Operation data analysis method and system of heat supply system

Abstract

The invention provides an operation data analysis method and system of a heat supply system, and relates to the technical field of building environment and energy application engineering.The method comprises the steps that if a diagnosis result shows that no abnormity exists, differential operation strategies are generated based on building historical room temperature, building characteristics and real-time meteorological parameters in a multi-source fusion data set; obtaining meteorological prediction data in a preset time period; based on the differentiated operation strategy and meteorological prediction data, predicting heat load demand change through a long short-term memory network; and calculating a target output parameter set value by taking the room temperature reaching the standard and the lowest energy consumption as an optimization target based on the change of the thermal load demand, and sending the set value to a boiler control system for execution. According to the invention, through multi-source data fusion and real-time analysis, closed-loop intelligent operation from abnormal accurate diagnosis to demand prediction regulation is realized, so that the reliability, energy efficiency and comfort level of the heat supply system are comprehensively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of building environment and energy application engineering, in particular to a running data analysis method and system of a heating system. BACKGROUND

[0002] With the acceleration of urbanization process and the continuous expansion of central heating scale in China, the structure of heating system is becoming increasingly complex, and the traditional operation and management mode relying on manual experience has been difficult to meet the fine and intelligent needs. Although some systems have deployed automatic monitoring equipment to realize basic data collection, due to the insufficient data processing and analysis capability, the system still has defects, which restricts the improvement of operation energy efficiency and service quality.

[0003] Taking the heating system of an old community in the north as an example, it serves 288 households (8 six-story buildings) and uses a 2.8MW gas hot water boiler. Three types of collection equipment are deployed: boiler side collects water temperature, pressure, etc. every minute (local storage), pipe network side collects building supply and return water parameters every 5 minutes (property platform storage), and 10% of the users upload room temperature every 30 minutes (including APP complaints). The system has the following defects: 1. Poor multi-source data fusion, boiler, pipe network and user data are stored in three independent systems, formats are different (binary, Excel, JSON) and there is no unified time sequence alignment, manual data comparison is needed when user room temperature is abnormal, which takes more than 2 hours and it is difficult to locate faults; 2. Insufficient real-time data analysis, offline batch processing is used every morning, and pipe network slight leakage (4 hours pressure drop 0.08MPa) needs to be found the next day, during which heat is lost and user room temperature drops, and a passive response is made; 3. Lack of personalized control basis, the system is run uniformly according to the average parameters of the community (such as boiler water temperature 75℃), without considering the differences such as 13% to 16% higher heat loss of street buildings and 2 to 2.5℃ lower room temperature of top floors, which may cause over-heating of middle buildings and only 72% of edge buildings to meet the standard, with 9% of gas waste. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a running data analysis method and system of a heating system, which realizes closed-loop intelligent operation from abnormal accurate diagnosis to demand prediction and control through multi-source data fusion and real-time analysis, thereby comprehensively improving the reliability, energy efficiency and comfort of the heating system.

[0005] To solve the above technical problems, the technical solutions of the present application are as follows: In a first aspect, a running data analysis method of a heating system, the method comprising: real-time collection of equipment operation data, pipe network operation data and user side data; time sequence alignment of the data based on a unified time base, and feature dimension reduction to construct a multi-source fusion data set; The calibration module is configured to calculate, based on the multi-source fusion data set, a key performance indicator of a heat supply inlet of each building in real time, dynamically determine a plurality of distributed reference nodes based on the key performance indicator, a pipe network topological structure and a real-time hydraulic working condition, construct a hydraulic balance reference model based on data of the reference nodes, extract a state feature of the hydraulic balance reference model, generate a dynamic calibration coefficient based on the state feature, and calibrate the key performance indicator by using the dynamic calibration coefficient to obtain a calibrated key performance indicator. The identification module is configured to compare a real-time value of the calibrated key performance indicator with a preset dynamic threshold interval to obtain an index deviation state vector, and identify an abnormal data cluster based on the index deviation state vector to obtain a diagnosis result containing an abnormal type and a location. The strategy generation module is configured to, if the diagnosis result is no abnormality, generate a differentiated operation strategy based on historical room temperatures of the buildings, building characteristics and real-time meteorological parameters in the multi-source fusion data set. The strategy generation module is configured to, if the diagnosis result is no abnormality, generate a differentiated operation strategy based on historical room temperatures of the buildings, building characteristics and real-time meteorological parameters in the multi-source fusion data set.

[0006] In a second aspect, an operation data analysis system of a heat supply system includes: The acquisition module is configured to acquire, in real time, equipment operation data, pipe network operation data and user-side data, perform time sequence alignment on the data based on a unified time base, and perform feature dimension reduction to construct a multi-source fusion data set. The calibration module is configured to calculate, based on the multi-source fusion data set, a key performance indicator of a heat supply inlet of each building in real time, dynamically determine a plurality of distributed reference nodes based on the key performance indicator, a pipe network topological structure and a real-time hydraulic working condition, construct a hydraulic balance reference model based on data of the reference nodes, extract a state feature of the hydraulic balance reference model, generate a dynamic calibration coefficient based on the state feature, and calibrate the key performance indicator by using the dynamic calibration coefficient to obtain a calibrated key performance indicator. The identification module is configured to compare a real-time value of the calibrated key performance indicator with a preset dynamic threshold interval to obtain an index deviation state vector, and identify an abnormal data cluster based on the index deviation state vector to obtain a diagnosis result containing an abnormal type and a location. The strategy generation module is configured to, if the diagnosis result is no abnormality, generate a differentiated operation strategy based on historical room temperatures of the buildings, building characteristics and real-time meteorological parameters in the multi-source fusion data set. An optimization control module is configured to acquire meteorological prediction data in a preset time period, predict heat load demand variation through a long short-term memory network based on a differentiated operation strategy and the meteorological prediction data, and calculate a target output parameter set value based on the heat load demand variation, with the optimization target being that the room temperature meets the standard and the energy consumption is the lowest, and send the set value to a boiler control system for execution.

[0007] In a third aspect, a computing device includes: one or more processors; a memory device storing one or more programs that, when executed by the one or more processors, cause the one or more processors to implement the method.

[0008] In a fourth aspect, a computer-readable storage medium stores a program that, when executed by a processor, implements the method.

[0009] The above scheme of the present application at least has the following beneficial effects: Real-time acquisition of multi-source data and alignment based on a unified time reference time sequence eliminate data time differences; dimension reduction through principal component analysis reduces data redundancy, constructs a high-quality multi-source fusion data set, and improves data integration and utilization efficiency; real-time calculation of key performance indicators based on the fusion data set, dynamic determination of reference nodes combined with pipe network topology and hydraulic working conditions, construction of a hydraulic model to generate dynamic calibration coefficients, calibration of indicators to improve indicator accuracy; deviation vectors are obtained by comparing calibrated indicators with dynamic thresholds, combined with clustering analysis to identify abnormal data clusters, determine abnormal types and locations, and improve the targeting of abnormal diagnosis; when there is no abnormality, the building historical room temperature, building characteristics and real-time meteorological parameters in the multi-source data are used to generate differentiated operation strategies through a pre-trained neural network, so that the strategies fit the building differences; meteorological prediction data is acquired, and the strategy is combined through a long short-term memory network to predict heat load variation, with the target being to calculate control parameters and send them to the system for execution to realize data-driven dynamic regulation and control and improve system operation adaptability. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 is a flowchart of a heat supply system operation data analysis method provided by an embodiment of the present application.

[0011] Figure 2 is a schematic diagram of a heat supply system operation data analysis system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0012] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in many forms and should not be limited by 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 present disclosure to those skilled in the art.

[0013] As shown in Figure 1 An embodiment of the present application proposes a method for analyzing operation data of a heating system, the method comprising the following steps: Step 100, real-time collection of equipment operation data, pipe network operation data and user side data; time series alignment of data based on a unified time reference, and feature dimension reduction, to construct a multi-source fusion data set; Step 200, based on the multi-source fusion data set, real-time calculation of key performance indicators of each building heat inlet; based on the key performance indicators, dynamic determination of a plurality of distributed reference nodes through the pipe network topology and real-time hydraulic working conditions; construction of a hydraulic balance reference model according to the data of the reference nodes, and extraction of state features, generation of dynamic calibration coefficients according to the state features; calibration of the key performance indicators using the dynamic calibration coefficients to obtain calibrated key performance indicators; Step 300, comparison of the real-time values of the calibrated key performance indicators with the preset dynamic threshold interval to obtain an indicator deviation state vector; based on the indicator deviation state vector, identification of abnormal data clusters to obtain a diagnosis result containing abnormal types and positions; Step 400, if the diagnosis result is no abnormality, generating a differentiated operation strategy based on the building historical room temperature, building characteristics and real-time meteorological parameters in the multi-source fusion data set; Step 500, obtaining meteorological prediction data for a preset time period; based on the differentiated operation strategy and the meteorological prediction data, predicting the heat load demand change through a long short-term memory network; based on the heat load demand change, taking room temperature compliance and minimum energy consumption as optimization objectives, calculating target output parameter set values, and sending the set values to a boiler control system for execution.

[0014] In the embodiments of the present application, multiple types of data are collected in real time, time alignment is achieved through a unified time reference, and time deviations of data from different sources are eliminated; principal component analysis is performed for feature dimension reduction, data dimension is reduced while core information is retained, data processing efficiency is improved, and a multi-source fusion data set is constructed to integrate multi-dimensional information of equipment, pipelines, and users; key performance indicators are calculated based on the multi-source data, reference nodes are dynamically determined based on pipeline topology and real-time working conditions, the reference nodes are more in line with actual operating conditions; a hydraulic balance model is constructed based on reference node data and dynamic calibration coefficients are generated to calibrate key indicators and improve the accuracy and relevance of the indicators; a deviation state vector is obtained by comparing calibrated indicators with dynamic thresholds, and the deviation degree of the indicators is quantified; abnormal data clusters are identified through clustering analysis, which can aggregate abnormal data with similar characteristics, clearly present the clustering characteristics of the abnormal data, and facilitate accurate judgment of the type and location of the abnormal data; based on historical room temperature, building characteristics, and real-time weather parameters in the multi-source data, a pre-trained neural network is used to generate differentiated strategies, fully utilize the correlation between multi-dimensional data, and make the strategies adaptable to the building characteristics and heating needs of different buildings, thereby achieving targeted regulation; meteorological prediction data are obtained, and a long short-term memory network is used to predict heat load changes based on the differentiated strategies, which can effectively capture the correlation between heat load, weather factors, and time series; and output parameters are calculated to meet the user's heating needs and control energy consumption.

[0015] In a preferred embodiment of the present application, the step 100 comprises: Step 101, collecting equipment operation data, pipe network operation data and user side data from three independent data sources of boiler control system, property monitoring platform and user service system in parallel, specifically including: starting the multi-source data parallel collection process, and connecting the three independent data sources of boiler control system, property monitoring platform and user service system; collecting equipment operation data from the boiler control system, including boiler outlet water temperature, boiler operating pressure, instantaneous gas consumption and boiler operating state parameters, keeping consistent with the original data collection frequency of the system, i.e. collecting once a minute, and recording the corresponding collection time information; collecting pipe network operation data from the property monitoring platform, specifically including the supply and return water temperature, supply and return water pressure and pipe network flow data of each building heat inlet, following the collection frequency of the platform every 5 minutes, synchronously recording the building number and collection time of the data; collecting user side data from the user service system, specifically including the indoor temperature data of users who have installed room temperature collectors and the room temperature overheating or overcooling complaint records submitted by users through the APP, collecting according to the frequency of the system uploading room temperature data every 30 minutes, and recording the building and floor information of the user; in the collection process, the original data format of each data source is preliminarily identified and saved to ensure that the binary format of the boiler data, the table format of the pipe network data and the JSON format of the user side data can be completely obtained.

