A multi-source data fusion-based intelligent heating system load prediction method

Abstract

The application relates to the technical field of heat supply, and discloses a load prediction method for a smart heat supply system based on multi-source data fusion, which comprises the following steps: acquiring multi-dimensional data containing multi-dimensional influence factors through a multi-dimensional data acquisition system; enhancing the quality of the multi-dimensional data and performing standardized conversion; establishing a space-time graph attention network to extract space-time joint features in the multi-dimensional data; constructing a multi-scale prediction model cluster to perform multi-scale load prediction on the space-time fusion features and output initial load prediction results; quantifying the uncertainty of the initial load prediction results to obtain load prediction values with confidence intervals; and performing online incremental learning and adaptive parameter updating on the prediction model according to actual operation feedback data. The method has the advantages of improving load prediction accuracy, enhancing system robustness, supporting real-time decision-making and adapting to working condition changes.

Description

Technical Field

[0001] This application relates to the field of heating technology, and for example to a method for load prediction of a smart heating system based on multi-source data fusion. Background Technology

[0002] With the rapid development of the smart energy industry, centralized heating systems, as a core component of urban energy consumption, directly impact energy utilization efficiency, user heating experience, and carbon emission control through their operational efficiency and control precision. Heat load forecasting is a crucial prerequisite for smart heating systems to achieve precise scheduling and energy conservation; its accuracy and real-time performance directly determine the operational quality of the heating system. Accurate load forecasting can avoid energy waste caused by excessive heat source output and prevent insufficient indoor temperature for users due to insufficient load. However, current heating system load forecasting technology faces multiple challenges in practical applications. At the data fusion level, existing methods collect relatively limited data dimensions, mostly confined to basic meteorological data such as historical loads and outdoor temperature and humidity. They fail to fully integrate key information such as building thermal properties parameters like the heat transfer coefficient of the building envelope, insulation level, and window-to-wall ratio, and also neglect dynamic factors such as user heating behavior data, including indoor temperature set patterns, occupancy rate fluctuations, and window opening frequency. Meanwhile, multi-dimensional heterogeneous data, such as pipeline hydraulic characteristics data (supply and return water temperature, pressure, flow rate, thermal inertia coefficient, and leakage rate) and heat source equipment operating data (boiler output status and energy efficiency), have not been effectively incorporated. Furthermore, multimodal data, including numerical, time-series, and remote sensing imagery, are only processed through simple stitching, lacking deep semantic fusion mechanisms, resulting in insufficient mining of potential correlations within the data. Regarding data quality, outliers and missing values ​​caused by sensor failures or communication interruptions are prevalent. Existing technologies rely solely on crude processing methods such as mean filling or data deletion, without establishing enhancement mechanisms based on multi-source cross-verification and spatiotemporal correlation, further weakening the reliability and predictive foundation of the data.

[0003] At the spatiotemporal feature modeling level, existing technologies generally neglect the spatiotemporal coupling characteristics of heating systems. The heat transfer process in heating networks is dynamically affected by pipeline distance, resistance characteristics, and flow relationships, but existing methods fail to construct a spatial graph structure reflecting the physical relationships of heat transfer, thus failing to capture the spatial dependencies between nodes. Simultaneously, load changes include both rapid responses caused by short-term meteorological fluctuations and slow evolutions due to long-term seasonal changes. Existing models are insufficient in extracting the long- and short-term temporal dependencies of the load, resulting in fragmented spatiotemporal features and making it difficult to accurately characterize the dynamic laws of heat load propagation and temporal evolution. In predictive model design, many adopt single neural network or static weight fusion architectures, failing to embed physical laws such as building heat balance equations as constraints into the model, and failing to coordinate the advantages of data-driven and mechanistic models. Purely data-driven methods have poor robustness in extreme weather or equipment failure scenarios, while purely physical models, due to their high computational complexity, cannot meet real-time requirements and cannot flexibly adapt to multi-scale needs such as short-term fine-grained predictions, medium-term periodic predictions, and long-term trend predictions.

[0004] At the engineering application level, existing prediction models generally suffer from a large number of parameters and slow inference speed, making them difficult to deploy on edge computing devices for real-time decision-making at the heating site. Load forecasting and scheduling are disconnected, lacking a hierarchical prediction system, resulting in prediction results that cannot support end-to-end collaborative optimization. After model training, parameters are fixed, lacking adaptability to seasonal changes, building aging, or evolving user habits, leading to a continuous decline in prediction accuracy over time. Furthermore, prediction results only output single values, failing to quantify uncertainties such as weather forecast errors and random user behavior, and cannot provide confidence intervals or risk probability distributions, leaving scheduling decisions without a scientific basis for potential risks. These shortcomings collectively make it difficult for existing technologies to meet the actual operational needs of smart heating systems in terms of prediction accuracy, system robustness, and engineering practicality.

[0005] Therefore, it is evident that how to meet the actual operational requirements of smart heating systems in terms of prediction accuracy, system robustness, and engineering practicality has become a technical problem that urgently needs to be solved by those skilled in the art.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments, but rather as a prelude to the detailed description that follows.

[0008] This disclosure provides a load prediction method for a smart heating system based on multi-source data fusion, in order to solve technical problems such as shallow data fusion, insufficient capture of spatiotemporal features, disconnect between physical mechanisms and data-driven approaches, poor engineering adaptability, and lack of quantification of uncertainties.

[0009] In some embodiments, the load forecasting method for a smart heating system based on multi-source data fusion includes: A multidimensional data acquisition system is used to obtain multidimensional data containing multiple influencing factors. Enhance the quality and standardize multidimensional data; Establish a spatiotemporal graph attention network to extract spatiotemporal joint features from multidimensional data; A multi-scale prediction model cluster is constructed, and the building heat balance equation and the pipe network thermal inertia model are embedded into the model cluster as physical constraints. Dynamic fusion weights are assigned to each model based on the scene perception strategy, and multi-scale load prediction is performed on the spatiotemporal fusion features to output the initial load prediction results. Uncertainty quantification is performed on the initial load forecast results to obtain load forecast values ​​with confidence intervals; The load forecast values ​​are aggregated and corrected step by step according to the hierarchical structure of heat source-pipeline network-heat exchange station-user. At the same time, the forecast model is subjected to online incremental learning and adaptive parameter updates based on actual operation feedback data.

[0010] This disclosure provides a method for load forecasting of a smart heating system based on multi-source data fusion, which can achieve the following technical effects: By acquiring multidimensional data containing multiple influencing factors, performing quality enhancement and standardization transformation, establishing a spatiotemporal graph attention network to extract features, constructing a multi-scale prediction model cluster embedded with physical constraints, quantifying uncertainty, and realizing hierarchical aggregation and online updates, the problems of insufficient data fusion, quality defects, missing spatiotemporal modeling, and limitations in engineering applications are solved. It has the advantages of improving load forecasting accuracy, enhancing system robustness, supporting real-time decision-making, and adapting to changes in operating conditions.

[0011] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0012] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations and drawings do not constitute a limitation on the embodiments. Elements having the same reference numerals in the drawings are shown as similar elements. The drawings are not to be scaled. And wherein: Figure 1 This is a schematic diagram of a smart heating system load prediction method based on multi-source data fusion provided in an embodiment of this disclosure; Figure 2This is a system block diagram of the intelligent heating system load prediction method based on multi-source data fusion provided in the embodiments of this disclosure. Detailed Implementation

[0013] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0014] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0015] Unless otherwise stated, the term "multiple" means two or more.

