Centralized heating load aggregation regulation and control method and system under virtual power plant framework

Abstract

The invention provides a centralized heating load aggregation regulation and control method and system under a virtual power plant framework. The method comprises the steps of collecting heat consumption behavior data, constructing a heat load prediction model, predicting a heat load curve in the next 24 hours, generating a thermal inertia compensation strategy and uploading the thermal inertia compensation strategy to a global layer. And then, constructing a digital twin heat network model, optimizing parameters of the regional layer heat load prediction model, generating a global heat load regulation and control instruction according to the model, issuing the global heat load regulation and control instruction to the regional layer, and dynamically adjusting heat supply parameters of a terminal user. According to the invention, the federated learning algorithm is used to realize the global optimization of the parameters of the cross-regional thermal load prediction model, and the dynamic simulation of the digital twin heat network model is combined to improve the thermal load prediction precision and the robustness of the regulation and control strategy. A user-side thermal inertia parameter library and an intelligent temperature control valve are used for dynamic compensation, global regulation and control instructions are coordinated for precise linkage, and on the premise that the comfort level of a user and the safety of a heat supply network are guaranteed, heat supply network fluctuation minimization, energy cost reduction and efficient aggregation regulation and control are achieved.

Description

Technical Field

[0001] This invention relates to the field of centralized heating load aggregation and control technology, and in particular to a method and system for centralized heating load aggregation and control under a virtual power plant framework. Background Technology

[0002] Currently, most centralized heating load control methods employ traditional approaches based on single-region prediction models. These models train regional-level heat load prediction models using locally collected user heating behavior data, generating 24-hour heat load curves to directly guide regional-level control. While some solutions introduce rolling prediction mechanisms or time-segmented peak-shaving strategies, they are limited by data silos. Each regional model is trained and optimized independently, lacking global-level collaborative optimization capabilities. This makes it difficult to balance heat load fluctuations between different regions, especially under extreme weather conditions or sudden heating demands, where traditional methods are prone to local heat source overload or insufficient heating. Furthermore, existing technologies do not fully utilize the multi-source data fusion capabilities within a virtual power plant framework, failing to deeply integrate GIS geographic information, heating network topology data, and user heating behavior data. This results in a lack of spatiotemporal dynamic adaptability in thermal inertia compensation strategies, hindering efficient aggregation and control of the global heat load. Moreover, existing technologies primarily rely on manual experience or fixed rules to set heat source output power and valve openings, failing to fully consider the dynamic characteristics of the heating network topology, the spatiotemporal differences in user-side thermal inertia, and the load aggregation effect between multiple regions. This leads to the risk of local optima but global imbalance in the control strategy. Summary of the Invention

[0003] This invention aims to at least address the technical problem of existing control strategies having the risk of local optima but global imbalance, and innovatively proposes a method and system for centralized heating load aggregation and control under a virtual power plant framework.

[0004] To achieve the above-mentioned objectives of this invention, this invention provides a method for centralized heating load aggregation and control under a virtual power plant framework, the method comprising: S1. Collect user heat usage behavior data at the end user and upload it to the regional layer; S2. Construct a heat load prediction model at the regional layer, and predict the heat load curve for the next 24 hours based on the user heat consumption behavior data; S3. Generate a thermal inertia compensation strategy based on the heat load curve and upload it to the global layer; S4. At the global level, a digital twin heating network model is constructed based on GIS geographic information and heating network topology data, and the parameters of the heat load prediction model for each regional layer are optimized through federated learning algorithm. S5. Based on the optimized heat load prediction model parameters and the digital twin heating network model, generate global heat load control instructions and send them to the regional layer. S6. After receiving the global heat load control command, the regional layer dynamically adjusts the heating parameters of each terminal user by combining the local heat source supply capacity and the user-side thermal inertia compensation strategy.

[0005] In another aspect, the present invention also provides a centralized heating load aggregation and control system under a virtual power plant framework, the system comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement the centralized heating load aggregation and control method under the virtual power plant framework when executing the executable instructions.

[0006] The beneficial effects of this invention are as follows: This invention achieves global optimization of cross-regional heat load prediction model parameters through federated learning algorithms, and combines digital twin heat network models with dynamic simulation of GIS geographic information and heat network topology. This breaks through the limitations of data silos and local optima in traditional methods, significantly improving the accuracy of heat load prediction and the robustness of control strategies. At the same time, it utilizes user-side thermal inertia parameter libraries and intelligent temperature control valves to achieve dynamic compensation for time-sharing peak shaving and valley filling, and coordinates global control commands to complete precise linkage between heat source, valve, and user. Ultimately, while ensuring user comfort and heat network safety, it achieves minimized heat network fluctuations, reduced energy costs, and efficient aggregation and control of multi-regional heat loads, effectively solving the problems of weak global coordination capabilities, poor dynamic adaptability, and low control efficiency in existing technologies.

[0007] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0008] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of a centralized heating load aggregation and control method under a virtual power plant framework according to the present invention. Detailed Implementation

[0009] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0010] Example 1 like Figure 1 As shown, a method for centralized heating load aggregation and control under a virtual power plant framework is described, the method comprising: S1. Collect user heat usage behavior data at the end user and upload it to the regional layer; In step S1, it is necessary to explain that IoT temperature sensors and smart valve controllers are deployed at the end user site to collect user heating behavior data in real time, including indoor temperature setpoint, actual room temperature, valve opening status, instantaneous supply / return water temperature and flow rate data; the collected data is encrypted and transmitted to the corresponding regional layer data processing center through an IoT communication gateway (such as LoRaWAN, NB-IoT or 5G); the regional layer preprocesses the received data, including data format verification, outlier removal (such as using the 3σ principle) and timestamp alignment, to ensure that the data quality meets the modeling requirements.

[0011] S2. Construct a heat load prediction model at the regional layer, and predict the heat load curve for the next 24 hours based on the user heat consumption behavior data; S3. Generate a thermal inertia compensation strategy based on the heat load curve and upload it to the global layer; S4. At the global level, a digital twin heating network model is constructed based on GIS geographic information and heating network topology data, and the parameters of the heat load prediction model for each regional layer are optimized through federated learning algorithm. S5. Based on the optimized heat load prediction model parameters and the digital twin heating network model, generate global heat load control instructions and send them to the regional layer. S6. After receiving the global heat load control command, the regional layer dynamically adjusts the heating parameters of each terminal user by combining the local heat source supply capacity and the user-side thermal inertia compensation strategy.

[0012] The principle of the centralized heating load aggregation and control method under the framework of a virtual power plant in this embodiment is as follows: First, in the regional layer (S1), real-time collection and preprocessing of end-user heat consumption behavior data are completed to construct a high-quality dataset. Based on this dataset, the regional layer (S2) trains a local heat load prediction model and outputs the heat load curve for the next 24 hours. Combining user type labels and fluctuation intensity (S3), time-segmented peak shaving, valley filling, and valve opening thermal inertia compensation strategies are generated and uploaded to the global layer. The global layer (S4) constructs a digital twin heat network model using GIS geographic information and heat network topology, and uses a federated learning algorithm to collaboratively optimize the model parameters of each regional layer: the federated server initializes global model parameters and downloads them to regional layer nodes; the regional layers perform model training based on local data, calculate gradients, and encrypt and upload model parameters; the global layer aggregates parameters to generate fusion weights, optimizes the digital twin model, and then sends it back to the regional layer. The global layer (S5) uses real-time environmental data as input to the optimized model to simulate the dynamic response of the heat network. Through multi-objective optimization, it generates a global control command set including valve opening commands and heat source power adjustments, and encrypts and sends it to the regional layers. After parsing the instructions at the regional layer (S6), the system dynamically calculates the valve opening adjustment values ​​for each terminal based on local heat source supply capacity and thermal inertia compensation strategies. These values ​​are then sent to intelligent temperature control equipment to dynamically adjust heating parameters. Real-time monitoring data is used to optimize subsequent strategies. Ultimately, this minimizes heating network fluctuations, reduces energy costs, and enables efficient aggregation and control of loads across multiple regions.

