Adaptive Constraints and Error Calibration: Short-Time Prediction Method and System for Building Cooling Load
By using an adaptive gray box prediction method, combined with building thermophysics mechanisms and deep learning technology, the problems of adaptability and error accumulation in the prediction of cooling load in public buildings are solved, achieving high-precision and physically consistent cooling load prediction, and improving the stability and adaptability of the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 华能置业有限公司
- Filing Date
- 2026-05-08
- Publication Date
- 2026-07-31
AI Technical Summary
Existing cooling load forecasting methods suffer from poor model adaptability, accumulation of prediction errors, and insufficient physical consistency in public buildings. In particular, they exhibit poor generalization performance under extreme weather conditions or sparse data conditions, and existing calibration mechanisms lack dynamic modeling.
An adaptive gray box prediction method is adopted, which integrates building thermophysical mechanisms and deep learning technology to construct a multi-source data preprocessing module, a cooling load prediction module, an adaptive physical information fusion module, and a real-time calibration module, thereby achieving high-precision and highly physically consistent cooling load prediction.
It improves the accuracy and robustness of cooling load forecasting, enhances the model's responsiveness to changes in operating conditions, reduces cumulative forecast errors, and provides reliable support for energy optimization control.
Smart Images
Figure CN122491585A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of building energy management and intelligent forecasting technology, and in particular to a method and system for short-term forecasting of building cooling load with adaptive constraints and error calibration. Background Technology
[0002] Short-term prediction of cooling load in public buildings is a core technology of building energy management systems, directly determining the effectiveness of air conditioning system optimization and indoor environmental comfort. Existing cooling load prediction methods are mainly divided into two categories: white-box models and black-box models, both of which have significant technical defects. White-box models are based on physical equations, but they require high accuracy of building thermophysical parameters. However, public buildings have complex structures and parameters with strong uncertainties, resulting in large prediction biases and difficulty in adapting to actual operating conditions. Black-box models (such as neural networks and machine learning models) rely on data-driven methods to capture nonlinear time-series relationships, but they ignore building thermophysical mechanisms. They have poor generalization performance in extreme weather, data sparse, or unseen operating conditions, and are prone to producing non-physical prediction results. In addition, traditional improvement methods often use fixed scene divisions and static physical constraint weights, which cannot adapt to the changing operating modes of buildings, resulting in poor model adaptability. Existing real-time calibration mechanisms are mostly simple threshold corrections, lacking dynamic modeling of the evolution of prediction error states, which easily leads to error accumulation and propagation, reducing prediction stability.
[0003] To address the aforementioned technical issues, this invention proposes an adaptive gray box prediction method that integrates building thermophysical mechanisms with deep learning technology to achieve high-precision, highly physical consistent, and robust short-term prediction of cooling load in public buildings. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for short-term prediction of building cooling load with adaptive constraints and error calibration. Following the technical approach of "high-quality data cleaning - adaptive scene modeling - dynamic physical gray box fusion - error state evolution calibration", four functional modules are set up to be executed sequentially and progressively. By deeply integrating building thermophysical mechanisms and deep learning technology, the shortcomings of existing technologies are improved.
[0005] To achieve the above objectives, this invention provides a short-term prediction method for building cooling load with adaptive constraints and error calibration, comprising the following steps: S1. Collect and preprocess multi-source data for predicting building cooling load; S2. Extract features from the multi-source data after S1 preprocessing to construct the input dataset; S3. Construct a cooling load prediction model and input the input dataset constructed in S2; S4. The prediction results of the cooling load prediction model based on the physical constraint function constraining S3 according to the building energy balance and heat conduction mechanism conform to physical laws. S5. Based on Kalman filtering, a prediction error state evolution model is constructed to realize dynamic modeling and real-time calibration of prediction errors.
