A method and system for load forecasting of a heating system

Abstract

The application relates to the technical field of heat load prediction, and discloses a heat supply system load prediction method and system, which comprises the following steps: collecting multi-source sensing data in real time, and collecting outdoor meteorological parameters and building structure information. Based on the building distribution, pipe network structure and user load characteristics of a heat supply area, a multi-level heat supply load prediction model is constructed. A heat topology graph model of the heat supply area is constructed based on a graph neural network. The parameters of a building-level edge prediction model, a heat exchange station-level aggregation prediction model and the heat topology graph model are periodically collected, global aggregation optimization is carried out, and the optimized edge node model is updated. According to the edge prediction result obtained by the optimized model, the heat source output power, the main pump rotating speed and the area valve opening degree are jointly controlled according to the edge prediction result and the heat source level scheduling prediction model. The application improves the description ability of the system to dynamic load changes and spatial heat conduction paths, ensures the generalization ability of model updating and the real-time responsiveness of edge deployment.

Description

Technical Field

[0001] This invention relates to the field of heat load prediction technology, and more specifically, to a method and system for predicting the load of a heating system. Background Technology

[0002] With the continuous development of smart energy and IoT technologies, heating systems are gradually evolving towards digitalization, networking, and intelligence. Especially in district-based centralized heating scenarios, accurate heat load forecasting and optimized control of system operating parameters have become key means to improve energy efficiency, reduce operating costs, and ensure heating quality. Currently, some advanced cities have deployed sensing terminals at heat users, heat exchange stations, and heat sources to collect operating parameters, including temperature, flow rate, and pressure, and use certain model algorithms for heating load analysis. However, existing load forecasting methods generally suffer from the following problems:

[0003] On the one hand, traditional methods often employ centralized modeling, lacking differentiated modeling and prediction of load characteristics at different levels, such as building level, heat exchange station level, and heat source level, making it difficult to achieve refined hierarchical control. On the other hand, existing models generally do not consider the spatial correlation characteristics of thermal influences between building units in the heating system, failing to effectively characterize the heat conduction paths and mutual coupling relationships between buildings, resulting in insufficient prediction accuracy. Furthermore, current systems often rely on centralized data processing architectures, which suffer from bandwidth bottlenecks, poor real-time performance, and privacy risks, and model updates lack federated learning mechanisms, making it difficult to achieve multi-node collaborative optimization.

[0004] Therefore, it is necessary to design a method and system for predicting the load of a heating system to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a method and system for predicting the load of a heating system, which aims to solve the problems of insufficient prediction accuracy caused by the difficulty in achieving refined hierarchical control and the failure to consider the spatial correlation characteristics of thermal influence between building units in the heating system.

[0006] In one aspect, the present invention proposes a method for predicting the load of a heating system, comprising:

[0007] Real-time acquisition of multi-source sensing data, including indoor temperature, user heat flow rate, valve opening, pipeline pressure, water flow velocity, temperature information of heat exchange inlet and return water, and operating parameters of heat source outlet and return water, as well as acquisition of outdoor meteorological parameters and building structure information;

[0008] Based on the building distribution, pipeline structure and user load characteristics of the heating area, a multi-level heating load prediction model is constructed. The multi-level heating load prediction model includes a building-level edge prediction model, a heat exchange station-level aggregation prediction model and a heat source-level scheduling prediction model.

[0009] A thermal topology graph model of the heating area is constructed based on a graph neural network, with building units as graph nodes, the connection relationship of the heating network as graph edges, and the multi-source sensing data mapped as attribute vectors of nodes and edges.

[0010] The parameters of the building-level edge prediction model, the heat exchange station-level aggregate prediction model, and the thermal topology model are periodically collected, and global aggregation optimization is performed. The models of each edge node are then updated and optimized.

[0011] The edge prediction results are obtained based on the optimized model, and the heat source output power, main pump speed and zone valve opening are jointly controlled based on the edge prediction results and the heat source level scheduling prediction model.

[0012] Furthermore, when collecting multi-source sensing data in real time, this includes:

[0013] On the heat user side, indoor temperature, user heat flow rate and valve opening are collected by temperature sensors, calorimeters and electric valve status monitors, respectively.

[0014] On the secondary pipeline side, pressure sensors and flow sensors are used to collect pipeline pressure values ​​and water flow velocity values;

[0015] Temperature information of the heat exchange inlet and outlet water is collected by supply water temperature sensors and return water temperature sensors at the heat exchange station side.

[0016] On the heat source side, the operating parameters of the heat source outlet water and return water are obtained through the main supply water temperature sensor and the main return water temperature sensor;

[0017] Each IoT terminal device collects corresponding physical quantity signals according to a preset data sampling period, and encodes and encapsulates the collection results locally. The results are then sent to the edge node via the MQTT communication protocol and Ethernet. The edge node performs time alignment, anomaly removal, and structural standardization on the collected data.

[0018] Furthermore, based on the building distribution, pipeline structure, and user load characteristics of the heating area, a multi-level heating load prediction model is constructed. The construction process of the building-level edge prediction model includes:

[0019] At each edge node corresponding to the target building, an input vector containing historical time series is constructed based on the indoor temperature, user heat flow, electric valve opening, building structure information and outdoor meteorological parameters within the target building.

[0020] A prediction model is established based on a neural network structure, wherein the neural network structure is a long short-term memory network with time recursive units, and a prediction model for predicting the change of building heat load within a preset time window is obtained by training according to the input vector.

[0021] The trained model is deployed on edge nodes to perform inference based on real-time sensing data and output building-level predicted heat load values.

[0022] Furthermore, the construction process of the heat exchange station-level aggregate prediction model includes:

[0023] At the edge node corresponding to each heat exchange station, the building-level predicted heat load value output by several building-level edge prediction models under the heat exchange station is received, and an aggregated input vector is constructed by combining the building structure information, number of users, historical water supply temperature, historical return water temperature and current outdoor meteorological parameters.

[0024] The network structure of the heat exchange station-level aggregated prediction model is based on a multi-layer feedforward neural network. The model is trained based on the aggregated input vector to predict the total heat load of the heat exchange station within a preset time window in the future.

[0025] The trained aggregated prediction model is deployed in the edge nodes of the heat exchange station, and the predicted heat load value of the heat exchange station is output based on the real-time prediction results of each building-level edge prediction model.

[0026] Furthermore, the construction process of the heat source-level scheduling prediction model includes:

[0027] On the central server, the predicted heat load values ​​of each heat exchange station are uploaded by edge nodes of each heat exchange station, and the supply water temperature, return water temperature, main pump speed and outdoor meteorological parameters are collected to construct the scheduling input vector.

[0028] A heat source-level scheduling prediction model is constructed based on a support vector regression model. The heat source-level scheduling prediction model is trained according to the scheduling input vector and is used to predict the output power change trend of the heat source within a preset time window in the future.

[0029] The heat source-level scheduling prediction model is deployed in the central server, and the inference results of the heat source-level scheduling prediction model are used as the basis for heat source output regulation, main pump frequency conversion control, and regional valve opening strategies.

[0030] Furthermore, when constructing a thermal topology graph model of the heating area based on a graph neural network, using building units as graph nodes and the connection relationships of the heating network as graph edges, and mapping the multi-source sensing data into attribute vectors of nodes and edges, the process includes:

[0031] Using each building unit within the heating area as a graph node and building pairs that are physically connected through the heating network as graph edges, an undirected graph structure containing a set of nodes and a set of edges is constructed.

[0032] The indoor temperature, user heat flow, valve opening and building structure information of each building unit are used as node attribute vectors, and the water flow velocity, pipeline pressure, heat exchange supply water temperature and return water temperature are used as edge attribute vectors.

[0033] A graph neural network model is constructed based on graph convolutional neural networks. The graph structure is embedded and learned, and a structured representation of each building node is generated.

[0034] Furthermore, the parameters of the building-level edge prediction model, the heat exchange station-level aggregated prediction model, and the thermal topology model are periodically collected for global aggregation optimization. When updating and optimizing the models of each edge node, this includes:

[0035] A federated learning module is deployed on the central server. The federated learning module sends aggregation requests to each edge node at a preset time period to collect local training parameters of the building-level edge prediction model, the heat exchange station-level aggregation prediction model, and the graph neural network model.

[0036] The federated learning module performs a weighted average strategy on the collected model parameters of each edge node to aggregate them globally, forming unified global model parameters.

[0037] The global model parameters are updated and synchronously distributed to each corresponding edge node to replace the local prediction model parameters.

[0038] Furthermore, when the federated learning module aggregates and optimizes the local parameters uploaded by each edge node, it adopts differentiated optimization strategies for different model types, including:

[0039] For the building-level edge prediction model and the heat exchange station-level aggregate prediction model, a federated average algorithm based on model weight sparsity constraints is used to regularize the parameter updates.

[0040] For the node embedding parameters of the graph neural network in the hot topology graph model, the weight distribution of highly associated edges is preserved during the aggregation process;

[0041] After aggregation is completed, the consistency loss function is used to check the degree of fit of the aggregation model to the edge prediction results. If the preset fitting threshold is met, the model synchronization process is carried out.

[0042] Furthermore, based on the edge prediction results obtained from the optimized model, when jointly controlling the heat source output power, main pump speed, and zone valve opening based on the edge prediction results and the heat source-level scheduling prediction model, the following is included:

[0043] Each building-level edge node and heat exchange station-level edge node outputs the predicted heat load values ​​of the building-level edge prediction model and the heat exchange station-level aggregate prediction model in real time within the prediction time window, respectively.

