Building heat load prediction method and system, readable storage medium and program product

Abstract

The invention relates to the technical field of machine learning, in particular to a building thermal load prediction method and system, a readable storage medium and a program product, and the method comprises the steps: extracting the feature information of building thermal load data; training an initial prediction model according to the feature information to obtain a target prediction model; the initial prediction model is constructed based on a large language model and a graph neural network; and building thermal load prediction is carried out according to the target prediction model. According to the technical scheme provided by the invention, a deep space-time analysis capability is introduced for building thermal load prediction by fusing a large language model based on Transform introducing an Attention mechanism and a graph neural network, and Bayesian optimization, an Adam optimizer, an advanced regularization technology and a two-stage training strategy are combined, so that the model training efficiency is improved, and the building thermal load prediction efficiency is improved. And the overall prediction precision and the generalization ability of the model are improved by finely adjusting the parameters of each stage, and overfitting is effectively prevented.

Description

Technical Field

[0001] The present disclosure relates to the technical field of machine learning, and particularly relates to a building heat load prediction method, system, readable storage medium and program product. Background Art

[0002] For existing buildings, reasonable energy planning and management to reduce operating energy consumption are the keys to building energy conservation. Traditional building heat load prediction methods are mainly divided into two categories: mechanism modeling and data-driven modeling. Mechanism modeling methods have high theoretical knowledge thresholds and significant engineering quantities when dealing with complex systems. In addition, the lack of information data such as building characteristics and heat pipe materials, as well as the great difficulty and low accuracy in collecting heat user social behavior data, will all lead to deviations between the simulation model and the actual physical system, and the calculation amount is extremely large in large-scale applications. Data-driven modeling methods establish the mapping relationship between input and output by inputting a large number of sample data into the model, so as to predict the future system behavior. Such methods are applicable to situations where data is rich and the internal structure of the system is complex or unknown. Especially in the case of lacking accurate physical parameter values, data-driven models show good adaptability and flexibility.

[0003] Although data-driven methods have achieved certain success in practical applications, there are still the following main problems:

[0004] (1) Strong data dependence: The performance of data-driven models depends greatly on the quality and quantity of input data. Incomplete, incorrect or deviated data will significantly affect the model accuracy.

[0005] (2) Insufficient generalization ability: Traditional data-driven models such as support vector machines (SVM) and artificial neural networks (ANN) perform well on specific data sets, but their adaptability to new environments or changing conditions that have not been seen before is limited.

[0006] (3) Poor model interpretability: Many efficient data-driven models, such as deep learning models, are often criticized as "black box" models. The opacity of the model decision-making process makes it difficult for users to trust the prediction results of the model.

[0007] Existing building heat load prediction methods often show insufficient processing and interpretation capabilities when dealing with complex time series data, which affects the prediction accuracy and the generalization ability of the model. In addition, different buildings and environmental conditions have adaptive requirements for the prediction model, but the adjustability and flexibility of existing technologies are limited and it is difficult to meet the diverse needs of practical applications. Therefore, there is an urgent need for a building heat load prediction method that can effectively process complex data, improve prediction accuracy and enhance model adaptability to improve the intelligent level of building energy management. Summary of the Invention

[0008] The present disclosure aims to solve at least one of the technical problems in the above technologies to this end, a building heat load prediction method is proposed, including:

[0009] Extracting the characteristic information of the building heat load data;

[0010] Training an initial prediction model according to the characteristic information to obtain a target prediction model; the initial prediction model is constructed based on a large language model and a graph neural network;

[0011] Performing building heat load prediction according to the target prediction model.

[0012] Further, the building heat load data includes: historical heat load data, meteorological data, building characteristic data, heating system data, and user behavior data.

[0013] Further, extracting the characteristic information of the building heat load data includes:

[0014] Extracting the time characteristics of the building heat load data based on the large language model;

[0015] Extracting the spatial characteristics of the building heat load data through the graph neural network;

[0016] Performing feature fusion on the time characteristics and the spatial characteristics.

[0017] Further, extracting the time characteristics of the building heat load data based on the large language model includes: extracting the long-term trend characteristics and periodic characteristics of the building heat load data through multi-scale pooling filtering; the multi-scale pooling filtering has the corresponding expression:

[0018] Further, the large language model contains an Attention mechanism; wherein,

[0019] The expression of the Attention mechanism is:

[0020]

[0021] wherein, Q represents a query vector; K represents a key vector; V represents a value vector; d k represents the dimension of the key vector; T represents the transpose of a matrix.