[0006] Preferably, the specific process of S1 is as follows: S11. Collect multi-source data of the building, including historical cooling load, indoor environmental parameters, outdoor meteorological parameters, and equipment operating status parameters. All multi-source data are set to uniform resampling with a sampling frequency of 15 minutes. S12. The Z-score and isolated forest joint detection method is used to identify mutation anomalies, latent anomalies, and transmission anomalies in multi-source data, and to remove or interpolate and repair abnormal data. S13. Use Synchronous Squeezing Wavelet Transform (SWT) to perform time-frequency decomposition and noise reduction on historical cooling loads, and extract intrinsic mode components.
[0007] Preferably, in S2, the cooling load sequence, outdoor temperature, solar radiation intensity, and chiller unit operating status are extracted from the preprocessed multi-source data as core features to construct an input dataset containing the lagging features of the previous 1-2 hours. That is, the input of the cooling load prediction model at the current moment contains historical feature data of the previous 4-8 15-minute time steps, so as to fully capture the temporal coupling relationship of the cooling load.
[0008] Preferably, the specific process of S3 is as follows: S31. The k-means clustering algorithm is used to perform cluster analysis on the input dataset of S2, automatically identifying 2-5 types of building operation mode scenarios, and optimizing the clustering effect through the contour coefficient. S32. For each type of building operation mode scenario identified in S31, a Transformer prediction sub-model is constructed. The multi-head self-attention mechanism is used to capture the temporal coupling relationship of multiple variables to achieve rolling prediction with a 15-minute time step.
[0009] Preferably, the specific process of S4 is as follows: S41. Construct a physical constraint loss function based on building energy balance and heat conduction mechanisms. The specific formula is as follows: ; in, This represents the total number of training samples. It is a linear rectification activation function, which generates a penalty term only when the predicted cooling load does not decrease monotonically with respect to the outdoor temperature over a 1-hour time lag, thus conforming to the well-known mechanism of building thermal inertia. For the first The predicted cooling load value for each sample For the first The predicted cooling load value for each sample For the first The outdoor temperature with a 1-hour time lag for each sample The time constant is set to 1 hour. For the first The outdoor temperature with a 1-hour time lag corresponding to each sample; S42. Set the weighting coefficients for physical constraints. The parameters are adaptive and dynamically adjusted based on the outdoor temperature change rate and the residual of the cooling load prediction model at the previous moment. S43. Using data-driven loss function and Weighted fusion into the total loss function Total loss function The formula is as follows: ; ; in, This represents the actual cooling load.
[0010] Preferably, physical constraint weighting coefficients are set in S42. base value According to the rate of change of outdoor temperature Residual of the previous time step compared with the cooling load prediction model The formula for dynamic adjustment is as follows: ; in, for The physical constraint weighting coefficients at time t and their range of values are . , for outdoor temperature change with a time lag of 1 hour , for Outdoor temperature with a 1-hour time lag. for Outdoor temperature with a 1-hour time lag. For time step, , These are the adjustment parameters obtained through optimization using the validation set.
[0011] Preferably, the specific process of S5 is as follows: S51. Construct a prediction error state evolution model based on Kalman filtering, and use modified Kalman filter MKF or ensemble Kalman filter EnKF to perform real-time state estimation of the prediction residuals. S52. Based on the uncertainty information predicted by the cooling load forecasting model, dynamically adjust the process noise covariance matrix. and measurement noise covariance matrix Adaptive update of Kalman gain ; S53. The optimal weighted fusion calibration of the predicted value and the real-time observation value is achieved through the state update equation to obtain the calibrated cooling load prediction result.
[0012] Preferably, the specific process of S53 is as follows: S531, Set residual threshold ,in This is the set absolute minimum load tolerance error constant; S532. Quantify and judge the predicted residuals when... When the residual is deemed too large, it is automatically increased. ;when When the system is deemed to be operating stably, the speed is automatically reduced. ; S533, The state update equation is as follows: ; in, The calibrated state value. To predict state values, For real-time observations, This is the observation matrix.
