Intelligent heat transfer recovery method for heat supply system based on LSTM neural network

Abstract

The invention relates to the technical field of heat supply system recovery, and discloses a heat supply system heat transfer intelligent recovery method based on an LSTM neural network. The method comprises the following steps: firstly, collecting thermodynamic parameter time sequences of water supply temperature, water return temperature, pipe section flow, environment temperature and the like of each monitoring node of the heat supply pipe network; performing standardization processing on the time sequences, eliminating influence of different physical quantity lines, and generating a standardized thermodynamic time sequence; then, the heat supply pipe network is divided into a plurality of functional areas according to the topological structure and the thermal characteristics, and each area comprises a plurality of associated monitoring nodes; the LSTM neural network is used for deeply analyzing processed data and divided areas, intelligent heat transfer recovery of the heat supply system is achieved, the dynamic change rule of parameters can be accurately captured, the method adapts to complex working conditions, heat transfer recovery is more targeted and reasonable, heat energy waste is reduced, and the operation efficiency of the heat supply system is improved.

Description

Technical Field

[0001] This invention relates to the field of heat recovery technology for heating systems, specifically to an intelligent heat transfer recovery method for heating systems based on LSTM neural networks. Background Technology

[0002] During winter heating, the energy consumption of heating systems accounts for a large proportion, and how to achieve efficient utilization and recovery of heat energy has always been a focus of industry attention. Traditional heat transfer and recovery methods in heating systems mostly rely on manual experience or simple mathematical models for regulation, which are difficult to cope with complex and ever-changing actual operating conditions.

[0003] Heating network systems are complex, containing numerous monitoring nodes, each with its own thermodynamic parameters that change dynamically over time. Parameters such as supply water temperature, return water temperature, pipe flow rate, and ambient temperature interact with each other and are further influenced by factors including building heat dissipation, user heating habits, and climate conditions. Traditional methods often fail to accurately capture the dynamic changes in these parameters, resulting in low heat transfer and recovery efficiency and wasted heat energy. Traditional heat transfer and recovery methods are not sufficiently rational in their regional division of heating network systems. They typically divide the systems simply based on geographical location or pipeline route, ignoring the thermal connections and functional differences between different areas. This makes it difficult to develop effective recovery strategies tailored to the characteristics of different areas, further impacting the overall recovery efficiency. With the acceleration of urbanization, the scale of heating systems is constantly expanding, and the requirements for the accuracy and intelligence of heat transfer and recovery are becoming increasingly higher. Traditional methods can no longer meet the needs of modern heating systems for energy saving and efficient operation. There is a need for an intelligent recovery method that can combine advanced algorithms and system characteristics to improve the effectiveness and adaptability of heat transfer and recovery. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent heat transfer recovery method for heating systems based on LSTM neural networks, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a method for intelligent heat recovery in heating systems based on LSTM neural networks, the method comprising: Collect time series of thermodynamic parameters of each monitoring node in the heating pipeline system. The thermodynamic parameters include at least the supply water temperature, return water temperature, pipeline flow rate and ambient temperature. The thermodynamic parameter time series is standardized to eliminate the influence of different physical dimensions and generate a standardized thermodynamic time series. The heating network system is divided into multiple functional areas according to its topology and thermal characteristics, and each functional area contains several interconnected monitoring nodes.

[0006] Preferably, the method further includes: An independent LSTM feature extraction network is constructed for each functional region. The LSTM feature extraction network includes forget gate, input gate and output gate structures. The standardized thermodynamic time series of each monitoring node in the corresponding functional area is input into the LSTM feature extraction network, and key time step features are filtered through a gating mechanism. Cell state vectors at each time step are extracted as regional dynamic thermodynamic feature vectors.

[0007] Preferably, the method further includes: Establish a cross-regional feature fusion module to receive regional dynamic thermal feature vectors from all functional areas; The correlation degree of feature vectors of different functional regions is calculated by an attention weighting mechanism, and the correlation degree is determined based on vector cosine similarity and feature entropy value. The regional dynamic thermal feature vectors are weighted and fused based on the correlation degree to generate a network-level spatiotemporal fusion feature matrix.

[0008] Preferably, the method further includes: A heat transfer potential assessment model is constructed, which adopts a stacked LSTM network architecture. The network-level spatiotemporal fusion feature matrix is ​​input into the heat transfer potential assessment model, and the feature dependencies are analyzed layer by layer through multi-layer LSTM units. Output the predicted heat loss distribution and recoverable heat potential energy gradient of each pipeline node within the future time window.

