A heating intelligent regulation method and system based on deep learning and hybrid intelligent algorithm

Abstract

The application provides a heating intelligent regulation and control method and system based on deep learning and a hybrid intelligent algorithm, and relates to the technical field of intelligent regulation and control. The method comprises the following steps: generating two characteristic straight lines passing through the original reference point by optimizing a particle swarm algorithm, and calculating the included angle to form a dynamic regulation and control included angle region with a determined boundary; setting a dynamic detection point in and outside the included angle region; optimizing the time sequence connection path of the detection point based on a genetic algorithm to generate an arc-shaped closed route representing the state evolution of a heating system; performing reinforcement learning evaluation on the geometric properties of the closed route to output a heating parameter correction value; and converting the heating parameter correction value into a cooperative regulation and control instruction of a heat source operation parameter, a pipe network circulating water volume and a valve opening degree, and issuing the instruction to a boiler group, a variable frequency pump and an electric regulating valve execution terminal. The application improves the precision and integrity of the state description of the heating system.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a heating intelligent control method and system based on deep learning and hybrid intelligent algorithms. Background Technology

[0002] In recent years, the industry has gradually introduced intelligent algorithms for heating regulation and optimization. However, current technologies often limit the application of intelligent algorithms to single aspects (such as load forecasting or parameter optimization), lacking a collaborative design for the entire heating system process, from "data perception to feature extraction to strategy generation to execution feedback."

[0003] On the one hand, insufficient fusion processing of multi-source information such as historical operating data and meteorological parameters leads to limited accuracy of feature representation; on the other hand, the generation of control strategies does not fully consider the dynamic boundary of the system and the coupling constraints of equipment, making it difficult to achieve global collaborative optimization of heat sources, pipelines, and terminals.

[0004] In addition, most methods rely on fixed mathematical models, which are poorly adaptable to the diversity of building types and the personalization of user behavior, and lack robustness in complex scenarios. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a heating intelligent control method and system based on deep learning and hybrid intelligent algorithms, which improves the accuracy and completeness of the heating system state description.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] Firstly, a method for intelligent heating control based on deep learning and hybrid intelligent algorithms, the method comprising:

[0008] Step 1: Collect historical operating data, meteorological parameters, building characteristics and user behavior data of the heating system in real time; use a deep neural network model to perform spatiotemporal feature fusion and dimensionality reduction on the data, and transform it into a dynamic point cloud dataset in a high-dimensional feature space.

[0009] Step 2: Based on the dynamic point cloud dataset generated in Step 1, the convex hull algorithm is used to determine the point cloud boundary, and the final symmetry center is calculated in the feature space based on the mirror algorithm as the original reference point; two feature lines passing through the original reference point are generated by the particle swarm optimization algorithm, and their included angle is calculated to form a dynamically adjustable included angle region with a defined boundary.

[0010] Step 3: Set a dynamic detection point inside and outside the included angle area constructed in Step 2; optimize the time series connection path of the detection point based on the genetic algorithm to generate an arc-shaped closed path that represents the state evolution of the heating system; evaluate the geometric properties of the closed path using reinforcement learning and output the heating parameter correction value.

[0011] Step 4: Convert the heating parameter correction values ​​generated in Step 3 into coordinated control commands for heat source operating parameters, pipeline circulating water volume, and valve opening, and send them to the boiler group, variable frequency pump, and electric regulating valve execution terminals.

[0012] Secondly, a heating intelligent control system based on deep learning and hybrid intelligent algorithms includes:

[0013] The data acquisition module is used to collect historical operating data, meteorological parameters, building characteristics and user behavior data of the heating system in real time; and to use a deep neural network model to perform spatiotemporal feature fusion and dimensionality reduction on the data, transforming it into a dynamic point cloud dataset in a high-dimensional feature space.

[0014] The calculation module is used to determine the point cloud boundary based on the dynamic point cloud dataset using the convex hull algorithm, and calculate the final symmetry center in the feature space based on the mirror algorithm as the original reference point; and to optimize and generate two feature lines passing through the original reference point through the particle swarm algorithm, and calculate their included angle to form a dynamically adjustable included angle region with a defined boundary.

[0015] The correction module is used to set a dynamic detection point both inside and outside the included angle area; optimize the time series connection path of the detection point based on the genetic algorithm to generate an arc-shaped closed path that characterizes the state evolution of the heating system; evaluate the geometric properties of the closed path through reinforcement learning and output the corrected heating parameter values.

[0016] The control module is used to convert the heating parameter correction values ​​into coordinated control commands for heat source operating parameters, pipeline circulating water volume and valve opening, and send them to the boiler group, variable frequency pump and electric regulating valve execution terminals.

[0017] Thirdly, a computing device, comprising:

[0018] One or more processors;

[0019] A storage device for storing one or more computer programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0020] Fourthly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0021] The above-described solution of the present invention has at least the following beneficial effects:

[0022] At the data processing level, deep neural networks are used to fuse spatiotemporal features of multi-source heterogeneous data, transforming high-dimensional information into dynamic point cloud datasets. This solves the problems of fragmented data representation and weak feature correlation in traditional methods, and improves the accuracy and completeness of the state description of the heating system.

[0023] In terms of constructing the control region, the benchmark point is determined by the collaborative use of the convex hull algorithm and the mirror algorithm, and the dynamic angle region is generated by the particle swarm optimization. This breaks through the limitations of fixed threshold control, enabling the control boundary to adapt to changes in system parameters and providing a constraint space that is more in line with the actual operating state for subsequent optimization.

[0024] In terms of dynamic optimization mechanism, a combination strategy of genetic algorithm and reinforcement learning is used to achieve global optimization of detection point path and real-time evaluation of geometric attributes. This overcomes the local optimum trap of single algorithm in complex scenarios, ensures that the heating parameter correction value can dynamically match the system evolution law, and improves the robustness of the control strategy.

[0025] At the collaborative execution level, the correction value is transformed into multi-device collaborative instructions to achieve linkage control of heat source, pipeline network and terminal, which solves the problems of asynchronous equipment response and energy waste in traditional decentralized control, and reduces the overall energy consumption of the system while ensuring the heating quality. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a heating intelligent control method based on deep learning and hybrid intelligent algorithms provided by an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of a heating intelligent control system based on deep learning and hybrid intelligent algorithms provided by an embodiment of the present invention. Detailed Implementation

[0028] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0029] like Figure 1 As shown, an embodiment of the present invention proposes a method for intelligent heating control based on deep learning and hybrid intelligent algorithms, the method comprising the following steps:

[0030] Step 1: Collect historical operating data, meteorological parameters, building characteristics and user behavior data of the heating system in real time; use a deep neural network model to perform spatiotemporal feature fusion and dimensionality reduction on the data, and transform it into a dynamic point cloud dataset in a high-dimensional feature space.

[0031] Step 2: Based on the dynamic point cloud dataset generated in Step 1, the convex hull algorithm is used to determine the point cloud boundary, and the final symmetry center is calculated in the feature space based on the mirror algorithm as the original reference point; two feature lines passing through the original reference point are generated by the particle swarm optimization algorithm, and their included angle is calculated to form a dynamically adjustable included angle region with a defined boundary.

[0032] Step 3: Set a dynamic detection point inside and outside the included angle area constructed in Step 2; optimize the time series connection path of the detection point based on the genetic algorithm to generate an arc-shaped closed path that represents the state evolution of the heating system; evaluate the geometric properties of the closed path using reinforcement learning and output the heating parameter correction value.

[0033] Step 4: Convert the heating parameter correction values ​​generated in Step 3 into coordinated control commands for heat source operating parameters, pipeline circulating water volume, and valve opening, and send them to the boiler group, variable frequency pump, and electric regulating valve execution terminals.

[0034] In this embodiment of the invention, real-time acquisition of multi-dimensional data covers key influencing factors of heating system operation, ensuring data comprehensiveness. Furthermore, utilizing a deep neural network model for spatiotemporal feature fusion and dimensionality reduction effectively solves the problems of inconsistent formats and complex spatiotemporal correlations in multi-source data. Transforming the data into a dynamic point cloud dataset in a high-dimensional feature space enables more accurate capture of potential patterns and features between data points, improving data usability and the accuracy of feature representation.

[0035] The convex hull algorithm was used to determine the point cloud boundary, accurately defining the data distribution range and avoiding interference from irrelevant data. The mirror algorithm was used to calculate the center of symmetry as the original reference point, providing a stable and reliable reference for subsequent control. A particle swarm optimization algorithm was employed to generate characteristic lines passing through the reference point and form a dynamic control angle region. This allows the control area to dynamically adjust according to changes in the point cloud data, adapting to the dynamic characteristics of the heating system and ensuring the rationality and adaptability of the control boundary.

[0036] Setting up dynamic monitoring points inside and outside the included angle area enables comprehensive monitoring of the system status both inside and outside the area, allowing for timely detection of anomalies. A genetic algorithm is used to optimize the time-series connection paths of the monitoring points, generating arc-shaped closed routes that effectively reflect the evolution of the heating system's status. This solves the local optima problem that may exist in traditional path planning, ensuring the global optimality of the path. Reinforcement learning is applied to evaluate the geometric properties of the closed path, continuously optimizing the evaluation criteria based on the system's real-time status. This results in more accurate heating parameter corrections, improving the targeting and effectiveness of regulation.