[0044] The predicted heat load value at the heat exchange station level and the building thermal correlation information output by the thermal topology model are jointly input into the heat source level scheduling prediction model to obtain the predicted heat source output power value within the target time window.

[0045] A global load distribution state map is constructed based on the predicted output power of the heat source, the predicted heat load of each heat exchange station, and the structured representation of building nodes generated by the graph neural network model as input features.

[0046] Based on the global load distribution state diagram, a rolling optimization objective function is established based on the model predictive control algorithm. The rolling optimization objective function is used to minimize the total energy consumption and pressure fluctuation of the system while meeting the heating demand of each region.

[0047] Under the constraints of the rolling optimization objective function, the main pump speed regulation signal and the regional valve opening regulation signal are obtained in real time, and multi-objective regulation commands are generated by combining the spatial gradient of heat demand of each node in the global load distribution state diagram.

[0048] The multi-target adjustment commands are encoded in an instruction set structure and sent to the main pump frequency converter and the control units of the execution valves in each area through the cloud-edge collaborative control interface.

[0049] Compared with existing technologies, the beneficial effects of this invention are as follows: By introducing a collaborative mechanism of "IoT sensing + hierarchical modeling + graph neural network + federated learning optimization," high-precision, hierarchical, and spatially coupled sensing and predictive control of the heating system's heat load is achieved. This application deploys multiple types of sensor devices on the heat user side, secondary pipe network side, heat exchange station side, and heat source side to comprehensively collect key operating parameters in the heat transmission and distribution process; by constructing multi-level prediction models at the building level, heat exchange station level, and heat source level, and combining graph neural networks to mine the thermal topological relationships between buildings, the system's ability to characterize dynamic load changes and spatial heat conduction paths is improved; the federated learning module further breaks down data silos, enabling collaborative optimization of each prediction model while protecting data privacy, ensuring the generalization ability of model updates and the real-time responsiveness of edge deployment; combined with the heat source-level scheduling model, it achieves linked control of heat source output, main pump speed, and valve opening, improving the energy efficiency and regulation flexibility of the heating system while ensuring prediction accuracy, effectively solving the problems of coarse prediction granularity, lack of spatial modeling, and bottlenecks in centralized computing architecture in current technologies.

[0050] On the other hand, this application also provides a heating system load forecasting system for applying the above-mentioned heating system load forecasting method, comprising:

[0051] The acquisition unit is configured to acquire multi-source sensing data in real time. The multi-source sensing data includes indoor temperature, user heat flow rate, valve opening, pipeline pressure, water flow velocity, temperature information of heat exchange inlet and return water, and operating parameters of heat source outlet and return water. It also acquires outdoor meteorological parameters and building structure information.

[0052] The first processing unit is configured to construct a multi-level heating load prediction model based on the building distribution, pipeline structure and user load characteristics of the heating area. The multi-level heating load prediction model includes a building-level edge prediction model, a heat exchange station-level aggregation prediction model and a heat source-level scheduling prediction model.

[0053] The second processing unit is configured to construct a thermal topology graph model of the heating area based on a graph neural network, using building units as graph nodes, the connection relationship of the heating network as graph edges, and mapping the multi-source sensing data into attribute vectors of nodes and edges.

[0054] The third processing unit is configured to periodically collect parameters from the building-level edge prediction model, the heat exchange station-level aggregate prediction model, and the thermal topology model, perform global aggregate optimization, and update and optimize the models of each edge node.

[0055] The adjustment unit is configured to obtain edge prediction results based on the optimized model, and to jointly control the heat source output power, main pump speed and zone valve opening based on the edge prediction results and the heat source-level scheduling prediction model.

[0056] It is understandable that the above-mentioned heating system load forecasting methods and systems have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0057] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0058] Figure 1 A flowchart of a heating system load prediction method provided in an embodiment of the present invention;

[0059] Figure 2 This is a functional block diagram of a heating system load prediction system provided in an embodiment of the present invention. Detailed Implementation

[0060] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0061] In traditional heating system load forecasting methods, centralized modeling architectures struggle to effectively differentiate the dynamic characteristics of loads at the building, heat exchange station, and heat source levels, resulting in forecasting models that cannot adapt to the operational characteristics of heating units at different levels. The heat conduction paths between building units and the network topology are not incorporated into the model input features, leading to spatial thermodynamic coupling effects and prediction biases. Centralized data processing causes frequent transmission of high-dimensional sensing data between edge nodes and the cloud, creating communication bandwidth pressure and privacy risks, while model parameter updates lack a collaborative optimization mechanism.

[0062] For example, in a centralized heating system comprising 32 heat exchange stations and 478 buildings, building-level heat load data is uploaded to the cloud server via a 4G network at a frequency of once per minute. Due to the lack of localized modeling capabilities at the edge, temperature fluctuations and valve status changes within buildings must be forwarded to the central platform through three layers of network nodes, resulting in an average transmission delay of 8 seconds. The thermal network connections between building units are stored in the form of a static topology table, without dynamic correlation mapping with real-time water flow velocity and network pressure. When a building experiences abnormal return water temperature due to valve malfunction, the central model cannot quickly identify the heat conduction impact of adjacent buildings, causing the predicted load at the heat exchange station level to deviate from actual demand by 12%. The central server performs global model retraining every 6 hours, during which each edge node maintains its original parameters, leading to a mismatch between control commands and real-time operating conditions.

[0063] If the above issues are not addressed, the differences in load characteristics across multiple levels will lead to a mismatch between heat source output power and local demand, resulting in hydraulic imbalance in the heating network and energy waste. The lack of inter-building thermal correlation characteristics causes abnormal operating conditions to spread further, increasing the risk of network pressure fluctuations. High-frequency data uploads exacerbate network congestion, and idle computing resources at edge nodes reduce real-time response. Asynchronous model updates cause control strategies to lag behind load changes, exacerbating heating temperature deviations and increasing user complaint rates.

[0064] For this, please refer to Figure 1 As shown, this application proposes a method for predicting the load of a heating system, including:

[0065] S100: Real-time acquisition of multi-source sensing data, including indoor temperature, user heat flow rate, valve opening, pipeline pressure, water flow velocity, temperature information of heat exchange inlet and outlet water, and operating parameters of heat source outlet and outlet water, as well as outdoor meteorological parameters and building structure information.

[0066] S200: Based on the building distribution, pipeline structure and user load characteristics of the heating area, a multi-level heating load prediction model is constructed. The multi-level heating load prediction model includes a building-level edge prediction model, a heat exchange station-level aggregation prediction model and a heat source-level scheduling prediction model.

[0067] S300: Based on graph neural networks, a thermal topology graph model of the heating area is constructed, with building units as graph nodes, the connection relationship of the heating network as graph edges, and multi-source sensing data mapped into attribute vectors of nodes and edges.

[0068] S400: Periodically collects parameters from the building-level edge prediction model, the heat exchange station-level aggregate prediction model, and the thermal topology model, performs global aggregation optimization, and updates and optimizes the models of each edge node.

[0069] S500: Obtain edge prediction results based on the optimized model, and jointly control the heat source output power, main pump speed and zone valve opening based on the edge prediction results and the heat source level scheduling prediction model.

[0070] Specifically, multi-source sensing data includes indoor temperature, user heat flow rate, valve opening, pipeline pressure, water flow velocity, temperature information of heat exchange inlet and return water, and operating parameters of heat source outlet and return water. This is achieved using temperature sensors, pressure sensors, flow sensors, calorimeters, and electric valve status monitors to comprehensively reflect the real-time operating status and heat transfer process of the heating system. Outdoor meteorological parameters and building structure information, including air temperature, humidity, wind speed, building materials, floor height, and wall insulation coefficient, are obtained from meteorological stations and building information model databases to correct for the impact of environmental factors on heat load. The multi-level heating load prediction model includes a building-level edge prediction model, a heat exchange station-level aggregation prediction model, and a heat source-level scheduling prediction model. These are constructed using long short-term memory networks, multi-layer feedforward neural networks, and support vector regression models, respectively, to achieve hierarchical load prediction from micro to macro levels. The thermal topology graph model uses a graph neural network to treat building units as graph nodes and heating network connections as graph edges. Specifically, it employs a graph convolutional neural network to embed node and edge attribute vectors, capturing the heat conduction paths and spatial relationships between buildings. Global aggregation optimization periodically collects edge node model parameters through a federated learning module and uses a weighted averaging strategy for parameter fusion. A cloud-edge collaborative architecture ensures synchronized model updates, addressing the collaborative optimization problem of distributed nodes. Joint control of heat source output power, main pump speed, and regional valve openings establishes a rolling optimization objective function using a model predictive control algorithm. This is combined with a global load distribution state diagram to generate multi-objective adjustment commands, achieving dynamic adjustment and energy consumption optimization.

[0071] This application constructs a dual architecture of a multi-level prediction model and a thermal topology model. It achieves distributed parameter aggregation through a federated learning mechanism, and combines spatial correlation characteristics with time series prediction to form a cloud-edge collaborative closed-loop optimization control system, which effectively improves the accuracy of heating load prediction and the efficiency of regulation.

[0072] The working process and principle of this application are as follows: real-time acquisition of multi-source sensing data, including indoor temperature, user heat flow rate, valve opening degree, pipeline pressure, water flow velocity, temperature information of heat exchange inlet and return water, and operating parameters of heat source outlet and return water, while simultaneously acquiring outdoor meteorological parameters and building structure information. This data provides input for subsequent prediction models.