[0013] As an optional embodiment of the present invention, the step S2 of predicting the heat load curve for the next 24 hours based on the user's heat consumption behavior data may include: S201. Preprocess the user's heat usage behavior data; In step S201, preprocessing includes cleaning the collected raw data to remove missing values, duplicate values, and outliers caused by sensor malfunctions or communication anomalies (e.g., using the DBSCAN algorithm based on density clustering to identify and remove outlier data points); standardizing the valid data (e.g., Z-Score standardization) to map indicators of different dimensions (temperature, flow rate, etc.) to a unified numerical range; and reducing the frequency of high-frequency sampled data using the sliding window averaging method to generate a minute-level time series dataset while retaining key event markers (e.g., the moment of a sudden change in valve opening). Preprocessing also includes extracting user heating behavior characteristics. Specifically, it involves identifying user heating habits by analyzing periodic patterns (e.g., daily cycle, weekly cycle) in the time series data, and classifying user type labels (e.g., office, residential, commercial) using cluster analysis (e.g., K-Means algorithm); extracting statistical features (mean, variance, peak value) and time-domain features (autocorrelation coefficient, volatility) to construct a three-dimensional feature matrix containing user type, time label, and feature vector. Furthermore, Principal Component Analysis (PCA) is used to reduce the dimensionality of high-dimensional features, retaining principal components with a cumulative contribution rate exceeding 95% to generate low-dimensional feature representations. An extended feature set is constructed by combining weather forecast data (such as outdoor temperature, wind speed, and humidity) and calendar information (weekdays / holidays). Interaction terms (such as outdoor temperature × user type) are generated through feature cross-referencing to enhance the model's adaptability to environmental changes. Finally, the processed feature matrix is ​​input into the heat load prediction model constructed at the regional layer.

[0014] S202. Construct a heat load prediction model based on the preprocessed user heat consumption behavior data; In step S202, it is necessary to explain in detail that the heat load prediction model is constructed using a time-series-based deep learning framework. Specifically, a Long Short-Term Memory (LSTM) network is selected as the core model architecture. This model includes an input layer, at least two LSTM hidden layers, and a fully connected output layer. The input layer receives a preprocessed feature matrix with dimensions T×F, where T represents the time window length (e.g., 24 hours), and F represents the number of features (including user type labels, outdoor temperature interaction terms, and features after principal component dimensionality reduction). The number of neurons in the LSTM hidden layer is set to 128, the activation function is ReLU, and a Dropout layer (dropout rate 0.2) is added to enhance generalization ability. The output layer generates a heat load prediction sequence for the next 24 hours, corresponding to 24 consecutive output nodes.

[0015] S203. Train and optimize the heat load prediction model; In step S203, it is necessary to explain in detail that the model training process includes: initializing weight parameters (such as Xavier initialization), and minimizing the mean squared error loss function (such as the cross-entropy loss function) using the Adam optimizer. The training data is divided into a training set (70%), a validation set (15%), and a test set (15%), with a batch size of 32 and a maximum training epoch of 200.

[0016] An early stopping mechanism is introduced to monitor the loss on the validation set; training is terminated if the loss does not decrease after 10 consecutive rounds. After training, the model performance is evaluated using a test set, and the root mean square error (RMSE) and mean absolute percentage error (MAPE) are calculated. Hyperparameters are optimized through cross-validation (e.g., the learning rate is set to 0.001). Finally, the model outputs a heat load prediction curve, which is used to generate control strategies in subsequent steps.

[0017] The preferred model, specifically an LSTM-based heat load prediction model, demonstrated strong feature extraction and time-series prediction capabilities during training. Its long short-term memory structure effectively captures the periodic patterns and time-varying characteristics of user heating behavior. For example, through dynamic adjustment of the input, forget, and output gates, the model adaptively retains key features from historical information (such as weekday morning and evening heating peaks) while filtering out noise interference. During the validation phase, the model's RMSE value on the test set was 32% lower than the traditional ARIMA model, and its MAPE value decreased to 4.8%, indicating a significant improvement in prediction accuracy for extreme weather (such as cold waves) and sudden heating demands (such as commercial area activities). Furthermore, by introducing an attention mechanism to assign higher weights to key time steps (such as nighttime off-peak periods), the model further optimized the prediction bias of peak and off-peak loads.

[0018] S204. Based on the trained heat load prediction model, a rolling prediction mechanism is executed. The input is updated based on the latest real-time data at preset intervals to dynamically generate the heat load curve for the next 24 hours.

[0019] In step S204, it is necessary to explain in detail that the rolling prediction mechanism triggers model updates by setting a fixed time interval (e.g., every hour). It replaces historical data in the feature matrix with the latest collected real-time data (including current room temperature, valve opening, supply and return water temperatures, etc.) and re-executes the preprocessing process. Specifically, the system obtains the latest minute-level data from the IoT communication gateway, removes outliers using the DBSCAN algorithm, and concatenates it with historical data to form an extended time window (e.g., data from the past 48 hours). An updated minute-level time series is generated using the sliding window averaging method, and user heat consumption behavior characteristics (e.g., daily cycle patterns, statistical features) are re-extracted. This is combined with real-time weather forecast data (e.g., predicted outdoor temperature for the next 3 hours) and calendar information (e.g., whether it is a holiday) to construct a dynamically expanded feature set. The updated feature matrix is ​​input into the trained LSTM model, outputting a heat load prediction curve for the next 24 hours, covering the time range of the original prediction results. For example, if the initial forecast covers the period from 8:00 AM on the current day to 8:00 AM the next day, after hourly rolling updates, the update at 7:00 AM the next day will generate a new forecast curve covering the period from 7:00 AM the following day to 7:00 AM the day after, ensuring that the forecast is always based on the latest environmental and heating conditions. This mechanism effectively corrects forecast deviations caused by sudden heating demands (such as temporary events in commercial areas) or environmental changes (such as an early cold wave) by continuously incorporating real-time data, making the heat load curve dynamically match the actual heating trend.

[0020] As an optional embodiment of the present invention, optionally, generating a thermal inertia compensation strategy based on the heat load curve in step S3 and uploading it to the global layer includes: S301. Extract the thermal inertia characteristic parameters of each terminal user based on the heat load curve, analyze the load fluctuation pattern, and identify the key time period and fluctuation intensity that need to be compensated. In step S301, it is necessary to explain in detail that the extraction of thermal inertia characteristic parameters is based on the dynamic change characteristics of the heat load curve, which is achieved by analyzing the load fluctuation patterns of end users at different times. Specifically, the system first performs first-order difference processing on the heat load curve, calculates the load change rate at adjacent time points, and identifies typical fluctuation patterns by combining user type labels (such as office type, residential type). For example, residential users exhibit a low-load stable state from 22:00 to 6:00 the next day, while office users experience peak load from 9:00 to 17:00 on weekdays; commercial users experience significant fluctuations due to business activities on weekends or holidays. Through cluster analysis (such as the DBSCAN algorithm), the load fluctuation patterns are divided into three categories: "peak filling valley", "valley heating", and "stable maintenance", and the key time periods requiring compensation (such as 1 hour before the peak and 2 hours after the valley) and the fluctuation intensity (high / medium / low) are marked. Simultaneously, by combining physical parameters such as building heat capacity and heat transfer coefficient in the user-side thermal inertia parameter library, the delay time and attenuation coefficient of load adjustment are calculated, and a thermal inertia feature vector containing the compensation period, target load adjustment amount and duration is generated.

[0021] S302. Based on the thermal inertia characteristic parameters, the key time periods requiring compensation, the fluctuation intensity, and the user type label, generate a thermal inertia compensation strategy for time-segmented peak shaving, valley filling, and dynamic valve opening control.

[0022] In step S302, it is necessary to explain in detail the construction of a compensation strategy generation model based on thermal inertia characteristic parameters, key time periods requiring compensation, fluctuation intensity, and user type labels. Specifically, a multi-objective optimization algorithm (such as NSGA-II) is used to design the thermal inertia compensation strategy: First, differentiated compensation weights are set according to user type labels (such as residential, office, and commercial). For example, residential users focus on filling in off-peak hours (such as late night to early morning), with a compensation weight of 0.6; office users focus on peak-hour adjustment (such as weekdays 9:00-11:00), with a compensation weight of 0.7; and the weights for commercial users are dynamically adjusted based on calendar information. Secondly, the compensation range is calculated based on the intensity of fluctuations (high / medium / low): For periods of high fluctuation (e.g., fluctuation rate exceeding 20%), a proportional-integral (PI) control algorithm (e.g., setting the proportional coefficient Kp=0.8 and the integral coefficient Ki=0.2) is used to generate valve opening adjustment values; for periods of medium to low fluctuation, a sliding time window is used to predict load trends based on the delay time and attenuation coefficient in the thermal inertia characteristic vector, and the valve opening is dynamically corrected (e.g., updated every 15 minutes). Furthermore, the time-segmented peak-shaving strategy limits the upper limit of valve opening (e.g., setting an opening threshold of 80%) through intelligent temperature control valves during peak load periods (e.g., 7:00-9:00) and adds thermal inertia compensation (e.g., increasing the load buffer by 10%); the off-peak filling strategy increases the lower limit of valve opening (e.g., minimum opening of 20%) during off-peak periods (e.g., 22:00-6:00) and calculates the maximum allowable heating rate based on user heat capacity parameters to ensure comfort constraints (e.g., room temperature change does not exceed ±1℃ / hour). Finally, a strategy matrix containing timestamps, target valve opening degrees, compensation duration, and priority indicators is generated and uploaded to the global layer in JSON format. This strategy is iteratively optimized through a real-time feedback mechanism (such as monitoring valve status and actual room temperature) to ensure minimal fluctuations in the heating network and improved energy efficiency.

