Airport terminal air conditioner cooling load prediction method based on improved iTransform

Abstract

The invention discloses an airport terminal air conditioner cooling load prediction method based on an improved iTransform. The airport terminal air conditioner cooling load prediction method comprises the following steps that cooling data are collected from an airport TES system; and de-noising the input variables by using the SWT. And then an improved iTransform model is used to carry out airport terminal cooling load prediction, and the model fuses a multi-scale time sequence convolutional network (TCN), Fourier transform and a differential attention mechanism. Local features of different time scales are extracted in parallel by using a multi-scale TCN module (MultiScaleTCN), and multi-scale fusion is realized through interpolation; then, through a Fourier block (Fourier block), the periodic mode modeling capability is enhanced in a frequency domain; finally, a differential multiple-head attention mechanism (MHDA) is introduced into the iTransform, a learnable differential weight and a causal mask are introduced, and the perception ability of the model to the dynamic change of the time sequence is improved. According to the method, the improved iTransform model is used for predicting the air conditioner cooling load of the airport terminal, and the cooling load prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of air conditioning cooling load forecasting, specifically a method for forecasting air conditioning cooling load in airport terminals based on an improved iTransformer. Background Technology

[0002] Existing building energy conservation measures mostly focus on optimizing the operation phase of heating, ventilation, and air conditioning systems. While this real-time feedback-based control strategy has some effect, it is essentially still a passive response mode and struggles to overcome the bottleneck of energy efficiency improvement. To achieve the leap from "passive response" to "proactive intelligence" in building energy conservation, it is essential to develop predictive energy management systems, which is particularly urgent for large public buildings with concentrated energy consumption and complex systems.

[0003] Airport terminals, as typical high-energy-consuming public buildings, exhibit distinctive energy consumption structures. Studies show that air conditioning systems account for 40%-60% of the total building energy consumption, with refrigeration systems making up a significant portion. This high energy consumption stems from the unique architectural structure of terminals, such as their large spaces, high ceilings, and glass curtain walls, as well as their high passenger volume, complex internal heat sources, and long operating hours. These energy consumption characteristics indicate that airport air conditioning systems, especially refrigeration systems, possess enormous energy-saving potential. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer, comprising the following steps:

[0005] Step 1) Collect cooling data from the airport's TES system and preprocess it to obtain cooling characteristic data;

[0006] Step 2) Use the synchronous squeezing wavelet transform method to denoise the cooling feature data, and divide the denoised data into training set and test set;

[0007] Step 3) Construct an airport terminal air conditioning cooling load prediction model based on the improved iTansformer; the prediction model integrates multi-scale temporal convolutional network Fourier transform and differential attention mechanism;

[0008] Step 4) Train and test the airport terminal air conditioning cooling load prediction model using the training set and test set respectively, and adjust the model parameters;

[0009] Step 5) Use the tested airport terminal air conditioning cooling load prediction model to predict the airport terminal air conditioning cooling load for the next time period t.

[0010] Furthermore, the preprocessing includes data alignment, missing value handling, and feature enhancement.

[0011] Furthermore, the method for handling missing values ​​is interpolation.

[0012] Furthermore, feature enhancement refers to using the Pearson correlation coefficient method to analyze the feature correlation of cooling data and enhance the three features of indoor dry-bulb temperature, outdoor wet-bulb temperature, and outdoor humidity respectively.

[0013] Furthermore, the steps for denoising the cooling feature data using the synchronous squeezing wavelet transform method are as follows:

[0014] Step 2.1) Using the selected mother wavelet function (t) Cooling characteristic data Perform continuous wavelet transform and calculate the wavelet coefficients. ,Right now:

[0015] (1)

[0016] (2)

[0017] In the formula, a represents the scaling factor, b represents the translation factor, and ψ* represents the conjugate wavelet;

[0018] Step 2.2) Calculate wavelet coefficients Instantaneous frequency on the scale-time plane ,Right now:

[0019] (3)

[0020] in, This indicates taking the partial derivative with respect to the translation parameter b;

[0021] Step 2.3) Set the dynamic noise threshold r to satisfy... The wavelet coefficients of r are treated as noise and discarded;

[0022] Step 2.4) Based on instantaneous frequency The effective wavelet coefficients, after threshold filtering, are redistributed to a series of preset frequency centers. Above, and calculate the time-frequency coefficient after compression, that is:

[0023] (4)

[0024] In the formula, In frequency The coefficient after compression at time b, For scale intervals, Frequency resolution;

[0025] Step 2.5) Calculate the time-frequency coefficients after extrusion. Performing an inverse transform, we reconstruct the denoised signal f(t), i.e.:

[0026] (5)

[0027] In the formula, Let Re be a normalization constant associated with the selected wavelet, and let Re denote the real part.