[0016] Step 102, time stamp alignment of equipment operation data, pipe network operation data and user side data based on a unified time reference, obtaining a time stamp aligned data sequence; for the time stamp aligned data sequence, time scale normalization processing is carried out by using dynamic time warping algorithm to obtain an intermediate time series data set with consistent time resolution, specifically including: first, establishing a unified time reference, taking the standard time stamp with millisecond level accuracy as the unified reference, and extracting the original collection time information from the equipment operation data, pipe network operation data and user side data collected from the three data sources; matching the original time information of each data with the unified millisecond level time stamp, correcting the data with time record deviation, making the three types of data correspond to the unified time reference, forming the time stamp aligned data sequence; aiming at the time resolution difference of different data sources in the time stamp aligned data sequence, starting the dynamic time warping algorithm processing flow; taking the time resolution of equipment operation data every minute as the target resolution, interpolating the time interval data of pipe network operation data every 5 minutes, calculating the intermediate time data value between the adjacent two collection points by algorithm, adjusting the time interval of pipe network data to 1 minute; resampling the time interval data of user side data every 30 minutes, according to the path matching rule of dynamic time warping algorithm, evenly distributing the user side data within 30 minutes to the time node of every minute, finally obtaining an intermediate time series data set with unified time scale and every minute as the time resolution of all data.

[0017] In step 103, based on the intermediate time series dataset, the characteristic vector and the characteristic value are calculated by the principal component analysis method to determine the main characteristic component and its contribution degree; the main characteristic component is screened according to a preset variance contribution rate threshold to obtain a main characteristic subspace, specifically including: standardizing the preprocessing of all data in the intermediate time series dataset to eliminate the influence of different types of data (such as temperature, pressure, flow) due to the dimensional difference, so that all types of data are in the same numerical magnitude; based on the standardized intermediate time series dataset, the principal component analysis method is applied to calculate the covariance matrix of all characteristic variables of the dataset, and then the characteristic vector and the corresponding characteristic value of the covariance matrix are solved by matrix operation; according to the size of the calculated characteristic value, the contribution degree of each characteristic vector to the original data information is determined, and the larger the characteristic value is, the higher the contribution degree of the corresponding characteristic component is; the variance contribution rate threshold is preset, such as the cumulative variance contribution rate is not less than 85%, the characteristic components are sequentially sorted from high to low according to the contribution degree, and the variance contribution rate of the characteristic components is gradually accumulated; when the accumulated variance contribution rate reaches the preset threshold, the screening of the characteristic components is stopped, and the screened characteristic components are combined to form a main characteristic subspace.

[0018] In step 104, the characteristic data in the main characteristic subspace is geometrically projected and transformed to map to a low-dimensional feature space to obtain the reduced dimension feature projection data, specifically including: determining the characteristic vector direction corresponding to each characteristic component in the main characteristic subspace, taking these characteristic vectors as the coordinate axes of the low-dimensional feature space, and constructing the coordinate system of the low-dimensional feature space; extracting all characteristic data in the main characteristic subspace, and regarding each characteristic data as a data point in the high-dimensional space; through geometric projection transformation operation, each data point in the high-dimensional space is projected into the low-dimensional feature space composed of the main characteristic component coordinate axes, and the relative position relationship between the data points is kept consistent with the original high-dimensional space during the projection process to ensure that the distribution characteristics of the data are not lost; the projected data is verified to check whether the core information in the main characteristic subspace is completely retained, and if there is information loss, the projection transformation parameter is adjusted for re-projection; after verification, the reduced dimension feature projection data is obtained.

[0019] At step 105, based on the dimension-reduced feature projection data, the feature information of the multi-source data stream is integrated by a feature weighting fusion method to construct a multi-source fusion data set with time sequence consistency, feature orthogonality and elimination of multicollinearity, specifically including: performing feature importance analysis on the dimension-reduced feature projection data, combining the heating system operation logic to determine the influence degree of different data source features on system operation state evaluation, for example, the boiler outlet water temperature feature in the equipment operation data, the supply and return water pressure difference feature in the pipe network operation data and the indoor temperature feature in the user side data are more critical to system analysis, and are given higher weights, and other auxiliary features are given relatively lower weights; based on the determined feature weights, starting the feature weighting fusion process, weighting and calculating the feature projection data corresponding to the equipment operation data, pipe network operation data and user side data according to the weights to obtain preliminary fused feature data; performing correlation test on the preliminary fused feature data, calculating the linear correlation coefficient between any two features, if the correlation coefficient exceeds a preset threshold, such as 0.8, it is determined that there is multicollinearity, and the feature with lower contribution degree is removed; repeating the correlation test and feature removal process until the linear correlation coefficient between all retained features is lower than the preset threshold; finally, the constructed multi-source fusion data set not only maintains the consistency of all data in the time dimension, but also realizes the orthogonality between the features, and eliminates the interference of multicollinearity on subsequent analysis.

[0020] In a preferred embodiment of the present application, the above step 200 comprises: In step 201, based on the multi-source fusion data set, the instantaneous supply and return water pressure difference, temperature difference and flow deviation rate key performance indicators of each building heat inlet are calculated in real time, which specifically includes: extracting the real-time supply and return water pressure data, supply and return water temperature data and actual flow data of each building heat inlet from the multi-source fusion data set, wherein the supply and return water pressure and temperature data are derived from the pipe network operation data collected by the original property monitoring platform, and the actual flow data is the integrated real-time monitoring data after multi-source fusion; for 8 residential buildings serving an old community in the north, the instantaneous supply and return water pressure difference of each building heat inlet is calculated one by one, specifically the supply water pipe pressure value minus the return water pipe pressure value of the building heat inlet at the same time; the instantaneous supply and return water temperature difference of each building heat inlet is calculated, specifically the supply water pipe temperature value minus the return water pipe temperature value of the building heat inlet at the same time; the instantaneous flow deviation rate of each building heat inlet is calculated, specifically the difference between the actual flow value and the design flow value of the building heat inlet at the same time divided by the design flow value, wherein the design flow value is determined in advance according to the rated output of the 2.8MW gas hot water boiler and the heat load demand of each building; the calculation process maintains the same time resolution as the multi-source fusion data set (i.e. calculation once every minute), ensuring real-time capture of changes in operating parameters of each building heat inlet, complete coverage of all heat inlets of 8 residential buildings, and formation of instantaneous supply and return water pressure difference, temperature difference and flow deviation rate data sequence corresponding to each minute of each building.

[0021] In step 202, based on the key performance indicators, the node feature parameters of the pipe network are extracted in combination with the pipe network topology and real-time hydraulic working condition data; the key performance indicators and the node feature parameters are standardized to generate a standardized feature matrix, specifically including: extracting the key performance indicators (i.e. the instantaneous supply and return water pressure difference, temperature difference and flow deviation rate of each building heat inlet) from the multi-source fusion data set, while calling the topology data of the heating pipe network of the community, including the pipe network connection relationship of the heat inlets of the eight residential buildings, the length and diameter parameters of each section of pipe, the pipe network branch distribution, and the real-time hydraulic working condition data, including the current boiler outlet pressure, total pipe network flow, and medium flow rate in each building pipe; based on the above data, the node feature parameters of the pipe network are extracted, specifically including the number of pipe connections of each building heat inlet node, the hierarchical position of the pipe network branch, the resistance coefficient of the corresponding pipe, and the distance between the node and the boiler heat source; the extracted key performance indicators and node feature parameters are standardized, the supply and return water pressure difference (value range 0.1 to 0.5 MPa), the supply and return water temperature difference (value range 5 to 20℃), the flow deviation rate (value range negative 20% to 20%) and the node feature parameters (such as pipe resistance coefficient) are all mapped to the numerical interval of 0 to 1 through linear transformation, eliminating the influence of the dimensional difference between different parameters; the standardized key performance indicators and node feature parameters are arranged by rows and columns, each row corresponds to a building heat inlet node, and each column corresponds to a standardized parameter, to form a standardized feature matrix, ensuring that the matrix covers the complete standardized data of all heat inlet nodes of the eight buildings.

[0022] Step 203, taking the standardized feature matrix as input, calculating the weighted Euclidean distance between each data sample by a clustering analysis algorithm, and dividing the pipe network nodes into multiple clusters by using the K-means clustering method to obtain the cluster division result, specifically including: taking the standardized feature matrix as input data, determining the weight coefficients of each parameter, wherein the supply and return water pressure difference and the supply and return water temperature difference have greater influence on the hydraulic state, the weight coefficient is set to 0.3, the flow deviation rate weight coefficient is set to 0.2, the pipe resistance coefficient and the node branch position weight coefficient in the node characteristic parameter are set to 0.15, and the weight coefficient of the remaining parameters is set to 0.05; based on the set weight coefficient, the weighted Euclidean distance between the data samples of each building heat inlet node is calculated, that is, the square root of the square sum of each parameter dimension difference after weighting is calculated; according to the pipe network distribution characteristics of the 8 buildings in the community, the number of clusters of K-means clustering is preset to 3, corresponding to the street buildings, the middle buildings and the top concentrated buildings respectively; initializing the centroids of the 3 clusters, assigning each data sample to the cluster where the nearest centroid is located, and calculating the mean of all samples in each cluster as a new centroid; repeating the sample assignment and centroid updating process until the centroid change in the continuous two iterations is less than a preset threshold, such as 0.001, stopping iteration, and obtaining the final cluster division result to determine the cluster group to which each building heat inlet node belongs.

[0023] Step 204, based on the cluster division result, combining the connectivity analysis of the pipe network topology structure, selecting the node closest to the centroid and representative in each cluster as a distributed reference node to complete the dynamic identification and determination of the reference node, specifically including: based on the cluster division result, calling the topology structure diagram of the heating pipe network of the community, analyzing the pipe network connectivity of each building heat inlet node in each cluster, judging whether the nodes are in the same pipe network branch, whether there is a direct pipe connection, whether they share the same water supply riser, and excluding isolated nodes with poor connectivity with other nodes in the cluster; for each cluster, calculate the Euclidean distance of all nodes in the cluster to the centroid of the cluster, and select 3 to 4 nodes closest to the centroid; from these nodes close in distance, combined with the representative requirements of the pipe network operation, select the nodes at the key position of the cluster pipe network branch, such as a cluster containing 3 street buildings, select the building heat inlet node located in the middle of the cluster pipe network branch, to ensure that the node can reflect the hydraulic state of other nodes in the cluster; for each cluster, the above screening is completed, and finally 2 to 3 distributed reference nodes are determined, covering all pipe network areas of the 8 buildings in the community, and the dynamic identification and determination of the reference node is completed.

[0024] Step 205, collect the real-time operation data of the distributed reference node; using the real-time operation data, the hydraulic balance reference model reflecting the overall hydraulic state is established by multiple linear regression method, specifically including: starting the data collection process, for the determined distributed reference node (covering the key area of the pipe network of 8 residential buildings in an old community in the north), collecting the real-time operation data of each reference node at a frequency of once per minute, specifically including the real-time temperature of the reference node water supply pipe, the real-time temperature of the return pipe, the real-time pressure of the water supply pipe, the real-time pressure of the return pipe, the instantaneous flow of the pipe medium, the real-time flow rate of the pipe medium, and synchronously recording the millisecond-level time stamp corresponding to each group of data during the collection process, ensuring that the time stamp is consistent with the unified time base of the multi-source fusion data set, and continuously collecting for 30 minutes to obtain 30 complete real-time operation data.

[0025] Based on the rated output parameters of the 2.8MW gas hot water boiler used in the heating system of the community, combined with the design heat load of each building (pre-calculated according to building area, insulation coefficient, and house type structure), the design hydraulic parameters of each distributed reference node are determined, specifically including the design supply-return water pressure difference of the reference node is 0.3MPa (to ensure sufficient medium circulation power), the design supply-return water temperature difference is 15℃ (to match the heat exchange demand under the rated output of the boiler), and the design flow rate is 8m 3 / h (determined according to the equal distribution of total flow rate according to the heat load of 8 buildings); based on the collected 30 groups of real-time operation data of the distributed reference node, the hydraulic balance reference model is constructed, first, the independent variables and dependent variables of the model are selected, the independent variables are the supply-return water pressure difference between each reference node (including the pressure difference of the water supply pipe between adjacent reference nodes, and the pressure difference of the water supply pipe of the same reference node itself), and the dependent variables are the instantaneous flow rate of each reference node (directly reflecting the hydraulic output state of the node, and is the core index for judging hydraulic balance), the correlation between the two is established by multiple linear regression method, first, the independent variables and dependent variables in the 30 groups of data are sorted, and the abnormal data (such as instantaneous flow rate suddenly being 0 or pressure difference suddenly rising or falling) outside the normal range are removed to ensure data effectiveness, then based on the sorted data, the correlation framework of the independent variables and dependent variables is preliminarily established, forming the basic structure of the model.