[0023] As an optional embodiment of the present invention, optionally, in step S4, a digital twin heating network model is constructed at the global layer based on GIS geographic information and heating network topology data, and the parameters of the heat load prediction model for each regional layer are optimized using a federated learning algorithm, including: S401. The global layer constructs a digital twin heating network model based on GIS geographic information, heating network topology data, and heat load prediction model parameters uploaded from each regional layer, and establishes a physical-digital parameter mapping relationship. In step S401, it is necessary to explain in detail that the global layer constructs a digital twin heating network model and establishes a physical-digital parameter mapping relationship based on GIS geographic information, heating network topology data, and heat load prediction model parameters uploaded from each regional layer. Specifically, the system first imports geographic information data from the GIS database, including the altitude, building distribution density, and land cover type of the heating network coverage area. Combined with the pipeline layout (such as main pipeline diameter, branch pipeline length, and valve location coordinates) and heat source nodes (such as thermal power plant coordinates and heat exchange station distribution) in the heating network topology data, a three-dimensional geometric model framework of the heating network is constructed. The physical-digital parameter mapping relationship is achieved by defining the mathematical relationship between key physical variables (such as pipeline thermal conductivity coefficient, fluid specific heat capacity, and insulation thermal resistance) and equivalent parameters in the digital model (such as heat loss coefficient and flow resistance factor). For example, a parameter calibration algorithm based on thermodynamic equations (such as Fourier's law of heat conduction) is used, and least squares fitting is performed using historical operating data (such as temperature sensor records and flow meter readings) to ensure that the model's dynamic response is consistent with the actual physical process. Furthermore, combining the heat load prediction model parameters uploaded from each regional layer (such as LSTM network weights and bias terms), the global layer initializes the input boundary conditions of the digital twin model (such as the initial heat load distribution) through the federated learning server and establishes a parameter mapping table to map regional layer features (such as user type labels and time series features) to the spatial heat distribution variables of the global model (such as heat source power and pipe temperature gradient). This process achieves dynamic updates of model parameters through automated scripts (such as Python API calls), ensuring that the digital twin can simulate the thermal behavior of the heating network under multi-regional collaboration in real time, providing a high-fidelity simulation environment for subsequent federated learning optimization. Finally, the mapping relationship library is stored in the global layer database, supporting the visual monitoring of the heating network status (such as heat map rendering) and the precise issuance of control commands.

[0024] S402. Construct a physical simulation kernel for a digital twin heat network based on the heat transfer equation and fluid dynamics equation. Verify the accuracy of the digital twin heat network model using historical measured data and trigger a parameter correction mechanism to meet the preset error threshold. In step S402, it is necessary to explain in detail that, based on the heat transfer equation and fluid dynamics equation, the physical simulation kernel of the digital twin heating network transforms the continuous thermodynamic and fluid processes into a computable numerical model through discretization methods (such as the finite difference method or the finite element method). Specifically, the heat transfer equation (such as the one-dimensional unsteady-state heat conduction equation) is used to simulate the heat exchange between the inner wall of the pipe and the surrounding environment, combined with fluid dynamics equations (such as the continuity equation and momentum equation) to describe the pressure distribution and velocity changes of water flow in the pipe. In the simulation kernel, the pipe is divided into multiple control volumes, and the dynamic changes of temperature, pressure, and flow rate are calculated within each control volume. The thermodynamic processes of each region are coupled through boundary conditions (such as the inlet temperature of the heat source and the flow rate demand at the user end). To verify the accuracy of the model, the system compares the simulation results with historical measured data (such as hourly pipe temperatures recorded by temperature sensors and branch pipe flow rates measured by flow meters) and calculates the root mean square error (RMSE) and mean absolute error (MAE). If the error exceeds a preset threshold (e.g., RMSE > 5℃ or MAE > 3%), a parameter correction mechanism is triggered: by adjusting physical parameters (e.g., pipe thermal resistance, fluid viscosity) or digital model parameters (e.g., heat loss coefficient calibration value), the parameter combination is optimized using gradient descent or genetic algorithms until the error between the model output and the measured data meets the threshold requirements. Furthermore, the simulation kernel integrates an anomaly detection module to preprocess missing values ​​or outliers in the measured data (e.g., abnormal temperature jumps caused by sensor malfunctions) using linear interpolation or DBSCAN algorithm removal, ensuring the reliability of the verification process. Finally, the verified digital twin model can accurately simulate the thermal response of the heating network under multiple operating conditions (e.g., extreme weather, sudden heating demand), providing a high-fidelity simulation foundation for federated learning optimization.

[0025] S403. Apply federated learning algorithm, set up federated server at global layer to coordinate distributed model training of nodes in each regional layer, including parameter initialization, local update and encrypted upload. In step S403, it is necessary to explain in detail that when applying the federated learning algorithm, the federated server set up in the global layer acts as the core coordination unit. First, it completes parameter initialization: assigning unified initial LSTM model parameters (such as weight matrices and bias vectors) to each regional layer node, and defining a global optimization objective (such as minimizing the joint loss function of heat load predictions for all regions). During the local update phase, each regional layer node independently trains its model based on local data (such as user heat usage records and meteorological data for that region), updates parameters through backpropagation, and simultaneously encrypts gradient information using differential privacy techniques (such as adding Gaussian noise) to prevent the leakage of sensitive data (such as user heat usage patterns). After the encrypted gradients are uploaded to the federated server, the server uses a secure aggregation protocol (such as a weighted average based on homomorphic encryption) to fuse the gradients from each region, generating the global model parameter update, and then distributes it to each node for the next iteration. During this process, the federated server ensures that model training takes into account the characteristics of both data-scarce and data-rich regions through a dynamic weight adjustment mechanism (such as allocating aggregation weights based on the amount of regional data or model contribution). Furthermore, to address the differences in data distribution across regions (such as different heating patterns in different climate zones), the server introduces a multi-task learning framework. This framework allows nodes to share underlying features (such as user type encoding) while retaining region-specific parameters (such as the weights of compensation strategies during cold waves), thereby improving the generalization ability of the global model. Finally, after multiple iterations (such as 50-100 iterations), the global model converges to a stable state, and each regional layer node obtains optimized personalized model parameters, achieving cross-regional knowledge transfer and collaborative optimization.

[0026] S404. Integrate the heat load prediction model parameters uploaded from each regional layer through the model aggregation mechanism, perform global optimization calculation, and generate the fused model weights. In step S404, it is necessary to explain in detail that when integrating the heat load prediction model parameters uploaded from each regional layer through the model aggregation mechanism, the federated server first decrypts and preprocesses the encrypted gradients or model parameters uploaded by each regional node, including parameter alignment (such as unifying the LSTM network weight matrices of different regional nodes to the same dimension) and outlier detection (such as removing missing parameters or extreme values ​​caused by communication failures). Subsequently, a weighted aggregation strategy is used to generate global model weights: the aggregation weights are allocated according to the data scale of each region (such as the number of users and the duration of historical data) and model performance (such as the prediction accuracy on the regional layer validation set). For example, regions with larger data volumes and lower prediction errors are given higher weights (such as weight coefficients of 0.3-0.5), while regions with scarce data or higher errors are given lower weights (such as 0.1-0.2). During the aggregation process, the server dynamically adjusts the weight allocation (such as recalculating the weights based on the model convergence speed after each iteration) to ensure that the global model takes into account the characteristics of different regions. Furthermore, to address the disparities in data distribution across regions (such as different heating patterns in northern and southern climate zones), an attention-based aggregation method is introduced. This method automatically adjusts aggregation weights by calculating the contribution of each region's parameters to the global objective (e.g., minimizing the joint loss function). For example, parameters from regions that show significant improvements in prediction accuracy during cold waves are given higher attention. After aggregation, the server performs global optimization calculations: using stochastic gradient descent (SGD) or adaptive moment estimation (Adam) algorithms, the model weights are updated based on the global training set (e.g., historical data from all regions) until the loss function converges to a preset threshold (e.g., a loss decrease of <0.1% over 10 consecutive iterations). The final generated global model weights are distributed to each regional layer node via an encrypted channel as initial parameters for the next round of local training, forming a closed-loop optimization process of "local training - encrypted upload - global aggregation - parameter distribution." This mechanism significantly improves the model's adaptability to complex operating conditions (e.g., cross-climate zone heating network coordination) by continuously integrating knowledge from multiple regions.

