A district heating load prediction method and system

Abstract

This invention provides a method and system for predicting district heating load, relating to the field of heating load prediction technology. The method includes: extracting features from historical heating load data using a Kaplan-Answer (KAN) network to form mapped transformation data, which serves as input information for a Gaussian process model; introducing thermodynamic laws to constrain the parameter learning process of the KAN network, using thermodynamic law-derived data and actual sensor-collected data to guide parameter learning; processing the mapped transformation data based on the Gaussian process model to obtain heating load prediction data and an estimate of the uncertainty in the prediction data; simultaneously learning the parameters of the KAN network and the Gaussian process model within a unified framework to obtain a trained KAN-Gaussian process model; and inputting the collected district heating load data into the trained KAN-Gaussian process model to obtain the prediction result. This invention can output prediction results with uncertainty quantification, achieving accurate prediction and reliable assessment of heating load.

Description

Technical Field

[0001] This invention relates to the field of heating load forecasting technology, and more specifically, to a method and system for forecasting district heating load. Background Technology

[0002] Heating systems are gradually developing towards large-scale and intelligent systems. Intelligent heating systems achieve more refined management through real-time monitoring and centralized scheduling, which can improve energy efficiency to a certain extent. However, with the continuous improvement of infrastructure and the expansion of heating coverage, heating systems face more complex demand fluctuations and energy consumption constraints during unified allocation. How to achieve energy conservation and emission reduction while ensuring the stability of residential heating has become a key issue for the further development of intelligent heating systems. Currently, commonly used energy scheduling methods include Support Vector Machines (SVM), Recurrent Neural Networks (RNN), Artificial Neural Networks (ANN), and Long Short-Term Memory Networks (LSTM). These methods can improve the accuracy of load forecasting to a certain extent by mining the implicit relationships in historical operating data. However, they are generally data-driven methods, overly relying on high-quality and high-completeness data, and failing to fully consider the physical laws and mechanistic characteristics of the heating system. When data contains noise, missing data, or distribution drift, these models are prone to prediction bias or even failure, lacking reliable generalization ability.

[0003] In the process of developing this invention, the applicant discovered that current research on achieving efficient and reliable operation of heating systems mainly focuses on two paths: physical-driven and data-driven approaches. In the physical-driven approach, researchers construct mechanistic models of the heating process based on physical laws such as thermodynamics and fluid mechanics. By describing factors such as pipe network heat loss, building heat transfer, and equipment characteristics, they achieve prediction and control of system behavior. This type of method has good interpretability and physical consistency, reflecting the actual operating mechanism of the system. However, its application in large-scale and nonlinear systems is limited by the complexity of model parameters and the constraints of on-site measurement conditions. The data-driven approach relies on artificial intelligence and machine learning technologies to establish a mapping relationship between heating load and external environmental variables through the mining and learning of historical operating data. This type of method can automatically extract potential patterns from large amounts of data without complex physical modeling, achieving significant progress in prediction accuracy. However, this type of method generally suffers from two main problems: firstly, it is highly dependent on data quality and completeness. Heating system operation data is often affected by factors such as sensor accuracy, data gaps, and external interference. If the input data contains noise or anomalies, it will directly lead to a decline in model performance. Secondly, most studies still remain at the level of providing a single predicted value without characterizing the uncertainty of the prediction results. This makes it difficult for the model to provide risk assessment and confidence interval support for actual scheduling, which is not conducive to the heating system achieving efficient energy distribution and safe operation while ensuring heat supply.

[0004] Therefore, how to achieve accurate prediction and reliable assessment of heating load has become a technical problem that needs to be solved. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art or related technologies, and discloses a method and system for predicting district heating load, which realizes accurate prediction and reliable assessment of heating load, thereby providing technical support for energy-saving scheduling and safe operation of intelligent heating systems.

[0006] The first aspect of this invention discloses a method for predicting district heating load, comprising: Mapping transformation: Features of historical heat load data are extracted through the KAN network to form mapping transformation data, which serves as input information for the Gaussian process model; Parameter learning constraints: Thermodynamic laws are introduced to constrain the parameter learning process of the KAN network, and thermodynamic law-derived data and actual sensor data are used to guide parameter learning. Gaussian process processing: Based on the Gaussian process model, the data is processed and transformed to obtain heating load prediction data and an estimate of the uncertainty of the prediction data; End-to-end training: Simultaneously learn the parameters of the KAN network and the Gaussian process model within a unified framework to obtain the trained KAN-Gaussian process model. Heating load prediction: The collected regional heating load data is input into the trained KAN-Gaussian process model to obtain prediction results, which include heating load values ​​and uncertainty estimates of the heating load values.

