A building envelope heat flow prediction method based on a KAN-film fusion model

Abstract

The present application relates to a kind of based on KAN-Film fusion model building envelope heat flow prediction method, comprising the following steps: 1), obtain the historical indoor and outdoor air temperature time series data of building envelope, and the thermal resistance, heat capacity of envelope, form training data set;2), model construction: model includes KAN network layer and Film modulation layer two parts, through KAN network layer and Film modulation layer collaborative constitute heat flow prediction model;3), model training: input KAN-Film model with training data, use mean square error as loss function to calculate the difference between predicted value and actual monitoring heat flow, and iteration optimization model parameter;4), heat flow prediction using the model after training: according to new outdoor environmental condition and envelope parameter input, calculate corresponding envelope heat flow prediction value.The present application can improve the accuracy, generalization and real-time of building envelope heat flow prediction, meet the demand of intelligent building energy consumption optimization control.

Description

TECHNICAL FIELD

[0001] The present application relates to a heat flow prediction method, in particular to a building envelope heat flow prediction method based on a KAN-Film fusion model, which is suitable for building energy saving design and intelligent control, and belongs to the field of building energy consumption prediction and artificial intelligence technology. BACKGROUND

[0002] Building envelope (such as wall) as the main barrier of indoor and outdoor heat exchange, its dynamic heat transfer process has a significant impact on the overall energy consumption of the building. Accurate prediction of the heat flow change of the building envelope is not only the basis for building energy efficiency evaluation, but also the key to realizing intelligent control and energy saving operation of air conditioning system. However, due to the combined action of multiple factors on the heat transfer process of building envelope, including outdoor temperature, solar radiation, indoor temperature and thermal performance of wall materials, the dynamic response of heat flow is complex and changeable, which makes the heat flow prediction a technical difficulty in the field of building physics and energy management for a long time.

[0003] Traditional methods mainly establish analytical models based on heat transfer mechanism, such as response factor method, thermal resistance-thermal capacity equivalent circuit model, etc., although they have certain physical explanation, but usually rely on complete thermal parameters acquisition and complex mathematical derivation, which is difficult to meet the requirements of real-time and adaptability of intelligent control system.

[0004] In recent years, data-driven machine learning methods (such as artificial neural network ANN, support vector machine SVM, etc.) have been gradually applied in the field of building energy consumption prediction, which learns the heat flow law through historical data. However, such models generally have the problems of large dependence on data volume, lack of physical constraints and poor interpretability. Once applied to scenarios other than training data (such as different climate conditions or wall structures), the prediction accuracy often decreases significantly, and the extrapolation ability is seriously insufficient. Considering the high diversity of building structures and use environment, it is actually difficult to provide sufficient training samples for each building, which greatly limits the popularization and application of pure data-driven methods in engineering practice.

[0005] To address these issues, some recent research has attempted to combine physical prior knowledge with data-driven models. For example, Reference 1 (“Physics-Informed Neural Networks for Heat Transfer Problems,” S. Cai, Z. Wang et al., Journal of Heat Transfer, Vol. 143, No. 6: 060801, 202106) describes the PINN method, which embeds physical boundary conditions into neural networks. This method improves the physical consistency of machine learning methods for heat transfer problems. However, this method is complex and difficult to effectively incorporate high-order thermal parameters, making it difficult to achieve universal predictions for arbitrary enclosure structures. Reference 2 (“A data-driven energy consumption prediction model using multilayer perceptron for building thermal performance,” K. Zhang, Y. Li et al., Energy, Vol. 272: 127078, 202303) proposes an energy consumption prediction method based on a multilayer perceptron (MLP) model. While this method performs well within the same climate region, its prediction error increases significantly when faced with changes in wall structure or cross-regional migration, exposing its limited generalization capabilities.

[0006] In summary, the existing technologies for thermal flow prediction of building envelope structures mainly have the following problems: first, the lack of physical constraints leads to insufficient model generalization ability; second, traditional physical models are difficult to quickly adapt to changing working conditions; third, high-precision predictions often rely on a large amount of historical data, which limits their practical application effect.

