A power load prediction method and system based on load temperature decoupling

Abstract

The present application relates to the technical field of electric power data processing, in particular to a power load prediction method and system based on load temperature decoupling. First, the load characteristic data of key industries, industry research data and physical rules are integrated to construct a multi-source fusion database containing an industry knowledge graph; then a deep learning network with a physical constraint layer is built, and a physical information load decoupling model is obtained by training by industry; the data to be analyzed is input into the model to complete load decoupling, and relevant evaluation indexes are output; after analyzing the load component characteristics, a comprehensive quality coefficient is obtained by fusing multiple indexes, high-quality basic load data is screened, and attribution analysis is performed combined with industry macroeconomic indicators to finally determine the driving factors of power load. The present application can realize the whole-process fine, interpretable and industry-customized analysis from load data decoupling to growth cause mining.

Description

Technical Field

[0001] This invention relates to the field of power data processing technology, and in particular to a power load forecasting method and system based on load-temperature decoupling. Background Technology

[0002] Existing power load analysis methods mostly focus on system-level macroscopic assessments or are based on purely data-driven statistical decomposition. For example, patent application CN202411413940.3 discloses a load resource regulation capacity analysis method based on knowledge graphs and data fusion. This method assesses the overall regulation capacity boundary of the system through data fusion and polynomial fitting, falling under the category of macroscopic statistics. While such methods can answer the question of how much load a system can withstand, they cannot deeply reveal the physical and economic mechanisms underlying load changes in a single industry. For load analysis of key industries, existing technologies often employ traditional time-series decomposition or regression models. The decoupled temperature-sensitive load components lack clear physical meaning, and the models do not incorporate industry-specific thermodynamics and production knowledge, resulting in poor interpretability of the decoupling results and difficulty in supporting in-depth attribution analysis of the driving forces of base load growth. Therefore, there is an urgent need for an analytical method that combines physical interpretability with industry specificity, enabling refined decoupling of load and the discovery of growth drivers at the mechanistic level. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies, such as poor interpretability of decoupling results and difficulty in supporting in-depth attribution analysis of the driving forces of base load growth, by providing a power load forecasting method and system based on load-temperature decoupling.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0005] In a first aspect, the present invention provides a power load forecasting method based on load-temperature decoupling, comprising: Obtain load characteristic data of key target industries, and integrate industry survey data with physical rules to construct a multi-source fusion database containing industry knowledge graphs; the industry survey data includes at least typical production processes, major energy-consuming equipment, and thermodynamic parameters; the load characteristic data includes at least historical load time series data, refined meteorological data, and industry production attribute data; Based on the multi-source fusion database, a deep learning network architecture including a physical constraint layer is constructed; the parameters of the deep learning network are initialized using an industry knowledge graph, and the deep learning network is trained for each industry to obtain a physical information load decoupling model for the target key industry, as well as the corresponding model confidence. Input the total load data of the target key industry for the period to be analyzed and the production-related characteristic data of the total load data into the physical information load decoupling model of the corresponding target key industry; the physical information load decoupling model automatically decomposes the total load data into basic load components and temperature-sensitive load components, and outputs the physical constraint compliance degree, which characterizes the physical consistency of the decoupling process, and the industry fit degree, which reflects the degree of matching of industry characteristics. The basic load component is subjected to characteristic analysis to calculate the growth trend intensity and fluctuation entropy value; the temperature-sensitive elasticity is calculated for the temperature-sensitive load component, and typical change patterns are extracted. By integrating the growth trend strength, the fluctuation entropy value, the temperature sensitivity elasticity, the physical constraint compliance degree, and the industry adaptability, a comprehensive quality coefficient characterizing the reliability of load data analysis for key target industries is obtained. The time-series data of the base load component corresponding to the comprehensive quality coefficient exceeding the preset threshold are used to conduct attribution analysis with the macroeconomic indicators of the target key industries to analyze and obtain the driving factors of power load.

[0006] Optionally, load characteristic data of key target industries can be obtained and integrated with industry survey data and physical rules to construct a multi-source fusion database containing industry knowledge graphs, including: The physical rules include at least: the nonlinear relationship between the energy consumption of temperature control equipment and the indoor-outdoor temperature difference, the simplified form of the heat conduction equation of the building envelope, and the reference power and start-stop logic of production equipment under the corresponding process flow. Based on the industry survey data, an industry knowledge graph is constructed, and the physical rules are bound to the corresponding entity nodes as attributes or constraints; wherein, the entity nodes of the industry knowledge graph include at least energy-consuming equipment, production workshops, and typical process sections; Simultaneously, production time sequence templates and equipment rated parameters are extracted from the industry production attribute data and injected into the corresponding entity nodes as attribute information to form a structured knowledge network, which serves as the multi-source fusion database.

[0007] Optionally, based on the multi-source fusion database, a deep learning network architecture including a physical constraint layer is constructed, including: A hierarchical neural network is constructed as the deep learning network. The hidden layers of the neural network contain at least one physical constraint sublayer, which is composed of differential operation units and is used to enforce the encoding of thermodynamic prior knowledge. The structure and initial parameters of the physical constraint sublayer are defined and initialized by extracting relevant entity nodes and relationship information from the industry knowledge graph.

[0008] Optionally, the parameters of the deep learning network are initialized using an industry knowledge graph, and the deep learning network is trained for each industry to obtain a physical information load decoupling model for the target key industry, as well as the corresponding model confidence, including: The deep learning network is trained under supervision using historical load time-series data from the multi-source fusion database as input and verified reference values ​​of the same-period basic load and temperature-sensitive load as training targets. The total loss function of the deep learning network consists of prediction error loss and physical consistency loss, wherein the physical consistency loss term is used to penalize the part of the deep learning network output that violates the rules encoded by the physical constraint sublayer. The deep learning network is trained on a preset training set and validation set until convergence, thereby obtaining the physical information load decoupling model. The model confidence level corresponding to the physical information load decoupling model is calculated based on the prediction error on the validation set and the physical consistency loss during the training process.

[0009] Optionally, the total load data of the target key industry for the period to be analyzed and the production-related characteristic data of the total load data are input into the physical information load decoupling model of the corresponding target key industry, including: The production-related characteristic data includes at least refined meteorological data, time stamp data, and generalized event tags for the period to be analyzed; The refined meteorological data and the total load data are synchronized and aligned on the time scale. The time-identification data is used to mark weekdays, weekends, holidays, and industry-specific production periods; The generalized event label is used to identify known major events that affect industry production.

[0010] Optionally, the physical information load decoupling model automatically decomposes the total load data into a basic load component and a temperature-sensitive load component, and simultaneously outputs a physical constraint compliance degree characterizing the physical consistency of the decoupling process, and an industry fit degree reflecting the degree of matching with industry characteristics, including: The physical information load decoupling model calculates the theoretical temperature control energy consumption baseline based on the input total load data of the target key industries and production-related characteristic data. The total load data is compared and separated from the temperature control energy consumption baseline, and combined with the capture of non-stationary modes by the time-series learning layer, the basic load component and the temperature-sensitive load component are jointly generated. The real-time theoretical output generated by the physical constraint layer for the current input is calculated, and the difference between the real-time theoretical output and the statistical mean calculated by the physical constraint layer on all outputs on the complete training dataset is analyzed. The difference is quantified as the degree of compliance with the physical constraint. Calculate the KL divergence between the total load data and production-related characteristic data of the target key industry currently input and the corresponding characteristic distribution of the target key industry used in the training of the physical information load decoupling model, and calculate the industry fit based on the KL divergence.

[0011] Optionally, characteristic analysis is performed on the basic load component to calculate the growth trend intensity and fluctuation entropy value; for the temperature-sensitive load component, temperature-sensitive elasticity is calculated, and typical change patterns are extracted, including: The strength of the growth trend is obtained by calculating the average slope of the corresponding long-term trend term during the analysis period. The fluctuation entropy value is obtained by calculating the sample entropy from the baseline load residual sequence after removing the long-term trend; A nonlinear regression method is used to fit the functional relationship between the temperature-sensitive load component and the corresponding time period temperature data; the average first derivative of the functional relationship in the typical operating temperature range is calculated as the temperature-sensitive elasticity. The characteristic curves of the temperature-sensitive load components are obtained, and the typical variation patterns are obtained based on the characteristic curves through cluster analysis.

[0012] Optionally, by integrating the growth trend strength, the fluctuation entropy value, the temperature-sensitive elasticity, the physical constraint compliance, and the industry adaptability, a comprehensive quality coefficient characterizing the reliability of load data analysis for the target key industry is obtained, including: The growth trend strength, fluctuation entropy value, temperature sensitivity elasticity, physical constraint compliance and industry adaptability are used as input factors, and preset weight coefficients are configured for each parameter in the input factors. A comprehensive score is calculated using a weighted summation formula and used as the comprehensive quality coefficient; wherein the weight coefficients of the physical constraint compliance degree and the industry fit degree are higher than the weight coefficients of the other three input factors.

[0013] Optionally, the time-series data of the base load component corresponding to when the comprehensive quality coefficient exceeds a preset threshold can be used for attribution analysis with macroeconomic indicators of the target key industries to analyze and obtain the driving factors of electricity load, including: An econometric model is established using time-series data of the basic load component that has passed the comprehensive quality coefficient test as the dependent variable and macroeconomic indicators of the selected target key industries as independent variables. By calculating the significance of the standardized regression coefficients of the independent variables in the econometric model and analyzing the explanatory contribution of the independent variables to the fluctuation of the dependent variable based on variance decomposition, the influence intensity and statistical reliability of different economic factors on the changes in the base load in the macroeconomic indicators are quantitatively assessed. Based on the intensity of the impact and statistical reliability, the driving factors and their corresponding weights that drive the growth of electricity load in the target key industries are identified and determined.

