Public building cold load short-time prediction method fusing physical information

Abstract

The invention discloses a public building cold load short-time prediction method fusing physical information, and the method comprises the steps: collecting and preprocessing the historical cooling capacity, indoor environment, outdoor weather and equipment operation state data of a public building at a fixed time interval, and obtaining multi-dimensional input features; respectively establishing a workday sub-model and a holiday sub-model according to workday and holiday scene division; the workday sub-model and the holiday sub-model jointly form a cold load prediction model, the workday sub-model adopts a long short-term memory (LSTM) network, and the holiday sub-model adopts a light gradient elevator (Light GBM); a physical constraint loss function based on building energy balance and heat conduction residual error is introduced in the training process, and the physical constraint loss is fused into a total loss function according to a weighting coefficient so as to constrain that the output of each sub-model accords with the law of energy conservation and thermal inertia; monitoring the prediction error MAPE in real time and performing online calibration; and outputting a short-time cold load prediction result.

Description

Technical Field

[0001] This invention belongs to the field of building energy management and prediction technology, specifically a method for short-term prediction of cooling load in public buildings that integrates physical information. Background Technology

[0002] Cooling load forecasting for public buildings is a crucial component of energy management systems, used to optimize air conditioning system operation, reduce energy consumption, and improve comfort. Existing technologies are mainly categorized into white-box models (such as energy balance models based on physical equations) and black-box models (such as data-driven neural networks). White-box models rely on precise building parameters, but the complex structures and uncertain parameters of public buildings lead to significant prediction biases. While black-box models can capture nonlinear relationships, they ignore physical laws, exhibit poor generalization performance, and are prone to overfitting. Traditional white-box models struggle to adapt to diverse building scenarios, while black-box models suffer from limited generalization when data is sparse. Furthermore, existing methods neglect real-time calibration and flexible user input, resulting in significant limitations in engineering applications. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for short-term prediction of cooling load in public buildings that integrates physical information. By modeling different scenarios, introducing physical constraint loss functions, online real-time calibration, and user parameter selectability mechanisms, this method achieves high-precision and highly generalizable short-term prediction of cooling load.

[0004] The objective of this invention is achieved through the following technical solution:

[0005] A method for short-term prediction of cooling load in public buildings that integrates physical information includes:

[0006] Historical cooling capacity, indoor environment, outdoor weather, and equipment operation status data of public buildings are collected and preprocessed at fixed time intervals to obtain multidimensional input features;

[0007] Based on the division of weekdays and holidays, a weekday model and a holiday sub-model are established respectively. The weekday model and the holiday sub-model together constitute the cold load prediction model. The weekday model adopts a long short-term memory network LSTM and the holiday sub-model adopts a light gradient booster machine LightGBM.

[0008] During the training process, a physical constraint loss function based on the principles of building thermodynamics is introduced, and this physical constraint loss function is incorporated into the total loss function with weighted coefficients to ensure that the output of the cooling load prediction model conforms to physical laws.

[0009] Real-time monitoring of errors; if the error exceeds a preset threshold, a real-time calibration mechanism is triggered.

[0010] Output short-term cooling load prediction results.

[0011] Further preprocessing includes: identifying and removing outliers in the collected data using Z-score and Isolation Forest; filling missing values ​​with forward fill or linear interpolation; and automatically generating identification fields based on timestamps to indicate whether it is a weekday, a holiday, or a working time period.

[0012] Furthermore, during data acquisition, all types of data are standardized to a 15-minute sampling frequency, and in the event of data mutations, latent anomalies, or data transmission anomalies, data is truncated or resampled according to rules to ensure the continuity of the time series.

[0013] Furthermore, the lag of the multidimensional input features is set to include at least the lag of the multidimensional input features from the previous hour, and a rolling window training and prediction strategy is adopted for each sub-model, with a rolling window prediction step of 15 minutes.

[0014] Furthermore, the physical constraint loss function L... physics Based on thermodynamic principles, the total loss function is expressed as: L total =(1-λ)×L data +λ×L physics L data For the data fitting loss term, the mean absolute error (MAE) is selected to measure the data fitting accuracy, and λ is the physical constraint weighting coefficient.