[0028] Furthermore, the dynamic noise threshold r is as follows:

[0029] (6)

[0030] In the formula, n is the signal length, and median represents the median function.

[0031] Furthermore, the parameters to be adjusted include the learning rate, batch size, and regularization parameter.

[0032] Furthermore, the airport terminal air conditioning cooling load prediction model based on the improved iTansformer includes a multi-scale temporal convolutional network, a frequency domain feature enhancement module, a differential attention mechanism, and a feedforward neural network module;

[0033] The multi-scale temporal convolutional network includes multiple parallel TCN branches and projection layers; each TCN branch is composed of multiple layers of dilated convolutions stacked together.

[0034] These TCN branches process the input data at different sampling rates, capturing features at different temporal granularities in the time series;

[0035] The outputs of these TCN branches are concatenated after being unified to the same time scale through linear interpolation, and then adaptively fused through a learnable projection layer to obtain temporal features.

[0036] The frequency domain feature enhancement module transforms the time domain features to the frequency domain space, performs feature transformation and selective filtering in the frequency domain, and then maps the processed features back to the time domain through inverse Fourier transform, forming a residual connection with the original time domain features.

[0037] The differential attention mechanism achieves feature interaction through dual-path attention calculation and uses RMSnorm for layer normalization, while maintaining autoregressive properties through causal masking.

[0038] The feedforward neural network module normalizes, nonlinearly transforms, and projects the features processed by the multi-scale temporal convolutional network, the frequency domain feature enhancement module, and the differential attention mechanism to generate the cold load value at the prediction time.

[0039] Furthermore, the frequency domain feature enhancement module stores a learnable frequency domain projection matrix;

[0040] The frequency domain projection matrix is ​​used to identify and enhance frequency components that are critical to the prediction task.

[0041] The frequency domain projection matrix is ​​in the form of a complex parameter matrix, which is randomly initialized and dynamically optimized during the network training process.

[0042] Furthermore, the differential attention mechanism includes a parallel main path and a differential path; the outputs of the two paths are fused through differential operations. Specifically, the main path uses standard dot product attention to calculate feature relevance, while the differential path introduces a learnable scalar λ to generate compensatory attention weights.

[0043] Differential paths map query vectors, key vectors, and value vectors to the output through a multi-head differential attention mechanism;

[0044] The multi-head differential attention mechanism uses the difference between two softmax attention maps for each head to cancel attention noise;

[0045] The output of the differential path is shown below:

[0046] (6)

[0047] , V= X

[0048] Among them, query vector Key vector Value vector V = X , λ is a parameter, and λ is a learnable scalar;

[0049] The learnable scalar λ is shown below: (7)

[0050] In the formula, It is a learnable scalar. (0,1) is a constant used to initialize λ.

[0051] The technical effects of this invention are undeniable, and its beneficial effects are as follows:

[0052] (1) This invention constructs dynamic time series samples by introducing a sliding time window to enhance the time series modeling capability of cooling load; combines the Pearson correlation coefficient algorithm to quantitatively screen key meteorological and operational characteristics strongly correlated with cooling load; and innovatively introduces synchronous wavelet transform to denoise the original load sequence, systematically improving the quality and reliability of input data, and providing high signal-to-noise ratio time series input for subsequent deep models.

[0053] (2) This invention designs a parallelized temporal convolutional network architecture and uses convolutional branches with different expansion rates (1 hour, 4 hours, 24 hours) to accurately extract the short-term fluctuations, medium-term changes and long-term periodic features contained in the cold load sequence. Then, it uses an adaptive interpolation algorithm to achieve scale alignment and deep fusion of multi-scale features, thereby constructing a feature vector with strong characterization ability for the complex dynamic characteristics of cold load.