[0016] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0017] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0018] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0019] Combination Figure 1-2 In some embodiments, a method for load forecasting of a smart heating system based on multi-source data fusion includes: S01, acquire multidimensional data containing multidimensional influencing factors through a multidimensional data acquisition system; S02, performs quality enhancement and standardization transformation on multidimensional data; S03, Establish a spatiotemporal graph attention network to extract spatiotemporal joint features from multidimensional data; S04. Construct a multi-scale prediction model cluster, embed the building heat balance equation and the pipe network thermal inertia model as physical constraints into the model cluster, and assign dynamic fusion weights to each model based on the scene perception strategy. Perform multi-scale load prediction on the spatiotemporal fusion features and output the initial load prediction results. S05, perform uncertainty quantification on the initial load forecast result to obtain the load forecast value with confidence interval; S06. The load forecast value is aggregated and corrected step by step according to the hierarchical structure of heat source-pipeline network-heat exchange station-user. At the same time, the forecast model is subjected to online incremental learning and adaptive parameter updates based on actual operation feedback data.

[0020] Among them, the multi-dimensional data acquisition system is used to acquire different types of data from all aspects of the heating system, including heat sources, pipe networks, heat exchange stations, user terminals, and external environments such as meteorological monitoring stations.

[0021] Multidimensional data contains a variety of factors that affect heating load, including meteorological data, building data, user data, pipeline network data, heat source data, time series data, and external data. The above data are heterogeneous and cover all aspects of the operation of the heating system.

[0022] Quality enhancement and standardization transformation involves preprocessing the original multidimensional data to improve its quality and make it meet the input requirements of subsequent models. This typically includes steps such as outlier removal, missing value completion, and data normalization.

[0023] Spatiotemporal graph attention network is a neural network model that can simultaneously process the time series features and spatial topological features of heating system data. This network introduces an attention mechanism, which can capture the complex relationships between different nodes in the system, such as heat sources, heat exchange stations, user buildings, and between different time steps.

[0024] Spatiotemporal joint features are deep information extracted from multidimensional data after quality enhancement and standardization. These features can simultaneously reflect the interaction and dynamic change patterns of the heating system in both time and space dimensions, providing more comprehensive information support for load forecasting.

[0025] A multi-scale forecasting model cluster is a collection of forecasting models designed for different time scales, such as short-term, medium-term, and long-term. Each model undertakes the task of load forecasting within a specific time range to meet different scheduling requirements.

[0026] Physical constraints are the mathematical embedding of fundamental physical laws such as energy conservation and heat and mass transfer equations of heating systems into the prediction model. Introducing physical constraints can improve the physical rationality of the model, enhance its robustness under complex operating conditions, and improve prediction accuracy.

[0027] Scene-aware strategy is a method that dynamically adjusts the parameters or fusion weights of each model in the model cluster based on the current operating status of the heating system, changes in the external environment, or specific prediction task requirements, so that the model can adapt to different operating scenarios.

[0028] Uncertainty quantification is a method for assessing the potential error range and confidence level of prediction results. It is usually presented in the form of confidence intervals or probability distributions and can provide information on the reliability of predictions and assessment of potential risks for decision-making.

[0029] A confidence interval is a statistical concept used to represent the range of predicted values ​​at a certain probability level. This interval reflects the reliability and accuracy of the prediction results; the narrower the interval, the higher the prediction accuracy usually is.

[0030] Online incremental learning and adaptive parameter updates refer to the ability of a prediction model to continuously learn and adjust its parameters using real-time feedback data during actual operation. This mechanism enables the model to adapt to long-term operating conditions such as seasonal changes, building aging, and changes in user heating habits, thus maintaining the long-term stability and accuracy of predictions.

[0031] In some embodiments, the load forecasting method for smart heating systems based on multi-source data fusion first acquires multi-dimensional data containing multiple influencing factors through a multi-dimensional data acquisition system. For example, the acquisition system can be deployed at various stages such as heat sources, pipe networks, heat exchange stations, and user terminals, and combined with external meteorological monitoring equipment to acquire various types of data related to heating operation. After acquiring the multi-dimensional data, it needs to undergo quality enhancement and standardization transformation. For example, data quality enhancement may include outlier detection and removal from the original data, and completion of missing data; standardization transformation can unify data of different dimensions to a similar numerical range, such as by using linear scaling or Z-score standardization. As one implementation method, outliers can be removed by simple threshold judgment, and missing values ​​can be filled using the mean or median; alternatively, standardization transformation can be completed by simply using maximum and minimum value normalization. Subsequently, a spatiotemporal graph attention network is constructed to extract spatiotemporal joint features from the multi-dimensional data. This network can handle the spatial correlation between nodes within the heating system and capture the dynamic dependence of load changes over time. For example, a topology graph representing the physical connections of the heating system can be constructed, and the multi-dimensional data can be input into the network as node features.

[0032] Furthermore, a multi-scale prediction model cluster is constructed, embedding the building heat balance equation and the pipe network thermal inertia model as physical constraints into the model cluster. This model cluster can assign dynamic fusion weights to each model based on a scene-aware strategy, perform multi-scale load prediction on spatiotemporal fusion features, and output initial load prediction results. For example, the model cluster can contain multiple sub-models, each optimized for different prediction time scales. The embedding of physical constraints ensures that the prediction results conform to fundamental physical laws such as energy conservation.

[0033] Uncertainty quantification is applied to the initial load forecast results to obtain load forecast values ​​with confidence intervals. Uncertainty quantification is used to assess the reliability of the forecast results and to provide the possible range of values ​​for the forecast. For example, statistical methods can be used to estimate the forecast error distribution and then calculate the confidence interval.

[0034] Finally, the load forecast values ​​are aggregated and corrected level by level according to the hierarchical structure of heat source, pipeline network, heat exchange station, and user. Simultaneously, online incremental learning and adaptive parameter updates are performed on the forecast model based on actual operational feedback data. Hierarchical aggregation and correction ensure consistency in load forecasts across different levels, while the online incremental learning mechanism allows the model to continuously optimize its performance based on new operational data. For example, when actual load data is fed back, the model can use the new data to fine-tune its parameters.

[0035] The method proposed in this embodiment effectively solves the problems of single data dimension and insufficient data quality in traditional methods through multi-dimensional data acquisition and quality enhancement. By constructing a spatiotemporal graph attention network to extract spatiotemporal joint features, it overcomes the deficiency of fragmented spatiotemporal features. Furthermore, by constructing a multi-scale prediction model cluster embedded with physical constraints and combining it with a scene-aware strategy to achieve dynamic fusion, it achieves a deep integration of physical mechanisms and data-driven approaches, improving prediction accuracy and robustness. In addition, uncertainty quantification provides a risk assessment basis for decision-making, and hierarchical aggregation and online adaptive update mechanisms significantly improve the engineering adaptability and long-term stability of the prediction system, providing reliable support for the precise scheduling of smart heating systems.

[0036] Optionally, multidimensional data containing multiple influencing factors specifically includes: Meteorological data, such as real-time / forecasted temperature and humidity, wind speed, wind direction, solar radiation, precipitation, air pressure, and extreme weather indicators, are crucial external environmental factors affecting heating load. Real-time / forecasted temperature and humidity directly determine the temperature difference between the inside and outside of a building and the heat demand. Wind speed and direction affect convective heat loss and infiltration ventilation. Solar radiation provides free heat and reduces heating demand. Precipitation and air pressure indirectly affect perceived temperature and building heat loss. Extreme weather indicators are used to identify abnormal load conditions. This data can be obtained through weather stations, satellite remote sensing, or professional meteorological service interfaces.