[0026] According to the principle of pipe network hydraulic balance, combined with the topological structure of the pipe network in the community (the connection relationship of each reference node, the pipeline direction and the branching condition), the balance equation set of the hydraulic balance reference model is constructed, which includes three types of core relationships: first, the inflow and outflow of each reference node are equal, for example, if a reference node is connected to two input pipes and one output pipe, it needs to satisfy that the sum of the real-time flow of the two input pipes is equal to the real-time flow of the output pipe, to ensure the single node flow conservation; second, the linear relationship between the pressure difference and the flow between the reference nodes, for each pair of adjacent reference nodes, a linear correlation equation between the change of the supply and return water pressure difference and the change of the pipeline flow is established, which reflects the driving effect of the pressure difference on the flow; third, the total flow balance relationship of the pipe network, that is, the total real-time flow of the boiler outlet water supply pipe is equal to the sum of the instantaneous flow of all distributed reference nodes, to ensure the flow balance of the entire pipe network system, and the three types of relationships together constitute the equation set system of the model.

[0027] Subsequently, model training and parameter determination are carried out, and the 30 groups of continuously collected real-time operation data are substituted into the constructed balance equation set one by one. First, the preliminary values of each coefficient in the equation set (such as the proportional coefficient of the linear relationship between the pressure difference and the flow, the node flow distribution coefficient) are calculated under each group of data; then the relative deviation (the difference between the theoretical value and the actual value divided by the actual value) between the theoretical flow value calculated by the equation set after each group of data is substituted and the instantaneous flow value actually collected is calculated; according to the size of the relative deviation, the coefficient parameters in the equation set are adjusted, if the deviation exceeds the preset 5% allowable range, the corresponding coefficient is increased or decreased to reduce the deviation, the data substitution, deviation calculation and coefficient adjustment process is repeated until the relative deviation between all theoretical flow values and actual flow values in the 30 groups of data is less than 5%; when the deviation meets the requirements, the final values of each coefficient in the equation set are determined, at this time the correlation relationship between the independent variables and the dependent variables established by the multiple linear regression is integrated with the adjusted balance equation set to form a complete hydraulic balance reference model, which can accurately reflect the overall hydraulic state of the community heating pipe network, and the construction and training of the model are completed.

[0028] Step 206, extracting the pressure distribution characteristics and flow balance characteristics in the hydraulic balance reference model as state characteristics, specifically including: calling the established hydraulic balance reference model, extracting the pressure distribution characteristics in the model, including the real-time supply and return water pressure values of each distributed reference node, the water supply pressure difference and return water pressure difference between adjacent reference nodes, the pressure gradient (i.e. the pressure difference value divided by the pipe length of the node and the boiler) of the reference node to the boiler heat source, the pressure fluctuation amplitude of the pipe network branch where each reference node is located; extracting the flow balance characteristics in the model, specifically including the real-time flow values of each distributed reference node, the proportion of the flow of each reference node to the total flow of the pipe network, the flow distribution proportion of each reference node in the same pipe network branch, and the trend data of the flow of the reference node over time; arranging the extracted pressure distribution characteristics and flow balance characteristics to ensure that each characteristic corresponds to a specific numerical value or data sequence and can reflect the overall hydraulic operation state of the district heating pipe network.

[0029] Step 207, correlating the state characteristics with the standard reference values to obtain the correlation analysis results; based on the correlation analysis results, fitting the mathematical relationship between the state characteristics and the reference values by the least squares method to obtain the fitting results, specifically including: determining the standard reference values of the state characteristics according to the design specifications and operation standards of the district heating system, wherein the standard reference values of the pressure distribution characteristics include the design supply and return water pressure difference 0.3 MPa, the adjacent node pressure gradient 0.02 MPa / m, and the pressure fluctuation amplitude not more than ±5%, the standard reference values of the flow balance characteristics include the design flow 8 m 3 / h, the reference node flow proportion 12.5% (averaged according to 8 buildings), and the flow distribution proportion deviation not more than ±10%; correlating the real-time state characteristic data extracted in step 206 with the above standard reference values, calculating the difference, ratio and correlation coefficient between each real-time state characteristic and the corresponding standard reference value to determine the numerical correlation relationship between them; based on the data obtained by the correlation analysis (including the corresponding data of multiple groups of real-time state characteristics and standard reference values), taking the real-time state characteristics as the independent variable and the standard reference values as the dependent variable, fitting by the least squares method; determining the coefficients of the fitting relationship by iterative calculation to make the error sum of squares of the predicted values calculated by the fitting relationship and the standard reference values minimum, and obtaining the final fitting result, i.e. the mathematical relationship expression between the real-time state characteristics and the standard reference values.

[0030] Step 208, based on the fitting result, the weight coefficient of each state feature is solved; based on the weight coefficient, the dynamic calibration coefficient is calculated, specifically including: based on the fitting result obtained in step 207, that is, the mathematical relationship expression, extracting the coefficients corresponding to each real-time state feature in the expression, taking these coefficients as the weight coefficients of each state feature, wherein the weight coefficients corresponding to the state features with greater influence on the standard reference value (such as adjacent node pressure gradient, reference node flow rate ratio) have higher values, and the weight coefficients corresponding to the features with smaller influence (such as pressure fluctuation amplitude, flow change trend) have lower values; according to the deviation degree of the real-time monitoring value and the standard reference value of each state feature, the deviation coefficient is calculated, that is, (real-time value-standard value) / standard value; the weight coefficient of each state feature is multiplied by the corresponding deviation coefficient to obtain the local calibration coefficient corresponding to the state feature; the local calibration coefficients of all state features are weighted and summed to calculate the final dynamic calibration coefficient, and each building heat inlet key performance indicator corresponds to a dynamic calibration coefficient matching the state of the pipe network region where it is located, ensuring that the calibration coefficient can be adjusted with the real-time hydraulic state of the pipe network.

[0031] Step 209, apply the dynamic calibration coefficient to the key performance indicator to dynamically calibrate and correct it to obtain the calibrated key performance indicator, specifically including: calling the instantaneous supply and return water pressure difference, temperature difference and flow deviation rate key performance indicator data of each building heat inlet, and calling the dynamic calibration coefficient corresponding to each building; multiplying the instantaneous supply and return water pressure difference value of each building by the corresponding dynamic calibration coefficient to obtain the calibrated instantaneous supply and return water pressure difference; multiplying the instantaneous supply and return water temperature difference value of each building by the corresponding dynamic calibration coefficient to obtain the calibrated instantaneous supply and return water temperature difference; multiplying the instantaneous flow deviation rate value of each building by the corresponding dynamic calibration coefficient to obtain the calibrated instantaneous flow deviation rate; during the calibration process, the numerical change before and after calibration and the corresponding time stamp are recorded synchronously to ensure the traceability of the data; after the calibration of the key performance indicators of the heat inlets of all 8 buildings is completed, the calibrated key performance indicator dataset is formed.

[0032] In the embodiment of the present application, the instantaneous supply and return water pressure difference, temperature difference and flow deviation rate of the building heat inlet are calculated in real time to provide key data support, capture dynamic changes in the running state, and ensure that the data reflects the current real situation; the node characteristics are extracted in combination with the pipe network topology and hydraulic working conditions, which is in line with the actual scene, the dimensional difference is eliminated through standardized processing, a regular feature matrix is generated, and the efficiency and accuracy of subsequent algorithm processing are improved; the weighted Euclidean distance highlights the important feature influence, the K-means clustering divides the node clusters according to the data similarity, reduces the redundant analysis of no difference, and provides a basis for the selection of reference nodes; the nodes are selected in combination with the connectivity of the pipe network, which is in line with the physical structure, the representative nodes near the centroid of the cluster group are selected, and the working condition changes are dynamically adapted; the reference node data is used for modeling to reduce the operation load, the multivariate linear regression is used to build a hydraulic balance model, and the overall hydraulic state is fully described; the core state characteristics of the model pressure and flow are extracted, irrelevant interference is excluded, accurate basis is provided for the subsequent correlation analysis and the generation of calibration coefficients; the correlation analysis quantifies the relationship between the state characteristics and the standard value, the least square method is used to fit the mathematical relationship, and the basis for solving the weight is laid; the weight coefficient is solved to reflect the influence degree of the characteristics on the hydraulic balance, the dynamic calibration coefficient is generated, and the actual hydraulic state is adapted; the dynamic calibration coefficient corrects the key performance indicators, adapts to the actual state of the pipe network, improves the accuracy of the indicators, and provides reliable data support for subsequent abnormal diagnosis.

[0033] In a preferred embodiment of the present application, the above step 300 comprises: In step 301, based on the calibrated key performance indicators, the real-time values thereof are compared with preset dynamic threshold intervals item by item, the relative deviation values of each indicator are calculated, and an indicator deviation state vector is obtained, specifically including: from the calibrated key performance indicator data set, the instantaneous supply and return water pressure difference, the instantaneous supply and return water temperature difference, and the instantaneous flow deviation rate data of each building of the eight residential buildings in the old community in the north per minute are extracted, wherein the calibrated supply and return water pressure difference ranges from 0.12 to 0.48 MPa, the calibrated supply and return water temperature difference ranges from 6 to 19℃, and the calibrated flow deviation rate ranges from negative 14% to 14%; the dynamic threshold interval of each indicator is preset, which is determined according to the design parameters and historical normal operation data of the 2.8 MW gas hot water boiler of the community, the supply and return water pressure difference threshold interval is 0.15 to 0.45 MPa, the supply and return water temperature difference threshold interval is 7 to 18℃, and the flow deviation rate threshold interval is negative 12% to 12%; for each building per minute, the three calibrated indicators are compared with the corresponding dynamic threshold interval one by one to determine whether the real-time value is within the threshold interval; the relative deviation value of each indicator is calculated, the relative deviation value of the supply and return water pressure difference is (real-time calibrated value-threshold interval median) divided by the threshold interval median, the relative deviation value of the supply and return water temperature difference is calculated in the same way, and the relative deviation value of the flow deviation rate is (real-time calibrated value-threshold interval median) divided by the threshold interval median, wherein the threshold interval median is 0.3 MPa for the supply and return water pressure difference, 12.5℃ for the supply and return water temperature difference, and 0% for the flow deviation rate; the corresponding supply and return water pressure difference relative deviation value, supply and return water temperature difference relative deviation value, and flow deviation rate relative deviation value of each building per minute are arranged in order to form an indicator deviation state vector of the building at that time, covering all time data of the eight buildings, ensuring that each data sample corresponds to a unique vector.

[0034] Step 302, taking the index deviation state vector as input data; calculating the reachable distance between each data sample point by performing a density clustering algorithm on the input data; identifying high-density regions and outliers according to the distribution characteristics of the reachable distance to obtain data point classification results; and forming a preliminary abnormal data cluster based on the data point classification results, specifically including: collecting all index deviation state vectors generated in step 301 as input data for the density clustering algorithm, and the data samples cover 8 buildings for 24 hours, with each minute corresponding to a vector (a total of 24x60x8=11520 samples); preprocessing the input data to remove null samples caused by data collection interruptions, ensuring that the remaining samples all contain complete three relative deviation values; performing the density clustering algorithm, first setting the algorithm parameters, with the neighborhood radius determined as 0.08 based on the average distance between samples, and the minimum sample number set as 5 (i.e., a region is considered a high-density region only if it contains at least 5 samples); calculating the reachable distance between each data sample point and all other sample points, i.e., the farthest distance that a sample point can indirectly reach through other samples within the neighborhood; dividing the regions according to the distribution characteristics of the reachable distance, identifying high-density regions with a reachable distance less than the neighborhood radius and a sample number exceeding the minimum sample number, and identifying outliers with a reachable distance greater than the neighborhood radius and a surrounding sample number less than the minimum sample number; obtaining data point classification results based on the above identification results, with samples in high-density regions classified as normal data points, and outliers and a small number of surrounding samples classified as potential abnormal data points; grouping the potential abnormal data point set according to the sample corresponding building and time correlation, grouping potential abnormal samples within the same building for 5 consecutive minutes, and grouping potential abnormal samples in different buildings but in the same pipe network branch and with similar time as a group to form a preliminary abnormal data cluster, each cluster containing sample number, corresponding building number, and time range information.

