Power distribution system source load operation state deduction method based on causal migration and related equipment

Abstract

The invention belongs to the technical field of power system operation and planning, and provides a power distribution system source load operation state deduction method based on causal migration and related equipment. The method comprises the following steps: constructing a candidate feature pool based on historical multi-source operation data, meteorological data and economic data of a power distribution system; performing nonlinear causal screening on the feature factors in the candidate feature pool based on the approximate transfer entropy to determine key feature factors influencing the operation state of the source load and corresponding causal intensity scores; constructing a deep causal gating cyclic memory network, and performing non-randomized initialization on an input gate weight of the deep causal gating cyclic memory network based on the key feature factors and the corresponding causal strength scores thereof to construct a time sequence deduction model; and dynamically deducing a source load operation state track of the power distribution system in a future planning period based on the time sequence deduction model. Through the technical scheme provided by the invention, the accuracy of predicting the source load operation state of the power distribution system can be improved.

Description

Technical Field

[0001] This application belongs to the field of power system operation and planning technology, and in particular relates to a method and related equipment for inferring the source-load operation status of a distribution system based on causal migration. Background Technology

[0002] Driven by the "dual carbon" goals, the large-scale integration of distributed new energy sources (such as photovoltaic and wind power) and interactive loads has resulted in highly stochastic and spatiotemporally coupled characteristics on both the source and load sides of the power distribution system. Traditional static planning methods mainly rely on typical daily data, severing the dynamic connection between the planning objectives and the actual operating status of the power distribution system, making it difficult to capture system boundary risks under extreme weather conditions or sudden changes in source and load. This makes it difficult for existing schemes to effectively quantify power supply adequacy and absorption capacity, causing power system planning schemes to frequently oscillate between excessive power redundancy and insufficient power safety margin. Therefore, how to improve the accuracy of power distribution system source and load operating status prediction and achieve dynamic coordination of power distribution system planning and operation to avoid excessive power redundancy and insufficient power safety margin in the formulated power system planning schemes is an urgent technical problem to be solved. Summary of the Invention

[0003] This application provides a method, related equipment, computer program product, computer-readable storage medium and electronic equipment for predicting the source and load operating status of a power distribution system based on causal transfer, which can improve the accuracy of predicting the source and load operating status of a power distribution system to a certain extent.

[0004] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0005] According to a first aspect of this application, a method for extrapolating the source-load operation status of a power distribution system based on causal transfer is provided. The method includes: constructing a candidate feature pool based on historical multi-source operation data, meteorological data, and economic data of the power distribution system, wherein the multi-source operation data includes power load data, photovoltaic output data, and wind power output data; performing nonlinear causal screening on feature factors in the candidate feature pool based on approximate transfer entropy to determine key feature factors affecting the source-load operation status and their corresponding causal strength scores; constructing a deep causal gated recurrent memory network, and performing non-randomized initialization of the input gate weights of the deep causal gated recurrent memory network based on the key feature factors and their corresponding causal strength scores to construct a time-series extrapolation model; and dynamically extrapolating the source-load operation status trajectory of the power distribution system within a future planning period based on the time-series extrapolation model.

[0006] In this application, based on the aforementioned scheme, the construction of a candidate feature pool based on historical multi-source operation data, meteorological data, and economic data of the power distribution system includes: obtaining the multi-source operation data, meteorological data, and economic data of the power distribution system at historical moments from the power distribution system database as raw data; using the raw data and its corresponding multi-order lag historical sequences as basic features, and using pre-released power load forecast data of the power system as prior features; integrating the basic features, the prior features, and the historical observations of the target variable to be predicted to obtain a candidate feature pool containing external driving factors and internal autoregressive factors, wherein the target variable includes power load, photovoltaic power output, and wind power output.

[0007] In this application, based on the aforementioned scheme, the nonlinear causal screening of feature factors in the candidate feature pool based on approximate transfer entropy to determine key feature factors affecting the source load operating state and their corresponding causal strength scores includes: using mutual information as an effective approximation of approximate transfer entropy, calculating the nonlinear correlation between the lagged sequences of each feature factor in the candidate feature pool and the future sequences of the target variable, quantifying the causal strength of each feature factor on the target variable as a positive driving force, and quantifying the causal strength of the target variable on each feature factor as a negative feedback force; based on the positive driving force and the negative feedback force, using a bidirectional net causal screening mechanism to calculate the net causal score of each feature factor, and screening feature factors with net causal scores higher than a preset score threshold from the candidate feature pool as key feature factors; and determining the net causal score of the key feature factor as the causal strength score of the key feature factor.

[0008] In this application, based on the aforementioned scheme, the construction of a deep causal gated recurrent memory network and the non-random initialization of the input gate weights of the deep causal gated recurrent memory network based on the key feature factors and their corresponding causal strength scores include: constructing a bidirectional temporal memory evolution layer, using a bidirectional long short-term memory network as the inference backbone, performing temporal modeling of the key feature factors through forward and reverse recursive chains to construct the deep causal gated recurrent memory network; mapping the causal strength scores corresponding to the key feature factors into weight gain matrices, and applying them to the input gate weight matrix of the first layer of the bidirectional long short-term memory network to achieve non-random initialization of the initial weights of the input gates.

[0009] In this application, based on the aforementioned scheme, the process of constructing the time series extrapolation model further includes: introducing an attention mechanism to dynamically weight the output state of the bidirectional long short-term memory network; constructing a lightweight gated recurrent unit network as a residual model, learning the residual sequence between the output result of the bidirectional long short-term memory network and the true value through the residual model, so as to superimpose the residual prediction value of the residual model onto the output result of the bidirectional long short-term memory network to obtain a high-fidelity intermediate result in the process of extrapolating the source load operation state, and using the network model corresponding to the intermediate result as the time series extrapolation model.

[0010] In this application, based on the aforementioned scheme, the step of dynamically extrapolating the source-load operating state trajectory of the power distribution system within a future planning period based on the time-series extrapolation model includes: constructing a multi-task quantile loss function and using an asymmetric bouncing ball loss to guide the time-series extrapolation model to learn the conditional probability distribution of the target variable; introducing into the multi-task quantile loss function Regularization terms are used to construct the objective function; the objective function is minimized using the gradient descent algorithm, and the time series extrapolation model is trained to output the source load operation state prediction results corresponding to the 10th quantile, 50th quantile, and 90th quantile; based on the source load operation state prediction results, the source load operation state trajectories under conservative, neutral, and aggressive scenarios are determined respectively, forming a set of source load operation state trajectories covering the full probability space within the future planning period.

[0011] In this application, based on the aforementioned scheme, after forming a set of source-load operating state trajectories covering the full probability space within the future planning period, the method further includes: conducting boundary pressure tests on the power supply adequacy and absorption capacity of the current planning scheme of the power distribution system based on the source-load operating state trajectories under conservative, neutral, and aggressive scenarios in the set of source-load operating state trajectories; quantitatively verifying the robustness of the current planning scheme based on the boundary pressure test results, and adjusting the planning parameters of the current planning scheme based on the verification results, so as to achieve dynamic coordination of the integrated planning and operation of the power distribution system.

[0012] According to a second aspect of this application, a device for extrapolating the source-load operation status of a power distribution system based on causal transfer is provided. The device includes: a first construction unit for constructing a candidate feature pool based on historical multi-source operation data, meteorological data, and economic data of the power distribution system, wherein the multi-source operation data includes power load data, photovoltaic output data, and wind power output data; a screening unit for performing nonlinear causal screening on feature factors in the candidate feature pool based on approximate transfer entropy to determine key feature factors affecting the source-load operation status and their corresponding causal strength scores; a second construction unit for constructing a deep causal-gated recurrent memory network and performing non-randomized initialization of the input gate weights of the deep causal-gated recurrent memory network based on the key feature factors and their corresponding causal strength scores to construct a time-series extrapolation model; and an extrapolation unit for dynamically extrapolating the source-load operation status trajectory of the power distribution system within a future planning period based on the time-series extrapolation model.

[0013] According to a third aspect of this application, a computer program product is provided, the computer program product including computer instructions stored in a computer-readable storage medium and adapted to be read and executed by a processor to cause a computer device having the processor to perform operations as described in the first aspect above.

[0014] According to a fourth aspect of this application, a computer-readable storage medium is provided, wherein at least one computer program instruction is stored therein, the at least one computer program instruction being loaded and executed by a processor to perform the operation as described in the first aspect above.

[0015] According to a fifth aspect of this application, an electronic device is provided, the electronic device including one or more processors and one or more memories, the one or more memories storing at least one computer program instruction, the at least one computer program instruction being loaded and executed by the one or more processors to perform the operation as described in the first aspect above.

