Heat supply prediction method based on spatial-temporal feature fusion deep learning

Abstract

The invention relates to a heat supply prediction method and system based on spatial-temporal feature fusion deep learning, and the method comprises the following steps: S1, carrying out the collection and fusion of multi-source heterogeneous data, and constructing an integrated data set; s2, preprocessing the data; s3, constructing a graph structure model of the heat supply system, and constructing a weighted undirected graph; s4, constructing and executing forward calculation of the space-time double-flow deep network; s5, designing a composite loss function including mean square error loss and physical constraint loss, and performing joint optimization training on the space-time double-flow deep network; and S6, performing multi-step heat supply load prediction by using the trained model, outputting a heat supply load curve of each heat exchange station in a specified time period in the future, and integrating a prediction result with a heat supply scheduling system. The method has the advantages that the prediction precision is improved compared with that of a traditional machine learning model by capturing the spatial-temporal characteristics at the same time, and the advantages are more remarkable in extreme weather.

Description

Technical Field

[0001] This invention relates to a heating prediction method based on spatiotemporal feature fusion deep learning, belonging to the interdisciplinary field of heating system automation and artificial intelligence. Background Technology

[0002] Currently, clean energy heating systems in northern cities generally employ control methods based on empirical thresholds or traditional machine learning, which have significant technical drawbacks:

[0003] (1) Existing methods treat different heat exchange stations as independent entities, ignoring the spatial correlation of heating load between geographically adjacent heat exchange stations, as well as the propagation effect of meteorological conditions in the region.

[0004] (2) Heating systems have large inertia and long delay characteristics. The current heating load is not only affected by the current weather, but also strongly correlated with the weather sequence and system state over the past few hours or even tens of hours. Traditional linear models or simple recursive networks are difficult to effectively capture this long-range time dependence;

[0005] (3) During extreme weather events such as cold waves and strong winds that exceed the historical average range, the accuracy of model predictions based on historical average patterns drops sharply. The lack of specialized learning and rapid response mechanisms for extreme patterns leads to scheduling delays.

[0006] The current solution is as follows:

[0007] (1) Traditional methods such as linear regression and ridge regression are simple to calculate but cannot capture nonlinear relationships and have limited prediction accuracy;

[0008] (2) Traditional machine learning models: Gradient boosting tree models perform well in point prediction, but they are essentially static models that cannot model dynamic spatiotemporal processes. Furthermore, the interpretability of these models is limited to feature importance and cannot provide spatiotemporal attribution analysis. Summary of the Invention

[0009] To overcome the shortcomings of existing technologies, this invention provides a heating prediction method based on spatiotemporal feature fusion deep learning. The technical solution of this invention is as follows:

[0010] A heating forecasting method based on spatiotemporal feature fusion deep learning includes the following steps:

[0011] S1. Acquisition and fusion of multi-source heterogeneous data to construct an integrated dataset;

[0012] S2. Preprocess the data, including performing sine-cosine transformation on the wind direction data to convert the wind direction data from a scalar to a vector representation, and using spatiotemporal Kriging interpolation on the missing data to preserve the spatial continuity of the data.

[0013] S3. Construct a graph structure model of the heating system, with each heat exchange station as a node, calculate the edge weights based on the geographical distance between stations and the topological connection of the pipeline network, and construct a weighted undirected graph.

[0014] S4. Construct and execute the forward computation of the spatiotemporal dual-stream deep network;

[0015] S5. Design a composite loss function that includes mean squared error loss and physical constraint loss, and perform joint optimization training on the spatiotemporal dual-stream deep network;

[0016] S6. Use the trained model to perform multi-step heating load prediction, output the heating load curves of each heat exchange station in the future specified time period, and integrate the prediction results with the heating dispatch system.

[0017] Step S1 specifically involves: collecting temperature, air pressure, wind speed, wind direction, and humidity data from meteorological stations within the region, primary heating, return heating, and pressure parameters from heat exchange stations, as well as the building age and insulation coefficient static attributes of the service area; integrating the above meteorological data, pipeline operation parameters, and building attribute data to construct an integrated dataset, providing a multi-source heterogeneous data foundation for subsequent spatiotemporal modeling.

