights for each time step; c This is a context vector used to represent the summary of historical information that is most critical to the current prediction.

8. A system for predicting the opening degree of an electronic expansion valve in a condensing heat recovery air conditioner, characterized in that, include: The data acquisition module is used to collect actual operating data during the operation of the condensing heat recovery air conditioner. The actual operating data includes temperature, humidity, pressure, and electronic expansion valve opening. A data preprocessing module, which is signal-connected to the data acquisition module, is used to perform outlier removal and feature filtering on the acquired running data. The decomposition module receives preprocessed data and is used to perform variational mode decomposition on the opening time series of the electronic expansion valve. The normalized sample construction module is used to normalize the modal components and air conditioning operating parameters obtained from the decomposition, and to complete the dataset partitioning and prediction sample construction. The prediction module integrates a gated recurrent unit network and an attention mechanism to perform multi-step opening prediction for each modal component. The reconstruction module is used to aggregate the prediction data of each modal component and complete the sequence reconstruction, and output the final predicted value of the electronic expansion valve opening. The control interface module is communicatively connected to the air conditioning controller and is used to transmit the opening prediction value to the air conditioning controller to realize closed-loop regulation of the electronic expansion valve opening.

9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the prediction method according to any one of claims 1 to 7.

Images