[0016] Based on the technical solution proposed in this application, a candidate feature pool is constructed by integrating multi-source historical data from the power distribution system, ensuring the comprehensiveness and richness of the data and providing a solid data foundation for subsequent causal screening and model construction. Nonlinear causal screening based on approximate transfer entropy can accurately identify key feature factors affecting the operating state of source loads, eliminate redundant information, reduce model computational complexity, and improve the model's ability to capture key information. By using the causal strength scores of key feature factors to perform non-randomized initialization of the input gate weights of the deep causal gated recurrent memory network, a time-series extrapolation model is constructed. This effectively integrates causal prior knowledge with the deep learning model, improving the model's training efficiency and generalization ability, and avoiding the problems of overfitting and slow convergence of traditional models in small sample or complex data scenarios. Based on this time-series extrapolation model, the trajectory of source load operating states within future planning cycles can be dynamically and accurately extrapolated, providing a reliable decision-making basis for power distribution system planning. This breaks the disconnect between traditional planning methods and actual operating states, laying the foundation for dynamic collaboration in the integrated planning and operation of power distribution systems.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings: Figure 1 A flowchart of the power distribution system source-load operation state deduction method based on causal migration in an embodiment of this application is shown; Figure 2 The diagram shows the verification of the causal screening effect in an embodiment of this application; Figure 3 The following diagram shows a comparison of model training convergence in the embodiments of this application; Figure 4 The following diagram illustrates the multi-task probability prediction and scene generation effects in an embodiment of this application. Figure 5 A block diagram of the equipment related to the source-load operation state inference of the power distribution system based on causal migration is shown in an embodiment of this application; Figure 6 A schematic diagram of the structure of an electronic device in an embodiment of this application is shown. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0021] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. It should also be noted that, for the sake of simplicity, certain components in the drawings that do not affect the interpretation of the technical solution of this application have been appropriately omitted.

[0022] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined. Therefore, the actual execution order may change depending on the actual situation.

[0023] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "multiple" means two or more.

[0024] Currently, the large-scale integration of distributed renewable energy sources (such as photovoltaic and wind power) and interactive loads has resulted in highly stochastic and spatiotemporally coupled characteristics on both the source and load sides of the power distribution system. Traditional static planning methods mainly rely on typical daily data, severing the dynamic connection between the planning objectives and the actual operating status of the power distribution system, making it difficult to capture system boundary risks under extreme weather conditions or sudden changes in source and load. For example, when planning a power distribution system in a certain region, only typical electricity consumption data and photovoltaic and wind power output data from a few days in summer were referenced, without considering the surge in electricity load caused by extreme high temperatures or the sharp drop in renewable energy output caused by continuous rainy days, leading to power supply shortages in actual operation.

[0025] The limitations of this traditional planning method make it difficult for existing schemes to effectively quantify power supply sufficiency and absorption capacity. Consequently, power system planning schemes often oscillate between excessive power redundancy and insufficient power safety margin. If the planning scheme is too conservative, it will lead to excessive redundancy in power facility construction, resulting in a waste of resources and funds. If the planning scheme is too aggressive, it will result in insufficient power safety margin, making it unable to cope with the impact of sudden changes in power supply and load, and affecting the stable operation of the distribution system.

[0026] In this context, this application proposes a source-load operation state prediction scheme based on causal migration of power distribution systems to improve the accuracy of source-load operation state prediction, thereby laying the foundation for dynamic coordination of integrated planning and operation of power distribution systems and avoiding situations where the formulated power system planning scheme has excessive power redundancy and insufficient power safety margin.

[0027] The implementation details of the technical solutions in the embodiments of this application are described below: Reference Figure 1 The diagram shows a flowchart of a power distribution system source-load operation state deduction method based on causal transition in an embodiment of this application. This power distribution system source-load operation state deduction method based on causal transition can be executed by a device with computational processing capabilities.

[0028] Reference Figure 1 As shown, the method for inferring the source-load operating state of a power distribution system based on causal transition includes at least steps 110 to 140, which are described in detail below: In step 110, a candidate feature pool is constructed based on the historical multi-source operation data, meteorological data, and economic data of the power distribution system. The multi-source operation data includes power load data, photovoltaic power output data, and wind power output data.

[0029] In this application, the historical data of the power distribution system can refer to the data continuously collected or recorded by the power distribution system over a period of time in the past, such as relevant data over a time span of 1 year or 3 years. Through the accumulation of data over a sufficiently long period of time, the changing patterns and trends of the data can be better reflected.

[0030] In this application, power load data refers to the total amount of electrical energy consumed by the power distribution system at different times. For example, the power load data of a city's power distribution system recorded every 15 minutes from 00:00 to 24:00 on June 1, 2023, may fluctuate between 3000MW and 5000MW. Photovoltaic output data refers to the electrical energy generated by photovoltaic power generation equipment in the power distribution system. For example, the power generation data of a distributed photovoltaic power station in a certain area at different times of the day throughout 2023. Due to the influence of sunlight intensity, the output data is usually higher at noon and zero in the early morning. Wind power output data refers to the electrical energy generated by wind power generation equipment. For example, the power generation data of a wind farm in each month of 2023. Due to the influence of wind speed, the output data is relatively higher in months with stronger winds.

[0031] In this application, meteorological data may include meteorological parameters related to the operation of the power distribution system's power sources and loads, such as temperature, humidity, wind speed, sunshine duration, and rainfall. Economic data may include economic indicators that can affect power load and the development of new energy sources, such as regional GDP, industrial added value, and enterprise operating rates.

[0032] In this application, the process of constructing a candidate feature pool can be an initial data set formed by collecting, organizing, and cleaning the aforementioned historical data, and integrating them together to provide basic data support for subsequent causal screening.

[0033] Specifically, in such Figure 1 In step 110, the construction of a candidate feature pool based on historical multi-source operational data, meteorological data, and economic data of the power distribution system can be performed according to steps 111 to 113 as follows: Step 111: Obtain the multi-source operation data, meteorological data, and economic data of the power distribution system at historical moments from the power distribution system database as raw data.

[0034] Step 112: The original data and its corresponding multi-stage lag historical sequence are used as basic features, and the pre-released power load forecast data of the power system are used as prior features.

[0035] Step 113: Integrate the basic features, the prior features, and the historical observations of the target variable to be predicted to obtain a candidate feature pool containing external driving factors and internal autoregressive factors. The target variable includes power load, photovoltaic power output, and wind power output.

[0036] In this application, the power distribution system database can be a dedicated database storing various operational data related to the power distribution system. This database can continuously record multi-source operational data, meteorological data, and economic data at certain time intervals (e.g., 15 minutes, 1 hour). For example, a power distribution system database stores daily 15-minute data on power load, photovoltaic output, and wind power output, daily average temperature, daily wind speed, and daily sunshine duration, as well as monthly economic data such as industrial added value and enterprise operating rates, for the period 2020-2023. When obtaining raw data, it is necessary to extract data from this database according to a preset time range (e.g., the past 3 years) and data type, and to perform preliminary screening and cleaning of the extracted data to remove obvious abnormal data (e.g., unreasonable data caused by sensor malfunctions) to ensure the accuracy and reliability of the raw data.

[0037] In this application, a multi-stage lag history sequence refers to a historical data sequence that traces back different time steps from a certain point in time. For example, for the power load data at 12:00 on June 1, 2023, the first-stage lag history data is the power load data at 11:45 on June 1, 2023, the second-stage lag history data is the power load data at 11:30 on June 1, 2023, and so on. The lag order can be set according to actual needs (e.g., 1-24 stages, corresponding to historical data within the past 6 hours). Using the original data and its multi-stage lag history sequence as basic features, the time series characteristics of the data can be fully utilized to capture the dynamic correlation between data. For example, by analyzing the multi-stage lag history sequence of power load data, the time inertia characteristics of power load can be discovered (e.g., if the load was high at the previous moment, the load at the current moment may also remain at a high level).

[0038] In this application, the pre-released power system load forecast data refers to power load forecast data for a specific region and time period released in advance by power management departments, professional forecasting agencies, etc., such as the power load forecast report for a certain province in July 2023 released by the State Grid Corporation of China, which includes the expected power load data for different time periods every day in July. Using this data as a priori features can provide additional reference information for the model, helping the model to better learn and predict the source load operating status. For example, when there is a large deviation between the power load data predicted by the model itself and the pre-released forecast data, the reasons for the deviation can be analyzed in conjunction with relevant factors, and the model prediction results can be adjusted.