[0009] Preferably, the method further includes: Design a heat recovery strategy generator to receive the predicted heat loss distribution and the recoverable heat potential energy gradient; Heat recovery constraints are established based on the pipe section material characteristics, insulation layer status parameters, and historical thermal efficiency data. Under the condition of satisfying the heat recovery constraints, a heat recovery execution strategy including pump and valve regulation schemes and heat exchange paths is generated.

[0010] Preferably, the method further includes: Establish a local strategy optimization module to receive the heat recovery execution strategy; Extract the key control device identifiers and their operating parameter sequences involved in the strategy; By combining the load rate, energy consumption curve and equipment health indicators in the real-time operation log of the equipment, the sequence of operating parameters is dynamically optimized. Output an optimized set of equipment-level heat recovery operation instructions.

[0011] Preferably, the method further includes: The configuration strategy verification unit receives the equipment-level heat recovery operation instruction set and simulates the thermodynamic response of the equipment-level heat recovery operation instruction set in the heating network system. The feasibility of the strategy is verified by comparing the simulation results with the preset safety threshold, and a verification report is generated.

[0012] Preferably, the method further includes: Design a dynamic adjustment mechanism to receive the verification report and the time series of newly added thermodynamic parameters collected in real time; When the validation report shows that the strategy deviates from the expected results, online fine-tuning of the LSTM feature extraction network parameters is triggered. The weight coefficients of the heat transfer potential assessment model are updated based on incremental learning.

[0013] Preferably, the method further includes: Deploy the strategy execution interface, receive the finally verified device-level heat recovery operation instruction set, and convert the device-level heat recovery operation instruction set into device control protocol instructions; The system sends a sequence of timestamped control commands to each pump and valve actuator via an industrial bus.

[0014] Preferably, the method further includes: Construct a heat recovery efficiency monitoring system to continuously collect the thermodynamic parameters of the pipeline network after performing heat recovery operations; Calculate the deviation between the actual heat recovery efficiency and the predicted value, and generate performance evaluation indicators; The performance evaluation index is fed back into the heat transfer potential evaluation model as training data increment.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By collecting time series data of thermodynamic parameters from various monitoring nodes in the heating network system, a comprehensive and continuous data foundation is provided for subsequent intelligent analysis. These parameters, covering supply water temperature, return water temperature, pipe flow rate, and ambient temperature, can fully reflect the real-time operating status and thermodynamic changes of the heating system. Standardizing the time series of thermodynamic parameters eliminates interference from different physical units, enabling analysis and calculation of various parameters under a unified standard. This improves data comparability and consistency, allowing LSTM neural networks to process data more accurately, reducing errors caused by differences in data units, and thus more accurately uncovering the intrinsic correlations and patterns of change between parameters.

[0016] The heating network system is divided into multiple functional zones based on its topology and thermal characteristics. Each zone contains several interconnected monitoring nodes. This division method fully considers the system's inherent structure and thermal characteristics. In this way, targeted heat transfer and recovery analyses can be performed based on the characteristics of different functional zones, avoiding analytical biases caused by regional differences in traditional overall analysis. This makes the regulation and control of heat transfer and recovery more targeted and rational. LSTM neural networks inherently possess the advantage of processing time-series data, effectively capturing the long-term dependencies of thermodynamic parameters over time. Applying them to heat transfer and recovery in heating systems allows for in-depth analysis of the changing trends and interactions of various parameters at different time scales, thereby more accurately predicting the heat transfer process. This provides a scientific reference for formulating heat transfer and recovery schemes, making the recovery process more consistent with the actual operating conditions of the system and adaptable to complex and changing working conditions. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent heat transfer recovery method for heating systems based on LSTM neural networks described in this invention. Figure 2 A flowchart for LSTM feature extraction and dynamic thermal feature generation; Figure 3 A flowchart for cross-regional feature fusion and spatiotemporal matrix generation; Figure 4 A flowchart for generating heat recovery strategies and applying constraints; Figure 5 The flowchart for dynamic adjustment and online model updates. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides a method for intelligent heat recovery in a heating system based on an LSTM neural network, the method comprising: The thermodynamic parameters of each monitoring node in the heating pipeline system are collected in time series. These thermodynamic parameters include at least the supply water temperature, return water temperature, pipe flow rate, and ambient temperature. The monitoring nodes are distributed at key locations in the heating pipeline network, including heat source outlets, main pipe segment points, branch inlets, and user-end interfaces. Parameters are collected every 5 minutes to form a continuous, multi-dimensional time series dataset.