[0015] Furthermore, the real-time calibration mechanism dynamically corrects the current prediction output based on the prediction error of the previous moment. It is divided into three levels according to the magnitude of the error and adopts different correction strategies for each level: when the error is small (less than 5% of the historical maximum cooling load), the model prediction is considered to be in the confidence range, the original prediction result is retained, and unnecessary intervention is avoided; when the error is moderate (between 5% and 35%), the original prediction value is weighted and fused using linear interpolation, and the fusion coefficient increases linearly with the increase of the error, so as to achieve a smooth transition from trusting the model to relying on the historical true value; when the error is large (more than 35% of the historical maximum cooling load), the true value of the previous moment is used as the current prediction value to quickly block the propagation of anomalies.

[0016] Furthermore, during the deployment phase of the cooling load prediction model, external parameters, including solar radiation, indoor humidity, outdoor humidity, personnel density, special event indicators, window opening status, and equipment maintenance, are allowed to be input. After preprocessing, these parameters are concatenated with the input features of the cooling load prediction model to participate in real-time prediction, and the above predictions can be completed without retraining the cooling load prediction model.

[0017] Preferably, the present invention also provides a short-term prediction device for cooling load of public buildings, comprising:

[0018] The data acquisition and preprocessing unit is used to collect and preprocess historical cooling capacity, indoor environment, outdoor weather and equipment operation status data of public buildings at fixed time intervals to obtain multi-dimensional input features.

[0019] The cooling load prediction model building unit is used to establish a working day model and a holiday sub-model according to the division of working day and holiday scenarios. The working day model and the holiday sub-model together constitute the cooling load prediction model. The working day model adopts a long short-term memory network LSTM and the holiday sub-model adopts a light gradient booster machine LightGBM.

[0020] The physical information fusion unit is used to introduce a physical loss function based on thermodynamic principles during the training process, and to incorporate this physical loss into the total loss function according to weighted coefficients, so as to constrain the output of the cold load prediction model to conform to physical laws.

[0021] The real-time calibration unit is used to monitor the prediction error in real time. When the error exceeds the preset threshold, the real-time calibration mechanism is triggered, and the prediction output is adaptively adjusted according to the error magnitude.

[0022] The output unit is used to output short-term cooling load prediction results.

[0023] Preferably, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the short-term prediction method for the cooling load of public buildings.

[0024] Preferably, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the short-term prediction method for the cooling load of public buildings.

[0025] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0026] 1. A uniform sampling frequency (15 minutes) ensures alignment of different data sources, facilitating the construction of a unified input feature matrix, improving the accuracy of feature temporal alignment, and reducing error sources; automatic generation of weekday / holiday / working time identifiers solves the feature representation problem under different operating modes, which is beneficial for subsequent scenario-based modeling; Z-score and isolation forest are used to remove outliers, and forward padding or linear interpolation is used to complete the sequence, which can restore the continuity of the sequence and avoid the interference of outliers on the model weights; improve data quality, reduce training noise, and make the model more stable and converge faster.

[0027] 2. The cooling load prediction model employs scenario-specific modeling (LSTM for weekdays and LightGBM for holidays) to specifically fit different patterns: LSTM excels at capturing long-term and short-term time-series dependencies, thermal inertia, and hysteresis effects; LightGBM can train quickly and avoid overfitting when there are few samples or high volatility. By selecting models to match scenarios, the fitting ability and generalization of each scenario are improved. Furthermore, the time-lag features of the first hour (4 steps) and rolling window prediction are introduced during model training to address the issues of cooling load thermal inertia and equipment response delay, thereby improving short-term prediction accuracy.

[0028] 3. By introducing a physical loss term L based on the monotonicity of the hysteresis response of cooling load and outdoor temperature. physics This approach embeds prior knowledge of building thermophysics into the model's total loss function in a differentiable form, constraining the predicted output's response trend to changes in outdoor temperature to conform to the laws of heat conduction. This loss term is constructed using ReLU and penalizes predicted samples that violate monotonicity constraints (i.e., cooling load should not decrease when outdoor temperature rises). Compared to purely data-driven methods, this design provides explicit physical consistency guidance for the model, effectively suppressing non-physical prediction behavior under extreme weather, data sparsity, or unseen scenarios, significantly improving the model's generalization ability, interpretability, and robustness.

[0029] At the same time, by balancing the data fitting term L data By using the mean absolute error (MAE) and the weights of the physical constraint terms (λ=0.1, automatically determined by optimization on the validation set), the model maintains good data fitting accuracy while outputting results that are closer to the actual physical process, thus enhancing the reliability and engineering applicability of cooling load prediction.