[0022] Further, the graph neural network is configured to: construct the dependency relationship between buildings through an adjacency matrix, and the corresponding expressions include:

[0023]

[0024] wherein, A i,jRepresents the weight relationship between building i and building j; dv i , v j Represents building v i and building v j The distance between them; σ represents the standard deviation of the Gaussian kernel; ò is used to control the sparsity.

[0025] Furthermore, training the initial prediction model according to the feature information includes: adjusting the hyperparameters of the initial prediction model based on Bayesian optimization.

[0026] The present disclosure also proposes a building heat load prediction system, including:

[0027] A feature extraction module, configured to:

[0028] Extract the feature information of the building heat load data;

[0029] A model training module, configured to:

[0030] Train an initial prediction model according to the feature information to obtain a target prediction model; the initial prediction model is constructed based on a large language model and a graph neural network;

[0031] A model application module, configured to:

[0032] Perform building heat load prediction according to the target prediction model.

[0033] The present disclosure also proposes a computer-readable storage medium, in which a computer program or instruction is stored. When the computer program or instruction is executed by a processor, it is at least used to implement the above-mentioned building heat load prediction method.

[0034] The present disclosure also proposes a computer program product, which is stored in a computer-readable storage medium. When the computer program product is executed by a processor, it is at least used to implement the above-mentioned building heat load prediction method.

[0035] Compared with the prior art, the beneficial effects of the present disclosure are:

[0036] (1) Innovative integration of deep spatio-temporal sequence modeling: By integrating a large language model based on Transformer and a graph neural network, deep spatio-temporal analysis capabilities are introduced for building heat load prediction. This combination not only uses the large language model to capture long-term dependencies in the time series, but also details the spatial interactions between buildings through the graph neural network, enabling the prediction model to comprehensively process the complex interaction relationships of time and space. This method overcomes the deficiencies of traditional methods in capturing long time spans and spatial dependencies between buildings.

[0037] (2) Intelligent Fusion of Multi-source Data: By integrating various influencing factors such as meteorological conditions, building physical properties, heating system operation, and user behavior patterns through advanced data processing techniques, complex factors affecting building heat load are comprehensively captured, forming a comprehensive data view, which significantly enhances the response speed of the model to environmental changes and prediction accuracy.

[0038] (3) Multi-scale Feature Decomposition and Attention Mechanism: The heat load sequence is decomposed into trend components and seasonal components through multi-scale pooling filters and modeled separately to reduce complexity. At the same time, the Attention mechanism is introduced to dynamically allocate weights according to the characteristics of the time series, enabling the model to focus on the most critical parts for prediction, thereby improving the prediction accuracy.

[0039] (4) Real-time Data Feedback and Model Self-optimization Mechanism: By continuously collecting new building operation data and updating model parameters in real time, this disclosure ensures continuous performance improvement and adaptability enhancement of the prediction system, which is particularly suitable for dynamically changing building environments.

[0040] (5) Efficient Model Training and Optimization Strategies: Combining Bayesian optimization, Adam optimizer, advanced regularization techniques, and two-stage training strategies not only improves the efficiency of model training, but also enhances the overall prediction accuracy and model generalization ability by fine-tuning parameters in each stage, and effectively prevents overfitting.

[0041] Other features and advantages of this disclosure will be described in the subsequent specification, and some of them will become obvious from the specification or be understood by implementing this disclosure. The objectives and other advantages of this disclosure can be achieved and obtained through the structures specifically pointed out in the written specification and the drawings. The technical solutions of this disclosure will be further described below through the drawings and embodiments. Description of the Drawings

[0042] The drawings are used to provide further understanding of this disclosure and constitute a part of the specification. They are used together with the embodiments of this disclosure to explain this disclosure and do not constitute a limitation to this disclosure. In the drawings:

[0043] Figure 1 Schematic diagram of the building heat load prediction method given for the embodiment;

[0044] Figure 2 Architecture diagram of the building heat load prediction system given for the embodiment;

[0045] Figure 3 Schematic diagram of the building heat load prediction system given for the embodiment;

[0046] Figure 4 Schematic diagram of the electronic device given for the embodiment;

[0047] Figure 5 Schematic diagram of a computer-readable storage medium given for the embodiment. Detailed implementation manners

[0048] The present disclosure will be described below with reference to the accompanying drawings. The preferred embodiments described herein are only for illustrating and explaining the present disclosure, and are not used to limit the present disclosure.