[0018] Step S2 specifically involves:

[0019] The specific formula for performing a sine-cosine transform on wind direction data is as follows:

[0020]

[0021] The calculation formula for the spatiotemporal Kriging interpolation is: Z(s0)=∑λ i Z(s i ),∑λ i =1;

[0022] Where Z(s0) is the point to be interpolated, Z(s i ) represents the nearest observation point, λ i It is the weighting coefficient.

[0023] Step S3 specifically involves calculating edge weights based on the geographical distance between stations and the pipeline topology, using each heat exchange station as a node, and constructing a weighted undirected graph. The edge weight W of the weighted undirected graph is... ij The calculation formula is: Where, d ij Let A be the Euclidean distance between stations. ij Let be the pipeline connection strength matrix, α and β be the balance coefficients, and σ be the distance scale parameter.

[0024] Step S4 specifically involves:

[0025] A spatiotemporal dual-stream deep network is constructed using a time-stream module and a spatial-stream module;

[0026] The time-flow module includes a one-dimensional convolutional layer and a long short-term memory network with an attention mechanism, used to extract local temporal patterns and focus on key historical moments;

[0027] The spatial flow module includes a multi-layer graph convolutional network for aggregating information from neighboring heat exchange stations and iteratively updating node feature representations;

[0028] The temporal features output by the temporal stream module and the spatial features output by the spatial stream module are spliced ​​and fused in the fully connected layer to complete the forward computation of the spatiotemporal dual-stream features.

[0029] Step S5 specifically involves:

[0030] Design a composite loss function Ltotal to perform joint optimization training on the graph structure model. Its expression is: L total =L MSE +λ·L physics ;

[0031] Among them, L MSE Mean squared error loss, used to measure the deviation between the predicted value and the true value; L physics λ represents the physical constraint loss, used to embed thermodynamic knowledge into the model in the form of soft constraints; λ is the adaptive weighting coefficient used to balance the two losses.

[0032] The physical constraint loss L physics Based on the heat balance equation, its expression is as follows:

[0033] Among them, Q pred The model predicts the heat load, where c is the specific heat capacity of water, m is the mass flow rate, and ΔT is the heat load. pred The predicted supply and return water temperature difference; during the training process, the adaptive weight coefficient λ is adjusted using an exponential decay strategy: λ=λ0·exp(-epoch / τ);

[0034] Where λ0 is the initial weight, epoch is the current training epoch, and τ is the decay constant.

[0035] Step 6 specifically includes:

[0036] The trained spatiotemporal dual-stream deep network is used for multi-step heating load prediction, and its prediction output is expressed as:

[0037] Where H = 72 is the prediction step size, representing the output heating load curves of each heat exchange station in the next 72 hours; T is the length of the input historical sequence; G is the heating system diagram structure; Θ is the model parameter;

[0038] The predicted heating load curve is integrated with the heating dispatching system through an API interface to enable advance control based on the prediction results.

[0039] A prediction system based on the aforementioned heating prediction method includes:

[0040] The data acquisition and fusion module is used to collect meteorological data, pipeline operation parameters, and building static attributes, and to build an integrated dataset.

[0041] The data preprocessing module is used to perform vector transformation on the wind direction data in the integrated dataset and to perform spatiotemporal interpolation on the missing data.

[0042] The graph structure modeling module is used to construct a weighted undirected graph to represent the spatial structure of the heating system based on the geographical location of each heat exchange station and the topology of the pipeline network.

[0043] The spatiotemporal feature extraction module includes a temporal flow submodule and a spatial flow submodule, which are used to extract temporal features from historical sequences and spatial features from graph structures, respectively, and then fuse the two.

[0044] The model training module is used to perform joint optimization training on the neural network corresponding to the spatiotemporal feature extraction module using a composite loss function that includes physical constraints;

[0045] The load forecasting module is used to output heating load forecast curves for multiple future steps based on real-time input data and the graph structure, using the trained model.

