A server and a cold and heat load prediction method based on a pre-trained prediction model

Abstract

The application discloses a server and a cold and heat load prediction method based on a pre-training prediction model. The method fuses multi-source original data of meteorological environment perception, heating and ventilation equipment operation and time sequence attribute characteristics, and calculates a cold and heat load sequence corresponding to the original data through a pre-training prediction model. The pre-training prediction model is trained by using a physical constraint loss function carrying multiple types of weights including heat balance, energy efficiency, monotonicity and boundary conditions, and is guided by thermodynamic laws throughout the process to optimize model parameters. The method not only overcomes the physical constraint failure, physical violation and cold start problem of a mechanism fusion scheme, but also solves the defects of a digital twin model, such as parameter solidification, lack of self-learning ability, poor precision and adaptability, and realizes all-round consideration of prediction accuracy, physical compliance, cold start adaptability and dynamic adjustment capability.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and building energy management technology, and in particular to a server and a method for predicting heating and cooling loads based on a pre-trained prediction model. Background Technology

[0002] Integrated energy stations (also known as centralized cooling and heating stations) serve as core facilities for centralized energy supply in building clusters. They primarily provide chilled water for summer cooling and hot water for winter heating to various building complexes such as office buildings, hospitals, and shopping malls. Their operational efficiency directly impacts energy supply quality, equipment wear and tear, and energy utilization efficiency. In the actual operation of integrated energy stations, core large-scale equipment such as chillers and boilers exhibit significant start-up and shutdown time delays. For example, a centrifugal chiller typically requires 30-60 minutes to reach a stable operating state after startup. This characteristic necessitates that energy station operators must predict the cooling and heating load demand for the next 24-48 hours in advance. Based on this prediction, they must plan equipment start-up and shutdown times, rationally allocate the load of each piece of equipment, and develop scientific energy storage strategies to ensure energy supply stability, reduce energy consumption, and extend equipment lifespan.

[0003] Currently, various technical methods exist in the industry for predicting the heating and cooling loads of integrated energy stations. Among them, the mechanism fusion method uses a mechanism model to obtain initial values ​​and data-driven correction of deviations, but physical constraints fail in neural network training, failing to address physical violations and cold start issues. The digital twin method predicts by comparing virtual models with measured data, but the model parameters of this method are fixed and cannot be updated adaptively, and it lacks artificial intelligence learning capabilities, resulting in insufficient adaptability and accuracy. It is evident that the existing methods for predicting the heating and cooling loads of integrated energy stations cannot simultaneously address prediction accuracy, physical rationality, cold start adaptability, and dynamic adjustment capabilities, making it difficult to meet the actual needs of refined operation and energy conservation and consumption reduction of energy stations. Therefore, there is an urgent need for a heating and cooling load prediction technology that solves the above pain points. Summary of the Invention

[0004] This application provides a server and a method for predicting hot and cold loads based on a pre-trained prediction model. This method addresses the technical problems of physical constraint failure in mechanism fusion methods and the inability of digital twin models to adaptively iterate.

[0005] In a first aspect, this application provides a server, comprising: The controller is configured to: acquire raw data, including meteorological environmental sensing data, HVAC equipment operation data, and time-series attribute feature data; input the raw data into a pre-trained prediction model to obtain the heating and cooling load sequence corresponding to the raw data output by the pre-trained prediction model; wherein, the pre-trained prediction model is trained based on historical sample data and a physical constraint loss function; the physical constraint loss function is used to guide the update direction of model parameters according to thermodynamic physical laws during model training; the physical constraint weights of the physical constraint loss function include at least one of thermal balance constraint weights, upper bound constraint weights of equipment performance coefficients, monotonicity constraint weights, and boundary condition constraint weights.

[0006] In some examples, the controller, in the process of inputting raw data into a pre-trained prediction model and obtaining the heating and cooling load sequence corresponding to the raw data output by the pre-trained prediction model, is configured to: convert meteorological environmental perception data into intermediate vectors with thermodynamic physical meaning through the physical feature engineering layer of the pre-trained prediction model; determine short-term dynamic features and long-term periodic features based on the intermediate vectors, HVAC equipment operation data, and time-series attribute feature data through the feature extraction network of the pre-trained prediction model; and determine the heating and cooling load sequence corresponding to the raw data based on the short-term dynamic features and long-term periodic features through the fusion network of the pre-trained prediction model.

[0007] In some embodiments, the controller, in the process of determining the cold and hot load sequence corresponding to the original data based on short-term dynamic features and long-term periodic features through the fusion network of the pre-trained prediction model, is configured to: perform weighted fusion of short-term dynamic features and long-term periodic features based on a gating mechanism through the fusion network of the pre-trained prediction model to obtain a fusion result; and perform smooth activation processing on the fusion result, mapping it to a cold and hot load sequence by a fully connected layer.

[0008] In some embodiments, before inputting raw data into a pre-trained prediction model to obtain the heating and cooling load sequence corresponding to the raw data output by the pre-trained prediction model, the controller is further configured to: determine the current operating condition based on the time range included in the time series attribute feature data; if the current operating condition is a cooling condition, activate the heat balance constraint weights, upper bound constraint weights of the performance coefficient of the cooling equipment, positive monotonicity constraint weights, and boundary condition constraint weights of the physical constraint loss function; if the current operating condition is a heating condition, activate the heat balance constraint weights, upper bound constraint weights of the performance coefficient of the heating equipment, negative monotonicity constraint weights, and boundary condition constraint weights of the physical constraint loss function; if the current operating condition is a transitional condition, activate the heat balance constraint weights of the physical constraint loss function.

[0009] In some embodiments, after inputting raw data into a pre-trained prediction model to obtain the heating and cooling load sequence corresponding to the raw data output by the pre-trained prediction model, the controller is further configured to: determine a prediction error factor based on the heating and cooling load sequence and the actual load sequence; determine a physical residual factor based on the heating and cooling load sequence and the estimated load sequence; wherein the estimated load sequence is calculated based on physical equations; calculate a data volume factor based on the data volume of the raw data; calculate a base value for the physical constraint weights based on the prediction error factor, the physical residual factor, and the data volume factor; and determine a physical constraint loss function for model training based on the base value.

[0010] In some embodiments, the controller, in determining the prediction error factor based on the heating / cooling load sequence and the actual load sequence, is configured to: calculate the prediction error based on the heating / cooling load sequence and the actual load sequence; the prediction error is used to measure the statistical deviation between the heating / cooling load sequence and the actual load sequence; calculate the mean prediction error over a first preset time period, the mean prediction error being used to represent the prediction accuracy over the first preset time period; normalize and scale the mean prediction error to obtain a first scaling result; and perform a hyperbolic tangent function transformation on the first scaling result to obtain the prediction error factor.

[0011] In some embodiments, the controller, in determining the physical residual factor based on the heating / cooling load sequence and the estimated load sequence, is configured to: calculate the physical residual based on the heating / cooling load sequence and the estimated load sequence; the physical residual is used to measure the degree of deviation between the heating / cooling load sequence and the estimated load sequence; calculate the mean physical residual over a second preset time period, the mean physical residual being used to represent the degree to which the pre-trained prediction model deviates from physical laws; normalize and scale the mean physical residual to obtain a second scaling result; and perform a hyperbolic tangent function transformation on the second scaling result to determine the physical residual factor.

[0012] In some embodiments, after calculating the mean prediction error for a first preset time period, or after calculating the mean physical residual for a second preset time period, the controller is further configured to: determine whether the mean prediction error is greater than or equal to a first threshold, or whether the mean physical residual is greater than or equal to a second threshold; if so, increase the physical constraint weights by a preset amount; and determine a physical constraint loss function for incremental training of the model based on the increased physical constraint weights.

[0013] Secondly, this application provides a method for predicting heating and cooling loads based on a pre-trained prediction model, comprising: acquiring raw data, including meteorological environmental sensing data, HVAC equipment operation data, and time-series attribute feature data; inputting the raw data into a pre-trained prediction model to obtain a heating and cooling load sequence corresponding to the raw data output by the pre-trained prediction model; wherein, the pre-trained prediction model is trained based on historical sample data and a physical constraint loss function; the physical constraint loss function is used to guide the update direction of model parameters according to thermodynamic physical laws during model training; the physical constraint weights of the physical constraint loss function include at least one of thermal balance constraint weights, upper bound constraint weights of equipment performance coefficients, monotonicity constraint weights, and boundary condition constraint weights.

[0014] In the above embodiments, a server and a method for predicting heating and cooling loads based on a pre-trained prediction model are provided. This method first integrates three types of core raw data—meteorological environmental perception data, HVAC equipment operation data, and time-series attribute feature data—as model inputs to achieve multi-dimensional, full-domain data-driven support. Simultaneously, a pre-trained prediction model is constructed based on historical sample data and trained using a physical constraint loss function. Thermodynamic physical laws are embedded into the entire model training process, and the physical constraint loss function standardizes the direction of model parameter updates, fundamentally preventing physical violations of thermal mechanisms in the prediction results. Furthermore, physical constraint weights are configured, including at least one of thermal balance constraint weights, upper bound constraint weights for equipment performance coefficients, monotonicity constraint weights, and boundary condition constraint weights, to achieve multi-dimensional, refined physical rule control. By using rigid physical laws to constrain model learning, the drawbacks of conventional mechanism fusion methods—physical constraint failure and the inability of digital twin models to adaptively iterate—are addressed, enabling the accurate output of heating and cooling load sequences that closely match actual operating conditions. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating an application scenario of a server provided in an embodiment of this application; Figure 2 A schematic diagram of the hardware architecture upon which the running server provided in this application embodiment relies; Figure 3 A flowchart illustrating a method for predicting heating and cooling loads based on a pre-trained prediction model, provided in an embodiment of this application; Figure 4 An adaptive variation curve of physical constraint weights provided for embodiments of this application; Figure 5 A flowchart illustrating another method for predicting heating and cooling loads based on a pre-trained prediction model provided in this application embodiment; Figure 6 This is a flowchart illustrating another method for predicting heating and cooling loads based on a pre-trained prediction model, provided in an embodiment of this application. Detailed Implementation

[0016] To make the objectives and implementation methods of this application clearer, the exemplary implementation methods of this application will be clearly and completely described below with reference to the accompanying drawings of the exemplary embodiments of this application. Obviously, the exemplary embodiments described are only some embodiments of this application, and not all embodiments.