[0049] As Figure 1 shown, the present disclosure provides a building heat load prediction method, including:

[0050] Extracting the feature information of the building heat load data;

[0051] Training an initial prediction model according to the feature information to obtain a target prediction model; the initial prediction model is constructed based on a large language model and a graph neural network;

[0052] Performing building heat load prediction according to the target prediction model.

[0053] According to some embodiments of the present disclosure, the building heat load prediction method includes:

[0054] (1) Training data acquisition: Collecting various data related to the building heat load, and cleaning and normalizing the collected data;

[0055] (2) Feature engineering: Extracting the key factors affecting the building heat load;

[0056] (3) Sequence modeling based on a large language model: Using a large language model to perform sequence modeling on the building heat load, and finding out the most influential part for the current prediction by calculating the weights of each input feature. In the sequence modeling based on the large language model, a graph neural network (GCN) is used to construct the spatial dependence relationship between buildings, capture the spatial correlation between buildings, especially the mutual influence of heat loads between adjacent buildings.

[0057] (4) Model training and optimization: Using the Adam optimizer to train the model, and adjusting the weights of the model to minimize the loss function. A two-stage training strategy is adopted, that is, first performing autoregressive fine-tuning to enable the model to better adapt to the time characteristics of the building heat load data, and then further optimizing the model performance through supervised learning.

[0058] Furthermore, the building heat load data includes: historical heat load data, meteorological data, building characteristic data, heating system data, and user behavior data.

[0059] According to some embodiments of the present disclosure, the training data acquisition includes the following steps:

[0060] Collect a variety of data related to building heat load, including: historical heat load data, meteorological data, building characteristic data, heating system data, and user behavior data. Among them, historical heat load data can reflect the thermal energy usage of the building over a past period; meteorological data includes temperature, humidity, wind speed, and air pressure, and these environmental factors have an important impact on building energy consumption; building characteristics such as wall insulation, window area, and building orientation, and these data can reflect the heat flow and dissipation characteristics of the building; heating system data includes supply water temperature and valve status, reflecting the working condition of the heating system; user behavior data records the heat usage habits and frequencies of users at different time periods, which helps to determine the changes in heat load demand.

[0061] Further, before extracting the feature information of the building heat load data, it also includes:

[0062] Perform data cleaning and normalization on the building heat load data.

[0063] According to some embodiments of the present disclosure, in order to ensure the quality of the data, the collected data is cleaned and standardized, including:

[0064] Normalization processing: Convert the data to the range of 0 to 1 to eliminate the dimension difference.

[0065] The corresponding expression for data normalization is as follows:

[0066]

[0067] Among them, X norm represents the normalized sample, X represents the sample to be normalized, X min represents the minimum value in the sample set where X is located, X max represents the maximum value in the sample set where X is located. Through data normalization processing, the model can maintain a consistent scale for all input data, thus better performing learning and prediction.

[0068] Data cleaning: Remove outliers and fill in missing values. For example, if the measurement data on a certain day deviates abnormally from other data, it may need to be processed or deleted.

[0069] Further, extracting the feature information of the building heat load data includes:

[0070] Extract the time features of the building heat load data based on the large language model;

[0071] Extract the spatial features of the building heat load data through the graph neural network;

[0072] Perform feature fusion on the time features and the spatial features.

[0073] Further, time features of the building heat load data are extracted based on the large language model, including: extracting long-term trend features and periodic features of the building heat load data through multi-scale pooling filtering.

[0074] Further, the multi-scale pooling filtering has the corresponding expression:

[0075] X T = Softmax(w(X))·f(X)

[0076] X S = X - X T

[0077] where Softmax represents the Softmax activation function; w(X) represents data-driven weights; f(X) represents multiple average pooling filters; X S represents periodic features; X represents the input data; X T represents long-term trend features.