[0007] The district heating load forecasting method disclosed in this invention preferably further includes: Data acquisition: Collect heat load data of the heating system and perform preprocessing.

[0008] According to the district heating load prediction method disclosed in this invention, preferably, the calculation process of the KAN network specifically includes: Let X be the historical data of the heating load of the heating system, Φ qp (x p ) is the basic mapping function of the KAN network, θ is the set of network parameters, and x p For a single dimension of X, i.e., X = {x} p}, p=1,2,…,n, where n represents the dimension of X; The KAN network uses piecewise polynomial basis functions to perform nonlinear mapping on the input data, through the node set {z} j The input data is divided into several intervals, and a polynomial function is expanded over each interval to obtain a mapping representation, thus determining Φ. qp (x p The value of ); where the piecewise polynomial basis functions can be expressed as (x p -z j ) k k is the order of the polynomial, and the coefficients of the polynomial are determined by the parameter set θ. Ultimately, historical data X is mapped to a new feature representation Φ(X) under the action of KAN, and Φ(X) is derived from Φ qp (x p The weighted combination of the values ​​is obtained.

[0009] According to the regional heating load prediction method disclosed in this invention, preferably, the step of processing the mapping transformation data based on the Gaussian process model specifically includes: The new feature representation Φ(X) is input into the Gaussian process model. Assuming that the heating load data are all Gaussian distributed, the output distribution under the corresponding representation Φ(X) is inferred by Bayes through the joint distribution between the data. The output distribution describes both the mean of the output and the uncertainty of the output.

[0010] According to the regional heating load prediction method disclosed in this invention, preferably, the thermodynamic laws specifically include: Using water as a common heat transfer medium, the amount of heat transferred by water per unit time is y. phy Through the relation y phy =c m The equation (t1-t2) is used to characterize the water, where c represents the specific heat capacity of water, m represents the mass of water, and t1 and t2 represent the supply water temperature and return water temperature, respectively.

[0011] A second aspect of the present invention discloses a district heating load prediction system, comprising: a memory for storing program instructions; and a processor for calling the program instructions stored in the memory to implement the district heating load prediction method as described in any of the above technical solutions.

[0012] Compared with existing technologies, the beneficial effects of this invention include at least the following: Compared with existing point prediction methods (which can only provide a single load value) and lack characterization of uncertainty, this invention proposes a prediction method based on a Gaussian process model, which can output prediction results with uncertainty quantification, achieving accurate prediction and reliable assessment of heating load. Specifically, this invention, by embedding thermodynamic constraints into the model training process, not only improves the stability and generalization ability of the model under conditions of insufficient data or high noise, but also effectively enhances the physical consistency of the model's prediction results. Simultaneously, the KAN network can perform adaptive nonlinear transformations on the input variables, significantly improving the expressive power of the Gaussian process in complex dynamic systems. Through the synergistic effect of physical priors and deep feature learning, this invention can output prediction results with uncertainty quantification, achieving accurate prediction and reliable assessment of heating load, thereby providing technical support for energy-saving scheduling and safe operation of intelligent heating systems. Attached Figure Description

[0013] Figure 1 A schematic flowchart of a district heating load prediction method according to an embodiment of the present invention is shown.

[0014] Figure 2 A schematic block diagram of a district heating load prediction system according to an embodiment of the present invention is shown. Detailed Implementation

[0015] To better understand the above-described objects, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention; however, the invention may be practiced in other ways different from those described herein, and therefore, the invention is not limited to the specific embodiments disclosed below.