[0037] Building data, such as building type, area, heat transfer coefficient of the building envelope, insulation grade, window-to-wall ratio, orientation, and year of construction, reflects the physical characteristics and thermal performance of the object being heated. Building type determines its inherent heat load pattern; area is the basis for calculating the total heat load; the heat transfer coefficient of the building envelope, insulation grade, and window-to-wall ratio directly affect the building's heat loss; orientation affects the amount of solar radiation received; and the year of construction may be related to factors such as structural aging and deterioration of insulation performance. This data is typically derived from architectural drawings, as-built documentation, or on-site surveys.

[0038] User data, such as indoor temperature settings, daily routines, window opening frequency, heating habits, indoor occupancy rate, and user complaint data, is crucial. User behavior is a significant uncertainty factor in heating load forecasting. Indoor temperature settings directly reflect users' comfort needs, while behavioral patterns such as daily routines, window opening frequency, heating habits, and indoor occupancy rate significantly influence actual heat consumption. User complaint data can serve as feedback on abnormal loads or system problems, assisting in model correction. This data can be obtained through smart thermostats, user surveys, smart home systems, or customer service records.

[0039] Pipeline network data, such as supply and return water temperature / pressure / flow rate, pipeline length / diameter / resistance, thermal inertia coefficient, leakage rate, and heat exchange station parameters, describes the physical state and system characteristics during heat transmission and distribution. Supply and return water temperature, pressure, and flow rate are key parameters for pipeline operation, directly reflecting heat transfer efficiency and load distribution. Pipeline length, diameter, and resistance determine hydraulic balance and heat loss, while thermal inertia coefficient and leakage rate reflect the dynamic response characteristics and operating losses of the pipeline network. Heat exchange station parameters affect the transfer of heat to users. This data can be obtained through pipeline sensors, SCADA systems, or design documents.

[0040] Heat source data, such as boiler / heat pump / peak-shaving equipment output, efficiency, energy consumption, start-up / shutdown status, and heating medium parameters, reflects the supply capacity and operating status of the heating system. The output, efficiency, and energy consumption of boilers, heat pumps, and peak-shaving equipment directly relate to the heat source's production capacity and economic efficiency. Start-up / shutdown status affects the system's dynamic response, while heating medium parameters determine the heat source's output characteristics. This data is typically collected by the heat source control system or energy management system.

[0041] Time-series data, such as historical load, date type, season, and time period labels, captures the regularity of heating load changes over time. Historical load is the basis for predicting future loads, reflecting the inherent periodicity and trends of the system. Date type, season, and time period labels are used to identify load patterns at different time scales. This data is typically extracted from historical operating databases.

[0042] External data, such as social activities, weather forecast errors, equipment failure logs, and outdoor thermal / remote sensing imagery, provides additional, non-traditional information that may affect heating loads. Social activities may cause regional load fluctuations, weather forecast errors can be used to assess the sensitivity of predictive models to uncertainties, equipment failure logs can explain abnormal loads or system outages, and outdoor thermal / remote sensing imagery can help assess the overall heat loss of a building complex or the status of district heating. This data can be obtained through public information, system logs, or specific sensors.

[0043] Optionally, the multidimensional data undergoes quality enhancement and standardization transformation, including: sequentially performing outlier removal, missing value completion, and standardization preprocessing on the multi-source data to obtain a high-quality multi-source dataset. Outlier removal employs a combination of the 3σ criterion and an isolation forest approach, while missing value completion utilizes a combination of multi-source data cross-validation and temporal interpolation.

[0044] Specifically, outlier removal identifies and removes observations that deviate from the normal range in the data. These outliers are often caused by sensor malfunctions, data transmission errors, or transient system disturbances. If left untreated, they introduce noise and reduce data purity. Besides combining the 3σ criterion with isolated forests, box plots and local anomaly factor algorithms can also be used to identify outlier data points. Missing value imputation fills in data gaps caused by sensor malfunctions, data acquisition interruptions, etc., maintaining data integrity and preventing bias in subsequent model training. In addition to multi-source data cross-validation and temporal interpolation, mean, median, and mode imputation, or imputation based on regression models and K-nearest neighbors algorithms, can also be used. Standardization preprocessing eliminates the influence of differences in units and numerical ranges of different features, ensuring all features are on the same order of magnitude, accelerating model convergence, and improving training stability and prediction accuracy. Common standardization methods include Min-Max scaling and Z-score standardization. The dataset obtained after these processing steps has higher accuracy, completeness, and consistency, providing reliable input for subsequent spatiotemporal graph attention networks and multi-scale prediction model clusters.

[0045] In the outlier removal stage, this application employs a combination of the 3σ criterion and the Isolation Forest method. The 3σ criterion is a statistical method applicable to data that follows or approximately follows a normal distribution. Its principle is that about 99.7% of data points in a normal distribution fall within the mean ± 3 standard deviations; data points outside this range are considered outliers. This method is simple and efficient, but it has certain requirements regarding the data distribution. The Isolation Forest is an unsupervised anomaly detection algorithm based on ensemble learning. Its core idea is that outliers are sparse and far from normal points in the data space, making them easier to "isolate." This algorithm constructs multiple isolation trees by randomly selecting features and split points, and outliers are usually isolated at shallower levels. Combining the 3σ criterion with the Isolation Forest leverages the advantages of both: the 3σ criterion effectively identifies numerical anomalies conforming to a normal distribution, while the Isolation Forest can handle complex multidimensional anomalies in non-normally distributed data, especially suitable for high-dimensional data. This combination improves the robustness and accuracy of outlier detection, reducing false positives and false negatives.

[0046] In the missing value completion stage, this application employs a combined approach of multi-source data cross-validation and time-series interpolation. Multi-source data cross-validation utilizes the inherent physical or logical relationships between different data sources to verify and fill in missing values. For example, when the water supply temperature data of a heat exchange station is missing, upstream pipeline water supply temperature, downstream user return water temperature, and meteorological data can be referenced. Inference and verification are completed by constructing a data relationship model. This method utilizes redundant information within the system to improve the rationality of the completion. Time-series interpolation utilizes the continuity and trend of time series data to complete the completion. Common methods include linear interpolation, spline interpolation, moving average, exponential smoothing, and predictive completion based on time series models such as ARIMA and LSTM. The combined approach of multi-source data cross-validation and time-series interpolation means that verification and preliminary completion are first completed based on the physical relationships between different data sources. If missing values ​​still exist or the verification results are uncertain, interpolation is then performed based on the inherent patterns of the time series. This combined strategy fully utilizes the spatial relationships and temporal continuity of the data, making the missing value completion results more accurate and closer to actual physical laws, avoiding the bias caused by simple completion.

[0047] The above processing methods effectively solve common quality problems in the multi-source data acquisition process of smart heating systems, such as outliers, missing values, and inconsistent dimensions. Combining the 3σ criterion with isolated forests for outlier removal comprehensively and accurately identifies and removes various types of abnormal data, significantly improving data purity. By combining multi-source data cross-validation and temporal interpolation to complete missing values, the physical correlation and temporal continuity of heating system data are fully utilized to ensure data integrity and rationality. Furthermore, standardized preprocessing eliminates dimensional differences between features, ensuring data consistency. This high-quality, standardized multi-source dataset provides reliable input for subsequent spatiotemporal graph attention network extraction of spatiotemporal joint features, thereby improving the training efficiency, convergence speed, and final load prediction accuracy and stability of the multi-scale prediction model cluster.