[0037] By converting heating parameter correction values ​​into coordinated control commands, an effective connection between parameter correction and actual execution is achieved. Corresponding control commands are issued to different execution terminals such as boiler groups, variable frequency pumps, and electric regulating valves, ensuring the coordinated operation of heat sources, pipeline circulating water volume, and valve opening. This avoids the incoordination problems that may occur when each device is controlled individually, improves the overall operating efficiency of the heating system, and ensures the stability and consistency of heating effects.

[0038] The specific implementation process of step 1 above is as follows:

[0039] First, real-time collection of multi-dimensional data is carried out. For the historical operation data of the heating system, sensors deployed on equipment such as boilers, heat exchangers, and circulating pumps are used to collect parameters such as outlet water temperature, return water temperature, heat supply, fuel consumption, and equipment operating power in real time. The collection frequency is set to once every 5 minutes. The data is stored in an industrial-grade database. The collected data needs to be preliminarily screened to remove abnormal values ​​that are obviously outside the normal operating range of the equipment.

[0040] In terms of meteorological parameters, data such as outdoor temperature, humidity, wind speed, and solar radiation intensity are obtained by integrating the API interface of the city-level meteorological monitoring network. The data is updated hourly. At the same time, local meteorological data collected by distributed micro meteorological stations set up in the heating area are combined to improve the regional specificity of meteorological parameters.

[0041] Building characteristic data is obtained through a combination of preliminary on-site surveys and Building Information Modeling (BIM). Specifically, it includes static parameters such as building area, floor height, thermal conductivity of wall materials, window type and area, and building orientation. After standardization, this data is entered into the system database as basic characteristic parameters.

[0042] User behavior data is collected through room temperature controllers and smart meters installed in users' rooms. This data includes dynamic information such as the user's set temperature, the actual indoor temperature, and the distribution of heating periods. It is also combined with data such as user occupancy rate and vacancy status provided by the property management system to form a comprehensive user behavior feature dataset.

[0043] In a preferred embodiment of the present invention, a deep neural network model is used to perform spatiotemporal feature fusion and dimensionality reduction on the data, transforming it into a dynamic point cloud dataset in a high-dimensional feature space, including:

[0044] The historical operational data, meteorological parameters, building characteristics and user behavior data collected in real time are divided into continuous time window data blocks according to a preset duration.

[0045] Based on continuous time window data blocks, the static correlation between boilers, pipe networks, buildings and users in the heating system is extracted through the first deep neural network channel to obtain the spatial topological feature vector of the time window;

[0046] Based on continuous time window data blocks, the trend and periodicity of heating parameters across time windows are extracted through the second deep neural network channel to obtain a dynamic evolution feature vector characterizing the state evolution of the intelligent system.

[0047] Based on the spatial topological feature vector and the dynamic evolution feature vector, a weighted fusion using an attention mechanism is adopted to generate a joint feature vector that combines spatial correlation and temporal dependence.

[0048] Based on the joint feature vector, it is mapped to three-dimensional space through an embedding layer, and compressed and reconstructed using an autoencoder to obtain the three-dimensional feature coordinates;

[0049] The three-dimensional feature coordinates are bound to the start timestamp of the corresponding time window and the actual value of the heating load to form a labeled data point in the feature space;

[0050] The labeled data points generated in all time windows are accumulated in chronological order to obtain a dynamic point cloud dataset that is updated in real time with the status of the heating system.

[0051] In this embodiment of the invention, the time window data block is divided as follows:

[0052] The preset duration is determined by setting the time window according to the fluctuation frequency of the heating system parameters. For example, the window corresponding to the boiler outlet water temperature (high frequency fluctuation) is set to 15 minutes, and the window corresponding to the meteorological temperature (low frequency change) is set to 1 hour. The continuous division of data blocks is realized through the sliding window mechanism, and adjacent windows overlap by 50% to preserve the continuity of time.

[0053] The data blocks are structured, with each time window containing four types of sub-data matrices. The historical operation data matrix is ​​arranged by equipment number (e.g., boilers 1-5, circulating pumps 1-3), with each row corresponding to a 5-minute sampling value of one equipment. The meteorological parameter matrix records hourly data such as outdoor temperature and humidity in chronological order. The building characteristic matrix stores static parameters using building ID as an index. The user behavior matrix records set temperature change sequences in groups by room number.

[0054] Spatial topological feature vector extraction (first deep neural network channel):

[0055] The input layer processes the data by combining the equipment operation data and building characteristic data within the time window to form an "equipment-building" correlation matrix. For example, the boiler outlet temperature and the thermal conductivity of the walls of the corresponding pipe network-covered buildings form a 20×30 dimension matrix (20 equipment nodes × 30 building nodes).

[0056] The graph convolution operation employs a 3-layer graph convolutional network (GCN). The first layer calculates the local features of nodes through an adjacency matrix (constructed based on the pipeline connection relationship, with the weight of directly connected nodes set to 0.8 and the weight of indirectly connected nodes set to 0.3). The second layer aggregates the second-order neighbor features, and the third layer outputs a 512-dimensional device-building spatial association vector.

[0057] Feature selection involves thresholding the 512-dimensional vector using the Sigmoid activation function (retaining features with activation values ​​> 0.6), ultimately generating a 128-dimensional spatial topological feature vector, which includes core features such as the flow matching degree between the boiler and the pipeline network, and the correlation strength between building insulation performance and heat supply.

[0058] Dynamic evolution feature vector extraction (second deep neural network channel):

[0059] The time series construction involves splicing the heating parameters (such as outlet water temperature and heating capacity) of six consecutive time windows along the time axis to form a 120×6 time series matrix (120 sampling points × 6 windows), which serves as the input sequence for the LSTM network.

[0060] The bidirectional LSTM processing employs a two-layer bidirectional LSTM. The first layer extracts upward trend features (such as the temperature rise in the morning and evening) and downward trend features (such as the temperature drop at night) in the reverse direction, outputting a 256-dimensional intermediate vector. The second layer focuses on periodic patterns and strengthens the feature weights of peak heating periods such as 8:00 and 18:00 every day through a gating mechanism.

[0061] Trend decomposition uses an attention weight matrix (32×256 dimensions) to weight the output vector. The attention weight of the periodic term is set to the peak at the 24-hour period node, and the weight of the trend term changes linearly with time. Finally, a 64-dimensional dynamic evolution feature vector is generated, which contains the predicted trend value of heat supply for the next 2 hours.

[0062] Joint feature vector generation (attention mechanism fusion):

[0063] Weight calculation involves constructing a mutual attention matrix between spatial and dynamic features. For example, the cosine similarity between "pipeline resistance features" in the spatial topology vector and "flow change trend" in the dynamic evolution vector is calculated and used as the fusion weight (range 0-1) for the feature pair.

[0064] Weighted fusion is performed by summing the 128-dimensional spatial vector and the 64-dimensional dynamic vector element by element, with spatial feature weights accounting for 60% (mainly static correlation) and dynamic feature weights accounting for 40% (secondarily time-varying characteristics), generating a 128-dimensional joint feature vector.

[0065] Feature enhancement is achieved through a two-layer MLP (128→256→128 nodes) nonlinear transformation, using the ReLU activation function to enhance the nonlinear expressive power of features, such as strengthening the interactive features of "building orientation" and "sunlight intensity variation".

[0066] 3D Feature Coordinate Mapping and Compression:

[0067] The embedding layer transformation maps the joint feature vector to a three-dimensional space through a fully connected embedding layer (128→3 nodes). Each dimension corresponds to the physical meaning of "heat source output intensity", "pipeline transmission efficiency" and "user demand matching degree". For example, the higher the value of the first dimension, the greater the boiler output.

[0068] Autoencoder training uses a convolutional autoencoder for compression. The encoder extracts local correlations of 3D features (such as matching features between heat sources and pipe networks) through a 3×3 convolutional kernel. The decoder reconstruction error is set to MSE (mean square error). Training stops when the error is less than 0.01, and the optimized 3D coordinates (x, y, z) are output. The coordinate values ​​are normalized to the range of [-1, 1].

[0069] Generate labeled data points:

[0070] Timestamp binding associates the UNIX timestamp (accurate to the second) at the start of the time window with three-dimensional coordinates. For example, the data point corresponding to 2023-12-01 08:00:00 is marked as (x1, y1, z1, 1672531200).

[0071] The load label calculation takes the average heat supply within the time window as the true load value, converts it into a label value between 0 and 1 through maximum and minimum normalization, and combines it with the three-dimensional coordinates to form a quadruple data point (x, y, z, label, timestamp).

[0072] Outlier filtering involves calculating the Euclidean distance between a data point and the historical mean for the same time period. Points with a distance greater than 3 times the standard deviation are marked as outliers and removed to ensure the reliability of point cloud data.

[0073] Construction of dynamic point cloud datasets:

[0074] Time-series queue management uses a FIFO queue to store point cloud data from the most recent 72 hours (approximately 288 data points). When new data enters each window, the oldest data point is automatically removed to maintain a stable dataset size. A KD-tree is used to create a spatial index for the 3D point cloud, stored hierarchically along the z-axis (user demand dimension) with a 0.1 interval between each layer, facilitating quick querying of the point cloud distribution within a specific load range. Point cloud updates are triggered every 15 minutes, synchronously calculating the spatial similarity between new data points and historical points. When the similarity is >0.9, the data points are merged into clusters to reduce redundant data.