[0017] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0018] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar or related objects or entities, and do not necessarily imply a specific order or sequence, unless otherwise specified. It should be understood that such terms are interchangeable where appropriate.

[0019] The terms “comprising” and “having”, and any variations thereof, are intended to cover but not exclude inclusion, for example, a product or device that includes a range of components is not necessarily limited to all of the components that are clearly listed, but may include other components that are not clearly listed or that are inherent to such product or device.

[0020] To more clearly illustrate the embodiments of this application, the technical terms used in the embodiments will be briefly introduced below.

[0021] An integrated energy station (also known as a centralized cooling and heating energy station) is a dedicated energy supply station that generates cooling and heating through centralized cooling and heating source equipment and delivers them to users via outdoor pipelines, providing unified cooling and heating for a specific area (building complex, park, street).

[0022] Heating, ventilation, and air conditioning (HVAC) is a general term for engineering technology systems that regulate and optimize temperature, humidity, airflow velocity, cleanliness, and air quality within a building or enclosed space to meet the comfort requirements of people, equipment operation, and process production.

[0023] Heat and cooling load forecasting is a technical process that uses multi-source information, including building characteristics, meteorological parameters, operational strategies, and user behavior, to quantitatively predict the amount of heat (heat load) or heat (cool load) required per unit time to maintain a set indoor thermal and humidity environment within a future period. Cooling load is the total heat that needs to be removed from the indoor environment per unit time; heating load is the total heat that needs to be added to the indoor environment per unit time.

[0024] Physics-Informed Neural Network (PINN) is a deep learning model that embeds prior physical information, such as physical laws, conservation equations, boundary conditions, and initial conditions, into the neural network structure and training process. By introducing physical constraints into the loss function, the network output simultaneously satisfies both data fitting accuracy and physical consistency, achieving dual-constraint modeling driven by both data and physics. Its core mechanism is to include both data fitting loss and physical residual loss in the loss function, minimizing both data error and violation of physical equations through gradient descent.

[0025] The building thermodynamics physical equations are a set of mathematical equations based on the fundamental laws of thermodynamics, heat and mass transfer mechanisms, and the thermal properties of building envelopes. They describe the temperature field, heat flux density, energy transfer and conversion laws within the building interior and envelope, and are used to quantitatively characterize building heat and moisture transfer, energy conservation, and thermal dynamic response processes.

[0026] Thermodynamic equations are mathematical equations that describe the relationships and changes among physical quantities such as energy, temperature, pressure, volume, entropy, and enthalpy in a substance or system during a thermodynamic process, based on the first law of thermodynamics, the second law of thermodynamics, and the equation of state. They are used to characterize the conservation of energy, energy conversion, and thermodynamic state evolution of a system.

[0027] The heat balance equation is a mathematical equation based on the law of conservation of energy. It characterizes the balance between the changes in total heat input, total heat output, and heat storage per unit time within a controlled body. It is applicable to the steady-state and dynamic thermal process analysis of objects such as buildings, energy systems, equipment, and fluids.

[0028] The physical limit of the coefficient of performance (COP) is the theoretically highest performance coefficient that a refrigeration or heat pump device can achieve under ideal reversible conditions, based on the second law of thermodynamics and the Carnot cycle theory, given the temperature of the heat source and the temperature of the heat sink. It is the thermodynamic limit that the actual equipment COP cannot exceed.

[0029] The upper bound constraint of the COP of the chiller is a thermodynamic feasible domain constraint imposed on the performance coefficient of the actual chiller unit, with the Carnot COP of the chiller as the theoretical upper limit. It is used to ensure that the COP of the chiller output by the control strategy, optimization algorithm or prediction model satisfies the physical laws and the rationality of actual operation.

[0030] The upper limit constraint of heat pump COP is a thermodynamic and physical constraint imposed on the coefficient of performance of heat pump units, based on the theoretical upper limit of the Carnot heat pump COP. It is used to ensure that the heat pump COP does not violate the laws of thermodynamics and has physical consistency and engineering feasibility in system optimization, load prediction and operation control.

[0031] Currently, various technical methods exist in the industry for predicting the heating and cooling loads of integrated energy stations. One approach is to "connect" physical laws with neural networks. Specifically, this involves first using physical simulation to obtain initial load prediction values, and then using a neural network to correct these initial values. However, in this method, the physical simulation and neural network operate independently and are not organically integrated. Besides this, related technologies can be mainly divided into the following four categories, all of which have significant drawbacks: The first category is purely data-driven methods, such as LSTM, Transformer, and XGBoost. These methods treat the prediction of heating and cooling loads as a purely statistical regression problem, and only achieve prediction by mining the input-output mapping relationship in historical data. They do not introduce any physical knowledge related to building thermal performance or equipment operation, resulting in a "black box" defect in the model. Furthermore, the prediction accuracy is extremely poor during the cold start period (i.e., the initial stage without historical data support), which cannot meet the actual operational needs.

[0032] The second category is pure mechanism simulation methods, such as simulation tools like EnergyPlus, DeST, and Green Building Swell. These methods are based on the physical equations of building heat balance and equipment performance curves, and calculate heating and cooling loads through hourly simulations. However, they rely on accurate building and equipment parameter inputs, which can easily lead to parameter deviations in practical applications. Furthermore, the models do not have the ability to learn autonomously and cannot adapt to load fluctuations caused by dynamic factors such as changes in building usage scenarios and equipment aging.

[0033] The third category is the mechanism fusion method. This type of method attempts to combine the advantages of the above two types of methods, using a mechanism model to generate initial load forecast values, and then correcting forecast biases through a data-driven approach. However, in this method, the physical mechanism only plays a role in the forecast preprocessing stage. During the neural network training process, the physical constraints no longer take effect, resulting in the model essentially remaining in the realm of pure data-driven approaches. This not only fails to fundamentally solve the problem of physical violations (such as forecast results not conforming to the building's thermal balance laws), but also fails to improve the poor forecast accuracy during the cold start period.

[0034] The fourth category is the digital twin method. This method establishes a virtual simulation model of the integrated energy station and compares and calibrates it with actual measured data to achieve load forecasting. However, the model parameters of this type of method are basically fixed, and it cannot be adaptively updated according to actual operating data. Furthermore, it lacks artificial intelligence learning capabilities, making it difficult to cope with complex and ever-changing energy supply scenarios and load fluctuations. Its forecasting accuracy and adaptability are both insufficient.

[0035] In summary, all the relevant integrated energy station cooling and heating load forecasting methods have their own shortcomings. They cannot take into account forecasting accuracy, physical rationality, cold start adaptability, and dynamic adjustment capabilities, and are therefore unable to meet the actual needs of refined operation and energy conservation and consumption reduction of energy stations.

[0036] To address all or part of the aforementioned technical problems, this application provides a server whose controller first integrates three core raw data types—meteorological environment perception data, HVAC equipment operation data, and time-series attribute feature data—as model inputs to achieve multi-dimensional, full-domain data-driven support. Simultaneously, it constructs a pre-trained prediction model based on historical sample data and trained using a physical constraint loss function. This embeds thermodynamic physical laws into the entire model training process, and the physical constraint loss function standardizes the direction of model parameter updates, fundamentally preventing physical violations of thermal mechanisms in prediction results. Furthermore, it configures physical constraint weights, including at least one of thermal balance constraint weights, upper bound constraint weights for equipment performance coefficients, monotonicity constraint weights, and boundary condition constraint weights, to achieve multi-dimensional, refined physical rule control. By constraining model learning with rigid physical laws, it overcomes the shortcomings of conventional mechanism fusion methods, such as physical constraint failure and the inability of digital twin models to adaptively iterate, and can accurately output heating and cooling load sequences that fit actual operating conditions.

[0037] like Figure 1 As shown, this server, designed for a centralized cooling and heating integrated energy station scenario, needs to operate in an intelligent energy station environment with automatic data acquisition, automatic equipment control, and real-time meteorological access. Through standard sensors 100 for temperature, flow, power, and meteorological data, operational data is uploaded hourly, providing a stable and reliable input data source for the prediction model. Simultaneously, server 200 is required to have edge computing or cloud computing capabilities to support the training, inference, and physical constraint calculation of the neural network model, meeting the needs for online prediction and automatic updates. The prediction results can be directly used for the start-up and shutdown of equipment 300 in the energy station, load allocation, and energy-saving optimization scheduling.

[0038] In some examples, the controller, in the process of inputting raw data into a pre-trained prediction model and obtaining the heating and cooling load sequence corresponding to the raw data output by the pre-trained prediction model, is configured to: convert meteorological environmental perception data into intermediate vectors with thermodynamic physical meaning through the physical feature engineering layer of the pre-trained prediction model; determine short-term dynamic features and long-term periodic features based on the intermediate vectors, HVAC equipment operation data, and time-series attribute feature data through the feature extraction network of the pre-trained prediction model; and determine the heating and cooling load sequence corresponding to the raw data based on the short-term dynamic features and long-term periodic features through the fusion network of the pre-trained prediction model.

