Heat supply network load prediction system based on heating station load time-varying characteristics

Abstract

The invention relates to the field of heat supply network load prediction and intelligent regulation and control, and discloses a heat supply network load prediction system based on a heating station load time-varying characteristic, and the system comprises a data collection and preprocessing module which is used for obtaining system parameters, meteorological parameters and time variables of a heat source and a heating station, and carrying out the cleaning and alignment; and the classification module is used for receiving the preprocessed system parameters, grouping based on thermal delay time and refining grouping based on conforming time sequence characteristics to finish two-stage classification. By introducing a dual classification mechanism and using a future moment load prediction value of each group of heating stations as an input feature, a heating delay effect and a load time sequence feature are accurately considered, so that the problems of single classification and a hysteresis effect in a traditional method are overcome, the accuracy, the real-time performance and the reliability of load prediction are remarkably improved, and the load prediction efficiency is improved. Therefore, the system can more quickly and accurately cope with the change of the heating power demand.

Description

Technical Field

[0001] The present invention relates to the field of heat network load prediction and intelligent regulation, and specifically to a heat network load prediction system based on the time-varying characteristics of heat substation loads. Background Art

[0002] With the increasing global attention to energy efficiency and environmental protection, the goal of energy conservation and carbon reduction has become the common pursuit of all industries, and the central heating system is also developing towards more refined regulation. Accurate hourly load prediction plays a crucial role in achieving this goal. As the hub connecting the heat source and heat substations, the accuracy of load prediction for the primary heat supply network is directly related to the optimal scheduling of the entire heating system and the rational utilization of energy.

[0003] Currently, load prediction technologies have gradually evolved from traditional time series prediction and regression prediction to intelligent prediction methods using machine learning algorithms. These intelligent prediction technologies include shallow machine learning algorithms based on feature selection and deep learning algorithms that integrate feature extraction and model learning. By collecting various data (such as meteorological data, system historical data, etc.) and performing anomaly processing on this data, high-quality input information can be effectively provided for subsequent prediction. However, in the existing technology, although relatively mature machine learning algorithms are applied to heat network load prediction, there are still multiple problems. Especially in the feature selection and classification links, existing methods often fail to fully extract all potential influencing factors, resulting in less than ideal prediction effects.

[0004] The challenges faced by load prediction for the primary heat supply network are particularly prominent. Due to the significant nonlinearity, time delay, thermal inertia, and time-variability of the primary network, there are large deviations in the prediction accuracy of traditional models. The scale of the primary network is large, and the thermal response time is long. Especially for remote heat substations, there is usually a delay of 4 - 12 hours in the response. Existing technologies have not fully considered this delay characteristic, resulting in the prediction model being unable to accurately reflect the load lag phenomenon of remote heat substations.

[0005] Another significant problem is that existing technologies have deficiencies in dealing with the time-varying characteristics and input features of different types of heat substations. The primary network covers various types of heat substations, and there are differences in the main service objects and load demands of these heat substations. Traditional unified feature selection methods cannot fully capture the time-varying characteristics of various heat substations, resulting in some key features not being extracted, thus affecting the accuracy of the prediction results. Summary of the Invention

[0006] In view of the deficiencies of the prior art, the present invention provides a heat network load prediction system based on the time-varying characteristics of the heat substation load, which solves the problem that the existing load prediction model of the primary heat supply network fails to fully consider the heat delay characteristics and the deficiencies of the time-varying characteristics of the load.

[0007] To achieve the above objectives, the present invention is realized through the following technical solutions: A heat network load prediction system based on the time-varying characteristics of the heat substation load includes:

[0008] A data acquisition and preprocessing module, which is used to obtain the system parameters, meteorological parameters and time variables of the heat source and heat substations and perform cleaning and alignment;

[0009] A classification module, which is used to receive the preprocessed system parameters, complete two-level classification based on grouping by heat delay time and refinement grouping based on load time series characteristics;

[0010] A load prediction model construction module, which is used to receive the grouping strategy and delay time parameters from the classification module, construct grouped load prediction sub-models, and construct a total heat network load prediction model. Through the delay time parameters, the load at future moments predicted by each sub-model is aligned backward to the current moment t, and the total load value is calculated by integration, and the real-time total load prediction result is transmitted to the prediction result output module;

[0011] A prediction result output module, which is used to receive the total model prediction result and generate a heat network regulation instruction in combination with preset rules.