[0030] 4. The real-time calibration mechanism sets an error threshold based on the historical maximum cooling load, ensuring that the discrimination standard has practical engineering significance and comparability. Compared with no calibration mechanism, this design effectively suppresses the accumulation of prediction errors and abnormal jumps, improves the stability and temporal continuity of prediction results, and avoids excessive intervention in transient processes. This significantly enhances the robustness of short-term cooling load prediction and improves the reliability and practicality of control decisions.

[0031] 5. Allow users to input optional parameters (such as solar radiation, indoor and outdoor humidity, personnel density, etc.) during the model deployment phase to improve the model's adaptability to special working conditions or short-term disturbances, and improve prediction accuracy without retraining; incorporating additional human or temporarily available information into the input features helps to capture the sources of short-term changes.

[0032] 6. This invention combines scenario-based modeling with physical constraint loss: the optimal model is used on weekdays and holidays respectively, and physical laws are constrained, which can simultaneously achieve high fitting ability and physical consistency, and significantly improve the generalization performance in extreme / unseen scenarios.

[0033] By combining preprocessing with real-time calibration: high-quality input data reduces the initial prediction error baseline, and real-time calibration ensures timely performance recovery when data drifts. The combination of the two ensures long-term stable operation and reduces energy waste.

[0034] Physical constraints combined with lightweight calibration: Physical constraints prevent serious physical deviations in the model output, while lightweight calibration uses a physically simplified model to correct the output when short-term emergencies occur (such as opening windows or a surge in people), ensuring both physical consistency and short-term response.

[0035] In summary, the method of this invention improves the accuracy, physical consistency and prediction generalization performance of short-term cooling load prediction by ensuring the quality of multi-source data, modeling in different scenarios, deeply fusing physical information and performing two-level real-time calibration, and combining user-selectable parameters. This reduces unnecessary energy consumption of the cooling load system and improves system stability and engineering adaptability. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the principle of the method of the present invention;

[0037] Figure 2 This is a schematic diagram of the physical information fusion module of the method of the present invention;

[0038] Figure 3 This is a flowchart of the real-time calibration mechanism of the method of the present invention;

[0039] Figure 4a This is a comparison chart of predicted and actual values ​​under ablation experiments conducted on weekdays and holidays without physical constraints.

[0040] Figure 4b This is a comparison chart of predicted and actual values ​​under ablation experiments with physical constraints, divided into weekday and holiday scenarios. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0042] Example 1

[0043] like Figure 1 As shown, this embodiment provides a method for short-term prediction of cooling load in public buildings by integrating physical information (hereinafter referred to as the method, see [link]). Figure 1The system comprises five parts: multi-source data preprocessing, cooling load prediction model construction, physical information fusion, real-time model calibration, and user input parameter selection. After preprocessing the historical cooling load data of public buildings, the system constructs scenario-specific models and fuses physical information, followed by real-time calibration and user parameter adjustment. These five parts work in a progressive manner to achieve high-precision short-term prediction of the cooling load of public buildings.

[0044] S110. Multi-source data preprocessing: By collecting historical cooling capacity, indoor environmental parameters (indoor temperature), outdoor meteorological parameters (outdoor temperature, weather type), and equipment operating status (chiller unit operating status) of public buildings, a unified sampling frequency of one data point every 15 minutes is used. Outliers are detected and corrected to obtain multi-dimensional input features. Due to sensor failures or transmission interference, the raw data often contains missing values, abrupt changes, and latent anomalies. This section first uses statistical methods (such as Z-score and isolated forest) to identify and remove obvious outliers. Then, forward imputation or linear interpolation is used to complete the missing values, ensuring the continuity and smoothness of the time series. At the same time, an identifier field is automatically generated based on the timestamp to indicate whether it is a weekday, a holiday, or within a working time period. This effectively improves data quality and provides a reliable foundation for subsequent modeling.

[0045] S120. Cooling load prediction model construction: Based on the differences in the operating patterns of public buildings on weekdays and holidays, two dedicated sub-models are established to jointly constitute the cooling load prediction model, namely the weekday model and the holiday sub-model.