[0054] (3) The present invention innovatively integrates Fourier blocks into the model to explicitly enhance the modeling ability of key periodic patterns such as daily and weekly cycles through frequency domain transformation; at the same time, it introduces a differential multi-head attention mechanism in the iTransformer architecture, enabling the model to focus on the first-order differential change trend of time series data, which significantly improves the model's perception sensitivity and prediction accuracy of the dynamic evolution of load. Attached Figure Description

[0055] Figure 1 For the prediction model structure;

[0056] Figure 2 The following are the specific implementation steps of this invention.

[0057] Figure 3 This is a graph showing the prediction results of this invention. Detailed Implementation

[0058] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0059] Example 1:

[0060] A method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer includes the following steps:

[0061] Step 1) Collect cooling data from the airport's TES system and preprocess it to obtain cooling characteristic data;

[0062] Step 2) Use the synchronous squeezing wavelet transform method to denoise the cooling feature data, and divide the denoised data into training set and test set;

[0063] Step 3) Construct an airport terminal air conditioning cooling load prediction model based on the improved iTansformer; the prediction model integrates multi-scale temporal convolutional network Fourier transform and differential attention mechanism;

[0064] Step 4) Train and test the airport terminal air conditioning cooling load prediction model using the training set and test set respectively, and adjust the model parameters;

[0065] Step 5) Use the tested airport terminal air conditioning cooling load prediction model to achieve real-time prediction of the airport terminal air conditioning cooling load.

[0066] Example 2:

[0067] A method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer is provided. The technical content is the same as in Embodiment 1. Further, the preprocessing includes data alignment, missing value processing, and feature enhancement.

[0068] Example 3:

[0069] A method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer, with the same technical content as any one of Examples 1-2, further wherein the method for handling missing values ​​is interpolation.

[0070] Example 4:

[0071] A method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer, with the same technical content as any one of Examples 1-3, further wherein feature enhancement refers to: using the Pearson correlation coefficient method to analyze the feature correlation of cooling data, and enhancing the three features of indoor dry-bulb temperature, outdoor wet-bulb temperature, and outdoor humidity respectively.

[0072] Example 5:

[0073] A method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer, with the same technical content as any one of embodiments 1-4, further comprising the following steps for denoising the cooling feature data using a synchronous squeezing wavelet transform method:

[0074] Step 2.1) Using the selected mother wavelet function (t) Cooling characteristic data Perform continuous wavelet transform and calculate the wavelet coefficients. ,Right now:

[0075] (1)

[0076] (2)

[0077] In the formula, a represents the scaling factor, b represents the translation factor, and ψ* represents the conjugate wavelet;

[0078] Step 2.2) Calculate wavelet coefficients Instantaneous frequency on the scale-time plane ,Right now:

[0079] (3)

[0080] in, This indicates taking the partial derivative with respect to the translation parameter b;

[0081] Step 2.3) Set the dynamic noise threshold r to satisfy... The wavelet coefficients of r are treated as noise and discarded;

[0082] Step 2.4) Based on instantaneous frequency The effective wavelet coefficients, after threshold filtering, are redistributed to a series of preset frequency centers. Above, and calculate the time-frequency coefficient after compression, that is:

[0083] (4)

[0084] In the formula, In frequency The coefficient after compression at time b, For scale intervals, Frequency resolution;

[0085] Step 2.5) Calculate the time-frequency coefficients after extrusion. Performing an inverse transform, we reconstruct the denoised signal f(t), i.e.:

[0086] (5)

[0087] In the formula, Let Re be a normalization constant associated with the selected wavelet, and let Re denote the real part.

[0088] Example 6:

[0089] A method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer, with the same technical content as any one of embodiments 1-5, further wherein the dynamic noise threshold r is as follows:

[0090] (6)

[0091] In the formula, n is the signal length, and median represents the median function.

[0092] Example 7:

[0093] A method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer, with the same technical content as any one of Examples 1-6, further comprising adjusting parameters including learning rate, batch size, and regularization parameters.