[0007] Therefore, in order to solve the technical problems of poor model extrapolation ability, insufficient physical consistency and high data demand in the existing technology, it is necessary to provide an innovative building envelope structure thermal flow prediction method based on the KAN-Film fusion model to overcome the above-mentioned defects in the existing technology. Summary of the Invention

[0008] The application aims to provide a building envelope heat flow prediction method based on a KAN-Film fusion model, which combines the structural characteristics of a Kolmogorov-Arnold Network (KAN) and a Feature-wise Linear Modulation (Film) mechanism, introduces prior knowledge of thermal parameters of a building envelope, and improves the accuracy, generalization and real-time performance of building envelope heat flow prediction to meet the actual needs of intelligent building energy consumption optimization control.

[0009] To achieve the above-mentioned purpose, the technical solution adopted by the application is as follows: a building envelope heat flow prediction method based on a KAN-Film fusion model, which comprises the following process steps:

[0010] 1) Data acquisition and preprocessing: obtaining historical outdoor air temperature time series data of a building envelope , indoor air temperature time series data , and thermal resistance R and heat capacity C thermal parameters of the envelope, forming a training data set for the prediction model;

[0011] 2) KAN-Film model construction: predefining the prediction model structure based on the KAN network principle, which includes a KAN network layer and a Film modulation layer, wherein the KAN network layer is used to extract the mapping relationship between the temperature data in step 1) and , and generate a basic prediction value; the Film modulation layer is based on the wall thermal resistance R and the heat capacity C, and the output of the activation function is scaled and offset modulated in the feature dimension to realize dynamic adaptation to the thermal structure change; the KAN network layer and the Film modulation layer cooperatively constitute a heat flow prediction model;

[0012] 3) Model training: input the training data obtained in step 1) into the KAN-Film model, use the mean square error as the loss function to calculate the difference between the prediction value and the actual monitored heat flow, and iteratively optimize the model parameters;

[0013] 4) Heat flow prediction using the trained KAN-Film prediction model: according to the new outdoor environmental conditions and envelope parameters, the corresponding envelope heat flow prediction value is calculated, wherein the Film modulation layer dynamically adjusts the KAN network layer output according to the new envelope parameters to realize the adaptation of the model to different envelopes.

[0014] The building envelope heat flow prediction method based on the KAN-Film fusion model further comprises that in step 1), the training data set is normalized, which specifically shows that:

[0015]

[0016]

[0017]

[0018] in, is the model training data set, and N is the number of training set subsamples;

[0019] and are the input and output vectors of the i-th training sample;

[0020] is the outdoor comprehensive temperature time series vector in the i-th training sample;

[0021] is the indoor air temperature time series vector in the i-th training sample;

[0022] is the wall heat flow time series vector in the i-th training sample;

[0023] , are the thermal resistance and thermal capacity values ​​in the i-th training sample.

[0024] The method for predicting heat flow of building envelope structures based on the KAN-Film fusion model of the present invention is further as follows: in the step 2), the KAN network layer adopts a single hidden layer structure with a variable width, including N input nodes and 1 output node, and the input nodes correspond to the outdoor equivalent temperature and indoor temperature data of N pre-selected time steps.

[0025] The heat flow prediction method for building envelope structures based on the KAN-Film fusion model of the present invention further comprises: in step 2), the heat flow prediction model calculates the heat flow data as follows:

[0026]

[0027] in, Predict wall heat flux data for the KAN-Film model;

[0028] and is the modulation factor of the Film layer;

[0029] Represents the KAN network layer's response to the input temperature time series data Output.

[0030] The method for predicting heat flow of building envelope structures based on the KAN-Film fusion model of the present invention further comprises: in step 3), the loss function is specifically:

[0031]

[0032] in, is the loss function;

[0033] The number of samples in the training batch;

[0034] Predict heat flow values ​​for the model;

[0035] is the actual heat flow value.