[0012] Preferably, the data acquisition and preprocessing module includes:

[0013] A sensor network unit, which is used to monitor and record in real time the supply water temperature, return water temperature, flow rate and historical heat supply of the heat source and heat substations;

[0014] A meteorological data interface unit, which is used to access the outdoor temperature, wind speed and solar radiation equivalent value of the meteorological data source;

[0015] A time variable generation unit, which generates date, time tags and holiday identifiers and encodes them into periodic features.

[0016] Preferably, the meteorological data interface unit obtains the predicted meteorological parameters for the next few hours in real time through the API.

[0017] Preferably, the classification module includes:

[0018] A first-level classification unit, which is used to calculate the heat delay time between the heat source and each heat substation through the supply water temperature time series and group the heat substations according to the heat delay time interval;

[0019] The second-level classification unit is used to perform secondary clustering on each group of heat stations based on the characteristics of load time-series fluctuations, and the number of clusters is optimized and determined through clustering evaluation indicators.

[0020] Preferably, the first-level classification unit groups the overall heat stations according to the heat delay time interval, satisfying:

[0021] N = N1 + N2 + … + N i + … + N H ;

[0022] where N i is the number of heat stations in the set with a heat delay time interval from i - 1 to i hours;

[0023] The second-level classification unit divides each group of heat stations into K subgroups based on the time-varying characteristics of the heat station load and using the clustering method. The number of subgroups k is determined by the clustering evaluation indicator, satisfying:

[0024] N i = N i1 + N i2 + … + N ij + … + N ik ;

[0025] where N ij is the number of heat stations in the jth subgroup of the ith group.

[0026] Preferably, the inputs of the load prediction model construction module include: historical meteorological parameters, time variable encoding, and future meteorological prediction values, and the output is the load prediction sequence for the next H hours corresponding to the heat station group.

[0027] Preferably, the output of the load prediction sub-model satisfies:

[0028] Q ij(t+0～H) = f(TW, VW, AAD, τ);

[0029] where τ is the encoded time variable, TW is the outdoor temperature, VW is the wind speed, AAD is the solar radiation equivalent value, and Q ij(t+0～H) is the load prediction value of the jth subgroup of the ith group of heat stations from the current moment to the next 1 - H hours;

[0030] Then the heat load of the ith group under the first-level classification is as follows:

[0031] Q i = Q i1 + Q i2 + … + Q ij + … + Q ik ;

[0032] where Q iRepresents the total load of heat sub - station i, Q ij Represents the load of the j - th subgroup under heat sub - station i, and k represents the total number of subgroups under heat sub - station i.

[0033] Preferably, the total heat network load prediction model calculates the total load by integrating the predicted values of each heat sub - station group at future moments, and it satisfies:

[0034] Q (t) = f(Q 1(t) , Q 1(t+1) , Q 2(t+1) , Q 2(t+2) , …, Q i(t+i-1) , Q i(t+i) , …, Q H(t+H-1) , Q H(t+H) )

[0035] Wherein, Q (t) Represents the predicted value of the total heat network load at time t, Q 1(t) Represents the predicted value of the load of the first group of heat sub - stations at the current moment, Q 1(t+1) Represents the predicted value of the load of the first group of heat sub - stations in the next 1 hour, Q 2(t+1) Represents the predicted value of the load of the second group of heat sub - stations in the next 1 hour, Q 2(t+2) Represents the predicted value of the load of the second group of heat sub - stations in the next 2 hours, Q i(t+i-1) Is the predicted value of the load of the i - th group of heat sub - stations in the next i - 1 hours, Q i(t+i) Represents the predicted value of the load of the i - th group of heat sub - stations in the next i hours, Q H(t+H-1) Represents the predicted value of the load of the H - th group of heat sub - stations in the next H - 1 hours, Q H(t+H) Represents the predicted value of the load of the H - th group of heat sub - stations in the next H hours.