[0046] The system integration interface is used to transmit the prediction curve output by the load prediction module to the heating dispatching system.

[0047] The advantages of this invention are:

[0048] (1) Significantly improved prediction accuracy: By capturing spatiotemporal features simultaneously, the prediction accuracy is improved compared to traditional machine learning models, and the advantages are more significant under extreme weather conditions.

[0049] (2) Enhanced predictability and decision support: It can predict the inflection point of heating load more accurately 3-6 hours in advance, leaving sufficient response time for dispatch.

[0050] (3) Excellent generalization ability and robustness: The embedding of physical information enables the model to not only learn the statistical laws of data, but also conform to physical principles, and perform more robustly in scenarios not covered by training data (such as new sites or abnormal working conditions).

[0051] (4) High engineering practicality: The end-to-end multi-step predictive output and standard API interface design facilitate rapid integration with existing heating dispatching systems, forming an intelligent closed loop from prediction to control. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the method of the present invention.

[0053] Figure 2 This is a block diagram of the main structure of the system of the present invention. Detailed Implementation

[0054] The present invention will be further described below with reference to specific embodiments, and the advantages and features of the present invention will become clearer as a result. However, these embodiments are merely exemplary and do not constitute any limitation on the scope of the present invention. Those skilled in the art should understand that modifications or substitutions can be made to the details and form of the technical solutions of the present invention without departing from the spirit and scope of the present invention, but all such modifications and substitutions fall within the protection scope of the present invention.

[0055] See Figure 1 and Figure 2 This invention relates to a heating forecasting method based on spatiotemporal feature fusion deep learning, comprising the following steps:

[0056] S1. Acquisition and fusion of multi-source heterogeneous data to construct an integrated dataset;

[0057] S2. Preprocess the data, including performing sine-cosine transformation on the wind direction data to convert the wind direction data from a scalar to a vector representation, and using spatiotemporal Kriging interpolation on the missing data to preserve the spatial continuity of the data.

[0058] S3. Construct a graph structure model of the heating system, with each heat exchange station as a node, calculate the edge weights based on the geographical distance between stations and the topological connection of the pipeline network, and construct a weighted undirected graph.

[0059] S4. Construct and execute the forward computation of the spatiotemporal dual-stream deep network;

[0060] S5. Design a composite loss function that includes mean squared error loss and physical constraint loss, and perform joint optimization training on the spatiotemporal dual-stream deep network;

[0061] S6. Use the trained model to perform multi-step heating load prediction, output the heating load curves of each heat exchange station in the future specified time period, and integrate the prediction results with the heating dispatch system.

[0062] Steps S1 to S6 together constitute a complete and coordinated technical solution, the advantages of which are specifically reflected in the following aspects:

[0063] 1. Data Completeness and High Quality (corresponding to steps S1 and S2): Through the fusion of multi-source heterogeneous data in S1, an integrated dataset covering "meteorology-pipeline network-building" was constructed, overcoming the limitation of single data sources in traditional methods and providing a more comprehensive information foundation for the model. Combined with the vector transformation of wind direction and interpolation processing for spatial continuity in S2, the quality and physical rationality of the input data were effectively improved, laying a solid foundation for subsequent high-precision modeling.

[0064] 2. Explicit and Quantitative Modeling of Spatial Relationships (corresponding to step S3): Step S3 abstracts the physical heating network into a mathematical graph structure and explicitly and quantitatively encodes the spatial influence relationships between heat exchange stations by integrating the weight formulas of geographical distance and pipeline topology. This fundamentally solves the key deficiency of existing methods that treat each station as an independent entity and ignore spatial correlations.

[0065] 3. Effective Collaborative Extraction of Spatiotemporal Dynamic Features (corresponding to step S4): The spatiotemporal dual-stream deep network designed in step S4 achieves collaborative capture of long-range temporal dependencies and complex spatial propagation effects through parallel processing and feature fusion of temporal and spatial streams. This architecture ensures the integrity of spatiotemporal features and significantly improves the model's ability to understand the dynamic processes of the heating system.