[0094] Example 8:

[0095] An airport terminal air conditioning cooling load prediction method based on improved iTransformer, with the same technical content as any one of embodiments 1-7. Further, the airport terminal air conditioning cooling load prediction model based on improved iTransformer includes a multi-scale temporal convolutional network, a frequency domain feature enhancement module, a differential attention mechanism, and a feedforward neural network module.

[0096] The multi-scale temporal convolutional network includes multiple parallel TCN branches and projection layers; each TCN branch is composed of multiple layers of dilated convolutions stacked together.

[0097] These TCN branches process the input data at different sampling rates, capturing features at different temporal granularities in the time series;

[0098] The outputs of these TCN branches are concatenated after being unified to the same time scale through linear interpolation, and then adaptively fused through a learnable projection layer to obtain temporal features.

[0099] The frequency domain feature enhancement module transforms the time domain features to the frequency domain space, performs feature transformation and selective filtering in the frequency domain, and then maps the processed features back to the time domain through inverse Fourier transform, forming a residual connection with the original time domain features.

[0100] The differential attention mechanism achieves feature interaction through dual-path attention computation and uses RMSnorm for layer normalization, while preserving autoregressive properties through causal masking.

[0101] In the improved iTransformer architecture, the final predicted air conditioning cooling load data is output by the feedforward neural network module at the end. Specifically, after feature extraction, frequency domain optimization, and feature fusion are completed sequentially by a multi-scale temporal convolutional network, a frequency domain feature enhancement module, and a differential attention mechanism, the processed high-level features are normalized by a LayerNorm layer, and finally, the feedforward neural network performs nonlinear transformation and projection to generate the cooling load value at the prediction time.

[0102] Example 9:

[0103] A method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer, with the same technical content as any one of embodiments 1-8, further wherein the frequency domain feature enhancement module stores a learnable frequency domain projection matrix.

[0104] The frequency domain projection matrix is ​​used to identify and enhance frequency components that are critical to the prediction task.

[0105] In the frequency domain feature enhancement module, the learnable frequency domain projection matrix adopts the form of a complex parameter matrix, which is randomly initialized and dynamically optimized during the network training process. This matrix is ​​applied to the frequency domain signal after Fourier transform through matrix multiplication, enabling it to autonomously learn the contribution weight of different frequency components to the cold load prediction, thereby achieving adaptive enhancement of key periodic features and suppression of noise frequencies, ultimately improving the prediction accuracy of the model.

[0106] Example 10:

[0107] A method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer, with the same technical content as any one of embodiments 1-9, further comprising the following: the differential attention mechanism includes a parallel main path and a differential path; the outputs of the two paths are fused through differential operations. Specifically, the main path uses standard dot product attention to calculate feature correlation, while the differential path introduces a learnable scalar λ to generate compensating attention weights.

[0108] Differential paths map query vectors, key vectors, and value vectors to the output through a multi-head differential attention mechanism;

[0109] The multi-head differential attention mechanism uses the difference between two softmax attention maps for each head to cancel attention noise;

[0110] The output of the differential path is shown below:

[0111] (6)

[0112] , V= X

[0113] Among them, query vector Key vector Value vector V = X , λ is a parameter, and λ is a learnable scalar;

[0114] The learnable scalar λ is shown below: (7)

[0115] In the formula, It is a learnable scalar. (0,1) is a constant used to initialize λ.

[0116] Example 11:

[0117] A method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer, with the same technical content as any one of embodiments 1-9, further wherein the prediction model receives historical cooling load data over a period of time at the current time T, including the values ​​of indoor dry-bulb temperature, outdoor wet-bulb temperature, and outdoor humidity at times T-3, T-2, T-1, T, T+1, T+2, and T+3 as input, and finally outputs the predicted cooling load value for the next 24 hours T+.

[0118] Example 12:

[0119] The specific steps of a method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer are as follows:

[0120] S100: Data preprocessing: Cooling data is collected from the airport TES system, and then missing values ​​are filled in using interpolation. The Pearson correlation coefficient method is used to analyze feature correlation. The three features of indoor dry-bulb temperature, outdoor wet-bulb temperature, and outdoor humidity are enhanced respectively, and the 10 features with the strongest correlation are output and saved before proceeding to the next step.