[0039] The above embodiments, through the model's built-in physical feature engineering layer, transform the originally ambiguous and simplistic meteorological environmental perception data into intermediate vectors with clear thermodynamic and physical meanings. This transformation breaks through the surface-level information limitations of the original data, allowing the model to directly understand the physical logic behind the data. This solves the problem of pure data-driven models lacking physical interpretability, laying a solid physical foundation for subsequent accurate predictions. The feature extraction network can simultaneously receive and process three types of key data: intermediate vectors with thermodynamic meaning, HVAC equipment operation data reflecting equipment operating status, and time-series attribute feature data describing time patterns. Through structured analysis, the network not only extracts short-term dynamic features reflecting short-term changes such as weather abrupt changes and equipment start-up and shutdown, but also learns long-term change patterns such as daily cycles, weekly cycles, and holiday patterns. This multi-source fusion feature extraction method ensures that the model's characterization of load changes includes both instantaneous fluctuations and trend patterns, improving the completeness of feature representation. The fusion network, through adaptive weight allocation, balances short-term fluctuation information and long-term pattern information in the prediction output. The final generated cold and heat load sequence not only strictly follows the physical laws contained in the input data, but also accurately matches the actual load change trend. From the architecture level, it ensures that the prediction results meet the high accuracy requirements of deep learning and satisfy the thermodynamic physical rules, completely eliminating the occurrence of unreasonable prediction results such as negative load values ​​and energy efficiency exceeding the limit.

[0040] In some embodiments, the controller, in the process of determining the cold and hot load sequence corresponding to the original data based on short-term dynamic features and long-term periodic features through the fusion network of the pre-trained prediction model, is configured to: perform weighted fusion of short-term dynamic features and long-term periodic features based on a gating mechanism through the fusion network of the pre-trained prediction model to obtain a fusion result; and perform smooth activation processing on the fusion result, mapping it to a cold and hot load sequence by a fully connected layer.

[0041] In the above embodiments, the fusion network employs a gating mechanism to intelligently weight and fuse short-term dynamic features and long-term periodic features. This dynamic weighting method avoids information redundancy or loss caused by simple feature splicing, ensuring that the fusion result simultaneously covers the instantaneous fluctuation details and long-term trend patterns of the load, thus ensuring a comprehensive and accurate representation of load changes. Smoothing activation processing of the fusion result effectively preserves the continuous gradient information of the features, preventing gradient vanishing or violent oscillations during model training. Simultaneously, combined with the mathematical properties of the smoothing activation function, the numerical range of the hot and cold load sequences can be constrained from the source, ensuring that the output load value is always non-negative. This directly eliminates invalid prediction results that violate fundamental thermodynamic laws, such as negative load values ​​and energy efficiency exceeding limits, at the network architecture level, improving the engineering usability of the prediction data.

[0042] In some embodiments, before inputting raw data into a pre-trained prediction model to obtain the heating and cooling load sequence corresponding to the raw data output by the pre-trained prediction model, the controller is further configured to: determine the current operating condition based on the time range included in the time series attribute feature data; if the current operating condition is a cooling condition, activate the heat balance constraint weights, upper bound constraint weights of the performance coefficient of the cooling equipment, positive monotonicity constraint weights, and boundary condition constraint weights of the physical constraint loss function; if the current operating condition is a heating condition, activate the heat balance constraint weights, upper bound constraint weights of the performance coefficient of the heating equipment, negative monotonicity constraint weights, and boundary condition constraint weights of the physical constraint loss function; if the current operating condition is a transitional condition, activate the heat balance constraint weights of the physical constraint loss function.

[0043] The above embodiments automatically identify three operating conditions—cooling, heating, and transitional season—based on time-series attribute feature data, and activate corresponding physical constraints accordingly. This design strictly follows the thermodynamic logic under different operating conditions, ensuring that each constraint conforms to the physical essence of the current scenario. For core operating conditions such as cooling and heating, the system fully activates four types of constraints: heat balance, equipment energy efficiency upper limit, monotonicity, and boundary conditions. Throughout the model training process, it continuously guides parameters to update in a direction that conforms to physical laws, preserving the high-precision fitting capability of deep learning while eliminating invalid predictions that violate thermodynamic laws, such as negative load values, energy efficiency exceeding physical limits, and temperature and load inversely changing. Under the transitional season condition, only the universal heat balance constraint is activated, while inapplicable constraint terms are turned off, avoiding the misleading influence of erroneous constraints on the model and allowing the model to autonomously learn the complex load change patterns of the transitional season. Through the linkage mechanism of automatic operating condition identification and dynamic constraint activation, there is no need for manual configuration of constraint parameters for different operating conditions; the switching of constraints can be completed automatically. This design significantly reduces the cost of model parameter tuning, while also reducing the computational overhead of invalid constraints, thus improving the efficiency of model training and online inference.

[0044] In some embodiments, after inputting raw data into a pre-trained prediction model to obtain the heating and cooling load sequence corresponding to the raw data output by the pre-trained prediction model, the controller is further configured to: determine a prediction error factor based on the heating and cooling load sequence and the actual load sequence; determine a physical residual factor based on the heating and cooling load sequence and the estimated load sequence; wherein the estimated load sequence is calculated based on physical equations; calculate a data volume factor based on the data volume of the raw data; calculate a base value for the physical constraint weights based on the prediction error factor, the physical residual factor, and the data volume factor; and determine a physical constraint loss function for model training based on the base value.

[0045] The above embodiments, after outputting the heating and cooling load sequences, simultaneously compare the actual operating load with the theoretical load calculated from the physical equations to form prediction error factors and physical residual factors, respectively. This allows for accurate determination of the model's fit deviation from real operating conditions and timely identification of deviations from thermodynamic laws in the prediction results. An additional data volume factor automatically matches the constraint level based on the richness of historical field data. The basic weights are calculated using the prediction error factor, physical residual factor, and data volume factor to ensure scientific and stable constraint adjustment. The final calculated basic weights are then incorporated into the loss function used for training, ensuring that each subsequent model iteration is simultaneously constrained by both actual data patterns and thermodynamic mechanisms. This effectively suppresses model drift, prevents bias from extreme samples of noisy data after long-term operation, and guarantees that the prediction results always balance accurate fitting and physical feasibility.

[0046] In some embodiments, the controller, in determining the prediction error factor based on the heating / cooling load sequence and the actual load sequence, is configured to: calculate the prediction error based on the heating / cooling load sequence and the actual load sequence; the prediction error is used to measure the statistical deviation between the heating / cooling load sequence and the actual load sequence; calculate the mean prediction error over a first preset time period, the mean prediction error being used to represent the prediction accuracy over the first preset time period; normalize and scale the mean prediction error to obtain a first scaling result; and perform a hyperbolic tangent function transformation on the first scaling result to obtain the prediction error factor.

[0047] In some embodiments, the controller, in determining the physical residual factor based on the heating / cooling load sequence and the estimated load sequence, is configured to: calculate the physical residual based on the heating / cooling load sequence and the estimated load sequence; the physical residual is used to measure the degree of deviation between the heating / cooling load sequence and the estimated load sequence; calculate the mean physical residual over a second preset time period, the mean physical residual being used to represent the degree to which the pre-trained prediction model deviates from physical laws; normalize and scale the mean physical residual to obtain a second scaling result; and perform a hyperbolic tangent function transformation on the second scaling result to determine the physical residual factor.

[0048] The above embodiments independently calculate the prediction error and physical residual based on the actual load sequence and the physically estimated load sequence, respectively. By statistically analyzing the mean error and mean residual over a fixed period, random deviations caused by short-term daily load fluctuations and instantaneous sensor noise are filtered out. This ensures that the evaluation indicators can stably reflect the overall performance of the model over a period of time, avoiding misjudgments of the model's quality based on single-point data. The mean error and mean residual are normalized and scaled to normalize deviation data of different magnitudes and dimensions into a unified range, providing a basis for comparison and linkage between the two types of indicators. This prevents excessive differences in the original values ​​from causing one indicator to dominate the weight calculation. After hyperbolic tangent function transformation, the final error factor and residual factor are constrained within a fixed and reasonable range. This not only sensitively reflects the risk level when deviations increase but also avoids sudden changes in factor values ​​and drastic fluctuations in constraint weights caused by individual extreme abnormal data, making the entire adjustment process continuous, smooth, and more resistant to interference.

[0049] In some embodiments, after calculating the mean prediction error for a first preset time period, or after calculating the mean physical residual for a second preset time period, the controller is further configured to: determine whether the mean prediction error is greater than or equal to a first threshold, or whether the mean physical residual is greater than or equal to a second threshold; if so, increase the physical constraint weights by a preset amount; and determine a physical constraint loss function for incremental training of the model based on the increased physical constraint weights.

[0050] The above embodiments rely on two core indicators—time-period average prediction error and time-period average physical residual—for judgment. If either indicator exceeds a set threshold, it can quickly detect inaccurate predictions or abnormal states that violate physical mechanisms. Once an anomaly is detected, the overall weight of physical constraints is proactively increased by a fixed preset amount, significantly enhancing the guiding role of thermodynamic mechanisms in subsequent model training. The increased constraint weights are directly substituted into the physical constraint loss function used for incremental training, allowing the model to prioritize updating parameters in a direction that aligns with physical laws and reduces prediction bias during subsequent fine-tuning iterations. This corrects model drift from the training source, ensuring that each incremental optimization has a clear corrective effect.

[0051] like Figure 2 As shown, the server operates based on a hardware architecture that coordinates the edge and the cloud. This hardware architecture includes a field perception layer, an edge control layer, a cloud computing layer, a communication module, and an integrated energy station.