[0013] To achieve the above objectives, the present invention also provides a short-term building cooling load prediction system with adaptive constraints and error calibration, including a multi-source data preprocessing module, a cooling load prediction module, an adaptive physical information fusion module, and a real-time calibration module; The multi-source data preprocessing module is responsible for collecting and processing multi-source data, and then constructing an input dataset which is fed into the cooling load prediction module. The cooling load prediction module performs scene clustering based on equipment operating status parameters and historical cooling load change rates to achieve adaptive identification of operating modes, and constructs a Transformer prediction sub-model to capture the temporal coupling relationship of multiple variables through a multi-head self-attention mechanism. The adaptive physical information fusion module introduces a physical constraint loss function to constrain the prediction results of the cooling load prediction model to conform to physical laws. The real-time calibration module performs real-time error calibration on the prediction results based on Kalman filtering.
[0014] Therefore, the adaptive constraint and error calibration method and system for short-term prediction of building cooling load in this invention, compared with the prior art, has the following advantages: 1. This application introduces SWT time-frequency decomposition and k-means adaptive clustering to achieve automatic identification of operating modes and dynamic matching of sub-models, reducing reliance on human experience and improving the adaptability of the cold load prediction model; it constructs an adaptive physical constraint mechanism, which drives the adjustment of physical weights through meteorological changes and error feedback, and combines monotonicity and energy conservation constraints to improve the physical consistency and generalization ability of the prediction results. 2. This application uses an error state modeling and real-time calibration mechanism based on Kalman filtering to achieve online correction of prediction errors, enhance the responsiveness of the cooling load prediction model to changes in operating conditions, and reduce cumulative prediction errors. It integrates multi-source data processing, physical constraint modeling, and data-driven prediction methods to improve the interpretability of results while ensuring prediction accuracy, providing reliable support for energy-saving operation and precise control of air conditioning systems in public buildings.
[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0016] Figure 1 This is a flowchart illustrating the principle of the adaptive constraint and error calibration method for short-term prediction of building cooling load in this invention. Figure 2 This is a schematic diagram of the adaptive physical information module of the building cooling load short-time prediction method with adaptive constraints and error calibration of the present invention; Figure 3 This is a schematic diagram of the real-time calibration mechanism of the adaptive constraint and error calibration method for short-term prediction of building cooling load in this invention; Figure 4 This is the overall framework diagram of the building cooling load short-term prediction system with adaptive constraints and error calibration of the present invention. Detailed Implementation
[0017] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed when in use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.
[0018] Example like Figures 1-3 As shown, the adaptive constraint and error calibration method for short-term prediction of building cooling load of the present invention includes the following steps: S1. Collect and preprocess multi-source data for predicting building cooling load; S11. Collect multi-source data of the building, including historical cooling load (the cooling load of the building is the measured cooling capacity data), indoor environmental parameters, outdoor meteorological parameters, and equipment operating status parameters. All multi-source data are set to uniform resampling and the sampling frequency is 15 minutes. S12. The Z-score and isolated forest joint detection method is used to identify mutation anomalies, latent anomalies, and transmission anomalies in multi-source data, and to remove or interpolate and repair abnormal data. S13. Synchronous squeeze wavelet transform (SWT) is used to decompose and denoise historical cooling loads in time and frequency, extract intrinsic mode components, and eliminate the influence of data non-stationarity. S2. Extract