[0039] In this application, the historical observations of the target variables to be predicted specifically refer to the actual historical observation data of the three target variables: power load, photovoltaic output, and wind power output. For example, the actual observations of power load, photovoltaic output, and wind power output at different times of the day over the past three years.

[0040] In this application, external driving factors can refer to features corresponding to factors outside the power distribution system that can influence changes in the target variable, including meteorological data, economic data, and pre-released power load forecast data. For example, increased temperature leads to increased power load (increased air conditioning electricity consumption), and increased wind speed leads to increased wind power output; these are all examples of the influence of external driving factors. Internal autoregressive factors can refer to features formed by the historical data of the target variable itself, including the original data of the target variable and its multi-stage lagged historical sequences. These reflect the time evolution and inertial characteristics of the target variable itself. For example, historical observations of power load can reflect the continuity of electricity consumption habits, and historical data of photovoltaic power output can reflect the influence of different seasons and time periods on photovoltaic power generation. By integrating external driving factors and internal autoregressive factors, a complete candidate feature pool is constructed, ensuring that the feature pool contains all relevant information that may affect the source load operating status. This provides comprehensive and complete feature data for subsequent nonlinear causal screening based on approximate transfer entropy, avoiding inaccurate causal relationship identification due to missing features.

[0041] Based on the technical solutions in steps 111 to 113 above, by clarifying the source of the original data (power distribution system database) and the data processing methods (screening and cleaning), the accuracy and reliability of the basic data can be ensured, providing high-quality raw materials for the construction of the candidate feature pool. Using the original data and its multi-stage lag historical sequences as basic features can fully explore the time series characteristics of the data, capture the dynamic correlation and time inertia characteristics of the target variable, and help improve the subsequent model's ability to learn the changing patterns of source load operation status. By introducing pre-released power load forecast data as prior features, additional reference information can be provided to the model, enriching the diversity and effectiveness of the features. By integrating the basic features, prior features, and historical observations of the target variable, a candidate feature pool containing external driving factors and internal autoregressive factors can be constructed, ensuring the comprehensiveness and completeness of the features. This ensures that all factors affecting the source load operation status are fully covered, laying a solid foundation for the subsequent accurate identification of key feature factors, thereby improving the accuracy and reliability of the entire source load operation status extrapolation method.

[0042] Continue to refer to Figure 1 In step 120, nonlinear causal screening is performed on the feature factors in the candidate feature pool based on approximate transfer entropy to determine the key feature factors affecting the source load operation state and their corresponding causal strength scores.

[0043] In this application, it should be noted that approximate transfer entropy is a nonlinear causal relationship quantification tool based on information theory, which can effectively capture the nonlinear causal dependency between two variables. Compared with traditional linear correlation analysis methods, it is more suitable for handling complex nonlinear data relationships in power distribution systems.

[0044] In this application, feature factors specifically refer to the variables corresponding to various types of data contained in the candidate feature pool, such as the power load variable corresponding to power load data and the temperature variable corresponding to temperature data. By calculating approximate transfer entropy, the degree of causal relationship between each feature factor and the source load operating state can be quantified, i.e., the causal strength score. This allows for the selection of key feature factors that have the most significant impact on the source load operating state, while eliminating irrelevant or less influential feature factors. This reduces the complexity of subsequent model calculations and improves the model's operating efficiency and inference accuracy.

[0045] Specifically, in such Figure 1 In step 120, the nonlinear causal screening of feature factors in the candidate feature pool based on approximate transfer entropy to determine the key feature factors affecting the operating state of the source load and their corresponding causal strength scores can be performed according to steps 121 to 123 as follows: Step 121: Using mutual information as an effective approximation of the approximate transfer entropy, calculate the nonlinear correlation between the lag sequence of each feature factor in the candidate feature pool and the future sequence of the target variable, so as to quantify the causal strength of each feature factor on the target variable as a positive driving force, and quantify the causal strength of the target variable on each feature factor as a negative feedback force.

[0046] Step 122: Based on the positive driving force and the negative feedback force, a bidirectional net causal screening mechanism is used to calculate the net causal score of each feature factor, and feature factors with net causal scores higher than a preset score threshold are selected from the candidate feature pool as key feature factors.

[0047] Step 123: The net causal score of the key feature factor is determined as the causal strength score of the key feature factor.

[0048] In this application, it should be noted that mutual information is an indicator that measures the degree of correlation between two random variables. It can effectively capture the nonlinear relationship between variables, and the larger the value, the stronger the correlation between the two variables. Since directly calculating the transfer entropy is highly complex, this application uses mutual information as an effective approximation of the transfer entropy. While ensuring computational accuracy, it can reduce computational complexity and improve the efficiency of causal screening.

[0049] In this application, the lagged sequences of feature factors can refer to the multi-order lagged historical sequences of each feature factor in the candidate feature pool, such as the first-order lagged sequence and second-order lagged sequence of the feature factor temperature. The future sequence of the target variable can refer to the data sequence of the target variable (electricity load, photovoltaic output, wind power output) over a period of time after the current moment. For example, taking the current moment as t, the future sequence of the target variable can be the target variable data at times t+1, t+2, ..., t+n (where n is a preset future time step). Calculating the nonlinear correlation between the lagged sequences of each feature factor and the future sequence of the target variable essentially analyzes the degree of influence of the feature factor data in the current and past periods on the target variable data in the future period. For example, calculating the nonlinear correlation between the lagged sequence of the temperature feature factor and the future sequence of the electricity load target variable; if the correlation is high, it indicates that the historical changes in temperature have a significant impact on the future changes in electricity load.

[0050] In this application, positive driving force refers to the strength of the causal influence of the characteristic factor on the target variable, such as the causal strength of the temperature characteristic factor on the target variable of power load, that is, to what extent temperature change can drive power load change. Reverse feedback force refers to the strength of the causal influence of the target variable on the characteristic factor, such as whether the change of power load will affect the temperature (in real-world scenarios, this reverse influence is usually weak, but this application can identify causal relationships more comprehensively and accurately through bidirectional calculation).

[0051] In this application, the two-way net causal screening mechanism refers to comprehensively considering the positive driving force of the characteristic factor on the target variable and the negative feedback force of the target variable on the characteristic factor, and obtaining an index (i.e., net causal score) that can truly reflect the degree of net causal influence of the characteristic factor on the target variable through a specific calculation method.

[0052] For example, the net causal score can be calculated using the formula: Net Causal Score = Positive Driving Force - λ × Reverse Feedback Force, where λ is the feedback penalty coefficient, which can be adjusted according to the actual situation (e.g., λ = 0.5). This formula means that a certain proportion of the influence of the reverse feedback force is deducted from the positive driving force to obtain the strength of the pure causal influence of the feature factor on the target variable.

[0053] In this application, the preset scoring threshold is a critical value pre-set based on actual application scenarios and requirements. It is used to determine whether a feature factor is a key feature factor. For example, if the scoring threshold is set to 0.6, and the net causality score of a feature factor is 0.7, which is higher than the threshold, then it is identified as a key feature factor. If the net causality score of a feature factor is 0.4, which is lower than the threshold, then it is rejected. The setting of the preset scoring threshold can be based on a large amount of experimental data and engineering experience to ensure that the selected key feature factors can accurately and effectively reflect the impact on the operating status of the source load.

[0054] In this application, the causal strength score is a quantitative representation of the degree to which key feature factors affect the operating state of source loads. The higher the score, the more significant the influence of the feature factor on the operating state of source loads, and the more important it is to focus on in the subsequent model construction process. For example, the net causal score of the temperature feature factor is 0.8, and the net causal score of the wind speed feature factor is 0.75, indicating that the influence of temperature on the operating state of source loads is slightly higher than that of wind speed.

[0055] Next, this application will describe some specific embodiments for implementing steps 121 to 123 above: In this application, according to the definition of generalized Granger causality, if the variable Historical information helps reduce variables The uncertainty of the future state is called yes The Granger causality. Within the information theory framework, this reduction in uncertainty corresponds to a decrease in information entropy. Therefore, characteristic variables can be defined. For target variable The transfer entropy for: in, Indicates from variable to variable The transition entropy is used to measure the known After reviewing historical information, The amount by which uncertainty in predicting future states is reduced; This represents Shannon Entropy. Indicates the time lag order; Indicates that in the known under conditions The conditional probability; Indicates that in the known and under conditions The conditional probability.

[0056] This application uses mutual information (MI) as an effective approximation of transfer entropy (ProxyTE) to quantify causal strength by calculating the nonlinear correlation between the feature lag sequence and the target future sequence.