[0078] According to some embodiments of the present disclosure, key factors affecting building heat load are extracted through feature engineering, including: meteorological parameters, such as environmental factors like temperature and humidity; building characteristics, such as wall insulation coefficient and window area; status parameters of the heating system, such as supply water temperature and flow rate; time factors, including time (day / night), weekend / weekday, etc.; user behavior characteristics, and user heat usage habits are extracted through historical data analysis.

[0079] The principle of feature decomposition is as follows:

[0080] The heat load time series data is decomposed into a trend component (long-term change, long-term trend feature) and a seasonal component (periodic change, periodic feature) to better understand and model different change characteristics of building energy consumption.

[0081] The multi-scale pooling filter is used to decompose the time series, so as to capture features on different time scales. The formula is as follows:

[0082] X T = Softmax(w(X))·f(X)

[0083] X S = X - X T

[0084] where Softmax represents the Softmax activation function; w(X) represents data-driven weights; f(X) represents multiple average pooling filters; X S represents the remaining periodic features, obtained by subtracting the long-term trend feature X TObtained; in some embodiments, X is an input feature matrix, including the original features in the time series data; X T represents the extracted long-term trend features, calculated by the Softmax activation function and filters.

[0085] Furthermore, the large language model contains an Attention mechanism; among them,

[0086] The expression of the Attention mechanism is:

[0087]

[0088] where Q represents the query vector; K represents the key vector; V represents the value vector; d k represents the dimension of the key vector; T represents the transpose of the matrix.

[0089] Furthermore, the graph neural network is configured to: construct the dependency relationship between buildings through the adjacency matrix, and the corresponding expressions include:

[0090]

[0091] where A i,j represents the weight relationship between building i and building j; dv i ,v j represents building v i and building v j the distance between; σ represents the standard deviation of the Gaussian kernel; ò is used to control the sparsity.

[0092] According to some embodiments of the present disclosure, as Figure 2 shown, a large language model based on Transformer (such as GPT3) is used in combination with a graph neural network to perform sequence modeling on building heat loads. This combination not only processes time series data, but also improves the prediction accuracy of the model for building heat load changes by introducing a graph neural network for spatial data processing.

[0093] Time series modeling:

[0094] Utilizing the deep self-attention mechanism of the large language model, the present disclosure accurately captures the long-distance dependencies in the time series data. The Attention mechanism processes the input sequence and finds the parts that have the most influence on the current prediction by calculating the weights of each input feature.

[0095] The Attention mechanism is implemented through the following formula:

[0096]

[0097] Among them, Q represents the query vector; K represents the key vector; V represents the value vector; d k represents the dimension of the key vector; T represents the transpose of the matrix. The Attention mechanism calculates the similarity scores at different time points in the input sequence, enabling the model to automatically find the most important inputs for prediction and generate new prediction representations based on these weights. The existence of the Attention mechanism allows the model to dynamically adjust the attention to different time points in the input sequence.

[0098] Spatial relationship modeling:

[0099] The graph neural network (GCN) is applied to capture the spatial dependency relationships between buildings. The spatial connections are expressed through the adjacency matrix, and the calculation method is as follows:

[0100]

[0101] Among them, A i,j represents the weight relationship between building i and building j; d 2 v i , v j represents the square of the distance between building v i and building v j ; σ represents the standard deviation of the Gaussian kernel, which is used to control the smoothness of the adjacency matrix, and ò is used to control the sparsity. In the formula, the Gaussian kernel emphasizes the influence of the distance between buildings through the sum of squares and exponential form, and can effectively express the characteristic that the influence of distance on the heat load decreases rapidly with the increase of distance. This attenuation method is very suitable for simulating the heat exchange or heat influence caused by distance between buildings, especially when the transfer and distribution of heat load are significantly affected by physical distance.

[0102] Spatio-temporal feature integration:

[0103] Spatial hint: GCN captures the mutual influence of heat loads between neighboring buildings and the spatial correlation between buildings by learning spatial relationships. The expression of spatial dependence strengthens the model's ability to analyze data in the physical space.

[0104] Time hint: Embed time information (such as day / night, weekend / weekday) into the model, allowing the model to understand and reflect the specific impact of time variables on the heat load.