[0046] For weekday data, Long Short-Term Memory (LSTM) network is used as the main prediction model to make full use of its ability to capture long-term series dependencies.

[0047] For holiday data, Lightweight Gradient Boosting Machine (LightGBM) is used as the main prediction model to better adapt to the characteristics of large load fluctuations and significant nonlinear features during holidays.

[0048] Both sub-models integrate multi-dimensional input features and consider significant time lag effects (data from the first hour), implementing rolling window predictions with a 15-minute step size. This dual-model strategy fully considers the two typical operating modes of public buildings, improving the model's applicability and prediction accuracy.

[0049] S130, Physical Information Fusion: For the cooling load prediction model, a physical constraint loss function based on building thermodynamics principles is introduced during the training process. Specifically, prior knowledge of building thermodynamics is introduced to transform the monotonicity constraint of the lag response between cooling load and outdoor temperature into a differentiable physical loss term. This loss term is constructed using ReLU form and is used to penalize prediction samples that violate the monotonicity constraint (i.e., when the outdoor temperature rises, the cooling load should not decrease). The total loss function is expressed as: L total =(1-λ)×L data +λ×L physics L data For the data fitting loss term, the mean absolute error (MAE) is selected to measure the data fitting accuracy. λ is the physical constraint weighting coefficient, automatically optimized to 0.1 using the validation set. Furthermore, the additional constraint prediction values ​​are non-negative to reflect the actual physical properties of the cooling load prediction model. This step transforms the purely data-driven black-box model into a more interpretable gray-box model, effectively improving the generalization ability of the cooling load prediction model under extreme weather or unseen conditions. See [link to relevant documentation]. Figure 2 .

[0050] S140. Real-time Model Calibration: For the output of the cooling load prediction model, the current output is dynamically corrected based on the prediction error feedback from the previous time step. The correction strategy is based on the absolute prediction error of the previous time step (t−1). = - The relative relationship with the preset threshold, its threshold benchmark. Take the historical maximum cooling load. Implement different correction logics in three tiers: If... >0.35× If the current prediction is significantly biased, the true value from the previous moment will be used instead. As the predicted value at the current moment If 0.05× ≤ ≤0.35× Then, linear interpolation is used to weight and correct the original predicted values: =(1-α)· +α· The fusion coefficient α increases linearly with increasing error, achieving a smooth transition from a trust model to a history-dependent model; if <0.05× The model prediction is considered to be in the high confidence interval, so the original output is retained. To avoid unnecessary intervention. See Figure 3 .

[0051] S150, Optional User Input Parameters: During the formal deployment phase of the cooling load prediction model, additional user-intervention interfaces are provided to meet common personalized needs in actual engineering projects. Users can input optional parameters such as solar radiation, indoor and outdoor humidity, occupancy density, special event indicators, and window opening status via API or interactive interface. These parameters, after being fused with the multidimensional input features in S110, directly participate in the prediction without retraining the cooling load prediction model. This significantly improves the applicability and ease of operation of the method under different building types and management needs, facilitating the integration and deployment of energy management systems for actual public buildings.

[0052] Preferably, the evaluation indicators for the cooling load prediction model in this embodiment are mainly: RMSE ≤ 10% of the historical maximum cooling capacity; MAE ≤ 8% of the historical maximum cooling capacity; R² ≥ 0.90; CV-RMSE ≤ 25%; |NMBE| ≤ 10%;

[0053] The superiority of the method of the present invention was further confirmed through ablation experiments, specifically:

[0054] (a) Segmented into weekday and holiday scenarios, but without physical constraints:

[0055] The prediction results of the test set (14 days) were comprehensively evaluated using the cooling load prediction model. The overall performance indicators are shown in Table 1, the daily performance indicators of the test set are shown in Table 2, and the comparison chart between predicted and actual values ​​is shown in the figure. Figure 4a As shown:

[0056] Table 1 Overall performance indicators of the model

[0057]

[0058] Table 2 Daily Performance Indicators of the Test Set

[0059]

[0060] (b) Segmented into weekday and holiday scenarios with physical constraints:

[0061] The prediction results of the test set (14 days) were comprehensively evaluated using the cooling load prediction model. Table 3 shows a comparison of the overall performance indicators before and after fusing the physical loss function. The graph comparing the predicted values ​​with the actual values ​​after fusing the physical loss function is shown below. Figure 4b As shown:

[0062] Table 3 Comparison of overall model performance indicators

[0063]

[0064] The specific hyperparameter selections are as follows:

[0065] (1) LSTM network architecture (working day model):

[0066] Long Short-Term Memory (LSTM) networks, by introducing forget gates, input gates, and output gates, can effectively learn long-range dependencies in time series, avoiding the gradient vanishing problem of traditional RNNs. The specific parameter configuration of the LSTM network used in this embodiment is shown in Table 4.