[0052] The field perception layer, composed of various high-precision sensing and acquisition devices, serves as the data source for the prediction system, responsible for acquiring raw physical quantity information. Temperature sensors collect indoor and outdoor ambient temperatures, chilled water supply and return temperatures, and hot water supply and return temperatures, providing basic temperature data for thermodynamic calculations and load projection. Humidity sensors collect indoor and outdoor relative humidity to calculate physical characteristics such as wet-bulb temperature and air enthalpy. Solar radiation and wind speed sensors acquire meteorological data to support calculations of solar heat gain and building envelope heat exchange. Flow sensors, installed at the main power pipe of the energy station, collect water flow data in real time, which, combined with temperature difference calculations, determines the actual cooling and heating loads. The power acquisition module monitors the power consumption of equipment such as chillers, heat pumps, and water pumps in real time, used to calculate the actual energy efficiency ratio of the system, providing a basis for verifying physical constraints.

[0053] The edge control layer consists of industrial-grade edge computing gateways deployed at the energy station site, undertaking tasks such as local inference, data preprocessing, and real-time control. The edge computing gateways are responsible for real-time acquisition of sensor and equipment operation data, performing data cleaning, outlier removal, and physical feature transformation, converting raw data into thermodynamic intermediate variables. The gateways run lightweight models locally to achieve rapid load prediction for short-term future periods, ensuring real-time scheduling of equipment start-up and shutdown. Simultaneously, they perform online verification of physical constraints to ensure that prediction results do not violate thermodynamic rules and maintain uninterrupted prediction functionality even during network outages.

[0054] The cloud computing layer, composed of high-performance servers and storage devices, is responsible for complex model training, iterative optimization of physical constraints, and long-term data management. The high-performance computing servers, equipped with graphics processing units (GPUs), handle offline training, incremental fine-tuning, and large-scale parallel computing of physical information neural networks, supporting dual-branch model structures and multi-physical-constraint loss function calculations. Data storage servers permanently store historical operational data, meteorological data, model parameters, and training logs, providing data support for adaptive weight adjustment and model retraining. Monitoring servers display real-time metrics such as prediction accuracy, physical residuals, and constraint weights, enabling system status visualization and anomaly alerts.

[0055] The communication module is responsible for data transmission between field devices, edge gateways, and cloud servers, ensuring real-time synchronization of meteorological, operational, and forecasting information.

[0056] The energy station receives load forecast results from the edge gateway and automatically executes operations such as equipment start-up and shutdown, load allocation, and energy storage scheduling based on the forecast values, forming a closed-loop hardware system of sensing, computing, and control.

[0057] like Figure 3 As shown, Figure 3 This is a flowchart illustrating a method for predicting heating and cooling loads based on a pre-trained prediction model, provided as an embodiment of this application. The method includes: S301. Obtain raw data, which includes meteorological and environmental perception data, HVAC equipment operation data, and time-series attribute characteristic data.

[0058] The raw data comes from the integrated energy station. Meteorological environmental sensing data includes temperature, humidity, solar radiation, and wind speed; HVAC equipment operation data includes chilled water temperature, hot water supply and return temperatures, flow rate, and equipment power; and time-series attribute data includes time of day, weekday, month, and whether it is a holiday.

[0059] Meteorological environmental data determines the intensity and trend of building heating and cooling demands. Temperature is a key factor affecting heat transfer in the building envelope and indoor heating and cooling loads, determining the overall baseline level of the load; humidity is used to calculate wet-bulb temperature and air enthalpy, supporting the physical calculations of fresh air load and unit energy efficiency limits; solar radiation is the main source of building heat gain, affecting the hourly load fluctuation range; wind speed, by changing the heat transfer coefficient of the building envelope surface, affects the intensity of heat transfer and dissipation, collectively constituting the external meteorological boundary conditions for load forecasting.

[0060] HVAC equipment operation data reflects the current physical operating status and actual energy load level of the integrated energy station. The supply and return water temperature difference and flow rate of chilled and hot water can accurately calculate the current real-time cooling and heating load, providing realistic labels and monitoring signals for model training. Equipment power is used in conjunction with the actual operating COP of the cooling and heating load computer group to provide measured evidence for thermodynamic constraints and physical rationality verification, ensuring that the prediction results do not deviate from the actual operating characteristics.

[0061] Temporal attribute data is used to characterize the periodic and rhythmic patterns of building energy consumption. Time periods can be used to determine intraday energy consumption fluctuations and distinguish between morning and evening peaks, off-peak periods, and low-end periods; weekdays can identify changes in personnel activity and equipment usage intensity between weekdays and weekends, adapting to different energy consumption patterns; months reflect the changes in weather and load trends brought about by seasonal changes, providing a time basis for operating condition judgment and constraint activation; whether it is a holiday can correct for energy consumption deviations on special dates, avoiding prediction distortions caused by unconventional work schedules, allowing the model to comprehensively learn the energy consumption patterns of the time dimension, and improving the stability of long-term predictions.

[0062] In some embodiments, the acquired raw data can be cleaned, outlier removed, and normalized to remove invalid data caused by equipment failure or sensor malfunction, unify data format and units, ensure that the input data meets the input requirements of the model, and provide a reliable basic input for subsequent predictions.

[0063] S302. Input the raw data into the pre-trained prediction model to obtain the cold and heat load sequence corresponding to the raw data output by the pre-trained prediction model.

[0064] The pre-trained prediction model is trained based on historical sample data and a physical constraint loss function. The historical sample data includes meteorological and environmental perception data, HVAC equipment operation data, and time-series attribute feature data for historical periods.

[0065] The physical constraint loss function is used to guide the update direction of model parameters according to thermodynamic physical laws during model training. The physical constraint weights of the physical constraint loss function include at least one of the following: thermal equilibrium constraint weight λ1, COP upper bound constraint weight λ2, monotonicity constraint weight λ3, and boundary condition constraint weight λ4.

[0066] In some embodiments, the physical constraint loss function is composed of the data fitting loss and multiple physical constraint penalty terms, used as a whole to balance data accuracy and physical plausibility during model training. It is based on the data fitting loss, and separately superimposed with constraint terms corresponding to the residuals of the heat balance equation, penalty terms corresponding to violations of the upper bound of equipment energy efficiency ratio, penalty terms corresponding to violations of temperature and load monotonicity, and constraint terms corresponding to the residuals of special operating conditions. Each constraint is weighted according to a set weight, with the heat balance constraint having the highest weight, followed by the upper bound of equipment energy efficiency ratio, then the monotonicity constraint, and finally the boundary condition constraint having the lowest weight. This weighted combination ensures that the model always follows the laws of thermodynamics and actual operating rules during training, guaranteeing that the prediction results do not violate physical laws.

[0067] Among them, the data fitting loss is used to measure the degree of closeness between the model prediction and the actual load, while the four physical constraint penalties correspond to the heat balance rule, the equipment energy efficiency upper limit rule, the monotonic change rule of load and temperature, and the special operating condition boundary rule, respectively. Each constraint participates in the loss calculation according to the set weight.

[0068] The first constraint is thermal balance. Based on the first law of thermodynamics, also known as the principle of energy conservation, the predicted total load must conform to the energy balance relationship composed of heat transfer from the building envelope, heat gain from solar radiation, heat dissipation from internal heat sources, fresh air load, and heat storage and release in the building. When the predicted value deviates too much from the estimated value calculated according to the physical equations, the system will penalize the model, forcing it to return to the reasonable range of energy conservation.

[0069] The second constraint is the upper limit of the equipment's energy efficiency ratio. Based on the second law of thermodynamics, the actual operating energy efficiency ratio of the equipment must not exceed the theoretical Carnot efficiency limit. The corresponding energy efficiency limits are calculated separately for cooling and heating conditions. When the energy efficiency ratio calculated by the model exceeds this safe limit, the system will impose a penalty to avoid physically impossible ultra-high energy efficiency results.

[0070] The third constraint is the monotonicity constraint between temperature and load. Based on the fundamental laws of heat transfer driving force, the higher the outdoor temperature in cooling mode, the greater the cooling load should be; conversely, the lower the outdoor temperature in heating mode, the greater the heat load should be. The system automatically calculates the gradient direction of load change with temperature to determine if it conforms to this law. If a reverse change occurs, a penalty is imposed to ensure that the load change trend aligns with common sense.

[0071] The fourth constraint is the boundary condition constraint for special operating conditions. Based on the theoretical load value under specific operating conditions, when the outdoor temperature equals the indoor design temperature and there is no solar radiation, the heat transfer of the building envelope is zero. At this time, the cooling load is basically equal to the heat generated by indoor personnel, equipment, and lighting. This constraint provides the model with a definite physical reference point, helping the model to stably learn reasonable load relationships.

[0072] For example, the physical constraint loss function = data fitting loss + λ1 × residual of heat balance equation + λ2 × penalty for violation of physical upper bound of COP + λ3 × penalty for violation of monotonicity + λ4 × residual of working condition boundary conditions.

[0073] In some embodiments, the pre-trained prediction model includes a physical feature engineering layer, a feature extraction network, and a fusion network.

[0074] The physical feature engineering layer transforms meteorological environmental perception data into intermediate vectors with thermodynamic physical meaning. This is because meteorological environmental perception data are basic observations without direct thermodynamic meaning, making it difficult for neural networks to automatically extract physical laws such as heat transfer and energy conservation. Therefore, this application uses a physical feature engineering layer to transform meteorological environmental perception data into intermediate variables that conform to HVAC thermodynamic laws and can be directly used by both short-term and long-term forecasting branches. These intermediate variables inherently contain the physical logic of building heat transfer, equipment efficiency, and energy balance, allowing the model to learn the relationship between load and meteorological conditions more quickly and accurately.