[0105] According to some embodiments of the present disclosure, the model given by the present disclosure captures the mutual relationships between buildings by introducing a localized spatial module:

[0106]

[0107] Among them, H(l) is the node feature matrix of the l-th layer, and W(l) is the weight matrix of the l-th layer; is the normalized adjacency matrix, and ReLU is the activation function. Through graph convolution operations, the model can learn the complex spatial dependencies between buildings.

[0108] Further, training the initial prediction model according to the feature information includes: adjusting hyperparameters of the initial prediction model based on Bayesian optimization.

[0109] Further, adjusting hyperparameters of the initial prediction model based on Bayesian optimization includes:

[0110] Randomly select a set of hyperparameters of the initial prediction model;

[0111] Calculate the loss function of the initial prediction model under the selected hyperparameters;

[0112] Update the relationship model between the selected hyperparameters and the loss function through Gaussian process;

[0113] Calculate the expected value of performance improvement corresponding to each hyperparameter based on the expected improvement acquisition function, and select the hyperparameter corresponding to the maximum expected value of performance improvement. Repeat the above method until the expected value of performance improvement is greater than the threshold or the preset number of iterations is reached.

[0114] According to some embodiments of the present disclosure, the process of Bayesian optimization for hyperparameter adjustment is as follows:

[0115] (1) Initialization: Select a set of random hyperparameters to start model evaluation.

[0116] (2) Iterative process:

[0117] a. Performance evaluation:

[0118] Run the model under the current hyperparameters and calculate the loss function using the validation set.

[0119] b. Update probability model:

[0120] Update the relationship model between hyperparameters and the loss function based on Gaussian process (GP). Gaussian process is a commonly used non-parametric probability model that can predict the performance of unexplored parameters.

[0121] c. Optimize acquisition function:

[0122] Use acquisition functions such as expected improvement (EI) to determine the next set of hyperparameters. The EI function calculates the expected value of the possible performance improvement for each candidate point, and selects the point with the largest improvement amount as the hyperparameter for the next step.

[0123] EI(x) = E[max(0, f(*) - f(x))]

[0124] Among them, x* represents the currently observed best input; max represents taking the maximum value; f represents the objective function, EI is the expected improvement; E represents the calculation of the expected value of a random variable.

[0125] (3) Convergence determination:

[0126] Repeat the above process until the preset number of iterations or improvement threshold is met.

[0127] Furthermore, the loss function is the mean squared error function with an L2 regularization term or the mean absolute error function with an L2 regularization term; among them,

[0128] The expression of the mean squared error function with an L2 regularization term is:

[0129] LF 1 = MSE + R(w)

[0130]

[0131] The expression of the mean absolute error function with an L2 regularization term is:

[0132] LF 2 = MAE + R(w)

[0133]

[0134] Among them, represents the model prediction value, y i is the actual value, N is the number of samples; R(w) represents the regularization term. MSE is used to calculate the sum of the squares of the errors, and MAE evaluates the model performance by calculating the absolute difference between the prediction value and the actual value. These metrics are used to measure the gap between the prediction value and the actual value, providing quantitative metrics for the optimization process.

[0135] Furthermore, after adjusting the hyperparameters of the initial prediction model based on Bayesian optimization, it further includes: optimizing the initial prediction model after parameter tuning based on the Adam optimizer.

[0136] According to some embodiments of the present disclosure, the Adam optimizer is integrated in the model training process. The Adam optimizer is an optimization algorithm with an adaptive learning rate, which adjusts the learning rate of each parameter based on the estimated first and second moments (i.e., the mean and the uncentered variance). This makes the Adam optimizer show excellent performance in practical applications, especially when dealing with non-stationary objectives and very large datasets or parameter spaces.

[0137] According to some embodiments of the present disclosure, in order to improve the generalization ability of the model and prevent overfitting, the present disclosure applies L2 regularization and Dropout techniques:

[0138] L2 Regularization (Weight Decay):

[0139] By adding a regularization term proportional to the L2 norm of the weight vector to the loss function, it helps control the model complexity, and its formula is:

[0140] R(w) = λw 2

[0141] where w represents the model weights and λ represents the regularization coefficient. This helps limit the magnitude of the model weights so that it does not overly rely on the noise in the training data.

[0142] Dropout: Randomly discard some neurons in the network during training to reduce the complex co-adaptation relationships between neurons. The Dropout rate is usually set to 0.5, that is, randomly discard half of the neurons in each training stage. Randomly discarding some neuron connections during training effectively increases the randomness and robustness of model training.