[0014] Secondly, the present invention provides a power load forecasting system based on load-temperature decoupling, comprising: The basic database construction module is used to acquire load characteristic data of key target industries and integrate industry survey data with physical rules to construct a multi-source fusion database containing an industry knowledge graph. The industry survey data includes at least typical production processes, major energy-consuming equipment, and thermodynamic parameters. The load characteristic data includes at least historical load time series data, refined meteorological data, and industry production attribute data. The decoupling model acquisition module is used to construct a deep learning network architecture including a physical constraint layer based on the multi-source fusion database; initialize the deep learning network parameters using industry knowledge graphs; train the deep learning network by industry to obtain a physical information load decoupling model for the target key industry, and the corresponding model confidence. The load data analysis module is used to input the total load data of the target key industry for the period to be analyzed and the production-related characteristic data of the total load data into the physical information load decoupling model of the corresponding target key industry. The physical information load decoupling model automatically decomposes the total load data into basic load components and temperature-sensitive load components, and outputs the physical constraint compliance degree, which characterizes the physical consistency of the decoupling process, and the industry fit degree, which reflects the degree of matching of industry characteristics. The load component analysis module is used to perform characteristic analysis on the basic load component, calculate the growth trend intensity and fluctuation entropy value; for the temperature-sensitive load component, calculate the temperature-sensitive elasticity and extract typical change patterns; The data quality analysis module is used to integrate the growth trend strength, the fluctuation entropy value, the temperature sensitivity elasticity, the physical constraint compliance degree, and the industry adaptability to obtain a comprehensive quality coefficient that characterizes the reliability of the target key industry load data analysis. The driving factor analysis module is used to perform attribution analysis on the time series data of the base load component corresponding to the comprehensive quality coefficient being higher than the preset threshold, and on the macroeconomic indicators of the target key industries to obtain the driving factors of power load.

[0015] By implementing this invention, it is possible to acquire load characteristic data of key target industries and integrate industry survey data with physical rules to construct a multi-source fusion database containing an industry knowledge graph. The industry survey data includes at least typical production processes, major energy-consuming equipment, and thermodynamic parameters. The load characteristic data includes at least historical load time-series data, refined meteorological data, and industry production attribute data. This breaks the limitations of single data types by deeply integrating industry production knowledge, thermodynamic physical rules, and load data. The structured design of the knowledge graph allows the model to recognize industry-specific entity relationships and constraints, avoiding the blindness of purely data-driven approaches. At the same time, it achieves the standardization and correlation of multi-source data, providing accurate and industry-specific data sources for subsequent industry-specific model training.

[0016] By implementing this invention, it is possible to construct a deep learning network architecture including a physical constraint layer based on the multi-source fusion database; initialize the deep learning network parameters using an industry knowledge graph, and train the deep learning network for each industry to obtain a physical information load decoupling model for the target key industry, as well as the corresponding model confidence; the design of the physical constraint layer encodes thermodynamic prior knowledge into the network, giving the deep learning model physical interpretability, which is different from traditional pure data-driven models; industry-specific training and knowledge graph parameter initialization make the model highly adaptable to the characteristics of the target industry, avoiding the generality deviation of cross-industry models; the output of the model confidence provides a reliability reference for subsequent model applications, while the physical consistency loss term ensures that the model output does not violate physical rules.

[0017] By implementing this invention, it is possible to input the total load data of the target key industry for the period to be analyzed and the production-related characteristic data of the total load data into the physical information load decoupling model of the corresponding target key industry. The physical information load decoupling model automatically decomposes the total load data into basic load components and temperature-sensitive load components, and outputs physical constraint compliance degree, which characterizes the physical consistency of the decoupling process, and industry fit degree, which reflects the degree of matching of industry characteristics. It achieves refined and mechanistic decoupling of load components. Compared with traditional time series decomposition methods, the decoupling results are highly matched with industry physical rules and production characteristics, and have clear physical meaning. The dual index output of physical constraint compliance degree and industry fit degree can evaluate the reliability of a single decoupling process in real time and promptly detect decoupling deviations caused by data anomalies or mismatch of industry characteristics.

[0018] By implementing this invention, it is possible to perform characteristic analysis on the base load component, calculate the growth trend intensity and fluctuation entropy value; calculate the temperature sensitivity elasticity of the temperature-sensitive load component, and extract typical change patterns; transform abstract load time series data into quantitative key characteristic indicators, realizing a quantitative description of the load component change law; the growth trend intensity and fluctuation entropy value can accurately characterize the long-term development and short-term fluctuation characteristics of the base load, and the temperature sensitivity elasticity and typical change patterns can clearly reflect the correlation law and change characteristics between temperature-sensitive load and air temperature, providing specific and analyzable characteristic basis for subsequent load analysis and forecasting.

[0019] By implementing this invention, a comprehensive quality coefficient characterizing the reliability of load data analysis for key target industries can be obtained by integrating the growth trend strength, fluctuation entropy value, temperature-sensitive elasticity, physical constraint compliance, and industry adaptability. Through weighted fusion of multi-dimensional indicators, a comprehensive reliability assessment of the entire load data analysis process is achieved, avoiding the bias of single-indicator evaluation. Differential weighting coefficients highlight the core importance of physical consistency and industry adaptability, ensuring that the assessment results align with the physical interpretability and industry-specific requirements of this method. The comprehensive quality coefficient sets a reliability threshold for subsequent attribution analysis, filtering out high-quality analytical data.

[0020] By implementing this invention, it is possible to perform attribution analysis on the time-series data of the base load component corresponding to a comprehensive quality coefficient higher than a preset threshold, and on macroeconomic indicators of key target industries. This analysis identifies the driving factors of power load. Attribution analysis based on high-quality base load data effectively avoids analytical biases caused by low-quality data, ensuring the accuracy and statistical reliability of driving factor identification. Furthermore, by using econometric models to quantitatively evaluate driving factors, not only can core driving factors be identified, but the influence intensity and weight of each factor can also be determined. This overcomes the limitations of traditional macroeconomic analysis in achieving in-depth attribution, providing precise economic mechanism support for power load planning and regulation.

[0021] In summary, by implementing this invention, a refined, interpretable, and industry-customized analysis can be achieved throughout the entire process, from decoupling load data to mining growth drivers. This ensures both the physical rationality and industry adaptability of the load analysis results, and deeply reveals the physical and economic mechanisms of power load changes in key industries. It provides reliable and guiding theoretical and data support for power load forecasting, resource regulation, and planning decisions. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a power load forecasting method based on load-temperature decoupling provided by the present invention; Figure 2This is a schematic diagram of a power load prediction system based on load-temperature decoupling provided by the present invention.

[0023] In the attached diagram, the components represented by each number are as follows: The modules are: Basic Database Construction Module 11, Decoupling Model Acquisition Module 12, Load Data Analysis Module 13, Load Component Analysis Module 14, Data Quality Analysis Module 15, and Driving Factor Analysis Module 16. Detailed Implementation

[0024] Example 1, as Figure 1 As shown, this embodiment of the invention provides a power load forecasting method based on load-temperature decoupling, including: S100: Obtain load characteristic data of key target industries, integrate industry survey data with physical rules, and construct a multi-source fusion database containing industry knowledge graphs; the industry survey data includes at least typical production processes, major energy-consuming equipment, and thermodynamic parameters; the load characteristic data includes at least historical load time series data, refined meteorological data, and industry production attribute data; S200: Based on the multi-source fusion database, construct a deep learning network architecture that includes a physical constraint layer; initialize the deep learning network parameters using an industry knowledge graph, and train the deep learning network by industry to obtain a physical information load decoupling model for the target key industry, as well as the corresponding model confidence. S300: Input the total load data of the target key industry for the period to be analyzed and the production-related characteristic data of the total load data into the physical information load decoupling model of the corresponding target key industry; the physical information load decoupling model automatically decomposes the total load data into basic load components and temperature-sensitive load components, and outputs the physical constraint compliance degree, which characterizes the physical consistency of the decoupling process, and the industry fit degree, which reflects the degree of matching of industry characteristics. S400: Perform characteristic analysis on the basic load component, calculate the growth trend intensity and fluctuation entropy value; for the temperature-sensitive load component, calculate the temperature-sensitive elasticity and extract typical change patterns; S500: By integrating the growth trend strength, the fluctuation entropy value, the temperature-sensitive elasticity, the physical constraint compliance degree, and the industry adaptability, a comprehensive quality coefficient characterizing the reliability of load data analysis for key target industries is obtained. S600: The time-series data of the base load component corresponding to the comprehensive quality coefficient when it is higher than the preset threshold are used to perform attribution analysis with the macroeconomic indicators of the target key industries to analyze and obtain the driving factors of power load.

[0025] In step S100 of this application embodiment, load characteristic data of the target key industry is obtained, and industry survey data and physical rules are integrated to construct a multi-source fusion database containing an industry knowledge graph, including: The physical rules include at least: the nonlinear relationship between the energy consumption of temperature control equipment and the indoor-outdoor temperature difference, the simplified form of the heat conduction equation of the building envelope, and the reference power and start-stop logic of production equipment under the corresponding process flow. Based on the industry survey data, an industry knowledge graph is constructed, and the physical rules are bound to the corresponding entity nodes as attributes or constraints; wherein, the entity nodes of the industry knowledge graph include at least energy-consuming equipment, production workshops, and typical process sections; Simultaneously, production time sequence templates and equipment rated parameters are extracted from the industry production attribute data and injected into the corresponding entity nodes as attribute information to form a structured knowledge network, which serves as the multi-source fusion database.