[0066] 4. Enhanced Model Generalization Ability and Physical Rationality (corresponding to step S5): Step S5 introduces a physical constraint loss based on the thermal balance equation and performs joint optimization with adaptive weights and data-driven loss, embedding domain knowledge into the model in the form of soft constraints. This mechanism significantly improves the model's predictive stability and the physical rationality of the output results under unseen extreme conditions such as cold waves, avoiding absurd predictions that might occur with purely data-driven models.

[0067] 5. Forward-looking decision support and ease of engineering use (corresponding to step S6): Step S6 enables multi-step load curve prediction up to 72 hours, providing ample early warning and response time for the scheduling system. Simultaneously, integration with existing systems through standard API interfaces demonstrates the advantages of modular design, giving this advanced prediction method good deployability and engineering application value.

[0068] In summary, these interconnected steps systematically address the core issues mentioned in the background technology, such as ignoring spatial correlation, difficulty in capturing long-range dependencies, and weak generalization under extreme conditions, from data, models, and optimization to the application level. Together, they achieve a comprehensive advantage in improving prediction accuracy, enhancing generalization ability, ensuring physical rationality, and providing decision support.

[0069] Step S1 specifically involves: collecting temperature, air pressure, wind speed, wind direction, and humidity data from meteorological stations within the region, primary heating, return heating, and pressure parameters from heat exchange stations, as well as the building age and insulation coefficient static attributes of the service area; integrating the above meteorological data, pipeline operation parameters, and building attribute data to construct an integrated "meteorological-pipeline-building" dataset, providing a complete data foundation for spatiotemporal modeling and overcoming the limitations of traditional methods that only consider a single data source.

[0070] Step S2 specifically involves:

[0071] The specific formula for performing a sine-cosine transform on wind direction data is as follows:

[0072]

[0073] The calculation formula for the spatiotemporal Kriging interpolation is: Z(s0)=∑λ i Z(s i ),∑λ i =1;

[0074] Where Z(s0) is the point to be interpolated, Z(s i ) represents the nearest observation point, λ i It is the weighting coefficient.

[0075] By performing a vector transformation that periodically encodes wind direction data, the ambiguity in model learning caused by traditional scalar representation is resolved. At the same time, spatiotemporal Kriging interpolation is used to maintain spatial and temporal continuity when repairing missing data, together providing high-quality, physically interpretable normalized input for subsequent spatiotemporal depth models.

[0076] Step S3 specifically involves calculating edge weights based on the geographical distance between stations and the pipeline topology, using each heat exchange station as a node, and constructing a weighted undirected graph. The edge weight W of the weighted undirected graph is... ij The calculation formula is: Where, d ij Let A be the Euclidean distance between stations. ij Let be the pipeline connection strength matrix, α and β be the balance coefficients, and σ be the distance scale parameter.

[0077] By constructing a weighted graph structure that integrates geographical distance and pipeline topology, we have for the first time achieved explicit and quantitative modeling of the spatial correlation between heat exchange stations in heating load forecasting. This fundamentally overcomes the shortcomings of traditional methods that treat each station as an independent entity and ignore spatial dependence, laying the foundation for subsequent graph neural networks to extract accurate spatial features.

[0078] Step S4 specifically involves:

[0079] A spatiotemporal dual-stream deep network is constructed using a time-stream module and a spatial-stream module;

[0080] The time-flow module includes a one-dimensional convolutional layer and a long short-term memory network with an attention mechanism, used to extract local temporal patterns and focus on key historical moments;

[0081] The spatial flow module includes a multi-layer graph convolutional network, which aggregates information from neighboring heat exchange stations and iteratively updates node feature representations. The temporal features output by the temporal flow module and the spatial features output by the spatial flow module are concatenated and fused in a fully connected layer to complete the forward computation of spatiotemporal dual-flow features.