[0075] Through the above refined processing, each step achieves a traceable transformation from raw data to feature point cloud. For example, the data points of a residential building during the morning peak period (x=0.8, y=0.6, z=0.9) can clearly correspond to the system status of "boiler operating at high load, good pipeline transmission efficiency, and high user demand matching degree".

[0076] This invention divides continuous, multi-dimensional data into structured data blocks by setting a time window duration, allowing subsequent feature extraction to focus on information within specific time segments and avoiding the inefficiency caused by the disordered processing of massive amounts of data. The first deep neural network channel is specifically designed to extract static relationships, clearly outlining the spatial topology between boilers, pipe networks, buildings, and users, such as the connection methods between pipe networks and buildings, and the boiler's impact on the heating of different buildings, enabling precise characterization of the correlation characteristics of each part of the system. The second channel targets the trend and periodic patterns across time windows, effectively uncovering the changing patterns of heating parameters over time, such as daily and weekly peak heating patterns, and the long-term impact trends of meteorological changes on heating parameters, comprehensively capturing the dynamic evolution characteristics of the system. The introduction of an attention mechanism allows for weighted fusion based on the importance of different features in the current system state. For example, during extreme weather events, the weight of meteorological-related dynamic evolution features increases, while during stable system operation, spatial topological features may receive more attention. This allows the generated joint feature vector to highlight key information, balancing spatial correlation and temporal dependence, thus enhancing the comprehensive representation capability of features. By mapping to three-dimensional space through an embedding layer and combining it with autoencoder compression and reconstruction, the data dimensionality is significantly reduced while preserving the key features in the original data to the greatest extent. The three-dimensional feature coordinates not only simplify data complexity but also provide an intuitive spatial carrier for subsequent conversion into a dynamic point cloud dataset. Labeled data points are bound to timestamps and actual heating load values, enabling the point cloud dataset to not only reflect the spatial distribution of system features but also their dynamic changes over time. Each data point has a clear practical reference, and the dynamic point cloud dataset formed by accumulating over time can track changes in the heating system status in real time.

[0077] In a preferred embodiment of the present invention, step 2, based on the dynamic point cloud dataset generated in step 1, uses the convex hull algorithm to determine the point cloud boundary, and calculates the final symmetry center in the feature space as the original reference point based on the mirror algorithm, including:

[0078] Step 21: Based on the dynamic point cloud dataset, use the convex hull algorithm to iteratively calculate the minimum convex polyhedron boundary of the point cloud in the three-dimensional feature space, and output the boundary vertex coordinate set of the current point cloud;

[0079] Step 22: Based on the boundary vertex coordinate set, generate a mirror point set of the boundary vertex set in the feature space; calculate the spatial distribution similarity between the original vertex set and the mirror point set, and take maximizing the similarity as the optimization objective; iteratively adjust the position of the candidate point of the symmetry center, and when the spatial distribution similarity reaches a preset threshold, output the optimal solution as the final coordinates of the symmetry center, i.e., the original reference point.

[0080] In this embodiment of the invention, step 21 involves selecting the maximum and minimum points (a total of 6 points) of the x-axis, y-axis, and z-axis coordinates in the three-dimensional feature space from the dynamic point cloud dataset, using them as the initial vertices of the convex hull; connecting these initial vertices to form triangular faces, thus constructing an initial convex polyhedron; traversing other points in the point cloud, determining whether each point is inside the current convex polyhedron (by checking whether the point is located inside all triangular faces); if a point is outside the convex polyhedron, removing the triangular face in the region where the point is located, connecting the point with adjacent vertices to form new triangular faces, and updating the convex polyhedron; repeating this process until all points are contained within the convex polyhedron; and extracting the coordinates of all vertices of the final convex polyhedron to form a boundary vertex coordinate set, which serves as the three-dimensional boundary of the point cloud.

[0081] In step 22 above, a candidate point for the center of symmetry is preset in the feature space (the initial value is set as the geometric center of the point cloud). Using this candidate point as the center of symmetry, the mirror point of each vertex in the boundary vertex coordinate set (i.e., the equidistant point of the extension line connecting the vertex and the candidate point) is calculated to form a mirror point set. The distribution characteristics of the original vertex set and the mirror point set in the three-dimensional space are statistically analyzed (such as the proportion of vertices in each coordinate interval, the average distance between vertices, the position of the vertex cluster center, etc.). By comparing the matching degree of these features, the spatial distribution similarity (range 0-1) is obtained. If the similarity does not reach the preset threshold (e.g., 0.9), the coordinates of the candidate point are finely adjusted along the x, y, and z axes (each fine adjustment is 1% of the current point cloud coordinate range), the mirror point set is regenerated and the similarity is calculated. The adjustment is repeated until the similarity is ≥ the threshold. At this time, the candidate point is the final center of symmetry (original reference point).

[0082] This invention constructs the minimum convex polyhedron boundary through an iterative convex hull algorithm, which can accurately delineate the spatial distribution range of point cloud data, eliminate noise points or outliers far from the core distribution, and provide a clear and reliable boundary reference for subsequent control area delineation, avoiding interference from irrelevant data. Based on the mirror algorithm and similarity optimization, the symmetry center is determined, making the benchmark point both objective and stable (unaffected by local data fluctuations), providing a unified and reliable spatial reference benchmark for subsequent feature line generation and control area construction, and improving the consistency and accuracy of the control logic.

[0083] In a preferred embodiment of the present invention, two feature lines passing through the original reference point are generated through particle swarm optimization, and their included angle is calculated to form a dynamically adjustable included angle region with defined boundaries, including:

[0084] Step 23: Based on the boundary vertex coordinate set and the final symmetry center coordinates, perform the following operations using the particle swarm optimization algorithm:

[0085] Initialize the parameters of two lines passing through the original reference point; minimize the sum of the weighted distances from the boundary vertex to the two lines as the optimization objective; iteratively update the line parameters until convergence, and output the optimized equations of the two feature lines;

[0086] Step 24: Based on the optimized equations of the two characteristic lines, calculate the spatial angle between the two lines, and output the dynamic control angle region containing the key state points of the heating system, with the original reference point as the vertex and the rays on both sides of the angle as the boundary.

[0087] In this embodiment of the invention, in step 23 above, the original reference point coordinates (e.g., (x0, y0, z0)) are known. The two lines to be optimized must always pass through this point. The direction of each line is described by a three-dimensional direction vector (a, b, c), where the values ​​of a, b, and c are limited to [-1, 1] (to ensure the normalization of the direction vector).

[0088] The particle swarm is initialized by creating a set of 50 "particles". Each particle represents a set of candidate solutions, namely the direction vector parameters of two lines (the first line (a1, b1, c1) + the second line (a2, b2, c2)). To ensure that the initial lines are evenly distributed around the reference point, the parameter values ​​are generated randomly and must cover different intervals within the range of [-1, 1] (for example, the a value of some particles is biased towards -1, some towards 0, and some towards 1).

[0089] Calculate the total weighted distance of individual particles (to evaluate the quality of the solution):

[0090] Obtain boundary vertex data. From the boundary vertex coordinate set obtained in step 21, extract the three-dimensional coordinates of all vertices (e.g., a total of 30 vertices with coordinates of (x1, y1, z1), (x2, y2, z2)...(x30, y30, z30)).

[0091] Calculate the perpendicular distance from a vertex to a line. For each boundary vertex, calculate its perpendicular distance to the two lines represented by the current particle (distance means: the length of the spatial line from the vertex to the nearest point on the line). For example, the distance from vertex (x1, y1, z1) to the first line is denoted as d1, and the distance to the second line is denoted as d2.

[0092] Vertex weights are introduced, and weight values ​​are assigned to vertices based on their importance (core region vertices have a weight of 1.2, and edge region vertices have a weight of 0.8). Core region vertices refer to vertices that are close to the geometric center of the point cloud and can represent the typical state of the system (such as vertices during normal stable operation), while edge region vertices refer to vertices that are far from the center and may correspond to extreme states.

[0093] To calculate the total weighted distance, for each vertex, add its distances to the two lines (d1 + d2), and then multiply by the weight of that vertex to get the weighted distance of a single vertex. Finally, sum the weighted distances of all 30 vertices to get the "total weighted distance" corresponding to the particle (the smaller the value, the closer the two lines represented by the particle are to the boundary vertex distribution, and the higher the quality of the solution).

[0094] Iterative particle swarm optimization (approaching the optimal solution step by step):

[0095] Record the initial optimal solution, calculate the total weighted distance of 50 particles, and retain the minimum value of each particle (individual historical best) and the minimum value among the 50 particles (population best, i.e. the current best global solution).

[0096] Adjusting particle parameters (updating direction vectors): Each particle updates its own direction vector parameters according to the following rules:

[0097] Referring to "individual historical best": if there was a better solution (smaller total weighted distance), then fine-tune the parameters in the direction of that solution (for example, the a1 value of a particle moves from 0.3 to the historical best of 0.2).

[0098] Refer to "population optimal": fine-tune the parameters in the direction of the current global optimal solution (for example, if the population optimal value of a2 is -0.5, the a2 value of a certain particle moves closer to -0.5 from -0.3).