[0036] The present invention provides a heat network load prediction system based on the time - varying characteristics of heat sub - station loads. It has the following beneficial effects:

[0037] 1. By introducing a dual - classification mechanism, the present invention first preliminarily groups heat sub - stations according to the heat delay time, and then refines the grouping according to the load time - series characteristics of each group of heat sub - stations. This innovative classification method enables each heat sub - station group to be accurately grouped according to its dynamic response characteristics, avoiding the problem of feature generalization caused by only using a single classification criterion in traditional methods. Compared with the prior art, this refined grouping strategy significantly improves the ability to capture complex time - series characteristics in load prediction and greatly enhances the prediction accuracy.

[0038] 2. In the present invention, the predicted future load values of each heat sub-station are used as input features to reflect the delay effect of the heat supply system in advance. Traditional load prediction models often fail to fully consider the heat supply delay, resulting in lagging prediction results. Especially in a system with a long response time like the primary pipe network, by introducing this innovative design, the present invention can significantly reduce the lagging effect in prediction, thereby improving the real-time performance and reliability of the load prediction results, enabling the entire system to respond more quickly and accurately to changes in heat demand. Brief Description of the Drawings

[0039] Figure 1 It is a framework diagram of the system of the present invention;

[0040] Figure 2 It is a flow chart of the system of the present invention;

[0041] Figure 3 It is a schematic diagram of the load prediction model for the primary heat supply pipe network of the present invention;

[0042] Figure 4 It is a schematic diagram of the primary heat supply pipe network of the present invention;

[0043] Figure 5 It is a schematic diagram of the heat supply pipe network system of the present invention. Detailed Embodiment

[0044] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0045] Please refer to the attached Figure 1 - attached Figure 2 , the embodiment of the present invention provides a heat network load prediction system based on the time-varying characteristics of heat sub-station loads, including:

[0046] A data acquisition and preprocessing module, which is used to obtain the system parameters, meteorological parameters and time variables of the heat source and heat sub-stations and perform cleaning and alignment;

[0047] Specifically, the data acquisition and preprocessing module is used to realize the acquisition and structured processing of multi-source heterogeneous data of the heat supply system, and provide input for downstream modules by establishing a data pool with a unified spatio-temporal reference.

[0048] In this embodiment, the construction of the module needs to solve the problem of data timeliness alignment under the topological association of the heat source - heat sub-station, and needs to be compatible with the hybrid processing mode of historical data backfilling and real-time data stream.

[0049] Specifically, the output data of this module needs to be synchronously transmitted to the system parameter parsing unit of the classification module and the feature input interface of the load forecasting model construction module to ensure that downstream tasks have a consistent input benchmark.

[0050] In some embodiments, the data acquisition and preprocessing module includes:

[0051] A sensor network unit that establishes a communication connection with the PLC controller or intelligent instrument of the heat source and the heat substation through the industrial Internet of Things protocol, and real-time monitors and records the water supply temperature, return water temperature, instantaneous flow rate, and cumulative heat supply data.

[0052] Generally, the communication protocol includes at least one of MQTT and ModbusTCP, and the device address mapping relationship is dynamically loaded through a preset configuration file. For example, in a possible implementation, for the temperature measurement requirements of a high-temperature water pipe network, the temperature sensor uploads data at a 1-minute interval, and the flow meter reports the cumulative value at a 5-minute interval in pulse accumulation mode.

[0053] A meteorological data interface unit that accesses future multi-period prediction data provided by a meteorological service provider through a RESTful API, including outdoor temperature, wind speed, and solar radiation equivalent value.

[0054] As an option, the unit has a built-in data caching mechanism that automatically switches to an alternative strategy of historical meteorological data in the same period when the API call fails.

[0055] Specifically, the outdoor temperature data is obtained at an hourly granularity, and the solar radiation equivalent value is generated by secondary conversion of the original cloud amount data through a surface radiation calculation model. For example, a calculation method that combines a clear sky radiation model based on astronomical formulas and a cloud attenuation coefficient is used.

[0056] A time variable generation unit that generates a date tag, an hourly timestamp, and a legal holiday identifier based on the system clock and encodes them into multi-dimensional features recognizable by a machine learning model.