[0020] The thermodynamic parameter time series is standardized to eliminate the influence of different physical dimensions, generating a standardized thermodynamic time series. The heating network system is divided into multiple functional regions according to its topology and thermodynamic characteristics, with each functional region containing several interconnected monitoring nodes. The topology division is based on the network's connectivity, dividing the main pipeline network, branch network, and end-user network into independent regions. The thermodynamic characteristic division is based on heat load density, pipe diameter, and insulation level, grouping pipe segments with similar heat dissipation characteristics into the same region, ultimately forming 5-8 functional regions, each containing 10-20 monitoring nodes.

[0021] Example 1: See Figure 2 An independent LSTM feature extraction network is constructed for each functional region, containing forget gates, input gates, and output gates. The forget gate determines whether to retain information from historical cell states. When processing short-term abnormal fluctuations in heating system data, the forget gate weakens the impact of this information; however, for long-term stable temperature trends, the forget gate retains relevant information. The input gate updates the cell state by filtering valid data from the currently input standardized thermodynamic time series and combining it with processed historical information to form new cell state content. The output gate generates output information for subsequent processing based on the current cell state and hidden states. This output information reflects the key thermodynamic features within the functional region at that time step.

[0022] The standardized thermodynamic time series data of each monitoring node within the corresponding functional area are input into the LSTM feature extraction network, and key time step features are filtered through a gating mechanism. During the input process, the standardized data of each monitoring node enters the network sequentially according to time, and the network analyzes the data at each time step. For example, during the period of rapid increase in heating load in the morning, the input gate focuses on the changes in water supply temperature and flow rate, incorporating this key information into the cell state; while for the period of lower load and smaller fluctuations at night, the forget gate appropriately ignores some redundant information, reducing unnecessary computation. Through this gating mechanism, the network can automatically identify and retain time step data that is important for thermodynamic feature extraction, filtering out noise and irrelevant information.

[0023] Cell state vectors at each time step are extracted as regional dynamic thermal feature vectors. The cell state vector is the part of the LSTM network used for long-term information storage, containing all important thermal information from the initial time step to the current time step. Each time step's cell state vector encompasses comprehensive features of parameters such as supply water temperature, return water temperature, pipe flow rate, and ambient temperature within the functional region. For example, at a certain time step, if the supply water temperature is significantly higher than the return water temperature and the pipe flow rate is at a high level, the cell state vector will reflect this high-load thermal state through numerical combinations. Conversely, when the ambient temperature drops sharply, the cell state vector will adjust accordingly, reflecting the impact of environmental factors on the regional thermal condition. These cell state vectors are arranged chronologically, forming a sequence that comprehensively reflects the dynamic changes in the thermal state of the functional region over time, providing fundamental data for subsequent cross-regional feature fusion and heat transfer potential assessment.

[0024] In actual operation, the LSTM feature extraction network for each functional area processes the data independently, ensuring that the thermal characteristics of different areas are accurately captured. For example, for the functional area where the backbone network is located, its LSTM network will focus more on the temperature attenuation characteristics under high flow and long-distance transmission; while for the functional area where the end-user network is located, the network will focus on extracting the temperature distribution characteristics under low flow and multi-branch conditions. This personalized processing approach for different functional areas makes the dynamic thermal feature vector of each area highly targeted and representative.

[0025] Furthermore, the structural parameters of the LSTM feature extraction network are adjusted according to the size of the functional region and the number of monitoring nodes. For functional regions with a large number of monitoring nodes, the hidden layer dimension of the network is increased accordingly to accommodate more feature information; while for regions with fewer nodes, the network structure is relatively simplified, improving computational efficiency while ensuring extraction performance. This flexible configuration ensures that each LSTM feature extraction network can adapt to the characteristics of its corresponding functional region, efficiently and accurately completing the feature extraction task.

[0026] The extracted regional dynamic thermodynamic feature vectors are temporarily stored, awaiting further processing in the cross-regional feature fusion module. These vectors not only contain the thermodynamic features of a single time step, but also implicitly contain the correlations between different time steps, such as the temperature change trends of adjacent time steps and the periodicity of flow fluctuations. This information is crucial for a comprehensive understanding of the thermodynamic behavior of the heating system.

[0027] Example 2: See Figure 3A cross-regional feature fusion module is established to receive regional dynamic thermal feature vectors from all functional regions. This module first preprocesses the feature vectors input from each region, including checking the consistency of vector dimensions and timestamp alignment. When the feature vector dimensions differ between different functional regions, a linear transformation is used to unify them to the same dimension. For cases of mismatched timestamps, interpolation is performed based on the feature values ​​of adjacent time points to ensure that the feature vectors of all regions remain synchronized over time. The processed feature vectors are temporarily stored in a buffer, awaiting the next step of correlation calculation.