[0073] Then, based on the building distribution, pipeline structure, and user load characteristics of the heating area, a multi-level heating load prediction model is constructed. This model includes a building-level edge prediction model, a heat exchange station-level aggregated prediction model, and a heat source-level scheduling prediction model. The building-level model runs on edge nodes to predict the heat load of a single building. The heat exchange station-level model aggregates the prediction results of multiple buildings. The heat source-level model is used for overall scheduling.

[0074] Next, a thermal topology model of the heating area is constructed based on a graph neural network. Building units are treated as graph nodes, the connections of the heating network are treated as graph edges, and multi-source sensing data is mapped as attribute vectors of nodes and edges. This allows for the capture of thermal coupling relationships between buildings.

[0075] The parameters of the prediction models and heatmap models at each level are periodically collected, globally aggregated and optimized, and the models of each edge node are updated. This federated learning approach can achieve continuous model optimization while protecting data privacy.

[0076] Finally, based on the optimized model, the edge prediction results are obtained, and combined with the heat source-level scheduling prediction model, the heat source output power, main pump speed, and zone valve opening are jointly controlled. This multi-level linkage control can achieve precise heating regulation.

[0077] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0078] During real-time acquisition of multi-source sensing data, indoor temperature is collected via temperature sensors, user heat flow is collected via calorimeters, and valve opening is collected via electric valve status monitors. On the secondary pipe network side, pressure sensors collect pipe network pressure values, and flow sensors collect water flow velocity values. On the heat exchange station side, supply water temperature sensors and return water temperature sensors collect temperature information of the heat exchange inlet and return water. On the heat source side, main supply water temperature sensors and main return water temperature sensors acquire operating parameters of the heat source outlet and return water. Simultaneously, data such as temperature, humidity, wind speed, and light intensity from outdoor weather stations, as well as building structural parameters such as area and insulation performance, are collected.

[0079] When constructing a multi-level heating load prediction model, the building-level edge prediction model adopts a long short-term memory network structure, with inputs including historical time series indoor temperature, heat flow rate, valve opening, building structure information, and outdoor meteorological parameters. The heat exchange station-level aggregate prediction model adopts a multi-layer feedforward neural network structure, with inputs including the prediction results of subordinate buildings, building information, number of users, and historical supply and return water temperatures. The heat source-level scheduling prediction model adopts a support vector regression structure, with inputs including the prediction results of each heat exchange station, supply and return water temperatures, and main pump speed.

[0080] The thermal topology graph model employs a graph convolutional neural network structure. Building units are treated as graph nodes, with node attributes including indoor temperature, heat flow rate, valve opening, and building structural information. The connections within the heating network are represented as graph edges, with edge attributes including water flow velocity, network pressure, and supply and return water temperatures. The structured representation of the nodes is learned through graph convolution operations.

[0081] A federated learning framework is used when periodically collecting model parameters. The central server periodically sends aggregation requests to the edge nodes to collect the locally trained model parameters. A weighted averaging strategy is used for global aggregation to form unified global model parameters. The updated parameters are then synchronously distributed to each edge node.

[0082] Finally, based on the edge prediction results and the heat source-level scheduling model, a global load distribution state diagram is constructed. A rolling optimization objective function is established based on the model predictive control algorithm to minimize energy consumption and pressure fluctuations while meeting heating demand. The solution yields the adjustment commands for the main pump speed and the opening of regional valves, which are then sent to the execution unit via the cloud-edge collaborative control interface.

[0083] Through the above scheme, this application achieves accurate prediction and coordinated control of multi-level heating loads. Edge computing and federated learning mechanisms reduce data transmission pressure and improve response speed. The graph neural network-based thermal topology model effectively captures the thermal coupling relationships between buildings, improving prediction accuracy. The multi-objective optimization control strategy achieves improved energy efficiency and stable operation of the heating system.

[0084] In some of the solutions described above in this application, during the real-time acquisition of multi-source sensing data, due to the dispersed data sources and the involvement of different physical quantity types, the sampling period, data format, and transmission protocol of each terminal device are different, resulting in low data integration efficiency and insufficient real-time performance, making it difficult to meet the high-precision input requirements of multi-level prediction models.

[0085] This application further proposes real-time acquisition of multi-source sensing data, including: on the heat user side, acquiring indoor temperature, user heat flow rate, and valve opening degree through temperature sensors, calorimeters, and electric valve status monitors, respectively; on the secondary pipeline side, acquiring pipeline pressure and water flow velocity values ​​through pressure sensors and flow sensors; on the heat exchange station side, acquiring the temperature information of the heat exchange inlet and return water through supply water temperature sensors and return water temperature sensors; and on the heat source side, acquiring the operating parameters of the heat source outlet and return water through main supply water temperature sensors and main return water temperature sensors. Each IoT terminal device acquires the corresponding physical quantity signal according to a preset data sampling period, and performs local encoding and format encapsulation of the acquisition results, sending them to the edge node via MQTT communication protocol and Ethernet. The edge node performs time alignment, anomaly removal, and structural standardization processing on the acquired data.

[0086] Temperature sensors, calorimeters, and electric valve status monitors are configured to monitor key operating parameters on the heat user side, respectively. Pressure and flow sensors capture the hydraulic status on the secondary pipe network side. Supply and return water temperature sensors cover the heat exchange process monitoring on the heat exchange station side, while main supply and return water temperature sensors collect operating parameters on the heat source side. IoT terminal devices employ a preset sampling period to ensure data acquisition synchronization. Local encoding and format encapsulation steps reduce transmission and parsing complexity through a unified data format. The combination of MQTT protocol and Ethernet transmission ensures real-time and reliable data transmission. Time alignment operations performed by edge nodes eliminate time discrepancies between multi-source data through timestamp matching. Anomaly removal filters invalid or erroneous data based on preset thresholds, and structural standardization converts heterogeneous data into a unified dimensional vector.

[0087] Specifically, on the heat user side, temperature sensors collect indoor temperature data at a sampling frequency of once per minute, calorimeters accumulate user heat flow using pulse counting, and electric valve status monitors detect valve opening via current signals. These three data sources are encapsulated into JSON format data packets at the local terminal and transmitted to the edge nodes via Ethernet. On the secondary pipeline side, pressure sensors monitor pipeline pressure fluctuations at a sampling rate of 0.5Hz, and flow sensors obtain water flow velocity values ​​using ultrasonic velocity measurement principles. Both data are verified by CRC and then published to the edge node subscription queue via the MQTT protocol. On the heat exchange station side, supply and return water temperature sensors use platinum resistance thermometers as sensitive elements, collecting temperature data every 30 seconds, which is then converted to a 4-20mA current signal and uploaded. On the heat source side, the main supply and return water temperature sensors employ distributed fiber optic temperature measurement technology, generating temperature distribution curves every 10 seconds. After receiving data, the edge node first performs millisecond-level time alignment according to the device clock synchronization protocol. Then, it uses the Z-score algorithm to detect and remove data points that deviate from the mean by three times the standard deviation. Finally, it arranges the parameters into a standardized tensor structure according to the time series for subsequent model training and inference.

[0088] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0089] On the heat user side, indoor temperature, heat flow rate, and valve opening are collected using temperature sensors, calorimeters, and electric valve status monitors. The temperature sensors are PT100 platinum resistance temperature sensors, installed at representative indoor locations, with a sampling period of 5 minutes. The calorimeter is an ultrasonic calorimeter, installed on the user's inlet pipe, with a sampling period of 15 minutes. The electric valve status monitor detects valve opening using Hall effect sensors, with a sampling period of 1 minute.

[0090] On the secondary pipeline network side, pressure and flow velocity values ​​are collected using pressure and flow sensors. The pressure sensors are piezoelectric pressure transmitters, installed at key nodes in the pipeline network, with a sampling period of 10 minutes. The flow sensors are electromagnetic flow meters, installed on the main pipeline, with a sampling period of 5 minutes.

[0091] Temperature information of the heat exchange inlet and outlet water is collected at the heat exchange station using supply water temperature sensors and return water temperature sensors. Thermocouple temperature sensors are used and are installed on the inlet and outlet water pipes of the heat exchange station, with a sampling period of 1 minute.

[0092] On the heat source side, operating parameters of the heat source outlet and return water are acquired using main supply water temperature sensors and main return water temperature sensors. The temperature sensors are RTD temperature sensors, installed on the heat source outlet and return water pipes, with a sampling period of 30 seconds.

[0093] Each IoT terminal device collects corresponding physical quantity signals according to a preset data sampling period, and encodes and encapsulates the collected results locally before sending them to edge nodes via MQTT communication protocol and Ethernet. Edge nodes utilize industrial-grade embedded computers to perform time alignment, anomaly removal, and structural standardization on the collected data. Time alignment uses linear interpolation to unify data from different sampling periods to a 1-minute interval. Anomaly removal employs the 3σ criterion to remove outliers outside the normal range. Structural standardization unifies data from different sources into JSON format.

[0094] Through the above technical solution, this application achieves efficient acquisition and preprocessing of multi-source heterogeneous data. The distributed data acquisition architecture reduces data transmission burden. Preliminary data processing at edge nodes reduces the computational pressure on the central server. Simultaneously, the use of standardized data formats and communication protocols enhances scalability and interoperability. Furthermore, outlier handling and time alignment of the acquired data improve data quality.