[0067] Table 4 LSTM Network Parameter Configuration

[0068]

[0069] (2) LightGBM model (holiday sub-model):

[0070] To address the variable load patterns during holidays, this embodiment employs the LightGBM gradient boosting tree as the holiday prediction model. LightGBM significantly improves the training rate while maintaining high accuracy through its histogram algorithm and leaf-first growth strategy. The specific parameter configurations of the LightGBM model used in this embodiment are shown in Table 5.

[0071] Table 5 LightGBM Model Parameter Configuration

[0072]

[0073] The cold load prediction model uses statistical features extracted from a 96-step historical sequence as input (mean, standard deviation, extreme values, and terminal values) to effectively capture the distribution characteristics of the sequence.

[0074] Example 2

[0075] Based on the same inventive concept, this application also provides a short-term cooling load prediction device for public buildings, which can be used to implement the method described in the above embodiments, as described in the following embodiments. Since the principle of the short-term cooling load prediction device for public buildings is similar to that of the short-term cooling load prediction method for public buildings, the implementation of this device can refer to the aforementioned method implementation, and repeated details will not be elaborated further. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0076] The embodiments of the present invention provide a specific implementation of a prediction device capable of realizing a short-term prediction method for the cooling load of public buildings, specifically including the following:

[0077] The data acquisition and preprocessing unit is used to collect and preprocess historical cooling capacity, indoor environment, outdoor weather and equipment operation status data of public buildings at fixed time intervals to obtain multi-dimensional input features.

[0078] The cooling load prediction model building unit is used to establish a working day model and a holiday sub-model according to the division of working day and holiday scenarios. The working day model and the holiday sub-model together constitute the cooling load prediction model. The working day model adopts a long short-term memory network LSTM and the holiday sub-model adopts a light gradient booster machine LightGBM.

[0079] The physical information fusion unit is used to introduce a physical constraint loss function based on the building energy balance and heat conduction residuals during the training process, and to incorporate the physical constraint loss into the total loss function according to the weighting coefficients, so as to constrain the output of each sub-model to conform to the laws of energy conservation and thermal inertia.

[0080] A real-time calibration unit is used to monitor the prediction error MAPE in real time and perform online calibration.

[0081] The output unit is used to output short-term cooling load prediction results.

[0082] Preferably, embodiments of this application also provide a specific implementation of an electronic device capable of implementing all steps in the short-term prediction method for cooling load of public buildings in the above embodiments. The electronic device specifically includes the following:

[0083] Processor, memory, communications interface, and bus;

[0084] The processor, memory, and communication interface communicate with each other via a bus; the communication interface is used to realize information transmission between server-side devices, metering devices, and user-side devices.

[0085] The processor is used to call the computer program in memory, and when the processor executes the computer program, it implements all the steps in the short-term prediction method for cooling load of public buildings in the above embodiments.

[0086] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the short-term prediction method for cooling load of public buildings in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the short-term prediction method for cooling load of public buildings in the above embodiments.

[0087] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, hardware + program embodiments are relatively simple in description because they are fundamentally similar to method embodiments; relevant parts can be referred to the descriptions in the method embodiments.

[0088] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0089] While this application provides method operation steps as shown in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the method can be executed in the order shown in the embodiments or drawings or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0090] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] This invention is not limited to the embodiments described above. The above description of specific embodiments is intended to illustrate and explain the technical solutions of this invention. The specific embodiments described above are merely illustrative and not restrictive. Without departing from the spirit and scope of the claims, those skilled in the art can make many specific modifications based on the teachings of this invention, and these modifications all fall within the scope of protection of this invention.