[0028] The correlation degree between feature vectors of different functional regions is calculated using an attention weighting mechanism. This correlation degree is determined based on vector cosine similarity and feature entropy. In the calculation process, firstly, cosine similarity analysis is performed on the feature vectors of any two functional regions to determine their directional consistency in the feature space. The closer the directions of the two vectors, the higher the similarity in the corresponding regions' thermal change trends. Simultaneously, the feature entropy value of each feature vector is calculated. A lower entropy value indicates more concentrated information in the feature vector, reflecting a stronger regularity in the region's thermal state; a higher entropy value indicates more dispersed information in the feature vector, suggesting potential fluctuations in the region's thermal state. Combining these two indicators, a comprehensive evaluation of the correlation degree between different functional regions is formed, creating a correlation matrix where each element corresponds to the degree of correlation between two regions.

[0029] The dynamic thermal feature vectors of the regions are weighted and fused based on their correlation to generate a network-level spatiotemporal fusion feature matrix. The fusion process proceeds sequentially along time steps. At each time step, the feature vectors of all functional regions are assigned corresponding weights according to their correlation. Regions with higher correlation have greater weights in the fusion process. For example, when two adjacent regions have a high correlation, their feature information will be referenced more during the fusion process to reflect their thermal interaction. After weighted calculation, a comprehensive feature vector is formed at each time step. These vectors are arranged in chronological order and combined with the spatial location information of each region to ultimately constitute a network-level spatiotemporal fusion feature matrix containing both temporal and spatial dimensions. This matrix fully preserves the thermal characteristics of different regions at different time points and their interrelationships.

[0030] A hot transfer potential assessment model is constructed, employing a stacked LSTM network architecture. This architecture consists of multiple layers of LSTM units connected sequentially. Each layer contains forget gates, input gates, and output gates, and the output of each layer serves as the input to the next. The bottom layers have a relatively large number of LSTM units to capture detailed information from the feature matrix; the upper layers have progressively fewer LSTM units, used for abstracting and integrating the features output from the bottom layers. A mechanism to prevent overfitting is implemented between layers by randomly and temporarily discarding connections of some neurons, improving the model's adaptability to unseen data.

[0031] The spatiotemporal fusion feature matrix at the pipeline network level is input into the heat transfer potential assessment model, and feature dependencies are analyzed layer by layer using multi-layer LSTM units. The bottom-layer LSTM units mainly handle short-term dependencies in the feature matrix, such as rapid changes in thermal characteristics of different regions within a few adjacent time steps; the middle-layer LSTM units are responsible for analyzing medium-term dependencies, such as periodic changes in thermal characteristics at different times of the day; and the top-layer LSTM units focus on long-term dependencies, capturing the evolution of thermal trends over several days. Through this layer-by-layer analysis, the model can gradually gain a deeper understanding of the complex thermal characteristic relationships in the pipeline network system, including the spatial influence between different regions and the dynamic changes of the same region at different times.

[0032] This program outputs the predicted heat loss distribution and recoverable heat potential energy gradient for each network node within a future time window. The predicted heat loss distribution, for each network node, provides the heat loss situation for that node over a future period, expressed numerically as the amount of heat loss per unit time. The recoverable heat potential energy gradient reflects the differences in heat potential energy between different nodes, represented as a vector indicating the direction and intensity of heat potential energy change from high to low, providing clear guidance for subsequent heat recovery strategy development. The prediction results will be presented in the form of data tables and graphs, with the graphical representation providing a more intuitive understanding of the spatial distribution of heat loss and the changing trend of heat potential energy.

[0033] During model operation, the allocation of computing resources is automatically adjusted based on the size of the input spatiotemporal fusion feature matrix to ensure that computation can be completed within a specified time while maintaining prediction accuracy. Simultaneously, the model performs internal validation of the output results. If significant anomalies are detected in the predicted values, some LSTM units are recalculated to correct any potential deviations.

[0034] Example 3: See Figure 4A heat recovery strategy generator was designed to receive the predicted heat loss distribution and the recoverable heat potential energy gradient. This generator comprises a data receiving module and a strategy generation core module. The data receiving module is responsible for converting the input predicted data into a unified format, ensuring data integrity and temporal consistency. The strategy generation core module internally stores the basic topology information of the heating system, including the connection relationships of each pipe segment, the installation locations of equipment, and the spatial layout of the pipe network. This information provides a spatial reference framework for strategy generation. In the initial stage of strategy generation, the predicted heat loss distribution is first divided into regions, marking areas with higher heat loss concentrations as key recovery areas. Simultaneously, based on the direction of the recoverable heat potential energy gradient, possible heat transfer paths are initially planned.