[0026] In this embodiment, the purpose of step S100 is to build a multi-source fusion database that combines industry physical characteristics with structured knowledge, providing a data foundation and prior knowledge support for the subsequent construction of a physical information load decoupling model. This integrates characteristic data such as load, production, and meteorology of key target industries, incorporates industry thermodynamic physical rules and production survey knowledge, and uses knowledge graphs to achieve structured association between data and rules. This allows subsequent model training to rely on industry-specific physical mechanisms and production patterns, solving the problems of traditional load analysis methods lacking physical interpretability and having weak industry specificity. Simultaneously, it provides industry knowledge basis for model parameter initialization.

[0027] To achieve the above objectives, it is first necessary to extract describable physical rules based on the basic principles of thermodynamics and the operation manuals of typical energy-consuming equipment in key industries. These physical rules include at least the following: the nonlinear relationship between the energy consumption of temperature control equipment and the indoor-outdoor temperature difference, the simplified form of the heat conduction equation of the building envelope, and the reference power and start-stop logic of production equipment under the corresponding process flow. Based on the fundamental principles of thermodynamics, and combined with the operation manuals of typical energy-consuming equipment in key target industries, describable rules that reflect the physical laws of energy consumption and production in these industries are extracted.

[0028] The extracted physical rules need to cover three core categories: the nonlinear relationship between the energy consumption of temperature control equipment and the indoor-outdoor temperature difference; the simplified form of the heat conduction equation of the building envelope; and the baseline power and start-stop logic of production equipment under the corresponding process flow. For example, the energy consumption of temperature control equipment such as central air conditioning in the chemical industry has a nonlinear relationship with the indoor-outdoor temperature difference, which is a quadratic function. For building envelopes such as industrial plants, the simplified heat conduction equation is used to characterize the heat transfer law, such as Q=λAΔT / δ. For production equipment such as blast furnaces in the steel industry, the baseline power under the ironmaking process is 2000kW, and the start-stop logic is to run at full load after 2 hours of preheating and to cool down for 1 hour before shutdown.

[0029] Then, the physical rules are formalized into differentiable functions or state transition constraints; The various physical rules to be extracted will be transformed into forms that can be recognized and computed by deep learning networks, specifically differentiable functions or state transition constraints. For example, the nonlinear relationship between the energy consumption of temperature control equipment and the indoor-outdoor temperature difference can be transformed into y=ax 2 Differentiable quadratic functions of the form +bx+c; the start-up and shutdown logic of production equipment is transformed into "the equipment needs to run continuously for t1 hours after startup to reach full load, and the shutdown requires the temperature to be below T1 as a state transition constraint".

[0030] Next, based on the industry survey data, the industry knowledge graph is constructed, and the physical rules are bound to the corresponding entity nodes as attributes or constraints; wherein, the entity nodes of the industry knowledge graph include at least energy-consuming equipment, production workshops, and typical process sections; Based on survey data from key target industries, an industry knowledge graph is built with energy-consuming equipment, production workshops, and typical process sections as core entity nodes, and the relationships between these entity nodes are analyzed.

[0031] The refined and formalized physical rules are then used as attribute information or constraints and bound to the corresponding entity nodes in the knowledge graph. For example, physical rules such as the base power and start-up / shutdown logic of a blast furnace are bound to the energy-consuming equipment entity node "blast furnace" in the knowledge graph; and the heat conduction rules of the workshop enclosure structure are bound to the production workshop entity node "steelmaking workshop".

[0032] Simultaneously, production time sequence templates and equipment rated parameters are extracted from the industry production attribute data and injected into the corresponding entity nodes as attribute information to form a structured knowledge network, which serves as the multi-source fusion database.

[0033] This involves extracting two types of key information from the production attribute data of the target key industries: first, production sequence templates, such as the two-shift production sequence in the machinery manufacturing industry and the continuous production sequence in the chemical industry; and second, equipment rated parameters, such as the rated power of an air compressor of 150kW and the rated flow rate of a water pump of 50m³ / h. 3 / h.

[0034] The extracted production timeline templates and equipment rated parameters are used as attribute information and injected into the corresponding entity nodes in the industry knowledge graph, so that each entity node has complete physical rules, production parameters, and timeline characteristic attributes.

[0035] Through the above operations, a multi-source fusion database is formed, which integrates load characteristic data, industry survey data, and physical rules, and uses knowledge graphs as the structured carrier, thus completing the integration and structured association of data and knowledge.

[0036] In step S200 of this application embodiment, based on the multi-source fusion database, a deep learning network architecture including a physical constraint layer is constructed, including: A hierarchical neural network is constructed as the deep learning network. The hidden layers of the neural network contain at least one physical constraint sublayer, which is composed of differential operation units and is used to enforce the encoding of thermodynamic prior knowledge. The structure and initial parameters of the physical constraint sublayer are defined and initialized by extracting relevant entity nodes and relationship information from the industry knowledge graph.

[0037] The purpose of step S200 in this embodiment is to construct a deep learning network architecture that incorporates prior thermodynamic knowledge of the industry, based on the established multi-source fusion database. The core is to add a physical constraint sublayer composed of differential operation units to the neural network, embedding industry-specific physical rules into the network structure and parameters in a computable form. This solves the black-box problem of traditional deep learning models, allowing the training and prediction processes of deep learning models to be constrained by physical laws, improving the physical interpretability and industry adaptability of the deep learning model's output, and avoiding unreasonable predictions that violate thermodynamic principles and deviate from industry realities.

[0038] To achieve the above objectives, a hierarchical neural network needs to be built first, which serves as the deep learning network. The hidden layers of the neural network contain at least one physical constraint sublayer, which is composed of differential operation units and is used to enforce the encoding of thermodynamic prior knowledge. First, a layered neural network is built as the core deep learning network. The key difference from ordinary neural networks is that at least one physical constraint sublayer is embedded in the hidden layers of the deep learning network. This physical constraint sublayer consists of differential operation units, and its core function is to enforce the encoding of prior thermodynamic knowledge. For example, it addresses the nonlinear relationship between the energy consumption of temperature control equipment and the indoor-outdoor temperature difference, y=ax. 2 +bx+c, the differential operation unit in the physical constraint sublayer can perform the derivative operation y'=2ax+b on this function, solidifying the differential relationship corresponding to this thermodynamic law into the network operation logic, ensuring that the deep learning model operation process conforms to this physical rule.

[0039] The structure and initial parameters of the physical constraint sublayer are defined and initialized by extracting relevant entity nodes and relationship information from the industry knowledge graph.

[0040] That is, from the industry knowledge graph of the multi-source fusion database, the entity nodes and relationship information related to thermodynamic rules are accurately extracted, and the structure of the physical constraint sublayer is defined and the initial parameters of the sublayer are initialized based on this.

[0041] For example, if entity nodes and relationship information for "steelmaking workshop - enclosure structure - heat conduction equation Q=λAΔT / δ" are extracted from the industry knowledge graph, the structure of the physical constraint sublayer will be configured with differential operation units capable of calculating this heat conduction equation, and the initial parameters will initialize the heat conduction coefficient λ to 0.8W / (m²). K), the heat transfer area A is initialized to 1000m². 2 Typical industry parameters extracted from the attributes of entity nodes in the knowledge graph.

[0042] For example, if the entity node and relationship information of "blast furnace - production equipment - base power 2000kW, start-stop logic (full load operation after 2 hours of preheating)" are extracted, the physical constraint sublayer will set up a state transition differential constraint unit that matches the start-stop logic. The initial parameter will use the base power of 2000kW as the basic threshold parameter for the unit operation to ensure that the initial operation of the physical constraint sublayer matches the physical operation law of the blast furnace.

[0043] In step S200 of this application embodiment, the parameters of the deep learning network are initialized using an industry knowledge graph, and the deep learning network is trained for each industry to obtain a physical information load decoupling model for the target key industry, as well as the corresponding model confidence, including: The deep learning network is trained under supervision using historical load time-series data from the multi-source fusion database as input and verified reference values ​​of the same-period basic load and temperature-sensitive load as training targets. The total loss function of the deep learning network consists of prediction error loss and physical consistency loss, wherein the physical consistency loss term is used to penalize the part of the deep learning network output that violates the rules encoded by the physical constraint sublayer. The deep learning network is trained on a preset training set and validation set until convergence, thereby obtaining the physical information load decoupling model. The model confidence level corresponding to the physical information load decoupling model is calculated based on the prediction error on the validation set and the physical consistency loss during the training process.

[0044] The purpose of step S200 in this embodiment is to initialize the parameters of the deep learning network based on industry knowledge graphs, and then adapt the network to the load characteristics and physical rules of the target key industries through industry-specific supervised training. This ultimately yields a physical information load decoupling model that is industry-specific and physically interpretable. Simultaneously, the prediction performance and compliance with physical rules of the physical information load decoupling model on the validation set are quantified, and the model confidence is calculated. This provides a reliability judgment basis for the subsequent practical application of the physical information load decoupling model, solving the problems of traditional model training lacking industry prior knowledge guidance, training results easily violating physical laws, and lacking quantitative indicators of model reliability.

[0045] To achieve the above objectives, the historical load time series data in the multi-source fusion database is first used as input, and the verified reference values ​​of the concurrent base load and temperature-sensitive load are used as training targets to supervise the training of the deep learning network. This involves first extracting prior industry knowledge, such as entity node attributes and physical rule parameters, from the industry knowledge graph of the multi-source fusion database to complete the initial parameter assignment of the deep learning network. This allows the network training to start from an initial state that aligns with the actual industry situation, rather than a meaningless random initialization.