[0027] S405. Optimize the digital twin heating network model based on the fused model weights; In step S405, it is necessary to explain in detail the optimization of the digital twin heating network model based on the fused model weights. Specifically, the system injects the global model weights (such as the LSTM network weight matrix and bias vector) generated by federated learning into the prediction module of the digital twin heating network model, replacing the original initial parameters of the regional layer. The optimization process includes the following steps: First, the heat load prediction engine of the digital twin model is initialized using the fused weights. By inputting real-time meteorological data (such as temperature, wind speed, and solar radiation intensity) and regional user type distribution, a high-precision prediction of the spatial distribution of heat load for future time periods (such as the next 24 hours) is generated. Second, the physical simulation kernel of the digital twin model is driven based on the prediction results: the predicted heat load is substituted as a boundary condition into the heat transfer equation and the fluid dynamics equation to dynamically simulate the changing trends of the temperature field, pressure field, and flow distribution of the entire network pipeline. Furthermore, the system uses a parameter mapping table to transform the regional collaborative knowledge (such as cross-regional load complementarity patterns) contained in the fused weights into key parameter calibration quantities for the physical model (such as adjusting the equivalent heat loss coefficient of pipelines in different climate zones), thereby improving the simulation kernel's response accuracy to complex collaborative control scenarios (such as multi-heat source linkage and cross-regional load transfer). Simultaneously, the optimized model supports a closed-loop iteration of "prediction-simulation-verification": the simulation output (such as predicted temperatures at key nodes) is compared with actual sensor data in real time. If the deviation exceeds a threshold, a weight fine-tuning mechanism is triggered—based on gradient backpropagation, the adaptive parameters of the fused weights under specific operating conditions (such as sudden cold waves) are corrected (such as the neuron weights in the LSTM network that are sensitive to sudden temperature drops). Finally, the optimized digital twin heat network model can generate heat network control commands that balance global efficiency and regional characteristics (such as heat source output allocation and pipeline valve coordination strategies), and output high-fidelity thermal state simulation results through a visualization platform, providing decision-making basis for the virtual power plant central dispatch system.

[0028] S406. The optimized digital twin heating network model parameters are transmitted to each regional layer to synchronously optimize the heat load prediction model parameters.

[0029] In step S406, it is necessary to explain in detail that the optimized digital twin heating network model parameters are transmitted to each regional layer to synchronously optimize the heat load prediction model parameters. Specifically, the system transmits the fused and optimized global model parameters (including LSTM network weights, bias terms, and attention mechanism weights) to each regional layer server through an encrypted communication protocol (such as TLS 1.3). After receiving the parameters, the regional layer performs local model synchronization: First, it uses the received global parameters to cover the basic architecture parameters of the local heat load prediction model (such as the recurrent layer weight matrix), while retaining the regional-specific parameter layers (such as the weights of the fully connected layer for local climate characteristics); second, it performs fine-tuning training based on local real-time data (such as the latest collected room temperature and valve opening records), and adapts to local dynamic conditions by limiting the gradient update range of backpropagation (such as adjusting only the top 2-3 neural network layers) while inheriting global knowledge. A version control mechanism is introduced during the synchronization optimization process, and a timestamp and global iteration round identifier are added to the transmitted parameters to ensure the consistency of model versions in each regional layer. Ultimately, the updated regional layer model output (such as the load forecast for the next hour) will serve as the input for the next round of federated learning, forming a dynamic optimization closed loop of "global optimization - parameter propagation - local fine-tuning - feedback upload". This mechanism enables the regional layer model to respond quickly to scenarios such as sudden cold waves and abrupt changes in user behavior through regular parameter synchronization (such as issuing incremental updates every 15 minutes), while maintaining cross-regional collaborative capabilities (such as including load complementarity patterns between adjacent regions in the issued parameters), significantly improving the robustness of network-wide regulation.

[0030] As an optional embodiment of the present invention, the expression of the heat transfer equation may be: ; ; in, Represents the heat flux density vector. Indicates the thermal conductivity of a material. Represents the temperature gradient. Indicates fluid density, This represents the specific heat capacity of a fluid at constant pressure. Represents the fluid velocity field. Indicates the heat source term; The expression for the fluid dynamics equation is: ; ; in, Indicates fluid pressure. Indicates dynamic viscosity. Represents gravitational acceleration. It represents an external force.

[0031] As an optional embodiment of the present invention, the expression of the federated learning algorithm is optionally: Region layer: , ; Global layer: , ; in, Represents the region layer The loss function value, Represents the region layer The model parameter vector, Represents the region layer The number of local samples, Represents the single-sample loss function. Represents the region layer The prediction model function, Represents the region layer The Each sample feature vector Represents the region layer The The true labels of each sample Represents the L2 regularization coefficient. Denotes the square of the L2 norm. Represents the region layer gradient vector, Represents the gradient operator, Represents the global model parameter vector. Indicates the total number of regional layers. Represents the region layer Sample weights, Represents the total number of samples. Represents the region layer In the Model parameters after training round Indicates the learning rate. Represents the region layer In the The gradient is calculated in rounds.

[0032] As an optional embodiment of the present invention, optionally, in step S5, based on the optimized heat load prediction model parameters and the digital twin heating network model, a global heat load control command is generated and issued to the regional layer, including: S501. Based on the optimized heat load prediction model parameters, input real-time environmental variables and heating network operation status data, perform heat load prediction simulation, and generate a heating network load distribution map for the next 24 hours. In step S501, it is necessary to explain in detail that the system first collects environmental variables (including air temperature, humidity, wind speed, and solar radiation intensity) and heating network operation status data (such as pipe inlet / outlet water temperature, valve opening, and pump speed) in real time from a distributed sensor network (such as temperature, pressure, and flow sensors deployed at heating stations). After preprocessing by a data cleaning module (such as outlier filtering based on Z-score), the data is input into the optimized heat load prediction model. Specifically, the model uses fusion weights (such as LSTM network parameters generated by federated learning) to perform spatiotemporal encoding on the input features: real-time data and historical data of the same period (such as records of the same time period in the past week) are concatenated into a time series tensor. Features are extracted layer by layer through a multi-layer recurrent neural network (such as a bidirectional LSTM), and an attention mechanism (such as calculating the contribution weight of different time steps to the prediction target) is applied to focus on key influencing factors (such as sudden temperature drops during cold waves).

[0033] The predictive simulation execution process includes the following steps: 1. Time-step iterative prediction: The heat load value for the next 24 hours (96 time points in total) is predicted sequentially at 15-minute intervals. The prediction for each time step is based on the current input and the previous hidden state, and outputs the instantaneous heat load demand of each regional node (such as a community heating station).

[0034] 2. Spatial Distribution Mapping: The prediction results are combined with Geographic Information System (GIS) topology data (such as pipeline layout and user density distribution) to generate a spatial distribution map of the heat load across the entire heating network using the Kriging interpolation algorithm. This distribution map is marked with key indicators (such as heat load intensity kW / m², peak load location) and the demand differences in different areas are visualized through color gradients (such as high load areas marked in red and low load areas marked in blue).

[0035] 3. Confidence assessment: Based on the model's historical performance on the validation set (such as the distribution of prediction errors), a confidence interval (such as ±5% error range at a 95% confidence level) is added to each prediction point and displayed as a semi-transparent color band on the distribution plot to improve the reliability of decision-making.

[0036] The final generated load distribution map is output to the virtual power plant central dispatch platform, serving as the core input for optimizing heating network control commands.

[0037] S502. Combining a digital twin heating network model, simulate the dynamic response process of the heating network, calculate the heat transfer efficiency and heat loss coefficient, and evaluate the feasibility and robustness of the control strategy. In step S502, it is necessary to explain in detail that the system inputs the heat load distribution map for the next 24 hours generated in step S501 into the digital twin heating network model, driving its physical simulation kernel to perform dynamic response simulation. The specific process includes: First, based on the predicted heat load as boundary conditions, the system substitutes the heat transfer equation and fluid dynamics equation, and uses a numerical solver (such as the finite volume method) to calculate the temperature field, pressure field, and flow distribution changes of the entire network pipeline in real time. During the simulation, the system dynamically tracks the parameter changes of key nodes (such as heat source outlets and pipeline junctions), generating heat transfer efficiency curves (such as the change in heat loss rate per unit length of pipeline over time) and heat loss coefficient distribution maps (such as the equivalent thermal resistance values ​​of different pipe diameters and insulation layers). Second, a control strategy evaluation module is introduced: preset control commands (such as adjusting heat source output and switching pipeline valve states) are substituted into the digital twin heating network model to calculate their impact on heating network operating indicators (such as supply and return water temperature difference and circulating pump energy consumption), and the Monte Carlo method is used to simulate the impact of uncertainties (such as user-side heat fluctuations and equipment failures) on the robustness of the strategy. Furthermore, the system generates a multi-dimensional evaluation report based on simulation results, including the improvement rate of heat transfer efficiency (e.g., a 15% reduction in heat loss after regulation), the pressure fluctuation range of key nodes (e.g., within ±0.2 MPa), and the failure probability of the strategy under extreme conditions (e.g., a 30% surge in heat load during a cold wave) (e.g., <5%). Finally, the evaluation report displays a three-dimensional thermodynamic cloud map of the heating network's dynamic response (e.g., the change in pipeline temperature over time and space) and the sensitivity analysis results of the regulation strategy (e.g., the weight of the impact of valve opening adjustment on heating quality) through a visualization interface, providing quantitative decision-making basis for the central dispatch platform.