[0016] like Figure 1As shown, according to one embodiment of the invention, a method for predicting district heating load is disclosed, comprising: Step 101: Collect heating system data and perform preprocessing: First, detect outliers in the water mass m, treating outliers as missing values, and use interpolation to fill in the missing and outlier data. Then, combining the processed water mass m with the supply water temperature t1 and return water temperature t2, calculate the physical heat load y. phy The calculation method is y phy =c m (t1-t2), where c represents the specific heat capacity of water. Based on this, the system operating efficiency is estimated by the relationship between the actual measured load and the physical heat load, and this efficiency is used to reconstruct missing or abnormal actual load data, thereby ensuring the physical consistency and rationality of the data repair process. Finally, to reduce the impact of differences in different feature scales on model training and accelerate model convergence, all data are normalized to improve overall modeling performance. Step 102: The historical input data is mapped to a new input space through the KAN network of the mapping transformer to form the mapping transformation data, and this data is used as the input of the Gaussian process model. In this step, KAN (Kolmogorov-Arnold Network) is a neural network architecture inspired by the Kolmogorov-Arnold representation theorem. Each weight parameter in KAN is a univariate function parameterized as a spline. The specific method for mapping and transforming historical input data using KAN is to extract deep-level features from historical heating data. The specific method used in KAN transformation is as follows: Let X be the historical load data of the heating system, Φ... qp (x p Let Φ represent the one-dimensional fundamental mapping function corresponding to the p-th connection in the q-th layer of a KAN. The fundamental mapping function of a KAN is specifically represented as Φ. qp (x p )=a j +b j (x p -z j )+c j (x p - z j ) 2 +…+d j (x p - z j ) k , where a j b j c j d j For network parameters, θ={aj ,b j ,c j ,d j} represents the set of network parameters. The essence of KAN lies in using piecewise polynomial basis functions to perform a nonlinear mapping on the original data. First, several nodes are generated evenly between the minimum and maximum load values. Each pair of adjacent nodes forms an interval, and different load values ​​will fall into different intervals. Each interval corresponds to a basic mapping function Φ. qp (x p The node set and interval partitioning ensure both the mathematical continuity and smoothness of the mapping function, and also enable the extraction of local nonlinear patterns for load values ​​of different magnitudes. This allows the feature Φ(X) after KAN mapping to fully reflect the changing characteristics of historical load data, providing reliable input for subsequent prediction or optimization. The Φ obtained from mapping all intervals and multiple layers... qp (x p The data are weighted and combined to form a continuous and smooth mapping function Φ(X). Ultimately, the original historical data X is mapped to a new feature representation Φ(X) under the action of KAN, and Φ(X) is derived from the previously obtained Φ... qp (x p The weighted combination of these values ​​is used to obtain the representation, which can fully capture the nonlinear patterns of historical loads.

[0017] Step 103: Input the new feature representation Φ(X) into the Gaussian process model, and let the Gaussian process model predict the data.

[0018] In this step, the Gaussian process assumes that the data are all Gaussian distributed. By using the joint distribution among the data, Bayesian inference is used to derive the output distribution corresponding to Φ(X). This distribution describes both the mean and uncertainty of the output.

[0019] Step 104: Introduce physical laws to constrain the parameter learning of the KAN network.

[0020] In this step, the specific method used is as follows: In a district heating system, heat transfer follows basic thermodynamic laws. Using water as a common heat transfer medium, its heat transfer rate y per unit time... phy It can be obtained through the relation y phy =c m The model is characterized by (t1-t2). Then, data derived from physical laws and actual sensor data are used to guide the model parameter learning. The model's objective function not only includes a data likelihood term but also adds a constraint term based on physical consistency to measure the deviation between the predicted results and the heat conservation equation. This ensures prediction accuracy while avoiding physically unreasonable predictions.

[0021] Step 105: Since the Gaussian process model requires a fixed input when learning its parameters, and changes in the KAN parameters alter the parameter learning process, a parameter learning method was designed to simultaneously learn the parameters of both the KAN network and the Gaussian process model within a unified framework. Specifically, the model parameters are divided into two parts: KAN transformation network parameters and Gaussian process model parameters. In each iteration, KAN is first used to perform nonlinear feature mapping on the input data. Then, the Gaussian process model is trained in the transformed feature space, and physical constraints are introduced to jointly construct the optimization objective. During parameter updates, the KAN parameters and Gaussian process parameters are alternately optimized, allowing feature representation learning and regression modeling to mutually reinforce each other. This strategy effectively enhances physical consistency and training stability while ensuring model prediction accuracy.

[0022] In this step, the end-to-end training process enhances the compatibility and synergy between feature transformation and probabilistic regression.