[0057] Variable entropy and joint entropy calculation assume characteristic sequence Lag The observation vector at time is Its probability distribution is ; target sequence exist The observation vector at time is Its probability distribution is Then their respective information entropy is defined as: in, Representing variables respectively and Edge information entropy. and Let represent the probability distributions of the feature sequence and the target sequence, respectively, and their joint probability distribution is: The joint entropy is defined as: in, Representation of features For the target The approximate causal strength is obtained. Based on equations (2)-(4), the feature is derived. For the target Approximate causal strength Calculation formula: in, Representation of features For the target The net causal score of this application in the lag interval Inner traversal calculation Select the one that maximizes the causal strength. This serves as the optimal lag order for the feature relative to the target.

[0058] In this application, a bidirectional net causality screening mechanism is designed, and ensemble fusion is performed for multi-task prediction scenarios. Features are defined. For the target Net causal score for: in, This represents the feedback penalty coefficient. The bidirectional net causality score is calculated for any candidate feature. and target variable ( ), not only calculating the positive driving force Simultaneously calculate the reverse feedback force .

[0059] Multi-target feature set fusion is performed on three target variables respectively. Perform filtering. Set a threshold. : in, Indicates the target The highest-quality features constitute a subset. This is the final model input feature set. It is the union of three subsets. For example... Figure 2 The diagram shows the verification of the causal screening effect in the embodiments of this application.

[0060] Based on the technical solutions in steps 121 to 123 above, by using mutual information as an effective approximation of the transfer entropy, the computational complexity can be reduced and the efficiency of causal screening can be improved while ensuring the accuracy of causal relationship identification, making this method applicable to large-scale multi-source data processing scenarios. By calculating the positive driving force of feature factors on the target variable and the negative feedback force of the target variable on the feature factors, a bidirectional net causal screening mechanism can be adopted to comprehensively and accurately identify the causal relationship between feature factors and the target variable, avoiding the misjudgment of causal relationships caused by traditional one-way causal analysis. By screening key feature factors through preset score thresholds, irrelevant or less influential feature factors can be accurately eliminated, while retaining key features that significantly affect the source load's operating state, reducing the computational load of subsequent models and improving the training efficiency and generalization ability of the model. The net causal score of the key feature factors is determined as the causal strength score, providing a quantitative basis for the initialization of the input gate weights of the subsequent deep causal gated recurrent memory network, enabling the model to allocate different attention weights according to the importance of feature factors, further improving the model's inference accuracy.

[0061] Continue to refer to Figure 1 In step 130, a deep causal gated recurrent memory network is constructed, and the input gate weights of the deep causal gated recurrent memory network are non-randomized based on the key feature factors and their corresponding causal strength scores to construct a time-series inference model.

[0062] In this application, it should be noted that the Deep Causal Gated Recurrent Memory Network (DC-GRMN) is a model architecture that integrates the advantages of causal relationship learning and deep neural networks. Based on the traditional gated recurrent memory network, it introduces a causal transfer learning mechanism, which can better utilize the causal information in the data to improve model performance.

[0063] In this application, it should also be noted that the input gate weights are important parameters in deep causal gated recurrent memory networks that control the input information entering the network memory units. They are initialized non-randomly by the causal strength scores corresponding to the key feature factors, that is, the input gate weights corresponding to the key feature factors with higher causal strength scores are set to relatively large values. This allows the model to focus on these key feature factors in the early stages of training, which is equivalent to injecting prior causal knowledge into the model. This can avoid the problems of low model training efficiency and easy getting trapped in local optima caused by traditional random initialization.

[0064] Specifically, in such Figure 1 In step 130, the construction of the deep causal gated recurrent memory network and the non-random initialization of the input gate weights of the deep causal gated recurrent memory network based on the key feature factors and their corresponding causal strength scores can be performed according to steps 131 to 132 as follows: Step 131: Construct a bidirectional temporal memory evolution layer, and use a bidirectional long short-term memory network as the main inference backbone. Perform temporal modeling on the key feature factors through forward and reverse recursive chains to construct the deep causal gated recurrent memory network.

[0065] Step 132: Map the causal strength score corresponding to the key feature factor to a weight gain matrix, and apply it to the input gate weight matrix of the first layer of the bidirectional long short-term memory network to achieve non-randomized initialization of the initial weights of the input gate.

[0066] In this application, the bidirectional temporal memory evolution layer is a core component of the deep causal gated recurrent memory network, which can process data from both the forward and reverse directions of the time series simultaneously, capturing the dependencies of data in different time dimensions.

[0067] In this application, a forward recursive chain refers to a chain that processes key feature factor data in chronological order from past to future, such as the order from time t-1, time t to time t+1, which can capture the forward temporal dependencies of the data. A backward recursive chain refers to a chain that processes key feature factor data in chronological order from future to past, such as the order from time t+1, time t to time t-1, which can capture the backward temporal dependencies of the data.

[0068] In this application, the Bi-LSTM network is a neural network model developed based on the traditional LSTM network. It possesses two independent LSTM structures, corresponding to the forward and reverse recursive chains respectively, enabling it to process both forward and reverse time-series data simultaneously. Compared to traditional LSTM, it can more comprehensively capture the contextual information and dependencies in time-series data. Using the Bi-LSTM network as the inference backbone, a deep causal gated recurrent memory network is constructed, which can fully utilize its powerful time-series modeling capabilities to accurately learn the dynamic causal relationship between key feature factors and the source load's operating state.

[0069] In this application, the weight gain matrix is ​​a matrix constructed based on the causal strength scores of key feature factors. Its dimension is consistent with the dimension of the input gate weight matrix of the first layer of the bidirectional long short-term memory network. Each element value in the matrix is ​​related to the causal strength score of the corresponding key feature factor. The higher the causal strength score, the larger the corresponding weight gain matrix element value.

[0070] For example, assuming there are three key feature factors with causal strength scores of 0.8, 0.7, and 0.6, the constructed weight gain matrix can be [[1.8],[1.7],[1.6]] (this is just an example; the actual matrix construction needs to be combined with the network structure and parameter settings). This weight gain matrix is ​​then applied to the input gate weight matrix, that is, each element in the input gate weight matrix is ​​multiplied by the corresponding element in the weight gain matrix to obtain a new input gate weight matrix.

[0071] In this application, non-randomized initialization indicates that the input gate weights are not randomly generated, but rather set based on the causal strength scores of key feature factors. This results in higher input gate weights for key feature factors with higher causal strength scores. Consequently, in the early stages of model training, the input gates give more attention to these key feature factors, enabling the model to capture important causal information more quickly. This avoids the problems of slow convergence and susceptibility to local optima in the early stages of model training caused by traditional random initialization.

[0072] Next, this application will describe some specific embodiments for implementing steps 131 to 132 above: In this application, the purpose of constructing a bidirectional temporal memory evolution layer is to accurately capture the positive physical inertia and negative smoothness dependency of the system state, and a bidirectional long short-term memory network is used as the backbone of the inference engine. The input tensor filtered in step 120 above is set as... Bidirectional Long Short-Term Memory (LSTM) networks process data through two recursive chains: forward and backward. in, This indicates that the input feature tensor after causal filtering is in The value at time; and These represent the hidden state vectors generated by the forward recursive chain and the reverse recursive chain, respectively. express Each time step contains a hidden state vector with complete contextual temporal logic. This structure directly models the original causal features temporally, ensuring the lossless transmission of physical information.

[0073] To address the cold start problem of deep models in small sample planning scenarios, this application utilizes the causal strength score extracted in step 120 above to perform non-randomized initialization of the input gate weights of the first layer of the bidirectional long short-term memory network.

[0074] Physical Mechanism: Input Gates of LSTM It determines how much of the input features at the current moment can enter the memory unit.

[0075] Weight injection: The feature causal scoring vector determined in step 120 above is weighted... Mapped to a weight gain matrix, which directly acts on the input weight matrix. : in, Indicates the index of the input feature dimension; Indicates the basic random weights; This represents the bidirectional net causality score. The initial weights of the input gate corresponding to strong causal features are explicitly amplified, enabling the model to prioritize key driving factors from the initial training phase. For example... Figure 3 The diagram shows a comparison of model training convergence in the embodiments of this application.

[0076] Based on the technical solutions in steps 131 to 132 above, by constructing a bidirectional temporal memory evolution layer, using forward and reverse recursive chains to perform temporal modeling of key feature factors, and using a bidirectional long short-term memory network as the inference backbone, the powerful temporal modeling capabilities of bidirectional LSTM can be fully utilized to comprehensively capture the forward and backward dependencies of key feature factors in the time dimension, accurately learn the dynamic causal relationship between key feature factors and the source load operating state, and provide reliable model architecture support for the accurate inference of the source load operating state. Mapping the causal strength scores of key feature factors to weight gain matrices and performing non-random initialization of the input gate weights of the first layer of the bidirectional long short-term memory network can effectively inject causal prior knowledge into the neural network model, enabling the model to focus on key feature factors that significantly affect the source load operating state in the early stages of training, accelerating the model's training convergence speed and improving training efficiency. Simultaneously, this non-random initialization method also helps improve the model's generalization ability and inference accuracy, avoiding the model instability problems caused by traditional random initialization, laying a solid foundation for the subsequent accurate inference of the source load operating state.