features from the multi-source data after S1 preprocessing to construct the input dataset; The core features are extracted from the preprocessed multi-source data, including the cooling load sequence, outdoor temperature, solar radiation intensity, and chiller unit operating status. An input dataset containing the lagging features of the previous 1-2 hours is constructed. That is, the input of the current cooling load prediction model contains historical feature data of the previous 4-8 15-minute time steps, which fully captures the temporal coupling relationship of the cooling load. S3. Construct a cooling load prediction model and input the input dataset constructed in S2; S31. The k-means clustering algorithm is used to perform cluster analysis on the input dataset of S2, automatically identifying 2-5 types of building operation mode scenarios, and optimizing the clustering effect through the contour coefficient. S32. For each type of building operation mode scenario identified in S31, a Transformer prediction sub-model is constructed. The multi-head self-attention mechanism is used to capture the temporal coupling relationship of multiple variables to achieve rolling prediction with a 15-minute time step. S4. The prediction results of the cooling load prediction model based on the physical constraint function constraining S3 according to the building energy balance and heat conduction mechanism conform to physical laws. S41. Construct a physical constraint loss function based on building energy balance and heat conduction mechanisms. The specific formula is as follows: ; in, This represents the total number of training samples. It is a linear rectification activation function, which generates a penalty term only when the predicted cooling load does not decrease monotonically with respect to the outdoor temperature over a 1-hour time lag, thus conforming to the well-known mechanism of building thermal inertia. For the first The predicted cooling load value for each sample For the first The predicted cooling load value for each sample For the first The outdoor temperature with a 1-hour time lag for each sample The time constant is set to 1 hour. For the first The outdoor temperature with a 1-hour time lag corresponding to each sample; S42. Set the weighting coefficients for physical constraints. The parameters are adaptive and dynamically adjusted based on the outdoor temperature change rate and the residual of the cooling load prediction model at the previous moment. The physical constraint weighting coefficient is set. base value According to the rate of change of outdoor temperature Residual of the previous time step compared with the cooling load prediction model The formula for dynamic adjustment is as follows: ; in, for The physical constraint weighting coefficients at time t and their range of values are . , for outdoor temperature change with a time lag of 1 hour , for Outdoor temperature with a 1-hour time lag. for Outdoor temperature with a 1-hour time lag. For time step, , These are the adjustment parameters obtained through optimization using the validation set; S43. Using data-driven loss function and Weighted fusion into the total loss function Total loss function The formula is as follows: ; ; in, This represents the actual cooling load. S5. Based on Kalman filtering, a prediction error state evolution model is constructed to realize dynamic modeling and real-time calibration of prediction errors; S51. Construct a prediction error state evolution model based on Kalman filtering, and use modified Kalman filter MKF or ensemble Kalman filter EnKF to perform real-time state estimation of the prediction residuals. S52. Based on the uncertainty information predicted by the cooling load forecasting model, dynamically adjust the process noise covariance matrix. and measurement noise covariance matrix Adaptive update of Kalman gain ; S53. The optimal weighted fusion calibration of the predicted value and the real-time observation value is achieved through the state update equation to obtain the calibrated cooling load prediction result. S531, Set residual threshold ,in This is the set absolute minimum load tolerance error constant; S532. Quantify and judge the predicted residuals when... When the residual is deemed too large, it is automatically increased. Increase the weight of real-time observations in calibration; when When the system is deemed to be operating stably, the speed is automatically reduced. This ensures the smoothness of the prediction results; S533, The state update equation is as follows: ; in, The calibrated state value. To predict state values, For real-time observations, This is the observation matrix.