[0057] In a possible implementation, the encoding method of the time variable includes:

[0058] Splitting the hour into sine / cosine periodic functions and performing Fourier transform encoding with a 24-hour period;

[0059] The holiday identifier uses one-hot encoding to distinguish among three states: working day, weekend, and legal holiday;

[0060] The date ordinal number is mapped to the annual cumulative number of days to represent the seasonal change trend.

[0061] Generally, the data cleaning and alignment process of the module includes the following steps:

[0062] Perform sliding window anomaly detection on the traffic data reported by the sensor network unit. When the data fluctuates beyond the preset threshold within three consecutive cycles, the linear interpolation method is used to repair the anomaly points.

[0063] Align the predicted meteorological parameters obtained by the meteorological data interface unit with the sensor data according to the acquisition timestamp. For data segments with a timestamp deviation exceeding 10 minutes, a realignment instruction is triggered.

[0064] The output of the time variable generation unit is forced to be synchronized with the above two types of data sources at the microsecond level according to the UTC time standard to generate a multi-dimensional data table with a unified timestamp.

[0065] As an option, the cleaned data output by the module is persisted in the Parquet columnar storage format, and a subset of system parameters is transmitted to the classification module through the Avro protocol, and a subset of meteorological and time variables is transmitted to the load forecasting model construction module through the Protobuf protocol. For example, the subset of system parameters includes at least fields such as the heat station ID, supply-return water temperature difference, instantaneous flow rate, and cumulative heat supply. The meteorological subset includes at least sequences of predicted temperature, wind speed, and radiation intensity every 15 minutes within the next 6 hours.

[0066] Please refer to Appendix Figure 3 - Appendix Figure 5 , a classification module, which is used to receive the preprocessed system parameters, calculate the thermal delay time and the load time series fluctuation characteristics, and complete two-level classification, including the delay time grouping calculated based on the mutual information method and the time series feature refinement grouping based on dynamic time warping;

[0067] The classification module includes a first-level classification unit and a second-level classification unit. The module realizes refined grouping through a two-level classification mechanism, and the specific implementation logic is as follows:

[0068] The first-level classification unit is used to perform initial grouping of heat stations according to the thermal delay time. Specifically, the unit calculates the mutual information between the heat source supply water temperature sequence and the supply water temperature sequences of each heat station to determine the thermal delay time of different heat stations. For example, the heat source temperature sequence is X, and the temperature sequence of a certain heat station is Y. The mutual information between the two is calculated at different time offsets τ, and the τ value that makes the mutual information reach the local minimum or maximum is selected as the delay time of the heat station. The delay time is divided into intervals according to a preset time granularity (such as 1 hour). Suppose the delay time of a certain heat station is 2.3 hours and the time granularity is 1 hour, then this heat station is classified into the group with a delay time interval of 2 to 3 hours (i.e., the second group). The initial grouping of all heat stations satisfies: the total number of heat stations N is equal to the sum of the number of heat stations in each group (N = N1 + N2 + … + N i + … + N H ), where N i represents the number of heat stations in the i-th group.

[0069] The second-level classification unit is used to refine the classification of heat stations within each initial group. Based on the load time-series fluctuation characteristics, the unit further divides the heat stations within the same initial group into several subgroups using a clustering algorithm. For example, for the second group with a delay time of 2 to 3 hours, if it contains 20 heat stations, the similarity is calculated according to the historical load curve shapes (such as peak time, fluctuation amplitude) of each heat station, and the optimal number of subgroups k (such as k = 3) is determined through silhouette coefficient optimization. The number of heat stations in each subgroup satisfies N i = N i1 + N i2 +…+ N ij +…+ N ik where N ij represents the number of heat stations in the jth subgroup of the ith group. The definition of the silhouette coefficient is as follows: The silhouette coefficient S of a certain heat station is the minimum average distance b to other subgroups minus the average distance a within its own subgroup, and then divided by the larger value of a and b. The average value of the silhouette coefficients of all heat stations is used as an evaluation index for the clustering quality.

[0070] In a possible implementation, the delay time interval of the first-level classification can be dynamically adjusted according to the pipe network topology parameters. For example, given that the pipe distance between a certain heat station and the heat source is 5 kilometers and the heat medium flow velocity is 2 m / s, the theoretical propagation time is 41 minutes, and the actual interval can be set to 30 to 60 minutes to accommodate the flow velocity fluctuations.