[0035] Step 303, based on the spatial distribution characteristics of the preliminary abnormal data cluster, combining the pipe network topology structure relationship, merging and optimizing the abnormal data cluster to determine the final abnormal data cluster classification, specifically including: Extract the spatial distribution characteristics of each preliminary abnormal data cluster, count the number of buildings covered by each cluster, the specific building number and the corresponding floor information, for example, a preliminary cluster covers building 1 (1-6 floors) and building 2 (1-6 floors), a preliminary cluster only covers a sample point on floor 5 of building 3, and a preliminary cluster covers building 5 (2-5 floors) and building 6 (1-4 floors); retrieve the topology data of the heating pipe network of the community, including the connection relationship of the heat inlet of the 8 buildings and the water supply riser and branch pipe, the pipe direction and length parameters, determine that building 1 and building 2 share a water supply riser (riser number L1), building 3 is independently connected to a water supply branch (branch number F1), building 4 is independently connected to a water supply branch (branch number F2), building 5 and building 6 share a water supply riser (riser number L2), and building 7 and building 8 share a water supply riser (riser number L3), based on these data, a pipe network topology graph is constructed, and the heat inlet of the building, the water supply riser and the branch pipe are taken as nodes in the graph, and the pipe connection relationship is taken as the edge between the nodes.

[0036] Map the building heat inlet node corresponding to each preliminary abnormal data cluster to the pipe network topology graph, determine the specific position of each preliminary cluster in the topology graph and the associated riser / branch node; start the spatial topology clustering process, calculate the topology distance between each preliminary cluster, which is defined as the shortest path length (in terms of pipe segment number) of the connection between the nodes corresponding to two preliminary clusters in the topology graph, for example, the preliminary cluster covering building 1 and the preliminary cluster covering building 2 are connected to riser L1, and the topology distance is 1 pipe segment (riser segment), the preliminary cluster covering building 3 and the preliminary cluster covering building 4 are connected to different branches, and the topology distance is 3 pipe segments (F1 branch segment + main pipe segment + F2 branch segment).

[0037] Combine the topology structure relationship of the pipe network to analyze the topology connectivity of the preliminary abnormal data cluster, determine whether all building heat inlet nodes in the same preliminary cluster belong to the same topology subgraph corresponding to the riser or branch, for example, the preliminary cluster covering building 5 and building 6 belongs to the topology subgraph corresponding to riser L2, and is determined to be topologically connected; determine whether the nodes of different preliminary clusters are in the same topology subgraph and the topology distance is less than a preset threshold (set to 2 pipe segments), for example, the preliminary cluster covering building 1 and the preliminary cluster covering building 2 are in the same topology subgraph and the topology distance is 1, which meets the threshold requirement.

[0038] The preliminary abnormal data clusters meeting the topological connectivity requirement and time range overlap (overlap duration greater than or equal to 5 minutes) are merged, for example, the preliminary cluster covering building 1 (time 8:00 to 8:10) and the preliminary cluster covering building 2 (time 8:02 to 8:11) are merged into a new cluster because they are in the same topological subgraph, the topological distance is 1 pipe, and the time overlap is 9 minutes; the preliminary clusters with topological distance greater than the threshold, or in different topological subgraphs, or without adjacent topological node support are screened, if the preliminary cluster only contains a single sample point and its corresponding node has no adjacent abnormal nodes (neither of the adjacent nodes has abnormal data) in the topological graph, for example, the preliminary cluster only covering building 3, floor 5, has no other abnormal nodes in branch F1 corresponding to the node, it is determined as a topological isolated cluster and is removed; if the preliminary cluster contains multiple sample points but is in an independent topological subgraph and has no associated abnormality, for example, the preliminary cluster covering building 4 (5 sample points), because branch F2 has no other abnormal nodes and is far from other branches in topological distance, combined with the historical fault data of this cell (branch F2 rarely has centralized abnormality), it is also determined as a topological isolated cluster and is removed.

[0039] After merging and removing, the final abnormal data cluster classification is determined, each final cluster needs to determine the building range and floor covered, the corresponding pipe network riser / branch (topological attribution), specific time interval, sample number and topological associated node information, for example, final cluster 1 covers building 1 (1 to 6 floors) and building 2 (1 to 6 floors), corresponds to riser L1, time 8:00 to 8:11, sample 22, and topological associated node is riser L1 and the pipes connected at both ends; final cluster 2 covers building 5 (2 to 5 floors) and building 6 (1 to 4 floors), corresponds to riser L2, time 14:30 to 14:40, sample 20, and topological associated node is riser L2 and the pipes connected at both ends.

[0040] At step 304, based on the final abnormal data cluster classification, the statistical characteristic parameters of each abnormal data cluster are extracted, including the average value of the index deviation degree in the cluster, the spatial distribution density, and the duration characteristic, specifically including: for each final abnormal data cluster, the index deviation degree data of all samples in the cluster is extracted, i.e. the relative deviation value of the supply and return water pressure difference, the relative deviation value of the supply and return water temperature difference, and the relative deviation value of the flow deviation rate; the average value of the index deviation degree in the cluster is calculated, the arithmetic mean value of the relative deviation value of the supply and return water pressure difference, the arithmetic mean value of the relative deviation value of the supply and return water temperature difference, and the arithmetic mean value of the relative deviation value of the flow deviation rate of all samples in the cluster are calculated, for example, the average value of the relative deviation of the supply and return water pressure difference of the final cluster 1 is negative 0.2, the average value of the relative deviation of the supply and return water temperature difference is 0.15, and the average value of the relative deviation of the flow deviation rate is negative 0.3; the spatial distribution density of the cluster is calculated, the proportion of the number of buildings covered by the final cluster to the total number of buildings in the community (8 buildings) is calculated, and the proportion of the number of floors covered to the total number of floors of the corresponding building (6 floors) is calculated, and the average value of the two is taken as the spatial distribution density, for example, the final cluster 1 covers 2 buildings, each building covers 6 floors, and the spatial distribution density is (2 / 8+6 / 6) / 2=0.625; the duration characteristic of the cluster is extracted, the time difference between the first appearance sample and the last appearance sample of the final cluster is determined, for example, the first sample time of the final cluster 1 is 8:00, the last sample time is 8:10, and the duration characteristic is 10 minutes; the average value of the index deviation degree in the cluster, the spatial distribution density, and the duration characteristic of each final abnormal data cluster are arranged as the statistical characteristic parameter set of the cluster, and each parameter corresponds to a specific value.

[0041] Step 305, based on the statistical characteristic parameters, the preset abnormal type discrimination rule is matched through the rule reasoning method, the abnormal type and the occurrence position are determined, the diagnostic result containing the abnormal type and the position is obtained, and specifically includes: the preset abnormal type discrimination rule is formulated according to the common heating fault type of the cell and historical fault data, the rule includes that the corresponding rule of the local blockage of the pipe network is that the average value of the relative deviation of the cluster internal flow deviation rate is less than or equal to-0.2, the average value of the relative deviation of the supply and return water pressure difference is less than or equal to-0.15, and the duration characteristic is greater than or equal to 5 minutes, the corresponding rule of the slight leakage of the pipe network is that the average value of the relative deviation of the cluster supply and return water pressure difference is less than or equal to-0.2, the average value of the relative deviation of the supply and return water temperature difference is greater than or equal to 0.1, and the duration characteristic is greater than or equal to 8 minutes, and the corresponding rule of the boiler heat efficiency decline is that the average value of the relative deviation of the supply and return water temperature difference of all the final clusters of the building is less than or equal to-0.1, and the average value of the relative deviation of the flow deviation rate is between-0.05 and 0.05; the statistical characteristic parameters of each final abnormal data cluster obtained in step 304 are called, and the parameters of each cluster are matched with the preset rule one by one; for example, the statistical characteristic parameters of the final cluster 1 are that the average value of the relative deviation of the flow deviation rate is-0.3, the average value of the relative deviation of the supply and return water pressure difference is-0.2, and the duration characteristic is 10 minutes, which meets the discrimination rule of the local blockage of the pipe network, and it is determined that the abnormal type corresponding to the cluster is the local blockage of the pipe network; according to the building range covered by the final abnormal data cluster and the corresponding pipe network topological structure, the abnormal occurrence position is determined, for example, the final cluster 1 covers the building 1 and the building 2 and corresponds to the water supply vertical pipe of the building 1 to the building 2, and it is determined that the abnormal occurrence position is the common water supply vertical pipe of the building 1 to the building 2 and the heat inlet of the two buildings; the abnormal type and the occurrence position corresponding to each final abnormal data cluster are sorted and summarized, the diagnostic result containing all the abnormal information is formed, the abnormal type, the involved building, the specific pipe network position and the abnormal duration are determined, and a determination basis is provided for subsequent fault processing.

[0042] In a preferred embodiment of the present application, the above step 400 comprises: Step 401, if the diagnostic result is normal, extract the historical room temperature, building characteristics and real-time meteorological parameters of each building from the multi-source fusion data set, specifically including: first, obtain the diagnostic result output in step 305, if the result shows that the heating system of 8 residential buildings in a certain old community in the north is normal (i.e. no pipe network blockage, leakage and other faults), start the data extraction process; extract the historical room temperature data of each building from the multi-source fusion data set, specifically 10% of the indoor temperature data collected every 30 minutes in the past 30 days (summarized by building, such as the average, maximum and minimum room temperature of No. 1 building from 6:00 to 22:00 every day), covering the complete historical room temperature sequence of each of the 8 buildings; extract the building characteristics data of each building, including the building age (all built around 2000), the external wall insulation coefficient (0.6 W / (m2·K) for buildings 1-2 and 7-8, 0.8 W / (m2·K) for buildings 3-6), whether it is street-facing (buildings 1-2 and 7-8 are street-facing, buildings 3-6 are in the middle), the number of floors (all 6 floors) and the area of the house type (average 80 m2 per household); extract real-time meteorological parameters, including current outdoor temperature (e.g. -8°C), outdoor wind speed (e.g. 2.5 m / s), solar radiation intensity (e.g. 200 W / m2), which are derived from real-time data of local meteorological platforms integrated in the multi-source fusion data set, ensuring that the three types of data extracted are one-to-one corresponding to each building without data confusion.

[0043] Step 402, combine the building historical room temperature, building characteristics and real-time meteorological parameters to construct the feature vector of each building, specifically including: feature preprocessing of the extracted data of each building, calculating the daily average room temperature, diurnal room temperature fluctuation amplitude (difference between daily maximum and minimum) and low temperature period (6:00-8:00) room temperature average of the historical room temperature data for the past 7 days; converting building characteristics data into quantifiable parameters, with 1 representing street-facing properties (buildings 1-2 and 7-8) and 0 representing middle buildings (buildings 3-6), the insulation coefficient is directly retained as the original value, and the number of floors is unified as 6; retaining the original values of current outdoor temperature and wind speed for real-time meteorological parameters, and taking the average of solar radiation intensity in the past 2 hours; constructing a feature vector for each building in the order of historical room temperature features, building characteristics parameters and real-time meteorological parameters, for example, the feature vector of building 1 (street-facing) includes: daily average room temperature 18.5°C for the past 7 days, diurnal fluctuation amplitude 3.2°C, low temperature period average 16.8°C, insulation coefficient 0.6 W / (m2·K), street-facing identifier 1, current outdoor temperature -8°C, current wind speed 2.5 m / s, and average solar radiation in the past 2 hours 200 W / m2; each of the 8 buildings constructs an independent feature vector, ensuring that each vector contains 8-dimensional feature data without missing or redundant dimensions.