[0046] For example, when training a deep learning network for the chemical industry, parameters such as the energy consumption temperature difference coefficient of chemical refrigeration equipment and the reference power of the reactor are extracted from the knowledge graph and assigned to the connection weights and bias terms of the deep learning network to complete the parameter initialization.

[0047] Subsequently, historical load time series data of key target industries were selected from the multi-source fusion database as the training input for the deep learning network. At the same time, verified reference values ​​of the baseline load and temperature-sensitive load of the same period were selected as training targets to carry out supervised training, so that the deep learning network could learn the decoupling rules from the total load to the baseline load and temperature-sensitive load.

[0048] For example, the historical total hourly load data of the chemical industry throughout 2023 is selected as the input, and the hourly basic load and temperature-sensitive load data of the chemical industry during the same period, after manual verification and data calibration, are selected as the training target.

[0049] The total loss function of the deep learning network is composed of prediction error loss and physical consistency loss, wherein the physical consistency loss term is used to penalize the part of the deep learning network output that violates the rules encoded by the physical constraint sublayer. This involves designing a total loss function for the deep learning network, which consists of two parts: prediction error loss and physical consistency loss. This allows the network training to simultaneously consider both the prediction accuracy achieved by load decoupling and the compliance with physical rules.

[0050] Prediction error loss measures the deviation between the network's output baseline load, temperature-sensitive load, and the training target reference value. Common loss methods include mean squared error (MSE) and mean absolute error (MAE). For example, MSE can be used to calculate the deviation between the predicted value and the reference value. , where y i Y is the load reference value. i These are the predicted values ​​from the deep learning network.

[0051] The physical consistency loss term is a core customized term specifically used to penalize parts of the network output that violate the thermodynamic rules and industry production physics rules encoded by the physical constraint sublayer. For example, if the physical constraint sublayer encodes the rule that "the energy consumption of temperature control equipment increases monotonically with the increase of indoor and outdoor temperature difference," then when the energy consumption of temperature control equipment output by the deep learning network decreases with the increase of temperature difference, the physical consistency loss term will immediately generate a large penalty value, causing the total loss function value to increase.

[0052] Then, the deep learning network is trained based on the preset training set and validation set until convergence, thereby obtaining the physical information load decoupling model; This involves pre-setting a reasonable ratio between the training and validation sets for the deep learning network. For example, historical load time series data of the chemical industry can be divided into training and validation sets in an 8:2 ratio. The training set is used for iterative updates of the deep learning network parameters, while the validation set is used to monitor the generalization ability of the deep learning network.

[0053] Following the principle of industry-specific training, the above-mentioned supervised training was carried out separately for different key industries. During the training process, based on the numerical change of the total loss function, the network parameters were continuously optimized through backpropagation until the numerical value of the total loss function tended to stabilize and no longer decreased significantly, that is, the network training reached the convergence state. At this point, the training was stopped, and the physical information load decoupling model for the target key industry was obtained.

[0054] For example, the deep learning network can be built using a temporal convolutional neural network to adapt to the characteristics of the load time-series data.

[0055] Optionally, in terms of network structure settings, the input layer dimension is set to 96, corresponding to 96 15-minute load time-series features per day; three temporal convolutional layers are set: the first layer has 64 convolutional kernels, a kernel size of 3, a stride of 1, and the same padding method; the second layer has 128 convolutional kernels, a kernel size of 3, and a stride of 1; the third layer has 256 convolutional kernels, a kernel size of 3, and a stride of 1; each convolutional layer is followed by a pooling layer with a pooling kernel size of 2 and a stride of 2; the hidden layer embeds a physical constraint sublayer, which consists of differential operation units, and the number of units matches the output dimension of the convolutional layer; the output layer has a dimension of 2, corresponding to the decoupling results of the basic load and the temperature-sensitive load.

[0056] Optionally, in the parameter settings of the deep learning network, the optimizer is selected as the adaptive moment estimation optimizer, the learning rate is set to 0.001, the weight decay coefficient is 0.0001, the batch size is 32, and the prediction error loss weight in the total loss function is 0.7 and the physical consistency loss weight is 0.3.

[0057] Optionally, in training the deep learning network, the training samples are sourced from the historical load time series data of the target industry over the past 3 years in a multi-source fusion database; the sample size is 100,000 records; the training rounds are preset to 100 rounds; the convergence criterion is that the fluctuation of the total loss function value of the validation set is less than 0.001 for 10 consecutive rounds, and the physical consistency loss value is less than 0.01. If these conditions are met, training is stopped, and the trained deep learning network is used as the physical information load decoupling model.

[0058] Finally, based on the prediction error on the validation set and the physical consistency loss during training, the model confidence level corresponding to the physical information load decoupling model is calculated.

[0059] In the training process of the physical information load decoupling model, two core indicators of the physical information load decoupling model on the validation set are extracted: the prediction error loss value and the physical consistency loss value. Based on these two indicators, they are comprehensively quantified through a preset calculation method to obtain the model confidence level of the physical information load decoupling model. The higher the confidence level, the higher the prediction accuracy of the physical information load decoupling model on the validation set, the better the compliance with physical rules, and the stronger the overall reliability of the physical information load decoupling model.

[0060] For example, the model confidence calculation formula is set as C=1-α×L1-β×L2, where L1 is the normalized prediction error loss on the validation set, L2 is the normalized physical consistency loss on the validation set, and α and β are weight coefficients, assigned values ​​of 0.6 and 0.4 respectively. The value calculated by this formula is the model confidence C, with a value range of 0 to 1.

[0061] In step S300 of the application embodiment, the total load data of the target key industry for the period to be analyzed and the production-related characteristic data of the total load data are input into the physical information load decoupling model of the corresponding target key industry, including: The production-related characteristic data includes at least refined meteorological data, time stamp data, and generalized event tags for the period to be analyzed; The refined meteorological data and the total load data are synchronized and aligned on the time scale. The time-identification data is used to mark weekdays, weekends, holidays, and industry-specific production periods; The generalized event label is used to identify known major events that affect industry production.

[0062] The purpose of step S300 in this embodiment is to input the total load data of the target key industry for the period to be analyzed, along with production-related characteristic data that reflects industry production, meteorological environment, and time characteristics, into the physical information load decoupling model of the corresponding industry. This provides complete and matching input data for the physical information load decoupling model to perform load decomposition. By supplementing key characteristic data such as refined meteorological, time markers, and generalized event labels, the physical information load decoupling model can combine actual production environment and scenario factors to decouple the load, improving the fit between the decoupling results and actual working conditions. At the same time, it ensures the consistency of input data in the time dimension, avoiding deviations in decoupling results due to data asynchrony, and ensuring that the physical information load decoupling model can accurately identify and decompose the basic load component and the temperature-sensitive load component.

[0063] To achieve the above objectives, it is first necessary to obtain production-related characteristics. The production-related characteristic data includes at least refined meteorological data, time stamp data, and generalized event tags for the period to be analyzed. Specifically, the method for obtaining refined meteorological data is to collect refined meteorological data of the target key industry's location within the period to be analyzed. The data must include meteorological indicators related to load changes, such as temperature, humidity, and wind speed. For example, real-time temperature and outdoor relative humidity data of the steel industry plant area are collected every 15 minutes throughout August 2024, with the temperature data accurate to 0.1℃.

[0064] Specifically, the refined meteorological data and the total load data are synchronized on the time scale. The collected refined meteorological data is matched with the time granularity of the total load data to be analyzed, ensuring that the two types of data are completely synchronized on the time scale. For example, if the total load data of the steel industry to be analyzed is one data point every 15 minutes, then the refined meteorological data also needs to be organized into one data point every 15 minutes, so that the total load data at 0:00 on August 1, 2024 corresponds to the meteorological data such as temperature and humidity at the same time, avoiding model input data matching errors due to inconsistent time granularity.

[0065] The time-identification data is used to mark weekdays, weekends, holidays, and industry-specific production periods. The method for obtaining the time-identification data is to mark each time node of the period to be analyzed with time-identification data. The marking content must cover weekdays, weekends, and holidays, and also mark industry-specific production periods. For example, for the time nodes in August 2024, August 10 is marked as a weekend, August 15 is marked as a weekday, and August 1 is marked as a holiday. For the three-shift production mode in the steel industry, 0:00-8:00 is marked as the night shift production period, 8:00-16:00 is marked as the day shift production period, and 16:00-24:00 is marked as the afternoon shift production period.

[0066] The generalized event tags are used to identify known major events affecting industry production. The method for obtaining generalized event tags is to identify known major events affecting the production of the target key industry within the period to be analyzed, and add generalized event tags to the corresponding time nodes. For example, if a steel plant undergoes annual equipment maintenance from August 20th to August 22nd, 2024, resulting in a significant drop in production load, then a generalized event tag for equipment maintenance is added to the time node for that period; if extreme weather with sustained high temperatures occurs in the area in mid-August, then a generalized event tag for extreme high-temperature weather is added to the corresponding period.

[0067] Finally, the total load data and refined meteorological data that have been synchronized and aligned with the time are integrated with the labeled time stamp data and the added generalized event labels to form a complete, dimension-matched input dataset to be analyzed.

[0068] The input dataset to be analyzed is then fed into a pre-trained physical information load decoupling model designed for the target key industry.