[0036] The heat flow prediction method of building envelope structure based on KAN-Film fusion model of the present invention is further as follows: in step 3), Adam optimization algorithm is used to iteratively update all learnable parameters, including the weight parameters of KAN network layer. and Film modulation network parameters ; Each training iteration step first calculates the current network output , and then calculate the loss function And the gradient of the loss function with respect to each parameter, and update the parameters according to the following algorithm rules:

[0037]

[0038]

[0039] in: is the weight parameter of the KAN network layer;

[0040] is the weight parameter of the Film network layer;

[0041] is the learning rate of the Adam algorithm;

[0042] and They are first-order moment estimation and second-order moment estimation respectively;

[0043] θ is the gradient; is the gradient θ update;

[0044] To prevent division by zero.

[0045] The building envelope heat flow prediction method based on the KAN-Film fusion model of the present application further comprises: the iteration optimization of the parameters in the model includes the weight function parameters in the KAN layer and the modulation parameters of the Film layer.

[0046] The building envelope heat flow prediction method based on the KAN-Film fusion model of the present application can also be: after the model training in step 3) is completed, the performance of the model in the small sample learning scene is evaluated; then, the trained model is applied to an independent test data set and external working conditions different from the training conditions to verify its generalization performance for unknown inputs.

[0047] Compared with the prior art, the present application has the following beneficial effects:

[0048] 1. Structure generalization: can accurately predict the wall heat flow under different thermal resistance and heat capacity combinations.

[0049] 2. Physically interpretable: the response function embedding is mapped as a network weight, which has a clear physical meaning.

[0050] 3. Strong small sample learning ability: 80% accuracy can be achieved when the training data is only 20%.

[0051] 4. High computational efficiency: single-layer structure model significantly reduces the computational resource consumption compared to traditional deep network.

[0052] 5. Greatly improve the generalization ability of the model and reduce the data requirement: by integrating physical parameters into the network structure, the model can adapt to changes in different regional climates and building envelopes, and still maintain high prediction accuracy in unseen working conditions. At the same time, due to the integration of physical prior knowledge, the model can also converge to good performance with only a small amount of training data, which is better than traditional black box models that rely on a large number of samples.

[0053] 6. Ensure physical consistency and interpretability: the present application pre-embeds the physical model form of building envelope heat transfer, and the KAN network layer uses linear activation to make the network structure equivalent to the first-order response factor model, thereby ensuring that the model output has a physically linear relationship with temperature changes. The unreasonable results that violate the law of heat conduction will not appear. Modulation parameters and respectively reflect the influence of different wall structures on the amplitude and reference value of the heat flow, so that the model has certain physical interpretability and is easier for engineers to understand and trust.

[0054] 7. Reduce data requirements and computational cost: the model structure of the present application is simple, the number of parameters is greatly reduced, and the training and inference calculation overhead is lower, which is suitable for real-time control scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 Schematic diagram of factors affecting the heat flow of the enclosure structure in step 1) of the present invention.

[0056] Figure 2 This is a flow chart of the KAN-Film model construction in step 2) of the present invention.

[0057] Figure 3 This is the prediction accuracy comparison result between the KAN-Film model of the present invention and the traditional MLP model under small sample training conditions.

[0058] Figure 4 It is the performance of the KAN-Film model of the present invention in extrapolating predictions on wall structures with different thermal capacities.

[0059] Figure 5 This is a comparison chart of the heat flow prediction performance of the KAN-Film model, MLP, and LSTM models of the present invention after cross-training and testing on different wall thickness datasets.

[0060] Figure 6 It is the meteorological data used in the specific implementation case of the present invention.

[0061] Figure 7 It is a model output result diagram of a specific implementation case of the present invention. DETAILED DESCRIPTION

[0062] The present invention is a method for predicting heat flow of building envelope structures based on the KAN-Film fusion model, which includes the following process steps:

[0063] 1) Data acquisition and preprocessing: Since the heat flow of building envelope is affected by factors such as temperature, radiation and wall thermal parameters, the historical outdoor air temperature (solar radiation has been converted into equivalent temperature) time series data of the building envelope is obtained. , indoor air temperature time series data , as well as the thermal parameters of the enclosure structure such as thermal resistance R and heat capacity C, form the training data set of the prediction model.