[0019] like Figure 4 As shown, the adaptive constraint and error calibration building cooling load short-term prediction system of the present invention includes a multi-source data preprocessing module, a cooling load prediction module, an adaptive physical information fusion module, and a real-time calibration module. The multi-source data preprocessing module is responsible for collecting and processing multi-source data, and then constructing an input dataset to be input into the cooling load prediction module. The cooling load prediction module performs scene clustering based on equipment operating status parameters and historical cooling load change rates to achieve adaptive identification of three types of operating modes: shutdown stage, start / stop transition stage, and stable start-up stage. It also constructs a Transformer prediction sub-model to capture the temporal coupling relationship of multiple variables through a multi-head self-attention mechanism. The adaptive physical information fusion module introduces a physical constraint loss function to constrain the prediction results of the cooling load prediction model to conform to physical laws. The real-time calibration module performs real-time error calibration on the prediction results based on Kalman filtering. Specific implementation examples: The multi-source data preprocessing process is as follows: Multi-source data acquisition: Collect historical cooling load, indoor temperature and humidity, outdoor temperature and humidity, solar radiation, day of the week, whether it is a holiday, time of day coding, and equipment operation status data of a public building. The collection period is 3 months. The multi-source data sampling frequency includes 5 minutes, 15 minutes, and 30 minutes. All data will be uniformly resampled to a 15-minute sampling frequency to ensure consistency of time scale. Anomaly detection and handling: The Z-score method is used to identify mutation anomalies in the data, and the isolated forest algorithm is used to identify latent anomalies and transmission anomalies (such as missing or duplicate data). The identified abnormal data is repaired using linear interpolation, and severely abnormal data is directly removed. Time-frequency decomposition denoising: Synchronous squeezed wavelet transform (SWT) is used to decompose the repaired historical cold load into multiple intrinsic mode components, remove noisy high-frequency components, reconstruct the denoised cold load sequence, eliminate the impact of data non-stationarity, and improve data quality.
[0021] The process of constructing the cooling load prediction model is as follows: Feature extraction: Extract the cooling load sequence, outdoor temperature, solar radiation intensity, and chiller unit operating status from the preprocessed data as core features, and construct an input dataset containing the lagging features of the previous 1-2 hours. That is, the input of the current cooling load prediction model contains historical feature data of the previous 4-8 15-minute time steps, so as to fully capture the temporal coupling relationship of the cooling load. Clustering of operating scenarios: The k-means clustering algorithm is used to perform clustering analysis on the input dataset. The initial number of clusters is set to 2-5. The silhouette coefficient is used to evaluate and optimize the clustering effect. The closer the silhouette coefficient is to 1, the better the clustering effect. Finally, the optimal number of clusters is determined to be 3, which are peak load condition, steady state operation condition and low load condition. Sub-model construction: For three types of operating conditions—peak load, steady-state operation, and low load—three Transformer prediction sub-models are constructed respectively. Each sub-model has a 3-layer encoder and 8 multi-head self-attention heads. The multi-head self-attention mechanism captures the temporal coupling relationship and lag features between multiple variables. Training parameters: learning rate 0.001, number of iterations 100 rounds, batch size 32, and the optimizer is Adam. The model output is the cold load prediction value 15 minutes later in the corresponding scenario, realizing scenario-adaptive rolling prediction of cold load.
[0022] The adaptive physical information fusion process is as follows: Loss function construction: During the training of the Transformer sub-model, a physical constraint loss function based on the building energy balance and heat conduction mechanism is introduced. The function includes a monotonic ReLU penalty term and an energy conservation soft constraint term. The monotonic ReLU penalty term ensures that the cooling load increases monotonically with the increase of outdoor temperature, avoiding non-physical prediction results. The energy conservation soft constraint term ensures the balance of indoor and outdoor heating and cooling energy expenditure, conforming to the actual thermophysical laws of buildings. The mean absolute error (MAE) is selected as the data-driven loss function. The deviation between the predicted and actual values of the cooling load forecasting model is quantified. Total Loss Function Fusion: Data-Driven Loss Function With physical constraint loss function Weighted fusion is used as the total loss function for training the cold load prediction model. , The value ranges from 0 to 1 and is used to adjust the weight ratio of physical constraints and data-driven factors in model training; Adaptive weight adjustment: Set the weighting coefficients for physical constraints base value According to the rate of change of outdoor temperature Residual predicted by the model at the previous time step The dynamic adjustment formula is: The adjustment parameters were obtained through validation set optimization. , Under unsteady operating conditions, such as a sudden rise or fall in outdoor temperature, Increase It automatically increases the physical constraint capability to ensure that the prediction results conform to physical laws; under steady-state conditions, Decrease Automatically reduce the weight of physical constraints to ensure the accuracy of data-driven predictions.