[0095] In some of the above-mentioned schemes in this application, a multi-level heating load prediction model was proposed to achieve graded regulation. However, in the process of building-level prediction, the building heat load prediction accuracy is limited due to the lack of full consideration of differences in building structure and dynamic time series characteristics, and it cannot accurately reflect the spatiotemporal variation of users' heating behavior.

[0096] This application further proposes constructing an input vector containing historical time series data at the edge nodes corresponding to each target building, based on indoor temperature, user heat flow, electric valve opening, building structure information, and outdoor meteorological parameters within the target building. A prediction model is established based on a neural network structure, specifically a Long Short-Term Memory network with temporal recursive units, and trained using the input vector to predict changes in building heat load within a predetermined time window. The trained model is deployed at the edge nodes, performing inference based on real-time sensing data to output building-level predicted heat load values.

[0097] The input vector integrates continuous sampling data from temperature sensors within the target building, instantaneous flow rates recorded by calorimeters, valve actuator opening feedback signals, and thermal resistance parameters of the building envelope. The time recursive unit employs a gating mechanism to filter features from historical heating data, and the Long Short-Term Memory (LSTM) network dynamically adjusts the state of its memory units through forget and input gates. The preset time window is set to the next two hours, and model training uses a mean squared error loss function combined with the Adam optimizer for parameter updates. During edge node deployment, the model inference cycle is synchronized with the data acquisition cycle; real-time perceived data, after preprocessing, is directly input into the local model to generate prediction results.

[0098] Specifically, the construction process of the building-level edge prediction model first involves collecting time-series indoor temperature data of the target building through edge nodes. This data records the temperature distribution of each room at a minute-by-minute frequency. User heat flow data is obtained through calorimeters, reflecting the building's overall instantaneous heat demand. Electric valve opening data is used to characterize the impact of user-initiated adjustments on the heat load. Building structural information, including wall thermal conductivity, window area ratio, and floor height parameters, is extracted through a building information model. Outdoor meteorological parameters, including temperature, wind speed, and solar radiation intensity data, are integrated. These multidimensional data are then time-aligned to form a timestamped input vector, with the input vector length set to a 24-hour historical data window. A Long Short-Term Memory (LSTM) network extracts temporal features from the input vector using a three-layer hidden layer structure, with the number of neurons in the hidden layers dynamically adjusted according to the building's heating area. During model training, a sliding window method is used to generate a training sample set, with each sample containing 24 hours of continuous input data and subsequent two hours of actual heat load labels. The trained model is then lightweighted and compressed through an edge computing platform, and the compressed model file is transmitted to the building's edge nodes via an encrypted channel. During the edge node deployment phase, the model is loaded into memory and runs, receiving preprocessed sensor data in real time and performing forward inference calculations to output the building's heat load change curve for the next two hours. The prediction results are temporarily stored through a local caching mechanism for subsequent use by the heat exchange station-level model. Edge computing enables localized data processing, avoiding the privacy risks caused by transmitting raw data externally, while also reducing the computational load on the central server.

[0099] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0100] At the edge node corresponding to each target building, an input vector containing historical time series is constructed based on indoor temperature, user heat flow, electric valve opening, building structure information, and outdoor meteorological parameters. Specifically, historical data from the past 24 hours can be selected as the time series length, with sampling occurring once per hour. Indoor temperature is collected by temperature sensors distributed across each floor of the building, user heat flow is measured by heat meters installed at the user's end, and electric valve opening is fed back through valve controllers. Building structure information includes static parameters such as the number of floors, building area, and insulation material grade. Outdoor meteorological parameters include outdoor temperature, humidity, and wind speed, which are obtained in real time from weather stations.

[0101] Furthermore, a prediction model is built based on a neural network structure. The neural network structure adopts a Long Short-Term Memory (LSTM) network with temporal recursive units. The LSTM network consists of one input layer, two hidden layers, and one output layer. The number of neurons in the input layer is consistent with the dimension of the input vector. The first hidden layer contains 64 LSTM units, the second hidden layer contains 32 LSTM units, and the output layer is a fully connected layer used to output the predicted value of the building's heat load for the next 24 hours. The LSTM network is trained based on the constructed input vector to obtain a prediction model for predicting changes in the building's heat load over the next 24 hours.

[0102] Therefore, the trained model is deployed on edge nodes. Edge nodes can be industrial computers or embedded devices deployed in building electrical distribution rooms. The model is packaged in Docker containers for easy remote updates. Inference is performed based on real-time sensing data, outputting building-level predicted heat load values ​​for the next 24 hours hourly.

[0103] Through the above technical solutions, this application achieves refined prediction of building-level heat load. By deploying the prediction model at edge nodes, the computational burden on the central server is reduced, improving real-time performance and scalability. The use of an LSTM network structure effectively captures the temporal characteristics of heat load changes, improving prediction accuracy. The input vector is constructed based on multi-source sensing data, comprehensively considering various factors affecting heat load, making the prediction results more accurate and reliable. The prediction time granularity is on the hourly level, providing a decision-making basis for the refined control of the heating system, and helping to improve heating quality and energy utilization efficiency.

[0104] In some of the above-mentioned schemes in this application, the building-level edge prediction model can only output the heat load prediction value of a single building. However, the heat load changes of multiple buildings under the heat exchange station have spatiotemporal differences and coupling relationships. Simply accumulating the building prediction values ​​is not enough to accurately reflect the dynamic change trend of the overall heat load of the heat exchange station, resulting in an increase in the prediction deviation of the total heat load at the station level, which affects the accuracy of the subsequent heat source scheduling strategy.

[0105] This application further proposes that, at the edge node corresponding to each heat exchange station, the building-level predicted heat load values ​​output by several building-level edge prediction models under the heat exchange station are received, and an aggregated input vector is constructed by combining the building structure information, number of users, historical supply water temperature, historical return water temperature, and current outdoor meteorological parameters. A multi-layer feedforward neural network is used as the network structure of the heat exchange station-level aggregated prediction model, and the model is trained based on the aggregated input vector to predict the total heat load of the heat exchange station within a preset time window. The trained aggregated prediction model is deployed in the heat exchange station-level edge nodes, and the heat exchange station-level predicted heat load value is output based on the real-time prediction results of each building-level edge prediction model.

[0106] The aggregated input vector includes building-level predicted heat load values, building structural parameters, number of users, historical temperature data, and real-time meteorological data, forming an input feature space through multi-dimensional data fusion. The multi-layer feedforward neural network employs a fully connected layer structure, with ReLU activation function for the hidden layers and a linear activation function for the output layer. The loss function is set as the mean squared error index. During training, the stochastic gradient descent algorithm is used to optimize the network weights, with a batch training sample size of 64 data sets per batch and an initial learning rate of 0.001.

[0107] Specifically, after receiving the predicted heat load values ​​from the building-level models, the edge nodes at the heat exchange station level normalize the building structure parameters of each building according to the proportions of wall thickness, floor height, and window area. The number of users is converted into a density index based on the actual number of people requiring heating. Historical supply and return water temperature data are processed by extracting the mean, variance, and trend characteristics from the time series data. Current outdoor meteorological parameters include temperature, wind speed, and humidity. After standardization, the above data is concatenated into a fixed-dimensional aggregated input vector, which is then input into a multi-layer feedforward neural network for nonlinear mapping. During the network training phase, the weight parameters are adjusted using a backpropagation algorithm, enabling the model to capture the combined effects of the cumulative heat load between buildings and the combined impact of heat loss from the pipe network. After deployment, the trained model receives building-level prediction results in real time and integrates environmental parameters, outputting the predicted total heat load value of the heat exchange station within the time window. This predicted value serves as the input parameter for the heat source scheduling model, used to optimize the heat allocation strategy.

[0108] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0109] At the edge node corresponding to each heat exchange station, the predicted heat load values ​​for each building are received from the edge prediction models of several buildings under the heat exchange station. For example, if a heat exchange station has 10 buildings, the edge prediction model for each building will output the predicted heat load value for the next 24 hours. These predicted values ​​are transmitted to the edge node of the heat exchange station.

[0110] Next, an aggregated input vector is constructed by combining the building structure information, number of users, historical supply water temperature, historical return water temperature, and current outdoor meteorological parameters. Specifically, building structure information includes the number of floors, building area, and insulation materials. The number of users refers to the number of heating users in each building. Historical supply and return water temperatures are taken from the data of the most recent 7 days. Current meteorological parameters include outdoor temperature, wind speed, and humidity. This information, together with the building-level predicted heat load value, constitutes the aggregated input vector.

[0111] Then, a multi-layer feedforward neural network was used as the network structure for the heat exchange station-level aggregated prediction model. This network contains three hidden layers, each with 64, 32, and 16 neurons respectively, using ReLU as the activation function. The input layer corresponds to the dimension of the aggregated input vector, and the output layer consists of one neuron, corresponding to the predicted heat load value of the heat exchange station.

[0112] Furthermore, the model is trained based on the aggregated input vector to predict the total heat load of the heat exchange station within a preset time window. The training data uses historical data from the past three months, with a batch size of 64, a learning rate of 0.001, and 100 training epochs. The preset time window is set to 24 hours, meaning the model outputs the predicted heat load of the heat exchange station for the next 24 hours.

[0113] Finally, the trained aggregated prediction model is deployed to the edge nodes at the heat exchange station level. The model performs inference at 1-hour intervals, outputting the predicted heat load value at the heat exchange station level based on the real-time prediction results of each building-level edge prediction model. The prediction results are used to guide the operation and adjustment of the heat exchange station.