[0035] Heat recovery constraints are established based on the pipe material characteristics, insulation layer condition parameters, and historical thermal efficiency data. Pipe material characteristics include thermal conductivity, high-temperature resistance, and compressive strength. Different materials result in different heat loss rates during heat transfer; for example, the difference in thermal conductivity between metal and plastic pipes directly affects the rate of heat loss. Therefore, a maximum allowable heat transfer amount needs to be set according to the material. Insulation layer condition parameters cover the insulation layer's thickness, thermal conductivity, and aging degree. The aging degree is represented by the periodically measured insulation performance degradation rate. When the insulation layer performance is poor, the heat transfer intensity in that area will be appropriately reduced to minimize additional losses. Historical thermal efficiency data includes system thermal efficiency values ​​under different operating conditions over a past period. Statistical analysis is used to determine a reasonable efficiency range, which serves as a reference benchmark for the constraints.

[0036] Under the constraint of heat recovery, a heat recovery execution strategy is generated, including pump and valve regulation schemes and heat exchange paths. The pump and valve regulation scheme specifies the operating frequency of each circulating water pump and the opening degree of each regulating valve. Frequency regulation aims to meet the flow requirements of the pipe section while avoiding pump operation in inefficient ranges. The opening degree of the regulating valves is set based on the target pressure difference to ensure that the pressure within the pipe section remains stable within a reasonable range. The planning of the heat exchange path is mainly based on the recoverable heat potential energy gradient, prioritizing paths with larger heat potential energy gradients, while also considering the path length and the load-bearing capacity of the pipe section to avoid excessive heat loss due to excessively long paths or overloaded pipe sections. The execution strategy also includes the execution time points for each operation. These time points are determined based on the load variation patterns of the heating system, typically adjusting during periods of low load fluctuation to minimize the impact on system stability.

[0037] A local strategy optimization module is established to receive heat recovery execution strategies. This module is connected to the real-time monitoring system of the heating system, enabling it to acquire real-time operating data of each pump and valve, including current operating parameters, load status, and fault warning information. The module contains an equipment characteristic database that stores performance curves, rated parameters, and operating limits for each type of pump and valve. This data provides equipment-level reference for strategy optimization.

[0038] The key control equipment identifiers and their operating parameter sequences involved in the strategy are extracted. Key control equipment includes the main circulating water pump, zone circulating water pumps, electric regulating valves, and plate heat exchangers. Each piece of equipment has a unique identifier for accurate location within the system. The operating parameter sequences are arranged chronologically and contain the target operating parameters of each equipment at different points in time, such as the target frequency of the circulating water pump and the target opening degree of the regulating valve. These parameter sequences constitute the core of the strategy.

[0039] By combining load rate, energy consumption curves, and equipment health indicators from the equipment's real-time operation logs, the sequence of operating parameters is dynamically optimized. Load rate refers to the ratio of the equipment's current load to its rated load. When the load rate is too high, the equipment's operating parameters are appropriately reduced to avoid overload. The energy consumption curve reflects the energy consumption variation pattern of the equipment under different operating parameters; during optimization, the parameter combination with lower energy consumption is selected with reference to the energy consumption curve. Equipment health indicators include the equipment's vibration amplitude, bearing temperature, and operating noise. When a certain indicator approaches the warning value, the parameters are adjusted to reduce equipment wear. The optimization process adopts a gradual adjustment method, with each adjustment controlled within a certain range to avoid sudden parameter changes impacting the system.

[0040] The system outputs an optimized set of equipment-level heat recovery operation instructions. The instruction set uses a standardized data format, including instruction number, equipment identifier, execution time, target parameter name, and target parameter value. Each instruction is accompanied by a checksum to ensure it is not tampered with during transmission. The instruction set is sent to the control systems of each device via a communication interface, enabling precise control of the equipment.

[0041] Throughout the process, it is necessary to calculate the real-time energy efficiency ratio of the equipment. The calculation formula is as follows:

[0042] in, This indicates the real-time energy efficiency ratio of the equipment. This indicates the amount of heat recovered per unit time. This indicates the energy consumption of the equipment during that time period. By calculating the energy efficiency ratio in real time, the operating efficiency of the equipment under the current operating parameters is determined. When the energy efficiency ratio falls below the set reasonable range, the local strategy optimization module is triggered to re-adjust the parameters until the energy efficiency ratio returns to a reasonable level. This feedback mechanism based on the real-time energy efficiency ratio ensures that the equipment always operates in a highly efficient state, improving the overall economic efficiency of the heat recovery process.