[0023] The real-time calibration process is as follows: Error Model Construction: A prediction error state evolution model is constructed based on Kalman filtering. The prediction residuals of the cooling load prediction model (the difference between the predicted value and the actual observed value) are used as state variables. Modified Kalman filtering (MKF) is used to estimate the state of the residuals in real time, accurately capturing the dynamic change law of the error. Dynamic adjustment of filtering parameters: The variance of the attention weights in the Transformer sub-model is used as information on model prediction uncertainty. When the variance of the attention weights is large, it is determined that the model prediction uncertainty is high, and the process noise covariance matrix is increased. When the observed data is of high quality and has low fluctuation, it is considered that the observed data has high accuracy, thus reducing the measurement noise covariance matrix. ,accomplish , Dynamic adaptive adjustment, and based on the adjusted and Preliminary calculation of Kalman gain ; Predicted value optimal fusion calibration: setting residual threshold Quantitatively assess the predicted residuals: when When the residual is deemed too large, it is automatically increased. Increase the weight of real-time observations in calibration; when When the system is deemed to be operating stably, the speed is automatically reduced. This ensures the smoothness of the prediction results; the optimal weighted fusion of the predicted value and the real-time observation value is achieved through the state update equation to obtain the calibrated cold load prediction value, thus realizing continuous and stable real-time dynamic calibration.
[0024] Therefore, the adaptive constraint and error calibration method and system for short-term prediction of building cooling load in this invention, which integrates building thermophysical mechanisms and deep learning technology, can effectively improve the accuracy, physical consistency and robustness of cooling load prediction, and provide reliable support for the energy optimization control of public buildings.
[0025] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A short-term prediction method for building cooling load with adaptive constraints and error calibration, characterized in that, Includes the following steps: S1. Collect and preprocess multi-source data for predicting building cooling load; S2. Extract features from the multi-source data after S1 preprocessing to construct the input dataset; S3. Construct a cooling load prediction model and input the input dataset constructed in S2; S4. The prediction results of the cooling load prediction model based on the physical constraint function constraining S3 according to the building energy balance and heat conduction mechanism conform to physical laws. S5. Based on Kalman filtering, a prediction error state evolution model is constructed to realize dynamic modeling and real-time calibration of prediction errors.
2. The short-term prediction method for building cooling load with adaptive constraints and error calibration according to claim 1, characterized in that: The specific process of S1 is as follows: S11. Collect multi-source data of the building, including historical cooling load, indoor environmental parameters, outdoor meteorological parameters, and equipment operating status parameters. All multi-source data are set to uniform resampling with a sampling frequency of 15 minutes. S12. The Z-score and isolated forest joint detection method is used to identify mutation anomalies, latent anomalies, and transmission anomalies in multi-source data, and to remove or interpolate and repair abnormal data. S13. Use Synchronous Squeezing Wavelet Transform (SWT) to perform time-frequency decomposition and noise reduction on historical cooling loads, and extract intrinsic mode components.
3. The short-term prediction method for building cooling load with adaptive constraints and error calibration according to claim 2, characterized in that: S2 extracts the cooling load sequence, outdoor temperature, solar radiation intensity, and chiller unit operating status from the preprocessed multi-source data as core features, and constructs an input dataset containing the lagging features of the previous 1-2 hours. That is, the input of the current cooling load prediction model contains historical feature data of the previous 4-8 15-minute time steps, which fully captures the temporal coupling relationship of the cooling load.
4. The short-term prediction method for building cooling load with adaptive constraints and error calibration according to claim 3, characterized in that: The specific process of S3 is as follows: S31. The k-means clustering algorithm is used to perform cluster analysis on the input dataset of S2, automatically identifying 2-5 types of building operation mode scenarios, and optimizing the clustering effect through the contour coefficient. S32. For each type of building operation mode scenario identified in S31, a Transformer prediction sub-model is constructed. The multi-head self-attention mechanism is used to capture the temporal coupling relationship of multiple variables to achieve rolling prediction with a 15-minute time step.