[0114] Through the above technical solution, this application achieves accurate prediction of heat load at the heat exchange station level. This allows for advance adjustment of the heat exchange station's operating parameters based on the prediction results, avoiding insufficient or excessive heating. Furthermore, this solution fully utilizes edge computing resources, reducing the computational burden on the central server. Simultaneously, by aggregating prediction results from multiple buildings, the accuracy and stability of heat exchange station-level prediction are improved. Moreover, the use of a neural network model enables adaptive learning of complex nonlinear relationships, exhibiting stronger generalization capabilities compared to traditional methods.

[0115] In some of the above-mentioned schemes in this application, the building-level edge prediction model and the heat exchange station-level aggregate prediction model have achieved hierarchical prediction. However, the heat source-level scheduling prediction needs to integrate data from multiple heat exchange stations and heat source operating parameters. Traditional centralized modeling is difficult to effectively capture the nonlinear correlation between multi-source heterogeneous data, resulting in insufficient prediction accuracy of heat source output power and affecting the dynamic control efficiency of main pump speed and valve opening.

[0116] This application further proposes a method on the central server to receive predicted heat load values ​​from edge nodes at each heat exchange station level, and to collect data on supply water temperature, return water temperature, main pump speed, and outdoor meteorological parameters to construct a scheduling input vector. A heat source-level scheduling prediction model is constructed based on a support vector regression model. This model is trained using the scheduling input vector to predict the output power change trend of the heat source within a preset time window. The heat source-level scheduling prediction model is deployed on the central server, and its inference results serve as the basis for heat source output regulation, main pump frequency conversion control, and regional valve opening strategies.

[0117] The scheduling input vector integrates the predicted heat load values, supply water temperature, return water temperature, main pump speed, and outdoor meteorological parameters uploaded by each heat exchange station to form a multi-dimensional feature combination. The support vector regression model transforms the input vector to a high-dimensional space through kernel function mapping, constructs an optimal hyperplane to fit the nonlinear relationship in the historical data, and updates the predicted output power change trend value through a sliding time window mechanism. The main pump speed and valve opening strategy is dynamically adjusted based on the predicted values, and the gradient descent method is used to optimize the objective function, balancing energy consumption and heating stability.

[0118] Specifically, the predicted heat load values ​​at each heat exchange station are uploaded to the central server and combined with real-time data collected on the supply and return water temperatures of the heat source side to form a scheduling input vector containing time-series characteristics. The support vector regression model processes the nonlinear features in the input vector using a radial basis function kernel. During the training phase, a sequential minimum optimization algorithm is used to solve a constrained quadratic programming problem to determine the support vectors and decision boundaries. After deployment, the model receives the latest input vector every five minutes and generates a predicted heat source output power curve for the next hour through forward inference. The prediction results are input to the frequency converter, which, combined with the current main pump speed and pipeline pressure feedback, generates a main pump speed adjustment command. Simultaneously, the opening of regional valves is controlled using proportional-integral-derivative (PID) control based on the spatial distribution of predicted power and building node heat demand, achieving dynamic matching between heat source output and load demand.

[0119] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0120] The central server receives predicted heat load values ​​for each heat exchange station from its edge nodes. Simultaneously, it collects supply water temperature, return water temperature, main pump speed, and outdoor meteorological parameters to construct a scheduling input vector. Specifically, predicted heat load values ​​for each heat exchange station are collected at 15-minute intervals, supply and return water temperatures at 1-minute intervals, main pump speed at 5-minute intervals, and outdoor meteorological parameters, including temperature, humidity, and wind speed, at 10-minute intervals.

[0121] Furthermore, a heat source-level scheduling prediction model is constructed based on a support vector regression model. This model uses a radial basis function as the kernel function, with a penalty factor C set to 100 and an epsilon parameter set to 0.1. The model input is the aforementioned scheduling input vector, and the output is a sequence of predicted heat source output power values ​​at 15-minute intervals over the next 4 hours.

[0122] Therefore, the heat source-level scheduling prediction model learns the nonlinear mapping relationship between the scheduling input vector and the heat source output power through training. The model training employs cross-validation, dividing historical data into training and test sets in a 7:3 ratio. During training, a grid search method is used to optimize the model's hyperparameters, with root mean square error (RMSE) used as the evaluation metric.

[0123] Specifically, the heat source-level scheduling prediction model is deployed on a central server. Model inference is performed every 15 minutes, generating a prediction of the heat source output power trend for the next 4 hours. For example, the prediction might show that the heat source output power will steadily increase by 10% over the next 2 hours, remain stable for the next hour, and then slowly decrease by 5% in the final hour.

[0124] As a preferred implementation, the inference results of the heat source-level scheduling prediction model are used as the basis for heat source output regulation, main pump frequency conversion control, and zone valve opening strategies. Specifically, based on the predicted trend of heat source output power changes, the fuel supply to the heat source boiler is adjusted in advance to achieve a smooth transition in heat source output. Simultaneously, feedforward control of the main pump speed is implemented based on the prediction results to avoid drastic fluctuations in heating system pressure. Furthermore, combined with the predicted heat load distribution, the valve openings in each zone are optimized to achieve a rational distribution of heat.

[0125] Through the above technical solution, this application enables accurate prediction and forward-looking control of heat source output power. Due to the use of a support vector regression model, this solution has strong generalization ability and can adapt to the characteristics of different heating systems. By accurately grasping the future trend of heat load changes, the output of the heat source can be adjusted in advance to avoid insufficient or excessive heating. Simultaneously, coordinated control of the main pump and zone valves based on the prediction results can improve operational stability and reduce pressure fluctuations and energy losses. This application can effectively improve the energy efficiency level of the heating system, reduce operating costs, and ensure heating quality.

[0126] In some of the above-mentioned schemes of this application, a multi-level heating load prediction model is proposed to achieve graded regulation. However, in the prediction process at the building level and heat exchange station level, due to the lack of spatial modeling of the connection relationship between the heat pipe network between building units, the model cannot accurately capture the heat conduction path and pipe network pressure transmission effect between adjacent buildings, resulting in a deviation between the prediction results and the actual heat load distribution.

[0127] This application further proposes constructing an undirected graph structure containing both a set of nodes and a set of edges, using each building unit within the heating area as a graph node and building pairs physically connected via a heating network as graph edges. The indoor temperature, user heat flow rate, valve opening, and building structure information of each building unit are used as node attribute vectors, while water flow velocity, network pressure, and heat exchange supply and return water temperatures are used as edge attribute vectors. A graph neural network model is constructed based on a graph convolutional neural network to embed the graph structure and generate a structured representation of each building node.

[0128] The undirected graph structure is constructed by parsing the physical connections in the heating network topology database. Each building unit corresponds to a unique node identifier, and two building units connected by heating pipes form a bidirectional edge. The node attribute vector consists of four-dimensional data, including indoor temperature measurements, instantaneous heat flow to the user, valve opening percentage, and building volume parameters. The edge attribute vector includes pipe water flow velocity, pressure sensor readings, and the difference between supply and return water temperatures. The graph convolutional neural network uses two graph convolutional layers. The first layer performs a linear transformation on the node and edge attributes, and the second layer generates node embedding vectors through neighborhood aggregation.

[0129] Specifically, in the thermal topology graph model, after building units are mapped to graph nodes, the valve opening data in the node attribute vector, together with the building structure information, reflects the building's heat demand characteristics. For example, larger building nodes correspond to higher base heat load coefficients. The water flow velocity and pipe network pressure values ​​in the edge attribute vector are used to characterize the transmission efficiency of the heat pipes. When the water flow velocity of a certain edge is lower than a threshold, its corresponding edge attribute vector will trigger the attention mechanism of the graph convolutional layer, suppressing the feature propagation of inefficient pipes. Through graph convolution operations, the embedding vector of each node integrates its own attributes and the thermal state information of neighboring nodes. For example, when the water supply temperature of a certain building node rises abnormally, the temperature components in the embedding vectors of its neighboring nodes will be updated synchronously. The generated node structured representation is further input into the federated learning module for joint optimization with the building-level prediction results, enabling the heat source-level scheduling model to dynamically adjust the regional valve opening based on spatial correlation characteristics.

[0130] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0131] When constructing a thermal topology graph model of a heating area based on a graph neural network, the first step is to use each building unit within the heating area as a graph node, and the building pairs that are physically connected to each other through the heating network as graph edges, thus constructing an undirected graph structure containing both a set of nodes and a set of edges. For example, for a heating area containing 100 building units, an undirected graph with 100 nodes can be constructed, where the connections between nodes are determined by the actual heating network topology.

[0132] Next, the indoor temperature, user heat flow, valve opening degree, and building structure information of each building unit are used as node attribute vectors, while the water flow velocity, pipe network pressure, and heat exchange supply and return water temperatures are used as edge attribute vectors. Specifically, for each building node, a multi-dimensional vector containing information such as average indoor temperature, cumulative heat consumption, percentage of main valve opening degree, building area, and number of floors can be constructed as the node attribute. For the edge connecting two building nodes, the average water flow velocity, pressure, and supply and return water temperatures of the corresponding pipe segment can be used to form the edge attribute vector.