[0064] Please refer to the instruction manual Figure 1As shown, it demonstrates the process of heat flow affected by wall structure, orientation, radiation, climate and indoor settings, etc. To build a heat flow prediction model with structural generalization ability, the system collects multiple sets of building envelope heat flow data covering different climate conditions, wall structures, orientation and use seasons in the data preparation stage. The data collection content specifically includes: different types of envelopes (such as concrete + EPS insulation layer, aerated concrete, double gypsum board + SPF insulation layer, etc.), a wide range of combinations of corresponding thermal resistance and heat capacity parameters; different azimuth orientations (east, south, west, north), corresponding to different angles of solar radiation incident intensity; different climate zones and seasons (including winter heating period and summer cooling period); different indoor temperature setting values and modes; each sample contains time series measured equivalent outdoor temperature , indoor temperature (including solar radiation processing). The above elements jointly constitute the comprehensive sample input space required for network training.

[0065] All original temperature and heat flow data are uniformly sampled, time-aligned, and constitute the training sample set of KAN-Film model: that is, the training data set is normalized, which is specifically manifested as:

[0066]

[0067]

[0068]

[0069] Wherein, is the model training data set, and N is the number of training set sub-samples;

[0070] and is the input and output vector of the i-th training sample;

[0071] is the outdoor comprehensive temperature time series vector in the i-th training sample;

[0072] is the indoor air temperature time series vector in the i-th training sample;

[0073] is the wall heat flow time series vector in the i-th training sample;

[0074] , is the thermal resistance and heat capacity value in the i-th training sample.

[0075] 2), KAN-Film model construction. Please refer to the attached Figure 2As shown in the figure, it shows the entire process from acquiring historical meteorological data and indoor temperature, input construction, neural network prediction, Film physical modulation, to the final output heat flow prediction value.

[0076] Specifically, the present invention predefines a prediction model structure based on the KAN network principle. The model includes two parts: the KAN network layer and the Film modulation layer. The two are parallel structures for receiving and processing model input samples. The KAN network layer is used to extract the temperature data of the input step 1). and The mapping relationship in [ ] is used to generate basic prediction values. The Film modulation layer uses the wall thermal resistance R and heat capacity C, and its activation function output performs scaling and offset modulation on the feature dimension to achieve dynamic adaptation to changes in thermal structure. The KAN network layer and the Film modulation layer work together to form a heat flow prediction model with interpretability and externalization capabilities.

[0077] Among them, the KAN network layer embeds the prior physical knowledge of the response factor method (Response Factor Method) of dynamic heat transfer of walls during the construction process. Specifically, the heat flow of the wall is not an instantaneous response to the temperature difference between indoor and outdoor, but a time-delay response process determined by the temperature difference at current and multiple historical moments in the form of linear convolution. Based on the above physical model, the KAN network is designed as a single hidden layer structure with variable width, including N input nodes and 1 output node. The input node corresponds to the outdoor equivalent temperature and indoor temperature data of N pre-selected time steps; the activation function of the KAN network is fixed as a linear function (Linear function) to convert the indoor temperature and outdoor equivalent temperature The time series delay characteristics of the wall are used as the dominant variable input. By mapping these historical characteristics, the network can achieve a preliminary prediction of the wall heat flow.

[0078] The film modulation layer receives the thermal resistance R, heat capacity C and other physical parameters of the enclosure structure as modulation input, and dynamically modulates the output of the KAN network layer through a shallow auxiliary multi-layer perceptron (MLP) neural network. The auxiliary network takes R and C as input and outputs a set of scaling factors. and offset factor By modulating the output of each activation function in the KAN network layer, the model can automatically adjust its output as the thermal performance of the building envelope changes. The model output is the predicted dynamic heat flow of the wall.

[0079] Based on the above theoretical basis, the specific calculation method of the heat flow prediction model for heat flow data is as follows:

[0080]

[0081] in, Predict wall heat flux data for the KAN-Film model;

[0082] and is the modulation factor of the Film layer;

[0083] Represents the KAN network layer's response to the input temperature time series data Output.

[0084] 3) Model training: Input the training data obtained in step 1) into the KAN-Film model constructed in step 2), use the mean square error as the loss function to calculate the difference between the predicted value and the actual monitored heat flow, and iteratively optimize the model parameters. The loss function is specifically:

[0085]

[0086] in, is the loss function;

[0087] The number of samples in the training batch;

[0088] Predict heat flow values ​​for the model;

[0089] is the actual heat flow value.