[0038] The method for calculating heat transfer efficiency and heat loss coefficient is as follows: First, for the calculation of heat transfer efficiency, the system adopts the heat flow balance method. Specifically, the total heat supply at the heat source outlet is obtained in real time through a digital twin heat network model, while the total heat return at each user's return water inlet is monitored. The heat transfer efficiency (η) is defined as the ratio of the effective heat actually absorbed by the user to the total heat provided by the heat source, i.e., η = (total heat return / total heat supply) × 100%. To improve calculation accuracy, the model introduces a temperature compensation mechanism during the data acquisition phase: the supply and return water temperature difference measured by the sensor is corrected according to the real-time ambient temperature (such as outdoor air temperature) to eliminate the interference of environmental heat loss on the measurement results. For example, when the outdoor air temperature is below 5℃, the system automatically increases the measured supply and return water temperature difference by 3% to compensate for the heat loss from the pipeline to the environment.

[0039] Secondly, the calculation of the heat loss coefficient combines pipeline physical parameters with simulation data. The system extracts pipe diameter, pipe length D, insulation layer thickness, and material thermal conductivity (δ) from the pipeline attribute library of the digital twin model. Simultaneously, it obtains the difference between the pipe surface temperature and the ambient temperature through simulation (ΔT = pipe surface temperature - ambient temperature). The heat loss coefficient is initially estimated using the empirical formula K = λ / (2δ) × ln(D / D0) (where D0 is the pipe inner diameter), and then back-calibrated through simulation: the model-predicted heat loss is compared with the actual monitored heat loss. If the deviation exceeds a threshold (e.g., ±5%), the K value is adjusted until the simulation and measured data match. For example, when the simulation shows that the heat loss of a certain pipe section is 8% higher than the measured value, the system iteratively corrects the K value by reducing it by 6.2% (based on the inverse relationship between thermal resistance and K value).

[0040] S503. Based on the evaluation results, apply a multi-objective optimization algorithm to solve for the optimal combination of control parameters with the objectives of minimizing the fluctuation range of the heating network, balancing the supply and demand of heat sources and reducing energy costs. In step S503, it should be explained in detail that the expression of the multi-objective optimization algorithm in this embodiment is as follows:

[0041] in, This represents a vector of decision variables, including control parameters such as valve opening and heat source output power. This represents the objective function vector, which contains three sub-objectives: minimizing the fluctuation amplitude of the heating network. Heat source supply and demand balance target and the goal of minimizing energy costs , This represents an inequality constraint, such as the upper limit of the output power of a heat source. Valve opening range Safe threshold for heating network temperature , Indicates the heat source. This indicates equality constraints, such as the requirement that the heat transfer efficiency of the heating network must meet a preset value, and the heat loss coefficient must conform to historical measured data.

[0042] It should also be noted that the multi-objective optimization algorithm system described above employs an optimization engine based on the Non-Dominated Sorting Genetic Algorithm (NSGA-II) to iteratively search the decision variable vector and generate a Pareto front solution set. The specific solution process includes: initializing the population (e.g., randomly generating 100 sets of control parameter combinations), calculating the objective function value of each solution (including the fluctuation range of the heating network, the supply and demand deviation of the heat source, and energy costs), applying fast non-dominated sorting and congestion distance calculation to select the optimal individual, updating the population through simulated binary crossover and polynomial mutation operations, and verifying inequality constraints (e.g., the output power of the heat source does not exceed the rated upper limit) and equality constraints (e.g., the heat transfer efficiency must reach a preset threshold of more than 85%) in each iteration. Finally, the optimal solution set is output for the scheduling platform to make decisions. This algorithm ensures that the control strategy reduces the peak load fluctuation of the heating network by 10% while optimizing the energy cost to below 92% of the historical benchmark, and automatically filters invalid solutions that violate the safety threshold, thereby improving the overall robustness of the system.

[0043] S504. Generate a global heat load control instruction set based on the optimal control parameter combination, including valve opening command, heat source output power adjustment and user-side compensation strategy. S505: Through an encrypted communication protocol, global thermal load control commands are sent to execution nodes at each regional layer.

[0044] In steps S504 and S505, it is necessary to explain in detail that the system, based on the optimal combination of control parameters output by the multi-objective optimization algorithm, refines the control instructions into a three-layer execution structure: The first layer is the heat source layer instruction, which includes the output power adjustment value of each heat source (such as increasing the output of heat source No. 1 from 50MW to 55MW) and the start / stop status (such as starting the standby heat source No. 2). The instruction is directly sent to the heat source PLC control system via the OPCUA protocol; the second layer is the pipeline layer instruction, which covers the adjustment amount of the opening of key valves (such as adjusting the opening of the main pipeline). The instructions for adjusting the opening of pipe valve V-103 from 60% to 75% and regulating the speed of circulating pump (such as reducing the speed of pump P-201 from 1450rpm to 1380rpm) are transmitted to the pipeline monitoring system via the Modbus TCP protocol. The third layer consists of user-side compensation instructions, including dynamic electricity price signals (such as a 20% increase in electricity price during peak hours) and interruptible load incentives (such as a subsidy of 0.5 yuan per kilowatt-hour for users participating in demand response). These instructions are pushed to end users via smart meters or user terminal APPs.

[0045] To ensure the accuracy of command execution, the system embeds a verification mechanism in the command set: a hash checksum (such as a 64-bit hexadecimal code generated by the SHA-256 algorithm) is added to the heat source output power command; a dual confirmation process is set for the valve opening command (requiring simultaneous feedback confirmation signals from both the area controller and the field actuator); and user-side compensation commands are timestamped and encrypted (e.g., embedding the command generation time into the RSA public key encrypted data packet). Simultaneously, the command set includes a fault-tolerant processing module: when a heat source fails to perform power adjustment, the system automatically triggers a backup strategy (e.g., allocating the unexecuted 5MW load to a nearby heat source) and recalculates the heat network balance state using a digital twin model; when the valve actuator reports an opening deviation exceeding a threshold (e.g., ±3%), the system immediately generates a correction command (e.g., a secondary adjustment of the opening to 78%) and records the fault log. The final generated global heat load control instruction set is encapsulated in XML format, including instruction type (such as heat source adjustment / valve control / user incentive), execution time (such as 2023-11-15T14:00:00+08:00), target device ID (such as heat source HS-001), parameter value (such as power 55MW), and digital signature (such as a 256-bit signature value based on the ECC algorithm). It is distributed to the regional layer execution system through the message middleware (such as Apache Kafka) of the virtual power plant central dispatch platform.

[0046] As an optional embodiment of the present invention, optionally, after the regional layer receives the global heat load control command in step S6, it dynamically adjusts the heating parameters of each terminal user by combining the local heat source supply capacity and the user-side thermal inertia compensation strategy, including: S601. Parse the received global heat load control command and extract the specific parameter set including valve opening command, heat source output power adjustment value and user-side compensation strategy; In step S601, it is necessary to explain in detail that the regional layer execution system receives XML format instruction packets from the central dispatch platform through a secure communication interface. First, it performs integrity verification: using a pre-set ECC public key to verify the digital signature, ensuring the instruction has not been tampered with; then, it parses the timestamp field to confirm the instruction's timeliness (e.g., the difference between the execution time and the current system time is ≤5 minutes). The instruction parsing module adopts a layered decoding mechanism: the first layer extracts heat source layer parameters (e.g., heat source ID is HS-001, output power 55MW) and writes them to the local heat source control system via the OPCUA client; the second layer parses pipeline layer instructions (e.g., valve V-103 target opening degree 75%) and forwards them to the regional valve controller via the ModbusTCP protocol; the third layer processes user-side compensation strategies (e.g., electricity price increase of 20%, subsidy of 0.5 yuan / kWh) and pushes them to smart meters or user apps via the MQTT protocol. The system records instruction processing logs, including key information such as reception time, parsing results, and equipment response status, and uploads them to the central dispatch platform via an encrypted channel (e.g., AES-256) for post-event auditing. Finally, the parsed parameter set is stored in the regional layer database.