[0133] Furthermore, a graph neural network model is constructed based on a graph convolutional neural network to perform embedding learning on the graph structure and generate a structured representation of each building node. Thus, a neural network architecture with multi-layer graph convolutional operations can be selected, and node features are aggregated and updated through a message-passing mechanism. In each layer of graph convolution, the feature vector of a node is weighted and combined with the features of its neighboring nodes to capture local topological information. After multiple layers of propagation, a low-dimensional embedding representation of each building node is finally obtained, which integrates information about the node's own attributes and the surrounding thermal network environment.

[0134] Through the above technical solution, this application can effectively capture the thermal coupling relationship between building units in a heating system and map the complex thermal network structure into a structured vector representation. This graph neural network-based modeling method overcomes the limitation of traditional methods that ignore the heat conduction paths between buildings, and can more accurately characterize the spatial distribution features of heat load. Simultaneously, by mapping multi-source sensing data into the node and edge attributes of a graph, a comprehensive representation of the operating status is achieved. This structured data representation provides richer feature information for subsequent load forecasting and control optimization, helping to improve prediction accuracy and control effectiveness.

[0135] In some of the solutions described above in this application, the multi-level heating load prediction model and the thermal topology model are deployed at edge nodes at different levels. Each edge node independently performs model training and parameter updates, which leads to local optimization deviations in the model parameters and makes it difficult to achieve global consistency. At the same time, frequent transmission of raw data to the central server may cause network bandwidth pressure and privacy leakage risks.

[0136] This application further proposes deploying a federated learning module on the central server. The federated learning module sends aggregation requests to each edge node at preset time intervals to collect local training parameters from the building-level edge prediction model, the heat exchange station-level aggregated prediction model, and the graph neural network model. The federated learning module performs a weighted average strategy on the collected model parameters from each edge node to perform global aggregation, forming unified global model parameters. These global model parameters are then updated and synchronously distributed to each corresponding edge node to replace the local prediction model parameters.

[0137] The federated learning module employs a distributed architecture, with preset time periods dynamically adjusted based on network load and model update frequency; for example, parameter aggregation is triggered every 30 minutes. The weighted averaging strategy assigns weight coefficients based on the data volume or prediction accuracy differences among edge nodes, with nodes having larger data volumes receiving a weight coefficient of 0.6 and nodes with smaller data volumes receiving a weight coefficient of 0.4. Global model parameters are distributed to edge nodes via an encrypted transmission channel, and historical parameter backups are retained when replacing local model parameters to ensure model rollback capability.

[0138] Specifically, after the federated learning module starts on the central server, it sends parameter aggregation commands to building-level and heat exchange station-level edge nodes at preset time intervals. Each edge node compresses and encodes its local model weight matrix and gradient information, and uploads it to the central server via a secure transmission protocol. The central server decodes and verifies the received local parameters, calculates the weight ratio based on the number of data samples corresponding to each node, and performs a weighted average calculation to generate global model parameters. After format conversion, the global model parameters are distributed to each edge node through the edge computing gateway. Before loading new parameters, the edge nodes verify the digital signature and overwrite the local model storage area after confirming the parameter integrity. After the model update is completed, each edge node performs the next round of local training based on the global parameters, while retaining the previous version of the parameters as a disaster recovery backup to prevent model failure due to aggregation anomalies.

[0139] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0140] A federated learning module is deployed on the central server. This module sends aggregation requests to each edge node at preset time intervals to collect local training parameters of the building-level edge prediction model, the heat exchange station-level aggregation prediction model, and the graph neural network model. Specifically, the federated learning module can send an aggregation request to the edge nodes every 6 hours to collect the training parameters of each model in the most recent period.

[0141] The federated learning module performs a weighted average strategy on the collected model parameters from each edge node to globally aggregate them, forming a unified global model parameter. For example, the FedAvg algorithm can be used to perform a weighted average of the model parameters from different nodes, with the weights determined based on the amount of data and computing power of each node.

[0142] The global model parameters are updated and synchronously distributed to each corresponding edge node to replace the local prediction model parameters. Specifically, the central server sends the aggregated global parameters to each edge node via a secure communication protocol, and the edge nodes update their local model parameters immediately upon receiving them.

[0143] Through the above technical solutions, this application achieves collaborative optimization between distributed edge nodes and the central server, improving the overall performance of the model while protecting user privacy. Simultaneously, the periodic parameter aggregation and update mechanism enables the model to continuously adapt to the dynamic changes of the heating system, improving prediction accuracy and adaptability. Furthermore, the federated learning-based distributed training method reduces the computational burden on the central server, improving scalability and real-time response capabilities.

[0144] In some of the schemes mentioned above in this application, if the federated learning module adopts a unified optimization strategy for different types of models when aggregating the local parameters uploaded by each edge node, it may cause parameter update conflicts between the building-level prediction model and the graph neural network model, affecting the adaptability of the global model to local data. At the same time, it is difficult to maintain the weight characteristics of highly correlated edges in the thermal topology graph, reducing the ability to represent the thermal conduction relationship between building nodes.

[0145] This application further proposes a differentiated optimization strategy for different model types when the federated learning module aggregates and optimizes the local parameters uploaded by each edge node. This includes: for building-level edge prediction models and heat exchange station-level aggregated prediction models, a federated averaging algorithm based on model weight sparsity constraints is used to regularize parameter updates. For the node embedding parameters of the graph neural network in the thermal topology graph model, the weight distribution of highly correlated edges is preserved during the aggregation process. After aggregation, the consistency loss function is used to verify the degree of fit of the aggregated model to the edge prediction results. If a preset fitting threshold is met, a model synchronization process is initiated.

[0146] Among them, the federated averaging algorithm based on model weight sparsity constraints limits the proportion of non-zero elements in the model weights by adding an L1 regularization term, making the parameter matrices of the building-level and heat exchange station-level models exhibit sparse characteristics, reducing communication overhead while retaining key feature dimensions. For the node embedding parameters of the graph neural network, the connection relationship between water flow velocity and pipe network pressure exceeding a set threshold is identified in the edge attribute vector during the aggregation stage, maintaining its weight update amplitude at no less than 85% of the original value, thereby strengthening the topological expression of high heat conduction paths between building units. The consistency loss function uses KL divergence to measure the distribution difference between the global model and the prediction results of each edge node; when the divergence value is below 0.05, the model is considered to have met the fitting criteria.

[0147] Specifically, during the parameter aggregation phase, the weight matrix of the building-level edge prediction model undergoes sparsification, retaining only the top 10% of parameters by absolute value for global averaging, while setting the remaining parameters to zero, effectively compressing the amount of transmitted data. The weight update process of the heat exchange station-level aggregation prediction model introduces a momentum factor, weightedly fusing historical gradient directions with the current gradient to improve model convergence speed. For the thermal topology graph model, when embedding parameters at aggregation nodes, the weighted score of water flow velocity and pipe network pressure for each edge attribute vector is calculated, selecting the top 20% of highly correlated edges. Their corresponding weights maintain their original numerical range during aggregation, avoiding feature decay of important heat conduction paths. After parameter aggregation is complete, the central server extracts the latest 100 sets of sample data collected from each edge node and inputs them into the global model. The difference in probability distribution between the prediction results and the local model output is calculated. If the mean KL divergence of three consecutive iterations is below a threshold, the global parameters are distributed to the edge nodes to replace the local model. This process, through differentiated parameter update mechanisms for different types of models, maintains the key topological characteristics of the thermal network while ensuring the model's generalization ability, enabling the global model to adapt to dynamic load changes in different building units.

[0148] As a preferred embodiment, the specific implementation of this application is as follows: The federated learning module processes the building-level edge prediction model parameters and heat exchange station-level aggregate prediction model parameters uploaded by the edge nodes using a federated averaging algorithm based on model weight sparsity constraints. Specifically, for the weight matrix of the long short-term memory network in the building-level edge prediction model, L1 regularization is used to filter out weight parameters with absolute values ​​greater than a preset threshold, and only the non-zero parameters after sparsification are retained for global aggregation. For the weight parameters of the multilayer feedforward neural network in the heat exchange station-level aggregate prediction model, a channel-level pruning strategy is adopted, and the top 30% of channel parameters are retained according to the L2 norm of each hidden layer channel. For the node embedding parameters of the graph neural network in the thermal topology graph model, the cosine similarity of each edge weight is calculated during the aggregation process, and the weight distribution corresponding to the edge connection relationship with a similarity higher than 0.85 is retained. After aggregation, the mean squared error is used as the consistency loss function to calculate the fitting error of the global model to the prediction results of each edge node after aggregation. When the error value is lower than 0.05, it is determined that the fitting threshold is met, and the model parameter synchronization process is triggered.

[0149] Through the above technical solutions, this application effectively solves the problem of low parameter aggregation efficiency caused by differences in model structure during federated learning. It reduces the amount of communication data and accelerates model convergence through differentiated optimization strategies. Specifically, sparse constraints and pruning operations reduce the transmission redundancy of non-critical parameters, the edge weight preservation mechanism maintains the spatial correlation characteristics of the hot topology graph, and consistency loss verification ensures the compatibility between the global model and edge data, thereby improving prediction accuracy while optimizing energy consumption.

[0150] In some of the schemes mentioned above in this application, the federated learning module adopts a unified optimization strategy when aggregating the model parameters of each edge node, without considering the impact of differences in different model structures on parameter aggregation. This leads to conflicts in parameter updates between the building-level edge prediction model and the heat topology model, affecting the convergence speed and prediction accuracy of the global model.