[0099] Adjustment range: Each parameter change shall not exceed 5% of the current range, that is, the single adjustment value of a, b, and c shall not exceed 0.1 (because the range length of [-1, 1] is 2, 2 × 5% = 0.1), to avoid parameter mutations that lead to a decrease in the stability of the solution.

[0100] Re-evaluate and update the optimal solution: After parameter adjustment, recalculate the total weighted distance of the particle. If the new distance is less than its individual historical best, update the individual best; if the new distance is less than the group best, update the group best.

[0101] Convergence determination: Repeat the adjustment-evaluation process described above, recording the total weighted distance of the population optimum after each iteration. If the change in the total weighted distance of the population optimum is less than 0.01 for 10 consecutive iterations (e.g., from 12.30 to 12.29, then to 12.295, with fluctuations within 0.01), then the iteration is considered to have converged, and optimization is stopped.

[0102] After convergence, the optimal particle parameters corresponding to the current population are taken, namely the direction vectors of the two lines (a1, b1, c1) and (a2, b2, c2). Since both lines pass through the original reference point (x0, y0, z0), the two final characteristic lines can be determined by combining the direction vectors (for example, the first line can be described as "starting from (x0, y0, z0) and extending along the direction (a1, b1, c1)").

[0103] In step 24 above, the direction vectors of the two feature lines obtained from step 23 are extracted and denoted as vector 1 (a1, b1, c1) and vector 2 (a2, b2, c2) (for example, vector 1 is (0.6, 0.3, 0.2) and vector 2 is (-0.4, 0.5, 0.3)). The angle between the two vectors is determined by comparing the "closeness of their directions". Specifically, the corresponding components of the two vectors are multiplied separately (e.g., a1×a2, b1×b2, c1×c2), and the results are added together to obtain a comprehensive value (this value reflects the correlation between the vector directions; the larger the value, the closer the directions are).

[0104] Based on the aforementioned correlation coefficient values, combined with the vector's own "length characteristic" (calculated by taking the square root of the sum of the squares of each component), a ratio value between 0 and 1 (i.e., the cosine value) is obtained. This ratio value corresponds to the cosine value of the angle between the vectors. For example, when the ratio value is 0.866, the corresponding angle is 30°; when the ratio value is 0.5, the corresponding angle is 60°. The final angle range is limited to 0°-90° (if the calculated result exceeds 90°, the supplementary angle is taken to ensure that the angle is acute). Taking the original reference point (x0, y0, z0) as the starting point, the two characteristic lines are respectively used as two boundary rays (for example, the reference point is (2, 3, 1), ray 1 extends along the direction of vector 1, and ray 2 extends along the direction of vector 2). The area between the two rays is the initial dynamic adjustment angle area. The method to determine whether a point is within the area is: draw a ray from the reference point to the point. If the angle between this ray and the two boundary rays is less than the previously calculated spatial angle (such as 30°), then the point is located inside the area.

[0105] From the generated dynamic point cloud dataset, key state points are selected, including:

[0106] Peak load point: The point with the highest heat supply (e.g., point cloud data during the coldest period of winter).

[0107] Stable system operating points: points where the heating supply is stable and the user demand is well matched (such as typical data during off-peak hours).

[0108] Extreme operating condition point: The point corresponding to extreme weather (such as blizzard) or minor equipment failure.

[0109] Check each of the key state points to ensure they are within the initial included angle region. For example, if a load peak point is found outside the region, it indicates that the initial included angle range is insufficient. If there are any omissions, expand the two boundary rays outward by a certain angle (each expansion should not exceed 2°, and the total expansion should be controlled within 5°), for example, from 30° to 33°, and then recheck whether all key state points are included. Once all key state points are within the included angle, stop fine-tuning, determine the final included angle (e.g., 33°), and output the dynamically adjustable included angle region using the two boundary rays and the reference point as the boundary.

[0110] This invention optimizes the feature line using a particle swarm optimization algorithm, ensuring that the line passes through the reference point (maintaining the stability of the reference point) and that the line conforms to the boundary vertex distribution to the maximum extent (minimizing the weighted distance), thus ensuring that the line can represent the core distribution trend of the point cloud. Iterative optimization avoids the problem of local optima, making the feature line more globally representative. The spatial angle calculation clarifies the boundary range of the control area. The angle region with the reference point as the vertex can dynamically adapt to changes in the point cloud distribution (changing with the adjustment of the line direction) and includes all key state points, providing a precise and controllable range framework for subsequent system state monitoring and parameter control, improving the targeting and boundary rationality of the control.

[0111] In a preferred embodiment of the present invention, step 3, setting a dynamic detection point both inside and outside the included angle area constructed in step 2, includes:

[0112] Step 31: Using the dynamically adjustable angle region output in Step 24 as the input space, filter all data points in the current point cloud dataset that are located within this region; identify the dense distribution center of the data points within the region based on the density clustering algorithm, and output the coordinates of the labeled data point closest to the dense center as the detection point within the region;

[0113] Step 32: Using the data points outside the region and the boundary ray of the region with the included angle from Step 24 as input, calculate the vertical distance from each data point outside the region to the two boundary rays; select the data point outside the region that maximizes the sum of the two vertical distances, project it onto the angle bisector of the two boundary rays, and output the coordinates of the projected point as the detection point outside the region.

[0114] Step 33: Pair the detection points within the region with the detection points outside the region, bind them to the start timestamp of the same time window, and output the coordinate set of the dual detection points with time sequence markers.

[0115] In this embodiment of the invention, step 31 above involves finding a detection point within the dynamically adjustable angle region that represents the typical state of the system (reflecting the core operating characteristics within the region).

[0116] Filter data points within the region:

[0117] From the generated dynamic point cloud dataset, check each data point one by one whether it is located within the included angle region defined in step 24. The determination method is as follows: draw a ray from the original reference point to the data point. If the included angle between this ray and the two boundary rays is less than the final included angle (e.g., 33°), it is determined to be a "data point within the region" (for example, if the included angle between the ray corresponding to a point and the boundary ray is 25° and 30° respectively, both less than 33°, then it belongs to the region).

[0118] Density clustering identifies dense centers:

[0119] Perform "density analysis" on the data points within the selected area:

[0120] Set a "density threshold": For example, count the number of other data points within a radius of 0.2 around each point (in three-dimensional space), and points with a number ≥ 5 are considered "high-density points";

[0121] Clustering high-density points: Grouping high-density points that are close together (spatial distance < 0.1) into one category to form several "dense clusters";

[0122] Determine the dense center: Calculate the average coordinates of all points in each dense cluster (such as the average x-coordinate, average y-coordinate, and average z-coordinate). The resulting coordinates are the "dense center" of the cluster (for example, if a cluster has 3 points (1.2, 0.8, 0.9), (1.3, 0.7, 0.8), and (1.1, 0.9, 1.0), the dense center is (1.2, 0.8, 0.9)).

[0123] Identify the detection points within the area:

[0124] Calculate the spatial distance (straight-line distance) between each data point in the region and the dense center. Select the data point that is closest to the center (which must have a timestamp and load label). Its coordinates are the "detection point in the region" (for example, if the coordinates of the point closest to the dense center are (1.2, 0.8, 0.9), then this will be the detection point in the region).

[0125] In step 32 above, outside the included area, a detection point that can reflect the system's abnormality or edge state is found (to form a contrast with the area within the region and assist in monitoring abnormalities).

[0126] Filter data points outside the region:

[0127] The judgment logic is the opposite of step 31. If the angle between the ray corresponding to a data point and any boundary ray is greater than or equal to the final angle (e.g., 33°), then it is determined to be a "data point outside the region" (for example, if the angle between the ray corresponding to a point and a boundary ray is 35°, then it belongs to the outside region).

[0128] Calculate the distance from a point outside the region to the boundary ray:

[0129] For each data point outside the region, calculate its "perpendicular distance" to the two boundary rays (i.e., the length of the perpendicular line drawn from that point to each ray). For example, if the perpendicular distance from a point outside the region to ray 1 is 0.5 and the perpendicular distance to ray 2 is 0.6, then the "sum of distances" is 1.1.

[0130] Select the target point and project it onto the angle bisector:

[0131] Find the data point outside the region with the "maximum distance" (this point is far from both boundary rays and can reflect the extreme state outside the region to the greatest extent).

[0132] Determine the "angle bisector" of the two boundary rays: that is, the ray that is located exactly in the middle of the two rays starting from the reference point (for example, if the angle between the two rays is 33°, then the angle bisector is 16.5° with each ray).

[0133] Projection operation: Draw a perpendicular line from the selected point outside the region to the angle bisector. The intersection of the perpendicular line and the angle bisector is the "projection point", and the coordinates of this point are the "outside-region detection point" (for example, if the coordinates of the outside-region point are (2.0, 1.5, 1.2), and it is projected onto the angle bisector to get (1.8, 1.4, 1.1), then this is the outside-region detection point).

[0134] Step 33 above links the detection points inside and outside the region to form time-stamped monitoring pairs, which facilitates tracking the system status during the same period.

[0135] Detection point pairing:

[0136] Ensure that the detection points within the region and the detection points outside the region come from the same time window (i.e., the same preset time segment divided in step 1, such as a 15-minute window), forming a "detection point pair" (for example, both corresponding to the time window of 2023-12-01 08:00-08:15).