[0143] According to some embodiments of the present disclosure, a two-stage training strategy is adopted during model training: First, perform autoregressive fine-tuning to enable the model to better adapt to the temporal characteristics of building heat load data; Second, further optimize the model performance through supervised learning.

[0144] According to some embodiments of the present disclosure, the model prediction and evaluation process is as follows:

[0145] Model Prediction:

[0146] Use the trained model to predict the building heat load for future time periods:

[0147] Y t+1：t+T = f(X t-L+1:t , A)

[0148] where X t-L+1:t is the input historical data, and A is the adjacency matrix describing the spatial dependence relationship between buildings. Y t+1:t+T is the predicted value for the next T time steps.

[0149] Performance Evaluation:

[0150] Evaluate the prediction accuracy of the model by calculating the Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). Through these metrics, the performance of the model on the test data can be measured, thereby determining its prediction accuracy and stability.

[0151] Based on the same technical concept, as Figure 3 shown, the present disclosure provides a building heat load prediction system, which includes:

[0152] A feature extraction module, configured to:

[0153] Extract the feature information of the building heat load data;

[0154] A model training module, configured to:

[0155] Train an initial prediction model according to the feature information to obtain a target prediction model; the initial prediction model is constructed based on a large language model and a graph neural network;

[0156] A model application module, configured to:

[0157] Perform building heat load prediction according to the target prediction model.

[0158] As Figure 4 shown, the present disclosure provides an electronic device, which includes a memory and a processor. A computer program or instruction is stored in the memory. When the computer program or instruction is executed by the processor, it is at least used to implement the above-mentioned building heat load prediction method.

[0159] As Figure 5 shown, the present disclosure provides a computer-readable storage medium, in which a computer program or instruction is stored. When the computer program or instruction is executed by the processor, it is at least used to implement the above-mentioned building heat load prediction method.

[0160] The present disclosure also provides a computer program product, which is stored in a computer-readable storage medium. When the computer program product is executed by a processor, it is at least used to implement the above-mentioned building heat load prediction method.

[0161] The working principle and beneficial effects of the above technical solutions:

[0162] First of all, through the application of Bayesian optimization and advanced regularization techniques, the present disclosure optimizes the learning process of the model, significantly improves the accuracy of long-term prediction, and allows the model to maintain a high degree of adaptability and stability under changing environmental conditions, thereby reducing energy waste and improving energy efficiency in practical applications. The application of this method is not limited to the energy consumption management of small-scale or single buildings, but can be extended to a wider range of building groups and even city-level energy efficiency systems. By integrating intelligent control systems, the present disclosure can perform more precise energy efficiency management and achieve the maximum efficiency and effect of energy use.

[0163] In summary, the present disclosure not only shows significant progress in technology, demonstrates its practicality and high efficiency through the application of actual data, but also provides strong technical support for the intelligentization and automation of future building energy efficiency management.

[0164] The present disclosure proposes an innovative building heat load prediction method, which overcomes the disadvantages of traditional prediction methods in capturing long-term dependencies and spatial correlations by integrating multi-source data, combining large language models and graph neural networks.

[0165] The innovative points of the technical solutions given in the present disclosure include:

[0166] (1) Innovative integration of deep spatio-temporal sequence modeling: By integrating a large language model based on Transformer and a graph neural network, deep spatio-temporal analysis capabilities are introduced for building heat load prediction. This combination not only uses the large language model to capture long-term dependencies in the time series, but also details the spatial interactions between buildings through the graph neural network, enabling the prediction model to comprehensively handle the complex interaction relationships of time and space. This method overcomes the deficiencies of traditional methods in capturing long time spans and spatial dependencies between buildings.

[0167] (2) Intelligent integration of multi-source data: By integrating various influencing factors such as meteorological conditions, building physical properties, heating system operation, and user behavior patterns through advanced data processing techniques, the complex factors affecting building heat load are comprehensively captured, forming a comprehensive data view, which significantly enhances the response speed and prediction accuracy of the model to environmental changes.

[0168] (3) Multi-scale feature decomposition and Attention mechanism: The heat load sequence is decomposed into trend components and seasonal components through multi-scale pooling filters and modeled separately to reduce complexity. At the same time, the Attention mechanism is introduced to dynamically assign weights according to the characteristics of the time series, enabling the model to focus on the most critical parts for prediction, thereby improving the prediction accuracy.