[0071] The following methods can be used for the similarity measurement in the second-level classification:

[0072] Extract the statistical characteristics of the load curve (such as peak time, daily average load rate, fluctuation variance), and calculate the difference through the Euclidean distance;

[0073] Divide the whole day into multiple time periods (such as morning, noon, evening), and calculate the load mean of each time period as the clustering feature respectively;

[0074] Normalize the load curve and directly compare the load change trends of adjacent time points.

[0075] The calculation of the silhouette coefficient needs to traverse multiple numbers of subgroups (such as k = 2 to 5), and select the k value that makes the silhouette coefficient the largest as the final number of groups. For example, when the silhouette coefficient of a certain initial group reaches the maximum value of 0.62 at k = 3, it is divided into 3 subgroups.

[0076] The grouped tags output by the module contain two levels of classification information. For example, the tag "Group 2 - Subgroup 3" indicates that this heat station belongs to the initial group with a delay time of 2 to 3 hours, and belongs to the 3rd subgroup in the second-level classification. The tags and the corresponding delay time parameters (such as the average delay time of 2.5 hours for Group 2) are passed to the load forecasting model construction module in a structured table form for the following purposes:

[0077] Sub-model training: Each subgroup corresponds to an independent load forecasting model, and the training data is divided according to the subgroup tags;

[0078] Time alignment compensation: The total model performs time offset on the predicted values of each subgroup according to the delay time parameter (for example, if the subgroup delay is 2.5 hours, the predicted future 2.5-hour load value is used to calculate the current total load).

[0079] The load forecasting model construction module is used to receive the grouping strategy and delay time parameters from the classification module, construct grouped load forecasting sub-models, and construct the total heat network load forecasting model. Through the delay time parameter, the load at future moments predicted by each sub-model is aligned backward to the current moment t, and the total load value is integrated and calculated, and the real-time total load forecasting result is passed to the forecasting result output module;

[0080] In the embodiment, the load forecasting model construction module is responsible for constructing the load forecasting sub-model for each group based on the heat station clustering grouping strategy and the heat delay time parameter, and finally constructing the total heat network load forecasting model. This module aligns the load at future moments predicted by each sub-model backward to the current moment, integrates the prediction results of each sub-model, and thus passes the real-time total load forecasting result to the forecasting result output module.

[0081] First, the load forecasting model construction module receives the grouping strategy and heat delay time parameters from the classification module. Based on these parameters, the module performs primary clustering grouping (Level-I classification) on the heat stations, and further refines the grouping according to the load time series characteristics (Level-II classification). Each clustering group is used as an independent unit to construct the heat load forecasting sub-model. Through this dual classification mechanism, this module can more precisely identify the dynamic response characteristics of each heat station, so as to design a suitable load forecasting model for each heat station group.

[0082] In the construction of the load forecasting sub-model, the long short-term memory neural network (LSTM) is selected in this embodiment. The LSTM network has strong sequence data modeling ability and can effectively learn the complex long-term dependencies in the time series. Through the input of historical meteorological data, time variables, and future meteorological prediction data, the LSTM model can predict the load value of each heat station group within the next H hours.

[0083] The inputs of the load forecasting sub-model include:

[0084] Historical meteorological parameters: including outdoor temperature (TW), wind speed (VW), equivalent solar radiation value (AAD), etc. These parameters reflect the impact of current environmental conditions on the load.

[0085] Time variable encoding: The time variable (τ) includes information such as date, hour, and holiday. These variables can help capture the periodic and seasonal variations of the load.

[0086] Future meteorological prediction values: including meteorological data such as predicted outdoor temperature, wind speed, and solar radiation at future times, which are used to improve the accuracy of prediction.

[0087] The output of the LSTM model is the load prediction value for each heat station group in the next H hours. Specifically, the output sequence form of the load prediction sub-model is:

[0088] Q ij(t+0～H) = f(TW, VW, AAD, τ);

[0089] Where: Q ij(t+0～H) is the load prediction value of the i-th group and j-th subgroup of heat stations at the current time and in the next 1 to H hours; TW is the outdoor temperature; VW is the wind speed; AAD is the equivalent solar radiation value; τ is the encoded time variable.