[0137] Bind timestamp:

[0138] The starting timestamp of the time window (e.g., UNIX timestamp 1672531200, corresponding to 2023-12-01 08:00:00) is bound to the coordinates of the detection point pair to form structured data containing "coordinates within the region, coordinates outside the region, and timestamp" (e.g., ((1.2, 0.8, 0.9), (1.8, 1.4, 1.1), 1672531200)).

[0139] Output the coordinate set of the two detection points:

[0140] All detection point pairs for each time window are stored in chronological order to form a "double detection point coordinate set with time sequence marking" (e.g., one pair is generated every 15 minutes, and data is stored continuously for 72 hours), providing basic data for subsequent path planning.

[0141] Through the above process, the detection points within the region are determined based on a dense center, which can accurately reflect the typical state of normal system operation (such as the core characteristics of stable heating and load matching). The detection points outside the region are filtered by maximizing distance and projected onto the angle bisector, which can effectively capture potential abnormal or extreme edge states (such as sudden load changes and equipment deviation from the optimal operating range). The combination of the two achieves comprehensive coverage of the system state. By binding the timestamps of the same time window, the detection points inside and outside the region form a time sequence pair, which can intuitively track the state correlation of the system in the core area and the edge area within the same time period (such as the coordinated changes of internal and external points during normal operation and the deviation trend during abnormalities). The setting of dual detection points provides clear monitoring anchors for subsequent path planning and parameter evaluation. It can ensure the stability of routine control through the internal points and provide timely warnings of abnormalities through the external points, so that the control strategy takes into account both normal optimization and dynamic emergency response, and improves the system's ability to cope with complex operating conditions.

[0142] In a preferred embodiment of the present invention, the time-series connection path of the detection points is optimized based on a genetic algorithm to generate an arc-shaped closed path characterizing the state evolution of the heating system, including:

[0143] Step 34: Based on the set of coordinates of the two detection points, arrange the two detection points of the consecutive time windows in the order of timestamps to form the initial path point sequence;

[0144] Step 35: Using the initial path point sequence as a chromosome, perform the following operations using a genetic algorithm:

[0145] The polar coordinate parameters of the path point sequence are used as genes;

[0146] With the optimization objectives of maximizing the overall curvature smoothness of the path and minimizing the total arc length, the selection, intersection, and mutation processes are iteratively executed until the final optimized path point sequence that satisfies the arc closure constraint is generated;

[0147] Step 36: Connect the final optimized path point sequence end to end and fit it into a continuous arc curve in the three-dimensional feature space, outputting an arc-shaped closed path that characterizes the state evolution of the heating system.

[0148] In this embodiment of the invention, step 34 above arranges the dual detection points of the continuous time window in chronological order to provide an initial path basis for the genetic algorithm.

[0149] Extract the time window sequence. From the set of coordinates of dual detection points with time sequence marking, extract continuous time windows (e.g., select 96 time windows in the most recent 24 hours, with each window spaced 15 minutes apart), and sort them from early to late by timestamp (e.g., t1 = 8:00, t2 = 8:15, ..., t96 = 8:00 the next day).

[0150] Integrating dual detection points: Each time window contains one detection point within the region (e.g., A1, A2, ..., A96) and one detection point outside the region (e.g., B1, B2, ..., B96). These points are arranged alternately in chronological order to form an initial path point sequence: A1→B1→A2→B2→...→A96→B96 (a total of 192 points, each with three-dimensional coordinates).

[0151] Step 35 above uses a genetic algorithm for iterative optimization to make the path smoother, shorter, and closure-compliant (connected at the beginning and end).

[0152] Define chromosomes and genes:

[0153] Chromosome: The initial path point sequence in step 34 is taken as a "chromosome" (i.e., a candidate path scheme). The entire genetic algorithm contains 50 such chromosomes (forming the initial population).

[0154] Gene: The polar coordinate parameters of each path point are used as the gene. The polar coordinates take the original reference point as the pole, and the parameters include the polar angle (the angle between the line connecting the point to the reference point and a certain reference axis) and the polar radius (the straight-line distance between the point and the reference point). For example, if the polar coordinate parameters of point A1 are (30°, 1.2), then these two values ​​are the gene corresponding to that point.

[0155] Set optimization goals:

[0156] Maximizing curvature smoothness: The smaller the change in the angle between the lines connecting two adjacent points in the path (e.g., the angle between the line connecting A1 to B1 and the line connecting B1 to A2 is ≤10°), the smoother the curvature;

[0157] Minimize total arc length: The sum of the arc lengths (curve distances) between all adjacent points should be as small as possible (e.g., the total arc length should be controlled within 20).

[0158] Combining the two objectives, a "fitness score" is calculated for each chromosome (chromosomes with higher smoothness and shorter arc lengths receive higher scores). The specific calculation process for the fitness score includes:

[0159] For each intermediate point in the path (e.g., A1, B1, A2, etc., excluding the beginning and end points), calculate the spatial angle between the two line segments preceding and following it (e.g., for point B1, calculate the angle between A1→B1 and B1→A2); convert each angle to a value in the range of 0-1. For example, if the angle is 0° (perfectly smooth), the corresponding value is 1; if the angle is 90°, the corresponding value is 0; other angles are calculated using linear interpolation (e.g., 45° corresponds to 0.5); sum the corresponding angle values ​​of all intermediate points in the path to obtain the total smoothness score of the path; for example, if there are 94 intermediate points in the path, and the sum of the corresponding angle values ​​of all points is 85, then the total smoothness score is 85; divide the total smoothness score by the "ideal maximum score" (i.e., the score when all angles are 0°). In this example, the value is 94), resulting in a normalized smoothness value (e.g., 85 / 94≈0.904). For each pair of adjacent points in the path (e.g., A1→B1, B1→A2, etc.), calculate the curve distance in three-dimensional space (if the two points are fitted with an arc, take the arc length; if they are connected by a straight line, take the straight line length). Sum the arc lengths between all adjacent points to obtain the total arc length of the path. For example, the sum of the arc lengths of all 95 segments is 18.6. Map the total arc length to the range of 0-1. Set an "ideal minimum arc length" (e.g., 15) and an "acceptable maximum arc length" (e.g., 25), and calculate the score through linear mapping: if the total arc length ≤ 15, the score is 1; if the total arc length ≥ 25, the score is 0; if the total arc length is 18.6, the score = (25-18.6) / (25-15) = 0.64.

[0160] The normalized smoothness value (e.g., 0.904) and arc length score (e.g., 0.64) are combined according to their weights. Assuming the smoothness weight is 0.6 and the arc length weight is 0.4, the overall score is: Fitness Score = 0.904 × 0.6 + 0.64 × 0.4 = 0.7984.

[0161] Iterative operations of genetic algorithm:

[0162] Selection: From 50 chromosomes, the top 20% of chromosomes in fitness score (10 chromosomes) are retained and directly enter the next generation; the remaining 40 chromosomes are selected by "roulette" (the higher the score of the chromosome, the higher the probability of it being selected), and finally 10 chromosomes are retained, for a total of 20 chromosomes to enter the next step.

[0163] Crossover: The 20 selected chromosomes are paired up (10 pairs). A "crossover point" (e.g., the 50th gene) is randomly selected from each pair of chromosomes, and all genes after the crossover point are exchanged. For example, if the gene sequence of chromosome 1 is [g1, g2, ..., g50, g51, ..., g192] and chromosome 2 is [h1, h2, ..., h50, h51, ..., h192], after crossover, two new chromosomes are generated: [g1, g2, ..., g50, h51, ..., h192] and [h1, h2, ..., h50, g51, ..., h192].

[0164] Mutation: For the chromosomes after crossover, 5% of the genes are randomly selected for fine-tuning (e.g., polar angle ±3°, polar radius ±0.05) to avoid the population getting trapped in local optima. For example, a gene with an original polar angle of 30° may be mutated to 32°.

[0165] Constraint check: After each iteration, check whether the path satisfies the "arc closure" (the spatial distance between the first and last points of the path is ≤0.1, ensuring that the beginning and end can be connected); if it does not satisfy the requirement, make targeted adjustments to the genes of the first and last points (such as shortening the polar diameter difference).

[0166] Termination condition: Stop iteration when the fitness score changes by ≤0.02 over 15 consecutive iterations, and output the current optimal path point sequence.

[0167] Step 36 above fits the optimized path point sequence into a continuous curve, which intuitively reflects the evolution of the system state.

[0168] First-to-last connection: Arrange the final optimized path point sequence output in step 35 in chronological order, and connect the last point to the first point (e.g., if the sequence is A1→B1→...→B96, connect B96→A1) to form a closed loop.

[0169] Smooth curve fitting: For points in a closed loop, a "piecewise arc fitting" method is used: adjacent points are connected by smooth arcs (instead of straight lines), ensuring that the curvature change of the arcs meets the optimized smoothness target (e.g., the curvature difference between adjacent arcs ≤ 0.05). For example, the curvature of the arc from A1 to B1 is 0.2, and the curvature of the arc from B1 to A2 is 0.22, showing a gradual change.

[0170] Output closed path: Finally, a continuous and smooth three-dimensional arc-shaped closed path is formed. The points on the path are arranged in chronological order, which can intuitively reflect the state evolution of the heating system from t1 to t96 and back to near t1 (such as the cycle process from stable operation → slight fluctuation → return to stability).