[0047] S602. Based on local heat source supply capacity data and specific parameter sets, assess the real-time availability range of heat source output power and calculate the heat source supply-demand balance deviation. In step S602, the regional layer execution system first obtains real-time supply capacity data from the local heat source monitoring module, including the current output of each heat source (e.g., actual output of heat source No. 1 is 48MW), rated power (e.g., rated power of heat source No. 1 is 60MW), remaining adjustable capacity (e.g., remaining capacity of heat source No. 1 is 12MW), and equipment operating status (e.g., normal operation / fault shutdown). Combined with the target power adjustment value in the central dispatch instruction (e.g., requiring heat source No. 1 to be increased to 55MW), the system performs a two-dimensional verification through the supply and demand balance calculation module: First, based on the remaining capacity of the heat sources, the feasibility of the instruction is verified. When the target power exceeds the remaining adjustable capacity (e.g., 55MW > 12MW + 48MW), a capacity shortage warning is triggered and a correction suggestion is generated (e.g., a suggestion to adjust to 52MW); Second, based on the heat network topology, the rationality of power allocation is verified. The adjusted heat flow distribution is simulated through a digital twin regional sub-model, and the pressure changes at key nodes (e.g., the expected pressure increase at the pipeline junction is 0.15MPa) are calculated to determine whether they exceed the safety threshold (e.g., ±0.2MPa). The system further calculates the supply-demand balance deviation of the heat sources using a dynamic weighting algorithm: the difference between the target power and the current output (e.g., 55MW-48MW=7MW) is used as the main deviation term, and normalized by combining the actual supply-demand fluctuation rate under historical operating conditions (e.g., ±3MW) to generate a balance deviation index of 0-100% (e.g., current deviation index = 7MW / 10MW×100%=70%, where 10MW is an empirical threshold). If the deviation index exceeds the warning value (e.g., >60%), the system activates the multi-heat source collaborative optimization module, which redistributes the load through a linear programming algorithm (e.g., transferring 3MW of load to the adjacent No. 2 heat source) to ensure balanced load rates of each heat source (e.g., target load rate difference ≤15%). Finally, the system generates an evaluation report including the real-time available range (e.g., the adjustable range of No. 1 heat source is 50-52MW), suggested correction value (e.g., 52MW), collaborative optimization scheme (e.g., adding 3MW to No. 2 heat source), and balance deviation index (e.g., 70%), which is fed back to the central dispatch platform as a basis for secondary verification of the command.

[0048] S603. Based on the heat source supply and demand balance deviation and the user-side thermal inertia compensation strategy, apply time-sharing peak shaving and valley filling algorithms to determine the dynamic valve opening adjustment value for each terminal user. In step S603, the regional layer execution system, based on the heat source supply and demand balance deviation report and combined with the user-side thermal inertia compensation strategy database, initiates the time-segmented peak shaving and valley filling algorithm module. This module first divides the entire day into three time periods: peak, flat, and valley (e.g., peak period 08:00-20:00, flat period 06:00-08:00 and 20:00-22:00, valley period 22:00-06:00 the next day). Differentiated control strategies are adopted for the characteristics of different time periods: during peak periods, priority is given to reducing the heat source load, and a linear decreasing algorithm (e.g., reducing valve opening by 2% every 15 minutes) is used to gradually adjust the valve opening in high-load areas from the current value (e.g., 75%) to the target value. For example, at 65%, interruptible load incentives are triggered on the user side (e.g., double subsidies are given to users participating in peak shaving); during off-peak hours, the valley filling algorithm is activated. Based on the remaining capacity of the heat source (e.g., 8MW remaining for heat source No. 2) and the user-side heat storage capacity (e.g., the capacity of thermal storage electric heaters), the optimal valve opening increase is calculated in 5-minute steps using a dynamic programming algorithm (e.g., increasing the valve opening in low-load areas from 50% to 58%), and a dynamic electricity price discount signal is pushed (e.g., a 30% reduction in off-peak electricity prices). The core of the algorithm adopts a thermal inertia compensation model: by analyzing users' historical heat consumption data (e.g., the average hourly heat consumption over the past 7 days), a heat demand prediction submodule is established. Combined with current indoor temperature sensor data (e.g., 22℃±1℃), the user-side thermal inertia delay time is calculated (e.g., the time difference from valve adjustment to indoor temperature change is 30 minutes), and the valve opening adjustment timing is corrected accordingly (e.g., adjustment is performed 15 minutes in advance). The system further introduces a constraint handling mechanism: when valve opening adjustment causes the temperature difference between the supply and return water on the user side to exceed the safety threshold (e.g., ΔT>15℃), a protective callback is automatically triggered (e.g., reducing the opening adjustment range from 8% to 5%), and the thermal flow stability after adjustment is verified through a digital twin regional sub-model; when synchronous adjustment of valves by multiple users causes hydraulic imbalance in the pipeline network, the pressure balance compensation method is activated to calculate the pressure compensation value of key nodes (e.g., 0.05MPa pressure needs to be added at the pipeline junction), and auxiliary valve adjustment instructions are generated (e.g., fine-tuning the opening of compensation valve V-201 by 2%). Ultimately, the system outputs a dynamic peak-shaving scheme that includes time period divisions, valve opening adjustment sequences (e.g., 73% opening from 08:00 to 08:15, 71% opening from 08:15 to 08:30), user-side incentive strategies (e.g., a peak-shaving subsidy of 1 yuan / kWh), and constraint processing records. This scheme is then distributed to each user's valve actuator via the regional control terminal. Simultaneously, the execution results (e.g., actual opening of 66%, indoor temperature of 21.5℃) are fed back to the central dispatch platform in real time for strategy optimization.

[0049] S604. Based on the valve opening adjustment value, an adjustment command is issued to the heating control equipment of each terminal user to perform dynamic adjustment of heating parameters in real time, including temperature setpoint and flow control. In step S604, the regional layer execution system generates refined adjustment instructions for each end user based on the valve opening adjustment values ​​determined in the dynamic peak shaving scheme. The instructions include the target valve opening, execution time window (e.g., 08:00-08:15), associated device identifier, and dynamic compensation parameters (e.g., thermal inertia delay time of 30 minutes). Instruction delivery employs a multi-channel transmission mechanism: for users with intelligent control terminals, encrypted instruction packets (containing AES-256 encrypted plaintext instructions and a timestamp signature) are directly pushed via the MQTT protocol; for users with traditional mechanical valves, a dual notification method of SMS + voice call is used, with the instruction content converted into plain language instructions after natural language processing (e.g., "Please rotate the heating valve clockwise by 1 / 4 turn at 8 o'clock sharp"). The device layer is configured with a dual-mode control interface: the intelligent valve actuator automatically completes the opening calibration (e.g., using a stepper motor drive to achieve ±1% accuracy control) by parsing XML format instruction packets, and simultaneously provides feedback on the execution status (e.g., current opening 66.2%, motor speed 15 rpm); mechanical valve users can obtain operation instructions by scanning a QR code via a mobile APP, and upload execution photos (e.g., a photo of the valve dial) to the system for visual verification (accuracy rate up to 98.7%). Temperature setpoint adjustment adopts a closed-loop control strategy: the system predicts the indoor temperature change trend (e.g., expected to drop to 21℃ in 30 minutes) based on the valve opening change (e.g., from 75% to 66%) and the building thermal model (e.g., wall thermal resistance 0.5 m²·K / W), and automatically generates a temperature compensation command (e.g., temporarily increasing the thermostat setpoint from 22℃ to 22.5℃), which is synchronized to the user's thermostat via the Zigbee protocol. Flow control is implemented in stages: Main pipe flow is monitored in real time by an ultrasonic flow meter (sampling frequency 1Hz). When the detected flow deviation exceeds the threshold (e.g., ±5%), the system activates a PID control algorithm (proportional coefficient 0.8, integral time 2min, derivative time 0.5min) to generate valve fine-tuning commands (e.g., adding an opening adjustment of 1.5%). Branch flow uses predictive control, pre-adjusting valve openings 2 hours in advance based on historical heating patterns (e.g., weekday / weekend differences) (e.g., pre-opening branch valves by 5% on weekend mornings). All adjustment commands include safety verification fields: physical limits are set for valve opening commands (e.g., 0%-100% mechanical stop), comfort range constraints are implemented for temperature setpoints (e.g., 18℃-24℃), and pipeline pressure verification is added to flow commands (e.g., not exceeding 1.6MPa design pressure). During execution, the system continuously collects feedback data (such as actual valve opening, indoor temperature, and instantaneous flow rate), eliminates measurement noise (such as ±0.5℃ error of temperature sensor) through Kalman filtering algorithm, generates an execution effect evaluation report (including indicators such as adjustment delay time, overshoot, and steady-state error), and uploads it to the central dispatch platform every 15 minutes for strategy iteration and optimization.When an abnormal execution situation is detected (such as valve jamming or temperature change), the system immediately triggers a three-level response mechanism: Level 1 response (within 5 seconds) suspends the execution of commands in the relevant area; Level 2 response (within 30 seconds) switches to the backup control channel (such as switching from MQTT to 4G); and Level 3 response (within 2 minutes) initiates the manual intervention process (such as notifying maintenance personnel to handle the situation on-site) to ensure the safe and stable operation of the heating system.

[0050] S605: Monitor the changes in heating parameters after adjustment and compare them with the predicted heat load curve to trigger a local feedback mechanism to optimize subsequent adjustment strategies.