[0044] Step 403, input the feature vector into the pre-trained neural network, and obtain the initial heating demand prediction value of each building through the forward propagation calculation of the neural network, specifically including: first, the construction and training of the pre-trained neural network model is carried out, and the model construction stage is based on the operation data of a certain old community in the north in the past 2 heating seasons, and the building feature vectors and actual heating demand data of 8 buildings in the period are collected; The dimension of the building feature vector is consistent with the vector constructed in step 402, and contains 8 dimensions of the average room temperature in the past 7 days, the diurnal room temperature fluctuation amplitude, the average room temperature in the low temperature period, the thermal insulation coefficient, the street identification, the average outdoor temperature, the average wind speed, and the average solar radiation; The actual heating demand data is the actual heat consumption of each building per hour (unit: kW), which is calculated by the boiler gas consumption, the temperature difference and flow data of the pipe network, and it is ensured that each feature vector corresponds to a unique actual heat consumption label.

[0045] The collected sample data is preprocessed, and the abnormal heat consumption data caused by equipment failure (such as heat consumption increasing to more than 3 times of the normal range) is excluded, and the missing feature values are filled with the average value of the same building at the same time period; The preprocessed samples are divided into training set and validation set in the ratio of 7:3, the training set is used for model parameter learning, and the validation set is used for monitoring the generalization ability of the model.

[0046] The neural network model structure is designed, the number of input layer neurons is set to 8, which matches the dimension of the feature vector; 3 hidden layers are set, the number of neurons in the first layer is 64, the second layer is 32, and the third layer is 16, and ReLU activation function is used for feature conversion in each hidden layer to capture nonlinear features; The output layer is set to one neuron, and the output value is the initial heating demand prediction value of the building (unit: kW), and the linear activation function is used to ensure that the output value meets the numerical range of heat consumption.

[0047] The model training process is started, the mean square error is selected as the loss function to measure the difference between the model prediction value and the actual heat consumption; Gradient descent optimizer is used, and appropriate learning rate is set (initial learning rate is 0.001, which is gradually attenuated with training iteration); The training set samples are input into the model in batches, each batch contains 32 samples, the prediction value is calculated through forward propagation, and the weights and bias parameters of each layer of neurons are updated through back propagation; After completing one round of training (traversing all training set samples), the validation loss of the model is calculated using the validation set, and if the validation loss does not decrease for 5 consecutive rounds, the training is stopped to avoid model overfitting; After training is completed, the optimal parameters of the model are saved, and a pre-trained neural network model that can be directly called is formed.

[0048] The trained pre-trained neural network model is called, and the feature vectors of the eight buildings constructed in step 402 are input into the model one by one to start forward propagation calculation; the feature vectors first pass through the input layer to the first hidden layer, are converted after ReLU activation function conversion to extract high-dimensional features, are then transmitted to the second hidden layer to further compress the feature dimension, and then are transmitted to the third hidden layer to complete final feature extraction; finally, the initial heating demand prediction value of each building is output through the output layer, with a unit of kW, for example, the initial prediction value of building No. 1 (street) is 120 kW, the initial prediction value of building No. 3 (middle) is 95 kW, and the initial prediction value of building No. 5 (middle) is 100 kW; the entire calculation process ensures that the prediction value of each building is generated based on its own exclusive feature vector, fully reflecting the difference in heating demand of different buildings due to building characteristics, historical heating conditions and real-time weather differences.

[0049] In step 404, hydraulic coupling analysis is performed on the initial heating demand prediction value of each building based on the topology of the pipe network to calculate the heat loss correction amount in the pipe network transmission process, specifically including: calling the topology data of the heating pipe network of the community to determine the connection relationship between the eight buildings and the boiler heat source: buildings No. 1 and No. 2 share a water supply vertical pipe (about 80 m away from the boiler, pipe diameter DN100), building No. 3 has an independent water supply pipe (about 60 m away from the boiler, diameter DN80), building No. 4 has an independent water supply pipe (about 65 m away from the boiler, diameter DN80), buildings No. 5 and No. 6 share a water supply vertical pipe (about 50 m away from the boiler, diameter DN100), and buildings No. 7 and No. 8 share a water supply vertical pipe (about 90 m away from the boiler, diameter DN100); based on the topology, hydraulic coupling analysis is performed to calculate the heat loss of the water supply pipeline of each building, and the heat loss calculation considers the pipe length (the longer the loss is greater), the diameter (the smaller the loss is greater), and the outdoor environment temperature (the lower the loss is greater), for example, the heat loss rate of the water supply pipeline of building No. 1 is 8%, that of building No. 3 is 6%, and that of building No. 7 is 9%; according to the initial heating demand prediction value and the heat loss rate, the heat loss correction amount of each building is calculated, correction amount = initial prediction value x heat loss rate, for example, the correction amount of building No. 1 is 120 kW x 8% = 9.6 kW, that of building No. 3 is 95 kW x 6% = 5.7 kW, and that of building No. 7 is 110 kW x 9% = 9.9 kW, ensuring that the correction amount fits the actual transmission loss of the pipe network.

[0050] Step 405, superimpose the initial heat supply demand prediction value and the heat loss correction amount to obtain the differentiated heat supply demand target value of each building, which specifically includes: for each building, superimpose the initial heat supply demand prediction value obtained in step 403 and the heat loss correction amount obtained in step 404, and the superposition formula is differentiated heat supply demand target value = initial prediction value + heat loss correction amount; complete the calculation of 8 buildings one by one, for example, the target value of No. 1 building = 120kW + 9.6kW = 129.6kW, the target value of No. 3 building = 95kW + 5.7kW = 100.7kW, the target value of No. 5 building = 100kW + 6.2kW = 106.2kW, and the target value of No. 7 building = 110kW + 9.9kW = 119.9kW; after the calculation is completed, the differentiated heat supply demand target value list of the 8 buildings is formed, the number of each building and the corresponding target value are marked, and it is ensured that the target value contains the heat demand of itself and covers the pipe network transmission loss, which meets the actual operation demand.

[0051] Step 406, according to the differentiated heat supply demand target value of each building, calculate the corresponding target water supply temperature setting value, form a differentiated operation strategy, which specifically includes: based on the rated parameters (rated outlet water temperature 85℃, rated return water temperature 60℃, rated heat load 2800kW) of the 2.8MW gas hot water boiler in the community, establish the correlation between the differentiated heat supply demand target value and the target water supply temperature, that is, according to the proportion of the target value in the total heat load of the boiler, combined with the pipe network return water temperature (set as the historical average value 58℃), the target water supply temperature of the corresponding building is calculated; for example, the target demand of No. 1 building is 129.6kW, which accounts for about 4.63% of the total load, and the target water supply temperature is calculated to be 78℃; the target demand of No. 3 building is 100.7kW, which accounts for about 3.59%, and the target water supply temperature is 72℃; the target demand of No. 5 building is 106.2kW, which accounts for about 3.79%, and the target water supply temperature is 73℃; the target demand of No. 7 building is 119.9kW, which accounts for about 4.28%, and the target water supply temperature is 76℃; the numbers of the 8 buildings and the corresponding target water supply temperatures are arranged into a table to determine the water supply temperature setting value that needs to be executed at the heat inlet of each building, and the strategy execution period (adjusted once every hour according to the real-time meteorological parameters) is marked, forming a complete differentiated operation strategy, which can be directly used to guide the regulation and control operation of the boiler and the pipe network.

[0052] In the embodiment of the present application, the data extraction process is started when the diagnosis result is normal, avoiding invalid resource consumption, and three types of key information, i.e., building historical room temperature, building characteristics and real-time weather, are extracted from the multi-source fusion data set, providing comprehensive and building-related data support for subsequent processing; the three types of data are combined into a structured feature vector specific to each building, avoiding data confusion and improving processing standardization and pertinence; the specific feature vector is input into a pre-trained neural network, the initial heating demand prediction value is generated through forward propagation by means of the existing rules of the model, and the individual differences of the buildings are reflected; the heat loss correction amount is calculated by combining the building location, pipe connection and other conditions of the pipe network topology, reflecting the loss differences of different buildings and improving the accuracy; the initial prediction value and the correction amount are superimposed, and the theoretical demand and the actual loss are integrated, so that the differentiated heating demand target value is closer to the actual situation, and the practicability is improved; the target water supply temperature of each building is calculated according to the target value, so that the strategy has landing property, and the temperature setting fits the actual demand of the building, which can directly guide the regulation and control of the boiler and the pipe network, and improve the operability of the strategy.

[0053] In a preferred embodiment of the present application, the above step 500 comprises: Step 501, acquiring meteorological prediction data of a future preset time period, including outdoor temperature, wind speed and solar radiation intensity parameters, specifically including: determining that the future preset time period is 24 hours, and the time resolution is set to once every hour, to ensure that the influence of meteorological changes on heat load can be tracked in real time; by connecting the open data interface of the local meteorological platform, the meteorological prediction data of the future 24 hours in the area where the northern old community is located is acquired; wherein the outdoor temperature prediction data covers the range of-15℃ to-8℃, and a prediction value is recorded every hour, such as-10℃ at 0 o'clock in the future, -11℃ at 1 o'clock, -15℃ at 6 o'clock, and-8℃ at 14 o'clock; the outdoor wind speed prediction data covers the range of 1.5m / s to 3.0m / s, and the wind speed in the area of the street buildings (No. 1 building to No. 2 building, No. 7 building to No. 8 building) is slightly higher, with a prediction value of 0.3 to 0.5m / s higher than that in the area of the middle buildings (No. 3 building to No. 6 building) every hour; the solar radiation intensity prediction data covers the range of 50W / ㎡ to 300W / ㎡, and only has a non-zero value during 6 o'clock to 18 o'clock, reaching a peak of 300W / ㎡ at about 12 o'clock; the three types of meteorological parameters obtained are arranged in chronological order to form structured data with one hour per piece, and each piece of data contains a time stamp, outdoor temperature, corresponding area wind speed and solar radiation intensity, to ensure that the data completely covers the future 24 hours and matches the distribution of the community buildings.

[0054] Step 502, based on the differentiated operation strategy and the weather prediction data, constructing a time series input feature, specifically including: extracting the target water supply temperature set value of each building from the differentiated operation strategy, wherein the target water supply temperature of the street buildings (1#-2# and 7#-8#) is 76-78℃, and the target water supply temperature of the middle buildings (3#-6#) is 72-73℃; constructing a time series input feature with each hour in the next 24 hours as a time step; the feature of each time step contains two parts, the first part is the weather prediction data (i.e. the outdoor temperature, wind speed and solar radiation intensity of the corresponding building area) of the corresponding time step, and the second part is the target water supply temperature of each building of the corresponding time step; arrange the features of the 24 time steps in time sequence to form a time series input feature, and the feature dimension of each time step is 5 (outdoor temperature 1, wind speed 1, solar radiation intensity 1, target water temperature of street buildings 1, target water temperature of middle buildings 1); for example, the time step feature of the future 0 hour is outdoor temperature -10℃, street area wind speed 2.8m / s, middle area wind speed 2.5m / s, solar radiation intensity 0W / ㎡, target water temperature of street buildings 78℃, and target water temperature of middle buildings 73℃, to ensure that the features of each time step can be associated with weather conditions and building operation strategy, and the time sequence logic is coherent.