[0171] Through the above process, the path optimized by the genetic algorithm can accurately reflect the temporal correlation of the detection points and intuitively present the periodic or cyclical evolution of the system state through the arc-shaped closure characteristic. After optimization, the detection point path arranged in time sequence can intuitively reflect the state changes of the heating system from one time window to the next (such as load fluctuations, efficiency changes, etc.), and the arc-shaped closure characteristic clearly presents the periodic cycle of the system state (such as the recurrence of day and night heating patterns). The genetic algorithm aims to "maximize curvature smoothness," avoiding abrupt turns in the path caused by the discrete distribution of detection points. The route is designed to better align with the continuous evolution of the system's state; minimizing the total arc length ensures a concise and efficient path, reducing redundant information interference; iterative optimization satisfies the "arc closure" constraint, ensuring the path forms a closed loop, which aligns with the cyclical nature of the heating system's operation (e.g., daily repetitive heating patterns); the smooth, closed arc route contains rich geometric features (e.g., curvature changes, arc length distribution), which can accurately map the health of the system's state (e.g., sudden curvature changes may correspond to abnormal fluctuations), making reinforcement learning evaluation more targeted and the output parameter correction values ​​more closely aligned with actual needs.

[0172] In a preferred embodiment of the present invention, the geometric properties of the closed path are evaluated using reinforcement learning, and corrected heating parameter values ​​are output, including:

[0173] Based on the arc-shaped closed path, the curvature distribution, total arc length, and first and last closure deviation values ​​of the path are extracted as the core geometric attribute set;

[0174] The core geometric attribute set is bound to the state parameters of the real-time heating system to construct the state space of reinforcement learning; the state parameters of the real-time heating system include the pipe network water temperature, the room temperature on the user side, and the valve opening feedback value.

[0175] An evaluation is performed based on the state space of the reinforcement learning to obtain a comprehensive reward value;

[0176] Based on the deep Q-network strategy, the action that maximizes the comprehensive reward value is selected to obtain the final action vector, which is then used as the correction value for the heating parameters.

[0177] In this embodiment of the invention, the arc-shaped closed path is sampled at fixed intervals (e.g., every 0.1 unit length) to obtain a discrete point set; for each sampled point, the angle between the line segments formed by the two points before and after it is calculated, and the angle is converted into a curvature value (the smaller the angle, the greater the curvature); all curvature values ​​are summarized, and the proportion of points in different curvature intervals (e.g., [0, 0.1), [0.1, 0.2), etc.) is statistically analyzed to form a curvature histogram; the straight-line distance between adjacent sampled points is accumulated (if the path has been fitted to an arc, the arc length is taken) to obtain The total path length is calculated; the magnitude of the difference between the three-dimensional coordinates of the path's starting and ending points is used as the closure deviation (ideally 0, the smaller the actual value, the better the closure); the proportion of each interval of the curvature histogram is directly used as a feature (e.g., [0.1, 0.3, 0.4, 0.15, 0.05]); the total arc length is divided by a preset reference length (e.g., the ideal path length), mapped to the [0, 1] interval; the closure deviation value is normalized using a preset maximum deviation value (e.g., 0.2), the smaller the deviation, the larger the value; real-time heating parameters are bound:

[0178] Pipeline water temperature: Normalized by a preset comfort range (e.g., 55-75℃), with a value of 1 within the range and a linear decay outside the range;

[0179] User-side room temperature: Normalized to a comfortable temperature range (e.g., 18-24℃), with a value of 1 within the range, otherwise reduced according to the degree of deviation;

[0180] Valve opening: Normalized by the economic operating range (e.g., 60%-80%), with a value of 1 within the range, otherwise penalized.

[0181] State vector concatenation:

[0182] The normalized geometric properties and heating parameters are concatenated in sequence to form a complete state vector (such as a 10-dimensional vector).

[0183] Calculate the overall reward value:

[0184] Geometric attribute bonus:

[0185] Curvature bonus: If the curvature is concentrated in the ideal range (e.g., [0.15, 0.25]), the bonus is +1; if there are high curvature points (>0.3), the penalty is -0.1 for each point;

[0186] Arc length bonus: When the total arc length is close to the ideal value, the bonus is +0.8; the greater the deviation, the greater the penalty.

[0187] Closure reward: +0.7 is awarded when the deviation is ≤0.1, otherwise the deviation is linearly penalized.

[0188] Heating parameter bonus:

[0189] Water temperature bonus: +1 for water temperature within the target range, -0.2 for every 1°C deviation.

[0190] Room temperature bonus: +1.2 if within the comfort zone (room temperature has higher priority), otherwise penalized;

[0191] Valve opening bonus: +0.8 is awarded within the economic range, otherwise a penalty is imposed.

[0192] The geometric attribute reward and the heating parameter reward are summed according to their weights (e.g., geometry accounts for 0.4 and heating accounts for 0.6) to obtain the final reward value.

[0193] The decision-making process of a deep Q-network is as follows:

[0194] Action space definition:

[0195] Define heating parameter correction actions, including:

[0196] Heat source power adjustment (e.g., {-5%, -2%, 0%, 2%, 5%}); circulating pump frequency adjustment (e.g., {-3Hz, -1Hz, 0Hz, 1Hz, 3Hz}); valve opening correction (e.g., {-5%, -2%, 0%, 2%, 5%}).

[0197] Q network structure:

[0198] Design a 3-layer fully connected neural network with a state vector (10-dimensional) as input and Q-values ​​for each action as output (a total of 5×5×5=125 Q-values).

[0199] ε-greedy strategy:

[0200] Initial phase (first 1000 steps): ε = 0.9 (high exploration rate);

[0201] Mid-training: ε decays linearly to 0.1;

[0202] Stable period: ε = 0.05 (mainly utilizing previously learned knowledge).

[0203] Experience replay:

[0204] The state, action, reward, and next state of each time step are stored in a buffer (capacity 10,000 entries), and 64 entries are randomly selected for learning in each training session.

[0205] The parameter correction values ​​are generated as follows:

[0206] Choose the optimal action:

[0207] In the current state, select the action combination with the highest Q value (such as power +2%, pump frequency +1Hz, valve -2%).

[0208] Generate correction instructions:

[0209] Convert the action into actual parameter correction values:

[0210] If the current heat source power is 80%, the corrected value is 82%; if the current pump frequency is 45Hz, the corrected value is 46Hz; if the current valve opening is 75%, the corrected value is 73%.

[0211] The correction value is sent to the heating control system to adjust the corresponding equipment parameters.

[0212] The error between the predicted Q-value and the target Q-value is calculated using mean squared error (MSE). The target Q-value is: Target Q-value = Reward + γ × max(Q-value of the next state), where γ is a discount factor (e.g., 0.9). The Adam optimizer is used with a learning rate of 0.001, which decays by 0.95 every 100 training steps. The parameters of the main network are copied to the target network every 500 steps to improve training stability. Through the above process, the system can dynamically adjust the heating parameters according to the path geometry and real-time status to achieve intelligent optimization control.

[0213] In a preferred embodiment of the present invention, step 4 involves converting the heating parameter correction value generated in step 3 into a coordinated control command for heat source operating parameters, pipeline circulating water volume, and valve opening, and then sending it to the boiler group, variable frequency pump, and electric regulating valve execution terminal, including:

[0214] Step 41: Based on the heating parameter correction values ​​generated in Step 3, analyze their vector structure and separate the heat source operating parameter components, the pipeline circulating water volume components, and the valve opening components.

[0215] Step 42: For the heat source operating parameter components, combined with the thermal efficiency characteristic curve of the boiler group and the real-time load rate, generate the fuel supply rate command and start-stop priority sequence of each boiler through linear mapping.

[0216] Step 43: For the circulating water component of the pipe network, the pipe network topology resistance coefficient and real-time differential pressure feedback are superimposed and converted into a speed adjustment command for the variable frequency pump through the hydraulic balance equation.

[0217] Step 44: For the valve opening component, integrate the user-side room temperature deviation and valve flow characteristics, and generate the opening increment command for each electric control valve through segmented proportional mapping.

[0218] Step 45: The fuel supply rate command generated in step 42, the start-stop sequence, the speed adjustment command generated in step 43, and the opening increment command generated in step 44 are synchronously encapsulated into a coordinated control command set according to the time window.

[0219] Step 46: The coordinated control command set is sent to the boiler group controller, variable frequency pump driver and electric regulating valve actuator through the industrial communication protocol to trigger the terminal equipment to perform linkage execution.

[0220] In this embodiment of the invention, step 41 above involves parsing the heating parameter correction value and separating its components:

[0221] The heating parameter correction value is a structured vector (such as containing 3 types of core adjustment information). The composition rules of the vector must be clarified first: the first half of the vector corresponds to the "heat source operation parameter adjustment" (such as the heat increase / decrease ratio), the middle part corresponds to the "pipeline circulating water volume adjustment" (such as the flow rate increase / decrease ratio), and the second half corresponds to the "valve opening adjustment" (such as the opening increase / decrease ratio).

[0222] For example: if the correction value vector is "+3%, -2%, +5%", then it will be separated according to the rules as follows:

[0223] Heat source operating parameter component: +3% (indicating a 3% increase in total heat supply);

[0224] Pipeline circulating water volume component: -2% (indicating a 2% reduction in total circulating water volume);

[0225] Valve opening component: +5% (meaning the overall valve opening needs to be increased by 5%).