[0069] In step S300 of this embodiment, the physical information load decoupling model automatically decomposes the total load data into a basic load component and a temperature-sensitive load component. Simultaneously, it outputs a physical constraint compliance degree characterizing the physical consistency of the decoupling process, and an industry fit degree reflecting the degree of matching with industry characteristics, including: The physical information load decoupling model calculates the theoretical temperature control energy consumption baseline based on the input total load data of the target key industries and production-related characteristic data. The total load data is compared and separated from the temperature control energy consumption baseline, and combined with the capture of non-stationary modes by the time-series learning layer, the basic load component and the temperature-sensitive load component are jointly generated. The real-time theoretical output generated by the physical constraint layer for the current input is calculated, and the difference between the real-time theoretical output and the statistical mean calculated by the physical constraint layer on all outputs on the complete training dataset is analyzed. The difference is quantified as the degree of compliance with the physical constraint. Calculate the KL divergence between the total load data and production-related characteristic data of the target key industry currently input and the corresponding characteristic distribution of the target key industry used in the training of the physical information load decoupling model, and calculate the industry fit based on the KL divergence.

[0070] The purpose of step S300 in this embodiment is to accurately decompose the total load data into basic load components and temperature-sensitive load components using a physical information load decoupling model, while quantifying and outputting two core evaluation indicators: physical constraint compliance and industry fit. The aim is to ensure that the load decomposition results align with industry physical laws and actual production characteristics, addressing the problems of poor interpretability and lack of objective evaluation dimensions in traditional decomposition methods. Physical constraint compliance measures whether the decomposition process conforms to preset thermodynamic and physical rules, while industry fit determines the degree of matching between the data to be analyzed and the industry characteristics used for model training. Both provide quantitative basis for judging the reliability of subsequent analysis results, ensuring that the decomposed load components truly reflect the electricity consumption patterns of the target key industries.

[0071] To achieve the above objectives, it is first necessary to calculate the theoretical temperature control energy consumption baseline based on the physical information load decoupling model, according to the input target key industry total load data and production-related characteristic data. For example, the physical information load decoupling model combines hourly temperature data for the period to be analyzed in the chemical industry, workshop temperature control standards, and the nonlinear function y=0.05x of refrigeration equipment energy consumption and temperature difference. 2 +0.2x+1.5, calculate the theoretical baseline temperature control energy consumption per hour for the period to be analyzed, with the baseline value ranging from 50kW to 200kW.

[0072] Next, the total load data is compared and separated with the temperature control energy consumption baseline, and combined with the capture of non-stationary modes by the time-series learning layer, the basic load component and the temperature-sensitive load component are jointly generated. The total load data for the period to be analyzed is compared and numerically separated with the calculated theoretical temperature control energy consumption baseline, initially dividing the energy consumption into temperature-related and non-temperature-related components. Simultaneously, relying on the time-series learning layer in the physical information load decoupling model, non-stationary patterns in the load data are captured, such as load abrupt changes caused by equipment start-up / shutdown or production shift switching. The initial separation results are then corrected and optimized, ultimately accurately generating the basic load component and the temperature-sensitive load component.

[0073] The basic load component consists of the core electrical loads in industry production that are not affected by temperature changes, while the temperature-sensitive load component consists of the loads generated by changes in the energy consumption of temperature control equipment due to temperature variations.

[0074] For example, the total load of the steel industry during a certain period is 2500kW, and the corresponding theoretical temperature control energy consumption baseline is 300kW. The time-series learning layer captures the stable mode during this period when there is no production shift switching. After separation and correction, it is determined that the temperature-sensitive load component during this period is 300kW and the basic load component is 2200kW.

[0075] Then, the real-time theoretical output generated by the physical constraint layer for the current input is calculated, and the difference between the real-time theoretical output and the statistical mean calculated by the physical constraint layer on all outputs on the complete training dataset is analyzed. The difference is quantified as the physical constraint compliance degree. First, the physical constraint layer in the information load decoupling model generates real-time theoretical output for the currently input data to be analyzed. This real-time theoretical output is the theoretically calculated value of the physical characteristics that the current load should exhibit, based on the encoded thermodynamic rules. Then, the statistical mean of all outputs from the physical constraint layer on the complete training dataset is calculated. This statistical mean represents the standard physical law output level during the training of the information load decoupling model. Next, the numerical difference between the real-time theoretical output and this statistical mean is analyzed. The difference is calculated using a preset quantification formula, and the result is the physical constraint compliance degree. The smaller the numerical difference, the higher the physical constraint compliance degree, indicating that the load decomposition process conforms more closely to the preset thermodynamic physical rules.

[0076] For example, the statistical mean of the output of the physical constraint layer on the complete training dataset is 100, and the real-time theoretical output for the current data to be analyzed is 98. Calculated using the quantization formula, the physical constraint compliance degree is 0.98. The closer the physical constraint compliance degree is to 1, the better the physical consistency.

[0077] Finally, calculate the KL divergence between the total load data and production-related characteristic data of the target key industry currently input and the corresponding characteristic distribution of the target key industry used in the training of the physical information load decoupling model, and calculate the industry fit based on the KL divergence.

[0078] That is, take the feature distribution of the total load data and production-related feature data of the target key industry currently input, and then retrieve the feature distribution of the corresponding feature data of the target key industry used when training the physical information load decoupling model.

[0079] Then, the KL divergence between these two feature distributions is calculated. KL divergence measures the degree of difference between two probability distributions; a smaller KL divergence value indicates a higher similarity between the two feature distributions. Finally, based on the KL divergence value, industry fit is calculated using a preset conversion formula. A smaller KL divergence value results in a higher industry fit, indicating a better match between the features of the data to be analyzed and the industry features trained on the model, and a stronger industry relevance to the decomposition results.

[0080] For example, the calculated KL divergence between the current data to be analyzed and the model training data is 0.15. Through the conversion formula, the industry fit is calculated to be 0.87. The closer the value is to 1, the higher the degree of matching of industry characteristics.

[0081] In step S400 of this application embodiment, feature analysis is performed on the basic load component to calculate the growth trend intensity and fluctuation entropy value; for the temperature-sensitive load component, temperature-sensitive elasticity is calculated, and typical change patterns are extracted, including: The strength of the growth trend is obtained by calculating the average slope of the corresponding long-term trend term during the analysis period. The fluctuation entropy value is obtained by calculating the sample entropy from the baseline load residual sequence after removing the long-term trend; A nonlinear regression method is used to fit the functional relationship between the temperature-sensitive load component and the corresponding time period temperature data; the average first derivative of the functional relationship in the typical operating temperature range is calculated as the temperature-sensitive elasticity. The characteristic curves of the temperature-sensitive load components are obtained, and the typical variation patterns are obtained based on the characteristic curves through cluster analysis.

[0082] The purpose of step S400 in this embodiment is to conduct targeted feature quantification and pattern mining on the decoupled base load component and temperature-sensitive load component. From dimensions such as growth trend, fluctuation characteristics, temperature sensitivity, and change patterns, core characteristic indicators and typical patterns of the two types of load components are extracted. This achieves both a quantitative description of the base load and temperature-sensitive load characteristics and uncovers the inherent patterns of temperature-sensitive load changes with temperature. It provides accurate feature data support for subsequent calculation of the comprehensive quality coefficient using multiple indicators and for attribution analysis of load growth drivers, solving the problems of insufficient characterization and pattern mining in traditional load analysis.

[0083] To achieve the above objective, the strength of the growth trend must first be obtained by calculating the average slope of the corresponding long-term trend term during the analysis period. This involves performing trend decomposition on the time-series data of the decoupled base load components, extracting the long-term trend term, and calculating the average slope of this long-term trend term over the analyzed period to obtain the strength of the growth trend. This indicator quantitatively reflects the overall growth or decline trend of the base load during the analysis period. A positive slope indicates an increasing trend in the base load, with a larger slope value indicating a stronger growth trend, while a negative slope indicates a declining trend. For example, the long-term trend term of the base load component for the chemical industry from January to June 2024 is a linear function y = 5x + 1000, where x represents the month and y represents the base load value in kW. The slope of this function is 5, meaning that the growth trend strength of the chemical industry's base load during this period is 5 kW / month, representing an average monthly increase of 5 kW in the base load.

[0084] Next, the fluctuation entropy value needs to be obtained by calculating the sample entropy of the basic load residual sequence after removing the long-term trend; First, the extracted long-term trend term is removed from the time-series data of the base load components to obtain the base load residual sequence. This sequence reflects the random fluctuation characteristics of the base load after removing the overall trend. Then, the sample entropy is calculated from this residual sequence. The resulting sample entropy value is the fluctuation entropy value. This fluctuation entropy value quantifies the complexity of the base load fluctuations. A larger fluctuation entropy value indicates more severe and complex random fluctuations in the base load, while a smaller value indicates more stable and regular fluctuations. For example, the sample entropy value of the base load residual sequence after removing the long-term trend for the steel industry is 0.25. This value is the fluctuation entropy value, indicating that the base load has a low degree of fluctuation and a stable operating state after removing the trend.

[0085] Then, a nonlinear regression method is needed to fit the functional relationship between the temperature-sensitive load component and the corresponding time period temperature data; the average first derivative of the functional relationship in the typical working temperature range is calculated as the temperature-sensitive elasticity. First, time-series data of the temperature-sensitive load component and corresponding temperature data for the period to be analyzed are collected. A nonlinear regression method is then used to fit the two sets of data, obtaining a nonlinear functional relationship between the temperature-sensitive load component and the temperature data. Next, the typical operating temperature range for the industry is determined. For example, the typical operating temperature range for a manufacturing industry is 15℃-35℃. Within this range, the first derivative of the fitted functional relationship is calculated, and the average value of all first derivative values ​​is taken. This average value is the temperature sensitivity elasticity. The temperature sensitivity elasticity quantitatively reflects the sensitivity of the temperature-sensitive load to changes in temperature. A larger value indicates a greater change in the temperature-sensitive load for every 1℃ change in temperature, and a stronger sensitivity to temperature.