[0082] By constructing a dual-stream network architecture that processes and merges temporal and spatial streams in parallel, this study achieves for the first time synchronous and collaborative modeling of the long-range temporal dependence and complex spatial correlation of heating systems. This effectively overcomes the fundamental deficiency of traditional static models or single-dimensional modeling methods in failing to capture dynamic spatiotemporal coupling processes, and provides feature representations for high-precision prediction.

[0083] Step S5 specifically involves:

[0084] Design a composite loss function Ltotal to perform joint optimization training on the graph structure model, its expression

[0085] Among them, L MSE Mean squared error loss, used to measure the deviation between the predicted value and the true value; L physics λ represents the physical constraint loss, used to embed thermodynamic knowledge into the model in the form of soft constraints; λ is the adaptive weighting coefficient used to balance the two losses.

[0086] The physical constraint loss L physics Based on the heat balance equation, its expression is as follows:

[0087] Among them, Q pred The model predicts the heat load, where c is the specific heat capacity of water, m is the mass flow rate, and ΔT is the heat load. pred The predicted supply and return water temperature difference; during the training process, the adaptive weight coefficient λ is adjusted using an exponential decay strategy: λ=λ0·exp(-epoch / τ);

[0088] Where λ0 is the initial weight, epoch is the current training epoch, and τ is the decay constant.

[0089] By designing a composite objective function that integrates data-driven loss and physical constraint loss, and by adopting an adaptive weight adjustment strategy, thermodynamic knowledge is embedded into the deep learning model in the form of differentiable soft constraints. This effectively enhances the model's prediction stability and physical rationality under extreme and untrained conditions, fundamentally alleviating the problem that pure data-driven models may produce outputs that violate physical laws.

[0090] Step 6 specifically includes:

[0091] The trained spatiotemporal dual-stream deep network is used for multi-step heating load prediction, and its prediction output is expressed as:

[0092] Where H = 72 is the prediction step size, representing the output heating load curves of each heat exchange station in the next 72 hours; T is the length of the input historical sequence; G is the heating system diagram structure; Θ is the model parameter;

[0093] The predicted heating load curve is integrated with the heating dispatching system through an API interface to enable advance control based on the prediction results.

[0094] By directly outputting the heating load curves of multiple sites for the next 72 hours and seamlessly integrating the forecast results with the existing dispatching system via API, a closed loop from accurate forecasting to proactive control is achieved. This provides the heating system with sufficient decision-making response time to cope with load changes, and significantly improves the foresight and automation level of dispatching decisions.

[0095] The present invention also relates to a prediction system based on the aforementioned heating prediction method, comprising:

[0096] Data acquisition and fusion module 1 is used to collect meteorological data, pipeline operation parameters and building static attributes, and to construct an integrated dataset;

[0097] Data preprocessing module 2 is used to perform vector transformation on the wind direction data in the integrated dataset and to perform spatiotemporal interpolation on the missing data;

[0098] Graph structure modeling module 3 is used to construct a weighted undirected graph to represent the spatial structure of the heating system based on the geographical location of each heat exchange station and the topology of the pipeline network.

[0099] The spatiotemporal feature extraction module 4 includes a temporal flow submodule and a spatial flow submodule, which are used to extract temporal features from historical sequences and spatial features from graph structures, respectively, and then fuse the two.

[0100] Model training module 5 is used to perform joint optimization training on the neural network corresponding to the spatiotemporal feature extraction module using a composite loss function that includes physical constraints;

[0101] The load prediction module 6 is used to output a heating load prediction curve for the next multiple steps based on the trained model, real-time input data, and the graph structure.

[0102] System integration interface 7 is used to transmit the prediction curve output by the load prediction module to the heating dispatching system.

[0103] The system comprehensively covers the entire process from raw data input to final scheduling instruction generation, realizing an automated pipeline for data preprocessing, model training, real-time prediction, and system integration. This greatly reduces manual intervention and improves the efficiency and response speed of prediction operations.