[0226] Step 42 above generates the fuel supply rate and start / stop priority for the boiler group:

[0227] The calculation needs to be performed in two steps, taking into account the boiler group's "thermal efficiency characteristic curve" (the efficiency variation pattern under different loads) and "real-time load rate" (the ratio of current total load to rated total load):

[0228] Allocate the heat adjustment ratio for each boiler:

[0229] Assume the boiler group consists of 3 boilers (rated power of 20MW, 30MW and 50MW respectively), with a real-time total load of 80MW (rated total load of 100MW, load rate of 80%). According to the thermal efficiency characteristic curve, the optimal efficiency range of the 3 boilers at 80% load rate is as follows: Boiler 1 (70%-90% load), Boiler 2 (60%-85% load), and Boiler 3 (75%-95% load).

[0230] The heat source operating parameter component is +3% (requiring an additional 2.4MW of heat), which is preferentially allocated to the boiler with the highest current efficiency: Boiler 3 has the highest efficiency and is allocated 1.4MW; Boiler 2 is next and is allocated 1.0MW; Boiler 1 has a lower current efficiency and is not allocated.

[0231] Calculate fuel supply rate and start-stop priority:

[0232] Fuel supply rate: Based on the boiler's "unit heat fuel consumption rate" (e.g., 80m³ of natural gas is required to generate 1MW of heat). 3 / h), boiler 3 needs to be increased by 1.4MW × 80 = 112m. 3 / h, Boiler 2 needs to be increased by 1.0MW × 80 = 80m² 3 / h generates the corresponding fuel valve opening command (e.g., increasing the natural gas valve opening from 60% to 65%).

[0233] Start-up and shutdown priority: If a standby boiler needs to be started (e.g., the total load rate exceeds 90%), it is sorted by "shortest cumulative running time and fewest recent failures". For example, standby boiler A has the shortest running time (priority 1), and boiler B is next (priority 2).

[0234] Step 43: Generate the speed adjustment command for the variable frequency pump:

[0235] The calculation process needs to combine the "pipeline topology resistance coefficient" (the resistance difference between different pipe sections, such as the main pipeline having low resistance and the branch pipeline having high resistance) and the "real-time differential pressure feedback" (the pressure difference between the supply and return water in the pipeline network, reflecting whether the current water volume meets the demand). The calculation process is as follows:

[0236] Correcting the matching relationship between circulating water volume and resistance:

[0237] The circulating water volume in the pipe network is -2% (requiring a 2% reduction in total water volume), but the resistance varies across different pipe sections: the resistance coefficient of the main pipe is 0.3 (low resistance), while the resistance coefficient of the branch pipe is 0.7 (high resistance). Therefore, priority should be given to reducing the water volume in the main pipe (where resistance is lower) to avoid insufficient flow in the branch pipes where resistance is higher. Thus, the water volume in the main pipe is reduced by 3%, and the water volume in the branch pipes is reduced by 1%.

[0238] Adjust the rotation speed based on real-time differential pressure:

[0239] The real-time differential pressure feedback is 0.3 MPa (standard differential pressure is 0.25 MPa; the current differential pressure is high, indicating excessive water flow). The variable frequency pump's speed is directly proportional to the circulating water volume (reducing speed decreases water flow), therefore:

[0240] For the main pipeline, the original speed of the frequency converter pump is 1450 r / min. The water volume needs to be reduced by 3%, and the speed needs to be adjusted to 1450 × (1-3%) = 1406.5 r / min.

[0241] For the branch pipeline corresponding to the frequency converter pump: the original speed is 1450 r / min, and the water volume needs to be reduced by 1%, so the speed is adjusted to 1450 × (1 - 1%) = 1435.5 r / min.

[0242] Step 44 above generates the opening increment command for the electric regulating valve:

[0243] It is necessary to integrate "user-side room temperature deviation" (the difference between the actual room temperature and the set value) and "valve flow characteristics" (the relationship between opening degree and flow rate, such as linear characteristics or equal percentage characteristics). The calculation process is as follows:

[0244] Adjust requirements according to room temperature deviation:

[0245] The user-side room temperature was set to 22℃, and the following data was collected in real time:

[0246] Area A: Actual room temperature 20℃ (deviation -2℃, flow rate needs to be increased);

[0247] Zone B: Actual room temperature 24℃ (deviation +2℃, flow rate needs to be reduced);

[0248] Zone C: Actual room temperature 22℃ (no deviation, no adjustment).

[0249] Calculate the valve opening increment based on valve characteristics:

[0250] The valve in Zone A has an "equal percentage characteristic" (a 10% increase in opening corresponds to a 20% increase in flow rate), requiring a 20% increase in flow rate to compensate for the -2℃ deviation, thus the opening increment is 10%; the valve in Zone B has a "linear characteristic" (a 10% decrease in opening corresponds to a 10% decrease in flow rate), requiring a 10% decrease in flow rate to reduce the +2℃ deviation, thus the opening increment is -10%; the valve in Zone C has a 0% opening increment.

[0251] Step 45 above involves synchronously encapsulating the collaborative control instruction set:

[0252] According to the time window set in step 1 (e.g., 15 minutes / window), bind the commands of each terminal to the same timestamp (e.g., "2023-12-01 08:00:00 to 08:15:00"), and check the command coordination:

[0253] If the boiler group increases its heating capacity (step 42), it is necessary to confirm whether the variable frequency pump increases the water volume synchronously (to avoid excessively high water temperature); if the valve opening is increased overall (step 44), it is necessary to confirm whether the variable frequency pump matches the water volume (to avoid a sudden drop in pressure).

[0254] For example, in the above example, the boiler increases heat, the water pump decreases water flow, and the valve is enlarged. After inspection, there is no conflict (because the total heat increase is small and the valve is only partially adjusted). It is encapsulated as a set containing "boiler fuel command, water pump speed command, and valve opening command".

[0255] In step 46 above, the instruction is sent to the terminal device:

[0256] The instruction set is sent to the corresponding terminal via industrial communication protocols (such as Modbus-RTU and Profinet):

[0257] The boiler group controller receives the "fuel supply rate (112m)" 3 / h、80m 3The system uses " / h)" and "start / stop priority (standby boiler A priority)" to control the fuel valve opening and boiler start / stop; the variable frequency pump driver receives "speed 1406.5r / min, 1435.5r / min" and adjusts the motor frequency to match the speed; the electric regulating valve actuator receives "opening +10%, -10%, 0%" and drives the valve to rotate to the target angle; after the terminal equipment executes, it provides real-time feedback on the execution results (such as actual fuel quantity, speed, and opening) for the calculation of the next round of correction values.

[0258] This invention analyzes the vector structure of heating parameter correction values ​​to precisely separate the adjustment amount into components of three dimensions: heat source, pipeline network, and valves, achieving independent and coordinated fine-grained control of each link. By combining boiler thermal efficiency curves with real-time load rate to allocate heat, it prioritizes the use of high-efficiency boilers to reduce overall energy consumption (e.g., prioritizing load allocation to boiler 3 in the example). It adjusts pump speed through pipeline resistance coefficient and differential pressure feedback to reduce pump group power consumption. It integrates user room temperature deviation with valve characteristics to adjust opening, improving the uniformity of end-point heating (e.g., increasing valve opening in areas with low room temperature). A synchronous instruction encapsulation and coordination check mechanism (step 45) ensures coordinated actions of the heat source, pipeline network, and valves, avoiding pressure fluctuations or abnormal water temperature. Industrial protocol-based instruction issuance enables automated control, reducing human intervention delays. Real-time feedback of execution results from terminal equipment provides data support for subsequent corrections, forming a closed-loop optimization. Through these optimizations, this method can improve the energy efficiency ratio, response speed, and user comfort of the heating system while reducing operation and maintenance costs.

[0259] like Figure 2 As shown, embodiments of the present invention also provide a heating intelligent control system based on deep learning and hybrid intelligent algorithms, comprising:

[0260] The data acquisition module is used to collect historical operating data, meteorological parameters, building characteristics and user behavior data of the heating system in real time; and to use a deep neural network model to perform spatiotemporal feature fusion and dimensionality reduction on the data, transforming it into a dynamic point cloud dataset in a high-dimensional feature space.

[0261] The calculation module is used to determine the point cloud boundary based on the dynamic point cloud dataset using the convex hull algorithm, and calculate the final symmetry center in the feature space based on the mirror algorithm as the original reference point; and to optimize and generate two feature lines passing through the original reference point through the particle swarm algorithm, and calculate their included angle to form a dynamically adjustable included angle region with a defined boundary.

[0262] The correction module is used to set a dynamic detection point both inside and outside the included angle area; optimize the time series connection path of the detection point based on the genetic algorithm to generate an arc-shaped closed path that characterizes the state evolution of the heating system; evaluate the geometric properties of the closed path through reinforcement learning and output the corrected heating parameter values.

[0263] The control module is used to convert the heating parameter correction values ​​into coordinated control commands for heat source operating parameters, pipeline circulating water volume and valve opening, and send them to the boiler group, variable frequency pump and electric regulating valve execution terminals.