[0048] Optionally, a spatiotemporal graph attention network is established to extract spatiotemporal joint features from multidimensional data. This includes extracting spatiotemporal joint features from multi-source datasets based on a constructed spatial topology map of the heating system, thereby obtaining spatiotemporal fusion features. The spatial topology map of the heating system uses heat sources, pipe networks, heat exchange stations, and buildings as nodes, and heat transfer resistance, pipe distance, and flow rate relationships as edge weights to construct a directed topology map. The spatiotemporal graph attention network extracts spatial correlation features using a graph convolutional neural network, extracts temporal dependent features using a temporal convolutional network combined with a self-attention mechanism, and completes spatiotemporal feature fusion through gating units.

[0049] Specifically, a heating system spatial topology diagram is a graph structure used to represent the connections and interactions between various physical entities in a heating system. This structure abstracts key components of the heating system as nodes and physical connections or logical relationships as edges, intuitively reflecting the complex internal structure of the heating system and laying the foundation for subsequent spatial association feature extraction. When constructing the topology diagram, heat sources, pipe networks, heat exchange stations, and buildings are defined as nodes. The heat source is the starting point of heating, responsible for generating heat energy; the pipe network is the heat energy transmission channel, connecting the heat source and the heat exchange station, and the heat exchange station and the building; the heat exchange station is the key link in heat energy distribution and conversion; and the building is the final heat-consuming terminal. Using these entities as nodes comprehensively covers the main components of the heating system, ensuring that the graph structure reflects the complete heating chain.

[0050] Edge weights in the spatial topology of a heating system are used to quantify the connection strength and characteristics between nodes. Heat transfer resistance reflects the obstacles encountered during heat transfer; the greater the resistance, the lower the heat loss or transfer efficiency. Pipe distance directly affects the time delay and losses along the way in heat transfer. Flow relationships reflect the flow characteristics and distribution of the heat medium in the pipe network. Using these physical quantities as edge weights allows the topology to more realistically reflect the internal physical characteristics and energy flow patterns of the heating system. For example, edge weights can be calculated and assigned based on actual measurement data or engineering parameters, and thermal resistance can be calculated using parameters such as pipe diameter, length, and material. Furthermore, heat transfer in a heating system has a clear direction, flowing from the heat source to the user. Therefore, constructing a directed topology can accurately represent unidirectional energy flow paths, and directed edges can clearly indicate the flow of heat from one node to another, such as from the heat source to the heat exchange station, or from the heat exchange station to the building. This directionality is crucial for understanding heat propagation paths, predicting the impact of downstream node loads on upstream nodes, and simulating the system's dynamic response.

[0051] Spatiotemporal graph attention networks are deep learning models that can simultaneously handle the spatial dependencies of graph-structured data and the temporal dependencies of time-series data. Combining the advantages of graph neural networks and temporal neural networks, they dynamically learn the importance of different nodes and time steps through attention mechanisms. This allows them to automatically capture hidden spatiotemporal joint features from complex multi-source heating data, providing more comprehensive information for load forecasting. Graph convolutional neural networks (GCNNs) are the core component for extracting spatial correlation features. By performing convolution operations on the graph structure, they aggregate neighbor information for each node and learn the spatial dependencies between nodes. In heating systems, GCNNs can capture the mutual influences between adjacent heat sources, pipe networks, heat exchange stations, and buildings, such as the impact of load changes at a heat exchange station on upstream pipe networks and downstream buildings. Temporal convolutional networks (TCNNs) are the component for extracting time-dependent features. They process time-series data through one-dimensional convolutional layers and expand the receptive field using dilated convolutions to capture long-term temporal dependencies. In heating systems, TCNNs can learn load patterns over time, such as daily, weekly, and seasonal variations, and capture the impact of historical loads and meteorological data on future loads. Self-attention mechanisms enable models to dynamically focus on information at different positions within a sequence when processing sequential data. In spatiotemporal graph attention networks, this is applied to the time dimension, allowing the model to automatically assign weights to different historical time steps based on the importance of the current prediction task, thus more efficiently capturing key temporal dependencies. Gating units are used to fuse spatial and temporal features, adaptively controlling the degree of spatiotemporal feature fusion based on the characteristics of the input data to generate spatiotemporal fused features containing rich spatiotemporal information.

[0052] The above approach effectively solves the problem of insufficient multi-dimensional data to fully capture complex spatial relationships and dynamic temporal dependencies in load forecasting of smart heating systems. Based on the spatial topology map of the heating system, key entities such as heat sources, pipe networks, heat exchange stations, and buildings are abstracted as nodes. A directed topology graph is constructed with heat transfer resistance, pipeline distance, and flow relationship as edge weights, enabling the model to accurately understand the internal physical connections and energy flow direction of the heating system, providing a realistic structured foundation for subsequent feature extraction. Building upon this, a spatiotemporal graph attention network efficiently extracts spatial relationship features between nodes through a graph convolutional neural network, capturing the mutual influence of various parts of the heating system. Simultaneously, a temporal convolutional network, combined with a self-attention mechanism, deeply mines temporal dependency patterns in historical data, dynamically focusing on key time step information. Finally, a gating unit intelligently fuses spatiotemporal features to generate a spatiotemporal fusion feature containing complete spatiotemporal information. This comprehensive spatiotemporal feature extraction capability significantly enhances the prediction model's ability to perceive dynamic changes in the heating system, improves the accuracy and robustness of load forecasting, provides high-quality input for subsequent multi-scale prediction model clusters, and thus improves the overall performance of smart heating system load forecasting.

[0053] Optionally, a multi-scale prediction model cluster is constructed, and the building heat balance equation and the pipe network thermal inertia model are embedded into the model cluster as physical constraints. Specifically, the method includes embedding the building heat balance equation and the pipe network thermal inertia model as physical constraints into the prediction model, and constructing a multi-scale prediction model cluster consisting of a short-term fine prediction sub-model, a medium-term periodic prediction sub-model, and a long-term trend prediction sub-model.

[0054] The multi-scale prediction model cluster consists of multiple independent prediction models optimized for different prediction time scales. This cluster structure is used to overcome the limitations of a single model in handling prediction tasks at different time scales, ensuring that high-precision predictions can be provided in each time dimension.

[0055] The building heat balance equation is a physical model that describes the heat balance inside a building. It integrates the influence of various factors such as solar radiation, internal heat sources, heat transfer from the building envelope, and ventilation on indoor temperature and heat load. By embedding these factors as physical constraints into the prediction model, it can ensure that the model's output load prediction values ​​conform to basic physical laws, thereby improving the physical interpretability of the prediction and its generalization ability under atypical conditions.

[0056] The pipeline thermal inertia model is a physical model that describes the dynamic characteristics of heat transfer and storage in a heating pipeline network. It considers the flow of the heat medium in the pipeline, heat loss, and the heat capacity effect of the pipe wall and the medium itself. It reflects the time delay and attenuation of heat from the heat source to the user end. Using this model as a physical constraint helps the prediction model to more accurately capture the dynamic response of the pipeline network, especially when the load fluctuates greatly, it can effectively improve the prediction accuracy.