[0121] S200: Use SWT to denoise the dataset and improve data quality; proceed to the next step;

[0122] S201: Perform continuous wavelet transform (CWT) on the input signal For the acquired input signal Using the selected mother wavelet function (t) Perform continuous wavelet transform and calculate its wavelet coefficients. The formula for calculating wavelet coefficients is as follows:

[0123]

[0124]

[0125] Where a represents the scaling factor, b represents the translation factor, and ψ* represents the conjugate wavelet.

[0126] S202: Calculate the instantaneous frequency. Based on the wavelet coefficients obtained in step S301. Calculate its instantaneous frequency in the scale-time plane. The formula for calculating instantaneous frequency is as follows:

[0127]

[0128] in, This indicates taking the partial derivative with respect to the translation parameter b.

[0129] S203: Set the noise threshold and filter the valid wavelet coefficients.

[0130] To suppress high-frequency noise, a dynamic noise threshold r is set. The formula for calculating the threshold r is:

[0131]

[0132] Where n is the signal length, and medium represents the median function, which satisfies... The wavelet coefficients of r are treated as noise and discarded.

[0133] S204: Perform synchronous extrusion operation.

[0134] The effective wavelet coefficients after threshold filtering are then analyzed based on their corresponding instantaneous frequencies. Reassigned to a series of preset frequency centers Above. The synchronous extrusion operation is defined by the following formula:

[0135]

[0136] in, In frequency The coefficient after compression at time b, For scale intervals, This refers to the frequency resolution.

[0137] S205: Reconstruct the denoised signal through inverse transform. The time-frequency coefficients obtained in step S204 after synchronous compression... Perform an inverse transform to reconstruct the denoised signal f(t). The reconstruction formula is as follows:

[0138]

[0139] in, Let Re be a normalization constant associated with the selected wavelet, and let Re denote the real part.

[0140] S300: Use the dataset processed by S200 as the training set and test set, with a ratio of 8:2 between the training set and the test set.

[0141] S400: Constructing an improved prediction model for iTansformer: This model integrates a multi-scale temporal convolutional network (TCN), Fourier transform, and differential attention mechanism, which can effectively capture long-term dependencies, local features, and periodic patterns in time series.

[0142] The specific construction method is as follows:

[0143] S401: Multi-scale Temporal Feature Extraction

[0144] This module employs an innovative MultiScale Temporal Convolutional Network (MultiScaleTCN) to capture feature patterns at different temporal granularities in time series. The module comprises three parallel TCN branches, processing the input sequence at sampling rates of [1, 4, 24], specifically designed to capture original high-frequency features, daily periodic patterns, and weekly periodic patterns. Each branch consists of multiple stacked dilated convolutions, with the dilation rate of the l-th layer set to 2^l, enabling the model to exponentially expand its receptive field and effectively model long-term dependencies. The features processed by each branch are unified to the same time scale through linear interpolation and then concatenated, followed by a learnable projection layer to achieve adaptive fusion of multi-scale features. This design allows the model to simultaneously focus on local details and global trends, significantly improving the expressive power of temporal features.

[0145] S402: Frequency Domain Feature Enhancement Module

[0146] To overcome the limitations of traditional time-domain modeling, this model introduces a frequency-domain feature enhancement mechanism. This module first transforms the time-domain features to the frequency domain using a Fast Fourier Transform (FFT), performing feature transformation and selective filtering in the frequency domain. Specifically, we design a learnable frequency-domain projection matrix that automatically identifies and enhances frequency components crucial for the prediction task. The processed features are then remapped back to the time domain using an Inverse Fourier Transform (iFFT) and form residual connections with the original features. This bidirectional frequency-time domain transformation mechanism enables the model to explicitly model periodic and seasonal patterns while maintaining the integrity of the original time-domain information, providing richer feature representations for subsequent processing.

[0147] S403: Differential Attention Mechanism

[0148] This module achieves more refined feature interaction through dual-path attention computation. The mechanism comprises two parallel paths: the main path uses standard dot-product attention to calculate feature relevance, while the differential path introduces a learnable λ parameter to generate compensatory attention weights. The outputs of the two paths are fused through differential operations, with the λ parameter being exponentially processed to ensure numerical stability. To ensure training effectiveness, the module uses RMSnorm for layer normalization and incorporates a causal mask to maintain autoregressive properties. This design not only enhances the model's ability to capture long-range dependencies but also automatically identifies and suppresses noisy attention, significantly improving the effectiveness and robustness of the attention mechanism.