[0090] In order to ensure the stability of network training and improve the generalization ability of the model, the present invention uses the Z-score standardization method to standardize each training sample. Temperature time series data in Normalization is performed as follows:

[0091]

[0092] in: For variables The mean of For variables The standard deviation of is the normalized value.

[0093] Furthermore, during iterative optimization, the present invention uses the Adam optimization algorithm to iteratively update all learnable parameters, including the weight parameters of the KAN network layer. and Film modulation network parameters ; Each training iteration step first calculates the current network output , and then calculate the loss function And the gradient of the loss function with respect to each parameter, and update the parameters according to the following algorithm rules:

[0094]

[0095]

[0096] in: is the weight parameter of the KAN network layer;

[0097] is the weight parameter of the Film network layer;

[0098] is the learning rate of the Adam algorithm;

[0099] and They are first-order moment estimation and second-order moment estimation respectively;

[0100] θ is the gradient; is the gradient θ update;

[0101] To prevent division by zero.

[0102] Furthermore, to avoid overfitting, the L1 regularization strategy is introduced in the above training process: sparse constraints are imposed on the network parameters, and an early stopping strategy is used to continuously monitor the loss changes on the validation set. If there is no improvement within several consecutive rounds, the training is terminated early to avoid overfitting.

[0103] After completing model training in step 3), further model validation and adjustment can be performed. This involves evaluating the model's performance in small-sample learning scenarios to address the lack of sensor data common in the construction industry. By limiting the number of training samples, the model's learning ability and prediction accuracy can be tested with only a small amount of data. The trained model is then applied to independent test datasets and external working conditions different from those used in the training process to verify its generalization to unknown inputs, with particular attention paid to the model's adaptability and prediction stability when wall structural parameters (such as thickness and heat capacity) vary.

[0104] To fully verify the practicality, robustness and structural extrapolation ability of the method of the present invention, a systematic performance evaluation experiment was carried out on the constructed KAN-Film model under various data conditions, covering three aspects: small sample learning ability, structural extrapolation ability and cross-structure transfer generalization ability.

[0105] 1. Small sample training and verification

[0106] In the case of limited training data, the prediction accuracy performance of the proposed model and the traditional MLP model is compared and analyzed. Figure 3 The paper presents a comparison of the prediction accuracy of the KAN-Film model and the traditional MLP model under small-sample training conditions. The results show that when the training set only accounts for 10% of the total data, KAN-Film maintains low error output, while the MLP model's prediction error increases significantly. This demonstrates the proposed method's strong small-sample learning capabilities, meeting the practical needs of building thermal monitoring applications where data is scarce.

[0107] 2. Structural extrapolation capability test

[0108] Please refer to the instruction manual Figure 4 As shown in Figure 2, in order to verify the generalization performance of the model on unseen structural parameter combinations, a set of heat capacity parameters and wall structure samples that did not appear in the training set were selected for extrapolation prediction. Figure 4 The results show that KAN-Film can accurately capture the changing trends of wall heat flux, demonstrating its predictive ability under unknown structural conditions and verifying the externalization advantages brought by the model's embedded physical knowledge. Even if the test data includes a combination of thermal resistance and heat capacity that the model has never seen, KAN-Film can still accurately predict the wall heat flux changes, indicating that it has good structural extrapolation capabilities.

[0109] 3. Evaluation of cross-structure transfer learning capabilities

[0110] Please refer to the instruction manual Figure 5 As shown in the figure, cross-training and testing experiments were conducted on datasets of multiple wall thicknesses to compare the generalization performance of KAN-Film with mainstream models such as MLP and LSTM. Figure 5 The figure shows the migration prediction errors between wall structures of different thicknesses (240mm, 250mm, 370mm, and 480mm). The figure presents the coefficient of determination (R²) of different models for each dataset combination in the form of a thermal matrix. The results show that KAN-Film outperforms the comparison models in all structural combinations, especially for high-thermal-capacity thick-wall structures. Its prediction error is significantly lower than that of LSTM and MLP, demonstrating its superior structural migration capabilities and prediction accuracy, further validating its adaptability to complex structural conditions.