[0057] Physical constraint embedding model clusters integrate the physical knowledge represented by building heat balance equations and pipe network thermal inertia models into data-driven prediction models in a specific form. This can be achieved by adding a penalty term to the model loss function to account for the difference between the predictions of the physical model and the data model, or by designing specific network structures and post-processing mechanisms to force the model output to satisfy physical laws. This embedding method allows the data-driven model to learn from historical data while adhering to fundamental physical principles, thereby improving model robustness and prediction reliability.

[0058] The short-term fine forecast sub-model aims to provide fine-grained load forecasts for the next few hours. This sub-model needs to be highly sensitive to real-time data changes to support timely operation scheduling and optimized control of the heating system.

[0059] The medium-term cycle forecasting sub-model typically forecasts over the next few days to a week, focusing on capturing the cyclical changes in heating load, such as daily and weekly cycles, to provide a basis for fuel procurement, equipment maintenance planning, and medium-term resource allocation.

[0060] The long-term trend forecasting sub-model can predict timeframes ranging from weeks and months to the entire heating season. It is primarily used to identify and predict long-term trends in heating load, seasonal variations, and the impact of macroeconomic factors such as climate change, providing support for strategic planning, capacity assessment, and investment decisions for heating systems.

[0061] The above approach effectively addresses the issues of insufficient physical interpretability and limited predictive capabilities across time scales in load forecasting of complex heating systems using purely data-driven models. The introduction of physical constraints allows the prediction model to better adhere to fundamental thermodynamic principles, enhancing the physical rationality and robustness of the prediction results. This is particularly beneficial when facing atypical operating conditions or sparse data, preventing the generation of predictions that do not conform to physical laws. Furthermore, the multi-scale model cluster, composed of short-term refined prediction sub-models, medium-term periodic prediction sub-models, and long-term trend prediction sub-models, can provide customized solutions for prediction needs at different time scales. This ensures high-precision and highly adaptable load forecasting results across all levels, including real-time operation scheduling, medium-term resource allocation, and long-term planning, comprehensively improving the operational efficiency and energy utilization rate of smart heating systems.

[0062] Optionally, the multi-scale prediction model cluster includes: a lightweight BiLSTM sub-model for 1–24h prediction, an XGBoost-TCN hybrid sub-model for 1–7 day prediction, and a physical mechanism fusion sub-model for long-term trend prediction.

[0063] Among them, the lightweight BiLSTM sub-model for 1-24h forecasting is designed for short-term load forecasting scenarios. BiLSTM is a bidirectional long short-term memory network, belonging to recurrent neural networks. Its bidirectional structure can simultaneously utilize past and future contextual information to predict the current load, effectively capturing the complex time dependencies in the heating load sequence. Lightweighting, while ensuring prediction accuracy, reduces model computational complexity and memory usage by optimizing the network structure, reducing the number of parameters, or employing model quantization, pruning, and other techniques, enabling it to respond quickly and meet the computational efficiency requirements of edge computing or real-time scheduling.

[0064] The XGBoost-TCN hybrid sub-model, used for 1-7 day forecasting, primarily serves medium-term load forecasting. XGBoost is an efficient and robust ensemble learning algorithm, adept at handling structured data and nonlinear relationships, effectively uncovering complex correlations between load and various influencing factors such as meteorology, construction, and user activity. TCN is a sequence model based on convolutional neural networks. Through techniques such as dilated convolution, it can process long-sequence data in parallel and capture dependencies over a longer time span, avoiding the vanishing / exploding gradient problem of traditional recurrent neural networks. The combination of XGBoost and TCN can be achieved through feature fusion or model ensemble, comprehensively leveraging XGBoost's powerful modeling capabilities for multi-source static / quasi-static features and TCN's excellent ability to capture temporal dynamic features, providing more comprehensive and accurate support for medium-term load forecasting.

[0065] A physical mechanism fusion sub-model for long-term trend prediction focuses on capturing long-term trends in heating load. Based on fundamental physical laws such as thermodynamics and fluid mechanics, the physical mechanism model fundamentally describes the energy transfer and conversion processes of the heating system, including building thermal inertia and network heat loss. This type of model possesses good interpretability and generalization ability under conditions of data sparsity and significant changes in the external environment. The fusion of the physical mechanism model and the data-driven model can be achieved by using the prediction results of the physical mechanism model as input features of the data-driven model, or by embedding physical mechanism constraints into the loss function of the data-driven model, guiding the model to learn prediction results that conform to physical laws. This fusion approach allows the model to learn patterns from historical data and combine them with the inherent physical characteristics of the heating system when predicting long-term trends, improving the stability and reliability of predictions and providing a solid basis for long-term planning and energy strategy formulation for heating systems.

[0066] To address the unique needs of heating system load forecasting at different time scales, this application constructs a multi-scale prediction model cluster consisting of a lightweight BiLSTM sub-model, an XGBoost-TCN hybrid sub-model, and a physical mechanism fusion sub-model, achieving an effective balance between prediction accuracy and computational efficiency. The lightweight BiLSTM sub-model provides rapid and accurate predictions of short-term load changes, meeting the needs of real-time scheduling and operation optimization. In medium-term predictions, the XGBoost-TCN hybrid sub-model, combining the advantages of ensemble learning and temporal convolutional networks, effectively captures complex time dependencies and multi-source features, providing a reliable basis for weekly operation plans. In long-term trend predictions, the physical mechanism fusion sub-model embeds the physical laws of the heating system, enhancing the model's generalization ability and interpretability, providing stable long-term trend predictions even with sparse data or drastic changes in the external environment. This multi-scale, specialized model configuration significantly improves the accuracy, robustness, and practicality of heating system load forecasting, enabling the system to more flexibly and efficiently address the operational management challenges at different time scales.

[0067] Optionally, the dynamic fusion weights of each sub-model are adaptively allocated using fuzzy inference or reinforcement learning based on the intensity of meteorological fluctuations, data completeness, and forecast period.

[0068] Specifically, dynamic fusion weights refer to parameters in a multi-scale prediction model cluster where the weighted combination of the outputs of different prediction sub-models is not fixed and can be adjusted according to environmental changes, data quality, and other factors. This dynamic adjustment mechanism allows the system to flexibly select or prioritize sub-models that perform better under current conditions, improving overall prediction accuracy and adaptability. For example, when meteorological conditions change drastically, the weight of sub-models that respond quickly to short-term changes can be increased; when data quality is poor, the weight of sub-models that rely on high-precision data can be reduced, shifting towards sub-models that are less sensitive to missing data and more robust.

[0069] Meteorological fluctuation intensity refers to the degree of drastic change in meteorological parameters affecting heating load, such as temperature, wind speed, and solar radiation, within a certain time window. It can be quantified as the rate of temperature change, the standard deviation of wind speed, or the frequency of extreme weather events per unit time. When meteorological fluctuation intensity is high, the external environment changes rapidly and uncertainty increases, requiring higher accuracy in short-term forecasts. Some sub-models may perform better or worse due to their sensitivity to rapid changes. By assessing meteorological fluctuation intensity, the system can determine the complexity of the current environment and guide the adjustment of fusion weights.

[0070] Data completeness refers to the proportion or quality level of valid, non-missing, and outlier-free observations in the input dataset used for model training and prediction. It can be measured by calculating the percentage of missing data, the number of outliers, or a data quality score. Low data completeness indicates bias or uncertainty in the model's input information, directly affecting the predictive reliability of each sub-model. By evaluating data completeness, the system can identify data quality issues and adjust the fusion weights of each sub-model accordingly, such as reducing the weight of sub-models sensitive to data quality and increasing the weight of robust sub-models.