[0151] This application further proposes a differentiated optimization strategy for different model types when the federated learning module aggregates and optimizes the local parameters uploaded by each edge node. This includes: for building-level edge prediction models and heat exchange station-level aggregated prediction models, a federated averaging algorithm based on model weight sparsity constraints is used to regularize parameter updates. For the node embedding parameters of the graph neural network in the thermal topology graph model, the weight distribution of highly correlated edges is preserved during the aggregation process. After aggregation, the consistency loss function is used to verify the degree of fit of the aggregated model to the edge prediction results. If a preset fitting threshold is met, a model synchronization process is initiated.

[0152] Specifically, for the building-level edge prediction model and the heat exchange station-level aggregated prediction model, the federated averaging algorithm based on model weight sparsity constraints aggregates by selecting key weight parameters. For example, L1 regularization constraints are applied to the weights of gated units in the LSTM network, retaining only parameters with absolute values ​​higher than a set threshold for global updates, thereby reducing communication overhead and improving model generalization ability. For the node embedding parameters of the thermal topology graph model, the weight distribution of highly correlated edges is determined by calculating the similarity of edge attribute vectors. For example, edges with a correlation between water flow velocity and network pressure higher than 0.8 are defined as highly correlated edges, maintaining consistency in their weight update direction during aggregation. The consistency loss function uses the mean squared error form to calculate the prediction error of the aggregated model for each heat exchange station-level predicted heat load value. Model synchronization is triggered when the error is lower than 0.05.

[0153] Specifically, within a preset time period, the federated learning module first collects the LSTM network weight matrix of the building-level edge prediction model and the multilayer feedforward network weight matrix of the heat exchange station-level aggregate prediction model. Through sparsification, it filters out weight parameters in each model whose contribution exceeds a set proportion. For the graph convolutional network parameters of the thermal topology graph model, it extracts node embedding vectors and edge weight matrices, identifies highly correlated edges based on the Pearson correlation coefficient between edge attribute vectors, and retains their weight update gradients. Subsequently, it performs a weighted average on the filtered building-level and heat exchange station-level model parameters, dynamically adjusting the weight ratio based on the historical prediction accuracy of each edge node; for example, the weight of nodes with prediction errors below 5% is increased to 1.2 times. For the graph neural network parameters, it uses weighted aggregation based on edge correlation, with highly correlated edge weights participating in the update at a proportion of 0.7. After aggregation, the global model parameters are distributed to each edge node, and the consistency loss is calculated based on the real-time collected heat exchange station-level predicted heat load values. If the loss value is below a threshold for three consecutive periods, the global parameters are synchronized to all nodes. As a result, the parameter update efficiency of building-level and heat exchange station-level models is improved by about 30%, and the graph neural network improves the accuracy of heat conduction path representation by 15%.

[0154] As a preferred embodiment, the specific implementation of this application is as follows: During the operation of the heating system, building-level edge nodes calculate the predicted building heat load for the next two hours in real time based on the trained LSTM model. Heat exchange station-level edge nodes aggregate the predicted values ​​of subordinate buildings through a multi-layer feedforward neural network to generate the station-level total heat load. The heat source-level scheduling prediction model receives the predicted data from each heat exchange station and the building node thermal correlation matrix output by the thermal topology model, and predicts the heat source output power trend through a support vector regression model. The global load distribution state diagram consists of the predicted heat source output power, the predicted heat exchange station load, and the three-dimensional node embedding vectors generated by the graph neural network. The node spatial coordinates reflect the building's geographical location, and the color gradient represents the real-time heat demand intensity. The model predictive control algorithm establishes an optimization objective function containing the network pressure variance term and the pump energy consumption quadratic term using a rolling time window approach. Under the condition of satisfying the heating temperature constraints of each area, it solves the main pump speed adjustment and valve opening control sequence through quadratic programming. The control commands are encoded in JSON format and transmitted to the main pump frequency converter via a RESTful API interface. At the same time, they are sent to the intelligent valve actuators distributed in the heating network via the OPC UA protocol.

[0155] Through the above technical solutions, this application effectively solves the problems of insufficient spatial correlation modeling and poor real-time control in traditional heating system load forecasting. By constructing a global load distribution state diagram, a unified representation of the multi-dimensional operating state of the heating system is achieved. Combined with model predictive control algorithms, multi-objective dynamic optimization of heating parameters is realized, reducing the amplitude of pipeline pressure fluctuations and the energy consumption level of the main pump while ensuring heating quality. The cloud-edge collaborative command transmission mechanism ensures millisecond-level response capability of control commands, avoiding the communication latency problem existing in traditional centralized control architectures.

[0156] The above embodiments, by introducing a collaborative mechanism of "IoT sensing + hierarchical modeling + graph neural network + federated learning optimization," achieve high-precision, hierarchical, and spatially coupled sensing and predictive control of the heating system's heat load. This application deploys multiple types of sensor devices on the heat user side, secondary pipe network side, heat exchange station side, and heat source side to comprehensively collect key operating parameters in the heat transmission and distribution process. By constructing multi-level prediction models at the building, heat exchange station, and heat source levels, and combining graph neural networks to mine the thermal topological relationships between buildings, the ability to characterize dynamic load changes and spatial heat conduction paths is improved. The federated learning module further breaks down data silos, enabling collaborative optimization of each prediction model while protecting data privacy, ensuring the generalization ability of model updates and the real-time responsiveness of edge deployment. Combined with a heat source-level scheduling model, it achieves linked control of heat source output, main pump speed, and valve opening, improving the heating system's energy efficiency and control flexibility while ensuring prediction accuracy, effectively solving the problems of coarse prediction granularity, lack of spatial modeling, and bottlenecks in centralized computing architecture in current technologies.

[0157] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides a heating system load forecasting system for applying the above-described heating system load forecasting method, including:

[0158] The data acquisition unit is configured to acquire multi-source sensing data in real time. The multi-source sensing data includes indoor temperature, user heat flow rate, valve opening, pipeline pressure, water flow velocity, temperature information of heat exchange inlet and return water, and operating parameters of heat source outlet and return water. It also acquires outdoor meteorological parameters and building structure information.

[0159] The first processing unit is configured to construct a multi-level heating load prediction model based on the building distribution, pipeline structure and user load characteristics of the heating area. The multi-level heating load prediction model includes a building-level edge prediction model, a heat exchange station-level aggregation prediction model and a heat source-level scheduling prediction model.

[0160] The second processing unit is configured to construct a thermal topology graph model of the heating area based on a graph neural network, using building units as graph nodes, the connection relationship of the heating network as graph edges, and mapping multi-source sensing data into attribute vectors of nodes and edges.

[0161] The third processing unit is configured to periodically collect parameters from the building-level edge prediction model, the heat exchange station-level aggregate prediction model, and the thermal topology model, perform global aggregate optimization, and update and optimize the models of each edge node.

[0162] The adjustment unit is configured to obtain edge prediction results based on the optimized model, and then jointly control the heat source output power, main pump speed and zone valve opening based on the edge prediction results and the heat source-level scheduling prediction model.

[0163] Specifically, the acquisition unit collects physical quantity signals through IoT terminal devices at preset intervals, completes local encoding and format encapsulation, and transmits them to edge nodes using the MQTT protocol. The edge nodes perform time alignment and anomaly removal on the data. The first processing unit deploys a long short-term memory network in the building-level edge nodes to receive historical time series input vectors and output predicted heat load values. A multi-layer feedforward neural network is deployed in the heat exchange station-level edge nodes to aggregate the prediction results of subordinate buildings and output the total heat load at the station level. A support vector regression model is deployed in the central server to receive predicted values ​​from each heat exchange station and generate predicted heat source output power. The second processing unit uses a graph convolutional neural network to embed the thermal topology map, generating a structured representation of building nodes and capturing the heat conduction paths and spatial correlation characteristics between buildings. The third processing unit uses a federated learning module to perform sparse constraint and regularization processing on the edge node model parameters, preserving the weight distribution of highly correlated edges, and uses a consistency loss function to verify the fit of the aggregated model. Once a threshold is met, the models of each node are updated synchronously. The adjustment unit inputs the global load distribution state diagram into the rolling optimization objective function, and solves the adjustment instructions that minimize energy consumption and pressure fluctuations under the constraint of meeting heating demand. The instruction set is encoded and sent to the frequency converter and valve control unit through the control interface to realize multi-objective joint control.

[0164] As a preferred embodiment, the solution of this application is implemented as follows: The acquisition unit deploys temperature sensors on the interior walls of the building, 1.5 meters above the ground; pressure sensors are installed 2 meters downstream of the bend in the main pipeline; and flow sensors are installed on the vertical section of the water supply pipeline. Data is transmitted to edge nodes via an RS485 bus. The edge nodes use a sliding window mechanism to perform data alignment and Z-score normalization to eliminate dimensional differences. The first processing unit constructs a long short-term memory network model at the building-level edge nodes. The input layer contains temperature sequences at 8 time steps, and the hidden layer has 32 neurons. The output is a predicted heat load value for the next 2 hours. The second processing unit constructs a graph convolutional network model on the central server, with a node embedding dimension of 64. A gated graph attention mechanism is used to update edge weights. The third processing unit triggers federated aggregation every 30 minutes, employing a dynamic weight allocation strategy to apply L2 regularization constraints to the parameters of the heat exchange station-level model. The adjustment unit solves for the optimal main pump speed using a quadratic programming algorithm, and control commands are sent to the frequency converter via the OPC UA protocol.