[0149] The differential attention mechanism maps query vectors, key vectors, and value vectors to the output. We use query and key vectors to compute attention scores, and then calculate a weighted sum of the value vectors. A key design feature is that we use a pair of softmax functions to eliminate noise in the attention scores. Specifically, given input X ∈ R, we first project them as query, key, and value vectors. , ∈ R, V ∈ R. Then, the differential attention operator DiffAttn(·) computes the output as follows:

[0150] , V= X

[0151] V

[0152] in , Here, λ is a parameter, and λ is a learnable scalar. To synchronously learn the dynamics, the scalar λ is reparameterized as follows:

[0153]

[0154] in It is a learnable scalar. (0,1) is a constant used to initialize λ.

[0155] Multi-head differential attention. Each head uses the difference between two softmax attention maps to cancel out attention noise. λ is a learnable scalar initialized to λ. GroupNorm applies normalization individually to each head. GroupNorm is followed by a fixed multiplier (1 - λ) that aligns the gradient flow with the diff-Transformer.

[0156] S500: Train the air conditioning cooling load prediction model using the training set in the prediction dataset; validate the prediction model and adjust the parameters using the test set; the adjustment parameters include the learning rate, scheduling strategy and regularization parameters;

[0157] We use mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and mean absolute percentage error (MAPE) as evaluation indicators, and their calculation formulas are as follows:

[0158]

[0159]

[0160]

[0161]

[0162] Where n represents the total sample size. and These represent the actual and predicted values ​​of power materials, respectively.

[0163] To demonstrate the effectiveness of the model of this invention in cooling load of airport terminal air conditioning, comparative experiments were conducted using RNN, MLP, LSTM, DLinear, LightTS, Reformer, Informer, Pyraformer, and iTransformer.

[0164] Experimental results show that this invention outperforms other algorithms in all three evaluation metrics: MAE, MSE, RMSE, and MAPE. In the air conditioning cooling load prediction task, the Transformer-based model performs exceptionally well overall. Among them, iTransformer, with its optimal MAE (1439.36) and RMSE (1958.94), is the best choice among the benchmark models, and its variable dimension attention mechanism effectively models the synergistic effects of multi-source features. Our model achieves significant improvements over iTransformer in all key metrics of air conditioning cooling load prediction. Specifically, in terms of error metrics, Our model reduces MAE by 10.90%, MSE by 32.98%, RMSE by 18.14%, and MAPE by 7.12%. These data fully demonstrate the superiority of Our model in prediction accuracy and stability, indicating that the model has stronger robustness in predicting extreme load values. This verifies the effectiveness of the proposed optimization strategy and provides reliable technical support for building energy conservation optimization.

[0165] Table 1 Evaluation Indicators for Different Models

[0166]

[0167] pass Figure 1 It can be seen that, compared with other model algorithms, this method has a significant advantage in both curve fitting accuracy and prediction accuracy at corresponding points.

Claims

1. A method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer, characterized in that, Includes the following steps: Step 1) Collect cooling data from the airport's TES system and preprocess it to obtain cooling characteristic data; Step 2) Use the synchronous squeezing wavelet transform method to denoise the cooling feature data, and divide the denoised data into training set and test set; Step 3) Construct an airport terminal air conditioning cooling load prediction model based on the improved iTansformer; the prediction model integrates multi-scale temporal convolutional network Fourier transform and differential attention mechanism; Step 4) Train and test the airport terminal air conditioning cooling load prediction model using the training set and test set respectively, and adjust the model parameters; Step 5) Use the tested airport terminal air conditioning cooling load prediction model to predict the airport terminal air conditioning cooling load for the next time period t.

2. The method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer as described in claim 1, characterized in that: The preprocessing includes data alignment, missing value handling, and feature enhancement.

3. The method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer as described in claim 2, characterized in that: The method for handling missing values ​​is interpolation.

4. The method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer as described in claim 2, characterized in that: Feature enhancement refers to using the Pearson correlation coefficient method to analyze the feature correlation of cooling data and enhance the three features of indoor dry-bulb temperature, outdoor wet-bulb temperature, and outdoor humidity respectively.