[0023] Step 106: Assuming the currently collected heating load data is U, input U into KAN to obtain the transformation Φ(U). Process Φ(U) using the Gaussian process model to obtain the distribution of the output data corresponding to Φ(U). This distribution is a Gaussian distribution, containing two parameter values: mean and variance. The confidence interval of the output distribution can also be calculated.

[0024] like Figure 2 As shown, according to another embodiment of the present invention, a district heating load prediction system 200 is also disclosed, including: a memory 201 for storing program instructions; and a processor 202 for calling the program instructions stored in the memory to implement the district heating load prediction method as described in the above embodiment.

[0025] In summary, this invention uses historical data from the heating system as input signals and performs a nonlinear mapping transformation via KAN to obtain a more expressive high-dimensional feature representation. These mapped features are then input into a Gaussian process model to model and predict heating load and its uncertainties. During training, the law of heat conservation is further introduced as a physical inference constraint, ensuring the model maintains compliance with energy balance while learning data-driven laws, thus avoiding predictions that violate physical laws. This method can obtain high-precision and physically consistent prediction results while ensuring heating demand, providing reliable support for energy scheduling in intelligent heating systems.

[0026] All or part of the steps in the various methods of the above embodiments can be implemented by a program controlling the relevant hardware. The program can be stored in a readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other readable medium that can be used to carry or store data.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting district heating load, characterized in that, include: Mapping transformation: Features of historical heating load data are extracted through the KAN network to form mapping transformation data, which serves as input information for the Gaussian process model; Parameter learning constraints: Thermodynamic laws are introduced to constrain the parameter learning process of the KAN network, and thermodynamic law-derived data and actual sensor data are used to guide parameter learning. Gaussian process processing: The mapping transformation data is processed based on the Gaussian process model to obtain heating load prediction data and an estimate of the uncertainty of the prediction data; End-to-end training: Simultaneously learn the parameters of the KAN network and the Gaussian process model within a unified framework to obtain the trained KAN-Gaussian process model. Heating load prediction: The collected regional heating load data is input into the trained KAN-Gaussian process model to obtain prediction results, which include heating load values ​​and uncertainty estimates of the heating load values.

2. The district heating load prediction method according to claim 1, characterized in that, Also includes: Data acquisition: Collect heat load data of the heating system and perform preprocessing.

3. The district heating load prediction method according to claim 1, characterized in that, The calculation process of the KAN network specifically includes: Let X be the historical data of the heating load of the heating system, Φ qp (x p ) is the basic mapping function of the KAN network, θ is the set of network parameters, and x p For a single dimension of X, i.e., X = {x} p }, p=1,2,…,n, where n represents the dimension of X; The KAN network uses piecewise polynomial basis functions to perform nonlinear mapping on the input data, through the node set {z} j The input data is divided into several intervals, and a polynomial function is expanded over each interval to obtain a mapping representation, thus determining Φ. qp (x p The value of ); wherein the piecewise polynomial basis function can be expressed in the form of (x p -z j ) k , k is the order of the polynomial, which takes values ​​between [1,5], and the coefficients of the polynomial are determined by the parameter set θ; Ultimately, historical data X is mapped to a new feature representation Φ(X) under the action of KAN, and Φ(X) is derived from Φ qp (x p The weighted combination of the values ​​is obtained.

4. The district heating load prediction method according to claim 3, characterized in that, The steps for processing the mapping transformation data based on the Gaussian process model specifically include: The new feature representation Φ(X) is input into the Gaussian process model. Assuming that the heating load data are all Gaussian distributed, the output distribution under the corresponding representation Φ(X) is inferred by Bayes through the joint distribution between the data. The output distribution describes both the mean of the output and the uncertainty of the output.

5. The district heating load prediction method according to claim 1, characterized in that, The laws of thermodynamics specifically include: Using water as a common heat transfer medium, the amount of heat transferred by water per unit time is y. phy Through the relation y phy =c m The equation (t1-t2) is used to characterize the water, where c represents the specific heat capacity of water, m represents the mass of water, and t1 and t2 represent the supply water temperature and return water temperature, respectively.

6. A district heating load prediction system, characterized in that, include: Memory, used to store program instructions; A processor is configured to invoke the program instructions stored in the memory to implement the district heating load prediction method as described in any one of claims 1 to 5.

Images