[0051] In step S605, the regional layer execution system collects adjusted heating parameter data in real time through an IoT sensor network deployed on the end-user side. This data includes key indicators such as actual valve opening, indoor temperature, supply and return water temperature, and instantaneous flow rate. The sampling frequency reaches 1Hz or higher to ensure data timeliness. The collected data is preprocessed by edge computing nodes, using a sliding window filtering algorithm to eliminate impulse noise, and timestamp synchronization technology to ensure time alignment of multi-source data. The system dynamically compares the real-time data with the predicted heat load curve issued by the central dispatch platform and calculates the deviation index: the root mean square error (RMSE) is used to quantify the overall deviation degree, and an RMSE ≤ 5% is considered an acceptable deviation; at the same time, the maximum deviation rate is calculated (e.g., the actual flow rate exceeds the predicted value by 18% at a certain moment). When it exceeds the safety threshold (e.g., ±15%), a local feedback mechanism is triggered. The local feedback mechanism adopts a two-layer control structure: the primary feedback generates an instantaneous correction value through a proportional-integral (PI) controller. For example, when the indoor temperature is detected to be 0.8℃ lower than the set value for 30 consecutive minutes, the valve opening compensation value is automatically increased by 2%. The advanced feedback calls the digital twin regional sub-model to perform simulation and simulation of heat flow changes under different correction strategies (such as the impact of increasing the valve opening of branch 2 by 5% on the pipeline pressure). The optimal adjustment scheme is generated through a multi-objective optimization algorithm (weight coefficients: thermal comfort 0.6, energy consumption 0.3, pipeline safety 0.1). The system initiates differentiated response procedures based on the type of deviation: For deviations attributable to user behavior (e.g., a user's indoor temperature is too high but the valve opening is normal), personalized suggestions are pushed through the user's app (e.g., "It is recommended to lower the thermostat setting by 1°C to save 15% of energy consumption"); for deviations caused by equipment failure (e.g., valve actuator jamming leading to insufficient opening), a maintenance work order is immediately generated and the system switches to the backup control channel; for systemic deviations (e.g., the overall regional heat load is lower than the predicted value), the corrected heat load curve is transmitted back to the central dispatch platform via the OPCUA protocol, triggering global strategy iteration. The feedback mechanism also includes a learning function. The system stores the deviation data, correction strategies, and execution effects of each adjustment in a historical database, uses a Long Short-Term Memory (LSTM) network to build a prediction model, and continuously trains to optimize the prediction accuracy of subsequent adjustment strategies. The final feedback report includes deviation analysis charts, correction strategy details, execution effect evaluation, and model optimization suggestions, and is uploaded to the central dispatch platform hourly to provide decision support for the next round of heat load control.

[0052] As an optional embodiment of the present invention, the method may further include: S7. Construct a thermal inertial parameter library for building types; In step S7, the system constructs a thermal inertial parameter library for building types using multi-source data fusion technology, specifically including the following processes: First, raw data is collected from the urban building information platform, historical heating database, and IoT sensor network, covering building physical properties (such as wall thermal resistance coefficient 0.35-0.65m²·K / W, window-to-wall ratio 15%-35%), functional classification (such as residential, office, and industrial), and historical thermal response records (such as hourly indoor temperature change data for the past 5 heating seasons). Data acquisition is achieved through a standardized API interface, with a sampling frequency of 1 time / hour to ensure coverage of typical operating conditions (such as extreme low temperature days in winter). Second, key thermal inertial parameters are extracted using feature engineering algorithms (such as principal component analysis PCA), including the thermal delay time constant τ (calculated as τ=RC, where R is thermal resistance and C is heat capacity), the thermal attenuation coefficient β (such as 0.05-0.2), and steady-state thermal conductivity. Data processing employs a multivariate regression model (e.g., least squares fitting) to generate baseline parameter values ​​based on building classification (e.g., mean τ for residential buildings: 45 minutes ± 15 minutes; mean τ for industrial buildings: 25 minutes ± 10 minutes), and calculates confidence intervals. The parameter library uses a distributed database architecture, storing key-value pairs in JSON format, supporting real-time querying and updates. The update mechanism is based on incremental learning: when new sensor data is added (e.g., the actual τ for a building is 50 minutes), the system calls the existing random forest algorithm to retrain the model, ensuring parameter accuracy (e.g., prediction error ≤ ± 5 minutes). Simultaneously, the parameter library integrates a verification function, simulating the building's thermal response using digital twin technology (e.g., predicting the room temperature change curve by inputting τ = 40 minutes) and comparing it with measured data (e.g., RMSE ≤ 3%). If the deviation exceeds a threshold (e.g., > 5%), parameter correction is triggered.

[0053] S8. Deploy intelligent temperature control valves at end users. Users can set flexible temperature zones according to their needs. Based on the thermal inertia parameter library of the building type, the heating temperature is increased by the intelligent temperature control valve when the power grid is in a low period and decreased by the intelligent temperature control valve when the power grid is in a high period. This achieves coordinated optimization of user-side thermal inertia compensation and power grid peak shaving and valley filling.

[0054] In step S8, it is important to explain in detail that the flexible temperature band in this embodiment refers to an acceptable range of indoor temperature fluctuations (e.g., 18℃-22℃) that users can set on the intelligent temperature control valve according to their own comfort needs and heating habits. This temperature band is not a fixed value, but is dynamically adjusted in conjunction with key parameters such as the thermal delay time constant τ and the thermal attenuation coefficient β from the building type thermal inertia parameter library. For example, for residential buildings with high thermal inertia (τ average 45 minutes), users can set the flexible temperature band width to ±2℃, allowing the system to start the heating temperature increase 15 minutes earlier during off-peak hours (e.g., 22:00-06:00 the next day) (e.g., from 20℃ to 22℃), utilizing the building's thermal inertia to store heat; and during peak hours (e.g., 08:00-20:00), the heating temperature is reduced 15 minutes later (e.g., from 20℃ to 18℃), releasing the stored heat through thermal inertia to achieve peak shaving and valley filling collaborative optimization. The specific control logic is as follows: The microprocessor built into the intelligent thermostatic valve reads indoor temperature sensor data in real time. When it detects that the current temperature is close to the upper limit of the elastic temperature range (e.g., 21.8℃) and the power grid is in a peak period, it automatically triggers a valve opening reduction command (e.g., reducing the opening by 1% every 5 minutes). At the same time, it combines the thermal inertia compensation model to predict the indoor temperature drop trend (e.g., predicting that it will drop to 20℃ in 30 minutes) to ensure that the temperature fluctuation is within the acceptable range for the user. When it detects that the current temperature is close to the lower limit of the elastic temperature range (e.g., 18.2℃) and the power grid is in a low period, it automatically triggers a valve opening increase command (e.g., increasing the opening by 1% every 5 minutes) and predicts the indoor temperature rise trend (e.g., predicting that it will rise to 20℃ in 30 minutes). The system uses digital twin technology to simulate and verify the control strategy: Inputting building type parameters (e.g., residential, τ=45 minutes), current indoor temperature (18.5℃), power grid time period (off-peak), and flexible temperature band setting (18℃-22℃), the system simulates and outputs a valve opening adjustment sequence (e.g., 00:00 65%, 00:05 66%...00:30 70%) and an indoor temperature change curve to ensure control accuracy (e.g., room temperature deviation ≤ ±0.5℃). Users can view the flexible temperature band setting, current room temperature, valve opening, and power grid time period information in real time via a mobile app or the thermostatic valve panel, and manually adjust the flexible temperature band range (e.g., from ±2℃ to ±1.5℃). The system will recalculate the control strategy based on the new settings. Furthermore, the system supports time-based flexible temperature band settings: users can configure different flexible temperature bands for weekdays and weekends in the app (e.g., 18℃-21℃ on weekdays, 19℃-22℃ on weekends), and the system automatically matches the date type and executes the corresponding strategy.All control commands include safety constraints: when valve opening adjustments may cause the supply and return water temperature difference to exceed a safety threshold (e.g., ΔT>12℃), a protective callback is automatically triggered (e.g., reducing the opening adjustment range from 2% to 1%), and the thermal flow stability after adjustment is verified through a digital twin model; when the indoor temperature exceeds the elastic temperature zone for more than 10 minutes, the system pushes an alarm message to the user (e.g., "Current room temperature is 22.5℃, exceeding the upper limit of the elastic temperature zone by 0.5℃, it is recommended to adjust the settings or check the door and window seals"). Finally, the intelligent thermostatic valve uploads the execution records (e.g., valve opening adjustment time, room temperature changes, and power grid time period matching results) to the regional layer execution system for optimizing the building type thermal inertia parameter library and elastic temperature zone control strategy.

[0055] Example 2 A centralized heating load aggregation and control system under a virtual power plant framework, the system comprising: processor; Memory used to store processor-executable instructions; The processor is configured to implement a centralized heating load aggregation and control method under a virtual power plant framework when executing executable instructions.