[0071] The forecast period refers to the time frame corresponding to the load forecasting task, such as short-term, medium-term, and long-term. Different forecast periods have different model focuses and performance requirements. Short-term forecasts emphasize real-time response and granularity, while long-term forecasts emphasize trend capture and stability. Although multi-scale forecasting model clusters already include sub-models for different time periods, the forecast period remains an important consideration during dynamic fusion. For example, when conducting short-term forecasts, even if the intensity of meteorological fluctuations is not large, higher weights can be assigned to the short-term granular forecasting sub-model to ensure the timeliness and accuracy of the forecast.

[0072] Fuzzy inference, a decision-making mechanism based on fuzzy logic, can handle imprecise or uncertain information. In dynamic fusion weight allocation, fuzzy inference can transform fuzzy language descriptions such as high meteorological fluctuation intensity, moderate data completeness, and short forecast periods into specific weight allocation strategies through fuzzy sets and fuzzy rules. For example, defining a fuzzy rule "if meteorological fluctuation intensity is high and data completeness is low, increase the weight of the BiLSTM sub-model and decrease the weight of the XGBoost-TCN sub-model" can better simulate the decision-making process of human experts in complex scenarios and achieve adaptive weight adjustment.

[0073] Reinforcement learning is a machine learning method that learns optimal decision-making strategies through interaction with the environment. In dynamic fusion weight allocation, the weight allocator can be viewed as an agent with the goal of learning a strategy. Given environmental conditions such as the intensity of meteorological fluctuations, data completeness, and prediction time periods, it selects the optimal sub-model fusion weights to maximize prediction accuracy or minimize prediction error. The agent continuously tries different weight combinations and updates its strategy based on reward or penalty signals obtained from actual prediction results, eventually converging to the optimal strategy that adaptively allocates weights.

[0074] The above approach effectively solves the problem of adaptively adjusting the fusion weights of sub-models in a multi-scale prediction model cluster in complex and variable heating environments. By sensing key factors such as the intensity of meteorological fluctuations, data completeness, and prediction time periods in real time, and using fuzzy inference or reinforcement learning to achieve dynamic weight allocation, the system can intelligently adjust the contribution ratio of each sub-model in the final load prediction result based on the current actual operating conditions. For example, when meteorological conditions change drastically, the system automatically increases the weight of short-term fine-grained prediction sub-models to capture rapidly changing load demands; when data quality deteriorates, the weight of data-sensitive models is reduced, relying on more robust models. This adaptive fusion mechanism significantly improves the accuracy, robustness, and environmental adaptability of load prediction, ensuring reliable load prediction values ​​under various complex operating conditions, and providing solid technical support for the optimized operation and energy conservation of smart heating systems.

[0075] Optionally, the initial load forecast results are subjected to uncertainty quantization to obtain load forecast values ​​with confidence intervals, and the forecast model is lightweighted through model pruning and quantization to meet the real-time inference requirements of the edge.

[0076] Specifically, uncertainty quantification is the process of identifying, characterizing, and quantifying the inherent uncertainty of the output results of a prediction model. In load forecasting scenarios, it is used to assess the degree to which the predicted value deviates from the true value, providing decision-makers with key information on the reliability of the prediction. The implementation methods include probabilistic methods, ensemble learning methods, Bayesian methods, etc., which can provide data support for subsequent risk assessment and decision optimization.

[0077] A load forecast with a confidence interval refers to the range within which the actual load value is likely to fall at a given confidence level. For example, a 95% confidence interval means that the actual load value has a 95% probability of falling within that range. Compared to a single load point forecast, a confidence interval provides a forecast range, more comprehensively reflecting the reliability and potential volatility of the forecast results. This helps heating system managers understand forecast risks and develop more robust operating strategies.

[0078] Bayesian inference is employed to quantify the uncertainty of initial load forecast results. Bayesian inference is a statistical inference method whose core idea is to update beliefs about model parameters or forecast results based on new observation data, relying on Bayes' theorem. In load forecasting, Bayesian inference can naturally obtain the posterior probability distribution of the predicted values ​​and directly derive the confidence interval. In practical implementation, a Bayesian neural network can be constructed, combined with techniques such as Monte Carlo Dropout, to model the uncertainty of model parameters and thus quantify the uncertainty of the forecast results. This method can effectively capture the uncertainty of model parameters, data, and structure, providing a more comprehensive uncertainty assessment.

[0079] Simultaneously, model pruning and quantization are used to achieve lightweight predictive models. Model pruning is a technique that removes redundant or unimportant connections or neurons from a neural network, reducing model size and computational complexity. For example, it involves identifying and removing irrelevant parts based on their weights or their impact on model performance. Model quantization is the process of reducing the numerical precision of model parameters from high to low. These two techniques can be used individually or in combination to significantly reduce model storage space, memory footprint, and inference computational resource consumption. Lightweight predictive models, through optimization techniques such as model pruning, quantization, knowledge distillation, or compact network structures, enable models to have smaller size, fewer parameters, and lower computational resource consumption while maintaining or slightly sacrificing performance. This makes them easier to deploy on resource-constrained hardware platforms such as edge computing devices.

[0080] Meeting the real-time inference requirements at the edge means that predictive models can quickly process new data and generate prediction results with low latency and high computational efficiency on edge devices close to the data source. In smart heating systems, this means that predictive models can run directly on local devices such as heat exchange station controllers and building management units, without having to upload all data to the cloud for processing. This significantly shortens decision response time, improves system autonomy and reliability, and reduces communication bandwidth and cloud computing costs.

[0081] By quantifying the uncertainty of initial load forecasts, the system can output load forecasts with confidence intervals. This allows heating managers to obtain both the central value of the forecast and an understanding of the reliability and potential fluctuation range of the forecast results. For example, when the forecast load is high and the confidence interval is narrow, the forecast result is highly reliable, and the corresponding heating strategy can be adopted boldly. When the confidence interval is wide, it indicates greater uncertainty, requiring more cautious decision-making or the introduction of additional risk mitigation measures. This uncertainty quantification method significantly improves the scientific nature and robustness of decision-making, avoiding insufficient or excessive heating caused by a single forecast deviation. At the same time, by using model pruning and quantization techniques to achieve lightweight forecast models, the storage and computational consumption of models is significantly reduced. Complex forecast models that originally required powerful server support can be efficiently deployed to edge controllers or embedded devices, meeting the real-time inference and rapid response requirements of smart heating systems, reducing system deployment and operating costs, and improving overall response speed and autonomous decision-making capabilities.

[0082] Optionally, the specific implementation of uncertainty quantification is to use a Bayesian neural network and Monte Carlo dropout to output load forecasts and risk probabilities with 95% confidence intervals.

[0083] Specifically, a Bayesian neural network (BNN) is a model that applies the Bayesian inference principle to neural networks. Unlike traditional neural networks that estimate output points, Bayesian neural networks model the probability distribution of network weight parameters, outputting a probability distribution of the prediction result rather than a single, fixed value. During training, instead of learning fixed weight values, it learns the posterior distribution of the weights, which can be achieved through variational inference, Markov chain Monte Carlo methods, or approximation methods. The core principle is to model the uncertainty of the model parameters to reflect the uncertainty of the prediction result.