[0104] This invention provides a heating forecasting method and system based on spatiotemporal feature fusion deep learning. Its working principle involves systematically integrating multi-source data, explicitly modeling spatiotemporal correlations, embedding physical knowledge constraints, and ultimately achieving forward-looking automated scheduling decisions. The entire process forms a closed-loop intelligent chain from data to decision-making. The specific working principle is described in the following steps:

[0105] Phase 1: Constructing a high-quality spatiotemporal data foundation

[0106] The system first aggregates data from multiple sources. For example, data from two weather stations was collected: wind speeds of 3.5 m / s and 4.1 m / s, wind directions of 30° and 210°, and an air temperature of -5°C. Operational data from three heat exchange stations (A, B, and C) was also collected: primary water supply temperatures of 85°C, 87°C, and 83°C, return water temperatures of 45°C, 47°C, and 44°C, and instantaneous loads of 18MW, 22MW, and 15MW, respectively. Simultaneously, the average building insulation coefficient of the service area was obtained as 0.7 W / (m²). 2 ·K).

[0107] In preprocessing, the system converts the wind direction scalar into a vector. For example, for a wind direction of 30°:

[0108] Wind direction_cos=cos(π / 180×(90-30))=cos(60°=0.5

[0109] Wind direction sin = sin(π / 180 × (90-30)) = sin(60°) ≈ 0.866. This transformation eliminates the numerical discontinuity between 0° and 360° wind direction. For missing temperature data, the system uses spatiotemporal kriging interpolation to estimate and fill in missing values ​​using data from neighboring stations at different times (e.g., -4°C one hour before and -6°C one hour after a neighboring station), ensuring spatiotemporal continuity of the data.

[0110] Phase Two: Quantitative Spatial Correlation and Co-spatial Feature Extraction

[0111] The system abstracts the physical heating network into a mathematical graph. Taking three heat exchange stations A, B, and C as nodes, and given their geographical coordinates (e.g., obtained via GPS), the Euclidean distance between the stations is calculated: d. AB =1.2km, d BC =0.8km,d AC =1.5km. Meanwhile, based on the pipeline topology, it is determined that A and B are directly connected, with a connection strength of 1; B and C are directly connected, with a strength of 1; A and C are not directly connected, with a strength of 0. Setting parameters α = 0.7, β = 0.3, σ = 1.0, the edge weights are calculated as follows:

[0112] W AB = 0.7 × exp(-1.2) 2 / 1 2 ) + 0.3 × 1 ≈ 0.7 × 0.237 + 0.3 = 0.466

[0113] W BC = 0.7 × exp(-0.8) 2 / 1 2 ) + 0.3 × 1 ≈ 0.7 × 0.527 + 0.3 = 0.669

[0114] This allows for the construction of a weighted graph that quantifies spatial relationships.

[0115] Simultaneously, the spatiotemporal dual-stream network is activated. The temporal stream receives historical weather and load sequences for each station over the past 24 hours, spaced at 1-hour intervals (a total of 24 time steps). One-dimensional convolution first extracts local patterns (e.g., identifying a cumulative temperature drop of 5°C over the past 3 hours), followed by an LSTM network with an attention mechanism focusing on historical moments most critical to the current prediction, such as "6 hours ago" and "18 hours ago". The spatial stream, based on the constructed graph, aggregates neighbor information through a two-layer graph convolutional network (GCN): for example, for station B, its representation incorporates approximately 46.6% of information from station A and 66.9% of information from station C. Finally, the dual-stream features are fused to form a joint feature vector that simultaneously encodes "station B's own historical load change pattern" and "the spatial influence of stations A and C".