[0090] Then the heat load of the i-th group under the first-level classification is as follows:

[0091] Q i = Q i1 + Q i2 +…+ Q ij +…+ Q ik ;

[0092] Where, Q i represents the total load of heat station i, Q ij represents the load of the j-th subgroup under heat station i, and k represents the total number of subgroups under heat station i.

[0093] After the load prediction sub-model predicts the load sequences of each group of heat stations, this module integrates these prediction results to form a total load prediction model. The total load prediction model calculates the total load of the primary pipe network by taking the future load prediction values of each group of heat stations as inputs.

[0094] The form of the total load prediction model is:

[0095] Q (t) = f(Q 1(t) , Q 1(t+1) , Q 2(t+1) , Q 2(t+2) ,…, Q i(t+i-1) , Q i(t+i) ,…, QH(t+H-1) , Q H(t+H) )

[0096] Among them:

[0097] Q (t) represents the predicted value of the total heat network load at time t;

[0098] Q 1(t) represents the predicted value of the load of the first group of heat substations at the current moment;

[0099] Q 1(t+1) represents the predicted value of the load of the first group of heat substations in the next 1 hour;

[0100] Q 2(t+1) represents the predicted value of the load of the second group of heat substations in the next 1 hour;

[0101] Q 2(t+2) represents the predicted value of the load of the second group of heat substations in the next 2 hours;

[0102] Q i(t+i-1) is the predicted value of the load of the i-th group of heat substations in the next i - 1 hours;

[0103] Q i(t+i) represents the predicted value of the load of the i-th group of heat substations in the next i hours;

[0104] Q H(t+H-1) represents the predicted value of the load of the H-th group of heat substations in the next H - 1 hours;

[0105] Q H(t+H) represents the predicted value of the load of the H-th group of heat substations in the next H hours.

[0106] Through this structure, the predicted load values of all heat substation groups are aligned in reverse and transmitted to the current moment, thereby obtaining the real-time total load prediction value.

[0107] The prediction result output module is used to receive the total model prediction result and generate a heat network regulation instruction in combination with a preset rule.

[0108] The prediction result output module first receives the predicted result of the total heat network load generated by the load prediction model, and this result includes the prediction of the total heat network load at future moments. Next, according to the preset rules, the prediction result output module generates specific regulation instructions. These regulation instructions are used to guide the heating adjustment of each heat substation, optimize the load distribution, so as to meet the actual demand and improve the operation efficiency of the system.

[0109] In this embodiment, the working process of the module is roughly as follows:

[0110] Input Reception: The prediction result output module receives the total load prediction result from the load prediction model, specifically manifested as the total load prediction value of the primary pipeline network. This value includes the prediction for a future period (such as the next H hours). Assume the total load prediction result of the heat network is:

[0111] Q (t) =f(Q 1(t) ,Q 1(t+1) ,Q 2(t+1) ,Q 2(t+2) ,…,Q i(t+i-1) ,Q i(t+i) ,…,Q H(t+H-1) ,Q H(t+H) )

[0112] Where:

[0113] Q (t) represents the predicted total load value of the heat network at time t;

[0114] Q 1(t) represents the predicted load value of the first group of heat substations at the current moment;

[0115] Q 1(t+1) represents the predicted load value of the first group of heat substations in the next 1 hour;

[0116] Q 2(t+1) represents the predicted load value of the second group of heat substations in the next 1 hour;

[0117] Q 2(t+2) represents the predicted load value of the second group of heat substations in the next 2 hours;

[0118] Q i(t+i-1) is the predicted load value of the i-th group of heat substations in the next i - 1 hours;

[0119] Q i(t+i) represents the predicted load value of the i-th group of heat substations in the next i hours;

[0120] Q H(t+H-1) represents the predicted load value of the H-th group of heat substations in the next H - 1 hours;

[0121] Q H(t+H) represents the predicted load value of the H-th group of heat substations in the next H hours.

[0122] Generating Regulation Instructions: The prediction result output module generates regulation instructions for each heat substation based on the total load prediction value through a series of preset rules. These rules adjust the heating parameters of the heat substations according to the operating state of the system, load demand, and actual response characteristics of the heat substations. Common regulation parameters include supply water temperature, flow rate, return water temperature, etc.