[0056] It should be noted that the computer device includes a processor, a memory, and may also include one or more of a multimedia component, an input / output (I / O) interface, and a communication component.

[0057] The processor controls the overall operation of the computer device to complete all or part of the steps in the centralized heating load aggregation and control method under the virtual power plant framework.

[0058] Memory is used to store various types of data to support the operation of the computer device. This data may include, for example, instructions for any application or method used to operate on the computer device, as well as application-related data. Memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0059] The multimedia component may include a screen and an audio component, wherein the screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals; for example, the audio component may include a microphone for receiving external audio signals, the received audio signals may be further stored in memory or transmitted via a communication component; the audio component may also include at least one speaker for outputting audio signals.

[0060] I / O interfaces provide interfaces between the processor and other interface modules, such as keyboards, mice, buttons, etc.; these buttons can be virtual buttons or physical buttons.

[0061] The communication component is used for wired or wireless communication between the computer device and other devices; wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G or 5G, or one or more combinations thereof, and the corresponding communication component may include: Wi-Fi module, Bluetooth module, NFC module, mobile communication module.

[0062] As a preferred embodiment, the computer device may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the centralized heating load aggregation and control method under the virtual power plant framework described above.

[0063] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for centralized heating load aggregation and control under a virtual power plant framework, characterized in that, The method includes: S1. Collect user heat usage behavior data at the end user and upload it to the regional layer; S2. Construct a heat load prediction model at the regional layer, and predict the heat load curve for the next 24 hours based on the user heat consumption behavior data; S3. Generate a thermal inertia compensation strategy based on the heat load curve and upload it to the global layer; S4. At the global level, a digital twin heating network model is constructed based on GIS geographic information and heating network topology data, and the parameters of the heat load prediction model for each regional layer are optimized through federated learning algorithm. S5. Based on the optimized heat load prediction model parameters and the digital twin heating network model, generate global heat load control instructions and send them to the regional layer. S6. After receiving the global heat load control command, the regional layer dynamically adjusts the heating parameters of each terminal user by combining the local heat source supply capacity and the user-side thermal inertia compensation strategy.

2. The method for centralized heating load aggregation and control under a virtual power plant framework as described in claim 1, characterized in that, In step S2, predicting the heat load curve for the next 24 hours based on the user's heat consumption behavior data includes: S201. Preprocess the user's heat usage behavior data; S202. Construct a heat load prediction model based on the preprocessed user heat consumption behavior data; S203. Train and optimize the heat load prediction model; S204. Based on the trained heat load prediction model, a rolling prediction mechanism is executed. The input is updated based on the latest real-time data at preset intervals to dynamically generate the heat load curve for the next 24 hours.

3. The method for centralized heating load aggregation and control under a virtual power plant framework as described in claim 1, characterized in that, In step S3, generating a thermal inertia compensation strategy based on the heat load curve and uploading it to the global layer includes: S301. Extract the thermal inertia characteristic parameters of each terminal user based on the heat load curve, analyze the load fluctuation pattern, and identify the key time period and fluctuation intensity that need to be compensated. S302. Based on the thermal inertia characteristic parameters, the key time periods requiring compensation, the fluctuation intensity, and the user type label, generate a thermal inertia compensation strategy for time-segmented peak shaving, valley filling, and dynamic valve opening control.

4. The method for centralized heating load aggregation and control under a virtual power plant framework as described in claim 1, characterized in that, In step S4, a digital twin heating network model is constructed at the global layer based on GIS geographic information and heating network topology data, and the parameters of the heat load prediction model for each regional layer are optimized using a federated learning algorithm, including: S401. The global layer constructs a digital twin heating network model based on GIS geographic information, heating network topology data, and heat load prediction model parameters uploaded from each regional layer, and establishes a physical-digital parameter mapping relationship. S402. Construct a physical simulation kernel for a digital twin heat network based on the heat transfer equation and fluid dynamics equation. Verify the accuracy of the digital twin heat network model using historical measured data and trigger a parameter correction mechanism to meet the preset error threshold. S403. Apply federated learning algorithm, set up federated server at global layer to coordinate distributed model training of nodes in each regional layer, including parameter initialization, local update and encrypted upload. S404. Integrate the heat load prediction model parameters uploaded from each regional layer through the model aggregation mechanism, perform global optimization calculation, and generate the fused model weights. S405. Optimize the digital twin heating network model based on the fused model weights; S406. The optimized digital twin heating network model parameters are transmitted to each regional layer to synchronously optimize the heat load prediction model parameters.

5. The method for centralized heating load aggregation and control under a virtual power plant framework as described in claim 1, characterized in that, The expression for the heat transfer equation is: ； ； in, Represents the heat flux density vector. Indicates the thermal conductivity of a material. Represents the temperature gradient. Indicates fluid density, This represents the specific heat capacity of a fluid at constant pressure. Represents the fluid velocity field. Indicates the heat source term; The expression for the fluid dynamics equation is: ； ； in, Indicates fluid pressure. Indicates dynamic viscosity. Represents gravitational acceleration. It represents an external force.

6. The method for centralized heating load aggregation and control under a virtual power plant framework as described in claim 4, characterized in that, The expression for the federated learning algorithm is: Region layer: , ; Global layer: , ; in, Represents the region layer The loss function value, Represents the region layer The model parameter vector, Represents the region layer The number of local samples, Represents the single-sample loss function. Represents the region layer The prediction model function, Represents the region layer The Each sample feature vector Represents the region layer The The true labels of each sample Represents the L2 regularization coefficient. Denotes the square of the L2 norm. Represents the region layer gradient vector, Represents the gradient operator. Represents the global model parameter vector. Indicates the total number of regional layers. Represents the region layer Sample weights, Represents the total number of samples. Represents the region layer In the Model parameters after training round Indicates the learning rate. Represents the region layer In the The gradient is calculated in rounds.

7. The method for centralized heating load aggregation and control under a virtual power plant framework as described in claim 1, characterized in that, In step S5, based on the optimized heat load prediction model parameters and the digital twin heat network model, a global heat load control command is generated and issued to the regional layer, including: S501. Based on the optimized heat load prediction model parameters, input real-time environmental variables and heating network operation status data, perform heat load prediction simulation, and generate a heating network load distribution map for the next 24 hours. S502. Combining a digital twin heating network model, simulate the dynamic response process of the heating network, calculate the heat transfer efficiency and heat loss coefficient, and evaluate the feasibility and robustness of the control strategy. S503. Based on the evaluation results, apply a multi-objective optimization algorithm to solve for the optimal combination of control parameters with the objectives of minimizing the fluctuation range of the heating network, balancing the supply and demand of heat sources and reducing energy costs. S504. Generate a global heat load control instruction set based on the optimal control parameter combination, including valve opening command, heat source output power adjustment and user-side compensation strategy. S505: Through an encrypted communication protocol, global thermal load control commands are sent to execution nodes at each regional layer.

8. The method for centralized heating load aggregation and control under a virtual power plant framework as described in claim 1, characterized in that, In step S6, after receiving the global heat load control command, the regional layer dynamically adjusts the heating parameters of each end user by combining the local heat source supply capacity and the user-side thermal inertia compensation strategy, including: S601. Parse the received global heat load control command and extract the specific parameter set including valve opening command, heat source output power adjustment value and user-side compensation strategy; S602. Based on local heat source supply capacity data and specific parameter sets, assess the real-time availability range of heat source output power and calculate the heat source supply-demand balance deviation. S603. Based on the heat source supply and demand balance deviation and the user-side thermal inertia compensation strategy, apply time-sharing peak shaving and valley filling algorithms to determine the dynamic valve opening adjustment value for each terminal user. S604. Based on the valve opening adjustment value, an adjustment command is issued to the heating control equipment of each terminal user to perform dynamic adjustment of heating parameters in real time, including temperature setpoint and flow control. S605: Monitor the changes in heating parameters after adjustment and compare them with the predicted heat load curve to trigger a local feedback mechanism to optimize subsequent adjustment strategies.

9. The method for centralized heating load aggregation and control under a virtual power plant framework as described in claim 1, characterized in that, The method further includes: S7. Construct a library of thermal inertia parameters for building types; S8. Deploy intelligent temperature control valves at end users. Users can set flexible temperature zones according to their needs. Based on the thermal inertia parameter library of the building type, the heating temperature is increased by the intelligent temperature control valve when the power grid is in a low period and decreased by the intelligent temperature control valve when the power grid is in a high period. This achieves coordinated optimization of user-side thermal inertia compensation and power grid peak shaving and valley filling.

10. A centralized heating load aggregation and control system under a virtual power plant framework, characterized in that, The system includes: processor; Memory used to store processor-executable instructions; The processor is configured to implement the centralized heating load aggregation and control method under the virtual power plant framework as described in any one of claims 1 to 9 when executing the executable instructions.

Images