[0084] Monte Carlo dropout is an approximate Bayesian inference technique in neural networks. It leverages the property of maintaining dropout layer activation during the testing phase, sampling different network structures through multiple forward propagations to obtain the distribution of prediction results. During the inference phase, dropout activation is maintained, and the same input is propagated forward multiple times. Each time, the presence of dropout produces slightly different outputs. These outputs are considered as samples of the model's posterior distribution, and the prediction mean and uncertainty are statistically obtained from these samples.

[0085] By combining a Bayesian neural network with Monte Carlo dropout, load forecasts and risk probabilities with 95% confidence intervals can be output. A 95% confidence interval indicates that, in multiple repeated forecasts, the actual load value has a 95% probability of falling within this interval. Using the obtained forecast distribution, intervals at any confidence level can be calculated; for example, the 2.5% and 97.5% quantiles of the forecast distribution can be used as the upper and lower limits for the 95% confidence interval. The risk probability can be defined based on the probability of the forecast value falling outside a preset threshold, such as the probability of the load forecast value exceeding the safe upper limit or falling below the safe lower limit. This is of great significance for the safe operation and scheduling of the heating system.

[0086] By combining Bayesian neural networks with Monte Carlo dropout, more refined and robust uncertainty quantification of load forecasting results can be achieved. Through model parameter uncertainty modeling and relying on Monte Carlo sampling to obtain the probability distribution of forecast results, a statistically significant 95% confidence interval is output, more accurately reflecting the range of load forecast fluctuations. Furthermore, explicitly providing risk probabilities allows the system to quantify the likelihood of forecast results deviating from expectations or safety thresholds, providing a more operational basis for heating system operation scheduling and risk management, avoiding potential risks from single forecast values, and improving the reliability and decision support capabilities of forecast results.

[0087] Optionally, the online adaptive update adopts an incremental learning approach, which automatically triggers parameter fine-tuning and weight recalibration when the prediction error exceeds a set threshold.

[0088] Online adaptive model updates refer to models that, after deployment and operation, continuously adjust and optimize their internal parameters and structure based on new data and feedback generated during actual operation, adapting to environmental changes. This differs from traditional offline training models that remain fixed after completion. Online adaptive updates ensure that the model always accurately reflects the current system state. For example, updates can be performed periodically at preset time points, or triggered when specific events such as system failures or large-scale changes in user behavior are detected.

[0089] Incremental learning is a machine learning paradigm that allows models to update their knowledge and parameters gradually by processing new data samples without retraining the entire dataset. Compared to batch learning, which requires retraining the entire historical dataset, incremental learning significantly saves computational resources and time, and is suitable for scenarios where data streams are continuously generated and real-time responses are required. In practical implementation, it can employ online optimization algorithms based on stochastic gradient descent or its variants, using only a small amount of new data for gradient calculation and parameter updates each time; or utilize knowledge distillation techniques to integrate knowledge learned from new data into the existing model; or use transfer learning to transfer the feature extraction capabilities of pre-trained models to new tasks or new data distributions.

[0090] Automatic triggering when prediction error exceeds a set threshold is an intelligent update strategy designed to improve the efficiency and timeliness of model updates. Prediction error can be the difference between predicted and observed values, such as absolute error, relative error, mean squared error, and mean absolute error. The set threshold is determined based on the system's requirements for prediction accuracy and the acceptable error range. When the prediction error is detected to continuously or cumulatively exceed the threshold, the system automatically initiates the model parameter fine-tuning and weight recalibration process to avoid unnecessary frequent updates and ensure timely intervention when model performance deteriorates.

[0091] Parameter fine-tuning and weight recalibration are specific operational steps in incremental learning. Parameter fine-tuning involves making small adjustments to the learnable parameters within the prediction model to reduce the current prediction error. Weight recalibration focuses more on adjusting the importance weights of different components of the model or different input features to better reflect the current data distribution or system dynamics. These adjustments are typically performed on new data using optimization algorithms to minimize the new loss function.

[0092] The predictive model can achieve continuous online adaptive updates through this method. When the prediction error exceeds a preset threshold, the system automatically triggers parameter fine-tuning and weight recalibration, without waiting for manual intervention or periodic updates. This incremental learning-based update method allows the model to efficiently absorb new operational feedback data, adjust internal parameters and weights in a timely manner, and effectively respond to dynamic changes in the heating system's operating environment, user behavior patterns, and external meteorological conditions. This method not only avoids prediction bias caused by model performance degradation, ensuring that load prediction values ​​maintain high accuracy and robustness over the long term, but also optimizes the utilization of computing resources through an on-demand triggering mechanism, providing reliable and continuously accurate decision-making basis for real-time optimization control of the smart heating system, thereby improving heating efficiency and user comfort.

[0093] In a specific example, a smart heating demonstration area includes a heat source, multiple heating pipelines, multiple heat exchange stations, and a large number of user buildings. In order to achieve precise regional heating scheduling and energy-saving operation, a smart heating system load prediction method based on multi-source data fusion is adopted.

[0094] First, a multi-dimensional data acquisition system deployed within the demonstration zone was used to acquire multi-dimensional data containing various influencing factors. This data encompassed meteorological data, building data, user data, pipeline network data, heat source data, time-series data, and external data. This comprehensive acquisition method overcomes the limitations of existing technologies, such as single data dimensions and insufficient information mining, laying the foundation for subsequent accurate forecasting.

[0095] Next, the collected multidimensional data underwent quality enhancement and standardization transformation, sequentially performing outlier removal, missing value completion, and standardization preprocessing to obtain a high-quality multi-source dataset. The outlier removal process combined the 3σ criterion and the isolated forest algorithm to effectively identify and handle abnormal data caused by sensor malfunctions or communication interruptions, avoiding the accuracy loss caused by simple deletion or mean filling in existing technologies. Missing values ​​were completed using a combination of multi-source data cross-validation and temporal interpolation to ensure data integrity and continuity. Subsequently, data standardization was performed to eliminate the influence of data with different dimensions, preparing the data for model input.

[0096] Then, a spatiotemporal graph attention network is constructed to extract spatiotemporal joint features from multidimensional data. Based on the actual physical connections in the demonstration area, a spatial topology map of the heating system is constructed, with heat sources, pipe networks, heat exchange stations, and buildings as nodes, and heat transfer resistance, pipeline distance, and flow rate as edge weights, to construct a directed topology map that accurately reflects the internal physical connections of the heating system. The multi-source dataset, after quality enhancement and standardization, is input into the spatiotemporal graph attention network. A graph convolutional neural network extracts the spatial connection features of each node in the heating system, and a temporal convolutional network combined with a self-attention mechanism captures the long and short temporal dependence features of the load data. Through deep fusion of spatiotemporal features using gating units, spatiotemporal fusion features are obtained, solving the problem of fragmented spatiotemporal features in existing technologies and their inability to accurately depict the dynamic changes in load.