[0043] Example 4: See Figure 5 The system includes a strategy verification unit that receives equipment-level heat recovery operation command sets and simulates the thermodynamic response of these commands within the heating network system. This unit incorporates a hydraulic-thermal coupling simulation model of the network, containing physical parameters for each pipe segment, such as pipe diameter, length, roughness, and connection methods for each node. When the operation command set is input, the simulation model sequentially simulates the actions of each pump and valve according to the time sequence specified in the command. For example, when a regulating valve is adjusted from 30% to 50%, the model calculates the resulting change in flow rate within the pipe segment and derives the impact on the temperature and pressure of upstream and downstream nodes. During the simulation, the thermodynamic parameters of each monitoring node are recorded every minute, including supply water temperature, return water temperature, pressure, and flow rate, forming a complete simulation response sequence.

[0044] By comparing simulation results with preset safety thresholds, the feasibility of the strategy is verified and a verification report is generated. The preset safety thresholds cover multiple aspects, such as the maximum allowable pressure value of the pipeline network; if the pressure at a node exceeds this value in the simulation, it is considered a risk of pipe burst. The minimum flow rate threshold is also included; if the flow rate of a pipe section is lower than this value, it may lead to localized overheating and should be marked as an anomaly. The verification report will list all parameters exceeding the thresholds in detail, such as "the water supply temperature at node A reached 85℃ at the 15th minute, exceeding the safety threshold of 80℃," and will also explain the source of the corresponding operating instruction, such as "caused by the frequency of circulating water pump P2 being increased from 40Hz to 45Hz." The report will also analyze the duration and scope of the threshold exceedances. If an anomaly lasts only 1 minute and its impact is limited to a single branch line, it is considered a minor risk; if it lasts for more than 5 minutes and involves the main pipeline, it is considered a high risk.

[0045] A dynamic adjustment mechanism is designed to receive verification reports and newly acquired thermodynamic parameter time series data in real time. This mechanism includes a data preprocessing module to clean the newly acquired thermodynamic parameter time series, removing obvious sensor fault data, such as instantaneously fluctuating temperature values, and then smoothing the data to reduce volatility. The dynamic adjustment mechanism is activated when the verification report indicates that the strategy has a high risk or slight risk accumulated more than three times. Simultaneously, if the deviation between the real-time acquired parameters and the predicted values ​​exceeds a certain range, for example, if the return water temperature deviation is greater than 3°C for three consecutive time steps, the adjustment process will also be triggered.

[0046] When the validation report shows that the strategy deviates from the expected results, online fine-tuning of the LSTM feature extraction network parameters is triggered. The fine-tuning process targets the weights and biases in the network, prioritizing adjustments to the neuron connection weights related to anomalous parameters. For example, if the validation report shows multiple instances of abnormal return water temperature, the LSTM unit parameters for processing the return water temperature sequence will be fine-tuned. Fine-tuning uses small batches of data for training, selecting newly added data from the most recent hour each time, and iteratively correcting the parameters to avoid excessively large adjustments in a single instance. During the adjustment process, changes in the feature vector output by the network are monitored in real time. If the fluctuation range of the feature vector exceeds the normal range, fine-tuning is paused and the network reverts to the previous state.

[0047] The weight coefficients of the heat transfer potential assessment model are updated using an incremental learning approach. Incremental learning only uses newly added thermodynamic parameter data, without retraining the entire model, to reduce computational resource consumption. New data is divided into time windows, each containing a 2-hour parameter sequence, and is sequentially input into the model using a sliding window mechanism. When updating the weight coefficients, a weighted average method is used, combining the newly calculated weights with the original weights, with a larger update magnitude for weight coefficients corresponding to recent data. For example, the weight update coefficient is set to 0.3 for new data within 3 days; and 0.1 for data from 3 to 7 days. After the update, the model's predicted values ​​are compared with the actual values ​​to determine if the weight coefficients are reasonable. If the deviation increases, the update magnitude is reduced and the model is recalculated.

[0048] The dynamic adjustment mechanism also includes an adjustment log recording function, which records in detail the triggering conditions for each adjustment, the name of the adjusted parameter, the values ​​before and after the adjustment, and the effect after the adjustment. These logs provide raw data for subsequent analysis of model stability and optimization of adjustment strategies. At the same time, the mechanism will periodically (e.g., daily) summarize the adjustment effect. If it finds that a certain type of adjustment is frequently triggered, it will suggest that the initial structural design of the LSTM feature extraction network or the architecture of the hot transfer potential evaluation model may need to be re-examined.