[0055] Step 503, input the time series input feature into the long short-term memory network, and predict the heat load demand change trend in the future time period through the time sequence memory unit of the network, specifically including: first, construct and train the pre-trained long short-term memory network model, and the model construction stage takes the hourly operation data of the past 2 heating seasons of a certain old community in the north as the training sample source; collect hourly weather data, including outdoor temperature, outdoor wind speed (distinguish street building area and middle building area), and solar radiation intensity; collect the actual value of the heat load of each building per hour, which is calculated from the hourly gas consumption of a 2.8MW gas-fired hot water boiler, the temperature difference and flow data of the pipe network, for example, calculate the total heat according to the gas consumption (m) and the gas calorific value (about 36MJ / m), and then split according to the heat load proportion of the street buildings (1#-2# and 7#-8#) and the middle buildings (3#-6#) to obtain the actual value of the hourly heat load of each building; preprocess the collected sample data, eliminate the heat load outliers caused by boiler failure and data interruption, such as data exceeding 2 times of the normal range, align the weather data and heat load data by timestamp, ensure that each hourly sample contains complete weather parameters and corresponding building heat load values, and then divide them into training set and validation set in the ratio of 7:3. 3 / h 3 The long short-term memory network model structure is designed, the number of input layer neurons is set to 5, which matches the dimension of the subsequent time series input features at each time step (outdoor temperature, street-side area wind speed, middle area wind speed, solar radiation intensity, and building target water temperature correlation features); two LSTM unit layers are set in the hidden layer, and the number of neurons in each unit layer is 32, which is used to capture the long-term dependence of data in the time dimension; the number of neurons in the output layer is set to 2, corresponding to the hourly heat load demand value of the street-side building and the middle building respectively, to ensure that the output result can distinguish the load difference of the two types of buildings; the Dropout mechanism (proportion set to 0.2) is used between each LSTM unit layer to prevent overfitting, and the linear activation function is used in the output layer to match the numerical range of the heat load.

[0056] The model training process is started, the mean square error is selected as the loss function to measure the difference between the predicted heat load value and the actual heat load value; the Adam optimizer is used, the initial learning rate is set to 0.001, and the learning rate is adjusted by 10% every 50 iterations; the training set samples are input into the model by 32 samples per batch, the predicted heat load value is calculated by forward propagation, and the model weight and bias parameters are updated by back propagation; after completing one round of training (traversing all training set samples), the validation loss of the model is calculated using the validation set, if the validation loss does not decrease (fluctuation less than 0.001) for 5 consecutive rounds, the training is stopped to avoid overfitting of the model; after the training is completed, the optimal parameters of the model are saved to form a pre-trained long short-term memory network model that can be directly called.

[0057] The pre-trained long short-term memory network model trained above is called, and the 24-hour time series input features (each time step contains outdoor temperature, street-side area wind speed, middle area wind speed, solar radiation intensity, and building target water temperature correlation features) constructed in step 502 are input into the model as a whole, and the time sequence memory unit of the model is started; the time sequence memory unit retains the key information of the historical time steps (such as the continuous influence of the previous 1-hour low outdoor temperature on the current heat load, and the lagging adjustment effect of the previous 3-hour solar radiation intensity on the heat load) through the gating mechanism of the input gate, the forget gate, and the output gate, and updates the feature information of the current time step; the input data of 24 time steps are processed in time sequence, and the time sequence correlation features are extracted step by step; after the processing is completed, the model outputs the future 24-hour hourly heat load demand trend through the output layer, for example, the street-side building heat load demand is 130kW and the middle building heat load demand is 105kW at 6 o'clock in the future (outdoor temperature is -15℃ and street-side area wind speed is 3.0m / s), the street-side building heat load demand is 110kW and the middle building heat load demand is 90kW at 14 o'clock in the future (outdoor temperature is -8℃ and solar radiation intensity is 280W / ㎡), to determine the fluctuation rule of the heat load with the change of meteorological conditions, and accurately distinguish the heat load difference between the street-side building and the middle building.

[0058] Step 504, based on the heat load demand trend, a multi-objective optimization function is used to maintain user room temperature and minimize total energy consumption. The particle swarm optimization algorithm is used to solve the final operating parameter combination, which specifically includes: setting the constraint conditions of the multi-objective optimization function, wherein the room temperature standard constraint is that the indoor temperature of all building users is maintained at 18-22°C, and the total energy consumption minimum constraint is that the hourly gas consumption of the 2.8MW gas hot water boiler is minimized; based on these constraint conditions, a multi-objective optimization function is constructed, which integrates the room temperature standard and the total energy consumption minimum optimization targets into a quantifiable comprehensive index, avoids the situation where only the room temperature standard is pursued, resulting in excessive energy consumption, or only the energy consumption is controlled, resulting in the room temperature not meeting the standard, and eliminates the conflict between the two targets, making the optimization direction more in line with the actual heating demand; the value generated by the function is a comprehensive score, which directly reflects the pros and cons of a set of operating parameter combinations. Specifically, the higher the room temperature standard rate and the lower the gas consumption, the higher the comprehensive score. This value is used to determine whether each parameter combination meets the optimization target.

[0059] Based on the heat load demand trend, the operating parameter range to be optimized is determined. The supply water temperature set value range is 65-85°C (not exceeding the rated outlet water temperature of the boiler 85°C), and the circulating pump frequency set value range is 30-50Hz (to ensure normal operation of the circulating pump and not to exceed the rated frequency). The particle swarm optimization algorithm is started, and the particle swarm size is initialized to 50 particles, each particle representing a set of supply water temperature and circulating pump frequency parameter combinations. During the iteration process of the algorithm, each particle adjusts the flight direction and speed according to its historical optimal position (i.e. the position corresponding to the highest comprehensive score calculated by the particle in the past) and the group optimal position (i.e. the position corresponding to the highest comprehensive score calculated by all particles in the past). At the same time, the target function value corresponding to each particle, i.e. the above-mentioned comprehensive score, is calculated. After 100 iterations, if the change amount of the group optimal position is less than 0.1°C (supply water temperature) and 0.5Hz (circulating pump frequency) for 10 consecutive iterations, the iteration is stopped. Finally, the parameter combination with the highest comprehensive score that meets the room temperature standard rate of 100% and the lowest gas consumption is selected, for example, when the heat load demand is 130kW, the supply water temperature is 78°C and the circulating pump frequency is 45Hz, which corresponds to the highest comprehensive score. When the heat load demand is 90kW, the supply water temperature is 72°C and the circulating pump frequency is 38Hz, which also meets the requirement of the highest comprehensive score.

[0060] Step 505, convert the final operating parameter combination into the target output parameter setting value of the boiler, including the water supply temperature setting value and the circulating pump frequency setting value, specifically including: according to the rated operating parameters (rated outlet water temperature 85℃, rated circulating water volume 50m) of the 2.8MW gas hot water boiler and the rated performance parameters (rated frequency 50Hz, rated flow 50m) of the circulating pump, the corresponding relationship between the operating parameter combination and the target output parameter is established; the final operating parameter combination obtained in step 504 is converted, wherein the water supply temperature setting value directly uses the optimized value, without additional conversion, for example, the optimized 78℃ is directly used as the boiler water supply temperature setting value; the circulating pump frequency setting value is directly determined according to the optimized frequency value, for example, the optimized 45Hz is directly used as the circulating pump frequency setting value; for the heat load demand of future 24 hours at different times, the target output parameter setting value of the corresponding period is respectively converted to form a 24-hour parameter setting sequence; for example, the future 6 hours correspond to the water supply temperature 78℃ and the circulating pump frequency 45Hz, the future 14 hours correspond to the water supply temperature 72℃ and the circulating pump frequency 38Hz, and the future 22 hours correspond to the water supply temperature 75℃ and the circulating pump frequency 42Hz; the converted parameter setting value is arranged in time sequence, and the water supply temperature and the circulating pump frequency corresponding to each time point are marked to ensure that the parameter value is within the rated operating range of the boiler and the circulating pump, and matches the heat load demand change trend. 3 / h 3 / h Step 506, send the target output parameter setting value to the boiler control system for real-time control, specifically including: establishing a communication connection with the boiler control system, which includes a water temperature control unit and a circulating pump variable frequency control unit to ensure the real-time and stability of data transmission; according to the time sequence of future 24 hours, the target output parameter setting value (water supply temperature, circulating pump frequency) corresponding to each time point is sent to the boiler control system in turn; during the sending process, a timing sending mechanism is adopted, and the parameter setting value of each hour is sent once an hour, for example, the water supply temperature 76℃ and the circulating pump frequency 43Hz from 0 o'clock to 1 o'clock in the future are sent at 0 o'clock, and the parameters from 1 o'clock to 2 o'clock are sent at 1 o'clock; after receiving the parameter setting value, the water temperature control unit of the boiler control system automatically adjusts the gas supply amount to maintain the set water supply temperature, and the circulating pump variable frequency control unit automatically adjusts the motor frequency to match the set frequency; at the same time, the system feeds back the actual operating parameters (actual water supply temperature, actual circulating pump frequency) to the data processing module every 10 minutes to ensure the execution effect of the setting value, if the deviation between the actual parameter and the setting value exceeds ±1℃ (water temperature) or ±1Hz (frequency), the mechanism of re-sending the setting value is triggered to realize the real-time control closed loop, avoid passive response problem, and ensure the stable operation of the heating system according to the optimized parameters.

[0061] In the embodiment of the present application, the multi-dimensional weather forecast data of outdoor temperature, wind speed and solar radiation intensity in the future preset time period is obtained, covering the key environmental factors affecting the heat load demand; the time series input features are constructed by combining the differentiated operation strategy (including building demand differences) and the weather forecast data (including future environmental changes), integrating static strategy data and dynamic time series data, so that the features have time correlation; the time series features formed are adapted to the long short-term memory network processing demand, laying a foundation for the network to capture the time series correlation of data and improving the adaptability of input data and the model; the time series memory unit of the long short-term memory network is used to process the time series input features, effectively capturing the time dimension dependence relationship of data (such as the lagging influence of weather on heat load, the correlation of heat load between time periods); the future heat load demand trend is output to ensure that the trend prediction reflects the short, medium and long term correlation in time series and fits the actual heat load change rule; the room temperature standard and the minimum total energy consumption are used as the multi-objective optimization function to determine the dual core optimization target; the particle swarm optimization algorithm is used to solve the operation parameter combination, which can efficiently search the optimal parameter space under multiple constraints and quickly find the parameters that consider both heat demand and energy consumption control, improving the comprehensive rationality of the parameters; the abstract operation parameter combination obtained by the particle swarm optimization is converted into specific setting values of the boiler executable water supply temperature and circulating pump frequency; the optimization result is converted from the theoretical parameter to the equipment operation value, eliminating the disconnection between the parameter and the equipment operation, and improving the landing nature of the optimization result; the target output parameter setting value is directly sent to the boiler control system to realize the seamless connection between the data processing result and the real-time control of the equipment; the delay of data processing and control execution is avoided to ensure that the optimization setting value acts on the boiler operation in time, quickly responds to the heat load change, and improves the timeliness and effectiveness of regulation and control.

[0062] In a preferred embodiment of the present application, the step 503 inputs the time series input features into the long short-term memory network, processes through the time series memory unit of the network, and predicts the heat load demand trend in the future time period, including: Step 5031, inputting the time sequence input feature into the long short-term memory network, specifically comprising: first determining the time sequence input feature to be input, which is the structured data of a certain old community in the north for every hour in the next 24 hours, and each time step contains 5 dimensional information, which are the current hour outdoor temperature (such as -10℃ at 0 o'clock in the future, -15℃ at 6 o'clock, -8℃ at 14 o'clock), the wind speed of the area where the street buildings (No. 1 building to No. 2 building, No. 7 building to No. 8 building) are located (such as 3.0 m / s at 6 o'clock, 2.0 m / s at 14 o'clock), the wind speed of the area where the middle buildings (No. 3 building to No. 6 building) are located (such as 2.7 m / s at 6 o'clock, 1.7 m / s at 14 o'clock), the current hour solar radiation intensity (such as 0 W / m2 at 6 o'clock, 280 W / m2 at 14 o'clock), and the building target water supply temperature correlation value (76 to 78℃ for street buildings, 72 to 73℃ for middle buildings) in the corresponding period; arranging the time sequence input feature in chronological order into a feature sequence of 24 continuous time steps, and inputting the whole into the pre-trained long short-term memory network; the network takes the hourly data of the community in the past 2 heating seasons as the training basis, the samples contain the meteorological data and the actual value of the heat load of each building (calculated by the gas consumption of a 2.8 MW gas-fired hot water boiler and the pipe network parameters) in the same period, the parameters have been optimized, and the network can adapt to the time sequence data processing needs of the community heating system.