[0075] The intermediate vector includes at least one of the following: indoor-outdoor temperature difference ΔT, wet-bulb temperature T. wb Cumulative solar radiation I accum Building thermal inertia index HI, Carnot efficiency upper limit COP carnot Outdoor enthalpy h out .

[0076] The indoor-outdoor temperature difference is obtained by subtracting the indoor set temperature from the outdoor temperature. It is the direct driving force for heat transfer in the building envelope; the greater the temperature difference, the more heat is transferred through the walls and windows. This characteristic is mainly used for thermal balance constraint calculations and for determining whether the load trend with temperature is correct.

[0077] The wet-bulb temperature is calculated from the dry-bulb temperature and relative humidity. It directly determines the outlet water temperature of the cooling tower, thus affecting the operating pressure differential and energy efficiency level of the chiller unit. The wet-bulb temperature is a core parameter constraining the upper limit of the chiller unit's energy efficiency ratio.

[0078] Cumulative solar radiation is a weighted sum of solar radiation values ​​over the past four hours, with higher weights for values ​​closer to the current time. It is used to reflect the lag effect of heat gain from building windows and heat storage in walls. It is mainly used in the calculation of the solar radiation heat gain component within thermal balance constraints.

[0079] The building thermal inertia index is calculated as the exponentially weighted average of the outdoor temperature difference over the past four hours. It is used to quantify the heat storage and release capacity of building walls and roofs, reflecting the degree to which the load lags behind meteorological changes. This index is used to correct the building heat storage and release term in the heat balance constraint.

[0080] The Carnot efficiency upper bound is calculated based on the unit's evaporation and condensation temperatures, representing the theoretically highest energy efficiency the unit can achieve under current meteorological conditions. It serves as a benchmark value for the upper limit constraint on the energy efficiency ratio, ensuring that model predictions do not violate thermodynamic laws.

[0081] The outdoor enthalpy, calculated from the dry-bulb temperature and air humidity, represents the total energy contained in the outdoor air, including both sensible and latent heat, and directly determines the load required for fresh air treatment. It is mainly used in the calculation of the fresh air load term in heat balance constraints.

[0082] The intermediate vectors are shown in Table 1.

[0083] Table 1

[0084] The above embodiments, through the model's built-in physical feature engineering layer, transform originally ambiguous and simplistic meteorological environmental perception data into intermediate vectors with clear thermodynamic and physical meanings. This transformation breaks through the limitations of the surface information in the original data, allowing the model to directly understand the physical logic behind the data. This solves the problem of purely data-driven models lacking physical interpretability and lays a solid physical foundation for subsequent accurate predictions.

[0085] In some embodiments, the current operating condition is identified based on time-series attribute feature data. It is determined whether the condition is a cooling condition, a heating condition, or a transitional season condition.

[0086] For example, the current operating condition is identified based on a preset time range. From May to September each year, it is determined to be the cooling condition, where the building's primary demand is cooling, and the chiller units are in operation. From November to March of the following year, it is determined to be the heating condition, where the building's primary demand is heating, and the boiler or heat pump equipment is in operation. April and October each year are determined to be the transitional season condition, during which both cooling and heating demands exist within the building, and the physical patterns are more complex. See Table 2 for details.

[0087] Table 2

[0088] Under cooling conditions, the weights of the thermal balance constraint, the upper limit constraint of the cooling COP, the positive monotonicity constraint, and the boundary condition constraint are activated; under heating conditions, the weights of the thermal balance constraint, the upper limit constraint of the heating COP, the negative monotonicity constraint, and the boundary condition constraint are activated; under transitional season conditions, the weight of the thermal balance constraint is activated. See Table 3.

[0089] Table 3

[0090] It should be noted that during the transitional season, the monotonicity constraint is disabled because, in the transitional season, there may be situations where south-facing rooms require cooling and north-facing rooms require heating simultaneously. The monotonic relationship between the total load and the outdoor temperature no longer holds. Forcing a monotonicity constraint in either direction under these circumstances would lead the model to learn incorrect physical laws. The boundary condition constraint is disabled because it relies on the theory that the load equals the internal heat when the outdoor temperature equals the indoor temperature. Simultaneous cooling and heating during the transitional season makes it difficult to define the boundary points; therefore, the boundary condition constraint is disabled.

[0091] The thermal balance constraint is always activated because the first law of thermodynamics (conservation of energy) holds true under any operating condition. Whether it is cooling, heating or transitional season, the balance between energy input and output remains unchanged.

[0092] The above embodiments automatically identify and selectively activate applicable physical constraints and deactivate inapplicable constraints for three different operating conditions: cooling, heating, and transitional seasons. This ensures the effectiveness of physical constraints under each operating condition and avoids misleading the model due to incorrect constraints. In particular, it solves the prediction distortion problem in scenarios with coexisting cooling and heating loads and complex physical relationships during transitional seasons, improving the robustness of the model under complex operating conditions such as seasonal changes and extreme weather. It automatically identifies seasonal changes and operating condition transitions over time and dynamically adjusts the constraint activation state to ensure that the model is always trained and predictive using physical rules adapted to the current operating conditions throughout the entire lifecycle of the energy station.

[0093] like Figure 4As shown, the adaptive variation curve of physical constraint weights over runtime exhibits an overall trend of gradually decreasing with runtime and eventually stabilizing at a low level. The model prediction accuracy and weight range at different stages are also labeled, and a temporary adjustment mechanism for extreme conditions is included.

[0094] In the initial stage of system operation, i.e., the cold start period, the basic weight of physical constraints is at its highest level, maintained between 1.5 and 2.0, corresponding to a mean prediction error (MAPE) of approximately 14%. During this stage, data accumulation is insufficient, and the model's prediction accuracy is limited. Therefore, high weights of physical constraints are needed to ensure the thermodynamic rationality of the prediction results and avoid outputs that violate physical laws. As the system enters the growth period, approximately one to three months of operation, the basic weight of physical constraints gradually decreases, stabilizing within the range of 0.7 to 1.2, and the mean prediction error (MAPE) of the model also decreases to approximately 8%. At this point, data gradually accumulates, and the model learns energy consumption patterns that better reflect reality. Therefore, the weight of physical constraints can be appropriately reduced to achieve a balance between physical rationality and data fitting accuracy. After the system has run for more than three months and entered the maturity period, the basic weight of physical constraints further decays, stabilizing at a low level of 0.15 to 0.3, and the mean prediction error (MAPE) of the model also decreases to approximately 5%. At this stage, sufficient data has been accumulated, and the model's prediction accuracy has reached a high level. Physical constraints serve only as a safety net, remaining with extremely low weights to ensure that the prediction results do not violate the fundamental laws of thermodynamics, while fully leveraging the fitting advantages of data-driven models.

[0095] In addition, at any stage of system operation, if extreme weather or special working conditions such as building renovation occur, the basic weight of physical constraints will be temporarily increased by 1.5 times to strengthen the constraint effect of physical constraints and avoid prediction distortion of the model under unconventional working conditions.

[0096] In some embodiments, after converting meteorological environmental perception data into intermediate vectors, the feature extraction network of a pre-trained prediction model is used to determine short-term dynamic features and long-term periodic features based on the intermediate vectors, HVAC equipment operation data, and time-series attribute feature data.

[0097] The feature extraction network can include a short-term prediction branch and a long-term prediction branch. The two branches are responsible for feature learning at different time scales, and the results are then combined through intelligent weighting to finally output a stable and reasonable load forecast.

[0098] The short-term prediction branch, such as the Long Short-Term Memory (LSTM) network, receives hourly data from a preset time period (e.g., 24 hours) to capture hourly load fluctuations, time lags caused by building thermal inertia, and rapid response patterns after sudden weather changes, outputting a feature vector containing short-term dynamic information. The long-term prediction branch (Transformer) receives hourly data from the past 7 days to learn cyclical patterns, differences between weekdays and rest days, and long-term seasonal trends, outputting a feature vector containing long-term cyclical information. See Table 4.

[0099] Table 4

[0100] For example, according to the input window rules preset by the model, the preprocessed historical data is input into two feature extraction branches: the short-term branch inputs hourly data of the past 24 hours to extract short-term dynamic features such as hourly load fluctuations, building thermal inertia delay, and weather change response; the long-term branch inputs hourly data of the past 7 days to extract long-term periodic features such as daily and weekly cycle patterns, differences between weekdays and holidays, and seasonal trends.

[0101] In the above embodiments, the feature extraction network can simultaneously receive and process three types of key data: intermediate vectors with thermodynamic significance, HVAC equipment operation data reflecting the equipment's operating status, and time-series attribute feature data describing time patterns. Through structured analysis, it not only extracts short-term dynamic features reflecting short-term changes such as sudden weather changes and equipment start-ups and shutdowns, but also learns long-term variation patterns such as daily cycles, weekly cycles, and holiday patterns. This multi-source fusion feature extraction method ensures that the model's characterization of load changes includes both instantaneous fluctuations and trend patterns, improving the completeness of feature representation.

[0102] The two branches do not directly output predicted data, but instead output feature vectors of fixed dimensions. In some embodiments, the cold and heat load sequence corresponding to the original data is determined by a fusion network of a pre-trained prediction model based on short-term dynamic features and long-term periodic features.