[0049] Example 5: Deploy a strategy execution interface to receive the finally verified device-level heat recovery operation instruction set and convert it into device control protocol instructions. This interface includes a protocol conversion module that supports multiple industrial control protocols, selecting the appropriate protocol for different types of equipment. For example, for circulating water pumps, the Modbus-RTU protocol is used for instruction conversion, converting the pump frequency parameters in the operation instruction set into register values ​​conforming to the protocol format; for intelligent control valves, the Profinet protocol is used, encapsulating the valve opening parameters into a data frame structure. During the conversion process, the integrity of the instructions is verified, checking whether necessary information such as device identifiers, execution times, and parameter values ​​are included. If any information is missing, an error message is returned and retransmission is required.

[0050] Control command sequences with timestamps are sent to each pump and valve actuator via an industrial bus. The industrial bus employs a dual-redundancy design, with the primary and backup buses operating simultaneously. If communication is interrupted on the primary bus, it automatically switches to the backup bus to ensure uninterrupted command transmission. The control command sequence is arranged chronologically according to execution time, and each command includes a timestamp accurate to milliseconds. Upon receiving the command, the actuator will perform the corresponding operation at the specified time based on the timestamp. For example, if the timestamp of a command is 10:05:30.120, the actuator will adjust the corresponding valve opening to the value required by the command at that time. During transmission, the transmission status of the commands is monitored in real time. If a command fails to be transmitted three times consecutively, it is marked as abnormal and retransmitted with priority. The reason for the failure is recorded, such as low signal strength or device offline.

[0051] A heat recovery efficiency monitoring system was constructed to continuously collect the thermodynamic parameters of the pipeline network after heat recovery operations. This system comprises a sensor network distributed across various monitoring nodes. The sensor types are determined based on the monitored parameters; for example, platinum resistance temperature sensors are used to collect supply and return water temperatures, electromagnetic flow meters are used to measure pipe flow rates, and ambient temperature is obtained through outdoor temperature and humidity sensors. The data acquisition frequency is consistent with the control command execution frequency to ensure that each operational step has corresponding parameter records. The collected data is transmitted wirelessly to a data center and stored in a time-series database, indexed by device identifiers and timestamps for easy subsequent querying and analysis.

[0052] The deviation between the actual heat recovery efficiency and the predicted value is calculated to generate performance evaluation indicators. The actual heat recovery efficiency is calculated by comparing the heat loss before and after the operation. For example, if the heat loss of a pipe section before the heat recovery operation is Q1, and the heat loss after the operation is Q2, then the actual recovered heat is Q1-Q2. The predicted value is obtained from the output of the heat transfer potential assessment model, and the deviation is the ratio of the difference between the actual recovered heat and the predicted recovered heat to the predicted recovered heat. Performance evaluation indicators also include the heat recovery duration, i.e., the time from operation execution to heat loss stabilizing at the target value; and the equipment response speed, i.e., the time from receiving the command to reaching the target parameter. These indicators are presented in tabular form, containing the specific values ​​and corresponding time ranges for each evaluation item.

[0053] The performance evaluation metrics are fed back to the heat transfer potential assessment model as incremental training data. The feedback process employs a weighted feedback mechanism, assigning different weights based on the importance of the metrics; for example, the weight of heat recovery efficiency deviation is higher than that of equipment response speed. More recent performance evaluation metrics have higher weights; for instance, data from the past 24 hours has twice the weight of data from the past 7 days. The feedback data is first standardized and converted to the format required for model training before being input into the model's training module to update the model's parameters. During the update process, the model's learning from historical data is retained; parameters are adjusted only based on the newly added feedback data, ensuring that the model adapts to new situations without losing previously learned patterns.

[0054] The heat recovery efficiency monitoring system also has an anomaly alarm function. When the collected thermodynamic parameters exceed the normal operating range, such as a sudden drop in water supply temperature of more than 10°C or a sudden decrease in flow rate to zero in a pipe section, the system will immediately issue an alarm signal and suspend the sending of relevant heat recovery operation commands until the fault is resolved. The alarm information includes the name, value, time of occurrence, and possible impact range of the abnormal parameter, and will remind maintenance personnel to handle the situation through audible and visual alarms and remote notifications.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent heat transfer recovery in a heating system based on an LSTM neural network, characterized in that, include: Collect time series of thermodynamic parameters of each monitoring node in the heating pipeline system. The thermodynamic parameters include at least the supply water temperature, return water temperature, pipeline flow rate and ambient temperature. The thermodynamic parameter time series is standardized to eliminate the influence of different physical dimensions and generate a standardized thermodynamic time series. The heating network system is divided into multiple functional areas according to its topology and thermal characteristics, and each functional area contains several interconnected monitoring nodes.