[0086] For example, the functional relationship between the temperature-sensitive load component in the construction industry and the air temperature was obtained by fitting the data as y=2x 2 +3x+50, with a typical operating temperature range of 10℃-30℃, the average value of the first derivative of the function within this typical operating temperature range is 63, which means that the temperature sensitivity elasticity of the temperature-sensitive load in this industry is 63kW / ℃, representing that for every 1℃ increase in temperature, the temperature-sensitive load increases by an average of 63kW.

[0087] Finally, the characteristic curves of the temperature-sensitive load components are obtained, and based on the characteristic curves, the typical change patterns are obtained through cluster analysis.

[0088] First, based on the time-series data of the decoupled temperature-sensitive load components, characteristic curves of the temperature-sensitive load changes with time and temperature are plotted. These curves fully represent the trajectory of the temperature-sensitive load changes during the analysis period. Then, an appropriate clustering analysis method is selected, such as K-means clustering or hierarchical clustering, using the shape, amplitude, peak and trough values ​​of the characteristic curves as clustering features to perform clustering analysis on the temperature-sensitive load characteristic curves. Curves with similar change characteristics are grouped into one class, and the change pattern corresponding to each class represents the typical change pattern of the temperature-sensitive load.

[0089] For example, K-means clustering was performed on the characteristic curves of temperature-sensitive loads in commercial complexes, resulting in three typical change patterns: a continuous high load pattern during the high-temperature period in summer, a step-like growth pattern during the low-temperature period in winter, and a stable low load pattern during the mild periods in spring and autumn, which fully covers the main change patterns of temperature-sensitive loads in this type of scenario.

[0090] In step S500 of this application embodiment, the comprehensive quality coefficient characterizing the reliability of load data analysis for the target key industry is obtained by integrating the growth trend intensity, the fluctuation entropy value, the temperature-sensitive elasticity, the physical constraint compliance degree, and the industry adaptability, including: The growth trend strength, fluctuation entropy value, temperature sensitivity elasticity, physical constraint compliance and industry adaptability are used as input factors, and preset weight coefficients are configured for each parameter in the input factors. A comprehensive score is calculated using a weighted summation formula and used as the comprehensive quality coefficient; wherein the weight coefficients of the physical constraint compliance degree and the industry fit degree are higher than the weight coefficients of the other three input factors.

[0091] The purpose of step S500 in this embodiment is to integrate and quantify five core indicators obtained from the analysis of base load and temperature-sensitive load: growth trend intensity, fluctuation entropy value, temperature sensitivity elasticity, and physical constraint compliance and industry adaptability obtained from the load decoupling process. A comprehensive quality coefficient is then calculated through weighted summation. This coefficient serves as a unified quantitative evaluation standard for the reliability of target key industry load data analysis. It comprehensively reflects the physical rationality of load decoupling, industry matching degree, and the effectiveness of sub-load characteristic analysis. This provides a clear basis for subsequent selection of reliable base load data for attribution analysis, avoiding distortion of subsequent driving factor analysis results due to the use of unreliable analytical data.

[0092] To achieve the above objectives, it is first necessary to use the growth trend strength, fluctuation entropy value, temperature sensitivity elasticity, physical constraint compliance and industry adaptability as input factors, and configure preset weight coefficients for each parameter in the input factors. The growth trend strength, volatility entropy, temperature sensitivity elasticity, physical constraint compliance, and industry fit are directly used as input factors for calculating the comprehensive quality coefficient. First, each input factor is standardized, mapping its values ​​to the 0-1 range to eliminate dimensional differences. Then, weights are assigned to each of the five input factors, specifically requiring that the weights of physical constraint compliance and industry fit are higher than the other three, with a total weight of 1. For example, growth trend strength is set to 0.1, volatility entropy to 0.1, temperature sensitivity elasticity to 0.1, physical constraint compliance to 0.35, and industry fit to 0.35.

[0093] Next, a comprehensive score needs to be calculated using a weighted summation formula, which will serve as the comprehensive quality coefficient. The overall quality coefficient is obtained by multiplying the standardized factor values ​​by their corresponding weights and then summing all the products. The result is between 0 and 1, with values ​​closer to 1 indicating higher reliability of the data analysis. For example, the calculated overall quality coefficient is 0.8475.

[0094] In step S600 of this embodiment, the time-series data of the base load component corresponding to when the comprehensive quality coefficient is higher than a preset threshold are subjected to attribution analysis with the macroeconomic indicators of the target key industry to obtain the driving factors of power load, including: An econometric model is established using time-series data of the basic load component that has passed the comprehensive quality coefficient test as the dependent variable and macroeconomic indicators of the selected target key industries as independent variables. By calculating the significance of the standardized regression coefficients of the independent variables in the econometric model and analyzing the explanatory contribution of the independent variables to the fluctuation of the dependent variable based on variance decomposition, the influence intensity and statistical reliability of different economic factors on the changes in the base load in the macroeconomic indicators are quantitatively assessed. Based on the intensity of the impact and statistical reliability, the driving factors and their corresponding weights that drive the growth of electricity load in the target key industries are identified and determined.

[0095] The purpose of step S600 in this embodiment is to screen the time-series data of reliable base load components that meet the comprehensive quality coefficient standard, and to conduct attribution analysis by combining them with macroeconomic indicators of key target industries. This quantitatively identifies the core factors driving changes in industry power load and the influence weights of each factor from an economic perspective. The core objective is to establish a quantitative correlation between base load changes and macroeconomic development, overcoming the limitations of traditional load analysis which only focuses on data and cannot deeply attribute growth drivers. This provides a data-supported and economically logical decision-making basis for power grid planning, energy strategy formulation, and industry power consumption management, while ensuring that the attribution analysis is based on reliable data, thus improving the scientific rigor and practicality of the analysis results.

[0096] To achieve the above objectives, it is first necessary to establish an econometric model using time-series data of the basic load components that have passed the comprehensive quality coefficient test as the dependent variable and macroeconomic indicators of the selected target key industries as independent variables. First, a preset threshold is set for the comprehensive quality coefficient. This preset threshold is a critical value for judging the reliability of load data analysis and is determined by industry characteristics and analysis accuracy requirements. For example, the preset threshold for the comprehensive quality coefficient of the chemical industry is set to 0.8, while the preset threshold for the steel industry is set to 0.85 because the production stability requirements are higher.

[0097] The comprehensive quality coefficient calculated in step S500 is compared with the preset threshold. Time-series data of the basic load components with comprehensive quality coefficients higher than the preset threshold are selected. Only this type of reliable data is used as the basis for subsequent attribution analysis, while unreliable data with coefficients lower than the threshold are discarded to avoid low-quality data affecting the attribution analysis results. For example, if the comprehensive quality coefficient of a certain analysis in the chemical industry is 0.86, which is higher than the preset threshold of 0.8, the corresponding basic load time-series data can be used for subsequent attribution analysis; if the coefficient is 0.78, then this batch of data is discarded.

[0098] Next, by calculating the significance of the standardized regression coefficients of the independent variables in the econometric model and analyzing the explanatory contribution of the independent variables to the fluctuation of the dependent variable based on variance decomposition, the influence intensity and statistical reliability of different economic factors on the changes in the base load in the macroeconomic indicators are quantitatively assessed. The time-series data of the basic load components that passed the screening and testing were used as the dependent variable, which reflects the changes in the core electricity load of the target key industries that are not affected by temperature.

[0099] Based on the characteristics of the target key industries, select highly correlated macroeconomic indicators as independent variables. The independent variables should be able to reflect the production, operation and development status of the industry. For example, for the manufacturing industry, select industrial output value, fixed asset investment, capacity utilization rate and number of employees as independent variables; for the commercial industry, select total retail sales of consumer goods, shop operating area and customer traffic as independent variables; for the chemical industry, select total output of chemical products, industry main business revenue and raw material purchase volume as independent variables.

[0100] Based on the selected dependent and independent variables, a suitable econometric model is constructed. The type of econometric model is selected according to the correlation characteristics between the variables. For example, if the variables are linearly correlated, a multiple linear regression model is selected; if there are nonlinear or dynamic correlation characteristics, a vector autoregression (VAR) model or an error correction model (ECM) is selected. This establishes a quantitative analysis model between changes in base load and macroeconomic indicators. The specific methods for obtaining the econometric model are existing technologies and will not be elaborated here.

[0101] Then, it is necessary to quantitatively assess the impact strength and statistical reliability of each economic factor.

[0102] Specifically, after solving and testing the established econometric model, the impact of each independent variable on the dependent variable is quantitatively assessed from two dimensions.

[0103] First, calculate the significance of the standardized regression coefficients: Standardization involves standardizing the regression coefficients of the econometric model to obtain standardized regression coefficients. These standardized coefficients eliminate the influence of the independent variable's dimensions, and their absolute values ​​represent the strength of the corresponding economic indicator's impact on changes in the base load; the larger the absolute value, the stronger the impact. Simultaneously, the statistical significance of the standardized regression coefficients is determined by calculating the p-value and t-statistic. Generally, a significance level of 0.05 is set. If the p-value < 0.05, it indicates that the independent variable's impact on changes in the base load is statistically significant and has reliable statistical meaning; if the p-value ≥ 0.05, it indicates that the impact is insignificant, and the independent variable can be considered for removal. For example, in the econometric model of the chemical industry, the standardized regression coefficient for total chemical product output is 0.75, and the p-value is 0.02 < 0.05, indicating that its impact on the base load is strong and statistically significant; the standardized regression coefficient for raw material purchases is 0.12, and the p-value is 0.65 > 0.05, indicating that its impact is weak and has no statistical reliability.