[0123] For example, when the predicted load of the system approaches the maximum carrying capacity, the regulation rule may require reducing the heating load of the remote heat sub-station; while in the case of low load, the heat supply in some areas may be increased to meet the demand.

[0124] Consider the thermal delay effect: In the heat network system, there is a strong thermal delay effect, that is, the response of the remote heat sub-station to load changes usually lags behind the changes of the heat source. Therefore, the prediction result output module also needs to consider this delay effect and adjust the heating strategy of the heat sub-station in advance according to the load prediction result.

[0125] Output of regulation instructions: Finally, the prediction result output module transmits the generated regulation instructions to the control systems of each heat sub-station. The control systems of each heat sub-station adjust the actual heating parameters according to these instructions to ensure the reasonable distribution of the entire heat network load and the stable operation of the system.

[0126] Through the foregoing steps, the prediction result output module can convert the total load prediction value calculated by the load prediction model into specific heat network regulation instructions. Specifically, the total load prediction value reflects the overall load demand of the heat network in a future period. Based on these demands, the regulation instructions will be adjusted in the following ways:

[0127] When the load is too high, the system will reduce the heating capacity of some heat sub-stations to reduce unnecessary load.

[0128] When the load is too low, the system can appropriately increase the heat supply to meet the demands of different areas.

[0129] Working principle: When the system of the present invention is running, first of all, the data acquisition and preprocessing module is responsible for collecting real-time data from the heat source, heat sub-station and meteorological sensors. These data cover various parameters of the heat network system, such as supply water temperature, return water temperature, flow rate, etc., as well as meteorological information such as outdoor temperature, wind speed, solar radiation, etc. Through the sensor network and the meteorological data interface unit, the system can ensure the comprehensiveness and timeliness of the data. In addition, the time variable generation unit captures factors such as date, time, and holidays through periodic coding, providing important time features for the prediction model.

[0130] The data after data collection will enter the classification module. This module analyzes the response characteristics of the heat stations and divides them into multiple groups. The system initially groups them by calculating the heat delay time of each heat station. The calculation of the heat delay time is based on the mutual information of the water supply temperature time series between the heat source and each heat station, so as to determine the differences in heat response among heat stations. Next, the similarity of the load time series characteristics within each group is calculated through the dynamic time warping method to refine the clustering of each group. On this basis, the load prediction model construction module starts to run. In this module, the load prediction sub-model adopts a long short-term memory neural network, which is good at processing complex data with time-dependent relationships. The load prediction model for each heat station group predicts the load change within the next H hours by inputting historical meteorological data, time variables, and future meteorological prediction values. The output of each prediction sub-model is the load prediction value at a future moment, and the prediction results are integrated to form a total heat network load prediction model.

[0131] The total heat network load prediction model not only considers the load predictions of each heat station group, but also aligns the predicted loads of each group at future moments back to the current moment. Finally, the prediction result output module receives the results from the total load prediction model. Based on these prediction data, the system generates heat network control instructions in combination with preset control rules. The control instructions ensure the balance of loads in each area by optimizing the heat supply of the heat stations. At high loads, the system may adjust the output of remote heat stations to avoid overload; while at low loads, the system will reasonably adjust the heating to avoid energy waste. The adjustment of each instruction is calculated in real time according to specific situations to ensure that the system always operates in an optimal state.

[0132] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A heat network load prediction system based on the time-varying characteristics of heat station load, characterized in that: include: Data acquisition and preprocessing module, used to obtain system parameters, meteorological parameters and time variables of heat sources and thermal power stations and perform cleaning and alignment; The classification module is used to receive the pre-processed system parameters, group them based on the thermal delay time and refine the groups based on the load timing characteristics, and complete the two-level classification; The load forecasting model building module is used to receive the grouping strategy and delay time parameters from the classification module, build the grouped load forecasting sub-model, and build the total load forecasting model of the heating network. Through the delay time parameter, the future load predicted by each sub-model is reversely aligned to the current time t, the total load value is integrated and calculated, and the real-time total load forecasting result is transmitted to the forecasting result output module; The prediction result output module is used to receive the overall model prediction results and generate heating network control instructions in combination with preset rules.