[0103] Optionally, short-term dynamic features and long-term cyclical features can be dynamically fused using a learnable gating weighting mechanism. The weights of the short-term dynamic features and long-term cyclical features can be determined based on the current operating state and pre-task, making short-term fluctuations and long-term cyclical patterns complementary. The fused features undergo smooth activation processing and are then mapped to a heating / cooling load sequence by a fully connected layer. It is important to emphasize that smooth activation processing mathematically guarantees that the predicted heating / cooling load values ​​are not negative, making it smoother and more stable than ordinary activation methods. This is equivalent to directly embedding physical constraints into the network structure, preventing invalid predictions that do not conform to actual operating conditions from the outset.

[0104] In some embodiments, the feature extraction network may also include only a Long Short-Term Memory (LSTM) network responsible for extracting short-term features. Such a feature extraction network is suitable for use on edge control devices with limited computing power. Although the prediction accuracy may decrease slightly, the overall computational load will be significantly reduced, making it faster and more resource-efficient.

[0105] Feature extraction networks can also employ more advanced long-sequence-specific models, such as the Beyond Efficient Transformer for Long Sequence Time-Series Forecasting (Informer) or the Patch Time Series Transformer (PatchTST), to replace the original Transformer structure. These models have stronger modeling capabilities for ultra-long-period time-series data and are suitable for application scenarios that require advance prediction of ultra-long-term loads of seven to fourteen days or more.

[0106] Feature extraction networks can also use Neural Basis Expansion Analysis for Interpretable TimeSeries Forecasting (N-BEATS) networks, which have the ability to decompose trends and seasons. This network can break down the final load results into trend and seasonal fluctuation components, making the forecasting process more transparent and easier to interpret. It is suitable for scenarios that require a clear explanation of the reasons for load changes.

[0107] Feature extraction networks can also use temporal convolutional networks as the backbone, which support parallel computing and have a training efficiency far superior to traditional long short-term memory networks, increasing training speed by several times. This makes them suitable for real-world operating environments where models require frequent incremental updates and retraining.

[0108] In the above embodiments, the feature extraction network can be flexibly replaced with a single-branch long short-term memory network, a long temporal Transformer variant, a temporal convolutional network, etc., which can adapt to different scenarios such as edge device deployment, ultra-long-term prediction, and frequent retraining.

[0109] In some embodiments, before outputting the final result, the pre-trained prediction model performs a full verification of the predicted values ​​based on preset thermodynamic physical constraints: checking whether the load values ​​conform to energy conservation, whether the equipment energy efficiency ratio does not exceed the physical upper limit, whether the load changes meet the monotonicity requirement, and whether they meet the extreme operating condition boundary conditions. Predicted values ​​that violate physical constraints are corrected to ensure that the output results fully conform to thermodynamic laws, and the final output is a compliant cooling and heating load sequence that can be used for energy station scheduling.

[0110] Regarding the pre-trained prediction model, this application also provides an offline training method for the prediction model. In the initialization phase of the prediction model, this method first acquires building-related design parameters, energy station equipment nameplate parameters, and historical sample data. The building-related design parameters include information such as building area, building orientation, and building envelope materials; the energy station equipment nameplate parameters, such as the chiller's rated energy efficiency ratio and the heat pump's rated power, are used to determine the basic calculation parameters in the physical constraint equations; the historical sample data covers meteorological environmental perception data, HVAC equipment operation data, and time-series attribute feature data for historical periods. Data cleaning and normalization are then performed to provide qualified input samples for model training.

[0111] Then, thermodynamic physical laws are transformed into loss terms that can be used for training. Based on the conventional data fitting loss, residual constraints of the heat balance equation, penalties for violations of the upper limit of equipment energy efficiency ratio, penalties for violations of temperature and load monotonicity, and residuals of special operating conditions are added. The weight of each constraint is allocated according to a two-level weighting rule. First, the basic weights are determined based on the amount of data, prediction error, and physical residuals. Then, the actual weights of each constraint are generated according to a fixed ratio. When there is insufficient data during the cold start phase, the weight of physical constraints is automatically increased, allowing physical knowledge to supplement the virtual data.

[0112] Subsequently, the processed data was trained using short-term and long-term branches respectively. The short-term branch input hourly data from the past 24 hours to learn hourly load fluctuations, building thermal inertia lag, and rapid response patterns to sudden weather changes. The long-term branch input hourly data from the past seven days to learn daily and weekly cycle patterns, differences between weekdays and weekends, holiday patterns, and seasonal trends. After extracting features independently from the two branches, the features were dynamically fused through a gating weighting mechanism to ensure effective complementarity between short-term dynamics and long-term patterns.

[0113] Next, iterative optimization training of the model is performed, using the constructed physical constraint loss function as the optimization objective. The network parameters are continuously updated, and each parameter update is simultaneously constrained by the data fitting effect and thermodynamic laws to avoid outputting results that violate physical laws. During the training process, the loss change is continuously monitored. When the loss converges and the prediction accuracy reaches the preset standard, iterative training is stopped, and the initially trained prediction model is obtained.

[0114] Further testing can be conducted using validation set data that was not used in training to check prediction accuracy and physical plausibility, confirming that the model output does not contain negative load values, energy efficiency ratio exceeding limits, or other violations. Once the test is passed, the model parameters are fixed as the base model for the online prediction phase, awaiting deployment.

[0115] like Figure 5 As shown in the embodiments of this application, another method for predicting hot and cold loads based on a pre-trained prediction model is also provided. This method can start with a basic model that has been trained offline and realize real-time load prediction and adaptive parameter updates during the model inference process, so as to ensure that the model maintains high accuracy and physical rationality in long-term operation.

[0116] S501. Determine the prediction error factor based on the heating and cooling load sequences and the actual load sequences.

[0117] In some embodiments, the prediction error is first calculated based on the heating and cooling load sequences and the actual load sequences. The prediction error is used to measure the statistical deviation between the model's predicted values ​​and the actual loads, reflecting the accuracy of the model. Then, the mean prediction error for a first preset time period is calculated, which represents the prediction accuracy for the first preset time period. Subsequently, the mean prediction error is normalized and scaled to obtain a first scaling result, and then the first scaling result is transformed using a hyperbolic tangent function to obtain the prediction error factor.

[0118] For example, the latest actual operating data is continuously collected on an hourly or daily basis, and the prediction error e1=|Q is calculated first. pred Q real | / Q real Q pred Q is the heating and cooling load sequence output by the pre-trained prediction model. real | This represents the actual load sequence. Then, the mean forecast error (MAPE) for the past seven days is calculated. 7d The mean prediction error (MAPE) of the statistical model over seven consecutive days is used to obtain a value representing the recent prediction accuracy; this mean prediction error (MAPE) 7d Dividing by a fixed reference value of 10 completes the normalization and scaling of the value, bringing the error value into a suitable range for function processing; finally, a hyperbolic tangent function transformation is applied to the scaled result to compress the calculation result into the range of zero to one, yielding the final prediction error factor f2 = tanh(MAPE). 7d / 10). MAPE 7d The larger the value, the larger f2, and the stronger the physical constraints, for example, MAPE. 7d When f = 5%, f2 ≈ 0.46; MAPE 7dWhen the physical constraint weight is 20%, f²≈0.96. If the prediction error remains excessively large over the past seven days, the weight of the physical constraint is automatically increased to prevent the model from deviating from a reasonable range.

[0119] S502. Determine the physical residual factor based on the heating and cooling load sequences and the estimated load sequences.

[0120] Among them, the physical residual is used to measure the degree of deviation between the cold and hot load sequence and the estimated load sequence. The estimated load sequence is obtained by estimating through thermodynamic physical equations.

[0121] In some embodiments, physical residuals are first calculated based on the heating and cooling load sequences and the estimated load sequences. These physical residuals measure the deviation between the model's predicted values ​​and the estimated values ​​from the physical equations, reflecting whether the results conform to physical laws. Then, the mean physical residual for a second preset time period is calculated. This mean physical residual represents the degree to which the pre-trained prediction model deviates from physical laws. The mean physical residual is then normalized and scaled to obtain a second scaling result. Finally, a hyperbolic tangent function transformation is applied to the second scaling result to determine the physical residual factor.

[0122] For example, first calculate the physical residual e2 = |Q pred Q physics | / Q physics Q physics To estimate the load sequence, the average physical residual over the most recent seven days can be calculated to obtain a value representing the degree to which the model deviates from physical laws. Next, this seven-day average physical residual is divided by a fixed reference value of 50 to scale and standardize the value, bringing the residual value into a suitable calculation range. Finally, a hyperbolic tangent function transformation is applied to the scaled result to constrain the calculation result to the interval between zero and one, ultimately obtaining the physical residual factor f3 = tanh(R). 7d / 50). R 7d The larger the value, the larger f3, and the stronger the physical constraints. When the deviation between the prediction results and the physical equations becomes larger, the constraints are automatically strengthened, allowing the model to revert to thermodynamic laws such as energy conservation.

[0123] In the above embodiments, the calculation of the prediction residual and the physical residual is shown in Table 5.

[0124] Table 5

[0125] The above embodiments, after outputting the heating and cooling load sequences, simultaneously compare the actual operating load with the theoretical load calculated from the physical equations, generating prediction error factors and physical residual factors respectively. This can accurately determine the fitting deviation between the model and the actual operating conditions, and promptly identify problems where the prediction results deviate from thermodynamic laws, avoiding the shortcomings of a single evaluation that only considers statistical accuracy and ignores physical rationality.

[0126] S503. Calculate the data volume factor based on the amount of raw data.

[0127] Wherein, the data volume factor f1 = exp( The number of samples (N / 2000) follows an exponential decay law, determined by the existing number of original samples. The larger the number of samples N, the smaller f1 is, and the weaker the physical constraint. For example, when N=2000 samples, f1≈0.37, and when N=34560 samples, f1≈0.00.