Claims

1. A method for short-term prediction of cooling load in public buildings that integrates physical information, characterized in that, include: Historical cooling capacity, indoor environment, outdoor weather, and equipment operation status data of public buildings are collected and preprocessed at fixed time intervals to obtain multidimensional input features; Based on the division of weekdays and holidays, a weekday model and a holiday sub-model are established respectively. The weekday model and the holiday sub-model together constitute the cold load prediction model. The weekday model adopts a long short-term memory network LSTM and the holiday sub-model adopts a light gradient booster machine LightGBM. During the training process, a physical constraint loss function based on the principles of building thermodynamics is introduced, and this physical constraint loss function is incorporated into the total loss function with weighted coefficients to ensure that the output of the cooling load prediction model conforms to physical laws. Real-time monitoring of errors; if the error exceeds a preset threshold, a real-time calibration mechanism is triggered. Output short-term cooling load prediction results.

2. The method for short-term prediction of cooling load in public buildings according to claim 1, characterized in that, Preprocessing includes: identifying and removing outliers in the collected data using Z-score and Isolation Forest; filling missing values ​​with forward fill or linear interpolation; and automatically generating identification fields based on timestamps to indicate whether it is a weekday, a holiday, or a working time period.

3. The method for short-term prediction of cooling load in public buildings according to claim 1, characterized in that, During data acquisition, all types of data are uniformly sampled at a frequency of 15 minutes. In the event of data mutations, latent anomalies, or data transmission anomalies, data is truncated or resampled according to rules to ensure the continuity of the time series.

4. The method for short-term prediction of cooling load in public buildings according to claim 1, characterized in that, The lag of the multidimensional input features should be included at least for the first hour, and a rolling window training and prediction strategy should be adopted for each sub-model, with a 15-minute step size for rolling window prediction.

5. The method for short-term prediction of cooling load in public buildings according to claim 1, characterized in that, The physical constraint loss function L physics Based on thermodynamic principles, the total loss function is expressed as: L total =(1-λ)×L data +λ×L physics L data The loss term for data fitting is the mean absolute error (MAE), which measures the accuracy of data fitting. λ is the weighting coefficient for physical constraints.

6. The method for short-term prediction of cooling load in public buildings according to claim 1, characterized in that, The real-time calibration mechanism dynamically corrects the current prediction output based on the prediction error of the previous moment. It is divided into three levels according to the magnitude of the error, and different correction strategies are adopted for each level: when the error is less than 5% of the historical maximum cooling load, the prediction value of the cooling load prediction model is considered to be in the confidence range, and the original prediction value is retained; when the error is between 5% and 35%, the original prediction value is weighted and fused using linear interpolation, and the fusion coefficient increases linearly with the increase of the error, so as to achieve a smooth transition from trusting the cooling load prediction model to relying on the historical true value; when the error exceeds 35% of the historical maximum cooling load, the true value of the previous moment is used as the current prediction value to quickly block the propagation of anomalies.

7. The method for short-term prediction of cooling load in public buildings according to claim 1, characterized in that, During the deployment phase of the cooling load prediction model, external parameters, including solar radiation, indoor humidity, outdoor humidity, personnel density, special event indicators, window opening status, and equipment maintenance, are allowed to be input. After preprocessing, these parameters are concatenated with the input features of the cooling load prediction model to participate in real-time prediction, and the above predictions can be completed without retraining the cooling load prediction model.

8. A short-term prediction device for cooling load of public buildings, characterized in that, include: The data acquisition and preprocessing unit is used to collect and preprocess historical cooling capacity, indoor environment, outdoor weather and equipment operation status data of public buildings at fixed time intervals to obtain multi-dimensional input features. The cooling load prediction model building unit is used to establish a working day model and a holiday sub-model according to the division of working day and holiday scenarios. The working day model and the holiday sub-model together constitute the cooling load prediction model. The working day model adopts a long short-term memory network LSTM and the holiday sub-model adopts a light gradient booster machine LightGBM. The physical information fusion unit is used to introduce a physical loss function based on thermodynamic principles during the training process, and to incorporate this physical loss into the total loss function according to weighted coefficients, so as to constrain the output of the cold load prediction model to conform to physical laws. The real-time calibration unit is used to monitor the prediction error in real time. When the error exceeds the preset threshold, the real-time calibration mechanism is triggered, and the prediction output is adaptively adjusted according to the error magnitude. The output unit is used to output short-term cooling load prediction results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the short-term prediction method for cooling load of public buildings according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the short-term prediction method for cooling load of public buildings as described in any one of claims 1 to 7.

Images