[0097] Building upon this foundation, a multi-scale prediction model cluster is constructed, embedding the building heat balance equation and the pipe network thermal inertia model as physical constraints. The model cluster includes short-term fine-grained prediction sub-models, medium-term periodic prediction sub-models, and long-term trend prediction sub-models. Specifically, the 1-24 hour prediction employs a lightweight BiLSTM sub-model to capture short-term load fluctuations; the 1-7 day prediction uses an XGBoost-TCN hybrid sub-model to handle medium-term load periodic changes; and the long-term trend prediction uses a physical mechanism fusion sub-model, incorporating the physical laws of the heating system to ensure long-term prediction stability. Physical constraints are embedded during model training, ensuring that the model, while data-driven, adheres to fundamental thermodynamic principles, improving robustness and physical interpretability under extreme scenarios, and compensating for the poor physical interpretability and susceptibility to failure inherent in purely data-driven models. Meanwhile, based on the scene-aware strategy, according to the intensity of meteorological fluctuations, data completeness, and forecast period, fuzzy reasoning or reinforcement learning is used to assign dynamic fusion weights to each sub-model, carry out multi-scale load forecasting on spatiotemporal fusion features, and output the initial load forecast results. This dynamic weight allocation mechanism enables the model cluster to adaptively adjust to the actual operating conditions, which is superior to the existing single model or static weight fusion architecture.

[0098] Subsequently, the uncertainty of the initial load forecast results is quantified to obtain load forecast values ​​with confidence intervals. Quantization is performed using Bayesian inference, implemented through a Bayesian neural network and Monte Carlo dropout, outputting load forecast values ​​with 95% confidence intervals and risk probabilities. This allows heating dispatch decision-makers to obtain not only the forecast value but also an understanding of the reliability range and potential risks of the forecast results, providing support for risk assessment and solving the problem of existing technologies that only output a single forecast value and cannot quantify uncertainty. Simultaneously, model pruning and quantization techniques are used to achieve a lightweight forecast model, meeting the real-time inference requirements of edge devices and improving the adaptability to engineering applications.

[0099] Finally, the load forecast values ​​are aggregated and corrected level by level according to the hierarchical structure of heat source, pipeline network, heat exchange station, and user. For example, the load of each building is predicted first, then aggregated to the heat exchange station level, then to the pipeline network area, and finally summarized to the total load of the heat source. During the aggregation process, the forecast values ​​are corrected based on actual operation feedback data to ensure that the load forecasts at each level are accurate and consistent. This hierarchical forecasting system achieves a close integration of load forecasting and scheduling, promotes full-link collaborative optimization, and overcomes the problem of disconnect between forecasting and scheduling in existing technologies. At the same time, the forecasting model is subjected to online incremental learning and adaptive parameter updates based on actual operation feedback data. When the forecast error exceeds the set threshold, the system automatically triggers parameter fine-tuning and weight recalibration, so that the model can self-adjust with seasonal changes, building aging, changes in user habits, and other operating conditions, avoiding the problem of fixed parameters and continuous decay of forecast accuracy after training in existing technology models, and ensuring long-term stable operation of the model.

[0100] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0101] The methods and products disclosed in the embodiments herein (including but not limited to devices and equipment) can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to implement this embodiment according to actual needs. In addition, the functional units in the embodiments of this disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0102] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A method for load forecasting of a smart heating system based on multi-source data fusion, characterized in that, include: A multidimensional data acquisition system is used to obtain multidimensional data containing multiple influencing factors. Enhance the quality and standardize multidimensional data; Establish a spatiotemporal graph attention network to extract spatiotemporal joint features from multidimensional data; A multi-scale prediction model cluster is constructed, and the building heat balance equation and the pipe network thermal inertia model are embedded into the model cluster as physical constraints. Dynamic fusion weights are assigned to each model based on the scene perception strategy, and multi-scale load prediction is performed on the spatiotemporal fusion features to output the initial load prediction results. Uncertainty quantification is performed on the initial load forecast results to obtain load forecast values ​​with confidence intervals; The load forecast values ​​are aggregated and corrected step by step according to the hierarchical structure of heat source, pipeline network, heat exchange station and user. At the same time, the forecast model is subjected to online incremental learning and adaptive parameter updates based on actual operation feedback data.

2. The method according to claim 1, characterized in that, Multidimensional data containing multiple influencing factors includes: Meteorological data: real-time / forecast temperature and humidity, wind speed, wind direction, solar radiation, precipitation, air pressure, extreme weather indicators; Building data: building type, area, heat transfer coefficient of building envelope, insulation grade, window-to-wall ratio, orientation, and year of construction; User data: indoor temperature settings, daily routines, frequency of window opening, heating habits, occupancy rate, and user complaint data; Pipeline data: supply and return water temperature / pressure / flow rate, pipeline length / diameter / resistance, thermal inertia coefficient, leakage rate, heat exchange station parameters; Heat source data: output, efficiency, energy consumption, start / stop status, and heating medium parameters of boilers / heat pumps / peak-shaving equipment; Time series data: historical load, date type, season, time period label; External data: social activities, weather forecast errors, equipment failure records, outdoor thermal imaging / remote sensing images.

3. The method according to claim 1, characterized in that, Multidimensional data undergoes quality enhancement and standardization transformation, including: The multi-source data is preprocessed sequentially by outlier removal, missing value completion, and standardization to obtain a high-quality multi-source dataset. Outlier removal adopts a combination of the 3σ criterion and the isolation forest method, while missing value completion is achieved by a combination of multi-source data cross-validation and temporal interpolation.

4. The method according to claim 1, characterized in that, Establish a spatiotemporal graph attention network to extract spatiotemporal joint features from multidimensional data, including: Based on the constructed spatial topology map of the heating system, spatiotemporal joint features are extracted from multi-source datasets through a spatiotemporal graph attention network to obtain spatiotemporal fusion features; The spatial topology of the heating system is constructed with heat source, pipe network, heat exchange station and building as nodes, and heat transmission resistance, pipe distance and flow relationship as edge weights. The spatiotemporal graph attention network extracts spatial correlation features using a graph convolutional neural network, extracts temporal dependent features using a temporal convolutional network combined with a self-attention mechanism, and completes spatiotemporal feature fusion through a gating unit.

5. The method according to claim 1, characterized in that, A multi-scale prediction model cluster is constructed, and the building heat balance equation and the pipe network thermal inertia model are embedded into the model cluster as physical constraints, including: The building heat balance equation and the pipe network thermal inertia model are embedded as physical constraints into the prediction model, and a multi-scale prediction model cluster is constructed, consisting of a short-term fine prediction sub-model, a medium-term cycle prediction sub-model, and a long-term trend prediction sub-model.

6. The method according to claim 5, characterized in that, The multi-scale prediction model cluster includes: a lightweight BiLSTM sub-model for 1–24h prediction, an XGBoost-TCN hybrid sub-model for 1–7 day prediction, and a physical mechanism fusion sub-model for long-term trend prediction.

7. The method according to claim 6, characterized in that, The dynamic fusion weights of each sub-model are adaptively allocated using fuzzy inference or reinforcement learning based on the intensity of meteorological fluctuations, data completeness, and forecast period.

8. The method according to any one of claims 1 to 7, characterized in that, Uncertainty quantification is performed on the initial load forecast results to obtain load forecast values ​​with confidence intervals, including: Bayesian inference is used to quantify the uncertainty of the initial load prediction results to obtain load prediction values ​​with confidence intervals. The prediction model is lightweighted through model pruning and quantization to meet the real-time inference requirements of the edge.

9. The method according to claim 8, characterized in that, Uncertainty quantification is achieved using a Bayesian neural network and Monte Carlo dropout, outputting load forecasts and risk probabilities with 95% confidence intervals.

10. The method according to any one of claims 1 to 7, characterized in that, The model's online adaptive update uses an incremental learning approach, automatically triggering parameter fine-tuning and weight recalibration when the prediction error exceeds a set threshold.

Images