[0063] Step 5032, based on the time sequence input feature, calculating the retention weight of the current input information and the decay weight of the historical memory state through the input gate structure and the forget gate structure of the network respectively, and obtaining the input gate output result and the forget gate output result respectively, specifically comprising: based on the input time sequence input feature, starting the gate structure calculation of the network for the data of each time step; analyzing the key information of the current time step through the input gate structure, for example, in the time step of 6 o'clock in the future (outdoor temperature -15℃, wind speed of street area 3.0 m / s), the input gate identifies that low temperature and high wind speed are the core factors affecting the heat load, and calculates that the retention weight of outdoor temperature in this period is 0.8, the retention weight of wind speed is 0.7, and the retention weight of solar radiation intensity (0 W / m2) is 0.1, and the input gate output result is obtained; analyzing the useful information in the historical memory state through the forget gate structure, for example, when processing the data of 6 o'clock in the future, the forget gate calls the heat load associated memory of the same kind of cold wave period (such as -14℃ period in the past heating season) in the history, and calculates that the decay weight of this kind of effective historical memory is 0.2 (i.e. retaining 80% effective information), and the decay weight of the historical memory of the non-cold wave period (such as 10℃ period in the past) is 0.8 (i.e. retaining only 20% redundant information), and the forget gate output result is obtained.

[0064] At step 5033, the memory cell state of the long short-term memory network is weighted and fused and updated based on the input gate output result and the forget gate output result to obtain an updated cell state, specifically including: obtaining the input gate output result and the forget gate output result, and updating the memory cell state of the long short-term memory network; multiplying the forget gate output result and the memory cell state at the previous time step to retain effective information in the history, for example, retaining the past cold wave period low temperature corresponding to high load associated memory; multiplying the input gate output result and the information after the current time step feature is processed by the activation function to filter out the key feature information of the current time step, for example, the heat load influence feature corresponding to future 6 o'clock negative 15°C + high wind speed; adding the above two results to complete the weighted fusion and update of the memory cell state; the updated cell state contains both the heat load law of the historical cold wave period and the meteorological features of the current time step, for example, the future 6 o'clock updated cell state can reflect the small district exclusive law of high heat loss and high load of street buildings in negative 15°C environment.

[0065] At step 5034, the hidden state at the current time is calculated by the output gate based on the updated cell state, and the hidden state is added to the hidden state sequence to obtain an updated continuous hidden state sequence, specifically including: starting the output gate structure calculation based on the updated cell state; the output gate processes the updated cell state by the activation function to filter out the key features to be output at the current time, for example, the low temperature high wind speed and street high load associated features in the future 6 o'clock cell state, and the moderate temperature + solar radiation and intermediate building low load associated features in the future 14 o'clock (outdoor temperature negative 8°C, solar radiation 280W / ㎡) cell state; the hidden state at the current time is calculated according to the filtering result, each time step corresponds to a hidden state, and the hidden state contains the core time sequence features related to the heat load at that time; each time step calculated hidden state is sequentially added to the hidden state sequence in time order, and finally a continuous hidden state sequence covering the next 24 hours is formed, for example, the 6th element (corresponding to the future 6 o'clock) in the sequence is low temperature high wind speed and high load features, and the 14th element (corresponding to the future 14 o'clock) is moderate temperature + solar radiation and low load features, ensuring that the sequence can reflect the change of heat load influencing factors at different time periods.

[0066] At step 5035, based on the continuous hidden state sequence, a regression calculation is performed through a fully connected layer to predict the heat load demand change trend in the future time period, specifically including: inputting the continuous hidden state sequence into the fully connected layer of the long short-term memory network; the fully connected layer performs linear conversion and integration on each element in the hidden state sequence through a weight matrix, converting the abstract time sequence features in the hidden state into specific heat load values; during the conversion process, the regression calculation is performed on the street-side buildings and the middle buildings respectively in combination with the differences between the buildings in the cell, for example, for the elements of low temperature and high wind speed in the hidden state sequence, the heat load demand values of the street-side buildings (such as 130 kW in the next 6 hours) and the heat load demand values of the middle buildings (such as 105 kW in the next 6 hours) are obtained by regression; for the elements of moderate temperature + solar radiation, the heat load demand values of the street-side buildings (such as 110 kW in the next 14 hours) and the heat load demand values of the middle buildings (such as 90 kW in the next 14 hours) are obtained by regression; the regression results of all time steps are arranged in chronological order to form the heat load demand change trend every hour in the next 24 hours, which reflects the load fluctuation in the time dimension, such as high in low temperature period and low in high temperature period, and also distinguishes the load difference between the street-side buildings and the middle buildings.

[0067] In a preferred embodiment of the present application, step 504 described above is based on the heat load demand change trend, and a multi-objective optimization function is used to maintain the user room temperature standard and minimize the total energy consumption, and a particle swarm optimization algorithm is used to solve the final operating parameter combination, including: Step 5041, based on the heat load demand trend, a multi-objective optimization function is constructed with the minimum deviation of user room temperature and the lowest total energy consumption, specifically including: first, determine the heat load demand trend, which covers the heat load of the old community in the north for every hour in the future 24 hours, among which the heat load demand of the street buildings (No. 1 to No. 2, No. 7 to No. 8) reaches 130kW during the cold wave period (such as the outdoor temperature is negative 15℃ in the future 6 hours), the heat load demand of the middle buildings (No. 3 to No. 6) is 105kW during the same period, and the heat load gradually decreases with the increase of outdoor temperature during other periods; based on the trend, a multi-objective optimization function is constructed, which includes two core parts, the first part is to minimize the deviation of user room temperature, based on the room temperature data uploaded by 10% of the households in the community, the room temperature standard range is set to 18℃ to 22℃, the absolute deviation value of the actual room temperature and the range is calculated, the deviation of the street buildings and the middle buildings is calculated respectively and then averaged to ensure that the room temperature of all buildings is covered; the second part is to minimize the total energy consumption, taking the hourly gas consumption of a 2.8MW gas-fired hot water boiler as the measurement index, collecting the consumption data through the boiler gas flow sensor and converting it into standard energy consumption unit (m); in order to balance the two targets, the weight of the room temperature standard deviation is set to 0.6, and the weight of the total energy consumption is set to 0.4, the weighted sum of the two parts is obtained to get the comprehensive output value of the optimization function, the smaller the value represents the better parameter combination, realizing the joint optimization of room temperature standard and energy consumption control. 3 / h In step 5042, based on the multi-objective optimization function, the particle swarm algorithm parameters are initialized, the particle swarm size is set, and the initial position vector and initial velocity vector containing the boiler feed water temperature set value and the circulating pump frequency set value are randomly generated for each particle to obtain the initialized particle swarm, specifically including: based on the multi-objective optimization function, the core parameters of the particle swarm algorithm are initialized; the particle swarm size is set to 50 particles, which can cover enough parameter search space to avoid incomplete search due to small size, and can also avoid increasing the operation load due to large size (adapt to the real-time regulation and control requirements of the small district heating system); the position vector of each particle contains two dimensions, which are the boiler feed water temperature set value and the circulating pump frequency set value, wherein the value range of the feed water temperature set value is 65-85°C, and the value range of the circulating pump frequency set value is 30-50Hz (to ensure the normal operation of the circulating pump and to adapt to the water circulation demand under different heat loads through frequency adjustment); the initial position vector is randomly generated for each particle, and the parameter differences between the street-side buildings and the middle buildings need to be considered during generation, for example, for the particles corresponding to the high load period of the street-side buildings, the initial feed water temperature can be randomly taken in the range of 75-82°C, and for the particles corresponding to the low load period of the middle buildings, the initial feed water temperature can be randomly taken in the range of 68-75°C; at the same time, the initial velocity vector is generated for each particle, and the two dimensions of the velocity vector correspond to the adjustment step of the feed water temperature and the circulating pump frequency, respectively, wherein the velocity range of the feed water temperature is 0.5-2°C / iteration, and the velocity range of the circulating pump frequency is 0.3-1 Hz / iteration, to avoid that the initial velocity is too large to cause the parameters to jump out of the feasible region or too small to cause low search efficiency; after the initial vector of all particles is generated, the initialized particle swarm is obtained.

[0068] In step 5043, based on the initialized particle swarm, in each iteration, the fitness value of each particle is calculated according to the current position vector of each particle by the multi-objective optimization function; based on the fitness value, the individual historical final position of each particle and the global final position of the particle swarm are updated; and according to the updated individual historical final position and global final position, the velocity vector and position vector of each particle are adjusted, specifically including: based on the initialized particle swarm, the total number of iterations is set to 100 times, and the iteration calculation is started; in each iteration, the current position vector of each particle (i.e. a set of water supply temperature and circulating pump frequency parameters) is first obtained, the parameters are substituted into the multi-objective optimization function of step 5041, the fitness value corresponding to the particle is calculated, for example, the position vector of a certain particle is water supply temperature 78°C and circulating pump frequency 45Hz, after substituting into the function, the room temperature standard deviation average value is 0.4°C and the energy consumption is 8.2m, the comprehensive fitness value is calculated according to the weight 0.4x0.6+8.2x0.4, and the smaller the fitness value is, the better the parameter combination is; based on the calculated fitness value, the individual historical optimal position of each particle is updated, i.e. the fitness value of the current particle is compared with the optimal fitness value of all iterations of the particle in the past, if the current value is better, the current position vector is updated as the individual historical optimal position; at the same time, the global optimal position of the particle swarm is updated, i.e. the fitness values corresponding to the individual historical optimal positions of all particles are compared, and the position vector with the smallest fitness value is selected as the global optimal position; according to the updated individual historical optimal position and global optimal position, the velocity vector of each particle is adjusted, and when adjusting, the directions of the individual optimal and global optimal are referred to, for example, if the current position of a certain particle is far away from the global optimal position, the velocity component of the particle in the direction of the global optimal position is increased, if it is close to the global optimal position, the velocity component is reduced, so as to avoid excessive oscillation of the particle; then the position vector of the particle is updated according to the adjusted velocity vector, to ensure that the new position vector is still in the feasible region of water supply temperature 65°C to 85°C and circulating pump frequency 30Hz to 50Hz, if it exceeds, the position vector is corrected to the boundary value of the feasible region; the above calculation, updating and adjusting process is repeated until the current iteration is completed, to ensure that the overall fitness value of the particle swarm gradually decreases after each iteration, and the optimal parameter region is approached. 3 / h At step 5044, based on the global final position, when a preset convergence condition is met, the parameter combination corresponding to the global final position is taken as the final operation parameter combination, specifically including: after each iteration is completed, the change of the global optimal position of the particle swarm is monitored, and the preset convergence condition is that in the last 10 iterations, the variation of the supply water temperature corresponding to the global optimal position is less than 0.1 DEG C and the variation of the circulating pump frequency is less than 0.5 Hz, and the convergence precision can meet the requirement that the room temperature control deviation is less than 1 DEG C and the energy consumption control precision; if the convergence condition is not met, the next iteration is continued to be performed until the total number of iterations is reached or the convergence condition is met; when the convergence condition is met, the iteration is stopped, and the parameter combination corresponding to the global optimal position at this time is determined as the final operation parameter combination; for example, in the iteration corresponding to the future 6 hours (outdoor temperature is negative 15 DEG C, and the heat load of the adjacent building is 130 kW), the global optimal position after convergence is 78 DEG C of the supply water temperature and 45 Hz of the circulating pump frequency, and after the parameter combination is substituted into the optimization function, the average value of the room temperature standard deviation is 0.3 DEG C (satisfying the range of 18 DEG C to 22 DEG C and the deviation is small), and the energy consumption is 8.0 m; in the iteration corresponding to the future 14 hours (outdoor temperature is negative 8 DEG C, and the heat load of the intermediate building is 90 kW), the global optimal position after convergence is 72 DEG C of the supply water temperature and 38 Hz of the circulating pump frequency, and the parameter combination can make the room temperature standard deviation of the intermediate building 0.5 DEG C, and the energy consumption is 6.5 m; the final operation parameter combinations corresponding to different heat load periods are sorted and summarized to form a parameter set covering 24 hours in the future, so that the parameter combination of each period can meet the target of meeting the room temperature standard and the lowest total energy consumption. 3 / h 3 / h As Figure 2 shown, the embodiment of the present application also provides an operation data analysis system of a heating system, comprising: An acquisition module is configured to acquire equipment operation data, pipe network operation data and user side data in real time; perform time series alignment on the data based on a unified time reference, and perform feature dimension reduction to construct a multi-source fusion data set; A calibration module is configured to calculate key performance indicators of heat inlet of each building based on the multi-source fusion data set in real time; determine a plurality of distributed reference nodes based on the key performance indicators, a pipe network topology structure and real-time hydraulic working conditions; construct a hydraulic balance reference model according to data of the reference nodes, extract state features thereof, and generate dynamic calibration coefficients according to the state features; and calibrate the key performance indicators by using the dynamic calibration coefficients to obtain calibrated key performance indicators; An identification module is configured to compare real-time values of the calibrated key performance indicators with preset dynamic threshold intervals to obtain an indicator deviation state vector; identify an abnormal data cluster based on the indicator deviation state vector to obtain a diagnosis result containing an abnormal type and a location; The strategy generation module is configured to generate a differentiated operation strategy based on building historical room temperature, building characteristics and real-time meteorological parameters in the multi-source fused data set if the diagnosis result is normal. The optimization control module is configured to acquire meteorological prediction data in a preset time period, predict heat load demand variation through a long short-term memory network based on the differentiated operation strategy and the meteorological prediction data, and calculate a target output parameter set value based on the heat load demand variation, with the optimization target being that the room temperature meets the standard and the energy consumption is the lowest, and send the set value to a boiler control system for execution.