[0077] Furthermore, in this application, during the process of constructing the time series extrapolation model, the following steps 133 to 134 may also be performed: Step 133: Introduce an attention mechanism to dynamically weight the output state of the bidirectional long short-term memory network.

[0078] Step 134: Construct a lightweight gated recurrent unit network as a residual model. Learn the residual sequence between the output of the bidirectional long short-term memory network and the true value through the residual model. Superimpose the residual prediction value of the residual model onto the output of the bidirectional long short-term memory network to obtain a high-fidelity intermediate result in the process of inferring the source load operation state. Use the network model corresponding to the intermediate result as the time series inference model.

[0079] In this application, it should be noted that the attention mechanism is a technique that simulates the human attention allocation mechanism. Its core idea is to assign different weights based on the importance of input data, giving more attention to important information and less attention to unimportant information. In this application, the output state of the bidirectional long short-term memory network contains temporal information of key feature factors at different time steps, and the importance of this information to the source load operation state prediction varies. By introducing the attention mechanism, the attention weight of the output state at each time step can be calculated. This weight reflects the importance of the information at the corresponding time step to the source load operation state prediction. For example, when predicting the power load during the summer peak electricity consumption period, the output state information of time steps adjacent to the peak consumption period (such as 1-2 hours prior) is more important, and the corresponding attention weight will be larger. Conversely, the output state information of time steps further away (such as the previous day) is less important, and the corresponding attention weight will be smaller. Then, the output state of the bidirectional long short-term memory network is dynamically weighted and summed according to the attention weights to obtain a feature vector that integrates important temporal information. This feature vector can more accurately reflect the influence of key feature factors on the source load operation state.

[0080] In this application, it should be noted that the Lightweight Gated Recurrent Unit (GRU) network is a recurrent neural network model with a relatively simple structure and high computational efficiency. While retaining the core functions of the gated recurrent memory network, it reduces the number of model parameters, thereby reducing computational complexity and making it suitable as a residual model for learning residual sequences.

[0081] In this application, the residual sequence refers to the sequence of differences between the output of the bidirectional long short-term memory network (BSSM) and the actual operating state of the source load. For example, if the BSSM predicts a power load of 4000MW at a certain moment, and the actual power load at that moment is 4050MW, then the residual is 50MW. The residuals from multiple moments constitute the residual sequence. The core function of the residual model is to learn and fit this residual sequence. Through training, the residual model can gradually grasp the deviation pattern of the BSSM prediction results and output the corresponding residual prediction values. The residual prediction values ​​of the residual model are superimposed on the output of the BSSM, thus correcting the prediction results of the BSSM. For example, if the BSSM prediction value is 4500MW and the residual prediction value of the residual model is 45MW, then the corrected result is 4545MW, which is closer to the actual value of 4050MW. By using this dual-model architecture of a main model (bidirectional long short-term memory network) and a residual model (lightweight GRU network), prediction errors can be effectively reduced and high-fidelity intermediate results can be obtained. The network model corresponding to the intermediate result (a network model that integrates attention mechanism, bidirectional LSTM and residual GRU) can be used as the final time series inference model, which can significantly improve its inference accuracy and reliability.

[0082] Next, this application will describe some specific embodiments for implementing steps 133 to 134 above: In this application, to address the issue of varying importance of features at different time steps, an attention mechanism is introduced to dynamically weight the output state of the bidirectional long short-term memory network, thereby identifying the key historical moments that have the greatest impact on future states.

[0083] Correlation score calculation: Normalized weight allocation and context aggregation: in, Used to measure history The extent to which time affects the future This represents the normalized attention weights.

[0084] In this application, a two-stage strategy of master model and residual model is adopted to eliminate the influence of random noise on the planning boundary. A lightweight GRU network is constructed to specifically learn the residual sequence of the master model. The residual prediction values ​​are then superimposed onto the main model output to obtain the final high-fidelity inference trajectory.

[0085] Based on the technical solutions in steps 133 to 134 above, by introducing an attention mechanism to dynamically weight the output state of the bidirectional long short-term memory network, corresponding attention weights can be assigned according to the importance of information at different time steps to the deduction of the source load's operating state. This strengthens the influence of key time step information and weakens the interference of irrelevant or secondary time step information, enabling the model to more accurately capture the temporal information that plays a decisive role in the source load's operating state. This further improves the model's efficiency in utilizing key information and its deduction accuracy. A lightweight gated recurrent unit network is constructed as a residual model. By learning the residual sequence between the output of the main model (bidirectional long short-term memory network) and the true value through the residual model, the prediction bias of the main model can be effectively fitted. The residual prediction values ​​are then used to correct the output of the main model, significantly reducing the overall prediction error and obtaining high-fidelity intermediate deduction results. This dual-model architecture of master model and residual model, combined with the application of attention mechanism, enables the final time series extrapolation model to have stronger learning and generalization capabilities. It can more accurately capture the changing patterns of source load operation status, providing stronger model support for the accurate extrapolation of source load operation status trajectory within future planning cycles, and further improving the practicality and reliability of the entire technical solution.

[0086] Continue to refer to Figure 1In step 140, based on the time-series simulation model, the source-load operation trajectory of the power distribution system in the future planning period is dynamically simulated.

[0087] In this application, the future planning period refers to the future time frame targeted by the power distribution system planning, such as the next 5 years or 10 years, with the specific duration determined based on actual planning needs. The source-load operating status trajectory refers to the trend and path of changes in the operating status of the power load, photovoltaic output, and wind power output of the power distribution system over time within the future planning period. Dynamically extrapolating the source-load operating status within the future planning period using a time-series model can provide comprehensive and accurate predictive information on the source-load operating status for power distribution system planning, such as predicting the peak power load during the summer peak electricity consumption period each year for the next 5 years, and the seasonal variations in photovoltaic and wind power output.

[0088] Specifically, in such Figure 1 In step 140, the construction of the deep causal gated recurrent memory network and the non-randomization initialization of the input gate weights of the deep causal gated recurrent memory network based on the key feature factors and their corresponding causal strength scores can be performed according to steps 141 to 144 as follows: Step 141: Construct a multi-task quantile loss function and use an asymmetric bouncing loss to guide the time series extrapolation model to learn the conditional probability distribution of the target variable.

[0089] Step 142, introduce the following into the multi-task quantile loss function Regularization terms are used to construct the objective function.

[0090] Step 143: Minimize the objective function using the gradient descent algorithm, train the time series extrapolation model to output the source load operation state prediction results corresponding to the 10th quantile, 50th quantile and 90th quantile.

[0091] Step 144: Based on the source load operation state prediction results, determine the source load operation state trajectories under conservative, neutral and aggressive scenarios respectively, forming a set of source load operation state trajectories covering the full probability space within the future planning period.

[0092] In this application, it should be noted that the multi-task quantile loss function is a loss function that can handle multiple prediction tasks simultaneously. In this application, the prediction tasks include three tasks: power load prediction, photovoltaic power output prediction, and wind power output prediction. This loss function can comprehensively consider and optimize the prediction errors of these three tasks.

[0093] In this application, it should be noted that the asymmetric pinball loss is a loss function used for quantile regression. Its characteristic is that it asymmetrically penalizes the deviation between the predicted and true values; that is, when the predicted value is lower than the true value (underestimation), the penalty coefficient is equal to the quantile value. When the predicted value is higher than the actual value (overestimation), the penalty coefficient is... This asymmetric penalty mechanism guides the model to learn the conditional probability distribution of the target variable at different quantiles.

[0094] For example, when When the ball loss is high, the penalty coefficient for overestimation (1-0.1=0.9) is much greater than the penalty coefficient for underestimation (0.1), which guides the model to learn the lower bound (conservative estimate) of the target variable. When the penalty coefficient for underestimating the ball loss (0.9) is much larger than the penalty coefficient for overestimating it (0.1), it will guide the model to learn the upper bound of the target variable (aggressive estimation). When the penalty coefficients are symmetrical (both are 0.5), they guide the model to learn the median (neutral estimate) of the target variable.