[0111] In summary, the KAN-Film model demonstrates superior modeling performance across multiple dimensions, particularly in real-world architectural scenarios with scarce samples and diverse structures. It demonstrates extrapolation capabilities and stability far exceeding those of traditional data-driven methods, fully aligning with the "physics-driven + data fusion" modeling concept proposed in this paper.

[0112] 4), and the heat flow prediction is performed by using the trained KAN-Film prediction model: according to the new outdoor environment conditions and the building envelope parameters, the corresponding building envelope heat flow prediction value is calculated, wherein the Film modulation layer dynamically adjusts the KAN network layer output according to the new building envelope parameters, so as to realize the adaptation of the model to different building envelopes.

[0113] Further, the obtained building envelope heat flow prediction value can be input into a building air conditioning control system as a reference signal of model predictive control, so as to guide the optimized adjustment of the air conditioning system, thereby reducing energy consumption fluctuation and improving indoor temperature control accuracy.

[0114] The KAN-Film prediction model of the present application is universal for different building structures and climate conditions, and can complete the prediction of a new building envelope by inputting corresponding thermal resistance and heat capacity parameters without modifying the model structure, wherein for a new building envelope working condition lacking measured data, model parameter correction can be completed by adding a small amount of sample data of the working condition in the model training set, so that the prediction error is controlled within the allowable range.

[0115] In order to more clearly show the application process and performance of the method of the present application in the actual building heat flow prediction task, the following will take the measured data of a small house from August 10, 2024 to August 19, 2024 as an example to explain the implementation steps and application effects of the present application in detail.

[0116] First, the collected outdoor air temperature and solar radiation intensity (test data are shown in the attached table) are calculated to obtain the outdoor comprehensive temperature. Figure 6 The red curve in the figure is the outdoor air temperature, and the blue curve is the solar radiation intensity.

[0117] The air temperature in the experimental house is selected as the indoor thermal boundary condition input, and the processed outdoor comprehensive temperature is used as the time series input vector together with the processed outdoor comprehensive temperature for the feedforward calculation of the heat flow prediction model. In addition, the physical thermal parameters of the building envelope of the experimental house are known, wherein the wall thermal resistance is about 1.25 m²K / W, and the heat capacity is 612000 J / m²K. The above thermal resistance and heat capacity are input into the Film layer of the KAN-Film model as modulation factors.

[0118] During the model execution phase, the constructed KAN-Film neural network model predicts the feature input at each time point and outputs the corresponding instantaneous wall heat flux value q(t). This heat flux prediction represents the thermal conductivity per unit area of ​​the building envelope and is a core parameter in building thermal environment regulation. The prediction process, as described in the model structure above, begins with the KAN network layer mapping the temperature time series to a preliminary heat flux estimate. The Film modulation layer then receives the R and C parameters and applies scaling and offset transformations, ultimately outputting a heat flux prediction that incorporates physical information.

[0119] In order to verify the effectiveness of the method proposed in this invention, the heat flow prediction value output by the model is compared point by point with the heat flow data measured in the experiment. Figure 7 As shown in the figure, the horizontal axis represents time (in hours), the vertical axis represents the heat flow value per unit area (in W / m²), the blue continuous curve is the measured heat flow data, and the black discontinuous curve is the prediction result of the model of the present invention. Figure 7 The wall heat flow curve predicted by the KAN-Film model is presented and compared with the measured data to verify the applicability and accuracy of the model under actual working conditions.

[0120] The figure clearly shows that the proposed KAN-Film model accurately predicts the dynamic heat flux trends of the wall in most time periods, especially during periods of intense heat flux fluctuations, such as the morning heating phase and the evening cooling phase. The model responds quickly, has a high degree of trend fit, and has a very small peak prediction error, demonstrating excellent dynamic tracking capabilities. Statistical analysis shows that the model's mean absolute error (MAE) over the entire test period is less than 0.43 W / m², and the correlation coefficient R 2 Reached 0.97.