[0104] Second, the contribution is explained based on variance decomposition analysis: Variance decomposition analysis (VAR) is used to calculate the explanatory contribution of each independent variable in the econometric model to the fluctuation of the dependent variable (base load). This explanatory contribution represents the proportion of base load fluctuations that a particular economic indicator can explain. The higher the value, the stronger the explanatory power of that indicator for base load fluctuations, and the more important it is as a driving factor for load changes. For example, VAR decomposition analysis on the chemical industry's econometric model shows that total chemical product output contributes 65% to the base load fluctuations, industry main business revenue contributes 20%, raw material purchases contribute 5%, and the remainder are random factors. This indicates that total chemical product output is the core indicator explaining base load fluctuations.

[0105] Through the analysis of the above two dimensions, the actual impact strength and statistical reliability of each macroeconomic indicator on the change of base load are comprehensively determined. Macroeconomic indicators with insignificant impact and low explanatory contribution are eliminated, and core indicators are retained.

[0106] Finally, based on the aforementioned impact strength and statistical reliability, the driving factors and corresponding weights of the electricity load growth in the target key industries are identified and determined. Specifically, based on the above quantitative assessment results, macroeconomic indicators with large absolute values ​​of standardized regression coefficients, high statistical significance, and high contribution from variance decomposition are selected, and these indicators are identified as the core driving factors for the electricity load growth in the target key industries.

[0107] Based on the proportion of the variance decomposition explained by each core driving factor or the standardized regression coefficient, the corresponding weight of each factor is calculated. The weight value reflects the importance of the factor in driving the growth of the base load, and the sum of the weights is 1. For example, after analysis, the total output of chemical products and the industry's main business revenue were identified as the core driving factors in the chemical industry. Their variance decomposition explained contributions were 65% and 20%, respectively, with a total contribution of 85%. After normalization, the weight of the total output of chemical products is 65% / 85%≈0.765, and the weight of the industry's main business revenue is 20% / 85%≈0.235. That is, 76.5% of the growth of the base load in the chemical industry is driven by the total output of chemical products, and 23.5% is driven by the industry's main business revenue.

[0108] Through the above steps, the core driving factors of electricity load growth in key target industries and the influence weight of each factor are finally identified, and an in-depth attribution analysis from load data to economic drivers is completed.

[0109] Example 2, as Figure 2 As shown, based on the same inventive concept as the load-temperature decoupling-based power load forecasting method provided in Embodiment 1, this embodiment of the invention also provides a load-temperature decoupling-based power load forecasting system, comprising: The basic database construction module 11 is used to acquire load characteristic data of key target industries and integrate industry survey data with physical rules to construct a multi-source fusion database containing an industry knowledge graph; the industry survey data includes at least typical production processes, major energy-consuming equipment, and thermodynamic parameters; the load characteristic data includes at least historical load time series data, refined meteorological data, and industry production attribute data. The decoupling model acquisition module 12 is used to construct a deep learning network architecture including a physical constraint layer based on the multi-source fusion database; initialize the deep learning network parameters using industry knowledge graphs; train the deep learning network by industry to obtain a physical information load decoupling model for the target key industry, and the corresponding model confidence. The load data analysis module 13 is used to input the total load data of the target key industry during the period to be analyzed and the production-related characteristic data of the total load data into the physical information load decoupling model of the corresponding target key industry. The physical information load decoupling model automatically decomposes the total load data into a basic load component and a temperature-sensitive load component, and outputs the physical constraint compliance degree, which characterizes the physical consistency of the decoupling process, and the industry fit degree, which reflects the degree of matching of industry characteristics. The load component analysis module 14 is used to perform characteristic analysis on the basic load component, calculate the growth trend intensity and fluctuation entropy value; for the temperature-sensitive load component, calculate the temperature-sensitive elasticity and extract typical change patterns; Data quality analysis module 15 is used to integrate the growth trend intensity, the fluctuation entropy value, the temperature sensitivity elasticity, the physical constraint compliance degree and the industry adaptability to obtain a comprehensive quality coefficient characterizing the reliability of the target key industry load data analysis. The driving factor analysis module 16 is used to perform attribution analysis on the time series data of the base load component corresponding to the comprehensive quality coefficient being higher than the preset threshold, and the macroeconomic indicators of the target key industries to obtain the driving factors of power load.

[0110] Furthermore, the basic database construction module 11 includes the following execution steps: The physical rules include at least: the nonlinear relationship between the energy consumption of temperature control equipment and the indoor-outdoor temperature difference, the simplified form of the heat conduction equation of the building envelope, and the reference power and start-stop logic of production equipment under the corresponding process flow. Based on the industry survey data, an industry knowledge graph is constructed, and the physical rules are bound to the corresponding entity nodes as attributes or constraints; wherein, the entity nodes of the industry knowledge graph include at least energy-consuming equipment, production workshops, and typical process sections; Simultaneously, production time sequence templates and equipment rated parameters are extracted from the industry production attribute data and injected into the corresponding entity nodes as attribute information to form a structured knowledge network, which serves as the multi-source fusion database.

[0111] Furthermore, the decoupling model acquisition module 12 includes the following execution steps: A hierarchical neural network is constructed as the deep learning network. The hidden layers of the neural network contain at least one physical constraint sublayer, which is composed of differential operation units and is used to enforce the encoding of thermodynamic prior knowledge. The structure and initial parameters of the physical constraint sublayer are defined and initialized by extracting relevant entity nodes and relationship information from the industry knowledge graph.

[0112] The deep learning network is trained under supervision using historical load time-series data from the multi-source fusion database as input and verified reference values ​​of the same-period basic load and temperature-sensitive load as training targets. The total loss function of the deep learning network consists of prediction error loss and physical consistency loss, wherein the physical consistency loss term is used to penalize the part of the deep learning network output that violates the rules encoded by the physical constraint sublayer. The deep learning network is trained on a preset training set and validation set until convergence, thereby obtaining the physical information load decoupling model. The model confidence level corresponding to the physical information load decoupling model is calculated based on the prediction error on the validation set and the physical consistency loss during the training process.

[0113] Furthermore, the load data analysis module 13 includes the following execution steps: The production-related characteristic data includes at least refined meteorological data, time stamp data, and generalized event tags for the period to be analyzed; The refined meteorological data and the total load data are synchronized and aligned on the time scale. The time-identification data is used to mark weekdays, weekends, holidays, and industry-specific production periods; The generalized event label is used to identify known major events that affect industry production.

[0114] The physical information load decoupling model calculates the theoretical temperature control energy consumption baseline based on the input total load data of the target key industries and production-related characteristic data. The total load data is compared and separated from the temperature control energy consumption baseline, and combined with the capture of non-stationary modes by the time-series learning layer, the basic load component and the temperature-sensitive load component are jointly generated. The real-time theoretical output generated by the physical constraint layer for the current input is calculated, and the difference between the real-time theoretical output and the statistical mean calculated by the physical constraint layer on all outputs on the complete training dataset is analyzed. The difference is quantified as the degree of compliance with the physical constraint. Calculate the KL divergence between the total load data and production-related characteristic data of the target key industry currently input and the corresponding characteristic distribution of the target key industry used in the training of the physical information load decoupling model, and calculate the industry fit based on the KL divergence.

[0115] Furthermore, the load component analysis module 14 includes the following execution steps: The strength of the growth trend is obtained by calculating the average slope of the corresponding long-term trend term during the analysis period. The fluctuation entropy value is obtained by calculating the sample entropy from the baseline load residual sequence after removing the long-term trend; A nonlinear regression method is used to fit the functional relationship between the temperature-sensitive load component and the corresponding time period temperature data; the average first derivative of the functional relationship in the typical operating temperature range is calculated as the temperature-sensitive elasticity. The characteristic curves of the temperature-sensitive load components are obtained, and the typical variation patterns are obtained based on the characteristic curves through cluster analysis.

[0116] Furthermore, the data quality analysis module 15 includes the following execution steps: The growth trend strength, fluctuation entropy value, temperature sensitivity elasticity, physical constraint compliance and industry adaptability are used as input factors, and preset weight coefficients are configured for each parameter in the input factors. A comprehensive score is calculated using a weighted summation formula and used as the comprehensive quality coefficient; wherein the weight coefficients of the physical constraint compliance degree and the industry fit degree are higher than the weight coefficients of the other three input factors.

[0117] Furthermore, the driving factor analysis module 16 includes the following execution steps: An econometric model is established using time-series data of the basic load component that has passed the comprehensive quality coefficient test as the dependent variable and macroeconomic indicators of the selected target key industries as independent variables. By calculating the significance of the standardized regression coefficients of the independent variables in the econometric model and analyzing the explanatory contribution of the independent variables to the fluctuation of the dependent variable based on variance decomposition, the influence intensity and statistical reliability of different economic factors on the changes in the base load in the macroeconomic indicators are quantitatively assessed. Based on the intensity of the impact and statistical reliability, the driving factors and their corresponding weights that drive the growth of electricity load in the target key industries are identified and determined.