[0095] In this application, Regularization is a regularization technique used to prevent model overfitting. It constrains the magnitude of model parameters by adding a sum-of-squares term to the loss function, preventing them from becoming excessively large and thus avoiding overfitting the training data, thereby improving the model's generalization ability. This is achieved by introducing... Regularization terms can effectively balance the model's fitting ability and generalization ability, ensuring that the model has a good fit on the training data while maintaining good predictive performance on unseen test data.

[0096] In this application, gradient descent is a commonly used machine learning optimization algorithm. Its core idea is to continuously adjust the model parameters along the gradient direction of the objective function to find the minimum point of the objective function. In this application, the parameters of the time-series model are iteratively updated using the gradient descent algorithm, causing the value of the objective function to gradually decrease until it converges to a preset threshold or reaches the maximum number of iterations.

[0097] In this application, after model training, the model can output source load operation status prediction results corresponding to three key quantiles. Specifically, the prediction result corresponding to the 10th quantile indicates that there is a 10% probability that the source load operation status will be lower than this value within the future planning period, representing a conservative scenario. The prediction result corresponding to the 50th quantile indicates that there is a 50% probability that the source load operation status will be lower than this value within the future planning period, representing a neutral scenario. The prediction result corresponding to the 90th quantile indicates that there is a 90% probability that the source load operation status will be lower than this value within the future planning period, representing an aggressive scenario.

[0098] For example, if a model predicts the power load for a certain future time period, and the 10th percentile prediction is 3500MW, the 50th percentile prediction is 4000MW, and the 90th percentile prediction is 5000MW, then it means that there is a 10% probability that the power load will be lower than 3500MW, a 50% probability that it will be lower than 4000MW, and a 90% probability that it will be lower than 5000MW during that time period.

[0099] In this application, the source-load operating state trajectory refers to the curve or path showing the change of the source-load operating state (electricity load, photovoltaic output, and wind power output) over time within a future planning period. Based on the prediction results corresponding to the 10th, 50th, and 90th quantiles, source-load operating state trajectories under conservative, neutral, and aggressive scenarios can be plotted respectively. For example, the electricity load trajectory under the conservative scenario is generally at a low level, the electricity load trajectory under the aggressive scenario is generally at a high level, and the electricity load trajectory under the neutral scenario is somewhere in between. These three trajectories together constitute a set of source-load operating state trajectories covering the full probability space within the future planning period, which can comprehensively reflect the uncertainty range of future source-load operating states.

[0100] Next, this application will describe some specific embodiments for implementing steps 141 to 144 above: In this application, in order to quantify the uncertainties on both the source and load sides of the power distribution system, the model can be designed as a multi-task, multi-output architecture to construct a multi-task probability prediction output space.

[0101] Define the output tensor of the model as ,in Indicates batch size. Task dimension. (Corresponding to the load, photovoltaic output, and wind power output of the power distribution system, respectively). Quantile dimension : These correspond to the 10th percentile of the probability distribution ( ), 50th percentile ( ), 90th percentile ( ).

[0102] Asymmetric pinball loss can be used to guide the model to learn the conditional probability distribution of the target variable.

[0103] For any objective variable At any moment The true value and quantiles Predicted values Define the loss function for: in, The asymmetric penalty of the function is applied when the true value is underestimated ( When ), the penalty coefficient is When the true value is overestimated ( When ), the penalty coefficient is Through this mechanism, The predicted values ​​will be pushed towards the upper bound of the true value distribution, while The predicted values ​​will be pushed to the lower bound of the distribution.

[0104] To collaboratively optimize multiple tasks and prevent a single task from dominating the gradient, a global comprehensive loss function is defined. For all time steps All tasks and the set of all quantiles Total losses: in, Represents the global comprehensive loss function; Indicates the time step index; Indicates the task dimension index; Represents the set of quantiles.

[0105] The goal of model training is to minimize the gradient descent algorithm. This forces the model output to cover the probability range of the true value.

[0106] Regularization Optimization and Model Training Strategy: To prevent overfitting of the deep model on limited historical data of the distribution network and to ensure the effective representation of causal features, this embodiment adopts an optimization strategy with weighted decay. Regularization is introduced into the loss function. Regularization terms are used to construct the final objective function: in, This represents all trainable parameters of the network. This represents the attenuation coefficient. This term constrains the norm of the network weights, forcing the model to learn a smoother mapping relationship and avoiding overfitting to noisy data.

[0107] Based on the trained model, the output quantile sequence is mapped to three typical scenarios for power distribution system planning and operation to aid decision-making. These three typical scenarios are: conservative scenario (…). ), neutral scene ( radical scenarios Through these three typical scenarios, this application constructs an 80% confidence interval for the target variable at future times (i.e., This represents a leap from single numerical prediction to probabilistic interval prediction, providing quantitative data support for robust planning of power distribution systems. For example... Figure 4 The diagram shows the effect of multi-task probability prediction and scene generation in an embodiment of this application.

[0108] Based on the technical solutions in steps 141 to 144 above, by constructing a multi-task quantile loss function and using asymmetric bouncing loss to guide the model in learning the conditional probability distribution of the target variable, the uncertainty of the source load operating state can be effectively handled, achieving probabilistic prediction of the source load operating state. Compared with traditional deterministic prediction methods, it provides more comprehensive and richer prediction information, better supporting risk decision-making in power distribution systems. Introducing an L2 regularization term into the multi-task quantile loss function to construct the objective function effectively prevents model overfitting, improves the model's generalization ability, and ensures stable and reliable performance in predicting the source load operating state within future planning periods. By minimizing the objective function using the gradient descent algorithm, the trained model outputs prediction results corresponding to the 10%, 50%, and 90% quantiles, corresponding to conservative, neutral, and aggressive scenarios respectively, comprehensively covering the uncertainty range of future source load operating states. Based on the prediction results of these three scenarios, the corresponding source and load operating state trajectories are determined, forming a trajectory set covering the entire probability space. This can provide comprehensive and accurate basic data for the boundary stress test and robustness verification of subsequent power distribution system planning schemes, which helps to achieve the scientific and rational planning of power distribution systems.

[0109] In this application, after step 144 above, that is, after forming a set of source load operating state trajectories covering the full probability space within the future planning period, steps 145 to 146 can also be performed: Step 145: Based on the source and load operating state trajectories under conservative, neutral, and aggressive scenarios in the source and load operating state trajectory set, conduct boundary pressure tests on the power supply adequacy and absorption capacity of the current planning scheme of the power distribution system.

[0110] Step 146: Quantitatively verify the robustness of the current planning scheme based on the boundary pressure test results, and adjust the planning parameters of the current planning scheme based on the verification results, so as to achieve dynamic coordination of the integrated planning and operation of the power distribution system.

[0111] In this application, the current planning scheme refers to the planning scheme that is currently being implemented or is planned to be implemented in the power distribution system, including power facility construction planning (such as new substation construction and line expansion), new energy access planning, load distribution planning, etc.

[0112] In this application, power supply adequacy refers to the extent to which the power distribution system can meet the electricity demand of users within a certain period of time, and its measurement indicators include power supply reliability, voltage qualification rate, and power outage time. Absorption capacity refers to the ability of the power distribution system to accept and absorb the output of new energy sources (photovoltaic and wind power), and its measurement indicators include new energy curtailment rate, line transmission capacity utilization rate, and transformer load rate.

[0113] In this application, boundary stress testing refers to placing the current planning scheme under extreme or boundary conditions to test its ability to operate normally and meet relevant performance requirements. Specifically, boundary stress testing in this application involves using the source-load operating trajectory under conservative, neutral, and aggressive scenarios as input conditions to simulate the operation of the power distribution system under these three scenarios, calculating relevant indicators of power supply adequacy and absorption capacity. For example, in an aggressive scenario, photovoltaic and wind power output is higher, while the power load is lower in a conservative scenario; this tests whether the current planning scheme can effectively absorb excess renewable energy power and whether there will be an excessively high renewable energy curtailment rate. Similarly, in a conservative scenario, photovoltaic and wind power output is lower, while the power load is higher in an aggressive scenario; this tests whether the current planning scheme can meet the power supply demand during peak loads and whether indicators such as power supply reliability and voltage qualification rate meet standards. Furthermore, in a neutral scenario, this tests the comprehensive performance of the current planning scheme under normal operating conditions.

[0114] In this application, robustness refers to the ability of the current planning scheme to maintain stable performance indicators when faced with uncertain changes in the operating status of the source load. The stronger the robustness, the stronger the adaptability of the current planning scheme to uncertainty.