[0169] (4) Real-time data feedback and model self-optimization mechanism: By continuously collecting new building operation data and updating model parameters in real time, the present disclosure ensures the continuous performance improvement and adaptability enhancement of the prediction system, which is particularly suitable for dynamically changing building environments.

[0170] (5) Efficient model training and optimization strategies: Combining Bayesian optimization, Adam optimizer, advanced regularization techniques, and a two-stage training strategy not only improves the efficiency of model training, but also enhances the overall prediction accuracy and model generalization ability by finely tuning the parameters of each stage, and effectively prevents overfitting.

[0171] In summary, the building heat load prediction method of the present disclosure integrates the sequence modeling capabilities of large language models and the spatial feature modeling capabilities of graph neural networks, and is applicable to the building energy consumption management needs of different types of buildings and various climate conditions. It can not only accurately predict the heat load of a single building, but also be extended to a wider range of building groups and complex scenarios, effectively supporting the intelligentization and high efficiency of building energy efficiency management.

[0172] It is obvious that those of ordinary skill in the art can make various changes and modifications to the present disclosure without departing from the spirit and scope of the present disclosure. Thus, if these modifications and variations of the present disclosure fall within the scope of the claims of the present disclosure and their equivalent technologies, the present disclosure is also intended to include these changes and modifications.

Claims

1. A method for predicting building heat load, characterized in that: include: Extract characteristic information of building heat load data; An initial prediction model is trained according to the feature information to obtain a target prediction model; the initial prediction model is constructed based on a large language model and a graph neural network; Building heat load prediction is carried out based on the target prediction model.

2. The building heat load prediction method according to claim 1, characterized in that: The building heat load data includes: historical heat load data, meteorological data, building characteristic data, heating system data and user behavior data.

3. The building heat load prediction method according to claim 1, characterized in that: Extract characteristic information of building heat load data, including: Extracting the time characteristics of the building heat load data based on the large language model; Extracting spatial features of the building heat load data through the graph neural network; The time feature and the space feature are fused.

4. The building heat load prediction method according to claim 3, characterized in that: Extracting the time characteristics of the building heat load data based on the large language model includes: extracting the long-term trend characteristics and periodic characteristics of the building heat load data through multi-scale pooling filtering; the multi-scale pooling filtering corresponds to the expression: X T =Softmax(w(X))·f(X) X S =X-X T Among them, Softmax represents the Softmax activation function; w(X) represents the data-driven weight; f(X) represents multiple average pooling filters; X S represents periodic characteristics; X represents input data; X T Indicates long-term trend characteristics.

5. The building heat load prediction method according to claim 4, characterized in that: The large language model contains an Attention mechanism; wherein, The expression of the Attention mechanism is: Where Q represents the query vector; K represents the key vector; V represents the numerical vector; d k represents the dimension of the key vector; T represents the transpose of the matrix.

6. The building heat load prediction method according to claim 4, characterized in that: The graph neural network is configured to construct dependency relationships between buildings through an adjacency matrix, and the corresponding expressions include: Among them, A i,j represents the weight relationship between building i and building j; dv i ,v j Indicates building v i and building v j The distance between them; σ represents the standard deviation of the Gaussian kernel; ò is used to control the sparsity.

7. The building heat load prediction method according to any one of claims 1 to 6, characterized in that: Training an initial prediction model according to the feature information includes: adjusting hyperparameters of the initial prediction model based on Bayesian optimization.

8. A building heat load prediction system, characterized in that: include: The feature extraction module is configured as follows: Extract characteristic information of building heat load data; The model training module is configured as follows: An initial prediction model is trained according to the feature information to obtain a target prediction model; the initial prediction model is constructed based on a large language model and a graph neural network; The model application module is configured as follows: Building heat load prediction is carried out based on the target prediction model.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program or instruction, and when the computer program or instruction is executed by the processor, it is used to implement at least the method according to any one of claims 1 to 7.

10. A computer program product, the computer program product being stored in a computer-readable storage medium, characterized in that: When the computer program product is executed by a processor, it is used to implement at least the method according to any one of claims 1 to 7.

Images