Claims

1. A power load forecasting method based on load-temperature decoupling, characterized in that, include: Obtain load characteristic data of key target industries, and integrate industry survey data with physical rules to construct a multi-source fusion database containing industry knowledge graphs; Based on the multi-source fusion database, a deep learning network architecture including a physical constraint layer is constructed; the parameters of the deep learning network are initialized using an industry knowledge graph, and the deep learning network is trained for each industry to obtain a physical information load decoupling model for the target key industry, as well as the corresponding model confidence. Input the total load data of the target key industry for the period to be analyzed and the production-related characteristic data of the total load data into the physical information load decoupling model of the corresponding target key industry; Obtain the basic load component and the temperature-sensitive load component, and output the physical constraint compliance and industry adaptability, including: The physical information load decoupling model calculates the theoretical temperature control energy consumption baseline based on the input total load data of the target key industries and production-related characteristic data. The total load data is compared and separated with the temperature control energy consumption baseline, and combined with the capture of non-stationary modes by the time series learning layer, the basic load component and the temperature-sensitive load component are generated together. The real-time theoretical output generated by the physical constraint layer for the current input is calculated, and the difference between the real-time theoretical output and the statistical mean calculated by the physical constraint layer on all outputs on the complete training dataset is analyzed. The difference is quantified as the degree of compliance with the physical constraint. Calculate the KL divergence between the total load data and production-related characteristic data of the target key industry currently input and the corresponding characteristic distribution of the target key industry used in the training of the physical information load decoupling model; and calculate the industry fit based on the KL divergence. The basic load component is subjected to feature analysis to calculate the growth trend intensity and fluctuation entropy value; the temperature-sensitive load component is calculated to calculate the temperature sensitivity elasticity and extract typical change patterns; the growth trend intensity, fluctuation entropy value, temperature sensitivity elasticity, physical constraint compliance, and industry adaptability are fused to obtain a comprehensive quality coefficient characterizing the reliability of load data analysis for the target key industry; the time series data of the basic load component corresponding to the comprehensive quality coefficient being higher than a preset threshold are subjected to attribution analysis with the macroeconomic indicators of the target key industry to analyze and obtain the driving factors of power load.

2. The power load forecasting method based on load-temperature decoupling according to claim 1, characterized in that, Obtain load characteristic data for key target industries, integrate industry survey data with physical rules, and construct a multi-source fusion database including an industry knowledge graph, including: The load characteristic data includes at least historical load time series data, refined meteorological data, and industry production attribute data; The physical rules include at least the nonlinear relationship between the energy consumption of temperature control equipment and the indoor-outdoor temperature difference, the simplified form of the heat conduction equation of the building envelope, and the reference power and start-stop logic of the production equipment under the corresponding process flow. Based on the industry survey data, an industry knowledge graph is constructed, and the physical rules are bound to the corresponding entity nodes as attributes or constraints; wherein, the entity nodes of the industry knowledge graph include at least energy-consuming equipment, production workshops, and typical process sections, and the industry survey data includes at least typical production processes, main energy-consuming equipment, and thermodynamic parameters; Simultaneously, production time sequence templates and equipment rated parameters are extracted from the industry production attribute data and injected into the corresponding entity nodes as attribute information to form a structured knowledge network, which serves as the multi-source fusion database.

3. The power load forecasting method based on load-temperature decoupling according to claim 1, characterized in that, Based on the aforementioned multi-source fusion database, a deep learning network architecture including a physical constraint layer is constructed, comprising: A hierarchical neural network is constructed as the deep learning network. The hidden layers of the neural network contain at least one physical constraint sublayer, which is composed of differential operation units and is used to enforce the encoding of thermodynamic prior knowledge. The structure and initial parameters of the physical constraint sublayer are defined and initialized by extracting relevant entity nodes and relationship information from the industry knowledge graph.

4. The power load forecasting method based on load-temperature decoupling according to claim 3, characterized in that, The parameters of a deep learning network are initialized using an industry knowledge graph, and the deep learning network is trained industry-specifically to obtain a physical information load decoupling model for the target key industry, along with the corresponding model confidence, including: The deep learning network is trained under supervision using historical load time-series data from the multi-source fusion database as input and verified reference values ​​of the same-period basic load and temperature-sensitive load as training targets. The total loss function of the deep learning network consists of prediction error loss and physical consistency loss, wherein the physical consistency loss term is used to penalize the part of the deep learning network output that violates the rules encoded by the physical constraint sublayer. The deep learning network is trained on a preset training set and validation set until convergence, thereby obtaining the physical information load decoupling model. The model confidence level corresponding to the physical information load decoupling model is calculated based on the prediction error on the validation set and the physical consistency loss during the training process.

5. The power load forecasting method based on load-temperature decoupling according to claim 1, characterized in that, Input the total load data of the target key industry for the period to be analyzed and the production-related characteristic data of the total load data into the physical information load decoupling model of the corresponding target key industry, including: The production-related characteristic data includes at least refined meteorological data, time-stamped data, and generalized event tags for the period to be analyzed; The refined meteorological data and the total load data are synchronized and aligned on the time scale. The time-identification data is used to mark weekdays, weekends, holidays, and industry-specific production periods; The generalized event label is used to identify known major events that affect industry production.

6. The power load forecasting method based on load-temperature decoupling according to claim 1, characterized in that, The basic load components are subjected to characteristic analysis to calculate the growth trend intensity and fluctuation entropy value; For the temperature-sensitive load component, calculate the temperature-sensitive elasticity and extract typical variation patterns, including: The strength of the growth trend is obtained by calculating the average slope of the corresponding long-term trend term during the analysis period. The fluctuation entropy value is obtained by calculating the sample entropy from the baseline load residual sequence after removing the long-term trend; A nonlinear regression method is used to fit the functional relationship between the temperature-sensitive load component and the corresponding time period temperature data; the average first derivative of the functional relationship in the typical operating temperature range is calculated as the temperature-sensitive elasticity. The characteristic curves of the temperature-sensitive load components are obtained, and the typical variation patterns are obtained based on the characteristic curves through cluster analysis.

7. The power load forecasting method based on load-temperature decoupling according to claim 1, characterized in that, By integrating the growth trend strength, the fluctuation entropy value, the temperature-sensitive elasticity, the physical constraint compliance, and the industry adaptability, a comprehensive quality coefficient characterizing the reliability of load data analysis for key target industries is obtained, including: The growth trend strength, fluctuation entropy value, temperature sensitivity elasticity, physical constraint compliance and industry adaptability are used as input factors, and preset weight coefficients are configured for each parameter in the input factors. A comprehensive score is calculated using a weighted summation formula and used as the comprehensive quality coefficient; wherein the weight coefficients of the physical constraint compliance degree and the industry fit degree are higher than the weight coefficients of the other three input factors.

8. The power load forecasting method based on load-temperature decoupling according to claim 1, characterized in that, The time-series data of the base load component corresponding to when the comprehensive quality coefficient exceeds a preset threshold are used to conduct attribution analysis with macroeconomic indicators of key target industries to obtain the driving factors of electricity load, including: An econometric model is established using time-series data of the basic load component that has passed the comprehensive quality coefficient test as the dependent variable and macroeconomic indicators of the selected target key industries as independent variables. By calculating the significance of the standardized regression coefficients of the independent variables in the econometric model and analyzing the explanatory contribution of the independent variables to the fluctuation of the dependent variable based on variance decomposition, the influence intensity and statistical reliability of different economic factors on the changes in the base load in the macroeconomic indicators are quantitatively assessed. Based on the intensity of the impact and statistical reliability, the driving factors and their corresponding weights that drive the growth of electricity load in the target key industries are identified and determined.

9. A power load forecasting system based on load-temperature decoupling, characterized in that, The system is used to implement a power load forecasting method based on load-temperature decoupling as described in any one of claims 1-8, comprising: The basic database construction module is used to acquire load characteristic data of key target industries and integrate industry survey data with physical rules to construct a multi-source fusion database containing an industry knowledge graph. The industry survey data includes at least typical production processes, major energy-consuming equipment, and thermodynamic parameters. The load characteristic data includes at least historical load time series data, refined meteorological data, and industry production attribute data. The decoupling model acquisition module is used to construct a deep learning network architecture including a physical constraint layer based on the multi-source fusion database; initialize the deep learning network parameters using industry knowledge graphs; train the deep learning network by industry to obtain a physical information load decoupling model for the target key industry, and the corresponding model confidence. The load data analysis module is used to input the total load data of the target key industry for the period to be analyzed and the production-related characteristic data of the total load data into the physical information load decoupling model of the corresponding target key industry. The physical information load decoupling model automatically decomposes the total load data into a basic load component and a temperature-sensitive load component, and outputs the physical constraint compliance degree, which characterizes the physical consistency of the decoupling process, and the industry fit degree, which reflects the degree of matching of industry characteristics, including: The physical information load decoupling model calculates the theoretical temperature control energy consumption baseline based on the input total load data of the target key industries and production-related characteristic data. The total load data is compared and separated with the temperature control energy consumption baseline, and combined with the capture of non-stationary modes by the time series learning layer, the basic load component and the temperature-sensitive load component are generated together. The real-time theoretical output generated by the physical constraint layer for the current input is calculated, and the difference between the real-time theoretical output and the statistical mean calculated by the physical constraint layer on all outputs on the complete training dataset is analyzed. The difference is quantified as the degree of compliance with the physical constraint. Calculate the KL divergence between the total load data and production-related characteristic data of the target key industry currently input and the corresponding characteristic distribution of the target key industry used in the training of the physical information load decoupling model; and calculate the industry fit based on the KL divergence. The load component analysis module is used to perform characteristic analysis on the basic load component, calculate the growth trend intensity and fluctuation entropy value; for the temperature-sensitive load component, calculate the temperature-sensitive elasticity and extract typical change patterns; The data quality analysis module is used to integrate the growth trend strength, the fluctuation entropy value, the temperature sensitivity elasticity, the physical constraint compliance degree, and the industry adaptability to obtain a comprehensive quality coefficient that characterizes the reliability of the target key industry load data analysis. The driving factor analysis module is used to perform attribution analysis on the time series data of the base load component corresponding to the comprehensive quality coefficient being higher than the preset threshold, and on the macroeconomic indicators of the target key industries to obtain the driving factors of power load.

Images