[0128] By adding a data volume factor, the constraint level can be automatically matched according to the richness of historical data on site. In the early stage of project commissioning, when data is scarce, the effect of physical constraints will be strengthened to make up for the fitting shortcomings caused by insufficient samples; in the later stage, when data is sufficient, the hard constraints will be gradually weakened, allowing the model to rely more on real operating data to optimize accuracy and adapt to the full life cycle changes of the energy station from cold start to mature operation.

[0129] S504. Calculate the basic value of the physical constraint weight based on the prediction error factor, physical residual factor, and data volume factor.

[0130] The base value λ of the physical constraint weight base =λ min +(λ max λ min )×(0.5×f1+0.3×f2+0.2×f3), which will be referred to as the basic weights for ease of description. Where λ min As a baseline weight lower limit, even with abundant data, some physical constraints will be retained as a safety net; they will not be completely eliminated. For example... Figure 4 As shown, Figure 4 The code also indicates that the lower bound of the physical constraints is 0.1, meaning that no matter how high the model accuracy, the physical constraints will never completely disappear, continuously providing a guarantee of the physical rationality of the prediction results. max The upper limit of the base weight ensures that there are sufficiently strong physical constraints during cold starts or abnormal operating conditions.

[0131] The base weight values ​​will gradually decrease from a high level to a low level as the system runs. During the cold start phase, the weight is maintained at a high level to ensure physical rationality. During the growth phase, the weight is gradually reduced to balance accuracy and constraints. During the maturity phase, the weight is maintained at a low level to ensure fitting effect. At the same time, the weight is temporarily increased under special conditions such as extreme weather or building renovation to strengthen physical constraints.

[0132] S505. Based on the basic values ​​of the physical constraint weights, determine the physical constraint loss function used for model training.

[0133] Based on the basic weights and using a fixed ratio, first calculate the thermal balance constraint weights λ1=λ base ×1.0、 COP Upper bound constraint weight λ2=λ base ×0.8, Monotonicity constraint weight λ3=λ base ×0.5, Boundary condition constraint weight λ4=λ base ×0.3; Among them, the weight of thermal balance constraint is the highest, followed by the weight of equipment energy efficiency upper limit constraint, then the weight of monotonicity constraint, and the weight of boundary condition constraint is the lowest. For example Figure 4 As shown, Figure 4 The document also shows the weight changes of different physical constraints: the weight of thermal balance constraint is based on the basic weight and remains at 1 times the coefficient; the weight of upper limit constraint of equipment energy efficiency ratio is 0.8 times the basic weight; the weight of monotonicity constraint is 0.5 times the basic weight; the weight of boundary condition constraint is 0.3 times the basic weight. The weights of each constraint are adjusted synchronously and adaptively with the basic weight, always following the weight allocation logic that thermal balance constraint has the highest priority and boundary condition constraint has the lowest priority.

[0134] It is important to emphasize that if the weights of the four constraints are adjusted individually, multiple parameters will be required for manual tuning. However, by adjusting the base weights uniformly and then distributing them to the four constraints in a fixed proportion, only one core parameter needs to be controlled. This allows the influence of all constraints to rise and fall synchronously, maintaining the order of importance of each constraint while greatly simplifying system settings.

[0135] The weights of each physical constraint are shown in Table 6.

[0136] Table 6

[0137] The weight ranges for each constraint are shown in Table 7.

[0138] Table 7

[0139] The above embodiments incorporate the calculated base weights into the loss function used for training, ensuring that each subsequent model iteration is simultaneously constrained by both actual data patterns and thermodynamic mechanisms. This effectively suppresses model drift, prevents bias from being skewed by noisy or extreme samples after long-term operation, and guarantees that the prediction results always balance accurate fitting and physical feasibility.

[0140] In some embodiments, based on the aforementioned core thermodynamic constraints, more physical constraints that fit the on-site operating rules can be added according to the actual equipment and system characteristics of the same energy station project, making the prediction results closer to the real working conditions and more practical for engineering.

[0141] For example, for projects equipped with cold or heat storage devices, energy conservation constraints can be added. This constraint requires that the total daily energy storage and release of the energy storage device cannot exceed the upper limit of the device's physical volume, preventing the model from predicting load transfer plans that exceed the actual storage capacity of the equipment and ensuring the execution of scheduling strategies. For projects requiring refined prediction of pump energy consumption, pipeline pressure drop constraints can be added. This constraint follows the basic laws of fluid mechanics; there is a fixed correspondence between chilled water flow rate and pipeline pressure difference, and changes in flow rate directly affect the magnitude of the pressure difference. By adding this physical relationship to the constraint, the model can more accurately calculate pump energy consumption while predicting load, achieving joint prediction of load and energy consumption. For projects with multiple chiller units operating in parallel, unit operation constraints can be added. This constraint requires that the load proportion undertaken by each chiller must be within a reasonable operating range; chillers should be shut down when the load is too low, and inefficient operation is not allowed. Adding this constraint can prevent the model from giving uneconomical scheduling suggestions that do not conform to equipment operating specifications. For projects that emphasize indoor comfort, temperature drift constraints can be added. This constraint requires that the cumulative prediction error of the model should not cause the indoor temperature to exceed the range of human comfort, ensuring that while optimizing load forecasting and system energy saving, the environmental comfort inside the building is always guaranteed, thus achieving a balance between energy saving and comfort.

[0142] It is important to emphasize that this application uses soft constraints instead of hard constraints. Penalties are only applied when the model output exceeds physically reasonable boundaries; numerical values ​​are not directly truncated. Because there are deviations between the actual physical parameters of a building and the design values, hard constraints would make the model too rigid. Soft constraints allow for errors within a reasonable range while continuously guiding the model to output predictions that conform to physical laws.

[0143] Subsequently, based on the latest measured data and the adjusted weights, the model network parameters were iteratively optimized in small batches to complete the continuous updating of the model and avoid the model's accuracy decay due to data drift and changes in operating conditions.

[0144] In some embodiments, after calculating the mean prediction error or the mean physical residual, it is determined whether the mean prediction error is greater than or equal to a first threshold, or whether the mean physical residual is greater than or equal to a second threshold. If so, the physical constraint weights are increased by a preset amount, and then the physical constraint loss function used for incremental training of the model is determined according to the increased physical constraint weights.

[0145] For example, determining the mean prediction error (MAPE) 7d Is it greater than or equal to 10%, or, determine the mean of the physical residuals R? 7dIs it greater than or equal to 20%? If so, increase the physical constraint weights by a factor of 1.5, as shown in Table 7. Then, determine the new physical constraint loss function according to the increased physical constraint weights, and then perform incremental retraining on the prediction model based on the new physical loss function.

[0146] If the mean prediction error (MAPE) over the past seven days remains consistently high, it indicates that the model's predictions are no longer accurate enough. If the residuals of the heat balance equation over the past seven days remain consistently large, it indicates that the model's output is beginning to deviate significantly from physical laws. If any of these anomalies occur, the overall base weights of the physical constraints will be amplified proportionally from the current value, allowing physical laws to play a more dominant role in model training. At this point, the model will be less affected by outlier data, extreme weather, or erroneous measurements, relying more on thermodynamic equations for judgment. This avoids being skewed by outlier samples, ensuring that the prediction results always conform to physical logic and preventing unreasonable or unusable load values.

[0147] Based on this, data from the most recent seven days can be selected as training samples for incremental fine-tuning training. Small updates are made to the model parameters instead of retraining from scratch. This preserves long-term patterns such as seasons and holidays already learned by the model, avoiding the forgetting of historical knowledge. Simultaneously, the training speed is fast, suitable for online real-time execution. The training samples cover the complete energy usage cycle of weekdays and weekends, focusing on the most recent operating status.

[0148] Subsequently, a dual-metric comparison and verification was conducted between the new and old models, using real-world test data from the past three days. The new model was only deployed if it outperformed the old model in both prediction accuracy and physical plausibility. Simultaneously, the entire retraining process was fully recorded in the operations log, including the trigger time, trigger reason, changes in metrics before and after training, and whether the new model was enabled, facilitating review and traceability by operations personnel.

[0149] If the mean prediction error and the mean physical residual both return to within the normal range for three consecutive days, the abnormal state is considered to have been eliminated. Afterward, the base weights will exit the temporary upward adjustment mode and return to the normal calculation state based on the three factors: data volume, prediction error, and physical residuals. The entire process is completely automatic and requires no manual operation or intervention.

[0150] For example, such as Figure 6As shown, the load forecasting and monitoring process is initiated hourly to ensure real-time awareness of the energy station's operational status. Upon execution, the process involves two aspects: firstly, real-time forecasting using a pre-trained prediction model, outputting the model's calculated heating and cooling load sequences; secondly, estimation of theoretical physical load values ​​based on real-time collected operational data using the heat balance equation, providing a physical reference for model prediction. Then, the predicted load values ​​output by the model are compared with the physical load values ​​estimated by the heat balance equation, and the residual between the two is calculated to determine if the mean prediction error exceeds a set threshold. If it does not exceed the threshold, the model maintains its current state and waits for the next hourly cycle to continue monitoring; if the error exceeds the threshold, an incremental retraining process is triggered. After triggering incremental retraining, the model is fine-tuned and optimized using the latest data from the last seven days. After training, the performance metrics of the new and old models are compared. The new model is only deployed if it outperforms the old model in both prediction accuracy and physical plausibility; if the new model's performance is substandard, the original model is retained, and the anomaly is recorded in the operation and maintenance log. After deploying the new model, the prediction error is continuously monitored for three consecutive days. Only when the error does not exceed the threshold for three consecutive days, and the abnormal state is determined to be eliminated, will the physical constraint base weights return to the normal calculation mode based on the three factors of data volume, prediction error, and physical residual. If the error still does not meet the standard, the weights will be increased to further strengthen the physical constraints, and the monitoring process will be re-executed in the next hourly cycle.