[0165] Through the above technical solutions, this application achieves multi-level load prediction and collaborative optimization control of the heating system. By constructing a hierarchical prediction model and thermal topology structure, the thermal coupling relationship between buildings is effectively captured, improving the accuracy of load prediction. A federated learning mechanism is adopted to achieve distributed updating of model parameters, optimizing global prediction performance while ensuring data privacy. Based on real-time prediction results, multi-objective adjustment commands are generated to achieve dynamic balance between heat source output power and pipeline hydraulic state, reducing system pressure fluctuations and energy loss.

[0166] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0167] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0168] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0169] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting the load of a heating system, characterized in that, include: Real-time acquisition of multi-source sensing data, including indoor temperature, user heat flow rate, valve opening, pipeline pressure, water flow velocity, temperature information of heat exchange inlet and return water, and operating parameters of heat source outlet and return water, as well as acquisition of outdoor meteorological parameters and building structure information; Based on the building distribution, pipeline structure and user load characteristics of the heating area, a multi-level heating load prediction model is constructed. The multi-level heating load prediction model includes a building-level edge prediction model, a heat exchange station-level aggregation prediction model and a heat source-level scheduling prediction model. A thermal topology graph model of the heating area is constructed based on a graph neural network, with building units as graph nodes, the connection relationship of the heating network as graph edges, and the multi-source sensing data mapped as attribute vectors of nodes and edges. The parameters of the building-level edge prediction model, the heat exchange station-level aggregate prediction model, and the thermal topology model are periodically collected, and global aggregation optimization is performed. The models of each edge node are then updated and optimized. The edge prediction results are obtained based on the optimized model. The edge prediction results are then combined with the heat source-level scheduling prediction model to control the heat source output power, main pump speed, and zone valve opening. The parameters of the building-level edge prediction model, the heat exchange station-level aggregated prediction model, and the thermal topology model are periodically collected, and global aggregation optimization is performed. When updating and optimizing the models of each edge node, the following steps are included: A federated learning module is deployed on the central server. The federated learning module sends aggregation requests to each edge node at a preset time period to collect local training parameters of the building-level edge prediction model, the heat exchange station-level aggregation prediction model, and the graph neural network model. The federated learning module performs a weighted average strategy on the collected model parameters of each edge node to aggregate them globally, forming unified global model parameters. The global model parameters are updated and synchronously distributed to each corresponding edge node to replace the local prediction model parameters. When the federated learning module aggregates and optimizes the local parameters uploaded by each edge node, it adopts differentiated optimization strategies for different model types, including: For the building-level edge prediction model and the heat exchange station-level aggregate prediction model, a federated average algorithm based on model weight sparsity constraints is used to regularize the parameter updates. For the node embedding parameters of the graph neural network in the hot topology graph model, the weight distribution of highly associated edges is preserved during the aggregation process; After aggregation is completed, the consistency loss function is used to check the degree of fit of the aggregation model to the edge prediction results. If the preset fitting threshold is met, the model synchronization process is carried out.

2. The heating system load prediction method according to claim 1, characterized in that, When collecting multi-source sensing data in real time, including: On the heat user side, indoor temperature, user heat flow rate and valve opening are collected by temperature sensors, calorimeters and electric valve status monitors, respectively. On the secondary pipeline side, pressure sensors and flow sensors are used to collect pipeline pressure values ​​and water flow velocity values; Temperature information of the heat exchange inlet and outlet water is collected by supply water temperature sensors and return water temperature sensors at the heat exchange station side. On the heat source side, the operating parameters of the heat source outlet water and return water are obtained through the main supply water temperature sensor and the main return water temperature sensor; Each IoT terminal device collects corresponding physical quantity signals according to a preset data sampling period, and encodes and encapsulates the collection results locally. The results are then sent to the edge node via the MQTT communication protocol and Ethernet. The edge node performs time alignment, anomaly removal, and structural standardization on the collected data.

3. The heating system load prediction method according to claim 1, characterized in that, Based on the building distribution, pipeline structure, and user load characteristics of the heating area, a multi-level heating load prediction model is constructed. The construction process of the building-level edge prediction model includes: At each edge node corresponding to the target building, an input vector containing historical time series is constructed based on the indoor temperature, user heat flow, electric valve opening, building structure information and outdoor meteorological parameters within the target building. A prediction model is established based on a neural network structure, wherein the neural network structure is a long short-term memory network with time recursive units, and a prediction model for predicting the change of building heat load within a preset time window is obtained by training according to the input vector. The trained model is deployed on edge nodes to perform inference based on real-time sensing data and output building-level predicted heat load values.

4. The heating system load prediction method according to claim 3, characterized in that, The construction process of the heat exchange station-level aggregated prediction model includes: At the edge node corresponding to each heat exchange station, the building-level predicted heat load value output by several building-level edge prediction models under the heat exchange station is received, and an aggregated input vector is constructed by combining the building structure information, number of users, historical water supply temperature, historical return water temperature and current outdoor meteorological parameters. The network structure of the heat exchange station-level aggregated prediction model is based on a multi-layer feedforward neural network. The model is trained based on the aggregated input vector to predict the total heat load of the heat exchange station within a preset time window in the future. The trained aggregated prediction model is deployed in the edge nodes of the heat exchange station, and the predicted heat load value of the heat exchange station is output based on the real-time prediction results of each building-level edge prediction model.

5. The heating system load prediction method according to claim 4, characterized in that, The construction process of the heat source-level scheduling prediction model includes: On the central server, the predicted heat load values ​​of each heat exchange station are uploaded by edge nodes of each heat exchange station, and the supply water temperature, return water temperature, main pump speed and outdoor meteorological parameters are collected to construct the scheduling input vector. A heat source-level scheduling prediction model is constructed based on a support vector regression model. The heat source-level scheduling prediction model is trained according to the scheduling input vector and is used to predict the output power change trend of the heat source within a preset time window in the future. The heat source-level scheduling prediction model is deployed in the central server, and the inference results of the heat source-level scheduling prediction model are used as the basis for heat source output regulation, main pump frequency conversion control, and regional valve opening strategies.

6. The heating system load prediction method according to claim 5, characterized in that, When constructing a thermal topology graph model of a heating area based on a graph neural network, using building units as graph nodes and heating network connections as graph edges, and mapping the multi-source sensing data into attribute vectors for nodes and edges, the following steps are included: Using each building unit within the heating area as a graph node and building pairs that are physically connected through the heating network as graph edges, an undirected graph structure containing a set of nodes and a set of edges is constructed. The indoor temperature, user heat flow, valve opening and building structure information of each building unit are used as node attribute vectors, and the water flow velocity, pipeline pressure, heat exchange supply water temperature and return water temperature are used as edge attribute vectors. A graph neural network model is constructed based on graph convolutional neural networks. The graph structure is embedded and learned, and a structured representation of each building node is generated.

7. The heating system load prediction method according to claim 1, characterized in that, When edge prediction results are obtained based on the optimized model, and the heat source output power, main pump speed, and zone valve opening are jointly controlled based on the edge prediction results and the heat source-level scheduling prediction model, the following is included: Each building-level edge node and heat exchange station-level edge node outputs the predicted heat load values ​​of the building-level edge prediction model and the heat exchange station-level aggregate prediction model in real time within the prediction time window, respectively. The predicted heat load value at the heat exchange station level and the building thermal correlation information output by the thermal topology model are jointly input into the heat source level scheduling prediction model to obtain the predicted heat source output power value within the target time window. A global load distribution state map is constructed based on the predicted output power of the heat source, the predicted heat load of each heat exchange station, and the structured representation of building nodes generated by the graph neural network model as input features. Based on the global load distribution state diagram, a rolling optimization objective function is established based on the model predictive control algorithm. The rolling optimization objective function is used to minimize the total energy consumption and pressure fluctuation of the system while meeting the heating demand of each region. Under the constraints of the rolling optimization objective function, the main pump speed regulation signal and the regional valve opening regulation signal are obtained in real time, and multi-objective regulation commands are generated by combining the spatial gradient of heat demand of each node in the global load distribution state diagram. The multi-target adjustment commands are encoded in an instruction set structure and sent to the main pump frequency converter and the control units of the execution valves in each area through the cloud-edge collaborative control interface.

8. A heating system load forecasting system, used to apply the heating system load forecasting method as described in any one of claims 1-7, characterized in that, include: The acquisition unit is configured to acquire multi-source sensing data in real time. The multi-source sensing data includes indoor temperature, user heat flow rate, valve opening, pipeline pressure, water flow velocity, temperature information of heat exchange inlet and return water, and operating parameters of heat source outlet and return water. It also acquires outdoor meteorological parameters and building structure information. The first processing unit is configured to construct a multi-level heating load prediction model based on the building distribution, pipeline structure and user load characteristics of the heating area. The multi-level heating load prediction model includes a building-level edge prediction model, a heat exchange station-level aggregation prediction model and a heat source-level scheduling prediction model. The second processing unit is configured to construct a thermal topology graph model of the heating area based on a graph neural network, using building units as graph nodes, the connection relationship of the heating network as graph edges, and mapping the multi-source sensing data into attribute vectors of nodes and edges. The third processing unit is configured to periodically collect parameters from the building-level edge prediction model, the heat exchange station-level aggregate prediction model, and the thermal topology model, perform global aggregate optimization, and update and optimize the models of each edge node. The adjustment unit is configured to obtain edge prediction results based on the optimized model, and to jointly control the heat source output power, main pump speed and zone valve opening based on the edge prediction results and the heat source-level scheduling prediction model.

Images