[0115] In this application, the robustness of the current planning scheme is quantitatively verified. Specifically, based on the power supply adequacy and absorption capacity indicators obtained from boundary stress tests, a specific evaluation method (such as weighted scoring or analytic hierarchy process) is used to quantitatively score the robustness of the current planning scheme. For example, the weight of power supply adequacy is set to 0.6, and the weight of absorption capacity is set to 0.4. Scores are assigned to the power supply adequacy and absorption capacity indicators under three different scenarios (out of 100). Then, a weighted average score is calculated as the robustness score of the current planning scheme; the higher the score, the stronger the robustness.

[0116] In this application, adjusting the planning parameters of the current planning scheme refers to modifying and optimizing the relevant planning parameters based on the robustness verification results, addressing any deficiencies or problems in the current planning scheme. For example, if insufficient power supply sufficiency is found in the boundary stress test under an aggressive scenario, the planning parameters can be adjusted to increase the substation construction capacity or line transmission capacity. If insufficient absorption capacity is found in a conservative scenario, the planning parameters for renewable energy access can be adjusted to optimize the allocation method of renewable energy output or increase the configuration capacity of energy storage equipment.

[0117] In this application, the dynamic coordination of integrated planning and operation of power distribution systems refers to breaking down the traditional separation between planning and operation, closely combining the results of source-load operation status projections with adjustments to the planning scheme. This enables the planning scheme to be dynamically optimized based on changes in operation status, while the planning scheme can also guide the formulation of operational strategies, forming a closed-loop coordination mechanism of operation projection, planning verification, feedback optimization, and operational guidance. For example, by projecting the peak power load under aggressive future scenarios through source-load operation status projections, the planning scheme can be adjusted accordingly to increase power supply capacity and ensure that the load demand can be met during operation. Simultaneously, based on the adjusted planning scheme, corresponding operational strategies (such as optimizing unit scheduling and rationally allocating load) are formulated, achieving dynamic coordination between planning and operation.

[0118] Based on the technical solutions in steps 145 and 146 above, and considering the source-load operating trajectories under conservative, neutral, and aggressive scenarios, boundary stress tests are conducted on the power supply adequacy and absorption capacity of the current planning scheme. This comprehensively and systematically verifies the performance of the planning scheme under different operating conditions, especially its adaptability in extreme scenarios, avoiding the incomplete considerations caused by traditional planning schemes that only test in typical scenarios. The robustness of the planning scheme is quantitatively verified based on the boundary stress test results, enabling an objective and accurate evaluation of its merits and demerits, providing a clear basis for adjustment and optimization. Adjusting the planning parameters based on the verification results allows for dynamic coordination of distribution system planning and operation, breaking down the traditional separation between planning and operation. This enables the planning scheme to be dynamically optimized based on the uncertainties in source-load operating states, improving its scientific rigor, rationality, and adaptability. It effectively avoids excessive power redundancy or insufficient safety margins in the planning scheme, ensuring the stable, efficient, and economical operation of the distribution system throughout the future planning cycle.

[0119] Overall, based on the technical solution proposed in this application, by integrating multi-source historical data from the power distribution system to construct a candidate feature pool, the comprehensiveness and richness of the data can be ensured, providing a solid data foundation for subsequent causal screening and model construction. Nonlinear causal screening based on approximate transfer entropy can accurately identify key feature factors affecting the operating state of source loads, eliminate redundant information, reduce model computational complexity, and improve the model's ability to capture key information. Using the causal strength scores of key feature factors to perform non-randomized initialization of the input gate weights of the deep causal gated recurrent memory network, a time-series extrapolation model is constructed. This effectively integrates causal prior knowledge with the deep learning model, improving the model's training efficiency and generalization ability, and avoiding the problems of overfitting and slow convergence of traditional models in small sample or complex data scenarios. Based on this time-series extrapolation model, the trajectory of source load operating states within future planning cycles can be dynamically and accurately extrapolated, providing a reliable decision-making basis for power distribution system planning, breaking the disconnect between traditional planning methods and actual operating states, and laying the foundation for achieving dynamic coordination of integrated power distribution system planning and operation.

[0120] The following describes relevant equipment embodiments of this application, which can be used to execute the power distribution system source-load operation state deduction method based on causal transition in the above embodiments of this application. For details not disclosed in the relevant equipment embodiments of this application, please refer to the embodiments of the power distribution system source-load operation state deduction method based on causal transition described above in this application.

[0121] See Figure 5 The diagram shows a block diagram of the equipment related to the inference of the source-load operation status of the power distribution system based on causal migration in an embodiment of this application.

[0122] like Figure 5 As shown, the power distribution system source-load operation state inference related device 500 based on causal migration according to the embodiment of this application includes: a first construction unit 501, a screening unit 502, a second construction unit 503 and an inference unit 504.

[0123] The system comprises the following components: a first construction unit 501, used to construct a candidate feature pool based on historical multi-source operation data, meteorological data, and economic data of the power distribution system, including power load data, photovoltaic output data, and wind power output data; a screening unit 502, used to perform nonlinear causal screening on the feature factors in the candidate feature pool based on approximate transfer entropy to determine the key feature factors affecting the source-load operation status and their corresponding causal strength scores; a second construction unit 503, used to construct a deep causal gated recurrent memory network and to perform non-randomized initialization of the input gate weights of the deep causal gated recurrent memory network based on the key feature factors and their corresponding causal strength scores to construct a time-series extrapolation model; and an extrapolation unit 504, used to dynamically extrapolate the source-load operation status trajectory of the power distribution system within a future planning period based on the time-series extrapolation model.

[0124] In some embodiments of this application, based on the foregoing scheme, the first construction unit 501 is configured to: obtain multi-source operation data, meteorological data, and economic data of the power distribution system at historical moments from the power distribution system database as raw data; use the raw data and its corresponding multi-order lag historical sequence as basic features, and use the pre-released power load forecast data of the power system as prior features; integrate the basic features, the prior features, and the historical observations of the target variable to be predicted to obtain a candidate feature pool containing external driving factors and internal autoregressive factors, wherein the target variable includes power load, photovoltaic power output, and wind power output.

[0125] In some embodiments of this application, based on the foregoing scheme, the screening unit 502 is configured to: use mutual information as an effective approximation of the approximate transfer entropy; calculate the nonlinear correlation between the lagged sequences of each feature factor in the candidate feature pool and the future sequences of the target variable; quantify the causal strength of each feature factor on the target variable as a positive driving force, and quantify the causal strength of the target variable on each feature factor as a negative feedback force; based on the positive driving force and the negative feedback force, use a bidirectional net causal screening mechanism to calculate the net causal score of each feature factor, and screen feature factors whose net causal scores are higher than a preset score threshold from the candidate feature pool as key feature factors; and determine the net causal score of the key feature factor as the causal strength score of the key feature factor.

[0126] In some embodiments of this application, based on the aforementioned scheme, the second construction unit 503 is configured to: construct a bidirectional temporal memory evolution layer, and use a bidirectional long short-term memory network as the inference backbone, perform temporal modeling of the key feature factors through forward recursive chains and reverse recursive chains, and construct the deep causal gated recurrent memory network; map the causal intensity scores corresponding to the key feature factors into weight gain matrices, and apply them to the input gate weight matrix of the first layer of the bidirectional long short-term memory network to achieve non-randomized initialization of the initial weights of the input gates.

[0127] In some embodiments of this application, based on the foregoing scheme, the second construction unit 503 is further configured to: introduce an attention mechanism to dynamically weight the output state of the bidirectional long short-term memory network; construct a lightweight gated recurrent unit network as a residual model; learn the residual sequence between the output result of the bidirectional long short-term memory network and the true value through the residual model; superimpose the residual prediction value of the residual model onto the output result of the bidirectional long short-term memory network to obtain a high-fidelity intermediate result in the process of inferring the source load operation state; and use the network model corresponding to the intermediate result as a time series inference model.

[0128] In some embodiments of this application, based on the foregoing scheme, the inference unit 504 is configured to: construct a multi-task quantile loss function, and use an asymmetric bouncing loss to guide the time-series inference model to learn the conditional probability distribution of the target variable; and introduce into the multi-task quantile loss function... Regularization terms are used to construct the objective function; the objective function is minimized using the gradient descent algorithm, and the time series extrapolation model is trained to output the source load operation state prediction results corresponding to the 10th quantile, 50th quantile, and 90th quantile; based on the source load operation state prediction results, the source load operation state trajectories under conservative, neutral, and aggressive scenarios are determined respectively, forming a set of source load operation state trajectories covering the full probability space within the future planning period.