[0121] The above specific implementation methods are only preferred embodiments of this creation and are not intended to limit this creation. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of this creation should be included in the scope of protection of this creation.

Claims

1. A method for predicting heat flow of building envelope structures based on the KAN-Film fusion model, characterized by: The process steps include: 1) Data acquisition and preprocessing: Obtain historical outdoor air temperature time series data of the building envelope , indoor air temperature time series data , as well as the thermal resistance R and heat capacity C thermal parameters of the enclosure structure, to form a training data set for the prediction model; 2) KAN-Film model construction: Based on the KAN network principle, the prediction model structure is predefined. The model consists of two parts: the KAN network layer and the Film modulation layer. The KAN network layer is used to extract the temperature data of the input step 1). and The mapping relationship in the [15] is used to generate basic prediction values. The Film modulation layer uses the wall thermal resistance R and heat capacity C as the basis, and the output of its activation function performs scaling and offset modulation on the feature dimension to achieve dynamic adaptation to changes in thermal structure. The KAN network layer and the Film modulation layer work together to form a heat flow prediction model. The heat flow prediction model calculates the heat flow data as follows: in, Predict wall heat flux data for the KAN-Film model; and is the modulation factor of the Film layer; Represents the KAN network layer's response to the input temperature time series data Output; 3) Model training: The training data obtained in step 1) is input into the KAN-Film model, the mean square error is used as the loss function to calculate the difference between the predicted value and the actual monitored heat flow, and the model parameters are iteratively optimized; 4) Use the trained KAN-Film prediction model to predict heat flow: Based on the new outdoor environmental conditions and enclosure parameter inputs, calculate the corresponding enclosure heat flow prediction value. The Film modulation layer dynamically adjusts the KAN network layer output according to the new enclosure parameters to achieve model adaptation to different enclosures.

2. The method for predicting heat flow of building envelope structures based on the KAN-Film fusion model according to claim 1, characterized in that: In step 1), the training data set is normalized, which is specifically performed as follows: in, is the model training data set, and N is the number of training set subsamples; and are the input and output vectors of the i-th training sample; is the outdoor comprehensive temperature time series vector in the i-th training sample; is the indoor air temperature time series vector in the i-th training sample; is the wall heat flow time series vector in the i-th training sample; , are the thermal resistance and thermal capacity values ​​in the i-th training sample.

3. The method for predicting heat flow of building envelope structures based on the KAN-Film fusion model according to claim 1, characterized in that: In step 2), the KAN network layer adopts a single hidden layer structure with variable width, including N input nodes and 1 output node, and the input node corresponds to the outdoor equivalent temperature and indoor temperature data of N pre-selected time steps.

4. The method for predicting heat flow of building envelope structures based on the KAN-Film fusion model according to claim 1, characterized in that: In step 3), the loss function is specifically: in, is the loss function; The number of samples in the training batch; Predict heat flow values ​​for the model; is the actual heat flow value.

5. The method for predicting heat flow of building envelope structures based on the KAN-Film fusion model according to claim 4, characterized in that: In step 3), the Adam optimization algorithm is used to iteratively update all learnable parameters, including the weight parameters of the KAN network layer. and Film modulation network parameters ; Each training iteration step first calculates the current network output , and then calculate the loss function And the gradient of the loss function with respect to each parameter, and update the parameters according to the following algorithm rules: in: is the weight parameter of the KAN network layer; is the weight parameter of the Film network layer; is the learning rate of the Adam algorithm; and They are first-order moment estimation and second-order moment estimation respectively; θ is the gradient; is the gradient θ update; To prevent division by zero.

6. The method for predicting heat flow of building envelope structures based on the KAN-Film fusion model according to claim 4, characterized in that: In step 3), the parameters in the iterative optimization model include the weight function parameters in the KAN layer and the modulation parameters in the Film layer.

7. The method for predicting heat flow of building envelope structures based on the KAN-Film fusion model according to claim 1, characterized in that: After completing the model training in step 3), evaluate the model's performance in a small-sample learning scenario. Subsequently, apply the trained model to an independent test dataset and external working conditions different from the training conditions to verify its generalization performance to unknown inputs.

Images