[0151] In the above embodiments, a server and a method for predicting hot and cold loads based on a pre-trained prediction model are described. This method embeds thermodynamic physical constraints throughout the model training process, using physical laws as supplementary virtual data. It can output prediction results that conform to physical logic during the cold start phase, reducing reliance on historical data and enabling newly built projects to be usable and accurate immediately upon commissioning. A multi-dimensional physical constraint loss function, including thermal balance, energy efficiency upper bound, monotonicity, and boundary conditions, is constructed. During training, it continuously guides model parameters to update in a direction consistent with physical laws, preserving the high-precision fitting capability of deep learning while ensuring the physical rationality of the prediction results from the source. Basic weights are dynamically calculated using three factors: data volume, prediction error, and physical residuals, and then allocated to each constraint item according to a fixed ratio. High weights are used to strengthen physical constraints during the cold start phase, data and mechanisms are gradually balanced during the growth phase, and low weights are used to ensure fitting accuracy during the maturity phase. Simultaneously, it supports temporary weight adjustments under extreme conditions, ensuring the model maintains optimal performance throughout the entire lifecycle of the energy station without the need for frequent manual parameter adjustments.

[0152] In addition, for three different operating conditions—cooling, heating, and transitional seasons—the system automatically identifies and selectively activates applicable physical constraints while disabling inapplicable ones. This ensures the effectiveness of physical constraints under each operating condition and avoids misleading the model due to incorrect constraints. In particular, it solves the prediction distortion problem in scenarios with coexisting cold and heat loads and complex physical relationships during transitional seasons, significantly improving the model's robustness under complex operating conditions such as seasonal changes and extreme weather.

[0153] To address the performance drift issue during long-term model operation, this method constructs a dual-path residual online monitoring system that simultaneously tracks prediction accuracy and physical plausibility, automatically triggering incremental retraining when deviations exceed limits. By using only recent data for minor adjustments, it avoids the catastrophic forgetting associated with full retraining while ensuring the model quickly adapts to changes such as building aging, equipment degradation, and climate change. This achieves self-maintenance and self-optimization throughout the model's lifecycle, reducing operational costs.

[0154] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0155] For ease of explanation, the above description has been provided in conjunction with specific embodiments. However, the above exemplary discussion is not intended to be exhaustive or to limit the embodiments to the specific forms disclosed above. Various modifications and variations can be obtained based on the above teachings. The selection and description of the above embodiments are for the purpose of better explaining the principles and practical applications, thereby enabling those skilled in the art to better utilize the described embodiments and various different variations of embodiments suitable for specific use considerations.

Claims

1. A server, characterized in that, include: The controller is configured as follows: Acquire raw data, which includes meteorological and environmental sensing data, HVAC equipment operation data, and time-series attribute characteristic data; The raw data is input into the pre-trained prediction model to obtain the cold and heat load sequence corresponding to the raw data output by the pre-trained prediction model. The pre-trained prediction model is trained based on historical sample data and a physical constraint loss function. The physical constraint loss function is used to guide the update direction of model parameters according to thermodynamic physical laws during model training. The physical constraint weights of the physical constraint loss function include at least one of thermal balance constraint weights, upper bound constraint weights of equipment performance coefficients, monotonicity constraint weights, and boundary condition constraint weights. The controller, after inputting the raw data into the pre-trained prediction model to obtain the cold and heat load sequence corresponding to the raw data output by the pre-trained prediction model, is further configured to: calculate the prediction error based on the cold and heat load sequence and the actual load sequence, and obtain the prediction error factor through time-period mean, normalization, and hyperbolic tangent transformation; calculate the physical residual based on the cold and heat load sequence and the estimated load sequence from the heat balance physical equation, and obtain the physical residual factor through time-period mean, normalization, and hyperbolic tangent transformation; calculate the data volume factor in exponential decay form based on the data volume of the raw data; calculate the base value of the physical constraint weights based on the prediction error factor, the physical residual factor, and the data volume factor; determine the physical constraint loss function used for model training based on the base value; calculate the mean prediction error and the mean physical residual for a preset time period; if the mean prediction error is greater than or equal to a first threshold, or the mean physical residual is greater than or equal to a second threshold, then increase all physical constraint weights by a preset magnitude, and perform incremental training on the pre-trained prediction model based on the updated physical constraint loss function.

2. The server according to claim 1, characterized in that, The controller is configured to: input the raw data into a pre-trained prediction model to obtain the heating and cooling load sequence corresponding to the raw data output by the pre-trained prediction model. The meteorological environment perception data is converted into an intermediate vector with thermodynamic physical meaning through the physical feature engineering layer of the pre-trained prediction model. The feature extraction network of the pre-trained prediction model determines short-term dynamic features and long-term periodic features based on the intermediate vector, the HVAC equipment operation data, and the time-series attribute feature data. The cold and heat load sequence corresponding to the original data is determined by the fusion network of the pre-trained prediction model based on the short-term dynamic features and long-term periodic features.

3. The server according to claim 2, characterized in that, The controller, in the process of determining the cold and heat load sequence corresponding to the original data based on the short-term dynamic features and long-term periodic features through the fusion network of the pre-trained prediction model, is configured as follows: The fusion network of the pre-trained prediction model is used to perform weighted fusion of the short-term dynamic features and the long-term periodic features based on a gating mechanism to obtain the fusion result. The fusion result is smoothed and activated, and mapped from the fully connected layer to the hot and cold load sequence.

4. The server according to claim 1, characterized in that, The controller is further configured to: input the raw data into the pre-trained prediction model to obtain the cold and heat load sequence corresponding to the raw data output by the pre-trained prediction model. The current operating condition is determined based on the time range included in the time-series attribute feature data; If the current operating condition is a cooling condition, then activate the heat balance constraint weight, the upper bound constraint weight of the performance coefficient of the cooling equipment, the positive monotonicity constraint weight, and the boundary condition constraint weight of the physical constraint loss function. If the current operating condition is a heating condition, then activate the heat balance constraint weight, the upper bound constraint weight of the performance coefficient of the heating equipment, the negative monotonicity constraint weight, and the boundary condition constraint weight of the physical constraint loss function. If the current operating condition is a transitional condition, then the thermal balance constraint weight of the physical constraint loss function is activated.

5. The server according to claim 1, characterized in that, The controller, in determining the prediction error factor based on the heating / cooling load sequence and the actual load sequence, is configured as follows: The prediction error is calculated based on the heating and cooling load sequence and the actual load sequence; the prediction error is used to measure the statistical deviation between the heating and cooling load sequence and the actual load sequence. Calculate the mean prediction error for the first preset time period, whereby the mean prediction error is used to represent the prediction accuracy for the first preset time period. The mean of the prediction error is normalized and scaled to obtain a first scaling result; The prediction error factor is obtained by performing a hyperbolic tangent function transformation on the first scaling result.

6. The server according to claim 1, characterized in that, The controller, in the process of determining the physical residual factor based on the heating and cooling load sequences and the estimated load sequences, is configured to: Based on the heating and cooling load sequence and the estimated load sequence, a physical residual is calculated; the physical residual is used to measure the degree of deviation between the heating and cooling load sequence and the estimated load sequence. Calculate the mean physical residual for the second preset time period, whereby the mean physical residual is used to represent the degree to which the pre-trained prediction model deviates from the physical laws. The mean of the physical residuals is normalized and scaled to obtain a second scaling result; The second scaling result is subjected to a hyperbolic tangent function transformation to determine the physical residual factor.

7. The server according to claim 2, characterized in that, The intermediate vector includes at least one of the following: indoor-outdoor temperature difference, wet-bulb temperature, cumulative solar radiation, thermal inertia index, upper limit of Carnot efficiency, and outdoor enthalpy.

8. A method for predicting heating and cooling loads based on a pre-trained prediction model, characterized in that, include: Acquire raw data, which includes meteorological and environmental sensing data, HVAC equipment operation data, and time-series attribute characteristic data; The raw data is input into a pre-trained prediction model to obtain the heating and cooling load sequence corresponding to the raw data output by the pre-trained prediction model; wherein, the pre-trained prediction model is trained based on historical sample data and a physical constraint loss function; the physical constraint loss function is used to guide the update direction of model parameters according to thermodynamic physical laws during model training; the physical constraint weights of the physical constraint loss function include at least one of thermal balance constraint weights, upper bound constraint weights of equipment performance coefficients, monotonicity constraint weights, and boundary condition constraint weights; After inputting the original data into the pre-trained prediction model to obtain the cold and heat load sequence corresponding to the original data output by the pre-trained prediction model, the method further includes: calculating the prediction error based on the cold and heat load sequence and the actual load sequence, and obtaining the prediction error factor through time-period mean, normalization, and hyperbolic tangent transformation; calculating the physical residual based on the cold and heat load sequence and the estimated load sequence from the heat balance physical equation, and obtaining the physical residual factor through time-period mean, normalization, and hyperbolic tangent transformation; calculating the data volume factor in exponential decay form based on the data volume of the original data; calculating the base value of the physical constraint weights based on the prediction error factor, the physical residual factor, and the data volume factor; determining the physical constraint loss function used for model training based on the base value; calculating the mean prediction error and the mean physical residual for a preset time period; if the mean prediction error is greater than or equal to a first threshold, or the mean physical residual is greater than or equal to a second threshold, then increasing all physical constraint weights by a preset magnitude, and performing incremental training on the pre-trained prediction model based on the updated physical constraint loss function.

Images