[0129] In some embodiments of this application, based on the foregoing scheme, the related equipment further includes: a testing unit, used to perform boundary pressure tests on the power supply adequacy and absorption capacity of the current planning scheme of the power distribution system based on the source and load operating state trajectories under conservative, neutral, and aggressive scenarios in the source and load operating state trajectory set after forming a set of source and load operating state trajectories covering the full probability space within the future planning period; and an adjustment unit, used to quantify and verify the robustness of the current planning scheme based on the boundary pressure test results, and adjust the planning parameters of the current planning scheme based on the verification results, so as to achieve dynamic coordination of the integrated planning and operation of the power distribution system.

[0130] Based on the same inventive concept, embodiments of this application provide a computer program product, the computer program product including computer instructions stored in a computer-readable storage medium and adapted to be read and executed by a processor, so as to cause a computer device having the processor to perform the operations performed by the power distribution system source-load operation state inference method based on causal migration as described above.

[0131] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing at least one computer program instruction, which is loaded and executed by a processor to perform the operations described above in the power distribution system source-load operation state deduction method based on causal transition.

[0132] Based on the same inventive concept, this application also provides an electronic device, see reference. Figure 6 The diagram shows a schematic of the structure of an electronic device in an embodiment of this application. The electronic device includes one or more memories 604, one or more processors 602, and at least one computer program (computer program instruction) stored in the memory 604 and executable on the processor 602. When the processor 602 executes the computer program, it implements the source-load operation state deduction method of the power distribution system based on causal migration as described above.

[0133] Among them, Figure 6 In this document, a bus architecture (represented by bus 600) is used. Bus 600 may include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 602 and memory represented by memory 604. Bus 600 may also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 605 provides an interface between bus 600 and receiver 601 and transmitter 603. Receiver 601 and transmitter 603 may be the same element, i.e., a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 602 is responsible for managing bus 600 and general processing, while memory 604 can be used to store data used by processor 602 during operation.

[0134] The functions described herein may be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored as one or more instructions or codes on or transmitted via a computer-readable medium. Other examples and embodiments are within the scope and spirit of this application and the appended claims. For example, due to the nature of software, the functions described above may be implemented using software executed by a processor, hardware, firmware, hardwired, or any combination thereof. Furthermore, the functional units may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit.

[0135] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The related device embodiments described above are merely illustrative. For example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection can be through some interfaces; the indirect coupling or communication connection between units or modules can be electrical or other forms.

[0136] The units described as separate components may or may not be physically separate. Similarly, the components of the control device may or may not be physical units; they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0137] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing computer program instructions, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0138] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for inferring the source-load operating state of a power distribution system based on causal transfer, characterized in that, The method includes: Based on historical multi-source operation data, meteorological data, and economic data of the power distribution system, a candidate feature pool is constructed. The multi-source operation data includes power load data, photovoltaic power output data, and wind power output data. Based on approximate transfer entropy, nonlinear causal screening is performed on the feature factors in the candidate feature pool to determine the key feature factors affecting the source load operation state and their corresponding causal strength scores. A deep causal gated recurrent memory network is constructed, and the input gate weights of the deep causal gated recurrent memory network are non-randomized based on the key feature factors and their corresponding causal strength scores to construct a time-series inference model. Based on the aforementioned time-series simulation model, the source-load operation trajectory of the power distribution system within the future planning period is dynamically simulated.

2. The method for inferring the source-load operating state of a power distribution system based on causal migration according to claim 1, characterized in that, The candidate feature pool is constructed based on historical multi-source operational data, meteorological data, and economic data of the power distribution system, including: The power distribution system's multi-source operational data, meteorological data, and economic data at historical moments are obtained from the power distribution system database as raw data. The original data and its corresponding multi-order lag historical sequence are used as basic features, and the pre-released power load forecast data of the power system are used as prior features. By integrating the basic features, the prior features, and the historical observations of the target variable to be predicted, a candidate feature pool containing external driving factors and internal autoregressive factors is obtained. The target variable includes power load, photovoltaic power output, and wind power output.

3. The method for inferring the source-load operating state of a power distribution system based on causal migration according to claim 2, characterized in that, The nonlinear causal screening of feature factors in the candidate feature pool based on approximate transfer entropy to determine the key feature factors affecting the operating state of the source load and their corresponding causal strength scores includes: Using mutual information as an effective approximation of the transfer entropy, the nonlinear correlation between the lag sequence of each feature factor in the candidate feature pool and the future sequence of the target variable is calculated. The causal strength of each feature factor on the target variable is quantified as a positive driving force, and the causal strength of the target variable on each feature factor is quantified as a negative feedback force. Based on the positive driving force and the negative feedback force, a bidirectional net causal screening mechanism is used to calculate the net causal score of each feature factor, and feature factors with net causal scores higher than a preset score threshold are selected from the candidate feature pool as key feature factors. The net causal score of the key feature factor is determined as the causal strength score of the key feature factor.

4. The method for inferring the source-load operating state of a power distribution system based on causal migration according to claim 3, characterized in that, The construction of a deep causal gated recurrent memory network, and the non-random initialization of the input gate weights of the deep causal gated recurrent memory network based on the key feature factors and their corresponding causal strength scores, includes: A bidirectional temporal memory evolution layer is constructed, and a bidirectional long short-term memory network is used as the main inference backbone. The key feature factors are temporally modeled through forward and reverse recursive chains to construct the deep causal gated recurrent memory network. The causal strength scores corresponding to the key feature factors are mapped to weight gain matrices and applied to the input gate weight matrix of the first layer of the bidirectional long short-term memory network to achieve non-randomized initialization of the input gate weights.

5. The method for inferring the source-load operating state of a power distribution system based on causal transition according to claim 4, characterized in that, The process of constructing the time series extrapolation model also includes: An attention mechanism is introduced to dynamically weight the output state of the bidirectional long short-term memory network; A lightweight gated recurrent unit network is constructed as a residual model. The residual sequence between the output of the bidirectional long short-term memory network and the true value is learned through the residual model. The residual prediction value of the residual model is superimposed on the output of the bidirectional long short-term memory network to obtain a high-fidelity intermediate result in the process of inferring the source load operation state. The network model corresponding to the intermediate result is used as the time series inference model.

6. The method for inferring the source-load operating state of a power distribution system based on causal migration according to claim 5, characterized in that, The dynamic simulation of the source-load operating state trajectory of the power distribution system within a future planning period based on the time-series simulation model includes: A multi-task quantile loss function is constructed, and an asymmetric bouncing loss is used to guide the time series extrapolation model to learn the conditional probability distribution of the target variable; Introduced in the multi-task quantile loss function Regularization terms are used to construct the objective function; The objective function is minimized using the gradient descent algorithm, and the time series extrapolation model is trained to output the source load operation status prediction results corresponding to the 10th quantile, 50th quantile, and 90th quantile. Based on the predicted source-load operation status, the source-load operation status trajectories under conservative, neutral, and aggressive scenarios are determined respectively, forming a set of source-load operation status trajectories covering the full probability space within the future planning period.

7. The method for inferring the source-load operating state of a power distribution system based on causal migration according to claim 6, characterized in that, After forming a set of source-load operating state trajectories covering the full probability space within the future planning period, the method further includes: Based on the source and load operating state trajectories under conservative, neutral, and aggressive scenarios in the source and load operating state trajectory set, boundary pressure tests are conducted on the power supply adequacy and absorption capacity of the current planning scheme of the power distribution system. The robustness of the current planning scheme is quantitatively verified based on the boundary pressure test results, and the planning parameters of the current planning scheme are adjusted based on the verification results to achieve dynamic coordination of the integrated planning and operation of the power distribution system.

8. A device for inferring the source-load operating state of a power distribution system based on causal transfer, characterized in that, The relevant equipment includes: The first construction unit is used to construct a candidate feature pool based on the historical multi-source operation data, meteorological data and economic data of the power distribution system. The multi-source operation data includes power load data, photovoltaic power output data and wind power output data. The filtering unit is used to perform nonlinear causal filtering on the feature factors in the candidate feature pool based on approximate transfer entropy, so as to determine the key feature factors affecting the source load operation state and their corresponding causal strength scores. The second construction unit is used to construct a deep causal gated recurrent memory network and to perform non-randomized initialization of the input gate weights of the deep causal gated recurrent memory network based on the key feature factors and their corresponding causal strength scores, so as to construct a time-series inference model. The simulation unit is used to dynamically simulate the source-load operation trajectory of the power distribution system within a future planning period based on the time-series simulation model.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one piece of program code, which is loaded and executed by a processor to perform the operations of the power distribution system source-load operation state inference method based on causal migration as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The electronic device includes one or more processors and one or more memories, wherein at least one piece of program code is stored in the one or more memories, and the at least one piece of program code is loaded and executed by the one or more processors to implement the source-load operation state inference method of power distribution system based on causal migration as described in any one of claims 1 to 7.

Images