2. The intelligent heat transfer recovery method for heating systems based on LSTM neural networks according to claim 1, characterized in that, Also includes: An independent LSTM feature extraction network is constructed for each functional region. The LSTM feature extraction network includes forget gate, input gate and output gate structures. The standardized thermodynamic time series of each monitoring node in the corresponding functional area is input into the LSTM feature extraction network, and key time step features are filtered through a gating mechanism. Cell state vectors at each time step are extracted as regional dynamic thermodynamic feature vectors.

3. The intelligent heat transfer recovery method for heating systems based on LSTM neural networks according to claim 2, characterized in that, Also includes: Establish a cross-regional feature fusion module to receive regional dynamic thermal feature vectors from all functional areas; The correlation degree of feature vectors of different functional regions is calculated by an attention weighting mechanism, and the correlation degree is determined based on vector cosine similarity and feature entropy value. The regional dynamic thermal feature vectors are weighted and fused based on the correlation degree to generate a network-level spatiotemporal fusion feature matrix.

4. The intelligent heat transfer recovery method for heating systems based on LSTM neural networks according to claim 3, characterized in that, Also includes: A heat transfer potential assessment model is constructed, which adopts a stacked LSTM network architecture. The network-level spatiotemporal fusion feature matrix is ​​input into the heat transfer potential assessment model, and the feature dependencies are analyzed layer by layer through multi-layer LSTM units. Output the predicted heat loss distribution and recoverable heat potential energy gradient of each pipeline node within the future time window.

5. The intelligent heat transfer recovery method for a heating system based on an LSTM neural network according to claim 4, characterized in that, Also includes: Design a heat recovery strategy generator to receive the predicted heat loss distribution and the recoverable heat potential energy gradient; Heat recovery constraints are established based on the pipe section material characteristics, insulation layer status parameters, and historical thermal efficiency data. Under the condition of satisfying the heat recovery constraints, a heat recovery execution strategy including pump and valve regulation schemes and heat exchange paths is generated.

6. The intelligent heat transfer recovery method for a heating system based on an LSTM neural network according to claim 5, characterized in that, Also includes: Establish a local strategy optimization module to receive the heat recovery execution strategy; Extract the key control device identifiers and their operating parameter sequences involved in the strategy; By combining the load rate, energy consumption curve and equipment health indicators in the real-time operation log of the equipment, the sequence of operating parameters is dynamically optimized. Output an optimized set of equipment-level heat recovery operation instructions.

7. The intelligent heat transfer recovery method for a heating system based on an LSTM neural network according to claim 6, characterized in that, Also includes: The configuration strategy verification unit receives the equipment-level heat recovery operation instruction set and simulates the thermodynamic response of the equipment-level heat recovery operation instruction set in the heating network system. The feasibility of the strategy is verified by comparing the simulation results with the preset safety threshold, and a verification report is generated.

8. The intelligent heat transfer recovery method for a heating system based on an LSTM neural network according to claim 7, characterized in that, Also includes: Design a dynamic adjustment mechanism to receive the verification report and the time series of newly added thermodynamic parameters collected in real time; When the validation report shows that the strategy deviates from the expected results, online fine-tuning of the LSTM feature extraction network parameters is triggered. The weight coefficients of the heat transfer potential assessment model are updated based on incremental learning.

9. The intelligent heat transfer recovery method for a heating system based on an LSTM neural network according to claim 8, characterized in that, Also includes: Deploy the strategy execution interface, receive the finally verified device-level heat recovery operation instruction set, and convert the device-level heat recovery operation instruction set into device control protocol instructions; The system sends a sequence of timestamped control commands to each pump and valve actuator via an industrial bus.

10. The intelligent heat transfer recovery method for a heating system based on an LSTM neural network according to claim 9, characterized in that, Also includes: Construct a heat recovery efficiency monitoring system to continuously collect the thermodynamic parameters of the pipeline network after performing heat recovery operations; Calculate the deviation between the actual heat recovery efficiency and the predicted value, and generate performance evaluation indicators; The performance evaluation index is fed back into the heat transfer potential evaluation model as training data increment.

Images