2. A heat network load prediction system based on the time-varying characteristics of heat station load according to claim 1, characterized in that: The data acquisition and preprocessing module includes: Sensor network unit, used to monitor and record the supply water temperature, return water temperature, flow rate and historical heating supply of heat source and thermal power station in real time; Meteorological data interface unit, used to access the outdoor temperature, wind speed and solar radiation equivalent value of the meteorological data source; The time variable generation unit generates date, time labels and holiday identifiers and encodes them as periodic features.

3. A heat network load prediction system based on the time-varying characteristics of heat station load according to claim 2, characterized in that: The meteorological data interface unit obtains the predicted meteorological parameters for the next several hours in real time through the API.

4. A heat network load prediction system based on the time-varying characteristics of heat station load according to claim 1, characterized in that: The classification module comprises: The first-level classification unit is used to calculate the thermal delay time between the heat source and each thermal power station through the water supply temperature time series, and to group the thermal power stations according to the thermal delay time interval; The second-level classification unit is used to perform secondary clustering on each group of thermal power stations based on the load time series fluctuation characteristics, and the number of clusters is determined by optimizing the clustering evaluation index.

5. A heat network load prediction system based on the time-varying characteristics of heat station load according to claim 4, characterized in that: The first-level classification unit groups the overall thermal power stations according to the thermal power delay time interval, satisfying: N=N1+N2+…+N i +…+N H ; Among them, N i is the number of heat stations with heat delay time interval from i-1 to i hours; The second-level classification unit divides each group of thermal power stations into K subgroups based on the time-varying characteristics of the thermal power station load and using a clustering method. The number of subgroups is determined by the clustering evaluation index and meets the following requirements: N i =N i1 +N i2 +…+N ij +…+N ik ; Among them, N ij is the number of thermal power stations in the jth subgroup of the i-th group.

6. A heat network load prediction system based on the time-varying characteristics of heat station load according to claim 1, characterized in that: The input of the load forecasting model building module includes: historical meteorological parameters, time variable codes and future meteorological forecast values, and the output is the load forecast sequence of the corresponding thermal power station group for the next H hours.

7. A heat network load prediction system based on the time-varying characteristics of heat station load according to claim 6, characterized in that: The load forecasting sub-model output satisfies: Its output satisfies: Q ij(t+0～H) =f(TW,VW,AAD,τ); Among them, τ is the encoded time variable, TW is the outdoor temperature, VW is the wind speed, AAD is the solar radiation equivalent value, Q ij(t+0～H) is the load forecast value of the jth subgroup thermal power station in the i-th group at the current time and in the next 1 to H hours; The heat load of group i under Class Ⅰ classification is as follows: Q i =Q i1 +Q i2 +…+Q ij +…+Q ik ; Among them, Q i represents the total load of thermal power station i, Q ij represents the j-th subgroup load under thermal power station i, and k represents the total number of subgroups under thermal power station i.

8. The heat network load prediction system based on the time-varying characteristics of heat station load according to claim 1 is characterized in that: The total load forecasting model of the heating network calculates the total load by integrating the future forecast values ​​of each heating station group, which satisfies: Q (t) =f(Q 1(t) ,Q 1(t+1) ,Q 2(t+1) ,Q 2(t+2) ,…,Q i(t+i-1) ,Q i(t+i) ,…,Q H(t+H-1) ,Q H(t+H) ) Among them, Q (t) represents the total load forecast value of the heating network at time t, Q 1(t) It represents the load forecast value of the first group of thermal power stations at the current moment, Q 1(t+1) It represents the load forecast value of the first group of thermal power stations in the next hour, Q 2(t+1) It represents the load forecast value of the second group of thermal power stations in the next hour, Q 2(t+2) It represents the load forecast value of the second group of thermal power stations in the next 2 hours, Q i(t+i-1) is the load forecast value of the i-th group of thermal power stations in the next i-1 hours, Q i(t+i) represents the load forecast value of the i-th group of thermal power stations in the next i hours, Q H(t+H-1) It represents the load forecast value of the Hth group of thermal power stations in the next H-1 hours, Q H(t+H) It indicates the load forecast value of the Hth group of thermal power stations in the next H hours.

Images