[0069] It should be noted that the system corresponds to the above method, and all implementation manners in the above method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0070] The above is the preferred embodiment of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should also be considered within the scope of protection of the present application.

Claims

1. A method of analyzing operation data of a heating system, characterized by, The method comprises: Step 100, collecting device operation data, pipe network operation data and user side data in real time; based on a unified time reference, the data is time aligned, and feature dimension reduction is performed to construct a multi-source fusion data set; Step 200, based on the multi-source fusion data set, real-time calculation of the key performance indicators of each building heat inlet; based on the key performance indicators, a plurality of distributed reference nodes are dynamically determined through the pipe network topological structure and the real-time hydraulic working condition; a hydraulic balance reference model is constructed according to the data of the reference nodes, and the state characteristics are extracted, and the dynamic calibration coefficient is generated according to the state characteristics; the key performance indicators are calibrated by using the dynamic calibration coefficient to obtain the calibrated key performance indicators; Step 300, comparing the real-time value of the calibrated key performance indicators with the preset dynamic threshold interval to obtain an index deviation state vector; based on the index deviation state vector, an abnormal data cluster is identified to obtain a diagnosis result containing an abnormal type and a location; Step 400, if the diagnosis result is no abnormality, generating a differentiated operation strategy based on the building historical room temperature, building characteristics and real-time meteorological parameters in the multi-source fusion data set; Step 500, obtaining meteorological prediction data of a preset time period; based on the differentiated operation strategy and the meteorological prediction data, predicting the heat load demand change through a long short-term memory network; based on the heat load demand change, taking the room temperature standard and the lowest energy consumption as the optimization goal, calculating the target output parameter set value, and sending the set value to the boiler control system for execution.

2. The operation data analysis method of a heating system according to claim 1, characterized by, Step 100 comprises: Collecting device operation data, pipe network operation data and user side data from three independent data sources, namely the boiler control system, the property monitoring platform and the user service system in parallel; Based on a unified time reference, the device operation data, pipe network operation data and user side data are time stamped and aligned to obtain a time stamped and aligned data sequence; the time stamped and aligned data sequence is subjected to time scale normalization processing by using a dynamic time warping algorithm to obtain an intermediate time series data set with consistent time resolution; Based on the intermediate time series data set, the characteristic vectors and characteristic values are calculated by a principal component analysis method to determine the main characteristic components and their contribution degrees; the main characteristic components are screened according to a preset variance contribution rate threshold to obtain a main characteristic subspace; The feature data in the main characteristic subspace is subjected to geometric projection transformation and mapped to a low-dimensional feature space to obtain the dimension-reduced feature projection data; Based on the dimension-reduced feature projection data, the feature information of the multi-source data streams is integrated by a feature weighted fusion method to construct a multi-source fusion data set with time sequence consistency, feature orthogonality and elimination of multicollinearity.

3. The operation data analysis method of a heating system according to claim 2, characterized by, Step 200 comprises: Based on the multi-source fusion data set, real-time calculation of the instantaneous supply and return water pressure difference, temperature difference and flow deviation rate key performance indicators of each building heat inlet; Based on the key performance indicators, the pipe network node characteristic parameters are extracted in combination with the pipe network topological structure and the real-time hydraulic working condition data; the key performance indicators and the node characteristic parameters are subjected to standardization processing to generate a standardized feature matrix; The standardized feature matrix is taken as input, a weighted Euclidean distance between each data sample is calculated by a clustering analysis algorithm, and a K-means clustering method is used to divide the pipe network nodes into multiple clusters to obtain a cluster division result; Based on the cluster division result, in combination with connectivity analysis of the pipe network topology structure, a node closest to a centroid and representative in each cluster is selected as a distributed reference node to complete dynamic identification and determination of the reference node.

4. The operation data analysis method of a heating system according to claim 3, characterized by, Step 200 further includes: Real-time operation data of the distributed reference node is collected; and a hydraulic balance reference model reflecting overall hydraulic state is established by using the real-time operation data and through a multivariate linear regression method; Pressure distribution features and flow balance features in the hydraulic balance reference model are extracted as state features; Correlation analysis is performed on the state features and standard reference values to obtain a correlation analysis result; based on the correlation analysis result, a mathematical relationship between the state features and the reference values is fitted by a least square method to obtain a fitting result; Based on the fitting result, weight coefficients of the state features are solved; and a dynamic calibration coefficient is calculated based on the weight coefficients; The dynamic calibration coefficient is applied to a key performance indicator to dynamically calibrate and correct the key performance indicator to obtain a calibrated key performance indicator.

5. The operation data analysis method of a heating system according to claim 4, characterized by, Step 300 includes: Based on the calibrated key performance indicator, real-time values thereof are compared with preset dynamic threshold intervals item by item, a relative deviation value of each indicator is calculated, and an indicator deviation state vector is obtained; The indicator deviation state vector is taken as input data; a reachable distance between each data sample point is calculated by performing a density clustering algorithm on the input data; a high-density area and an outlier are identified according to a distribution feature of the reachable distance to obtain a data point classification result; and based on the data point classification result, a preliminary abnormal data cluster is formed; Based on a spatial distribution feature of the preliminary abnormal data cluster, in combination with a pipe network topology structure relationship, the abnormal data cluster is subjected to merging and optimization processing to determine a final abnormal data cluster classification; Based on the final abnormal data cluster classification, statistical feature parameters of each abnormal data cluster are extracted, including an average value of a cluster-internal indicator deviation degree, a spatial distribution density and a duration feature; Based on the statistical feature parameters, a preset abnormal type judgment rule is matched by a rule reasoning method to determine an abnormal type and an occurrence position, and a diagnosis result containing the abnormal type and the position is obtained.

6. The operation data analysis method of a heating system according to claim 5, characterized by, Step 400 includes: If the diagnosis result is no abnormality, historical room temperatures, building characteristics and real-time meteorological parameters of each building are extracted from the multi-source fusion data set; The building historical room temperatures, the building characteristics and the real-time meteorological parameters are combined to construct a feature vector of each building; The feature vector is input into a pre-trained neural network, and an initial heating demand prediction value of each building is obtained through forward propagation calculation of the neural network; Based on the pipe network topology structure, a hydraulic coupling analysis is performed on the initial heating demand prediction value of each building to calculate a heat loss correction amount in a pipe network delivery process; and Superimpose the initial heat supply demand prediction value and the heat loss correction amount to obtain the differentiated heat supply demand target value of each building; According to the differentiated heat supply demand target value of each building, the corresponding target water supply temperature set value is calculated to form a differentiated operation strategy.

7. The operation data analysis method of a heating system according to claim 6, characterized by, Step 500 comprises: Obtain meteorological prediction data of a future preset time period, including outdoor temperature, wind speed and solar radiation intensity parameters; Based on the differentiated operation strategy and the meteorological prediction data, time series input features are constructed; The time series input features are input into a long short-term memory network, and the change trend of the heat load demand in the future time period is predicted through the time sequence memory unit of the network; Based on the change trend of the heat load demand, a multi-objective optimization function is constructed to maintain user room temperature standard and minimize total energy consumption, and a particle swarm optimization algorithm is used to solve the final operation parameter combination; The final operation parameter combination is converted into target output parameter set values of the boiler, including water supply temperature set value and circulating pump frequency set value; The target output parameter set values are sent to the boiler control system for real-time control.

8. The operation data analysis method of a heating system according to claim 7, characterized by, The time series input features are input into a long short-term memory network, and the change trend of the heat load demand in the future time period is predicted through the time sequence memory unit of the network, including: The time series input features are input into a long short-term memory network; Based on the time series input features, the retention weight of the current input information is calculated through the input gate structure of the network, and the decay weight of the historical memory state is calculated through the forget gate structure, and the input gate output result and the forget gate output result are obtained respectively; Based on the input gate output result and the forget gate output result, the memory cell state of the long short-term memory network is weighted and fused and updated to obtain the updated cell state; Based on the updated cell state, the hidden state at the current time is calculated through the output gate, and the hidden state is added to the hidden state sequence to obtain the updated continuous hidden state sequence; Based on the continuous hidden state sequence, regression calculation is performed through the fully connected layer to predict the change trend of the heat load demand in the future time period.

9. The operation data analysis method of a heating system according to claim 8, characterized by, Based on the change trend of the heat load demand, a multi-objective optimization function is constructed to maintain user room temperature standard and minimize total energy consumption, and a particle swarm optimization algorithm is used to solve the final operation parameter combination, including: Based on the change trend of the heat load demand, a multi-objective optimization function is constructed to maintain user room temperature standard and minimize total energy consumption; Based on the multi-objective optimization function, the particle swarm optimization algorithm parameters are initialized, the particle swarm size is set, and the initial position vector and initial velocity vector containing the boiler water supply temperature set value and circulating pump frequency set value are randomly generated for each particle to obtain the initialized particle swarm; Based on the initialized particle swarm, in each iteration, the fitness value of each particle is calculated through the multi-objective optimization function according to the current position vector of each particle; based on the fitness value, the individual historical final position and the global final position of the particle swarm are updated; according to the updated individual historical final position and global final position, the velocity vector and position vector of each particle are adjusted; Based on the global final position, when a preset convergence condition is met, a parameter combination corresponding to the global final position is taken as a final operation parameter combination.

10. A system for analyzing operational data of a heating system, the system implementing the method of any one of claims 1 to 9, characterized in that, Comprise: The acquisition module is used for collecting equipment operation data, pipe network operation data and user side data in real time; Based on the unified time reference, the data is time-aligned and the feature dimension is reduced to construct a multi-source fusion data set; The calibration module is used for calculating the key performance indicators of each building heat inlet in real time based on the multi-source fusion data set; based on the key performance indicators, a plurality of distributed reference nodes are dynamically determined through the pipe network topology and the real-time hydraulic working condition; a hydraulic balance reference model is constructed according to the data of the reference nodes, and the state characteristics are extracted, and the dynamic calibration coefficient is generated according to the state characteristics; the key performance indicators are calibrated by using the dynamic calibration coefficient to obtain the calibrated key performance indicators; The identification module is used for comparing the real-time values of the calibrated key performance indicators with the preset dynamic threshold interval to obtain an index deviation state vector; based on the index deviation state vector, an abnormal data cluster is identified to obtain a diagnosis result containing an abnormal type and a location; The strategy generation module is used for generating a differentiated operation strategy based on the building historical room temperature, building characteristics and real-time meteorological parameters in the multi-source fusion data set if the diagnosis result is normal; The optimization control module is used for obtaining meteorological prediction data of a preset time period; based on the differentiated operation strategy and the meteorological prediction data, the long short-term memory network is used to predict the heat load demand change; based on the heat load demand change, the target output parameter set value is calculated with the room temperature reaching the standard and the energy consumption being the lowest as the optimization target, and the set value is sent to the boiler control system for execution.

Images