5. The airport terminal air conditioning cooling load prediction method based on the improved iTransformer according to claim 1, characterized in that, The steps for denoising cooling feature data using the synchronous squeezing wavelet transform method are as follows: Step 2.1) Using the selected mother wavelet function (t) Cooling characteristic data Perform continuous wavelet transform and calculate the wavelet coefficients. ,Right now: （1） （2） In the formula, a represents the scaling factor, b represents the translation factor, and ψ* represents the conjugate wavelet; Step 2.2) Calculate wavelet coefficients Instantaneous frequency on the scale-time plane ,Right now: （3） in, This indicates taking the partial derivative with respect to the translation parameter b; Step 2.3) Set the dynamic noise threshold r to satisfy... The wavelet coefficients of r are treated as noise and discarded; Step 2.4) Based on instantaneous frequency The effective wavelet coefficients, after threshold filtering, are redistributed to a series of preset frequency centers. Above, and calculate the time-frequency coefficient after compression, that is: （4） In the formula, In frequency The coefficient after compression at time b, For scale intervals, Frequency resolution; Step 2.5) Calculate the time-frequency coefficients after extrusion. Performing an inverse transform, we reconstruct the denoised signal f(t), i.e.: （5） In the formula, Let Re be a normalization constant associated with the selected wavelet, and let Re denote the real part.

6. The airport terminal air conditioning cooling load prediction method based on the improved iTransformer according to claim 5, characterized in that, The dynamic noise threshold r is shown below: （6） In the formula, n is the signal length, and median represents the median function.

7. The method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer as described in claim 1, characterized in that: The parameters to be adjusted include the learning rate, batch size, and regularization parameter.

8. The method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer as described in claim 1, characterized in that: The airport terminal air conditioning cooling load prediction model based on the improved iTansformer includes a multi-scale temporal convolutional network, a frequency domain feature enhancement module, a differential attention mechanism, and a feedforward neural network module. The multi-scale temporal convolutional network includes multiple parallel TCN branches and projection layers; each TCN branch is composed of multiple layers of dilated convolutions stacked together. These TCN branches process the input data at different sampling rates, capturing features at different temporal granularities in the time series; The outputs of these TCN branches are concatenated after being unified to the same time scale through linear interpolation, and then adaptively fused through a learnable projection layer to obtain temporal features. The frequency domain feature enhancement module transforms the time domain features to the frequency domain space, performs feature transformation and selective filtering in the frequency domain, and then maps the processed features back to the time domain through inverse Fourier transform, forming a residual connection with the original time domain features. The differential attention mechanism achieves feature interaction through dual-path attention calculation and uses RMSNorm for layer normalization, while maintaining autoregressive properties through causal masking. The feedforward neural network module normalizes, nonlinearly transforms, and projects the features processed by the multi-scale temporal convolutional network, the frequency domain feature enhancement module, and the differential attention mechanism to generate the cold load value at the prediction time.

9. The airport terminal air conditioning cooling load prediction method based on the improved iTransformer according to claim 8, characterized in that: The frequency domain feature enhancement module stores a learnable frequency domain projection matrix; the frequency domain projection matrix adopts the form of a complex parameter matrix, which is randomly initialized and dynamically optimized during the network training process. The frequency domain projection matrix is ​​used to identify and enhance frequency components that are critical to the prediction task.

10. A method for predicting the cooling load of airport terminal air conditioning based on an improved iTransformer as described in claim 8, characterized in that: The differential attention mechanism includes a parallel main path and a differential path; the outputs of the two paths are fused through differential operations; wherein, the main path uses standard dot product attention to calculate feature correlation, and the differential path introduces a learnable scalar λ to generate compensatory attention weights. Differential paths map query vectors, key vectors, and value vectors to the output through a multi-head differential attention mechanism; The multi-head differential attention mechanism uses the difference between two softmax attention maps for each head to cancel attention noise; The output of the differential path is shown below: （6） , , V= X Among them, query vector Key vector Value vector V = X , λ is a parameter, and λ is a learnable scalar; The learnable scalar λ is shown below: (7) In the formula, It is a learnable scalar. (0,1) is a constant used to initialize λ.

Images