[0116] Phase 3: Model Training and Optimization Integrating Physical Laws

[0117] Model training uses a composite loss function: L total =L MSE +λ·L physics For example, for a prediction at a certain moment: the model predicts the heat load Q at station B. pred =20.5MW, predicted supply and return water temperature difference ΔT pred=40℃, given that the specific heat capacity of water is c≈4.2kJ / (kg·℃), and the estimated mass flow rate m is 100kg / s, then the theoretical heat load should be c·m·ΔT. pred ≈4.2×100×40=16800kW=16.8MW. The physical loss term calculates the difference between the predicted value of 20.5MW and the theoretical value of 16.8MW (this is an example; in actual training, it will encourage the prediction to move closer to the theoretical value), penalizing results that deviate significantly from the physical laws.

[0118] In the initial training phase, the initial value of λ is set to λ0 = 0.5, and the decay constant τ = 10. As the number of training epochs increases, the value of λ decays exponentially; for example, in the 10th epoch, λ = 0.5 × exp(-10 / 10) ≈ 0.184. This strategy allows the model to focus on learning physical laws in the early stages and on optimizing data fitting accuracy in the later stages.

[0119] Phase 4: Long-term forecasting and system integration regulation

[0120] The trained model performs multi-step rolling forecasts. For example, at the current time t, the model is input with the latest 24-hour historical sequence and a fixed graph structure G, and directly outputs the load forecast curves for each hour for the next 72 hours (H=72).

[0121] The forecast results show that, affected by the upcoming cold wave, the load of station A will rise from the current 18MW to a peak of 32MW in 18 hours; the load of station B will rise from 22MW to a peak of 28MW in 20 hours. The system has verified through physical constraints that this upward trend matches the predicted temperature difference change, which is reasonable.

[0122] These specific, quantitative forecast curves are pushed to the existing heating dispatching system in real time via an interface in JSON format. The dispatching system then generates control instructions based on these instructions: for example, 12 hours before the peak load of a cold wave, the system automatically raises the outlet water temperature of the heat source plant by 3°C and increases the frequency of the circulating pumps in the ring network where stations A and B are located by 5Hz.

[0123] In summary, this invention transforms an abstract deep learning framework into a quantifiable, verifiable, and executable engineering solution by introducing specific numerical processing (such as wind direction vector transformation and graph weight calculation), spatiotemporal modeling (such as 24-hour history and 72-hour prediction), and physical constraints (such as calculation based on specific heat capacity and temperature difference). This systematically solves the problems of spatiotemporal fragmentation, long-range dependence, and generalization to extreme conditions in heating forecasting.

[0124] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A heating forecasting method based on spatiotemporal feature fusion deep learning, characterized in that, Includes the following steps: S1. Acquisition and fusion of multi-source heterogeneous data to construct an integrated dataset; S2. Preprocess the data, including performing sine-cosine transformation on the wind direction data to convert the wind direction data from a scalar to a vector representation, and using spatiotemporal Kriging interpolation on the missing data to preserve the spatial continuity of the data. S3. Construct a graph structure model of the heating system, with each heat exchange station as a node, calculate the edge weights based on the geographical distance between stations and the topological connection of the pipeline network, and construct a weighted undirected graph. S4. Construct and execute the forward computation of the spatiotemporal dual-stream deep network; S5. Design a composite loss function that includes mean squared error loss and physical constraint loss, and perform joint optimization training on the spatiotemporal dual-stream deep network; S6. Use the trained model to perform multi-step heating load prediction, output the heating load curves of each heat exchange station in the future specified time period, and integrate the prediction results with the heating dispatch system.

2. The heating forecasting method based on spatiotemporal feature fusion deep learning according to claim 1, characterized in that, The specific steps of step S1 are as follows: collecting temperature, air pressure, wind speed, wind direction and humidity data from meteorological stations in the area, primary heating, return heating and pressure parameters of heat exchange stations, and static attributes of building age and insulation coefficient of the service area. By integrating the aforementioned meteorological data, pipeline operation parameters, and building attribute data, an integrated dataset is constructed, providing a multi-source heterogeneous data foundation for subsequent spatiotemporal modeling.