5. The short-term prediction method for building cooling load with adaptive constraints and error calibration according to claim 4, characterized in that: The specific process of S4 is as follows: S41. Construct a physical constraint loss function based on building energy balance and heat conduction mechanisms. The specific formula is as follows: ； in, This represents the total number of training samples. It is a linear rectification activation function, which generates a penalty term only when the predicted cooling load does not decrease monotonically with respect to the outdoor temperature over a 1-hour time lag, thus conforming to the well-known mechanism of building thermal inertia. For the first The predicted cooling load value for each sample For the first The predicted cooling load value for each sample For the first The outdoor temperature with a 1-hour time lag for each sample The time constant is set to 1 hour. For the first The outdoor temperature with a 1-hour time lag corresponding to each sample; S42. Set the weighting coefficients for physical constraints. The parameters are adaptive and dynamically adjusted based on the outdoor temperature change rate and the residual of the cooling load prediction model at the previous moment. S43. Using data-driven loss function and Weighted fusion into the total loss function Total loss function The formula is as follows: ； ； in, This represents the actual cooling load.
6. The short-term prediction method for building cooling load with adaptive constraints and error calibration according to claim 5, characterized in that: Set the physical constraint weighting coefficient in S42 base value According to the rate of change of outdoor temperature Residual of the previous time step compared with the cooling load prediction model The formula for dynamic adjustment is as follows: ； in, for The physical constraint weighting coefficients at time t and their range of values are . , for outdoor temperature change with a time lag of 1 hour , for Outdoor temperature with a 1-hour time lag. for Outdoor temperature with a 1-hour time lag. For time step, , These are the adjustment parameters obtained through optimization using the validation set.
7. The short-term prediction method for building cooling load with adaptive constraints and error calibration according to claim 6, characterized in that: The specific process of S5 is as follows: S51. Construct a prediction error state evolution model based on Kalman filtering, and use modified Kalman filter MKF or ensemble Kalman filter EnKF to perform real-time state estimation of the prediction residuals. S52. Based on the uncertainty information predicted by the cooling load forecasting model, dynamically adjust the process noise covariance matrix. and measurement noise covariance matrix Adaptive update of Kalman gain ; S53. The optimal weighted fusion calibration of the predicted value and the real-time observation value is achieved through the state update equation to obtain the calibrated cooling load prediction result.
8. The short-term prediction method for building cooling load with adaptive constraints and error calibration according to claim 7, characterized in that: The specific process of S53 is as follows: S531, Set residual threshold ,in This is the set absolute minimum load tolerance error constant; S532. Quantify and judge the predicted residuals when... When the residual is deemed too large, it is automatically increased. ;when When the system is deemed to be operating stably, the speed is automatically reduced. ; S533, The state update equation is as follows: ； in, The calibrated state value. To predict state values, For real-time observations, This is the observation matrix.
9. A short-term prediction system for building cooling load with adaptive constraints and error calibration, characterized in that: The building cooling load short-term prediction method using adaptive constraint and error calibration as described in any one of claims 1-8 includes a multi-source data preprocessing module, a cooling load prediction module, an adaptive physical information fusion module, and a real-time calibration module. The multi-source data preprocessing module is responsible for collecting and processing multi-source data, and then constructing an input dataset which is fed into the cooling load prediction module. The cooling load prediction module performs scene clustering based on equipment operating status parameters and historical cooling load change rates to achieve adaptive identification of operating modes, and constructs a Transformer prediction sub-model to capture the temporal coupling relationship of multiple variables through a multi-head self-attention mechanism. The adaptive physical information fusion module introduces a physical constraint loss function to constrain the prediction results of the cooling load prediction model to conform to physical laws. The real-time calibration module performs real-time error calibration on the prediction results based on Kalman filtering.