[0264] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent heating control based on deep learning and hybrid intelligent algorithms, characterized in that, The method includes: Step 1: Real-time collection of historical operating data, meteorological parameters, building characteristics, and user behavior data of the heating system; use a deep neural network model to perform spatiotemporal feature fusion and dimensionality reduction on the data, transforming it into a dynamic point cloud dataset in a high-dimensional feature space; embedding layer transformation: through a fully connected embedding layer, the joint feature vector is mapped to a three-dimensional space, with each dimension corresponding to the physical meaning of heat source output intensity, pipeline transmission efficiency, and user demand matching degree. Step 2: Based on the dynamic point cloud dataset generated in Step 1, the convex hull algorithm is used to determine the point cloud boundary, and the final symmetry center is calculated in the feature space based on the mirror algorithm as the original reference point; two feature lines passing through the original reference point are generated by the particle swarm optimization algorithm, and their included angle is calculated to form a dynamically adjustable included angle region with a defined boundary. Step 3: Set up a dynamic detection point both inside and outside the included angle area constructed in Step 2; optimize the time-series connection path of the detection points based on a genetic algorithm to generate an arc-shaped closed path representing the state evolution of the heating system; extract the curvature distribution, total arc length, and first-to-last closure deviation value of the path as the core geometric attribute set based on the arc-shaped closed path; bind the core geometric attribute set with the real-time heating system state parameters to construct a reinforcement learning state space, including the pipe network water temperature, user-side room temperature, and valve opening feedback value; evaluate based on the reinforcement learning state space to obtain a comprehensive reward value; select the action that maximizes the comprehensive reward value based on a deep Q-network strategy to obtain the final action vector, and use it as the heating parameter correction value, including: Step 31: Using the dynamically adjustable angle region output in Step 24 as the input space, filter all data points in the current point cloud dataset that are located within this region; identify the dense distribution center of the data points within the region based on the density clustering algorithm, and output the coordinates of the labeled data point closest to the dense center as the detection point within the region; Step 32: Using the data points outside the region and the boundary ray of the region with the included angle from Step 24 as input, calculate the vertical distance from each data point outside the region to the two boundary rays; select the data point outside the region that maximizes the sum of the two vertical distances, project it onto the angle bisector of the two boundary rays, and output the coordinates of the projected point as the detection point outside the region. Step 33: Pair the detection points within the region with the detection points outside the region, bind the start timestamp of the same time window, and output the coordinate set of the dual detection points with time sequence markers; Step 4: Convert the heating parameter correction values ​​generated in Step 3 into coordinated control commands for heat source operating parameters, pipeline circulating water volume, and valve opening, and send them to the boiler group, variable frequency pump, and electric regulating valve execution terminals.

2. The intelligent heating control method based on deep learning and hybrid intelligent algorithms according to claim 1, characterized in that, The data is fused and dimensionality reduced using a deep neural network model, transforming it into a dynamic point cloud dataset in a high-dimensional feature space, including: The historical operational data, meteorological parameters, building characteristics and user behavior data collected in real time are divided into continuous time window data blocks according to a preset duration. Based on continuous time window data blocks, the static correlation between boilers, pipe networks, buildings and users in the heating system is extracted through the first deep neural network channel to obtain the spatial topological feature vector of the time window; Based on continuous time window data blocks, the trend and periodicity of heating parameters across time windows are extracted through the second deep neural network channel to obtain a dynamic evolution feature vector characterizing the state evolution of the intelligent system. Based on the spatial topological feature vector and the dynamic evolution feature vector, a weighted fusion using an attention mechanism is adopted to generate a joint feature vector that combines spatial correlation and temporal dependence. Based on the joint feature vector, it is mapped to three-dimensional space through an embedding layer, and compressed and reconstructed using an autoencoder to obtain the three-dimensional feature coordinates; The three-dimensional feature coordinates are bound to the start timestamp of the corresponding time window and the actual value of the heating load to form a labeled data point in the feature space; The labeled data points generated in all time windows are accumulated in chronological order to obtain a dynamic point cloud dataset that is updated in real time with the status of the heating system.

3. The intelligent heating control method based on deep learning and hybrid intelligent algorithms according to claim 2, characterized in that, Step 2: Based on the dynamic point cloud dataset generated in Step 1, the convex hull algorithm is used to determine the point cloud boundary, and the final center of symmetry in the feature space is calculated using the mirror algorithm as the original reference point, including: Step 21: Based on the dynamic point cloud dataset, use the convex hull algorithm to iteratively calculate the minimum convex polyhedron boundary of the point cloud in the three-dimensional feature space, and output the boundary vertex coordinate set of the current point cloud; Step 22: Based on the boundary vertex coordinate set, generate a mirror point set of the boundary vertex coordinate set in the feature space; calculate the spatial distribution similarity between the original vertex set and the mirror point set, and take maximizing the similarity as the optimization objective; iteratively adjust the position of the candidate point of the symmetry center, and when the spatial distribution similarity reaches a preset threshold, output the optimal solution as the final symmetry center coordinate, i.e., the original reference point.

4. The intelligent heating control method based on deep learning and hybrid intelligent algorithms according to claim 3, characterized in that, Two feature lines passing through the original reference point are generated using a particle swarm optimization algorithm, and their included angle is calculated to form a dynamically adjustable included angle region with defined boundaries, including: Step 23: Based on the boundary vertex coordinate set and the final symmetry center coordinates, perform the following operations using the particle swarm optimization algorithm: Initialize the parameters of two lines passing through the original reference point; minimize the sum of the weighted distances from the boundary vertex to the two lines as the optimization objective; iteratively update the line parameters until convergence, and output the optimized equations of the two feature lines; Step 24: Based on the optimized equations of the two characteristic lines, calculate the spatial angle between the two lines, and output the dynamic control angle region containing the key state points of the heating system, with the original reference point as the vertex and the rays on both sides of the angle as the boundary.

5. The intelligent heating control method based on deep learning and hybrid intelligent algorithms according to claim 4, characterized in that, Based on a genetic algorithm, the time-series connection paths of the detection points are optimized to generate an arc-shaped closed path representing the state evolution of the heating system, including: Step 34: Based on the set of coordinates of the two detection points, arrange the two detection points of the consecutive time windows in the order of timestamps to form the initial path point sequence; Step 35: Using the initial path point sequence as a chromosome, perform the following operations using a genetic algorithm: The polar coordinate parameters of the path point sequence are used as genes; With the optimization objectives of maximizing the overall curvature smoothness of the path and minimizing the total arc length, the selection, intersection, and mutation processes are iteratively executed until the final optimized path point sequence that satisfies the arc closure constraint is generated; Step 36: Connect the final optimized path point sequence end to end and fit it into a continuous arc curve in the three-dimensional feature space, outputting an arc-shaped closed path that characterizes the state evolution of the heating system.

6. The intelligent heating control method based on deep learning and hybrid intelligent algorithms according to claim 5, characterized in that, Step 4: Convert the heating parameter correction values ​​generated in Step 3 into coordinated control commands for heat source operating parameters, pipeline circulating water volume, and valve opening, and send them to the boiler group, variable frequency pump, and electric regulating valve execution terminals, including: Step 41: Based on the heating parameter correction values ​​generated in Step 3, analyze their vector structure and separate the heat source operating parameter components, the pipeline circulating water volume components, and the valve opening components. Step 42: For the heat source operating parameter components, combined with the thermal efficiency characteristic curve of the boiler group and the real-time load rate, generate the fuel supply rate command and start-stop priority sequence of each boiler through linear mapping. Step 43: For the circulating water component of the pipe network, the pipe network topology resistance coefficient and real-time differential pressure feedback are superimposed and converted into a speed adjustment command for the variable frequency pump through the hydraulic balance equation. Step 44: For the valve opening component, integrate the user-side room temperature deviation and valve flow characteristics, and generate the opening increment command for each electric control valve through segmented proportional mapping. Step 45: The fuel supply rate command generated in step 42, the start-stop sequence, the speed adjustment command generated in step 43, and the opening increment command generated in step 44 are synchronously encapsulated into a coordinated control command set according to the time window. Step 46: The coordinated control command set is sent to the boiler group controller, variable frequency pump driver and electric regulating valve actuator through the industrial communication protocol to trigger the terminal equipment to perform linkage execution.

7. A heating intelligent control system based on deep learning and hybrid intelligent algorithms, characterized in that, The system is used to perform the method as described in any one of claims 1 to 6, the system comprising: The data acquisition module is used to collect historical operating data, meteorological parameters, building characteristics and user behavior data of the heating system in real time; and to use a deep neural network model to perform spatiotemporal feature fusion and dimensionality reduction on the data, transforming it into a dynamic point cloud dataset in a high-dimensional feature space. The calculation module is used to determine the point cloud boundary based on the dynamic point cloud dataset using the convex hull algorithm, and calculate the final symmetry center in the feature space based on the mirror algorithm as the original reference point; and to optimize and generate two feature lines passing through the original reference point through the particle swarm algorithm, and calculate their included angle to form a dynamically adjustable included angle region with a defined boundary. The correction module is used to set a dynamic detection point both inside and outside the included angle area; optimize the time series connection path of the detection point based on the genetic algorithm to generate an arc-shaped closed path that represents the state evolution of the heating system; evaluate the geometric properties of the closed path through reinforcement learning and output the corrected heating parameter values. The control module is used to convert the heating parameter correction values ​​into coordinated control commands for heat source operating parameters, pipeline circulating water volume and valve opening, and send them to the boiler group, variable frequency pump and electric regulating valve execution terminals.

Images