3. The heating forecasting method based on spatiotemporal feature fusion deep learning according to claim 1, characterized in that, Step S2 specifically involves: The specific formula for performing a sine-cosine transform on wind direction data is as follows: The calculation formula for the spatiotemporal Kriging interpolation is: Z(s0)=∑λ i Z(s i ),∑λ i =1; Where Z(s0) is the point to be interpolated, Z(s i ) represents the nearest observation point, λ i It is the weighting coefficient.

4. The heating forecasting method based on spatiotemporal feature fusion deep learning according to claim 3, characterized in that, Step S3 specifically involves calculating edge weights based on the geographical distance between stations and the pipeline topology, using each heat exchange station as a node, and constructing a weighted undirected graph. The edge weight W of the weighted undirected graph is... ij The calculation formula is: Where, d ij Let A be the Euclidean distance between stations. ij Let be the pipeline connection strength matrix, α and β be the balance coefficients, and σ be the distance scale parameter.

5. The heating forecasting method based on spatiotemporal feature fusion deep learning according to claim 4, characterized in that, Step S4 specifically involves: A spatiotemporal dual-stream deep network is constructed using a time-stream module and a spatial-stream module; The time-flow module includes a one-dimensional convolutional layer and a long short-term memory network with an attention mechanism, used to extract local temporal patterns and focus on key historical moments; The spatial flow module includes a multi-layer graph convolutional network for aggregating information from neighboring heat exchange stations and iteratively updating node feature representations; The temporal features output by the temporal stream module and the spatial features output by the spatial stream module are spliced ​​and fused in the fully connected layer to complete the forward computation of the spatiotemporal dual-stream features.

6. The heating forecasting method based on spatiotemporal feature fusion deep learning according to claim 5, characterized in that, Step S5 specifically involves: Design a composite loss function Ltotal to perform joint optimization training on the graph structure model. Its expression is: L total =L MSE +λ·L physics ; Among them, L MSE Mean squared error loss, used to measure the deviation between the predicted value and the true value; L physics λ represents the physical constraint loss, used to embed thermodynamic knowledge into the model in the form of soft constraints; λ is the adaptive weighting coefficient used to balance the two losses. The physical constraint loss L physics Based on the heat balance equation, its expression is as follows: Among them, Q pred The model predicts the heat load, where c is the specific heat capacity of water, m is the mass flow rate, and ΔT is the heat load. pred The predicted supply and return water temperature difference; during the training process, the adaptive weight coefficient λ is adjusted using an exponential decay strategy: λ=λ0·exp(-epoch / τ); Where λ0 is the initial weight, epoch is the current training epoch, and τ is the decay constant.

7. The heating forecasting method according to claim 6, characterized in that, Step 6 specifically includes: The trained spatiotemporal dual-stream deep network is used for multi-step heating load prediction, and its prediction output is expressed as: Where H = 72 is the prediction step size, representing the output heating load curves of each heat exchange station in the next 72 hours; T is the length of the input historical sequence; G is the heating system diagram structure; Θ is the model parameter; The predicted heating load curve is integrated with the heating dispatching system through an API interface to enable advance control based on the prediction results.

8. A prediction system based on the heating prediction method according to any one of claims 1 to 7, characterized in that, include: The data acquisition and fusion module is used to collect meteorological data, pipeline operation parameters, and building static attributes, and to build an integrated dataset. The data preprocessing module is used to perform vector transformation on the wind direction data in the integrated dataset and to perform spatiotemporal interpolation on the missing data. The graph structure modeling module is used to construct a weighted undirected graph to represent the spatial structure of the heating system based on the geographical location of each heat exchange station and the topology of the pipeline network. The spatiotemporal feature extraction module includes a temporal flow submodule and a spatial flow submodule, which are used to extract temporal features from historical sequences and spatial features from graph structures, respectively, and then fuse the two. The model training module is used to perform joint optimization training on the neural network corresponding to the spatiotemporal feature extraction module using a composite loss function that includes physical constraints; The load forecasting module is used to output heating load forecast curves for multiple future steps based on real-time input data and the graph structure, using the trained model. The system integration interface is used to transmit the prediction curve output by the load prediction module to the heating dispatching system.

Images